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 set_compound_head(p, page);
480 #ifdef CONFIG_DEBUG_PAGEALLOC
481 unsigned int _debug_guardpage_minorder;
482 bool _debug_pagealloc_enabled __read_mostly;
483 bool _debug_guardpage_enabled __read_mostly;
485 static int __init early_debug_pagealloc(char *buf)
490 if (strcmp(buf, "on") == 0)
491 _debug_pagealloc_enabled = true;
495 early_param("debug_pagealloc", early_debug_pagealloc);
497 static bool need_debug_guardpage(void)
499 /* If we don't use debug_pagealloc, we don't need guard page */
500 if (!debug_pagealloc_enabled())
506 static void init_debug_guardpage(void)
508 if (!debug_pagealloc_enabled())
511 _debug_guardpage_enabled = true;
514 struct page_ext_operations debug_guardpage_ops = {
515 .need = need_debug_guardpage,
516 .init = init_debug_guardpage,
519 static int __init debug_guardpage_minorder_setup(char *buf)
523 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
524 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
527 _debug_guardpage_minorder = res;
528 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
531 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
533 static inline void set_page_guard(struct zone *zone, struct page *page,
534 unsigned int order, int migratetype)
536 struct page_ext *page_ext;
538 if (!debug_guardpage_enabled())
541 page_ext = lookup_page_ext(page);
542 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
544 INIT_LIST_HEAD(&page->lru);
545 set_page_private(page, order);
546 /* Guard pages are not available for any usage */
547 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
550 static inline void clear_page_guard(struct zone *zone, struct page *page,
551 unsigned int order, int migratetype)
553 struct page_ext *page_ext;
555 if (!debug_guardpage_enabled())
558 page_ext = lookup_page_ext(page);
559 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
561 set_page_private(page, 0);
562 if (!is_migrate_isolate(migratetype))
563 __mod_zone_freepage_state(zone, (1 << order), migratetype);
566 struct page_ext_operations debug_guardpage_ops = { NULL, };
567 static inline void set_page_guard(struct zone *zone, struct page *page,
568 unsigned int order, int migratetype) {}
569 static inline void clear_page_guard(struct zone *zone, struct page *page,
570 unsigned int order, int migratetype) {}
573 static inline void set_page_order(struct page *page, unsigned int order)
575 set_page_private(page, order);
576 __SetPageBuddy(page);
579 static inline void rmv_page_order(struct page *page)
581 __ClearPageBuddy(page);
582 set_page_private(page, 0);
586 * This function checks whether a page is free && is the buddy
587 * we can do coalesce a page and its buddy if
588 * (a) the buddy is not in a hole &&
589 * (b) the buddy is in the buddy system &&
590 * (c) a page and its buddy have the same order &&
591 * (d) a page and its buddy are in the same zone.
593 * For recording whether a page is in the buddy system, we set ->_mapcount
594 * PAGE_BUDDY_MAPCOUNT_VALUE.
595 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
596 * serialized by zone->lock.
598 * For recording page's order, we use page_private(page).
600 static inline int page_is_buddy(struct page *page, struct page *buddy,
603 if (!pfn_valid_within(page_to_pfn(buddy)))
606 if (page_is_guard(buddy) && page_order(buddy) == order) {
607 if (page_zone_id(page) != page_zone_id(buddy))
610 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
615 if (PageBuddy(buddy) && page_order(buddy) == order) {
617 * zone check is done late to avoid uselessly
618 * calculating zone/node ids for pages that could
621 if (page_zone_id(page) != page_zone_id(buddy))
624 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
632 * Freeing function for a buddy system allocator.
634 * The concept of a buddy system is to maintain direct-mapped table
635 * (containing bit values) for memory blocks of various "orders".
636 * The bottom level table contains the map for the smallest allocatable
637 * units of memory (here, pages), and each level above it describes
638 * pairs of units from the levels below, hence, "buddies".
639 * At a high level, all that happens here is marking the table entry
640 * at the bottom level available, and propagating the changes upward
641 * as necessary, plus some accounting needed to play nicely with other
642 * parts of the VM system.
643 * At each level, we keep a list of pages, which are heads of continuous
644 * free pages of length of (1 << order) and marked with _mapcount
645 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
647 * So when we are allocating or freeing one, we can derive the state of the
648 * other. That is, if we allocate a small block, and both were
649 * free, the remainder of the region must be split into blocks.
650 * If a block is freed, and its buddy is also free, then this
651 * triggers coalescing into a block of larger size.
656 static inline void __free_one_page(struct page *page,
658 struct zone *zone, unsigned int order,
661 unsigned long page_idx;
662 unsigned long combined_idx;
663 unsigned long uninitialized_var(buddy_idx);
665 unsigned int max_order = MAX_ORDER;
667 VM_BUG_ON(!zone_is_initialized(zone));
668 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
670 VM_BUG_ON(migratetype == -1);
671 if (is_migrate_isolate(migratetype)) {
673 * We restrict max order of merging to prevent merge
674 * between freepages on isolate pageblock and normal
675 * pageblock. Without this, pageblock isolation
676 * could cause incorrect freepage accounting.
678 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
680 __mod_zone_freepage_state(zone, 1 << order, migratetype);
683 page_idx = pfn & ((1 << max_order) - 1);
685 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
686 VM_BUG_ON_PAGE(bad_range(zone, page), page);
688 while (order < max_order - 1) {
689 buddy_idx = __find_buddy_index(page_idx, order);
690 buddy = page + (buddy_idx - page_idx);
691 if (!page_is_buddy(page, buddy, order))
694 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
695 * merge with it and move up one order.
697 if (page_is_guard(buddy)) {
698 clear_page_guard(zone, buddy, order, migratetype);
700 list_del(&buddy->lru);
701 zone->free_area[order].nr_free--;
702 rmv_page_order(buddy);
704 combined_idx = buddy_idx & page_idx;
705 page = page + (combined_idx - page_idx);
706 page_idx = combined_idx;
709 set_page_order(page, order);
712 * If this is not the largest possible page, check if the buddy
713 * of the next-highest order is free. If it is, it's possible
714 * that pages are being freed that will coalesce soon. In case,
715 * that is happening, add the free page to the tail of the list
716 * so it's less likely to be used soon and more likely to be merged
717 * as a higher order page
719 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
720 struct page *higher_page, *higher_buddy;
721 combined_idx = buddy_idx & page_idx;
722 higher_page = page + (combined_idx - page_idx);
723 buddy_idx = __find_buddy_index(combined_idx, order + 1);
724 higher_buddy = higher_page + (buddy_idx - combined_idx);
725 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
726 list_add_tail(&page->lru,
727 &zone->free_area[order].free_list[migratetype]);
732 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
734 zone->free_area[order].nr_free++;
737 static inline int free_pages_check(struct page *page)
739 const char *bad_reason = NULL;
740 unsigned long bad_flags = 0;
742 if (unlikely(page_mapcount(page)))
743 bad_reason = "nonzero mapcount";
744 if (unlikely(page->mapping != NULL))
745 bad_reason = "non-NULL mapping";
746 if (unlikely(atomic_read(&page->_count) != 0))
747 bad_reason = "nonzero _count";
748 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
749 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
750 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
753 if (unlikely(page->mem_cgroup))
754 bad_reason = "page still charged to cgroup";
756 if (unlikely(bad_reason)) {
757 bad_page(page, bad_reason, bad_flags);
760 page_cpupid_reset_last(page);
761 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
762 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
767 * Frees a number of pages from the PCP lists
768 * Assumes all pages on list are in same zone, and of same order.
769 * count is the number of pages to free.
771 * If the zone was previously in an "all pages pinned" state then look to
772 * see if this freeing clears that state.
774 * And clear the zone's pages_scanned counter, to hold off the "all pages are
775 * pinned" detection logic.
777 static void free_pcppages_bulk(struct zone *zone, int count,
778 struct per_cpu_pages *pcp)
783 unsigned long nr_scanned;
785 spin_lock(&zone->lock);
786 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
788 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
792 struct list_head *list;
795 * Remove pages from lists in a round-robin fashion. A
796 * batch_free count is maintained that is incremented when an
797 * empty list is encountered. This is so more pages are freed
798 * off fuller lists instead of spinning excessively around empty
803 if (++migratetype == MIGRATE_PCPTYPES)
805 list = &pcp->lists[migratetype];
806 } while (list_empty(list));
808 /* This is the only non-empty list. Free them all. */
809 if (batch_free == MIGRATE_PCPTYPES)
810 batch_free = to_free;
813 int mt; /* migratetype of the to-be-freed page */
815 page = list_entry(list->prev, struct page, lru);
816 /* must delete as __free_one_page list manipulates */
817 list_del(&page->lru);
819 mt = get_pcppage_migratetype(page);
820 /* MIGRATE_ISOLATE page should not go to pcplists */
821 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
822 /* Pageblock could have been isolated meanwhile */
823 if (unlikely(has_isolate_pageblock(zone)))
824 mt = get_pageblock_migratetype(page);
826 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
827 trace_mm_page_pcpu_drain(page, 0, mt);
828 } while (--to_free && --batch_free && !list_empty(list));
830 spin_unlock(&zone->lock);
833 static void free_one_page(struct zone *zone,
834 struct page *page, unsigned long pfn,
838 unsigned long nr_scanned;
839 spin_lock(&zone->lock);
840 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
842 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
844 if (unlikely(has_isolate_pageblock(zone) ||
845 is_migrate_isolate(migratetype))) {
846 migratetype = get_pfnblock_migratetype(page, pfn);
848 __free_one_page(page, pfn, zone, order, migratetype);
849 spin_unlock(&zone->lock);
852 static int free_tail_pages_check(struct page *head_page, struct page *page)
857 * We rely page->lru.next never has bit 0 set, unless the page
858 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
860 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
862 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
866 if (unlikely(!PageTail(page))) {
867 bad_page(page, "PageTail not set", 0);
870 if (unlikely(compound_head(page) != head_page)) {
871 bad_page(page, "compound_head not consistent", 0);
876 clear_compound_head(page);
880 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
881 unsigned long zone, int nid)
883 set_page_links(page, zone, nid, pfn);
884 init_page_count(page);
885 page_mapcount_reset(page);
886 page_cpupid_reset_last(page);
888 INIT_LIST_HEAD(&page->lru);
889 #ifdef WANT_PAGE_VIRTUAL
890 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
891 if (!is_highmem_idx(zone))
892 set_page_address(page, __va(pfn << PAGE_SHIFT));
896 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
899 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
902 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
903 static void init_reserved_page(unsigned long pfn)
908 if (!early_page_uninitialised(pfn))
911 nid = early_pfn_to_nid(pfn);
912 pgdat = NODE_DATA(nid);
914 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
915 struct zone *zone = &pgdat->node_zones[zid];
917 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
920 __init_single_pfn(pfn, zid, nid);
923 static inline void init_reserved_page(unsigned long pfn)
926 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
929 * Initialised pages do not have PageReserved set. This function is
930 * called for each range allocated by the bootmem allocator and
931 * marks the pages PageReserved. The remaining valid pages are later
932 * sent to the buddy page allocator.
934 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
936 unsigned long start_pfn = PFN_DOWN(start);
937 unsigned long end_pfn = PFN_UP(end);
939 for (; start_pfn < end_pfn; start_pfn++) {
940 if (pfn_valid(start_pfn)) {
941 struct page *page = pfn_to_page(start_pfn);
943 init_reserved_page(start_pfn);
945 /* Avoid false-positive PageTail() */
946 INIT_LIST_HEAD(&page->lru);
948 SetPageReserved(page);
953 static bool free_pages_prepare(struct page *page, unsigned int order)
955 bool compound = PageCompound(page);
958 VM_BUG_ON_PAGE(PageTail(page), page);
959 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
961 trace_mm_page_free(page, order);
962 kmemcheck_free_shadow(page, order);
963 kasan_free_pages(page, order);
966 page->mapping = NULL;
967 bad += free_pages_check(page);
968 for (i = 1; i < (1 << order); i++) {
970 bad += free_tail_pages_check(page, page + i);
971 bad += free_pages_check(page + i);
976 reset_page_owner(page, order);
978 if (!PageHighMem(page)) {
979 debug_check_no_locks_freed(page_address(page),
981 debug_check_no_obj_freed(page_address(page),
984 arch_free_page(page, order);
985 kernel_map_pages(page, 1 << order, 0);
990 static void __free_pages_ok(struct page *page, unsigned int order)
994 unsigned long pfn = page_to_pfn(page);
996 if (!free_pages_prepare(page, order))
999 migratetype = get_pfnblock_migratetype(page, pfn);
1000 local_irq_save(flags);
1001 __count_vm_events(PGFREE, 1 << order);
1002 free_one_page(page_zone(page), page, pfn, order, migratetype);
1003 local_irq_restore(flags);
1006 static void __init __free_pages_boot_core(struct page *page,
1007 unsigned long pfn, unsigned int order)
1009 unsigned int nr_pages = 1 << order;
1010 struct page *p = page;
1014 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1016 __ClearPageReserved(p);
1017 set_page_count(p, 0);
1019 __ClearPageReserved(p);
1020 set_page_count(p, 0);
1022 page_zone(page)->managed_pages += nr_pages;
1023 set_page_refcounted(page);
1024 __free_pages(page, order);
1027 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1028 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1030 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1032 int __meminit early_pfn_to_nid(unsigned long pfn)
1034 static DEFINE_SPINLOCK(early_pfn_lock);
1037 spin_lock(&early_pfn_lock);
1038 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1041 spin_unlock(&early_pfn_lock);
1047 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1048 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1049 struct mminit_pfnnid_cache *state)
1053 nid = __early_pfn_to_nid(pfn, state);
1054 if (nid >= 0 && nid != node)
1059 /* Only safe to use early in boot when initialisation is single-threaded */
1060 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1062 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1067 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1071 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1072 struct mminit_pfnnid_cache *state)
1079 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1082 if (early_page_uninitialised(pfn))
1084 return __free_pages_boot_core(page, pfn, order);
1087 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1088 static void __init deferred_free_range(struct page *page,
1089 unsigned long pfn, int nr_pages)
1096 /* Free a large naturally-aligned chunk if possible */
1097 if (nr_pages == MAX_ORDER_NR_PAGES &&
1098 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1099 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1100 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1104 for (i = 0; i < nr_pages; i++, page++, pfn++)
1105 __free_pages_boot_core(page, pfn, 0);
1108 /* Completion tracking for deferred_init_memmap() threads */
1109 static atomic_t pgdat_init_n_undone __initdata;
1110 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1112 static inline void __init pgdat_init_report_one_done(void)
1114 if (atomic_dec_and_test(&pgdat_init_n_undone))
1115 complete(&pgdat_init_all_done_comp);
1118 /* Initialise remaining memory on a node */
1119 static int __init deferred_init_memmap(void *data)
1121 pg_data_t *pgdat = data;
1122 int nid = pgdat->node_id;
1123 struct mminit_pfnnid_cache nid_init_state = { };
1124 unsigned long start = jiffies;
1125 unsigned long nr_pages = 0;
1126 unsigned long walk_start, walk_end;
1129 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1130 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1132 if (first_init_pfn == ULONG_MAX) {
1133 pgdat_init_report_one_done();
1137 /* Bind memory initialisation thread to a local node if possible */
1138 if (!cpumask_empty(cpumask))
1139 set_cpus_allowed_ptr(current, cpumask);
1141 /* Sanity check boundaries */
1142 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1143 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1144 pgdat->first_deferred_pfn = ULONG_MAX;
1146 /* Only the highest zone is deferred so find it */
1147 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1148 zone = pgdat->node_zones + zid;
1149 if (first_init_pfn < zone_end_pfn(zone))
1153 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1154 unsigned long pfn, end_pfn;
1155 struct page *page = NULL;
1156 struct page *free_base_page = NULL;
1157 unsigned long free_base_pfn = 0;
1160 end_pfn = min(walk_end, zone_end_pfn(zone));
1161 pfn = first_init_pfn;
1162 if (pfn < walk_start)
1164 if (pfn < zone->zone_start_pfn)
1165 pfn = zone->zone_start_pfn;
1167 for (; pfn < end_pfn; pfn++) {
1168 if (!pfn_valid_within(pfn))
1172 * Ensure pfn_valid is checked every
1173 * MAX_ORDER_NR_PAGES for memory holes
1175 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1176 if (!pfn_valid(pfn)) {
1182 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1187 /* Minimise pfn page lookups and scheduler checks */
1188 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1191 nr_pages += nr_to_free;
1192 deferred_free_range(free_base_page,
1193 free_base_pfn, nr_to_free);
1194 free_base_page = NULL;
1195 free_base_pfn = nr_to_free = 0;
1197 page = pfn_to_page(pfn);
1202 VM_BUG_ON(page_zone(page) != zone);
1206 __init_single_page(page, pfn, zid, nid);
1207 if (!free_base_page) {
1208 free_base_page = page;
1209 free_base_pfn = pfn;
1214 /* Where possible, batch up pages for a single free */
1217 /* Free the current block of pages to allocator */
1218 nr_pages += nr_to_free;
1219 deferred_free_range(free_base_page, free_base_pfn,
1221 free_base_page = NULL;
1222 free_base_pfn = nr_to_free = 0;
1225 first_init_pfn = max(end_pfn, first_init_pfn);
1228 /* Sanity check that the next zone really is unpopulated */
1229 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1231 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1232 jiffies_to_msecs(jiffies - start));
1234 pgdat_init_report_one_done();
1238 void __init page_alloc_init_late(void)
1242 /* There will be num_node_state(N_MEMORY) threads */
1243 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1244 for_each_node_state(nid, N_MEMORY) {
1245 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1248 /* Block until all are initialised */
1249 wait_for_completion(&pgdat_init_all_done_comp);
1251 /* Reinit limits that are based on free pages after the kernel is up */
1252 files_maxfiles_init();
1254 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1257 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1258 void __init init_cma_reserved_pageblock(struct page *page)
1260 unsigned i = pageblock_nr_pages;
1261 struct page *p = page;
1264 __ClearPageReserved(p);
1265 set_page_count(p, 0);
1268 set_pageblock_migratetype(page, MIGRATE_CMA);
1270 if (pageblock_order >= MAX_ORDER) {
1271 i = pageblock_nr_pages;
1274 set_page_refcounted(p);
1275 __free_pages(p, MAX_ORDER - 1);
1276 p += MAX_ORDER_NR_PAGES;
1277 } while (i -= MAX_ORDER_NR_PAGES);
1279 set_page_refcounted(page);
1280 __free_pages(page, pageblock_order);
1283 adjust_managed_page_count(page, pageblock_nr_pages);
1288 * The order of subdivision here is critical for the IO subsystem.
1289 * Please do not alter this order without good reasons and regression
1290 * testing. Specifically, as large blocks of memory are subdivided,
1291 * the order in which smaller blocks are delivered depends on the order
1292 * they're subdivided in this function. This is the primary factor
1293 * influencing the order in which pages are delivered to the IO
1294 * subsystem according to empirical testing, and this is also justified
1295 * by considering the behavior of a buddy system containing a single
1296 * large block of memory acted on by a series of small allocations.
1297 * This behavior is a critical factor in sglist merging's success.
1301 static inline void expand(struct zone *zone, struct page *page,
1302 int low, int high, struct free_area *area,
1305 unsigned long size = 1 << high;
1307 while (high > low) {
1311 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1313 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1314 debug_guardpage_enabled() &&
1315 high < debug_guardpage_minorder()) {
1317 * Mark as guard pages (or page), that will allow to
1318 * merge back to allocator when buddy will be freed.
1319 * Corresponding page table entries will not be touched,
1320 * pages will stay not present in virtual address space
1322 set_page_guard(zone, &page[size], high, migratetype);
1325 list_add(&page[size].lru, &area->free_list[migratetype]);
1327 set_page_order(&page[size], high);
1332 * This page is about to be returned from the page allocator
1334 static inline int check_new_page(struct page *page)
1336 const char *bad_reason = NULL;
1337 unsigned long bad_flags = 0;
1339 if (unlikely(page_mapcount(page)))
1340 bad_reason = "nonzero mapcount";
1341 if (unlikely(page->mapping != NULL))
1342 bad_reason = "non-NULL mapping";
1343 if (unlikely(atomic_read(&page->_count) != 0))
1344 bad_reason = "nonzero _count";
1345 if (unlikely(page->flags & __PG_HWPOISON)) {
1346 bad_reason = "HWPoisoned (hardware-corrupted)";
1347 bad_flags = __PG_HWPOISON;
1349 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1350 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1351 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1354 if (unlikely(page->mem_cgroup))
1355 bad_reason = "page still charged to cgroup";
1357 if (unlikely(bad_reason)) {
1358 bad_page(page, bad_reason, bad_flags);
1364 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1369 for (i = 0; i < (1 << order); i++) {
1370 struct page *p = page + i;
1371 if (unlikely(check_new_page(p)))
1375 set_page_private(page, 0);
1376 set_page_refcounted(page);
1378 arch_alloc_page(page, order);
1379 kernel_map_pages(page, 1 << order, 1);
1380 kasan_alloc_pages(page, order);
1382 if (gfp_flags & __GFP_ZERO)
1383 for (i = 0; i < (1 << order); i++)
1384 clear_highpage(page + i);
1386 if (order && (gfp_flags & __GFP_COMP))
1387 prep_compound_page(page, order);
1389 set_page_owner(page, order, gfp_flags);
1392 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1393 * allocate the page. The expectation is that the caller is taking
1394 * steps that will free more memory. The caller should avoid the page
1395 * being used for !PFMEMALLOC purposes.
1397 if (alloc_flags & ALLOC_NO_WATERMARKS)
1398 set_page_pfmemalloc(page);
1400 clear_page_pfmemalloc(page);
1406 * Go through the free lists for the given migratetype and remove
1407 * the smallest available page from the freelists
1410 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1413 unsigned int current_order;
1414 struct free_area *area;
1417 /* Find a page of the appropriate size in the preferred list */
1418 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1419 area = &(zone->free_area[current_order]);
1420 if (list_empty(&area->free_list[migratetype]))
1423 page = list_entry(area->free_list[migratetype].next,
1425 list_del(&page->lru);
1426 rmv_page_order(page);
1428 expand(zone, page, order, current_order, area, migratetype);
1429 set_pcppage_migratetype(page, migratetype);
1438 * This array describes the order lists are fallen back to when
1439 * the free lists for the desirable migrate type are depleted
1441 static int fallbacks[MIGRATE_TYPES][4] = {
1442 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1443 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1444 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1446 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1448 #ifdef CONFIG_MEMORY_ISOLATION
1449 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1454 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1457 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1460 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1461 unsigned int order) { return NULL; }
1465 * Move the free pages in a range to the free lists of the requested type.
1466 * Note that start_page and end_pages are not aligned on a pageblock
1467 * boundary. If alignment is required, use move_freepages_block()
1469 int move_freepages(struct zone *zone,
1470 struct page *start_page, struct page *end_page,
1475 int pages_moved = 0;
1477 #ifndef CONFIG_HOLES_IN_ZONE
1479 * page_zone is not safe to call in this context when
1480 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1481 * anyway as we check zone boundaries in move_freepages_block().
1482 * Remove at a later date when no bug reports exist related to
1483 * grouping pages by mobility
1485 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1488 for (page = start_page; page <= end_page;) {
1489 /* Make sure we are not inadvertently changing nodes */
1490 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1492 if (!pfn_valid_within(page_to_pfn(page))) {
1497 if (!PageBuddy(page)) {
1502 order = page_order(page);
1503 list_move(&page->lru,
1504 &zone->free_area[order].free_list[migratetype]);
1506 pages_moved += 1 << order;
1512 int move_freepages_block(struct zone *zone, struct page *page,
1515 unsigned long start_pfn, end_pfn;
1516 struct page *start_page, *end_page;
1518 start_pfn = page_to_pfn(page);
1519 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1520 start_page = pfn_to_page(start_pfn);
1521 end_page = start_page + pageblock_nr_pages - 1;
1522 end_pfn = start_pfn + pageblock_nr_pages - 1;
1524 /* Do not cross zone boundaries */
1525 if (!zone_spans_pfn(zone, start_pfn))
1527 if (!zone_spans_pfn(zone, end_pfn))
1530 return move_freepages(zone, start_page, end_page, migratetype);
1533 static void change_pageblock_range(struct page *pageblock_page,
1534 int start_order, int migratetype)
1536 int nr_pageblocks = 1 << (start_order - pageblock_order);
1538 while (nr_pageblocks--) {
1539 set_pageblock_migratetype(pageblock_page, migratetype);
1540 pageblock_page += pageblock_nr_pages;
1545 * When we are falling back to another migratetype during allocation, try to
1546 * steal extra free pages from the same pageblocks to satisfy further
1547 * allocations, instead of polluting multiple pageblocks.
1549 * If we are stealing a relatively large buddy page, it is likely there will
1550 * be more free pages in the pageblock, so try to steal them all. For
1551 * reclaimable and unmovable allocations, we steal regardless of page size,
1552 * as fragmentation caused by those allocations polluting movable pageblocks
1553 * is worse than movable allocations stealing from unmovable and reclaimable
1556 static bool can_steal_fallback(unsigned int order, int start_mt)
1559 * Leaving this order check is intended, although there is
1560 * relaxed order check in next check. The reason is that
1561 * we can actually steal whole pageblock if this condition met,
1562 * but, below check doesn't guarantee it and that is just heuristic
1563 * so could be changed anytime.
1565 if (order >= pageblock_order)
1568 if (order >= pageblock_order / 2 ||
1569 start_mt == MIGRATE_RECLAIMABLE ||
1570 start_mt == MIGRATE_UNMOVABLE ||
1571 page_group_by_mobility_disabled)
1578 * This function implements actual steal behaviour. If order is large enough,
1579 * we can steal whole pageblock. If not, we first move freepages in this
1580 * pageblock and check whether half of pages are moved or not. If half of
1581 * pages are moved, we can change migratetype of pageblock and permanently
1582 * use it's pages as requested migratetype in the future.
1584 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1587 unsigned int current_order = page_order(page);
1590 /* Take ownership for orders >= pageblock_order */
1591 if (current_order >= pageblock_order) {
1592 change_pageblock_range(page, current_order, start_type);
1596 pages = move_freepages_block(zone, page, start_type);
1598 /* Claim the whole block if over half of it is free */
1599 if (pages >= (1 << (pageblock_order-1)) ||
1600 page_group_by_mobility_disabled)
1601 set_pageblock_migratetype(page, start_type);
1605 * Check whether there is a suitable fallback freepage with requested order.
1606 * If only_stealable is true, this function returns fallback_mt only if
1607 * we can steal other freepages all together. This would help to reduce
1608 * fragmentation due to mixed migratetype pages in one pageblock.
1610 int find_suitable_fallback(struct free_area *area, unsigned int order,
1611 int migratetype, bool only_stealable, bool *can_steal)
1616 if (area->nr_free == 0)
1621 fallback_mt = fallbacks[migratetype][i];
1622 if (fallback_mt == MIGRATE_TYPES)
1625 if (list_empty(&area->free_list[fallback_mt]))
1628 if (can_steal_fallback(order, migratetype))
1631 if (!only_stealable)
1642 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1643 * there are no empty page blocks that contain a page with a suitable order
1645 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1646 unsigned int alloc_order)
1649 unsigned long max_managed, flags;
1652 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1653 * Check is race-prone but harmless.
1655 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1656 if (zone->nr_reserved_highatomic >= max_managed)
1659 spin_lock_irqsave(&zone->lock, flags);
1661 /* Recheck the nr_reserved_highatomic limit under the lock */
1662 if (zone->nr_reserved_highatomic >= max_managed)
1666 mt = get_pageblock_migratetype(page);
1667 if (mt != MIGRATE_HIGHATOMIC &&
1668 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1669 zone->nr_reserved_highatomic += pageblock_nr_pages;
1670 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1671 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1675 spin_unlock_irqrestore(&zone->lock, flags);
1679 * Used when an allocation is about to fail under memory pressure. This
1680 * potentially hurts the reliability of high-order allocations when under
1681 * intense memory pressure but failed atomic allocations should be easier
1682 * to recover from than an OOM.
1684 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1686 struct zonelist *zonelist = ac->zonelist;
1687 unsigned long flags;
1693 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1695 /* Preserve at least one pageblock */
1696 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1699 spin_lock_irqsave(&zone->lock, flags);
1700 for (order = 0; order < MAX_ORDER; order++) {
1701 struct free_area *area = &(zone->free_area[order]);
1703 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1706 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1710 * It should never happen but changes to locking could
1711 * inadvertently allow a per-cpu drain to add pages
1712 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1713 * and watch for underflows.
1715 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1716 zone->nr_reserved_highatomic);
1719 * Convert to ac->migratetype and avoid the normal
1720 * pageblock stealing heuristics. Minimally, the caller
1721 * is doing the work and needs the pages. More
1722 * importantly, if the block was always converted to
1723 * MIGRATE_UNMOVABLE or another type then the number
1724 * of pageblocks that cannot be completely freed
1727 set_pageblock_migratetype(page, ac->migratetype);
1728 move_freepages_block(zone, page, ac->migratetype);
1729 spin_unlock_irqrestore(&zone->lock, flags);
1732 spin_unlock_irqrestore(&zone->lock, flags);
1736 /* Remove an element from the buddy allocator from the fallback list */
1737 static inline struct page *
1738 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1740 struct free_area *area;
1741 unsigned int current_order;
1746 /* Find the largest possible block of pages in the other list */
1747 for (current_order = MAX_ORDER-1;
1748 current_order >= order && current_order <= MAX_ORDER-1;
1750 area = &(zone->free_area[current_order]);
1751 fallback_mt = find_suitable_fallback(area, current_order,
1752 start_migratetype, false, &can_steal);
1753 if (fallback_mt == -1)
1756 page = list_entry(area->free_list[fallback_mt].next,
1759 steal_suitable_fallback(zone, page, start_migratetype);
1761 /* Remove the page from the freelists */
1763 list_del(&page->lru);
1764 rmv_page_order(page);
1766 expand(zone, page, order, current_order, area,
1769 * The pcppage_migratetype may differ from pageblock's
1770 * migratetype depending on the decisions in
1771 * find_suitable_fallback(). This is OK as long as it does not
1772 * differ for MIGRATE_CMA pageblocks. Those can be used as
1773 * fallback only via special __rmqueue_cma_fallback() function
1775 set_pcppage_migratetype(page, start_migratetype);
1777 trace_mm_page_alloc_extfrag(page, order, current_order,
1778 start_migratetype, fallback_mt);
1787 * Do the hard work of removing an element from the buddy allocator.
1788 * Call me with the zone->lock already held.
1790 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1791 int migratetype, gfp_t gfp_flags)
1795 page = __rmqueue_smallest(zone, order, migratetype);
1796 if (unlikely(!page)) {
1797 if (migratetype == MIGRATE_MOVABLE)
1798 page = __rmqueue_cma_fallback(zone, order);
1801 page = __rmqueue_fallback(zone, order, migratetype);
1804 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1809 * Obtain a specified number of elements from the buddy allocator, all under
1810 * a single hold of the lock, for efficiency. Add them to the supplied list.
1811 * Returns the number of new pages which were placed at *list.
1813 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1814 unsigned long count, struct list_head *list,
1815 int migratetype, bool cold)
1819 spin_lock(&zone->lock);
1820 for (i = 0; i < count; ++i) {
1821 struct page *page = __rmqueue(zone, order, migratetype, 0);
1822 if (unlikely(page == NULL))
1826 * Split buddy pages returned by expand() are received here
1827 * in physical page order. The page is added to the callers and
1828 * list and the list head then moves forward. From the callers
1829 * perspective, the linked list is ordered by page number in
1830 * some conditions. This is useful for IO devices that can
1831 * merge IO requests if the physical pages are ordered
1835 list_add(&page->lru, list);
1837 list_add_tail(&page->lru, list);
1839 if (is_migrate_cma(get_pcppage_migratetype(page)))
1840 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1843 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1844 spin_unlock(&zone->lock);
1850 * Called from the vmstat counter updater to drain pagesets of this
1851 * currently executing processor on remote nodes after they have
1854 * Note that this function must be called with the thread pinned to
1855 * a single processor.
1857 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1859 unsigned long flags;
1860 int to_drain, batch;
1862 local_irq_save(flags);
1863 batch = READ_ONCE(pcp->batch);
1864 to_drain = min(pcp->count, batch);
1866 free_pcppages_bulk(zone, to_drain, pcp);
1867 pcp->count -= to_drain;
1869 local_irq_restore(flags);
1874 * Drain pcplists of the indicated processor and zone.
1876 * The processor must either be the current processor and the
1877 * thread pinned to the current processor or a processor that
1880 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1882 unsigned long flags;
1883 struct per_cpu_pageset *pset;
1884 struct per_cpu_pages *pcp;
1886 local_irq_save(flags);
1887 pset = per_cpu_ptr(zone->pageset, cpu);
1891 free_pcppages_bulk(zone, pcp->count, pcp);
1894 local_irq_restore(flags);
1898 * Drain pcplists of all zones on the indicated processor.
1900 * The processor must either be the current processor and the
1901 * thread pinned to the current processor or a processor that
1904 static void drain_pages(unsigned int cpu)
1908 for_each_populated_zone(zone) {
1909 drain_pages_zone(cpu, zone);
1914 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1916 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1917 * the single zone's pages.
1919 void drain_local_pages(struct zone *zone)
1921 int cpu = smp_processor_id();
1924 drain_pages_zone(cpu, zone);
1930 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1932 * When zone parameter is non-NULL, spill just the single zone's pages.
1934 * Note that this code is protected against sending an IPI to an offline
1935 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1936 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1937 * nothing keeps CPUs from showing up after we populated the cpumask and
1938 * before the call to on_each_cpu_mask().
1940 void drain_all_pages(struct zone *zone)
1945 * Allocate in the BSS so we wont require allocation in
1946 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1948 static cpumask_t cpus_with_pcps;
1951 * We don't care about racing with CPU hotplug event
1952 * as offline notification will cause the notified
1953 * cpu to drain that CPU pcps and on_each_cpu_mask
1954 * disables preemption as part of its processing
1956 for_each_online_cpu(cpu) {
1957 struct per_cpu_pageset *pcp;
1959 bool has_pcps = false;
1962 pcp = per_cpu_ptr(zone->pageset, cpu);
1966 for_each_populated_zone(z) {
1967 pcp = per_cpu_ptr(z->pageset, cpu);
1968 if (pcp->pcp.count) {
1976 cpumask_set_cpu(cpu, &cpus_with_pcps);
1978 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1980 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1984 #ifdef CONFIG_HIBERNATION
1986 void mark_free_pages(struct zone *zone)
1988 unsigned long pfn, max_zone_pfn;
1989 unsigned long flags;
1990 unsigned int order, t;
1991 struct list_head *curr;
1993 if (zone_is_empty(zone))
1996 spin_lock_irqsave(&zone->lock, flags);
1998 max_zone_pfn = zone_end_pfn(zone);
1999 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2000 if (pfn_valid(pfn)) {
2001 struct page *page = pfn_to_page(pfn);
2003 if (!swsusp_page_is_forbidden(page))
2004 swsusp_unset_page_free(page);
2007 for_each_migratetype_order(order, t) {
2008 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2011 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2012 for (i = 0; i < (1UL << order); i++)
2013 swsusp_set_page_free(pfn_to_page(pfn + i));
2016 spin_unlock_irqrestore(&zone->lock, flags);
2018 #endif /* CONFIG_PM */
2021 * Free a 0-order page
2022 * cold == true ? free a cold page : free a hot page
2024 void free_hot_cold_page(struct page *page, bool cold)
2026 struct zone *zone = page_zone(page);
2027 struct per_cpu_pages *pcp;
2028 unsigned long flags;
2029 unsigned long pfn = page_to_pfn(page);
2032 if (!free_pages_prepare(page, 0))
2035 migratetype = get_pfnblock_migratetype(page, pfn);
2036 set_pcppage_migratetype(page, migratetype);
2037 local_irq_save(flags);
2038 __count_vm_event(PGFREE);
2041 * We only track unmovable, reclaimable and movable on pcp lists.
2042 * Free ISOLATE pages back to the allocator because they are being
2043 * offlined but treat RESERVE as movable pages so we can get those
2044 * areas back if necessary. Otherwise, we may have to free
2045 * excessively into the page allocator
2047 if (migratetype >= MIGRATE_PCPTYPES) {
2048 if (unlikely(is_migrate_isolate(migratetype))) {
2049 free_one_page(zone, page, pfn, 0, migratetype);
2052 migratetype = MIGRATE_MOVABLE;
2055 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2057 list_add(&page->lru, &pcp->lists[migratetype]);
2059 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2061 if (pcp->count >= pcp->high) {
2062 unsigned long batch = READ_ONCE(pcp->batch);
2063 free_pcppages_bulk(zone, batch, pcp);
2064 pcp->count -= batch;
2068 local_irq_restore(flags);
2072 * Free a list of 0-order pages
2074 void free_hot_cold_page_list(struct list_head *list, bool cold)
2076 struct page *page, *next;
2078 list_for_each_entry_safe(page, next, list, lru) {
2079 trace_mm_page_free_batched(page, cold);
2080 free_hot_cold_page(page, cold);
2085 * split_page takes a non-compound higher-order page, and splits it into
2086 * n (1<<order) sub-pages: page[0..n]
2087 * Each sub-page must be freed individually.
2089 * Note: this is probably too low level an operation for use in drivers.
2090 * Please consult with lkml before using this in your driver.
2092 void split_page(struct page *page, unsigned int order)
2097 VM_BUG_ON_PAGE(PageCompound(page), page);
2098 VM_BUG_ON_PAGE(!page_count(page), page);
2100 #ifdef CONFIG_KMEMCHECK
2102 * Split shadow pages too, because free(page[0]) would
2103 * otherwise free the whole shadow.
2105 if (kmemcheck_page_is_tracked(page))
2106 split_page(virt_to_page(page[0].shadow), order);
2109 gfp_mask = get_page_owner_gfp(page);
2110 set_page_owner(page, 0, gfp_mask);
2111 for (i = 1; i < (1 << order); i++) {
2112 set_page_refcounted(page + i);
2113 set_page_owner(page + i, 0, gfp_mask);
2116 EXPORT_SYMBOL_GPL(split_page);
2118 int __isolate_free_page(struct page *page, unsigned int order)
2120 unsigned long watermark;
2124 BUG_ON(!PageBuddy(page));
2126 zone = page_zone(page);
2127 mt = get_pageblock_migratetype(page);
2129 if (!is_migrate_isolate(mt)) {
2130 /* Obey watermarks as if the page was being allocated */
2131 watermark = low_wmark_pages(zone) + (1 << order);
2132 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2135 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2138 /* Remove page from free list */
2139 list_del(&page->lru);
2140 zone->free_area[order].nr_free--;
2141 rmv_page_order(page);
2143 set_page_owner(page, order, __GFP_MOVABLE);
2145 /* Set the pageblock if the isolated page is at least a pageblock */
2146 if (order >= pageblock_order - 1) {
2147 struct page *endpage = page + (1 << order) - 1;
2148 for (; page < endpage; page += pageblock_nr_pages) {
2149 int mt = get_pageblock_migratetype(page);
2150 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2151 set_pageblock_migratetype(page,
2157 return 1UL << order;
2161 * Similar to split_page except the page is already free. As this is only
2162 * being used for migration, the migratetype of the block also changes.
2163 * As this is called with interrupts disabled, the caller is responsible
2164 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2167 * Note: this is probably too low level an operation for use in drivers.
2168 * Please consult with lkml before using this in your driver.
2170 int split_free_page(struct page *page)
2175 order = page_order(page);
2177 nr_pages = __isolate_free_page(page, order);
2181 /* Split into individual pages */
2182 set_page_refcounted(page);
2183 split_page(page, order);
2188 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2191 struct page *buffered_rmqueue(struct zone *preferred_zone,
2192 struct zone *zone, unsigned int order,
2193 gfp_t gfp_flags, int alloc_flags, int migratetype)
2195 unsigned long flags;
2197 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2199 if (likely(order == 0)) {
2200 struct per_cpu_pages *pcp;
2201 struct list_head *list;
2203 local_irq_save(flags);
2204 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2205 list = &pcp->lists[migratetype];
2206 if (list_empty(list)) {
2207 pcp->count += rmqueue_bulk(zone, 0,
2210 if (unlikely(list_empty(list)))
2215 page = list_entry(list->prev, struct page, lru);
2217 page = list_entry(list->next, struct page, lru);
2219 list_del(&page->lru);
2222 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2224 * __GFP_NOFAIL is not to be used in new code.
2226 * All __GFP_NOFAIL callers should be fixed so that they
2227 * properly detect and handle allocation failures.
2229 * We most definitely don't want callers attempting to
2230 * allocate greater than order-1 page units with
2233 WARN_ON_ONCE(order > 1);
2235 spin_lock_irqsave(&zone->lock, flags);
2238 if (alloc_flags & ALLOC_HARDER) {
2239 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2241 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2244 page = __rmqueue(zone, order, migratetype, gfp_flags);
2245 spin_unlock(&zone->lock);
2248 __mod_zone_freepage_state(zone, -(1 << order),
2249 get_pcppage_migratetype(page));
2252 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2253 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2254 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2255 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2257 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2258 zone_statistics(preferred_zone, zone, gfp_flags);
2259 local_irq_restore(flags);
2261 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2265 local_irq_restore(flags);
2269 #ifdef CONFIG_FAIL_PAGE_ALLOC
2272 struct fault_attr attr;
2274 bool ignore_gfp_highmem;
2275 bool ignore_gfp_reclaim;
2277 } fail_page_alloc = {
2278 .attr = FAULT_ATTR_INITIALIZER,
2279 .ignore_gfp_reclaim = true,
2280 .ignore_gfp_highmem = true,
2284 static int __init setup_fail_page_alloc(char *str)
2286 return setup_fault_attr(&fail_page_alloc.attr, str);
2288 __setup("fail_page_alloc=", setup_fail_page_alloc);
2290 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2292 if (order < fail_page_alloc.min_order)
2294 if (gfp_mask & __GFP_NOFAIL)
2296 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2298 if (fail_page_alloc.ignore_gfp_reclaim &&
2299 (gfp_mask & __GFP_DIRECT_RECLAIM))
2302 return should_fail(&fail_page_alloc.attr, 1 << order);
2305 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2307 static int __init fail_page_alloc_debugfs(void)
2309 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2312 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2313 &fail_page_alloc.attr);
2315 return PTR_ERR(dir);
2317 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2318 &fail_page_alloc.ignore_gfp_reclaim))
2320 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2321 &fail_page_alloc.ignore_gfp_highmem))
2323 if (!debugfs_create_u32("min-order", mode, dir,
2324 &fail_page_alloc.min_order))
2329 debugfs_remove_recursive(dir);
2334 late_initcall(fail_page_alloc_debugfs);
2336 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2338 #else /* CONFIG_FAIL_PAGE_ALLOC */
2340 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2345 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2348 * Return true if free base pages are above 'mark'. For high-order checks it
2349 * will return true of the order-0 watermark is reached and there is at least
2350 * one free page of a suitable size. Checking now avoids taking the zone lock
2351 * to check in the allocation paths if no pages are free.
2353 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2354 unsigned long mark, int classzone_idx, int alloc_flags,
2359 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2361 /* free_pages may go negative - that's OK */
2362 free_pages -= (1 << order) - 1;
2364 if (alloc_flags & ALLOC_HIGH)
2368 * If the caller does not have rights to ALLOC_HARDER then subtract
2369 * the high-atomic reserves. This will over-estimate the size of the
2370 * atomic reserve but it avoids a search.
2372 if (likely(!alloc_harder))
2373 free_pages -= z->nr_reserved_highatomic;
2378 /* If allocation can't use CMA areas don't use free CMA pages */
2379 if (!(alloc_flags & ALLOC_CMA))
2380 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2384 * Check watermarks for an order-0 allocation request. If these
2385 * are not met, then a high-order request also cannot go ahead
2386 * even if a suitable page happened to be free.
2388 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2391 /* If this is an order-0 request then the watermark is fine */
2395 /* For a high-order request, check at least one suitable page is free */
2396 for (o = order; o < MAX_ORDER; o++) {
2397 struct free_area *area = &z->free_area[o];
2406 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2407 if (!list_empty(&area->free_list[mt]))
2412 if ((alloc_flags & ALLOC_CMA) &&
2413 !list_empty(&area->free_list[MIGRATE_CMA])) {
2421 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2422 int classzone_idx, int alloc_flags)
2424 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2425 zone_page_state(z, NR_FREE_PAGES));
2428 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2429 unsigned long mark, int classzone_idx)
2431 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2433 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2434 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2436 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2441 static bool zone_local(struct zone *local_zone, struct zone *zone)
2443 return local_zone->node == zone->node;
2446 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2448 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2451 #else /* CONFIG_NUMA */
2452 static bool zone_local(struct zone *local_zone, struct zone *zone)
2457 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2461 #endif /* CONFIG_NUMA */
2463 static void reset_alloc_batches(struct zone *preferred_zone)
2465 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2468 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2469 high_wmark_pages(zone) - low_wmark_pages(zone) -
2470 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2471 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2472 } while (zone++ != preferred_zone);
2476 * get_page_from_freelist goes through the zonelist trying to allocate
2479 static struct page *
2480 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2481 const struct alloc_context *ac)
2483 struct zonelist *zonelist = ac->zonelist;
2485 struct page *page = NULL;
2487 int nr_fair_skipped = 0;
2488 bool zonelist_rescan;
2491 zonelist_rescan = false;
2494 * Scan zonelist, looking for a zone with enough free.
2495 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2497 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2501 if (cpusets_enabled() &&
2502 (alloc_flags & ALLOC_CPUSET) &&
2503 !cpuset_zone_allowed(zone, gfp_mask))
2506 * Distribute pages in proportion to the individual
2507 * zone size to ensure fair page aging. The zone a
2508 * page was allocated in should have no effect on the
2509 * time the page has in memory before being reclaimed.
2511 if (alloc_flags & ALLOC_FAIR) {
2512 if (!zone_local(ac->preferred_zone, zone))
2514 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2520 * When allocating a page cache page for writing, we
2521 * want to get it from a zone that is within its dirty
2522 * limit, such that no single zone holds more than its
2523 * proportional share of globally allowed dirty pages.
2524 * The dirty limits take into account the zone's
2525 * lowmem reserves and high watermark so that kswapd
2526 * should be able to balance it without having to
2527 * write pages from its LRU list.
2529 * This may look like it could increase pressure on
2530 * lower zones by failing allocations in higher zones
2531 * before they are full. But the pages that do spill
2532 * over are limited as the lower zones are protected
2533 * by this very same mechanism. It should not become
2534 * a practical burden to them.
2536 * XXX: For now, allow allocations to potentially
2537 * exceed the per-zone dirty limit in the slowpath
2538 * (spread_dirty_pages unset) before going into reclaim,
2539 * which is important when on a NUMA setup the allowed
2540 * zones are together not big enough to reach the
2541 * global limit. The proper fix for these situations
2542 * will require awareness of zones in the
2543 * dirty-throttling and the flusher threads.
2545 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2548 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2549 if (!zone_watermark_ok(zone, order, mark,
2550 ac->classzone_idx, alloc_flags)) {
2553 /* Checked here to keep the fast path fast */
2554 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2555 if (alloc_flags & ALLOC_NO_WATERMARKS)
2558 if (zone_reclaim_mode == 0 ||
2559 !zone_allows_reclaim(ac->preferred_zone, zone))
2562 ret = zone_reclaim(zone, gfp_mask, order);
2564 case ZONE_RECLAIM_NOSCAN:
2567 case ZONE_RECLAIM_FULL:
2568 /* scanned but unreclaimable */
2571 /* did we reclaim enough */
2572 if (zone_watermark_ok(zone, order, mark,
2573 ac->classzone_idx, alloc_flags))
2581 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2582 gfp_mask, alloc_flags, ac->migratetype);
2584 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2588 * If this is a high-order atomic allocation then check
2589 * if the pageblock should be reserved for the future
2591 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2592 reserve_highatomic_pageblock(page, zone, order);
2599 * The first pass makes sure allocations are spread fairly within the
2600 * local node. However, the local node might have free pages left
2601 * after the fairness batches are exhausted, and remote zones haven't
2602 * even been considered yet. Try once more without fairness, and
2603 * include remote zones now, before entering the slowpath and waking
2604 * kswapd: prefer spilling to a remote zone over swapping locally.
2606 if (alloc_flags & ALLOC_FAIR) {
2607 alloc_flags &= ~ALLOC_FAIR;
2608 if (nr_fair_skipped) {
2609 zonelist_rescan = true;
2610 reset_alloc_batches(ac->preferred_zone);
2612 if (nr_online_nodes > 1)
2613 zonelist_rescan = true;
2616 if (zonelist_rescan)
2623 * Large machines with many possible nodes should not always dump per-node
2624 * meminfo in irq context.
2626 static inline bool should_suppress_show_mem(void)
2631 ret = in_interrupt();
2636 static DEFINE_RATELIMIT_STATE(nopage_rs,
2637 DEFAULT_RATELIMIT_INTERVAL,
2638 DEFAULT_RATELIMIT_BURST);
2640 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2642 unsigned int filter = SHOW_MEM_FILTER_NODES;
2644 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2645 debug_guardpage_minorder() > 0)
2649 * This documents exceptions given to allocations in certain
2650 * contexts that are allowed to allocate outside current's set
2653 if (!(gfp_mask & __GFP_NOMEMALLOC))
2654 if (test_thread_flag(TIF_MEMDIE) ||
2655 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2656 filter &= ~SHOW_MEM_FILTER_NODES;
2657 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2658 filter &= ~SHOW_MEM_FILTER_NODES;
2661 struct va_format vaf;
2664 va_start(args, fmt);
2669 pr_warn("%pV", &vaf);
2674 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2675 current->comm, order, gfp_mask);
2678 if (!should_suppress_show_mem())
2682 static inline struct page *
2683 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2684 const struct alloc_context *ac, unsigned long *did_some_progress)
2686 struct oom_control oc = {
2687 .zonelist = ac->zonelist,
2688 .nodemask = ac->nodemask,
2689 .gfp_mask = gfp_mask,
2694 *did_some_progress = 0;
2697 * Acquire the oom lock. If that fails, somebody else is
2698 * making progress for us.
2700 if (!mutex_trylock(&oom_lock)) {
2701 *did_some_progress = 1;
2702 schedule_timeout_uninterruptible(1);
2707 * Go through the zonelist yet one more time, keep very high watermark
2708 * here, this is only to catch a parallel oom killing, we must fail if
2709 * we're still under heavy pressure.
2711 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2712 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2716 if (!(gfp_mask & __GFP_NOFAIL)) {
2717 /* Coredumps can quickly deplete all memory reserves */
2718 if (current->flags & PF_DUMPCORE)
2720 /* The OOM killer will not help higher order allocs */
2721 if (order > PAGE_ALLOC_COSTLY_ORDER)
2723 /* The OOM killer does not needlessly kill tasks for lowmem */
2724 if (ac->high_zoneidx < ZONE_NORMAL)
2726 /* The OOM killer does not compensate for IO-less reclaim */
2727 if (!(gfp_mask & __GFP_FS)) {
2729 * XXX: Page reclaim didn't yield anything,
2730 * and the OOM killer can't be invoked, but
2731 * keep looping as per tradition.
2733 *did_some_progress = 1;
2736 if (pm_suspended_storage())
2738 /* The OOM killer may not free memory on a specific node */
2739 if (gfp_mask & __GFP_THISNODE)
2742 /* Exhausted what can be done so it's blamo time */
2743 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2744 *did_some_progress = 1;
2746 mutex_unlock(&oom_lock);
2750 #ifdef CONFIG_COMPACTION
2751 /* Try memory compaction for high-order allocations before reclaim */
2752 static struct page *
2753 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2754 int alloc_flags, const struct alloc_context *ac,
2755 enum migrate_mode mode, int *contended_compaction,
2756 bool *deferred_compaction)
2758 unsigned long compact_result;
2764 current->flags |= PF_MEMALLOC;
2765 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2766 mode, contended_compaction);
2767 current->flags &= ~PF_MEMALLOC;
2769 switch (compact_result) {
2770 case COMPACT_DEFERRED:
2771 *deferred_compaction = true;
2773 case COMPACT_SKIPPED:
2780 * At least in one zone compaction wasn't deferred or skipped, so let's
2781 * count a compaction stall
2783 count_vm_event(COMPACTSTALL);
2785 page = get_page_from_freelist(gfp_mask, order,
2786 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2789 struct zone *zone = page_zone(page);
2791 zone->compact_blockskip_flush = false;
2792 compaction_defer_reset(zone, order, true);
2793 count_vm_event(COMPACTSUCCESS);
2798 * It's bad if compaction run occurs and fails. The most likely reason
2799 * is that pages exist, but not enough to satisfy watermarks.
2801 count_vm_event(COMPACTFAIL);
2808 static inline struct page *
2809 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2810 int alloc_flags, const struct alloc_context *ac,
2811 enum migrate_mode mode, int *contended_compaction,
2812 bool *deferred_compaction)
2816 #endif /* CONFIG_COMPACTION */
2818 /* Perform direct synchronous page reclaim */
2820 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2821 const struct alloc_context *ac)
2823 struct reclaim_state reclaim_state;
2828 /* We now go into synchronous reclaim */
2829 cpuset_memory_pressure_bump();
2830 current->flags |= PF_MEMALLOC;
2831 lockdep_set_current_reclaim_state(gfp_mask);
2832 reclaim_state.reclaimed_slab = 0;
2833 current->reclaim_state = &reclaim_state;
2835 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2838 current->reclaim_state = NULL;
2839 lockdep_clear_current_reclaim_state();
2840 current->flags &= ~PF_MEMALLOC;
2847 /* The really slow allocator path where we enter direct reclaim */
2848 static inline struct page *
2849 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2850 int alloc_flags, const struct alloc_context *ac,
2851 unsigned long *did_some_progress)
2853 struct page *page = NULL;
2854 bool drained = false;
2856 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2857 if (unlikely(!(*did_some_progress)))
2861 page = get_page_from_freelist(gfp_mask, order,
2862 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2865 * If an allocation failed after direct reclaim, it could be because
2866 * pages are pinned on the per-cpu lists or in high alloc reserves.
2867 * Shrink them them and try again
2869 if (!page && !drained) {
2870 unreserve_highatomic_pageblock(ac);
2871 drain_all_pages(NULL);
2879 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2884 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2885 ac->high_zoneidx, ac->nodemask)
2886 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2890 gfp_to_alloc_flags(gfp_t gfp_mask)
2892 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2894 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2895 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2898 * The caller may dip into page reserves a bit more if the caller
2899 * cannot run direct reclaim, or if the caller has realtime scheduling
2900 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2901 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2903 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2905 if (gfp_mask & __GFP_ATOMIC) {
2907 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2908 * if it can't schedule.
2910 if (!(gfp_mask & __GFP_NOMEMALLOC))
2911 alloc_flags |= ALLOC_HARDER;
2913 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2914 * comment for __cpuset_node_allowed().
2916 alloc_flags &= ~ALLOC_CPUSET;
2917 } else if (unlikely(rt_task(current)) && !in_interrupt())
2918 alloc_flags |= ALLOC_HARDER;
2920 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2921 if (gfp_mask & __GFP_MEMALLOC)
2922 alloc_flags |= ALLOC_NO_WATERMARKS;
2923 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2924 alloc_flags |= ALLOC_NO_WATERMARKS;
2925 else if (!in_interrupt() &&
2926 ((current->flags & PF_MEMALLOC) ||
2927 unlikely(test_thread_flag(TIF_MEMDIE))))
2928 alloc_flags |= ALLOC_NO_WATERMARKS;
2931 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2932 alloc_flags |= ALLOC_CMA;
2937 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2939 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2942 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2944 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2947 static inline struct page *
2948 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2949 struct alloc_context *ac)
2951 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2952 struct page *page = NULL;
2954 unsigned long pages_reclaimed = 0;
2955 unsigned long did_some_progress;
2956 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2957 bool deferred_compaction = false;
2958 int contended_compaction = COMPACT_CONTENDED_NONE;
2961 * In the slowpath, we sanity check order to avoid ever trying to
2962 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2963 * be using allocators in order of preference for an area that is
2966 if (order >= MAX_ORDER) {
2967 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2972 * We also sanity check to catch abuse of atomic reserves being used by
2973 * callers that are not in atomic context.
2975 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
2976 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
2977 gfp_mask &= ~__GFP_ATOMIC;
2980 * If this allocation cannot block and it is for a specific node, then
2981 * fail early. There's no need to wakeup kswapd or retry for a
2982 * speculative node-specific allocation.
2984 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
2988 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
2989 wake_all_kswapds(order, ac);
2992 * OK, we're below the kswapd watermark and have kicked background
2993 * reclaim. Now things get more complex, so set up alloc_flags according
2994 * to how we want to proceed.
2996 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2999 * Find the true preferred zone if the allocation is unconstrained by
3002 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3003 struct zoneref *preferred_zoneref;
3004 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3005 ac->high_zoneidx, NULL, &ac->preferred_zone);
3006 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3009 /* This is the last chance, in general, before the goto nopage. */
3010 page = get_page_from_freelist(gfp_mask, order,
3011 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3015 /* Allocate without watermarks if the context allows */
3016 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3018 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3019 * the allocation is high priority and these type of
3020 * allocations are system rather than user orientated
3022 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3024 page = get_page_from_freelist(gfp_mask, order,
3025 ALLOC_NO_WATERMARKS, ac);
3029 if (gfp_mask & __GFP_NOFAIL)
3030 wait_iff_congested(ac->preferred_zone,
3031 BLK_RW_ASYNC, HZ/50);
3032 } while (gfp_mask & __GFP_NOFAIL);
3035 /* Caller is not willing to reclaim, we can't balance anything */
3036 if (!can_direct_reclaim) {
3038 * All existing users of the deprecated __GFP_NOFAIL are
3039 * blockable, so warn of any new users that actually allow this
3040 * type of allocation to fail.
3042 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3046 /* Avoid recursion of direct reclaim */
3047 if (current->flags & PF_MEMALLOC)
3050 /* Avoid allocations with no watermarks from looping endlessly */
3051 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3055 * Try direct compaction. The first pass is asynchronous. Subsequent
3056 * attempts after direct reclaim are synchronous
3058 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3060 &contended_compaction,
3061 &deferred_compaction);
3065 /* Checks for THP-specific high-order allocations */
3066 if (is_thp_gfp_mask(gfp_mask)) {
3068 * If compaction is deferred for high-order allocations, it is
3069 * because sync compaction recently failed. If this is the case
3070 * and the caller requested a THP allocation, we do not want
3071 * to heavily disrupt the system, so we fail the allocation
3072 * instead of entering direct reclaim.
3074 if (deferred_compaction)
3078 * In all zones where compaction was attempted (and not
3079 * deferred or skipped), lock contention has been detected.
3080 * For THP allocation we do not want to disrupt the others
3081 * so we fallback to base pages instead.
3083 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3087 * If compaction was aborted due to need_resched(), we do not
3088 * want to further increase allocation latency, unless it is
3089 * khugepaged trying to collapse.
3091 if (contended_compaction == COMPACT_CONTENDED_SCHED
3092 && !(current->flags & PF_KTHREAD))
3097 * It can become very expensive to allocate transparent hugepages at
3098 * fault, so use asynchronous memory compaction for THP unless it is
3099 * khugepaged trying to collapse.
3101 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3102 migration_mode = MIGRATE_SYNC_LIGHT;
3104 /* Try direct reclaim and then allocating */
3105 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3106 &did_some_progress);
3110 /* Do not loop if specifically requested */
3111 if (gfp_mask & __GFP_NORETRY)
3114 /* Keep reclaiming pages as long as there is reasonable progress */
3115 pages_reclaimed += did_some_progress;
3116 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3117 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3118 /* Wait for some write requests to complete then retry */
3119 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3123 /* Reclaim has failed us, start killing things */
3124 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3128 /* Retry as long as the OOM killer is making progress */
3129 if (did_some_progress)
3134 * High-order allocations do not necessarily loop after
3135 * direct reclaim and reclaim/compaction depends on compaction
3136 * being called after reclaim so call directly if necessary
3138 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3140 &contended_compaction,
3141 &deferred_compaction);
3145 warn_alloc_failed(gfp_mask, order, NULL);
3151 * This is the 'heart' of the zoned buddy allocator.
3154 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3155 struct zonelist *zonelist, nodemask_t *nodemask)
3157 struct zoneref *preferred_zoneref;
3158 struct page *page = NULL;
3159 unsigned int cpuset_mems_cookie;
3160 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3161 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3162 struct alloc_context ac = {
3163 .high_zoneidx = gfp_zone(gfp_mask),
3164 .nodemask = nodemask,
3165 .migratetype = gfpflags_to_migratetype(gfp_mask),
3168 gfp_mask &= gfp_allowed_mask;
3170 lockdep_trace_alloc(gfp_mask);
3172 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3174 if (should_fail_alloc_page(gfp_mask, order))
3178 * Check the zones suitable for the gfp_mask contain at least one
3179 * valid zone. It's possible to have an empty zonelist as a result
3180 * of __GFP_THISNODE and a memoryless node
3182 if (unlikely(!zonelist->_zonerefs->zone))
3185 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3186 alloc_flags |= ALLOC_CMA;
3189 cpuset_mems_cookie = read_mems_allowed_begin();
3191 /* We set it here, as __alloc_pages_slowpath might have changed it */
3192 ac.zonelist = zonelist;
3194 /* Dirty zone balancing only done in the fast path */
3195 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3197 /* The preferred zone is used for statistics later */
3198 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3199 ac.nodemask ? : &cpuset_current_mems_allowed,
3200 &ac.preferred_zone);
3201 if (!ac.preferred_zone)
3203 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3205 /* First allocation attempt */
3206 alloc_mask = gfp_mask|__GFP_HARDWALL;
3207 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3208 if (unlikely(!page)) {
3210 * Runtime PM, block IO and its error handling path
3211 * can deadlock because I/O on the device might not
3214 alloc_mask = memalloc_noio_flags(gfp_mask);
3215 ac.spread_dirty_pages = false;
3217 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3220 if (kmemcheck_enabled && page)
3221 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3223 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3227 * When updating a task's mems_allowed, it is possible to race with
3228 * parallel threads in such a way that an allocation can fail while
3229 * the mask is being updated. If a page allocation is about to fail,
3230 * check if the cpuset changed during allocation and if so, retry.
3232 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3237 EXPORT_SYMBOL(__alloc_pages_nodemask);
3240 * Common helper functions.
3242 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3247 * __get_free_pages() returns a 32-bit address, which cannot represent
3250 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3252 page = alloc_pages(gfp_mask, order);
3255 return (unsigned long) page_address(page);
3257 EXPORT_SYMBOL(__get_free_pages);
3259 unsigned long get_zeroed_page(gfp_t gfp_mask)
3261 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3263 EXPORT_SYMBOL(get_zeroed_page);
3265 void __free_pages(struct page *page, unsigned int order)
3267 if (put_page_testzero(page)) {
3269 free_hot_cold_page(page, false);
3271 __free_pages_ok(page, order);
3275 EXPORT_SYMBOL(__free_pages);
3277 void free_pages(unsigned long addr, unsigned int order)
3280 VM_BUG_ON(!virt_addr_valid((void *)addr));
3281 __free_pages(virt_to_page((void *)addr), order);
3285 EXPORT_SYMBOL(free_pages);
3289 * An arbitrary-length arbitrary-offset area of memory which resides
3290 * within a 0 or higher order page. Multiple fragments within that page
3291 * are individually refcounted, in the page's reference counter.
3293 * The page_frag functions below provide a simple allocation framework for
3294 * page fragments. This is used by the network stack and network device
3295 * drivers to provide a backing region of memory for use as either an
3296 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3298 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3301 struct page *page = NULL;
3302 gfp_t gfp = gfp_mask;
3304 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3305 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3307 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3308 PAGE_FRAG_CACHE_MAX_ORDER);
3309 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3311 if (unlikely(!page))
3312 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3314 nc->va = page ? page_address(page) : NULL;
3319 void *__alloc_page_frag(struct page_frag_cache *nc,
3320 unsigned int fragsz, gfp_t gfp_mask)
3322 unsigned int size = PAGE_SIZE;
3326 if (unlikely(!nc->va)) {
3328 page = __page_frag_refill(nc, gfp_mask);
3332 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3333 /* if size can vary use size else just use PAGE_SIZE */
3336 /* Even if we own the page, we do not use atomic_set().
3337 * This would break get_page_unless_zero() users.
3339 atomic_add(size - 1, &page->_count);
3341 /* reset page count bias and offset to start of new frag */
3342 nc->pfmemalloc = page_is_pfmemalloc(page);
3343 nc->pagecnt_bias = size;
3347 offset = nc->offset - fragsz;
3348 if (unlikely(offset < 0)) {
3349 page = virt_to_page(nc->va);
3351 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3354 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3355 /* if size can vary use size else just use PAGE_SIZE */
3358 /* OK, page count is 0, we can safely set it */
3359 atomic_set(&page->_count, size);
3361 /* reset page count bias and offset to start of new frag */
3362 nc->pagecnt_bias = size;
3363 offset = size - fragsz;
3367 nc->offset = offset;
3369 return nc->va + offset;
3371 EXPORT_SYMBOL(__alloc_page_frag);
3374 * Frees a page fragment allocated out of either a compound or order 0 page.
3376 void __free_page_frag(void *addr)
3378 struct page *page = virt_to_head_page(addr);
3380 if (unlikely(put_page_testzero(page)))
3381 __free_pages_ok(page, compound_order(page));
3383 EXPORT_SYMBOL(__free_page_frag);
3386 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3387 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3388 * equivalent to alloc_pages.
3390 * It should be used when the caller would like to use kmalloc, but since the
3391 * allocation is large, it has to fall back to the page allocator.
3393 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3397 page = alloc_pages(gfp_mask, order);
3398 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3399 __free_pages(page, order);
3405 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3409 page = alloc_pages_node(nid, gfp_mask, order);
3410 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3411 __free_pages(page, order);
3418 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3421 void __free_kmem_pages(struct page *page, unsigned int order)
3423 memcg_kmem_uncharge(page, order);
3424 __free_pages(page, order);
3427 void free_kmem_pages(unsigned long addr, unsigned int order)
3430 VM_BUG_ON(!virt_addr_valid((void *)addr));
3431 __free_kmem_pages(virt_to_page((void *)addr), order);
3435 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3439 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3440 unsigned long used = addr + PAGE_ALIGN(size);
3442 split_page(virt_to_page((void *)addr), order);
3443 while (used < alloc_end) {
3448 return (void *)addr;
3452 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3453 * @size: the number of bytes to allocate
3454 * @gfp_mask: GFP flags for the allocation
3456 * This function is similar to alloc_pages(), except that it allocates the
3457 * minimum number of pages to satisfy the request. alloc_pages() can only
3458 * allocate memory in power-of-two pages.
3460 * This function is also limited by MAX_ORDER.
3462 * Memory allocated by this function must be released by free_pages_exact().
3464 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3466 unsigned int order = get_order(size);
3469 addr = __get_free_pages(gfp_mask, order);
3470 return make_alloc_exact(addr, order, size);
3472 EXPORT_SYMBOL(alloc_pages_exact);
3475 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3477 * @nid: the preferred node ID where memory should be allocated
3478 * @size: the number of bytes to allocate
3479 * @gfp_mask: GFP flags for the allocation
3481 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3484 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3486 unsigned int order = get_order(size);
3487 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3490 return make_alloc_exact((unsigned long)page_address(p), order, size);
3494 * free_pages_exact - release memory allocated via alloc_pages_exact()
3495 * @virt: the value returned by alloc_pages_exact.
3496 * @size: size of allocation, same value as passed to alloc_pages_exact().
3498 * Release the memory allocated by a previous call to alloc_pages_exact.
3500 void free_pages_exact(void *virt, size_t size)
3502 unsigned long addr = (unsigned long)virt;
3503 unsigned long end = addr + PAGE_ALIGN(size);
3505 while (addr < end) {
3510 EXPORT_SYMBOL(free_pages_exact);
3513 * nr_free_zone_pages - count number of pages beyond high watermark
3514 * @offset: The zone index of the highest zone
3516 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3517 * high watermark within all zones at or below a given zone index. For each
3518 * zone, the number of pages is calculated as:
3519 * managed_pages - high_pages
3521 static unsigned long nr_free_zone_pages(int offset)
3526 /* Just pick one node, since fallback list is circular */
3527 unsigned long sum = 0;
3529 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3531 for_each_zone_zonelist(zone, z, zonelist, offset) {
3532 unsigned long size = zone->managed_pages;
3533 unsigned long high = high_wmark_pages(zone);
3542 * nr_free_buffer_pages - count number of pages beyond high watermark
3544 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3545 * watermark within ZONE_DMA and ZONE_NORMAL.
3547 unsigned long nr_free_buffer_pages(void)
3549 return nr_free_zone_pages(gfp_zone(GFP_USER));
3551 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3554 * nr_free_pagecache_pages - count number of pages beyond high watermark
3556 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3557 * high watermark within all zones.
3559 unsigned long nr_free_pagecache_pages(void)
3561 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3564 static inline void show_node(struct zone *zone)
3566 if (IS_ENABLED(CONFIG_NUMA))
3567 printk("Node %d ", zone_to_nid(zone));
3570 void si_meminfo(struct sysinfo *val)
3572 val->totalram = totalram_pages;
3573 val->sharedram = global_page_state(NR_SHMEM);
3574 val->freeram = global_page_state(NR_FREE_PAGES);
3575 val->bufferram = nr_blockdev_pages();
3576 val->totalhigh = totalhigh_pages;
3577 val->freehigh = nr_free_highpages();
3578 val->mem_unit = PAGE_SIZE;
3581 EXPORT_SYMBOL(si_meminfo);
3584 void si_meminfo_node(struct sysinfo *val, int nid)
3586 int zone_type; /* needs to be signed */
3587 unsigned long managed_pages = 0;
3588 pg_data_t *pgdat = NODE_DATA(nid);
3590 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3591 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3592 val->totalram = managed_pages;
3593 val->sharedram = node_page_state(nid, NR_SHMEM);
3594 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3595 #ifdef CONFIG_HIGHMEM
3596 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3597 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3603 val->mem_unit = PAGE_SIZE;
3608 * Determine whether the node should be displayed or not, depending on whether
3609 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3611 bool skip_free_areas_node(unsigned int flags, int nid)
3614 unsigned int cpuset_mems_cookie;
3616 if (!(flags & SHOW_MEM_FILTER_NODES))
3620 cpuset_mems_cookie = read_mems_allowed_begin();
3621 ret = !node_isset(nid, cpuset_current_mems_allowed);
3622 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3627 #define K(x) ((x) << (PAGE_SHIFT-10))
3629 static void show_migration_types(unsigned char type)
3631 static const char types[MIGRATE_TYPES] = {
3632 [MIGRATE_UNMOVABLE] = 'U',
3633 [MIGRATE_MOVABLE] = 'M',
3634 [MIGRATE_RECLAIMABLE] = 'E',
3635 [MIGRATE_HIGHATOMIC] = 'H',
3637 [MIGRATE_CMA] = 'C',
3639 #ifdef CONFIG_MEMORY_ISOLATION
3640 [MIGRATE_ISOLATE] = 'I',
3643 char tmp[MIGRATE_TYPES + 1];
3647 for (i = 0; i < MIGRATE_TYPES; i++) {
3648 if (type & (1 << i))
3653 printk("(%s) ", tmp);
3657 * Show free area list (used inside shift_scroll-lock stuff)
3658 * We also calculate the percentage fragmentation. We do this by counting the
3659 * memory on each free list with the exception of the first item on the list.
3662 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3665 void show_free_areas(unsigned int filter)
3667 unsigned long free_pcp = 0;
3671 for_each_populated_zone(zone) {
3672 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3675 for_each_online_cpu(cpu)
3676 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3679 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3680 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3681 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3682 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3683 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3684 " free:%lu free_pcp:%lu free_cma:%lu\n",
3685 global_page_state(NR_ACTIVE_ANON),
3686 global_page_state(NR_INACTIVE_ANON),
3687 global_page_state(NR_ISOLATED_ANON),
3688 global_page_state(NR_ACTIVE_FILE),
3689 global_page_state(NR_INACTIVE_FILE),
3690 global_page_state(NR_ISOLATED_FILE),
3691 global_page_state(NR_UNEVICTABLE),
3692 global_page_state(NR_FILE_DIRTY),
3693 global_page_state(NR_WRITEBACK),
3694 global_page_state(NR_UNSTABLE_NFS),
3695 global_page_state(NR_SLAB_RECLAIMABLE),
3696 global_page_state(NR_SLAB_UNRECLAIMABLE),
3697 global_page_state(NR_FILE_MAPPED),
3698 global_page_state(NR_SHMEM),
3699 global_page_state(NR_PAGETABLE),
3700 global_page_state(NR_BOUNCE),
3701 global_page_state(NR_FREE_PAGES),
3703 global_page_state(NR_FREE_CMA_PAGES));
3705 for_each_populated_zone(zone) {
3708 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3712 for_each_online_cpu(cpu)
3713 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3721 " active_anon:%lukB"
3722 " inactive_anon:%lukB"
3723 " active_file:%lukB"
3724 " inactive_file:%lukB"
3725 " unevictable:%lukB"
3726 " isolated(anon):%lukB"
3727 " isolated(file):%lukB"
3735 " slab_reclaimable:%lukB"
3736 " slab_unreclaimable:%lukB"
3737 " kernel_stack:%lukB"
3744 " writeback_tmp:%lukB"
3745 " pages_scanned:%lu"
3746 " all_unreclaimable? %s"
3749 K(zone_page_state(zone, NR_FREE_PAGES)),
3750 K(min_wmark_pages(zone)),
3751 K(low_wmark_pages(zone)),
3752 K(high_wmark_pages(zone)),
3753 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3754 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3755 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3756 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3757 K(zone_page_state(zone, NR_UNEVICTABLE)),
3758 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3759 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3760 K(zone->present_pages),
3761 K(zone->managed_pages),
3762 K(zone_page_state(zone, NR_MLOCK)),
3763 K(zone_page_state(zone, NR_FILE_DIRTY)),
3764 K(zone_page_state(zone, NR_WRITEBACK)),
3765 K(zone_page_state(zone, NR_FILE_MAPPED)),
3766 K(zone_page_state(zone, NR_SHMEM)),
3767 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3768 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3769 zone_page_state(zone, NR_KERNEL_STACK) *
3771 K(zone_page_state(zone, NR_PAGETABLE)),
3772 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3773 K(zone_page_state(zone, NR_BOUNCE)),
3775 K(this_cpu_read(zone->pageset->pcp.count)),
3776 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3777 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3778 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3779 (!zone_reclaimable(zone) ? "yes" : "no")
3781 printk("lowmem_reserve[]:");
3782 for (i = 0; i < MAX_NR_ZONES; i++)
3783 printk(" %ld", zone->lowmem_reserve[i]);
3787 for_each_populated_zone(zone) {
3789 unsigned long nr[MAX_ORDER], flags, total = 0;
3790 unsigned char types[MAX_ORDER];
3792 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3795 printk("%s: ", zone->name);
3797 spin_lock_irqsave(&zone->lock, flags);
3798 for (order = 0; order < MAX_ORDER; order++) {
3799 struct free_area *area = &zone->free_area[order];
3802 nr[order] = area->nr_free;
3803 total += nr[order] << order;
3806 for (type = 0; type < MIGRATE_TYPES; type++) {
3807 if (!list_empty(&area->free_list[type]))
3808 types[order] |= 1 << type;
3811 spin_unlock_irqrestore(&zone->lock, flags);
3812 for (order = 0; order < MAX_ORDER; order++) {
3813 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3815 show_migration_types(types[order]);
3817 printk("= %lukB\n", K(total));
3820 hugetlb_show_meminfo();
3822 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3824 show_swap_cache_info();
3827 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3829 zoneref->zone = zone;
3830 zoneref->zone_idx = zone_idx(zone);
3834 * Builds allocation fallback zone lists.
3836 * Add all populated zones of a node to the zonelist.
3838 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3842 enum zone_type zone_type = MAX_NR_ZONES;
3846 zone = pgdat->node_zones + zone_type;
3847 if (populated_zone(zone)) {
3848 zoneref_set_zone(zone,
3849 &zonelist->_zonerefs[nr_zones++]);
3850 check_highest_zone(zone_type);
3852 } while (zone_type);
3860 * 0 = automatic detection of better ordering.
3861 * 1 = order by ([node] distance, -zonetype)
3862 * 2 = order by (-zonetype, [node] distance)
3864 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3865 * the same zonelist. So only NUMA can configure this param.
3867 #define ZONELIST_ORDER_DEFAULT 0
3868 #define ZONELIST_ORDER_NODE 1
3869 #define ZONELIST_ORDER_ZONE 2
3871 /* zonelist order in the kernel.
3872 * set_zonelist_order() will set this to NODE or ZONE.
3874 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3875 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3879 /* The value user specified ....changed by config */
3880 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3881 /* string for sysctl */
3882 #define NUMA_ZONELIST_ORDER_LEN 16
3883 char numa_zonelist_order[16] = "default";
3886 * interface for configure zonelist ordering.
3887 * command line option "numa_zonelist_order"
3888 * = "[dD]efault - default, automatic configuration.
3889 * = "[nN]ode - order by node locality, then by zone within node
3890 * = "[zZ]one - order by zone, then by locality within zone
3893 static int __parse_numa_zonelist_order(char *s)
3895 if (*s == 'd' || *s == 'D') {
3896 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3897 } else if (*s == 'n' || *s == 'N') {
3898 user_zonelist_order = ZONELIST_ORDER_NODE;
3899 } else if (*s == 'z' || *s == 'Z') {
3900 user_zonelist_order = ZONELIST_ORDER_ZONE;
3903 "Ignoring invalid numa_zonelist_order value: "
3910 static __init int setup_numa_zonelist_order(char *s)
3917 ret = __parse_numa_zonelist_order(s);
3919 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3923 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3926 * sysctl handler for numa_zonelist_order
3928 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3929 void __user *buffer, size_t *length,
3932 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3934 static DEFINE_MUTEX(zl_order_mutex);
3936 mutex_lock(&zl_order_mutex);
3938 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3942 strcpy(saved_string, (char *)table->data);
3944 ret = proc_dostring(table, write, buffer, length, ppos);
3948 int oldval = user_zonelist_order;
3950 ret = __parse_numa_zonelist_order((char *)table->data);
3953 * bogus value. restore saved string
3955 strncpy((char *)table->data, saved_string,
3956 NUMA_ZONELIST_ORDER_LEN);
3957 user_zonelist_order = oldval;
3958 } else if (oldval != user_zonelist_order) {
3959 mutex_lock(&zonelists_mutex);
3960 build_all_zonelists(NULL, NULL);
3961 mutex_unlock(&zonelists_mutex);
3965 mutex_unlock(&zl_order_mutex);
3970 #define MAX_NODE_LOAD (nr_online_nodes)
3971 static int node_load[MAX_NUMNODES];
3974 * find_next_best_node - find the next node that should appear in a given node's fallback list
3975 * @node: node whose fallback list we're appending
3976 * @used_node_mask: nodemask_t of already used nodes
3978 * We use a number of factors to determine which is the next node that should
3979 * appear on a given node's fallback list. The node should not have appeared
3980 * already in @node's fallback list, and it should be the next closest node
3981 * according to the distance array (which contains arbitrary distance values
3982 * from each node to each node in the system), and should also prefer nodes
3983 * with no CPUs, since presumably they'll have very little allocation pressure
3984 * on them otherwise.
3985 * It returns -1 if no node is found.
3987 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3990 int min_val = INT_MAX;
3991 int best_node = NUMA_NO_NODE;
3992 const struct cpumask *tmp = cpumask_of_node(0);
3994 /* Use the local node if we haven't already */
3995 if (!node_isset(node, *used_node_mask)) {
3996 node_set(node, *used_node_mask);
4000 for_each_node_state(n, N_MEMORY) {
4002 /* Don't want a node to appear more than once */
4003 if (node_isset(n, *used_node_mask))
4006 /* Use the distance array to find the distance */
4007 val = node_distance(node, n);
4009 /* Penalize nodes under us ("prefer the next node") */
4012 /* Give preference to headless and unused nodes */
4013 tmp = cpumask_of_node(n);
4014 if (!cpumask_empty(tmp))
4015 val += PENALTY_FOR_NODE_WITH_CPUS;
4017 /* Slight preference for less loaded node */
4018 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4019 val += node_load[n];
4021 if (val < min_val) {
4028 node_set(best_node, *used_node_mask);
4035 * Build zonelists ordered by node and zones within node.
4036 * This results in maximum locality--normal zone overflows into local
4037 * DMA zone, if any--but risks exhausting DMA zone.
4039 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4042 struct zonelist *zonelist;
4044 zonelist = &pgdat->node_zonelists[0];
4045 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4047 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4048 zonelist->_zonerefs[j].zone = NULL;
4049 zonelist->_zonerefs[j].zone_idx = 0;
4053 * Build gfp_thisnode zonelists
4055 static void build_thisnode_zonelists(pg_data_t *pgdat)
4058 struct zonelist *zonelist;
4060 zonelist = &pgdat->node_zonelists[1];
4061 j = build_zonelists_node(pgdat, zonelist, 0);
4062 zonelist->_zonerefs[j].zone = NULL;
4063 zonelist->_zonerefs[j].zone_idx = 0;
4067 * Build zonelists ordered by zone and nodes within zones.
4068 * This results in conserving DMA zone[s] until all Normal memory is
4069 * exhausted, but results in overflowing to remote node while memory
4070 * may still exist in local DMA zone.
4072 static int node_order[MAX_NUMNODES];
4074 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4077 int zone_type; /* needs to be signed */
4079 struct zonelist *zonelist;
4081 zonelist = &pgdat->node_zonelists[0];
4083 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4084 for (j = 0; j < nr_nodes; j++) {
4085 node = node_order[j];
4086 z = &NODE_DATA(node)->node_zones[zone_type];
4087 if (populated_zone(z)) {
4089 &zonelist->_zonerefs[pos++]);
4090 check_highest_zone(zone_type);
4094 zonelist->_zonerefs[pos].zone = NULL;
4095 zonelist->_zonerefs[pos].zone_idx = 0;
4098 #if defined(CONFIG_64BIT)
4100 * Devices that require DMA32/DMA are relatively rare and do not justify a
4101 * penalty to every machine in case the specialised case applies. Default
4102 * to Node-ordering on 64-bit NUMA machines
4104 static int default_zonelist_order(void)
4106 return ZONELIST_ORDER_NODE;
4110 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4111 * by the kernel. If processes running on node 0 deplete the low memory zone
4112 * then reclaim will occur more frequency increasing stalls and potentially
4113 * be easier to OOM if a large percentage of the zone is under writeback or
4114 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4115 * Hence, default to zone ordering on 32-bit.
4117 static int default_zonelist_order(void)
4119 return ZONELIST_ORDER_ZONE;
4121 #endif /* CONFIG_64BIT */
4123 static void set_zonelist_order(void)
4125 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4126 current_zonelist_order = default_zonelist_order();
4128 current_zonelist_order = user_zonelist_order;
4131 static void build_zonelists(pg_data_t *pgdat)
4134 nodemask_t used_mask;
4135 int local_node, prev_node;
4136 struct zonelist *zonelist;
4137 unsigned int order = current_zonelist_order;
4139 /* initialize zonelists */
4140 for (i = 0; i < MAX_ZONELISTS; i++) {
4141 zonelist = pgdat->node_zonelists + i;
4142 zonelist->_zonerefs[0].zone = NULL;
4143 zonelist->_zonerefs[0].zone_idx = 0;
4146 /* NUMA-aware ordering of nodes */
4147 local_node = pgdat->node_id;
4148 load = nr_online_nodes;
4149 prev_node = local_node;
4150 nodes_clear(used_mask);
4152 memset(node_order, 0, sizeof(node_order));
4155 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4157 * We don't want to pressure a particular node.
4158 * So adding penalty to the first node in same
4159 * distance group to make it round-robin.
4161 if (node_distance(local_node, node) !=
4162 node_distance(local_node, prev_node))
4163 node_load[node] = load;
4167 if (order == ZONELIST_ORDER_NODE)
4168 build_zonelists_in_node_order(pgdat, node);
4170 node_order[i++] = node; /* remember order */
4173 if (order == ZONELIST_ORDER_ZONE) {
4174 /* calculate node order -- i.e., DMA last! */
4175 build_zonelists_in_zone_order(pgdat, i);
4178 build_thisnode_zonelists(pgdat);
4181 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4183 * Return node id of node used for "local" allocations.
4184 * I.e., first node id of first zone in arg node's generic zonelist.
4185 * Used for initializing percpu 'numa_mem', which is used primarily
4186 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4188 int local_memory_node(int node)
4192 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4193 gfp_zone(GFP_KERNEL),
4200 #else /* CONFIG_NUMA */
4202 static void set_zonelist_order(void)
4204 current_zonelist_order = ZONELIST_ORDER_ZONE;
4207 static void build_zonelists(pg_data_t *pgdat)
4209 int node, local_node;
4211 struct zonelist *zonelist;
4213 local_node = pgdat->node_id;
4215 zonelist = &pgdat->node_zonelists[0];
4216 j = build_zonelists_node(pgdat, zonelist, 0);
4219 * Now we build the zonelist so that it contains the zones
4220 * of all the other nodes.
4221 * We don't want to pressure a particular node, so when
4222 * building the zones for node N, we make sure that the
4223 * zones coming right after the local ones are those from
4224 * node N+1 (modulo N)
4226 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4227 if (!node_online(node))
4229 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4231 for (node = 0; node < local_node; node++) {
4232 if (!node_online(node))
4234 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4237 zonelist->_zonerefs[j].zone = NULL;
4238 zonelist->_zonerefs[j].zone_idx = 0;
4241 #endif /* CONFIG_NUMA */
4244 * Boot pageset table. One per cpu which is going to be used for all
4245 * zones and all nodes. The parameters will be set in such a way
4246 * that an item put on a list will immediately be handed over to
4247 * the buddy list. This is safe since pageset manipulation is done
4248 * with interrupts disabled.
4250 * The boot_pagesets must be kept even after bootup is complete for
4251 * unused processors and/or zones. They do play a role for bootstrapping
4252 * hotplugged processors.
4254 * zoneinfo_show() and maybe other functions do
4255 * not check if the processor is online before following the pageset pointer.
4256 * Other parts of the kernel may not check if the zone is available.
4258 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4259 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4260 static void setup_zone_pageset(struct zone *zone);
4263 * Global mutex to protect against size modification of zonelists
4264 * as well as to serialize pageset setup for the new populated zone.
4266 DEFINE_MUTEX(zonelists_mutex);
4268 /* return values int ....just for stop_machine() */
4269 static int __build_all_zonelists(void *data)
4273 pg_data_t *self = data;
4276 memset(node_load, 0, sizeof(node_load));
4279 if (self && !node_online(self->node_id)) {
4280 build_zonelists(self);
4283 for_each_online_node(nid) {
4284 pg_data_t *pgdat = NODE_DATA(nid);
4286 build_zonelists(pgdat);
4290 * Initialize the boot_pagesets that are going to be used
4291 * for bootstrapping processors. The real pagesets for
4292 * each zone will be allocated later when the per cpu
4293 * allocator is available.
4295 * boot_pagesets are used also for bootstrapping offline
4296 * cpus if the system is already booted because the pagesets
4297 * are needed to initialize allocators on a specific cpu too.
4298 * F.e. the percpu allocator needs the page allocator which
4299 * needs the percpu allocator in order to allocate its pagesets
4300 * (a chicken-egg dilemma).
4302 for_each_possible_cpu(cpu) {
4303 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4305 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4307 * We now know the "local memory node" for each node--
4308 * i.e., the node of the first zone in the generic zonelist.
4309 * Set up numa_mem percpu variable for on-line cpus. During
4310 * boot, only the boot cpu should be on-line; we'll init the
4311 * secondary cpus' numa_mem as they come on-line. During
4312 * node/memory hotplug, we'll fixup all on-line cpus.
4314 if (cpu_online(cpu))
4315 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4322 static noinline void __init
4323 build_all_zonelists_init(void)
4325 __build_all_zonelists(NULL);
4326 mminit_verify_zonelist();
4327 cpuset_init_current_mems_allowed();
4331 * Called with zonelists_mutex held always
4332 * unless system_state == SYSTEM_BOOTING.
4334 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4335 * [we're only called with non-NULL zone through __meminit paths] and
4336 * (2) call of __init annotated helper build_all_zonelists_init
4337 * [protected by SYSTEM_BOOTING].
4339 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4341 set_zonelist_order();
4343 if (system_state == SYSTEM_BOOTING) {
4344 build_all_zonelists_init();
4346 #ifdef CONFIG_MEMORY_HOTPLUG
4348 setup_zone_pageset(zone);
4350 /* we have to stop all cpus to guarantee there is no user
4352 stop_machine(__build_all_zonelists, pgdat, NULL);
4353 /* cpuset refresh routine should be here */
4355 vm_total_pages = nr_free_pagecache_pages();
4357 * Disable grouping by mobility if the number of pages in the
4358 * system is too low to allow the mechanism to work. It would be
4359 * more accurate, but expensive to check per-zone. This check is
4360 * made on memory-hotadd so a system can start with mobility
4361 * disabled and enable it later
4363 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4364 page_group_by_mobility_disabled = 1;
4366 page_group_by_mobility_disabled = 0;
4368 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4369 "Total pages: %ld\n",
4371 zonelist_order_name[current_zonelist_order],
4372 page_group_by_mobility_disabled ? "off" : "on",
4375 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4380 * Helper functions to size the waitqueue hash table.
4381 * Essentially these want to choose hash table sizes sufficiently
4382 * large so that collisions trying to wait on pages are rare.
4383 * But in fact, the number of active page waitqueues on typical
4384 * systems is ridiculously low, less than 200. So this is even
4385 * conservative, even though it seems large.
4387 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4388 * waitqueues, i.e. the size of the waitq table given the number of pages.
4390 #define PAGES_PER_WAITQUEUE 256
4392 #ifndef CONFIG_MEMORY_HOTPLUG
4393 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4395 unsigned long size = 1;
4397 pages /= PAGES_PER_WAITQUEUE;
4399 while (size < pages)
4403 * Once we have dozens or even hundreds of threads sleeping
4404 * on IO we've got bigger problems than wait queue collision.
4405 * Limit the size of the wait table to a reasonable size.
4407 size = min(size, 4096UL);
4409 return max(size, 4UL);
4413 * A zone's size might be changed by hot-add, so it is not possible to determine
4414 * a suitable size for its wait_table. So we use the maximum size now.
4416 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4418 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4419 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4420 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4422 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4423 * or more by the traditional way. (See above). It equals:
4425 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4426 * ia64(16K page size) : = ( 8G + 4M)byte.
4427 * powerpc (64K page size) : = (32G +16M)byte.
4429 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4436 * This is an integer logarithm so that shifts can be used later
4437 * to extract the more random high bits from the multiplicative
4438 * hash function before the remainder is taken.
4440 static inline unsigned long wait_table_bits(unsigned long size)
4446 * Initially all pages are reserved - free ones are freed
4447 * up by free_all_bootmem() once the early boot process is
4448 * done. Non-atomic initialization, single-pass.
4450 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4451 unsigned long start_pfn, enum memmap_context context)
4453 pg_data_t *pgdat = NODE_DATA(nid);
4454 unsigned long end_pfn = start_pfn + size;
4457 unsigned long nr_initialised = 0;
4459 if (highest_memmap_pfn < end_pfn - 1)
4460 highest_memmap_pfn = end_pfn - 1;
4462 z = &pgdat->node_zones[zone];
4463 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4465 * There can be holes in boot-time mem_map[]s
4466 * handed to this function. They do not
4467 * exist on hotplugged memory.
4469 if (context == MEMMAP_EARLY) {
4470 if (!early_pfn_valid(pfn))
4472 if (!early_pfn_in_nid(pfn, nid))
4474 if (!update_defer_init(pgdat, pfn, end_pfn,
4480 * Mark the block movable so that blocks are reserved for
4481 * movable at startup. This will force kernel allocations
4482 * to reserve their blocks rather than leaking throughout
4483 * the address space during boot when many long-lived
4484 * kernel allocations are made.
4486 * bitmap is created for zone's valid pfn range. but memmap
4487 * can be created for invalid pages (for alignment)
4488 * check here not to call set_pageblock_migratetype() against
4491 if (!(pfn & (pageblock_nr_pages - 1))) {
4492 struct page *page = pfn_to_page(pfn);
4494 __init_single_page(page, pfn, zone, nid);
4495 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4497 __init_single_pfn(pfn, zone, nid);
4502 static void __meminit zone_init_free_lists(struct zone *zone)
4504 unsigned int order, t;
4505 for_each_migratetype_order(order, t) {
4506 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4507 zone->free_area[order].nr_free = 0;
4511 #ifndef __HAVE_ARCH_MEMMAP_INIT
4512 #define memmap_init(size, nid, zone, start_pfn) \
4513 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4516 static int zone_batchsize(struct zone *zone)
4522 * The per-cpu-pages pools are set to around 1000th of the
4523 * size of the zone. But no more than 1/2 of a meg.
4525 * OK, so we don't know how big the cache is. So guess.
4527 batch = zone->managed_pages / 1024;
4528 if (batch * PAGE_SIZE > 512 * 1024)
4529 batch = (512 * 1024) / PAGE_SIZE;
4530 batch /= 4; /* We effectively *= 4 below */
4535 * Clamp the batch to a 2^n - 1 value. Having a power
4536 * of 2 value was found to be more likely to have
4537 * suboptimal cache aliasing properties in some cases.
4539 * For example if 2 tasks are alternately allocating
4540 * batches of pages, one task can end up with a lot
4541 * of pages of one half of the possible page colors
4542 * and the other with pages of the other colors.
4544 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4549 /* The deferral and batching of frees should be suppressed under NOMMU
4552 * The problem is that NOMMU needs to be able to allocate large chunks
4553 * of contiguous memory as there's no hardware page translation to
4554 * assemble apparent contiguous memory from discontiguous pages.
4556 * Queueing large contiguous runs of pages for batching, however,
4557 * causes the pages to actually be freed in smaller chunks. As there
4558 * can be a significant delay between the individual batches being
4559 * recycled, this leads to the once large chunks of space being
4560 * fragmented and becoming unavailable for high-order allocations.
4567 * pcp->high and pcp->batch values are related and dependent on one another:
4568 * ->batch must never be higher then ->high.
4569 * The following function updates them in a safe manner without read side
4572 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4573 * those fields changing asynchronously (acording the the above rule).
4575 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4576 * outside of boot time (or some other assurance that no concurrent updaters
4579 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4580 unsigned long batch)
4582 /* start with a fail safe value for batch */
4586 /* Update high, then batch, in order */
4593 /* a companion to pageset_set_high() */
4594 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4596 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4599 static void pageset_init(struct per_cpu_pageset *p)
4601 struct per_cpu_pages *pcp;
4604 memset(p, 0, sizeof(*p));
4608 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4609 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4612 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4615 pageset_set_batch(p, batch);
4619 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4620 * to the value high for the pageset p.
4622 static void pageset_set_high(struct per_cpu_pageset *p,
4625 unsigned long batch = max(1UL, high / 4);
4626 if ((high / 4) > (PAGE_SHIFT * 8))
4627 batch = PAGE_SHIFT * 8;
4629 pageset_update(&p->pcp, high, batch);
4632 static void pageset_set_high_and_batch(struct zone *zone,
4633 struct per_cpu_pageset *pcp)
4635 if (percpu_pagelist_fraction)
4636 pageset_set_high(pcp,
4637 (zone->managed_pages /
4638 percpu_pagelist_fraction));
4640 pageset_set_batch(pcp, zone_batchsize(zone));
4643 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4645 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4648 pageset_set_high_and_batch(zone, pcp);
4651 static void __meminit setup_zone_pageset(struct zone *zone)
4654 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4655 for_each_possible_cpu(cpu)
4656 zone_pageset_init(zone, cpu);
4660 * Allocate per cpu pagesets and initialize them.
4661 * Before this call only boot pagesets were available.
4663 void __init setup_per_cpu_pageset(void)
4667 for_each_populated_zone(zone)
4668 setup_zone_pageset(zone);
4671 static noinline __init_refok
4672 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4678 * The per-page waitqueue mechanism uses hashed waitqueues
4681 zone->wait_table_hash_nr_entries =
4682 wait_table_hash_nr_entries(zone_size_pages);
4683 zone->wait_table_bits =
4684 wait_table_bits(zone->wait_table_hash_nr_entries);
4685 alloc_size = zone->wait_table_hash_nr_entries
4686 * sizeof(wait_queue_head_t);
4688 if (!slab_is_available()) {
4689 zone->wait_table = (wait_queue_head_t *)
4690 memblock_virt_alloc_node_nopanic(
4691 alloc_size, zone->zone_pgdat->node_id);
4694 * This case means that a zone whose size was 0 gets new memory
4695 * via memory hot-add.
4696 * But it may be the case that a new node was hot-added. In
4697 * this case vmalloc() will not be able to use this new node's
4698 * memory - this wait_table must be initialized to use this new
4699 * node itself as well.
4700 * To use this new node's memory, further consideration will be
4703 zone->wait_table = vmalloc(alloc_size);
4705 if (!zone->wait_table)
4708 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4709 init_waitqueue_head(zone->wait_table + i);
4714 static __meminit void zone_pcp_init(struct zone *zone)
4717 * per cpu subsystem is not up at this point. The following code
4718 * relies on the ability of the linker to provide the
4719 * offset of a (static) per cpu variable into the per cpu area.
4721 zone->pageset = &boot_pageset;
4723 if (populated_zone(zone))
4724 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4725 zone->name, zone->present_pages,
4726 zone_batchsize(zone));
4729 int __meminit init_currently_empty_zone(struct zone *zone,
4730 unsigned long zone_start_pfn,
4733 struct pglist_data *pgdat = zone->zone_pgdat;
4735 ret = zone_wait_table_init(zone, size);
4738 pgdat->nr_zones = zone_idx(zone) + 1;
4740 zone->zone_start_pfn = zone_start_pfn;
4742 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4743 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4745 (unsigned long)zone_idx(zone),
4746 zone_start_pfn, (zone_start_pfn + size));
4748 zone_init_free_lists(zone);
4753 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4754 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4757 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4759 int __meminit __early_pfn_to_nid(unsigned long pfn,
4760 struct mminit_pfnnid_cache *state)
4762 unsigned long start_pfn, end_pfn;
4765 if (state->last_start <= pfn && pfn < state->last_end)
4766 return state->last_nid;
4768 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4770 state->last_start = start_pfn;
4771 state->last_end = end_pfn;
4772 state->last_nid = nid;
4777 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4780 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4781 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4782 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4784 * If an architecture guarantees that all ranges registered contain no holes
4785 * and may be freed, this this function may be used instead of calling
4786 * memblock_free_early_nid() manually.
4788 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4790 unsigned long start_pfn, end_pfn;
4793 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4794 start_pfn = min(start_pfn, max_low_pfn);
4795 end_pfn = min(end_pfn, max_low_pfn);
4797 if (start_pfn < end_pfn)
4798 memblock_free_early_nid(PFN_PHYS(start_pfn),
4799 (end_pfn - start_pfn) << PAGE_SHIFT,
4805 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4806 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4808 * If an architecture guarantees that all ranges registered contain no holes and may
4809 * be freed, this function may be used instead of calling memory_present() manually.
4811 void __init sparse_memory_present_with_active_regions(int nid)
4813 unsigned long start_pfn, end_pfn;
4816 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4817 memory_present(this_nid, start_pfn, end_pfn);
4821 * get_pfn_range_for_nid - Return the start and end page frames for a node
4822 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4823 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4824 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4826 * It returns the start and end page frame of a node based on information
4827 * provided by memblock_set_node(). If called for a node
4828 * with no available memory, a warning is printed and the start and end
4831 void __meminit get_pfn_range_for_nid(unsigned int nid,
4832 unsigned long *start_pfn, unsigned long *end_pfn)
4834 unsigned long this_start_pfn, this_end_pfn;
4840 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4841 *start_pfn = min(*start_pfn, this_start_pfn);
4842 *end_pfn = max(*end_pfn, this_end_pfn);
4845 if (*start_pfn == -1UL)
4850 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4851 * assumption is made that zones within a node are ordered in monotonic
4852 * increasing memory addresses so that the "highest" populated zone is used
4854 static void __init find_usable_zone_for_movable(void)
4857 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4858 if (zone_index == ZONE_MOVABLE)
4861 if (arch_zone_highest_possible_pfn[zone_index] >
4862 arch_zone_lowest_possible_pfn[zone_index])
4866 VM_BUG_ON(zone_index == -1);
4867 movable_zone = zone_index;
4871 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4872 * because it is sized independent of architecture. Unlike the other zones,
4873 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4874 * in each node depending on the size of each node and how evenly kernelcore
4875 * is distributed. This helper function adjusts the zone ranges
4876 * provided by the architecture for a given node by using the end of the
4877 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4878 * zones within a node are in order of monotonic increases memory addresses
4880 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4881 unsigned long zone_type,
4882 unsigned long node_start_pfn,
4883 unsigned long node_end_pfn,
4884 unsigned long *zone_start_pfn,
4885 unsigned long *zone_end_pfn)
4887 /* Only adjust if ZONE_MOVABLE is on this node */
4888 if (zone_movable_pfn[nid]) {
4889 /* Size ZONE_MOVABLE */
4890 if (zone_type == ZONE_MOVABLE) {
4891 *zone_start_pfn = zone_movable_pfn[nid];
4892 *zone_end_pfn = min(node_end_pfn,
4893 arch_zone_highest_possible_pfn[movable_zone]);
4895 /* Adjust for ZONE_MOVABLE starting within this range */
4896 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4897 *zone_end_pfn > zone_movable_pfn[nid]) {
4898 *zone_end_pfn = zone_movable_pfn[nid];
4900 /* Check if this whole range is within ZONE_MOVABLE */
4901 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4902 *zone_start_pfn = *zone_end_pfn;
4907 * Return the number of pages a zone spans in a node, including holes
4908 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4910 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4911 unsigned long zone_type,
4912 unsigned long node_start_pfn,
4913 unsigned long node_end_pfn,
4914 unsigned long *ignored)
4916 unsigned long zone_start_pfn, zone_end_pfn;
4918 /* When hotadd a new node from cpu_up(), the node should be empty */
4919 if (!node_start_pfn && !node_end_pfn)
4922 /* Get the start and end of the zone */
4923 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4924 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4925 adjust_zone_range_for_zone_movable(nid, zone_type,
4926 node_start_pfn, node_end_pfn,
4927 &zone_start_pfn, &zone_end_pfn);
4929 /* Check that this node has pages within the zone's required range */
4930 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4933 /* Move the zone boundaries inside the node if necessary */
4934 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4935 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4937 /* Return the spanned pages */
4938 return zone_end_pfn - zone_start_pfn;
4942 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4943 * then all holes in the requested range will be accounted for.
4945 unsigned long __meminit __absent_pages_in_range(int nid,
4946 unsigned long range_start_pfn,
4947 unsigned long range_end_pfn)
4949 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4950 unsigned long start_pfn, end_pfn;
4953 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4954 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4955 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4956 nr_absent -= end_pfn - start_pfn;
4962 * absent_pages_in_range - Return number of page frames in holes within a range
4963 * @start_pfn: The start PFN to start searching for holes
4964 * @end_pfn: The end PFN to stop searching for holes
4966 * It returns the number of pages frames in memory holes within a range.
4968 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4969 unsigned long end_pfn)
4971 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4974 /* Return the number of page frames in holes in a zone on a node */
4975 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4976 unsigned long zone_type,
4977 unsigned long node_start_pfn,
4978 unsigned long node_end_pfn,
4979 unsigned long *ignored)
4981 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4982 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4983 unsigned long zone_start_pfn, zone_end_pfn;
4985 /* When hotadd a new node from cpu_up(), the node should be empty */
4986 if (!node_start_pfn && !node_end_pfn)
4989 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4990 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4992 adjust_zone_range_for_zone_movable(nid, zone_type,
4993 node_start_pfn, node_end_pfn,
4994 &zone_start_pfn, &zone_end_pfn);
4995 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4998 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4999 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5000 unsigned long zone_type,
5001 unsigned long node_start_pfn,
5002 unsigned long node_end_pfn,
5003 unsigned long *zones_size)
5005 return zones_size[zone_type];
5008 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5009 unsigned long zone_type,
5010 unsigned long node_start_pfn,
5011 unsigned long node_end_pfn,
5012 unsigned long *zholes_size)
5017 return zholes_size[zone_type];
5020 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5022 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5023 unsigned long node_start_pfn,
5024 unsigned long node_end_pfn,
5025 unsigned long *zones_size,
5026 unsigned long *zholes_size)
5028 unsigned long realtotalpages = 0, totalpages = 0;
5031 for (i = 0; i < MAX_NR_ZONES; i++) {
5032 struct zone *zone = pgdat->node_zones + i;
5033 unsigned long size, real_size;
5035 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5039 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5040 node_start_pfn, node_end_pfn,
5042 zone->spanned_pages = size;
5043 zone->present_pages = real_size;
5046 realtotalpages += real_size;
5049 pgdat->node_spanned_pages = totalpages;
5050 pgdat->node_present_pages = realtotalpages;
5051 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5055 #ifndef CONFIG_SPARSEMEM
5057 * Calculate the size of the zone->blockflags rounded to an unsigned long
5058 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5059 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5060 * round what is now in bits to nearest long in bits, then return it in
5063 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5065 unsigned long usemapsize;
5067 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5068 usemapsize = roundup(zonesize, pageblock_nr_pages);
5069 usemapsize = usemapsize >> pageblock_order;
5070 usemapsize *= NR_PAGEBLOCK_BITS;
5071 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5073 return usemapsize / 8;
5076 static void __init setup_usemap(struct pglist_data *pgdat,
5078 unsigned long zone_start_pfn,
5079 unsigned long zonesize)
5081 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5082 zone->pageblock_flags = NULL;
5084 zone->pageblock_flags =
5085 memblock_virt_alloc_node_nopanic(usemapsize,
5089 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5090 unsigned long zone_start_pfn, unsigned long zonesize) {}
5091 #endif /* CONFIG_SPARSEMEM */
5093 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5095 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5096 void __paginginit set_pageblock_order(void)
5100 /* Check that pageblock_nr_pages has not already been setup */
5101 if (pageblock_order)
5104 if (HPAGE_SHIFT > PAGE_SHIFT)
5105 order = HUGETLB_PAGE_ORDER;
5107 order = MAX_ORDER - 1;
5110 * Assume the largest contiguous order of interest is a huge page.
5111 * This value may be variable depending on boot parameters on IA64 and
5114 pageblock_order = order;
5116 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5119 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5120 * is unused as pageblock_order is set at compile-time. See
5121 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5124 void __paginginit set_pageblock_order(void)
5128 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5130 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5131 unsigned long present_pages)
5133 unsigned long pages = spanned_pages;
5136 * Provide a more accurate estimation if there are holes within
5137 * the zone and SPARSEMEM is in use. If there are holes within the
5138 * zone, each populated memory region may cost us one or two extra
5139 * memmap pages due to alignment because memmap pages for each
5140 * populated regions may not naturally algined on page boundary.
5141 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5143 if (spanned_pages > present_pages + (present_pages >> 4) &&
5144 IS_ENABLED(CONFIG_SPARSEMEM))
5145 pages = present_pages;
5147 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5151 * Set up the zone data structures:
5152 * - mark all pages reserved
5153 * - mark all memory queues empty
5154 * - clear the memory bitmaps
5156 * NOTE: pgdat should get zeroed by caller.
5158 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5161 int nid = pgdat->node_id;
5162 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5165 pgdat_resize_init(pgdat);
5166 #ifdef CONFIG_NUMA_BALANCING
5167 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5168 pgdat->numabalancing_migrate_nr_pages = 0;
5169 pgdat->numabalancing_migrate_next_window = jiffies;
5171 init_waitqueue_head(&pgdat->kswapd_wait);
5172 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5173 pgdat_page_ext_init(pgdat);
5175 for (j = 0; j < MAX_NR_ZONES; j++) {
5176 struct zone *zone = pgdat->node_zones + j;
5177 unsigned long size, realsize, freesize, memmap_pages;
5179 size = zone->spanned_pages;
5180 realsize = freesize = zone->present_pages;
5183 * Adjust freesize so that it accounts for how much memory
5184 * is used by this zone for memmap. This affects the watermark
5185 * and per-cpu initialisations
5187 memmap_pages = calc_memmap_size(size, realsize);
5188 if (!is_highmem_idx(j)) {
5189 if (freesize >= memmap_pages) {
5190 freesize -= memmap_pages;
5193 " %s zone: %lu pages used for memmap\n",
5194 zone_names[j], memmap_pages);
5197 " %s zone: %lu pages exceeds freesize %lu\n",
5198 zone_names[j], memmap_pages, freesize);
5201 /* Account for reserved pages */
5202 if (j == 0 && freesize > dma_reserve) {
5203 freesize -= dma_reserve;
5204 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5205 zone_names[0], dma_reserve);
5208 if (!is_highmem_idx(j))
5209 nr_kernel_pages += freesize;
5210 /* Charge for highmem memmap if there are enough kernel pages */
5211 else if (nr_kernel_pages > memmap_pages * 2)
5212 nr_kernel_pages -= memmap_pages;
5213 nr_all_pages += freesize;
5216 * Set an approximate value for lowmem here, it will be adjusted
5217 * when the bootmem allocator frees pages into the buddy system.
5218 * And all highmem pages will be managed by the buddy system.
5220 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5223 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5225 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5227 zone->name = zone_names[j];
5228 spin_lock_init(&zone->lock);
5229 spin_lock_init(&zone->lru_lock);
5230 zone_seqlock_init(zone);
5231 zone->zone_pgdat = pgdat;
5232 zone_pcp_init(zone);
5234 /* For bootup, initialized properly in watermark setup */
5235 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5237 lruvec_init(&zone->lruvec);
5241 set_pageblock_order();
5242 setup_usemap(pgdat, zone, zone_start_pfn, size);
5243 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5245 memmap_init(size, nid, j, zone_start_pfn);
5246 zone_start_pfn += size;
5250 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5252 unsigned long __maybe_unused start = 0;
5253 unsigned long __maybe_unused offset = 0;
5255 /* Skip empty nodes */
5256 if (!pgdat->node_spanned_pages)
5259 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5260 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5261 offset = pgdat->node_start_pfn - start;
5262 /* ia64 gets its own node_mem_map, before this, without bootmem */
5263 if (!pgdat->node_mem_map) {
5264 unsigned long size, end;
5268 * The zone's endpoints aren't required to be MAX_ORDER
5269 * aligned but the node_mem_map endpoints must be in order
5270 * for the buddy allocator to function correctly.
5272 end = pgdat_end_pfn(pgdat);
5273 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5274 size = (end - start) * sizeof(struct page);
5275 map = alloc_remap(pgdat->node_id, size);
5277 map = memblock_virt_alloc_node_nopanic(size,
5279 pgdat->node_mem_map = map + offset;
5281 #ifndef CONFIG_NEED_MULTIPLE_NODES
5283 * With no DISCONTIG, the global mem_map is just set as node 0's
5285 if (pgdat == NODE_DATA(0)) {
5286 mem_map = NODE_DATA(0)->node_mem_map;
5287 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5288 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5290 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5293 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5296 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5297 unsigned long node_start_pfn, unsigned long *zholes_size)
5299 pg_data_t *pgdat = NODE_DATA(nid);
5300 unsigned long start_pfn = 0;
5301 unsigned long end_pfn = 0;
5303 /* pg_data_t should be reset to zero when it's allocated */
5304 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5306 reset_deferred_meminit(pgdat);
5307 pgdat->node_id = nid;
5308 pgdat->node_start_pfn = node_start_pfn;
5309 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5310 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5311 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5312 (u64)start_pfn << PAGE_SHIFT,
5313 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5315 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5316 zones_size, zholes_size);
5318 alloc_node_mem_map(pgdat);
5319 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5320 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5321 nid, (unsigned long)pgdat,
5322 (unsigned long)pgdat->node_mem_map);
5325 free_area_init_core(pgdat);
5328 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5330 #if MAX_NUMNODES > 1
5332 * Figure out the number of possible node ids.
5334 void __init setup_nr_node_ids(void)
5336 unsigned int highest;
5338 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5339 nr_node_ids = highest + 1;
5344 * node_map_pfn_alignment - determine the maximum internode alignment
5346 * This function should be called after node map is populated and sorted.
5347 * It calculates the maximum power of two alignment which can distinguish
5350 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5351 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5352 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5353 * shifted, 1GiB is enough and this function will indicate so.
5355 * This is used to test whether pfn -> nid mapping of the chosen memory
5356 * model has fine enough granularity to avoid incorrect mapping for the
5357 * populated node map.
5359 * Returns the determined alignment in pfn's. 0 if there is no alignment
5360 * requirement (single node).
5362 unsigned long __init node_map_pfn_alignment(void)
5364 unsigned long accl_mask = 0, last_end = 0;
5365 unsigned long start, end, mask;
5369 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5370 if (!start || last_nid < 0 || last_nid == nid) {
5377 * Start with a mask granular enough to pin-point to the
5378 * start pfn and tick off bits one-by-one until it becomes
5379 * too coarse to separate the current node from the last.
5381 mask = ~((1 << __ffs(start)) - 1);
5382 while (mask && last_end <= (start & (mask << 1)))
5385 /* accumulate all internode masks */
5389 /* convert mask to number of pages */
5390 return ~accl_mask + 1;
5393 /* Find the lowest pfn for a node */
5394 static unsigned long __init find_min_pfn_for_node(int nid)
5396 unsigned long min_pfn = ULONG_MAX;
5397 unsigned long start_pfn;
5400 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5401 min_pfn = min(min_pfn, start_pfn);
5403 if (min_pfn == ULONG_MAX) {
5405 "Could not find start_pfn for node %d\n", nid);
5413 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5415 * It returns the minimum PFN based on information provided via
5416 * memblock_set_node().
5418 unsigned long __init find_min_pfn_with_active_regions(void)
5420 return find_min_pfn_for_node(MAX_NUMNODES);
5424 * early_calculate_totalpages()
5425 * Sum pages in active regions for movable zone.
5426 * Populate N_MEMORY for calculating usable_nodes.
5428 static unsigned long __init early_calculate_totalpages(void)
5430 unsigned long totalpages = 0;
5431 unsigned long start_pfn, end_pfn;
5434 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5435 unsigned long pages = end_pfn - start_pfn;
5437 totalpages += pages;
5439 node_set_state(nid, N_MEMORY);
5445 * Find the PFN the Movable zone begins in each node. Kernel memory
5446 * is spread evenly between nodes as long as the nodes have enough
5447 * memory. When they don't, some nodes will have more kernelcore than
5450 static void __init find_zone_movable_pfns_for_nodes(void)
5453 unsigned long usable_startpfn;
5454 unsigned long kernelcore_node, kernelcore_remaining;
5455 /* save the state before borrow the nodemask */
5456 nodemask_t saved_node_state = node_states[N_MEMORY];
5457 unsigned long totalpages = early_calculate_totalpages();
5458 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5459 struct memblock_region *r;
5461 /* Need to find movable_zone earlier when movable_node is specified. */
5462 find_usable_zone_for_movable();
5465 * If movable_node is specified, ignore kernelcore and movablecore
5468 if (movable_node_is_enabled()) {
5469 for_each_memblock(memory, r) {
5470 if (!memblock_is_hotpluggable(r))
5475 usable_startpfn = PFN_DOWN(r->base);
5476 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5477 min(usable_startpfn, zone_movable_pfn[nid]) :
5485 * If movablecore=nn[KMG] was specified, calculate what size of
5486 * kernelcore that corresponds so that memory usable for
5487 * any allocation type is evenly spread. If both kernelcore
5488 * and movablecore are specified, then the value of kernelcore
5489 * will be used for required_kernelcore if it's greater than
5490 * what movablecore would have allowed.
5492 if (required_movablecore) {
5493 unsigned long corepages;
5496 * Round-up so that ZONE_MOVABLE is at least as large as what
5497 * was requested by the user
5499 required_movablecore =
5500 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5501 required_movablecore = min(totalpages, required_movablecore);
5502 corepages = totalpages - required_movablecore;
5504 required_kernelcore = max(required_kernelcore, corepages);
5508 * If kernelcore was not specified or kernelcore size is larger
5509 * than totalpages, there is no ZONE_MOVABLE.
5511 if (!required_kernelcore || required_kernelcore >= totalpages)
5514 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5515 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5518 /* Spread kernelcore memory as evenly as possible throughout nodes */
5519 kernelcore_node = required_kernelcore / usable_nodes;
5520 for_each_node_state(nid, N_MEMORY) {
5521 unsigned long start_pfn, end_pfn;
5524 * Recalculate kernelcore_node if the division per node
5525 * now exceeds what is necessary to satisfy the requested
5526 * amount of memory for the kernel
5528 if (required_kernelcore < kernelcore_node)
5529 kernelcore_node = required_kernelcore / usable_nodes;
5532 * As the map is walked, we track how much memory is usable
5533 * by the kernel using kernelcore_remaining. When it is
5534 * 0, the rest of the node is usable by ZONE_MOVABLE
5536 kernelcore_remaining = kernelcore_node;
5538 /* Go through each range of PFNs within this node */
5539 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5540 unsigned long size_pages;
5542 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5543 if (start_pfn >= end_pfn)
5546 /* Account for what is only usable for kernelcore */
5547 if (start_pfn < usable_startpfn) {
5548 unsigned long kernel_pages;
5549 kernel_pages = min(end_pfn, usable_startpfn)
5552 kernelcore_remaining -= min(kernel_pages,
5553 kernelcore_remaining);
5554 required_kernelcore -= min(kernel_pages,
5555 required_kernelcore);
5557 /* Continue if range is now fully accounted */
5558 if (end_pfn <= usable_startpfn) {
5561 * Push zone_movable_pfn to the end so
5562 * that if we have to rebalance
5563 * kernelcore across nodes, we will
5564 * not double account here
5566 zone_movable_pfn[nid] = end_pfn;
5569 start_pfn = usable_startpfn;
5573 * The usable PFN range for ZONE_MOVABLE is from
5574 * start_pfn->end_pfn. Calculate size_pages as the
5575 * number of pages used as kernelcore
5577 size_pages = end_pfn - start_pfn;
5578 if (size_pages > kernelcore_remaining)
5579 size_pages = kernelcore_remaining;
5580 zone_movable_pfn[nid] = start_pfn + size_pages;
5583 * Some kernelcore has been met, update counts and
5584 * break if the kernelcore for this node has been
5587 required_kernelcore -= min(required_kernelcore,
5589 kernelcore_remaining -= size_pages;
5590 if (!kernelcore_remaining)
5596 * If there is still required_kernelcore, we do another pass with one
5597 * less node in the count. This will push zone_movable_pfn[nid] further
5598 * along on the nodes that still have memory until kernelcore is
5602 if (usable_nodes && required_kernelcore > usable_nodes)
5606 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5607 for (nid = 0; nid < MAX_NUMNODES; nid++)
5608 zone_movable_pfn[nid] =
5609 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5612 /* restore the node_state */
5613 node_states[N_MEMORY] = saved_node_state;
5616 /* Any regular or high memory on that node ? */
5617 static void check_for_memory(pg_data_t *pgdat, int nid)
5619 enum zone_type zone_type;
5621 if (N_MEMORY == N_NORMAL_MEMORY)
5624 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5625 struct zone *zone = &pgdat->node_zones[zone_type];
5626 if (populated_zone(zone)) {
5627 node_set_state(nid, N_HIGH_MEMORY);
5628 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5629 zone_type <= ZONE_NORMAL)
5630 node_set_state(nid, N_NORMAL_MEMORY);
5637 * free_area_init_nodes - Initialise all pg_data_t and zone data
5638 * @max_zone_pfn: an array of max PFNs for each zone
5640 * This will call free_area_init_node() for each active node in the system.
5641 * Using the page ranges provided by memblock_set_node(), the size of each
5642 * zone in each node and their holes is calculated. If the maximum PFN
5643 * between two adjacent zones match, it is assumed that the zone is empty.
5644 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5645 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5646 * starts where the previous one ended. For example, ZONE_DMA32 starts
5647 * at arch_max_dma_pfn.
5649 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5651 unsigned long start_pfn, end_pfn;
5654 /* Record where the zone boundaries are */
5655 memset(arch_zone_lowest_possible_pfn, 0,
5656 sizeof(arch_zone_lowest_possible_pfn));
5657 memset(arch_zone_highest_possible_pfn, 0,
5658 sizeof(arch_zone_highest_possible_pfn));
5659 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5660 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5661 for (i = 1; i < MAX_NR_ZONES; i++) {
5662 if (i == ZONE_MOVABLE)
5664 arch_zone_lowest_possible_pfn[i] =
5665 arch_zone_highest_possible_pfn[i-1];
5666 arch_zone_highest_possible_pfn[i] =
5667 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5669 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5670 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5672 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5673 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5674 find_zone_movable_pfns_for_nodes();
5676 /* Print out the zone ranges */
5677 pr_info("Zone ranges:\n");
5678 for (i = 0; i < MAX_NR_ZONES; i++) {
5679 if (i == ZONE_MOVABLE)
5681 pr_info(" %-8s ", zone_names[i]);
5682 if (arch_zone_lowest_possible_pfn[i] ==
5683 arch_zone_highest_possible_pfn[i])
5686 pr_cont("[mem %#018Lx-%#018Lx]\n",
5687 (u64)arch_zone_lowest_possible_pfn[i]
5689 ((u64)arch_zone_highest_possible_pfn[i]
5690 << PAGE_SHIFT) - 1);
5693 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5694 pr_info("Movable zone start for each node\n");
5695 for (i = 0; i < MAX_NUMNODES; i++) {
5696 if (zone_movable_pfn[i])
5697 pr_info(" Node %d: %#018Lx\n", i,
5698 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5701 /* Print out the early node map */
5702 pr_info("Early memory node ranges\n");
5703 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5704 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5705 (u64)start_pfn << PAGE_SHIFT,
5706 ((u64)end_pfn << PAGE_SHIFT) - 1);
5708 /* Initialise every node */
5709 mminit_verify_pageflags_layout();
5710 setup_nr_node_ids();
5711 for_each_online_node(nid) {
5712 pg_data_t *pgdat = NODE_DATA(nid);
5713 free_area_init_node(nid, NULL,
5714 find_min_pfn_for_node(nid), NULL);
5716 /* Any memory on that node */
5717 if (pgdat->node_present_pages)
5718 node_set_state(nid, N_MEMORY);
5719 check_for_memory(pgdat, nid);
5723 static int __init cmdline_parse_core(char *p, unsigned long *core)
5725 unsigned long long coremem;
5729 coremem = memparse(p, &p);
5730 *core = coremem >> PAGE_SHIFT;
5732 /* Paranoid check that UL is enough for the coremem value */
5733 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5739 * kernelcore=size sets the amount of memory for use for allocations that
5740 * cannot be reclaimed or migrated.
5742 static int __init cmdline_parse_kernelcore(char *p)
5744 return cmdline_parse_core(p, &required_kernelcore);
5748 * movablecore=size sets the amount of memory for use for allocations that
5749 * can be reclaimed or migrated.
5751 static int __init cmdline_parse_movablecore(char *p)
5753 return cmdline_parse_core(p, &required_movablecore);
5756 early_param("kernelcore", cmdline_parse_kernelcore);
5757 early_param("movablecore", cmdline_parse_movablecore);
5759 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5761 void adjust_managed_page_count(struct page *page, long count)
5763 spin_lock(&managed_page_count_lock);
5764 page_zone(page)->managed_pages += count;
5765 totalram_pages += count;
5766 #ifdef CONFIG_HIGHMEM
5767 if (PageHighMem(page))
5768 totalhigh_pages += count;
5770 spin_unlock(&managed_page_count_lock);
5772 EXPORT_SYMBOL(adjust_managed_page_count);
5774 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5777 unsigned long pages = 0;
5779 start = (void *)PAGE_ALIGN((unsigned long)start);
5780 end = (void *)((unsigned long)end & PAGE_MASK);
5781 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5782 if ((unsigned int)poison <= 0xFF)
5783 memset(pos, poison, PAGE_SIZE);
5784 free_reserved_page(virt_to_page(pos));
5788 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5789 s, pages << (PAGE_SHIFT - 10), start, end);
5793 EXPORT_SYMBOL(free_reserved_area);
5795 #ifdef CONFIG_HIGHMEM
5796 void free_highmem_page(struct page *page)
5798 __free_reserved_page(page);
5800 page_zone(page)->managed_pages++;
5806 void __init mem_init_print_info(const char *str)
5808 unsigned long physpages, codesize, datasize, rosize, bss_size;
5809 unsigned long init_code_size, init_data_size;
5811 physpages = get_num_physpages();
5812 codesize = _etext - _stext;
5813 datasize = _edata - _sdata;
5814 rosize = __end_rodata - __start_rodata;
5815 bss_size = __bss_stop - __bss_start;
5816 init_data_size = __init_end - __init_begin;
5817 init_code_size = _einittext - _sinittext;
5820 * Detect special cases and adjust section sizes accordingly:
5821 * 1) .init.* may be embedded into .data sections
5822 * 2) .init.text.* may be out of [__init_begin, __init_end],
5823 * please refer to arch/tile/kernel/vmlinux.lds.S.
5824 * 3) .rodata.* may be embedded into .text or .data sections.
5826 #define adj_init_size(start, end, size, pos, adj) \
5828 if (start <= pos && pos < end && size > adj) \
5832 adj_init_size(__init_begin, __init_end, init_data_size,
5833 _sinittext, init_code_size);
5834 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5835 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5836 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5837 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5839 #undef adj_init_size
5841 pr_info("Memory: %luK/%luK available "
5842 "(%luK kernel code, %luK rwdata, %luK rodata, "
5843 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5844 #ifdef CONFIG_HIGHMEM
5848 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5849 codesize >> 10, datasize >> 10, rosize >> 10,
5850 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5851 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5852 totalcma_pages << (PAGE_SHIFT-10),
5853 #ifdef CONFIG_HIGHMEM
5854 totalhigh_pages << (PAGE_SHIFT-10),
5856 str ? ", " : "", str ? str : "");
5860 * set_dma_reserve - set the specified number of pages reserved in the first zone
5861 * @new_dma_reserve: The number of pages to mark reserved
5863 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5864 * In the DMA zone, a significant percentage may be consumed by kernel image
5865 * and other unfreeable allocations which can skew the watermarks badly. This
5866 * function may optionally be used to account for unfreeable pages in the
5867 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5868 * smaller per-cpu batchsize.
5870 void __init set_dma_reserve(unsigned long new_dma_reserve)
5872 dma_reserve = new_dma_reserve;
5875 void __init free_area_init(unsigned long *zones_size)
5877 free_area_init_node(0, zones_size,
5878 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5881 static int page_alloc_cpu_notify(struct notifier_block *self,
5882 unsigned long action, void *hcpu)
5884 int cpu = (unsigned long)hcpu;
5886 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5887 lru_add_drain_cpu(cpu);
5891 * Spill the event counters of the dead processor
5892 * into the current processors event counters.
5893 * This artificially elevates the count of the current
5896 vm_events_fold_cpu(cpu);
5899 * Zero the differential counters of the dead processor
5900 * so that the vm statistics are consistent.
5902 * This is only okay since the processor is dead and cannot
5903 * race with what we are doing.
5905 cpu_vm_stats_fold(cpu);
5910 void __init page_alloc_init(void)
5912 hotcpu_notifier(page_alloc_cpu_notify, 0);
5916 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5917 * or min_free_kbytes changes.
5919 static void calculate_totalreserve_pages(void)
5921 struct pglist_data *pgdat;
5922 unsigned long reserve_pages = 0;
5923 enum zone_type i, j;
5925 for_each_online_pgdat(pgdat) {
5926 for (i = 0; i < MAX_NR_ZONES; i++) {
5927 struct zone *zone = pgdat->node_zones + i;
5930 /* Find valid and maximum lowmem_reserve in the zone */
5931 for (j = i; j < MAX_NR_ZONES; j++) {
5932 if (zone->lowmem_reserve[j] > max)
5933 max = zone->lowmem_reserve[j];
5936 /* we treat the high watermark as reserved pages. */
5937 max += high_wmark_pages(zone);
5939 if (max > zone->managed_pages)
5940 max = zone->managed_pages;
5941 reserve_pages += max;
5943 * Lowmem reserves are not available to
5944 * GFP_HIGHUSER page cache allocations and
5945 * kswapd tries to balance zones to their high
5946 * watermark. As a result, neither should be
5947 * regarded as dirtyable memory, to prevent a
5948 * situation where reclaim has to clean pages
5949 * in order to balance the zones.
5951 zone->dirty_balance_reserve = max;
5954 dirty_balance_reserve = reserve_pages;
5955 totalreserve_pages = reserve_pages;
5959 * setup_per_zone_lowmem_reserve - called whenever
5960 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5961 * has a correct pages reserved value, so an adequate number of
5962 * pages are left in the zone after a successful __alloc_pages().
5964 static void setup_per_zone_lowmem_reserve(void)
5966 struct pglist_data *pgdat;
5967 enum zone_type j, idx;
5969 for_each_online_pgdat(pgdat) {
5970 for (j = 0; j < MAX_NR_ZONES; j++) {
5971 struct zone *zone = pgdat->node_zones + j;
5972 unsigned long managed_pages = zone->managed_pages;
5974 zone->lowmem_reserve[j] = 0;
5978 struct zone *lower_zone;
5982 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5983 sysctl_lowmem_reserve_ratio[idx] = 1;
5985 lower_zone = pgdat->node_zones + idx;
5986 lower_zone->lowmem_reserve[j] = managed_pages /
5987 sysctl_lowmem_reserve_ratio[idx];
5988 managed_pages += lower_zone->managed_pages;
5993 /* update totalreserve_pages */
5994 calculate_totalreserve_pages();
5997 static void __setup_per_zone_wmarks(void)
5999 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6000 unsigned long lowmem_pages = 0;
6002 unsigned long flags;
6004 /* Calculate total number of !ZONE_HIGHMEM pages */
6005 for_each_zone(zone) {
6006 if (!is_highmem(zone))
6007 lowmem_pages += zone->managed_pages;
6010 for_each_zone(zone) {
6013 spin_lock_irqsave(&zone->lock, flags);
6014 tmp = (u64)pages_min * zone->managed_pages;
6015 do_div(tmp, lowmem_pages);
6016 if (is_highmem(zone)) {
6018 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6019 * need highmem pages, so cap pages_min to a small
6022 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6023 * deltas control asynch page reclaim, and so should
6024 * not be capped for highmem.
6026 unsigned long min_pages;
6028 min_pages = zone->managed_pages / 1024;
6029 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6030 zone->watermark[WMARK_MIN] = min_pages;
6033 * If it's a lowmem zone, reserve a number of pages
6034 * proportionate to the zone's size.
6036 zone->watermark[WMARK_MIN] = tmp;
6039 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6040 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6042 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6043 high_wmark_pages(zone) - low_wmark_pages(zone) -
6044 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6046 spin_unlock_irqrestore(&zone->lock, flags);
6049 /* update totalreserve_pages */
6050 calculate_totalreserve_pages();
6054 * setup_per_zone_wmarks - called when min_free_kbytes changes
6055 * or when memory is hot-{added|removed}
6057 * Ensures that the watermark[min,low,high] values for each zone are set
6058 * correctly with respect to min_free_kbytes.
6060 void setup_per_zone_wmarks(void)
6062 mutex_lock(&zonelists_mutex);
6063 __setup_per_zone_wmarks();
6064 mutex_unlock(&zonelists_mutex);
6068 * The inactive anon list should be small enough that the VM never has to
6069 * do too much work, but large enough that each inactive page has a chance
6070 * to be referenced again before it is swapped out.
6072 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6073 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6074 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6075 * the anonymous pages are kept on the inactive list.
6078 * memory ratio inactive anon
6079 * -------------------------------------
6088 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6090 unsigned int gb, ratio;
6092 /* Zone size in gigabytes */
6093 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6095 ratio = int_sqrt(10 * gb);
6099 zone->inactive_ratio = ratio;
6102 static void __meminit setup_per_zone_inactive_ratio(void)
6107 calculate_zone_inactive_ratio(zone);
6111 * Initialise min_free_kbytes.
6113 * For small machines we want it small (128k min). For large machines
6114 * we want it large (64MB max). But it is not linear, because network
6115 * bandwidth does not increase linearly with machine size. We use
6117 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6118 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6134 int __meminit init_per_zone_wmark_min(void)
6136 unsigned long lowmem_kbytes;
6137 int new_min_free_kbytes;
6139 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6140 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6142 if (new_min_free_kbytes > user_min_free_kbytes) {
6143 min_free_kbytes = new_min_free_kbytes;
6144 if (min_free_kbytes < 128)
6145 min_free_kbytes = 128;
6146 if (min_free_kbytes > 65536)
6147 min_free_kbytes = 65536;
6149 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6150 new_min_free_kbytes, user_min_free_kbytes);
6152 setup_per_zone_wmarks();
6153 refresh_zone_stat_thresholds();
6154 setup_per_zone_lowmem_reserve();
6155 setup_per_zone_inactive_ratio();
6158 module_init(init_per_zone_wmark_min)
6161 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6162 * that we can call two helper functions whenever min_free_kbytes
6165 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6166 void __user *buffer, size_t *length, loff_t *ppos)
6170 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6175 user_min_free_kbytes = min_free_kbytes;
6176 setup_per_zone_wmarks();
6182 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6183 void __user *buffer, size_t *length, loff_t *ppos)
6188 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6193 zone->min_unmapped_pages = (zone->managed_pages *
6194 sysctl_min_unmapped_ratio) / 100;
6198 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6199 void __user *buffer, size_t *length, loff_t *ppos)
6204 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6209 zone->min_slab_pages = (zone->managed_pages *
6210 sysctl_min_slab_ratio) / 100;
6216 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6217 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6218 * whenever sysctl_lowmem_reserve_ratio changes.
6220 * The reserve ratio obviously has absolutely no relation with the
6221 * minimum watermarks. The lowmem reserve ratio can only make sense
6222 * if in function of the boot time zone sizes.
6224 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6225 void __user *buffer, size_t *length, loff_t *ppos)
6227 proc_dointvec_minmax(table, write, buffer, length, ppos);
6228 setup_per_zone_lowmem_reserve();
6233 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6234 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6235 * pagelist can have before it gets flushed back to buddy allocator.
6237 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6238 void __user *buffer, size_t *length, loff_t *ppos)
6241 int old_percpu_pagelist_fraction;
6244 mutex_lock(&pcp_batch_high_lock);
6245 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6247 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6248 if (!write || ret < 0)
6251 /* Sanity checking to avoid pcp imbalance */
6252 if (percpu_pagelist_fraction &&
6253 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6254 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6260 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6263 for_each_populated_zone(zone) {
6266 for_each_possible_cpu(cpu)
6267 pageset_set_high_and_batch(zone,
6268 per_cpu_ptr(zone->pageset, cpu));
6271 mutex_unlock(&pcp_batch_high_lock);
6276 int hashdist = HASHDIST_DEFAULT;
6278 static int __init set_hashdist(char *str)
6282 hashdist = simple_strtoul(str, &str, 0);
6285 __setup("hashdist=", set_hashdist);
6289 * allocate a large system hash table from bootmem
6290 * - it is assumed that the hash table must contain an exact power-of-2
6291 * quantity of entries
6292 * - limit is the number of hash buckets, not the total allocation size
6294 void *__init alloc_large_system_hash(const char *tablename,
6295 unsigned long bucketsize,
6296 unsigned long numentries,
6299 unsigned int *_hash_shift,
6300 unsigned int *_hash_mask,
6301 unsigned long low_limit,
6302 unsigned long high_limit)
6304 unsigned long long max = high_limit;
6305 unsigned long log2qty, size;
6308 /* allow the kernel cmdline to have a say */
6310 /* round applicable memory size up to nearest megabyte */
6311 numentries = nr_kernel_pages;
6313 /* It isn't necessary when PAGE_SIZE >= 1MB */
6314 if (PAGE_SHIFT < 20)
6315 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6317 /* limit to 1 bucket per 2^scale bytes of low memory */
6318 if (scale > PAGE_SHIFT)
6319 numentries >>= (scale - PAGE_SHIFT);
6321 numentries <<= (PAGE_SHIFT - scale);
6323 /* Make sure we've got at least a 0-order allocation.. */
6324 if (unlikely(flags & HASH_SMALL)) {
6325 /* Makes no sense without HASH_EARLY */
6326 WARN_ON(!(flags & HASH_EARLY));
6327 if (!(numentries >> *_hash_shift)) {
6328 numentries = 1UL << *_hash_shift;
6329 BUG_ON(!numentries);
6331 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6332 numentries = PAGE_SIZE / bucketsize;
6334 numentries = roundup_pow_of_two(numentries);
6336 /* limit allocation size to 1/16 total memory by default */
6338 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6339 do_div(max, bucketsize);
6341 max = min(max, 0x80000000ULL);
6343 if (numentries < low_limit)
6344 numentries = low_limit;
6345 if (numentries > max)
6348 log2qty = ilog2(numentries);
6351 size = bucketsize << log2qty;
6352 if (flags & HASH_EARLY)
6353 table = memblock_virt_alloc_nopanic(size, 0);
6355 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6358 * If bucketsize is not a power-of-two, we may free
6359 * some pages at the end of hash table which
6360 * alloc_pages_exact() automatically does
6362 if (get_order(size) < MAX_ORDER) {
6363 table = alloc_pages_exact(size, GFP_ATOMIC);
6364 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6367 } while (!table && size > PAGE_SIZE && --log2qty);
6370 panic("Failed to allocate %s hash table\n", tablename);
6372 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6375 ilog2(size) - PAGE_SHIFT,
6379 *_hash_shift = log2qty;
6381 *_hash_mask = (1 << log2qty) - 1;
6386 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6387 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6390 #ifdef CONFIG_SPARSEMEM
6391 return __pfn_to_section(pfn)->pageblock_flags;
6393 return zone->pageblock_flags;
6394 #endif /* CONFIG_SPARSEMEM */
6397 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6399 #ifdef CONFIG_SPARSEMEM
6400 pfn &= (PAGES_PER_SECTION-1);
6401 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6403 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6404 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6405 #endif /* CONFIG_SPARSEMEM */
6409 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6410 * @page: The page within the block of interest
6411 * @pfn: The target page frame number
6412 * @end_bitidx: The last bit of interest to retrieve
6413 * @mask: mask of bits that the caller is interested in
6415 * Return: pageblock_bits flags
6417 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6418 unsigned long end_bitidx,
6422 unsigned long *bitmap;
6423 unsigned long bitidx, word_bitidx;
6426 zone = page_zone(page);
6427 bitmap = get_pageblock_bitmap(zone, pfn);
6428 bitidx = pfn_to_bitidx(zone, pfn);
6429 word_bitidx = bitidx / BITS_PER_LONG;
6430 bitidx &= (BITS_PER_LONG-1);
6432 word = bitmap[word_bitidx];
6433 bitidx += end_bitidx;
6434 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6438 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6439 * @page: The page within the block of interest
6440 * @flags: The flags to set
6441 * @pfn: The target page frame number
6442 * @end_bitidx: The last bit of interest
6443 * @mask: mask of bits that the caller is interested in
6445 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6447 unsigned long end_bitidx,
6451 unsigned long *bitmap;
6452 unsigned long bitidx, word_bitidx;
6453 unsigned long old_word, word;
6455 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6457 zone = page_zone(page);
6458 bitmap = get_pageblock_bitmap(zone, pfn);
6459 bitidx = pfn_to_bitidx(zone, pfn);
6460 word_bitidx = bitidx / BITS_PER_LONG;
6461 bitidx &= (BITS_PER_LONG-1);
6463 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6465 bitidx += end_bitidx;
6466 mask <<= (BITS_PER_LONG - bitidx - 1);
6467 flags <<= (BITS_PER_LONG - bitidx - 1);
6469 word = READ_ONCE(bitmap[word_bitidx]);
6471 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6472 if (word == old_word)
6479 * This function checks whether pageblock includes unmovable pages or not.
6480 * If @count is not zero, it is okay to include less @count unmovable pages
6482 * PageLRU check without isolation or lru_lock could race so that
6483 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6484 * expect this function should be exact.
6486 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6487 bool skip_hwpoisoned_pages)
6489 unsigned long pfn, iter, found;
6493 * For avoiding noise data, lru_add_drain_all() should be called
6494 * If ZONE_MOVABLE, the zone never contains unmovable pages
6496 if (zone_idx(zone) == ZONE_MOVABLE)
6498 mt = get_pageblock_migratetype(page);
6499 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6502 pfn = page_to_pfn(page);
6503 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6504 unsigned long check = pfn + iter;
6506 if (!pfn_valid_within(check))
6509 page = pfn_to_page(check);
6512 * Hugepages are not in LRU lists, but they're movable.
6513 * We need not scan over tail pages bacause we don't
6514 * handle each tail page individually in migration.
6516 if (PageHuge(page)) {
6517 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6522 * We can't use page_count without pin a page
6523 * because another CPU can free compound page.
6524 * This check already skips compound tails of THP
6525 * because their page->_count is zero at all time.
6527 if (!atomic_read(&page->_count)) {
6528 if (PageBuddy(page))
6529 iter += (1 << page_order(page)) - 1;
6534 * The HWPoisoned page may be not in buddy system, and
6535 * page_count() is not 0.
6537 if (skip_hwpoisoned_pages && PageHWPoison(page))
6543 * If there are RECLAIMABLE pages, we need to check
6544 * it. But now, memory offline itself doesn't call
6545 * shrink_node_slabs() and it still to be fixed.
6548 * If the page is not RAM, page_count()should be 0.
6549 * we don't need more check. This is an _used_ not-movable page.
6551 * The problematic thing here is PG_reserved pages. PG_reserved
6552 * is set to both of a memory hole page and a _used_ kernel
6561 bool is_pageblock_removable_nolock(struct page *page)
6567 * We have to be careful here because we are iterating over memory
6568 * sections which are not zone aware so we might end up outside of
6569 * the zone but still within the section.
6570 * We have to take care about the node as well. If the node is offline
6571 * its NODE_DATA will be NULL - see page_zone.
6573 if (!node_online(page_to_nid(page)))
6576 zone = page_zone(page);
6577 pfn = page_to_pfn(page);
6578 if (!zone_spans_pfn(zone, pfn))
6581 return !has_unmovable_pages(zone, page, 0, true);
6586 static unsigned long pfn_max_align_down(unsigned long pfn)
6588 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6589 pageblock_nr_pages) - 1);
6592 static unsigned long pfn_max_align_up(unsigned long pfn)
6594 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6595 pageblock_nr_pages));
6598 /* [start, end) must belong to a single zone. */
6599 static int __alloc_contig_migrate_range(struct compact_control *cc,
6600 unsigned long start, unsigned long end)
6602 /* This function is based on compact_zone() from compaction.c. */
6603 unsigned long nr_reclaimed;
6604 unsigned long pfn = start;
6605 unsigned int tries = 0;
6610 while (pfn < end || !list_empty(&cc->migratepages)) {
6611 if (fatal_signal_pending(current)) {
6616 if (list_empty(&cc->migratepages)) {
6617 cc->nr_migratepages = 0;
6618 pfn = isolate_migratepages_range(cc, pfn, end);
6624 } else if (++tries == 5) {
6625 ret = ret < 0 ? ret : -EBUSY;
6629 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6631 cc->nr_migratepages -= nr_reclaimed;
6633 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6634 NULL, 0, cc->mode, MR_CMA);
6637 putback_movable_pages(&cc->migratepages);
6644 * alloc_contig_range() -- tries to allocate given range of pages
6645 * @start: start PFN to allocate
6646 * @end: one-past-the-last PFN to allocate
6647 * @migratetype: migratetype of the underlaying pageblocks (either
6648 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6649 * in range must have the same migratetype and it must
6650 * be either of the two.
6652 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6653 * aligned, however it's the caller's responsibility to guarantee that
6654 * we are the only thread that changes migrate type of pageblocks the
6657 * The PFN range must belong to a single zone.
6659 * Returns zero on success or negative error code. On success all
6660 * pages which PFN is in [start, end) are allocated for the caller and
6661 * need to be freed with free_contig_range().
6663 int alloc_contig_range(unsigned long start, unsigned long end,
6664 unsigned migratetype)
6666 unsigned long outer_start, outer_end;
6670 struct compact_control cc = {
6671 .nr_migratepages = 0,
6673 .zone = page_zone(pfn_to_page(start)),
6674 .mode = MIGRATE_SYNC,
6675 .ignore_skip_hint = true,
6677 INIT_LIST_HEAD(&cc.migratepages);
6680 * What we do here is we mark all pageblocks in range as
6681 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6682 * have different sizes, and due to the way page allocator
6683 * work, we align the range to biggest of the two pages so
6684 * that page allocator won't try to merge buddies from
6685 * different pageblocks and change MIGRATE_ISOLATE to some
6686 * other migration type.
6688 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6689 * migrate the pages from an unaligned range (ie. pages that
6690 * we are interested in). This will put all the pages in
6691 * range back to page allocator as MIGRATE_ISOLATE.
6693 * When this is done, we take the pages in range from page
6694 * allocator removing them from the buddy system. This way
6695 * page allocator will never consider using them.
6697 * This lets us mark the pageblocks back as
6698 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6699 * aligned range but not in the unaligned, original range are
6700 * put back to page allocator so that buddy can use them.
6703 ret = start_isolate_page_range(pfn_max_align_down(start),
6704 pfn_max_align_up(end), migratetype,
6710 * In case of -EBUSY, we'd like to know which page causes problem.
6711 * So, just fall through. We will check it in test_pages_isolated().
6713 ret = __alloc_contig_migrate_range(&cc, start, end);
6714 if (ret && ret != -EBUSY)
6718 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6719 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6720 * more, all pages in [start, end) are free in page allocator.
6721 * What we are going to do is to allocate all pages from
6722 * [start, end) (that is remove them from page allocator).
6724 * The only problem is that pages at the beginning and at the
6725 * end of interesting range may be not aligned with pages that
6726 * page allocator holds, ie. they can be part of higher order
6727 * pages. Because of this, we reserve the bigger range and
6728 * once this is done free the pages we are not interested in.
6730 * We don't have to hold zone->lock here because the pages are
6731 * isolated thus they won't get removed from buddy.
6734 lru_add_drain_all();
6735 drain_all_pages(cc.zone);
6738 outer_start = start;
6739 while (!PageBuddy(pfn_to_page(outer_start))) {
6740 if (++order >= MAX_ORDER) {
6741 outer_start = start;
6744 outer_start &= ~0UL << order;
6747 if (outer_start != start) {
6748 order = page_order(pfn_to_page(outer_start));
6751 * outer_start page could be small order buddy page and
6752 * it doesn't include start page. Adjust outer_start
6753 * in this case to report failed page properly
6754 * on tracepoint in test_pages_isolated()
6756 if (outer_start + (1UL << order) <= start)
6757 outer_start = start;
6760 /* Make sure the range is really isolated. */
6761 if (test_pages_isolated(outer_start, end, false)) {
6762 pr_info("%s: [%lx, %lx) PFNs busy\n",
6763 __func__, outer_start, end);
6768 /* Grab isolated pages from freelists. */
6769 outer_end = isolate_freepages_range(&cc, outer_start, end);
6775 /* Free head and tail (if any) */
6776 if (start != outer_start)
6777 free_contig_range(outer_start, start - outer_start);
6778 if (end != outer_end)
6779 free_contig_range(end, outer_end - end);
6782 undo_isolate_page_range(pfn_max_align_down(start),
6783 pfn_max_align_up(end), migratetype);
6787 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6789 unsigned int count = 0;
6791 for (; nr_pages--; pfn++) {
6792 struct page *page = pfn_to_page(pfn);
6794 count += page_count(page) != 1;
6797 WARN(count != 0, "%d pages are still in use!\n", count);
6801 #ifdef CONFIG_MEMORY_HOTPLUG
6803 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6804 * page high values need to be recalulated.
6806 void __meminit zone_pcp_update(struct zone *zone)
6809 mutex_lock(&pcp_batch_high_lock);
6810 for_each_possible_cpu(cpu)
6811 pageset_set_high_and_batch(zone,
6812 per_cpu_ptr(zone->pageset, cpu));
6813 mutex_unlock(&pcp_batch_high_lock);
6817 void zone_pcp_reset(struct zone *zone)
6819 unsigned long flags;
6821 struct per_cpu_pageset *pset;
6823 /* avoid races with drain_pages() */
6824 local_irq_save(flags);
6825 if (zone->pageset != &boot_pageset) {
6826 for_each_online_cpu(cpu) {
6827 pset = per_cpu_ptr(zone->pageset, cpu);
6828 drain_zonestat(zone, pset);
6830 free_percpu(zone->pageset);
6831 zone->pageset = &boot_pageset;
6833 local_irq_restore(flags);
6836 #ifdef CONFIG_MEMORY_HOTREMOVE
6838 * All pages in the range must be isolated before calling this.
6841 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6845 unsigned int order, i;
6847 unsigned long flags;
6848 /* find the first valid pfn */
6849 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6854 zone = page_zone(pfn_to_page(pfn));
6855 spin_lock_irqsave(&zone->lock, flags);
6857 while (pfn < end_pfn) {
6858 if (!pfn_valid(pfn)) {
6862 page = pfn_to_page(pfn);
6864 * The HWPoisoned page may be not in buddy system, and
6865 * page_count() is not 0.
6867 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6869 SetPageReserved(page);
6873 BUG_ON(page_count(page));
6874 BUG_ON(!PageBuddy(page));
6875 order = page_order(page);
6876 #ifdef CONFIG_DEBUG_VM
6877 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6878 pfn, 1 << order, end_pfn);
6880 list_del(&page->lru);
6881 rmv_page_order(page);
6882 zone->free_area[order].nr_free--;
6883 for (i = 0; i < (1 << order); i++)
6884 SetPageReserved((page+i));
6885 pfn += (1 << order);
6887 spin_unlock_irqrestore(&zone->lock, flags);
6891 #ifdef CONFIG_MEMORY_FAILURE
6892 bool is_free_buddy_page(struct page *page)
6894 struct zone *zone = page_zone(page);
6895 unsigned long pfn = page_to_pfn(page);
6896 unsigned long flags;
6899 spin_lock_irqsave(&zone->lock, flags);
6900 for (order = 0; order < MAX_ORDER; order++) {
6901 struct page *page_head = page - (pfn & ((1 << order) - 1));
6903 if (PageBuddy(page_head) && page_order(page_head) >= order)
6906 spin_unlock_irqrestore(&zone->lock, flags);
6908 return order < MAX_ORDER;