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
272 * Determine how many pages need to be initialized durig early boot
273 * (non-deferred initialization).
274 * The value of first_deferred_pfn will be set later, once non-deferred pages
275 * are initialized, but for now set it ULONG_MAX.
277 static inline void reset_deferred_meminit(pg_data_t *pgdat)
279 phys_addr_t start_addr, end_addr;
280 unsigned long max_pgcnt;
281 unsigned long reserved;
284 * Initialise at least 2G of a node but also take into account that
285 * two large system hashes that can take up 1GB for 0.25TB/node.
287 max_pgcnt = max(2UL << (30 - PAGE_SHIFT),
288 (pgdat->node_spanned_pages >> 8));
291 * Compensate the all the memblock reservations (e.g. crash kernel)
292 * from the initial estimation to make sure we will initialize enough
295 start_addr = PFN_PHYS(pgdat->node_start_pfn);
296 end_addr = PFN_PHYS(pgdat->node_start_pfn + max_pgcnt);
297 reserved = memblock_reserved_memory_within(start_addr, end_addr);
298 max_pgcnt += PHYS_PFN(reserved);
300 pgdat->static_init_pgcnt = min(max_pgcnt, pgdat->node_spanned_pages);
301 pgdat->first_deferred_pfn = ULONG_MAX;
304 /* Returns true if the struct page for the pfn is uninitialised */
305 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
307 int nid = early_pfn_to_nid(pfn);
309 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
315 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
317 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
324 * Returns false when the remaining initialisation should be deferred until
325 * later in the boot cycle when it can be parallelised.
327 static inline bool update_defer_init(pg_data_t *pgdat,
328 unsigned long pfn, unsigned long zone_end,
329 unsigned long *nr_initialised)
331 /* Always populate low zones for address-contrained allocations */
332 if (zone_end < pgdat_end_pfn(pgdat))
334 /* Initialise at least 2G of the highest zone */
336 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
337 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
338 pgdat->first_deferred_pfn = pfn;
345 static inline void reset_deferred_meminit(pg_data_t *pgdat)
349 static inline bool early_page_uninitialised(unsigned long pfn)
354 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
359 static inline bool update_defer_init(pg_data_t *pgdat,
360 unsigned long pfn, unsigned long zone_end,
361 unsigned long *nr_initialised)
368 void set_pageblock_migratetype(struct page *page, int migratetype)
370 if (unlikely(page_group_by_mobility_disabled &&
371 migratetype < MIGRATE_PCPTYPES))
372 migratetype = MIGRATE_UNMOVABLE;
374 set_pageblock_flags_group(page, (unsigned long)migratetype,
375 PB_migrate, PB_migrate_end);
378 #ifdef CONFIG_DEBUG_VM
379 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
383 unsigned long pfn = page_to_pfn(page);
384 unsigned long sp, start_pfn;
387 seq = zone_span_seqbegin(zone);
388 start_pfn = zone->zone_start_pfn;
389 sp = zone->spanned_pages;
390 if (!zone_spans_pfn(zone, pfn))
392 } while (zone_span_seqretry(zone, seq));
395 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
396 pfn, zone_to_nid(zone), zone->name,
397 start_pfn, start_pfn + sp);
402 static int page_is_consistent(struct zone *zone, struct page *page)
404 if (!pfn_valid_within(page_to_pfn(page)))
406 if (zone != page_zone(page))
412 * Temporary debugging check for pages not lying within a given zone.
414 static int bad_range(struct zone *zone, struct page *page)
416 if (page_outside_zone_boundaries(zone, page))
418 if (!page_is_consistent(zone, page))
424 static inline int bad_range(struct zone *zone, struct page *page)
430 static void bad_page(struct page *page, const char *reason,
431 unsigned long bad_flags)
433 static unsigned long resume;
434 static unsigned long nr_shown;
435 static unsigned long nr_unshown;
437 /* Don't complain about poisoned pages */
438 if (PageHWPoison(page)) {
439 page_mapcount_reset(page); /* remove PageBuddy */
444 * Allow a burst of 60 reports, then keep quiet for that minute;
445 * or allow a steady drip of one report per second.
447 if (nr_shown == 60) {
448 if (time_before(jiffies, resume)) {
454 "BUG: Bad page state: %lu messages suppressed\n",
461 resume = jiffies + 60 * HZ;
463 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
464 current->comm, page_to_pfn(page));
465 dump_page_badflags(page, reason, bad_flags);
470 /* Leave bad fields for debug, except PageBuddy could make trouble */
471 page_mapcount_reset(page); /* remove PageBuddy */
472 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
476 * Higher-order pages are called "compound pages". They are structured thusly:
478 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
480 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
481 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
483 * The first tail page's ->compound_dtor holds the offset in array of compound
484 * page destructors. See compound_page_dtors.
486 * The first tail page's ->compound_order holds the order of allocation.
487 * This usage means that zero-order pages may not be compound.
490 static void free_compound_page(struct page *page)
492 __free_pages_ok(page, compound_order(page));
495 void prep_compound_page(struct page *page, unsigned int order)
498 int nr_pages = 1 << order;
500 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
501 set_compound_order(page, order);
503 for (i = 1; i < nr_pages; i++) {
504 struct page *p = page + i;
505 set_page_count(p, 0);
506 set_compound_head(p, page);
510 #ifdef CONFIG_DEBUG_PAGEALLOC
511 unsigned int _debug_guardpage_minorder;
512 bool _debug_pagealloc_enabled __read_mostly;
513 bool _debug_guardpage_enabled __read_mostly;
515 static int __init early_debug_pagealloc(char *buf)
520 if (strcmp(buf, "on") == 0)
521 _debug_pagealloc_enabled = true;
525 early_param("debug_pagealloc", early_debug_pagealloc);
527 static bool need_debug_guardpage(void)
529 /* If we don't use debug_pagealloc, we don't need guard page */
530 if (!debug_pagealloc_enabled())
536 static void init_debug_guardpage(void)
538 if (!debug_pagealloc_enabled())
541 _debug_guardpage_enabled = true;
544 struct page_ext_operations debug_guardpage_ops = {
545 .need = need_debug_guardpage,
546 .init = init_debug_guardpage,
549 static int __init debug_guardpage_minorder_setup(char *buf)
553 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
554 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
557 _debug_guardpage_minorder = res;
558 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
561 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
563 static inline void set_page_guard(struct zone *zone, struct page *page,
564 unsigned int order, int migratetype)
566 struct page_ext *page_ext;
568 if (!debug_guardpage_enabled())
571 page_ext = lookup_page_ext(page);
572 if (unlikely(!page_ext))
575 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
577 INIT_LIST_HEAD(&page->lru);
578 set_page_private(page, order);
579 /* Guard pages are not available for any usage */
580 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
583 static inline void clear_page_guard(struct zone *zone, struct page *page,
584 unsigned int order, int migratetype)
586 struct page_ext *page_ext;
588 if (!debug_guardpage_enabled())
591 page_ext = lookup_page_ext(page);
592 if (unlikely(!page_ext))
595 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
597 set_page_private(page, 0);
598 if (!is_migrate_isolate(migratetype))
599 __mod_zone_freepage_state(zone, (1 << order), migratetype);
602 struct page_ext_operations debug_guardpage_ops = { NULL, };
603 static inline void set_page_guard(struct zone *zone, struct page *page,
604 unsigned int order, int migratetype) {}
605 static inline void clear_page_guard(struct zone *zone, struct page *page,
606 unsigned int order, int migratetype) {}
609 static inline void set_page_order(struct page *page, unsigned int order)
611 set_page_private(page, order);
612 __SetPageBuddy(page);
615 static inline void rmv_page_order(struct page *page)
617 __ClearPageBuddy(page);
618 set_page_private(page, 0);
622 * This function checks whether a page is free && is the buddy
623 * we can do coalesce a page and its buddy if
624 * (a) the buddy is not in a hole &&
625 * (b) the buddy is in the buddy system &&
626 * (c) a page and its buddy have the same order &&
627 * (d) a page and its buddy are in the same zone.
629 * For recording whether a page is in the buddy system, we set ->_mapcount
630 * PAGE_BUDDY_MAPCOUNT_VALUE.
631 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
632 * serialized by zone->lock.
634 * For recording page's order, we use page_private(page).
636 static inline int page_is_buddy(struct page *page, struct page *buddy,
639 if (!pfn_valid_within(page_to_pfn(buddy)))
642 if (page_is_guard(buddy) && page_order(buddy) == order) {
643 if (page_zone_id(page) != page_zone_id(buddy))
646 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
651 if (PageBuddy(buddy) && page_order(buddy) == order) {
653 * zone check is done late to avoid uselessly
654 * calculating zone/node ids for pages that could
657 if (page_zone_id(page) != page_zone_id(buddy))
660 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
668 * Freeing function for a buddy system allocator.
670 * The concept of a buddy system is to maintain direct-mapped table
671 * (containing bit values) for memory blocks of various "orders".
672 * The bottom level table contains the map for the smallest allocatable
673 * units of memory (here, pages), and each level above it describes
674 * pairs of units from the levels below, hence, "buddies".
675 * At a high level, all that happens here is marking the table entry
676 * at the bottom level available, and propagating the changes upward
677 * as necessary, plus some accounting needed to play nicely with other
678 * parts of the VM system.
679 * At each level, we keep a list of pages, which are heads of continuous
680 * free pages of length of (1 << order) and marked with _mapcount
681 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
683 * So when we are allocating or freeing one, we can derive the state of the
684 * other. That is, if we allocate a small block, and both were
685 * free, the remainder of the region must be split into blocks.
686 * If a block is freed, and its buddy is also free, then this
687 * triggers coalescing into a block of larger size.
692 static inline void __free_one_page(struct page *page,
694 struct zone *zone, unsigned int order,
697 unsigned long page_idx;
698 unsigned long combined_idx;
699 unsigned long uninitialized_var(buddy_idx);
701 unsigned int max_order;
703 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
705 VM_BUG_ON(!zone_is_initialized(zone));
706 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
708 VM_BUG_ON(migratetype == -1);
709 if (likely(!is_migrate_isolate(migratetype)))
710 __mod_zone_freepage_state(zone, 1 << order, migratetype);
712 page_idx = pfn & ((1 << MAX_ORDER) - 1);
714 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
715 VM_BUG_ON_PAGE(bad_range(zone, page), page);
718 while (order < max_order - 1) {
719 buddy_idx = __find_buddy_index(page_idx, order);
720 buddy = page + (buddy_idx - page_idx);
721 if (!page_is_buddy(page, buddy, order))
724 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
725 * merge with it and move up one order.
727 if (page_is_guard(buddy)) {
728 clear_page_guard(zone, buddy, order, migratetype);
730 list_del(&buddy->lru);
731 zone->free_area[order].nr_free--;
732 rmv_page_order(buddy);
734 combined_idx = buddy_idx & page_idx;
735 page = page + (combined_idx - page_idx);
736 page_idx = combined_idx;
739 if (max_order < MAX_ORDER) {
740 /* If we are here, it means order is >= pageblock_order.
741 * We want to prevent merge between freepages on isolate
742 * pageblock and normal pageblock. Without this, pageblock
743 * isolation could cause incorrect freepage or CMA accounting.
745 * We don't want to hit this code for the more frequent
748 if (unlikely(has_isolate_pageblock(zone))) {
751 buddy_idx = __find_buddy_index(page_idx, order);
752 buddy = page + (buddy_idx - page_idx);
753 buddy_mt = get_pageblock_migratetype(buddy);
755 if (migratetype != buddy_mt
756 && (is_migrate_isolate(migratetype) ||
757 is_migrate_isolate(buddy_mt)))
761 goto continue_merging;
765 set_page_order(page, order);
768 * If this is not the largest possible page, check if the buddy
769 * of the next-highest order is free. If it is, it's possible
770 * that pages are being freed that will coalesce soon. In case,
771 * that is happening, add the free page to the tail of the list
772 * so it's less likely to be used soon and more likely to be merged
773 * as a higher order page
775 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
776 struct page *higher_page, *higher_buddy;
777 combined_idx = buddy_idx & page_idx;
778 higher_page = page + (combined_idx - page_idx);
779 buddy_idx = __find_buddy_index(combined_idx, order + 1);
780 higher_buddy = higher_page + (buddy_idx - combined_idx);
781 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
782 list_add_tail(&page->lru,
783 &zone->free_area[order].free_list[migratetype]);
788 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
790 zone->free_area[order].nr_free++;
793 static inline int free_pages_check(struct page *page)
795 const char *bad_reason = NULL;
796 unsigned long bad_flags = 0;
798 if (unlikely(page_mapcount(page)))
799 bad_reason = "nonzero mapcount";
800 if (unlikely(page->mapping != NULL))
801 bad_reason = "non-NULL mapping";
802 if (unlikely(atomic_read(&page->_count) != 0))
803 bad_reason = "nonzero _count";
804 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
805 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
806 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
809 if (unlikely(page->mem_cgroup))
810 bad_reason = "page still charged to cgroup";
812 if (unlikely(bad_reason)) {
813 bad_page(page, bad_reason, bad_flags);
816 page_cpupid_reset_last(page);
817 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
818 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
823 * Frees a number of pages from the PCP lists
824 * Assumes all pages on list are in same zone, and of same order.
825 * count is the number of pages to free.
827 * If the zone was previously in an "all pages pinned" state then look to
828 * see if this freeing clears that state.
830 * And clear the zone's pages_scanned counter, to hold off the "all pages are
831 * pinned" detection logic.
833 static void free_pcppages_bulk(struct zone *zone, int count,
834 struct per_cpu_pages *pcp)
839 unsigned long nr_scanned;
841 spin_lock(&zone->lock);
842 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
844 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
848 struct list_head *list;
851 * Remove pages from lists in a round-robin fashion. A
852 * batch_free count is maintained that is incremented when an
853 * empty list is encountered. This is so more pages are freed
854 * off fuller lists instead of spinning excessively around empty
859 if (++migratetype == MIGRATE_PCPTYPES)
861 list = &pcp->lists[migratetype];
862 } while (list_empty(list));
864 /* This is the only non-empty list. Free them all. */
865 if (batch_free == MIGRATE_PCPTYPES)
866 batch_free = to_free;
869 int mt; /* migratetype of the to-be-freed page */
871 page = list_entry(list->prev, struct page, lru);
872 /* must delete as __free_one_page list manipulates */
873 list_del(&page->lru);
875 mt = get_pcppage_migratetype(page);
876 /* MIGRATE_ISOLATE page should not go to pcplists */
877 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
878 /* Pageblock could have been isolated meanwhile */
879 if (unlikely(has_isolate_pageblock(zone)))
880 mt = get_pageblock_migratetype(page);
882 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
883 trace_mm_page_pcpu_drain(page, 0, mt);
884 } while (--to_free && --batch_free && !list_empty(list));
886 spin_unlock(&zone->lock);
889 static void free_one_page(struct zone *zone,
890 struct page *page, unsigned long pfn,
894 unsigned long nr_scanned;
895 spin_lock(&zone->lock);
896 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
898 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
900 if (unlikely(has_isolate_pageblock(zone) ||
901 is_migrate_isolate(migratetype))) {
902 migratetype = get_pfnblock_migratetype(page, pfn);
904 __free_one_page(page, pfn, zone, order, migratetype);
905 spin_unlock(&zone->lock);
908 static int free_tail_pages_check(struct page *head_page, struct page *page)
913 * We rely page->lru.next never has bit 0 set, unless the page
914 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
916 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
918 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
922 if (unlikely(!PageTail(page))) {
923 bad_page(page, "PageTail not set", 0);
926 if (unlikely(compound_head(page) != head_page)) {
927 bad_page(page, "compound_head not consistent", 0);
932 clear_compound_head(page);
936 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
937 unsigned long zone, int nid)
939 set_page_links(page, zone, nid, pfn);
940 init_page_count(page);
941 page_mapcount_reset(page);
942 page_cpupid_reset_last(page);
944 INIT_LIST_HEAD(&page->lru);
945 #ifdef WANT_PAGE_VIRTUAL
946 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
947 if (!is_highmem_idx(zone))
948 set_page_address(page, __va(pfn << PAGE_SHIFT));
952 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
955 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
958 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
959 static void init_reserved_page(unsigned long pfn)
964 if (!early_page_uninitialised(pfn))
967 nid = early_pfn_to_nid(pfn);
968 pgdat = NODE_DATA(nid);
970 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
971 struct zone *zone = &pgdat->node_zones[zid];
973 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
976 __init_single_pfn(pfn, zid, nid);
979 static inline void init_reserved_page(unsigned long pfn)
982 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
985 * Initialised pages do not have PageReserved set. This function is
986 * called for each range allocated by the bootmem allocator and
987 * marks the pages PageReserved. The remaining valid pages are later
988 * sent to the buddy page allocator.
990 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
992 unsigned long start_pfn = PFN_DOWN(start);
993 unsigned long end_pfn = PFN_UP(end);
995 for (; start_pfn < end_pfn; start_pfn++) {
996 if (pfn_valid(start_pfn)) {
997 struct page *page = pfn_to_page(start_pfn);
999 init_reserved_page(start_pfn);
1001 /* Avoid false-positive PageTail() */
1002 INIT_LIST_HEAD(&page->lru);
1004 SetPageReserved(page);
1009 static bool free_pages_prepare(struct page *page, unsigned int order)
1011 bool compound = PageCompound(page);
1014 VM_BUG_ON_PAGE(PageTail(page), page);
1015 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1017 trace_mm_page_free(page, order);
1018 kmemcheck_free_shadow(page, order);
1019 kasan_free_pages(page, order);
1022 page->mapping = NULL;
1023 bad += free_pages_check(page);
1024 for (i = 1; i < (1 << order); i++) {
1026 bad += free_tail_pages_check(page, page + i);
1027 bad += free_pages_check(page + i);
1032 reset_page_owner(page, order);
1034 if (!PageHighMem(page)) {
1035 debug_check_no_locks_freed(page_address(page),
1036 PAGE_SIZE << order);
1037 debug_check_no_obj_freed(page_address(page),
1038 PAGE_SIZE << order);
1040 arch_free_page(page, order);
1041 kernel_map_pages(page, 1 << order, 0);
1046 static void __free_pages_ok(struct page *page, unsigned int order)
1048 unsigned long flags;
1050 unsigned long pfn = page_to_pfn(page);
1052 if (!free_pages_prepare(page, order))
1055 migratetype = get_pfnblock_migratetype(page, pfn);
1056 local_irq_save(flags);
1057 __count_vm_events(PGFREE, 1 << order);
1058 free_one_page(page_zone(page), page, pfn, order, migratetype);
1059 local_irq_restore(flags);
1062 static void __init __free_pages_boot_core(struct page *page,
1063 unsigned long pfn, unsigned int order)
1065 unsigned int nr_pages = 1 << order;
1066 struct page *p = page;
1070 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1072 __ClearPageReserved(p);
1073 set_page_count(p, 0);
1075 __ClearPageReserved(p);
1076 set_page_count(p, 0);
1078 page_zone(page)->managed_pages += nr_pages;
1079 set_page_refcounted(page);
1080 __free_pages(page, order);
1083 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1084 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1086 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1088 int __meminit early_pfn_to_nid(unsigned long pfn)
1090 static DEFINE_SPINLOCK(early_pfn_lock);
1093 spin_lock(&early_pfn_lock);
1094 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1096 nid = first_online_node;
1097 spin_unlock(&early_pfn_lock);
1103 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1104 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1105 struct mminit_pfnnid_cache *state)
1109 nid = __early_pfn_to_nid(pfn, state);
1110 if (nid >= 0 && nid != node)
1115 /* Only safe to use early in boot when initialisation is single-threaded */
1116 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1118 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1123 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1127 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1128 struct mminit_pfnnid_cache *state)
1135 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1138 if (early_page_uninitialised(pfn))
1140 return __free_pages_boot_core(page, pfn, order);
1143 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1144 static void __init deferred_free_range(struct page *page,
1145 unsigned long pfn, int nr_pages)
1152 /* Free a large naturally-aligned chunk if possible */
1153 if (nr_pages == MAX_ORDER_NR_PAGES &&
1154 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1155 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1156 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1160 for (i = 0; i < nr_pages; i++, page++, pfn++)
1161 __free_pages_boot_core(page, pfn, 0);
1164 /* Completion tracking for deferred_init_memmap() threads */
1165 static atomic_t pgdat_init_n_undone __initdata;
1166 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1168 static inline void __init pgdat_init_report_one_done(void)
1170 if (atomic_dec_and_test(&pgdat_init_n_undone))
1171 complete(&pgdat_init_all_done_comp);
1174 /* Initialise remaining memory on a node */
1175 static int __init deferred_init_memmap(void *data)
1177 pg_data_t *pgdat = data;
1178 int nid = pgdat->node_id;
1179 struct mminit_pfnnid_cache nid_init_state = { };
1180 unsigned long start = jiffies;
1181 unsigned long nr_pages = 0;
1182 unsigned long walk_start, walk_end;
1185 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1186 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1188 if (first_init_pfn == ULONG_MAX) {
1189 pgdat_init_report_one_done();
1193 /* Bind memory initialisation thread to a local node if possible */
1194 if (!cpumask_empty(cpumask))
1195 set_cpus_allowed_ptr(current, cpumask);
1197 /* Sanity check boundaries */
1198 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1199 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1200 pgdat->first_deferred_pfn = ULONG_MAX;
1202 /* Only the highest zone is deferred so find it */
1203 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1204 zone = pgdat->node_zones + zid;
1205 if (first_init_pfn < zone_end_pfn(zone))
1209 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1210 unsigned long pfn, end_pfn;
1211 struct page *page = NULL;
1212 struct page *free_base_page = NULL;
1213 unsigned long free_base_pfn = 0;
1216 end_pfn = min(walk_end, zone_end_pfn(zone));
1217 pfn = first_init_pfn;
1218 if (pfn < walk_start)
1220 if (pfn < zone->zone_start_pfn)
1221 pfn = zone->zone_start_pfn;
1223 for (; pfn < end_pfn; pfn++) {
1224 if (!pfn_valid_within(pfn))
1228 * Ensure pfn_valid is checked every
1229 * MAX_ORDER_NR_PAGES for memory holes
1231 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1232 if (!pfn_valid(pfn)) {
1238 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1243 /* Minimise pfn page lookups and scheduler checks */
1244 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1247 nr_pages += nr_to_free;
1248 deferred_free_range(free_base_page,
1249 free_base_pfn, nr_to_free);
1250 free_base_page = NULL;
1251 free_base_pfn = nr_to_free = 0;
1253 page = pfn_to_page(pfn);
1258 VM_BUG_ON(page_zone(page) != zone);
1262 __init_single_page(page, pfn, zid, nid);
1263 if (!free_base_page) {
1264 free_base_page = page;
1265 free_base_pfn = pfn;
1270 /* Where possible, batch up pages for a single free */
1273 /* Free the current block of pages to allocator */
1274 nr_pages += nr_to_free;
1275 deferred_free_range(free_base_page, free_base_pfn,
1277 free_base_page = NULL;
1278 free_base_pfn = nr_to_free = 0;
1281 first_init_pfn = max(end_pfn, first_init_pfn);
1284 /* Sanity check that the next zone really is unpopulated */
1285 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1287 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1288 jiffies_to_msecs(jiffies - start));
1290 pgdat_init_report_one_done();
1294 void __init page_alloc_init_late(void)
1298 /* There will be num_node_state(N_MEMORY) threads */
1299 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1300 for_each_node_state(nid, N_MEMORY) {
1301 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1304 /* Block until all are initialised */
1305 wait_for_completion(&pgdat_init_all_done_comp);
1307 /* Reinit limits that are based on free pages after the kernel is up */
1308 files_maxfiles_init();
1310 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1313 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1314 void __init init_cma_reserved_pageblock(struct page *page)
1316 unsigned i = pageblock_nr_pages;
1317 struct page *p = page;
1320 __ClearPageReserved(p);
1321 set_page_count(p, 0);
1324 set_pageblock_migratetype(page, MIGRATE_CMA);
1326 if (pageblock_order >= MAX_ORDER) {
1327 i = pageblock_nr_pages;
1330 set_page_refcounted(p);
1331 __free_pages(p, MAX_ORDER - 1);
1332 p += MAX_ORDER_NR_PAGES;
1333 } while (i -= MAX_ORDER_NR_PAGES);
1335 set_page_refcounted(page);
1336 __free_pages(page, pageblock_order);
1339 adjust_managed_page_count(page, pageblock_nr_pages);
1344 * The order of subdivision here is critical for the IO subsystem.
1345 * Please do not alter this order without good reasons and regression
1346 * testing. Specifically, as large blocks of memory are subdivided,
1347 * the order in which smaller blocks are delivered depends on the order
1348 * they're subdivided in this function. This is the primary factor
1349 * influencing the order in which pages are delivered to the IO
1350 * subsystem according to empirical testing, and this is also justified
1351 * by considering the behavior of a buddy system containing a single
1352 * large block of memory acted on by a series of small allocations.
1353 * This behavior is a critical factor in sglist merging's success.
1357 static inline void expand(struct zone *zone, struct page *page,
1358 int low, int high, struct free_area *area,
1361 unsigned long size = 1 << high;
1363 while (high > low) {
1367 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1369 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1370 debug_guardpage_enabled() &&
1371 high < debug_guardpage_minorder()) {
1373 * Mark as guard pages (or page), that will allow to
1374 * merge back to allocator when buddy will be freed.
1375 * Corresponding page table entries will not be touched,
1376 * pages will stay not present in virtual address space
1378 set_page_guard(zone, &page[size], high, migratetype);
1381 list_add(&page[size].lru, &area->free_list[migratetype]);
1383 set_page_order(&page[size], high);
1388 * This page is about to be returned from the page allocator
1390 static inline int check_new_page(struct page *page)
1392 const char *bad_reason = NULL;
1393 unsigned long bad_flags = 0;
1395 if (unlikely(page_mapcount(page)))
1396 bad_reason = "nonzero mapcount";
1397 if (unlikely(page->mapping != NULL))
1398 bad_reason = "non-NULL mapping";
1399 if (unlikely(atomic_read(&page->_count) != 0))
1400 bad_reason = "nonzero _count";
1401 if (unlikely(page->flags & __PG_HWPOISON)) {
1402 bad_reason = "HWPoisoned (hardware-corrupted)";
1403 bad_flags = __PG_HWPOISON;
1405 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1406 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1407 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1410 if (unlikely(page->mem_cgroup))
1411 bad_reason = "page still charged to cgroup";
1413 if (unlikely(bad_reason)) {
1414 bad_page(page, bad_reason, bad_flags);
1420 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1425 for (i = 0; i < (1 << order); i++) {
1426 struct page *p = page + i;
1427 if (unlikely(check_new_page(p)))
1431 set_page_private(page, 0);
1432 set_page_refcounted(page);
1434 arch_alloc_page(page, order);
1435 kernel_map_pages(page, 1 << order, 1);
1436 kasan_alloc_pages(page, order);
1438 if (gfp_flags & __GFP_ZERO)
1439 for (i = 0; i < (1 << order); i++)
1440 clear_highpage(page + i);
1442 if (order && (gfp_flags & __GFP_COMP))
1443 prep_compound_page(page, order);
1445 set_page_owner(page, order, gfp_flags);
1448 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1449 * allocate the page. The expectation is that the caller is taking
1450 * steps that will free more memory. The caller should avoid the page
1451 * being used for !PFMEMALLOC purposes.
1453 if (alloc_flags & ALLOC_NO_WATERMARKS)
1454 set_page_pfmemalloc(page);
1456 clear_page_pfmemalloc(page);
1462 * Go through the free lists for the given migratetype and remove
1463 * the smallest available page from the freelists
1466 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1469 unsigned int current_order;
1470 struct free_area *area;
1473 /* Find a page of the appropriate size in the preferred list */
1474 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1475 area = &(zone->free_area[current_order]);
1476 if (list_empty(&area->free_list[migratetype]))
1479 page = list_entry(area->free_list[migratetype].next,
1481 list_del(&page->lru);
1482 rmv_page_order(page);
1484 expand(zone, page, order, current_order, area, migratetype);
1485 set_pcppage_migratetype(page, migratetype);
1494 * This array describes the order lists are fallen back to when
1495 * the free lists for the desirable migrate type are depleted
1497 static int fallbacks[MIGRATE_TYPES][4] = {
1498 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1499 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1500 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1502 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1504 #ifdef CONFIG_MEMORY_ISOLATION
1505 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1510 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1513 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1516 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1517 unsigned int order) { return NULL; }
1521 * Move the free pages in a range to the free lists of the requested type.
1522 * Note that start_page and end_pages are not aligned on a pageblock
1523 * boundary. If alignment is required, use move_freepages_block()
1525 int move_freepages(struct zone *zone,
1526 struct page *start_page, struct page *end_page,
1531 int pages_moved = 0;
1533 #ifndef CONFIG_HOLES_IN_ZONE
1535 * page_zone is not safe to call in this context when
1536 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1537 * anyway as we check zone boundaries in move_freepages_block().
1538 * Remove at a later date when no bug reports exist related to
1539 * grouping pages by mobility
1541 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1544 for (page = start_page; page <= end_page;) {
1545 if (!pfn_valid_within(page_to_pfn(page))) {
1550 /* Make sure we are not inadvertently changing nodes */
1551 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1553 if (!PageBuddy(page)) {
1558 order = page_order(page);
1559 list_move(&page->lru,
1560 &zone->free_area[order].free_list[migratetype]);
1562 pages_moved += 1 << order;
1568 int move_freepages_block(struct zone *zone, struct page *page,
1571 unsigned long start_pfn, end_pfn;
1572 struct page *start_page, *end_page;
1574 start_pfn = page_to_pfn(page);
1575 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1576 start_page = pfn_to_page(start_pfn);
1577 end_page = start_page + pageblock_nr_pages - 1;
1578 end_pfn = start_pfn + pageblock_nr_pages - 1;
1580 /* Do not cross zone boundaries */
1581 if (!zone_spans_pfn(zone, start_pfn))
1583 if (!zone_spans_pfn(zone, end_pfn))
1586 return move_freepages(zone, start_page, end_page, migratetype);
1589 static void change_pageblock_range(struct page *pageblock_page,
1590 int start_order, int migratetype)
1592 int nr_pageblocks = 1 << (start_order - pageblock_order);
1594 while (nr_pageblocks--) {
1595 set_pageblock_migratetype(pageblock_page, migratetype);
1596 pageblock_page += pageblock_nr_pages;
1601 * When we are falling back to another migratetype during allocation, try to
1602 * steal extra free pages from the same pageblocks to satisfy further
1603 * allocations, instead of polluting multiple pageblocks.
1605 * If we are stealing a relatively large buddy page, it is likely there will
1606 * be more free pages in the pageblock, so try to steal them all. For
1607 * reclaimable and unmovable allocations, we steal regardless of page size,
1608 * as fragmentation caused by those allocations polluting movable pageblocks
1609 * is worse than movable allocations stealing from unmovable and reclaimable
1612 static bool can_steal_fallback(unsigned int order, int start_mt)
1615 * Leaving this order check is intended, although there is
1616 * relaxed order check in next check. The reason is that
1617 * we can actually steal whole pageblock if this condition met,
1618 * but, below check doesn't guarantee it and that is just heuristic
1619 * so could be changed anytime.
1621 if (order >= pageblock_order)
1624 if (order >= pageblock_order / 2 ||
1625 start_mt == MIGRATE_RECLAIMABLE ||
1626 start_mt == MIGRATE_UNMOVABLE ||
1627 page_group_by_mobility_disabled)
1634 * This function implements actual steal behaviour. If order is large enough,
1635 * we can steal whole pageblock. If not, we first move freepages in this
1636 * pageblock and check whether half of pages are moved or not. If half of
1637 * pages are moved, we can change migratetype of pageblock and permanently
1638 * use it's pages as requested migratetype in the future.
1640 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1643 unsigned int current_order = page_order(page);
1646 /* Take ownership for orders >= pageblock_order */
1647 if (current_order >= pageblock_order) {
1648 change_pageblock_range(page, current_order, start_type);
1652 pages = move_freepages_block(zone, page, start_type);
1654 /* Claim the whole block if over half of it is free */
1655 if (pages >= (1 << (pageblock_order-1)) ||
1656 page_group_by_mobility_disabled)
1657 set_pageblock_migratetype(page, start_type);
1661 * Check whether there is a suitable fallback freepage with requested order.
1662 * If only_stealable is true, this function returns fallback_mt only if
1663 * we can steal other freepages all together. This would help to reduce
1664 * fragmentation due to mixed migratetype pages in one pageblock.
1666 int find_suitable_fallback(struct free_area *area, unsigned int order,
1667 int migratetype, bool only_stealable, bool *can_steal)
1672 if (area->nr_free == 0)
1677 fallback_mt = fallbacks[migratetype][i];
1678 if (fallback_mt == MIGRATE_TYPES)
1681 if (list_empty(&area->free_list[fallback_mt]))
1684 if (can_steal_fallback(order, migratetype))
1687 if (!only_stealable)
1698 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1699 * there are no empty page blocks that contain a page with a suitable order
1701 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1702 unsigned int alloc_order)
1705 unsigned long max_managed, flags;
1708 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1709 * Check is race-prone but harmless.
1711 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1712 if (zone->nr_reserved_highatomic >= max_managed)
1715 spin_lock_irqsave(&zone->lock, flags);
1717 /* Recheck the nr_reserved_highatomic limit under the lock */
1718 if (zone->nr_reserved_highatomic >= max_managed)
1722 mt = get_pageblock_migratetype(page);
1723 if (mt != MIGRATE_HIGHATOMIC &&
1724 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1725 zone->nr_reserved_highatomic += pageblock_nr_pages;
1726 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1727 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1731 spin_unlock_irqrestore(&zone->lock, flags);
1735 * Used when an allocation is about to fail under memory pressure. This
1736 * potentially hurts the reliability of high-order allocations when under
1737 * intense memory pressure but failed atomic allocations should be easier
1738 * to recover from than an OOM.
1740 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1742 struct zonelist *zonelist = ac->zonelist;
1743 unsigned long flags;
1749 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1751 /* Preserve at least one pageblock */
1752 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1755 spin_lock_irqsave(&zone->lock, flags);
1756 for (order = 0; order < MAX_ORDER; order++) {
1757 struct free_area *area = &(zone->free_area[order]);
1759 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1762 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1766 * In page freeing path, migratetype change is racy so
1767 * we can counter several free pages in a pageblock
1768 * in this loop althoug we changed the pageblock type
1769 * from highatomic to ac->migratetype. So we should
1770 * adjust the count once.
1772 if (get_pageblock_migratetype(page) ==
1773 MIGRATE_HIGHATOMIC) {
1775 * It should never happen but changes to
1776 * locking could inadvertently allow a per-cpu
1777 * drain to add pages to MIGRATE_HIGHATOMIC
1778 * while unreserving so be safe and watch for
1781 zone->nr_reserved_highatomic -= min(
1783 zone->nr_reserved_highatomic);
1787 * Convert to ac->migratetype and avoid the normal
1788 * pageblock stealing heuristics. Minimally, the caller
1789 * is doing the work and needs the pages. More
1790 * importantly, if the block was always converted to
1791 * MIGRATE_UNMOVABLE or another type then the number
1792 * of pageblocks that cannot be completely freed
1795 set_pageblock_migratetype(page, ac->migratetype);
1796 move_freepages_block(zone, page, ac->migratetype);
1797 spin_unlock_irqrestore(&zone->lock, flags);
1800 spin_unlock_irqrestore(&zone->lock, flags);
1804 /* Remove an element from the buddy allocator from the fallback list */
1805 static inline struct page *
1806 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1808 struct free_area *area;
1809 unsigned int current_order;
1814 /* Find the largest possible block of pages in the other list */
1815 for (current_order = MAX_ORDER-1;
1816 current_order >= order && current_order <= MAX_ORDER-1;
1818 area = &(zone->free_area[current_order]);
1819 fallback_mt = find_suitable_fallback(area, current_order,
1820 start_migratetype, false, &can_steal);
1821 if (fallback_mt == -1)
1824 page = list_entry(area->free_list[fallback_mt].next,
1827 steal_suitable_fallback(zone, page, start_migratetype);
1829 /* Remove the page from the freelists */
1831 list_del(&page->lru);
1832 rmv_page_order(page);
1834 expand(zone, page, order, current_order, area,
1837 * The pcppage_migratetype may differ from pageblock's
1838 * migratetype depending on the decisions in
1839 * find_suitable_fallback(). This is OK as long as it does not
1840 * differ for MIGRATE_CMA pageblocks. Those can be used as
1841 * fallback only via special __rmqueue_cma_fallback() function
1843 set_pcppage_migratetype(page, start_migratetype);
1845 trace_mm_page_alloc_extfrag(page, order, current_order,
1846 start_migratetype, fallback_mt);
1855 * Do the hard work of removing an element from the buddy allocator.
1856 * Call me with the zone->lock already held.
1858 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1859 int migratetype, gfp_t gfp_flags)
1863 page = __rmqueue_smallest(zone, order, migratetype);
1864 if (unlikely(!page)) {
1865 if (migratetype == MIGRATE_MOVABLE)
1866 page = __rmqueue_cma_fallback(zone, order);
1869 page = __rmqueue_fallback(zone, order, migratetype);
1872 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1877 * Obtain a specified number of elements from the buddy allocator, all under
1878 * a single hold of the lock, for efficiency. Add them to the supplied list.
1879 * Returns the number of new pages which were placed at *list.
1881 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1882 unsigned long count, struct list_head *list,
1883 int migratetype, bool cold)
1887 spin_lock(&zone->lock);
1888 for (i = 0; i < count; ++i) {
1889 struct page *page = __rmqueue(zone, order, migratetype, 0);
1890 if (unlikely(page == NULL))
1894 * Split buddy pages returned by expand() are received here
1895 * in physical page order. The page is added to the callers and
1896 * list and the list head then moves forward. From the callers
1897 * perspective, the linked list is ordered by page number in
1898 * some conditions. This is useful for IO devices that can
1899 * merge IO requests if the physical pages are ordered
1903 list_add(&page->lru, list);
1905 list_add_tail(&page->lru, list);
1907 if (is_migrate_cma(get_pcppage_migratetype(page)))
1908 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1911 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1912 spin_unlock(&zone->lock);
1918 * Called from the vmstat counter updater to drain pagesets of this
1919 * currently executing processor on remote nodes after they have
1922 * Note that this function must be called with the thread pinned to
1923 * a single processor.
1925 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1927 unsigned long flags;
1928 int to_drain, batch;
1930 local_irq_save(flags);
1931 batch = READ_ONCE(pcp->batch);
1932 to_drain = min(pcp->count, batch);
1934 free_pcppages_bulk(zone, to_drain, pcp);
1935 pcp->count -= to_drain;
1937 local_irq_restore(flags);
1942 * Drain pcplists of the indicated processor and zone.
1944 * The processor must either be the current processor and the
1945 * thread pinned to the current processor or a processor that
1948 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1950 unsigned long flags;
1951 struct per_cpu_pageset *pset;
1952 struct per_cpu_pages *pcp;
1954 local_irq_save(flags);
1955 pset = per_cpu_ptr(zone->pageset, cpu);
1959 free_pcppages_bulk(zone, pcp->count, pcp);
1962 local_irq_restore(flags);
1966 * Drain pcplists of all zones on the indicated processor.
1968 * The processor must either be the current processor and the
1969 * thread pinned to the current processor or a processor that
1972 static void drain_pages(unsigned int cpu)
1976 for_each_populated_zone(zone) {
1977 drain_pages_zone(cpu, zone);
1982 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1984 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1985 * the single zone's pages.
1987 void drain_local_pages(struct zone *zone)
1989 int cpu = smp_processor_id();
1992 drain_pages_zone(cpu, zone);
1998 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2000 * When zone parameter is non-NULL, spill just the single zone's pages.
2002 * Note that this code is protected against sending an IPI to an offline
2003 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2004 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2005 * nothing keeps CPUs from showing up after we populated the cpumask and
2006 * before the call to on_each_cpu_mask().
2008 void drain_all_pages(struct zone *zone)
2013 * Allocate in the BSS so we wont require allocation in
2014 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2016 static cpumask_t cpus_with_pcps;
2019 * We don't care about racing with CPU hotplug event
2020 * as offline notification will cause the notified
2021 * cpu to drain that CPU pcps and on_each_cpu_mask
2022 * disables preemption as part of its processing
2024 for_each_online_cpu(cpu) {
2025 struct per_cpu_pageset *pcp;
2027 bool has_pcps = false;
2030 pcp = per_cpu_ptr(zone->pageset, cpu);
2034 for_each_populated_zone(z) {
2035 pcp = per_cpu_ptr(z->pageset, cpu);
2036 if (pcp->pcp.count) {
2044 cpumask_set_cpu(cpu, &cpus_with_pcps);
2046 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2048 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2052 #ifdef CONFIG_HIBERNATION
2054 void mark_free_pages(struct zone *zone)
2056 unsigned long pfn, max_zone_pfn;
2057 unsigned long flags;
2058 unsigned int order, t;
2059 struct list_head *curr;
2061 if (zone_is_empty(zone))
2064 spin_lock_irqsave(&zone->lock, flags);
2066 max_zone_pfn = zone_end_pfn(zone);
2067 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2068 if (pfn_valid(pfn)) {
2069 struct page *page = pfn_to_page(pfn);
2071 if (!swsusp_page_is_forbidden(page))
2072 swsusp_unset_page_free(page);
2075 for_each_migratetype_order(order, t) {
2076 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2079 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2080 for (i = 0; i < (1UL << order); i++)
2081 swsusp_set_page_free(pfn_to_page(pfn + i));
2084 spin_unlock_irqrestore(&zone->lock, flags);
2086 #endif /* CONFIG_PM */
2089 * Free a 0-order page
2090 * cold == true ? free a cold page : free a hot page
2092 void free_hot_cold_page(struct page *page, bool cold)
2094 struct zone *zone = page_zone(page);
2095 struct per_cpu_pages *pcp;
2096 unsigned long flags;
2097 unsigned long pfn = page_to_pfn(page);
2100 if (!free_pages_prepare(page, 0))
2103 migratetype = get_pfnblock_migratetype(page, pfn);
2104 set_pcppage_migratetype(page, migratetype);
2105 local_irq_save(flags);
2106 __count_vm_event(PGFREE);
2109 * We only track unmovable, reclaimable and movable on pcp lists.
2110 * Free ISOLATE pages back to the allocator because they are being
2111 * offlined but treat RESERVE as movable pages so we can get those
2112 * areas back if necessary. Otherwise, we may have to free
2113 * excessively into the page allocator
2115 if (migratetype >= MIGRATE_PCPTYPES) {
2116 if (unlikely(is_migrate_isolate(migratetype))) {
2117 free_one_page(zone, page, pfn, 0, migratetype);
2120 migratetype = MIGRATE_MOVABLE;
2123 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2125 list_add(&page->lru, &pcp->lists[migratetype]);
2127 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2129 if (pcp->count >= pcp->high) {
2130 unsigned long batch = READ_ONCE(pcp->batch);
2131 free_pcppages_bulk(zone, batch, pcp);
2132 pcp->count -= batch;
2136 local_irq_restore(flags);
2140 * Free a list of 0-order pages
2142 void free_hot_cold_page_list(struct list_head *list, bool cold)
2144 struct page *page, *next;
2146 list_for_each_entry_safe(page, next, list, lru) {
2147 trace_mm_page_free_batched(page, cold);
2148 free_hot_cold_page(page, cold);
2153 * split_page takes a non-compound higher-order page, and splits it into
2154 * n (1<<order) sub-pages: page[0..n]
2155 * Each sub-page must be freed individually.
2157 * Note: this is probably too low level an operation for use in drivers.
2158 * Please consult with lkml before using this in your driver.
2160 void split_page(struct page *page, unsigned int order)
2165 VM_BUG_ON_PAGE(PageCompound(page), page);
2166 VM_BUG_ON_PAGE(!page_count(page), page);
2168 #ifdef CONFIG_KMEMCHECK
2170 * Split shadow pages too, because free(page[0]) would
2171 * otherwise free the whole shadow.
2173 if (kmemcheck_page_is_tracked(page))
2174 split_page(virt_to_page(page[0].shadow), order);
2177 gfp_mask = get_page_owner_gfp(page);
2178 set_page_owner(page, 0, gfp_mask);
2179 for (i = 1; i < (1 << order); i++) {
2180 set_page_refcounted(page + i);
2181 set_page_owner(page + i, 0, gfp_mask);
2184 EXPORT_SYMBOL_GPL(split_page);
2186 int __isolate_free_page(struct page *page, unsigned int order)
2188 unsigned long watermark;
2192 BUG_ON(!PageBuddy(page));
2194 zone = page_zone(page);
2195 mt = get_pageblock_migratetype(page);
2197 if (!is_migrate_isolate(mt)) {
2198 /* Obey watermarks as if the page was being allocated */
2199 watermark = low_wmark_pages(zone) + (1 << order);
2200 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2203 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2206 /* Remove page from free list */
2207 list_del(&page->lru);
2208 zone->free_area[order].nr_free--;
2209 rmv_page_order(page);
2211 set_page_owner(page, order, __GFP_MOVABLE);
2213 /* Set the pageblock if the isolated page is at least a pageblock */
2214 if (order >= pageblock_order - 1) {
2215 struct page *endpage = page + (1 << order) - 1;
2216 for (; page < endpage; page += pageblock_nr_pages) {
2217 int mt = get_pageblock_migratetype(page);
2218 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2219 set_pageblock_migratetype(page,
2225 return 1UL << order;
2229 * Similar to split_page except the page is already free. As this is only
2230 * being used for migration, the migratetype of the block also changes.
2231 * As this is called with interrupts disabled, the caller is responsible
2232 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2235 * Note: this is probably too low level an operation for use in drivers.
2236 * Please consult with lkml before using this in your driver.
2238 int split_free_page(struct page *page)
2243 order = page_order(page);
2245 nr_pages = __isolate_free_page(page, order);
2249 /* Split into individual pages */
2250 set_page_refcounted(page);
2251 split_page(page, order);
2256 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2259 struct page *buffered_rmqueue(struct zone *preferred_zone,
2260 struct zone *zone, unsigned int order,
2261 gfp_t gfp_flags, int alloc_flags, int migratetype)
2263 unsigned long flags;
2265 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2267 if (likely(order == 0)) {
2268 struct per_cpu_pages *pcp;
2269 struct list_head *list;
2271 local_irq_save(flags);
2272 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2273 list = &pcp->lists[migratetype];
2274 if (list_empty(list)) {
2275 pcp->count += rmqueue_bulk(zone, 0,
2278 if (unlikely(list_empty(list)))
2283 page = list_entry(list->prev, struct page, lru);
2285 page = list_entry(list->next, struct page, lru);
2287 list_del(&page->lru);
2290 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2292 * __GFP_NOFAIL is not to be used in new code.
2294 * All __GFP_NOFAIL callers should be fixed so that they
2295 * properly detect and handle allocation failures.
2297 * We most definitely don't want callers attempting to
2298 * allocate greater than order-1 page units with
2301 WARN_ON_ONCE(order > 1);
2303 spin_lock_irqsave(&zone->lock, flags);
2306 if (alloc_flags & ALLOC_HARDER) {
2307 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2309 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2312 page = __rmqueue(zone, order, migratetype, gfp_flags);
2313 spin_unlock(&zone->lock);
2316 __mod_zone_freepage_state(zone, -(1 << order),
2317 get_pcppage_migratetype(page));
2320 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2321 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2322 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2323 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2325 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2326 zone_statistics(preferred_zone, zone, gfp_flags);
2327 local_irq_restore(flags);
2329 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2333 local_irq_restore(flags);
2337 #ifdef CONFIG_FAIL_PAGE_ALLOC
2340 struct fault_attr attr;
2342 bool ignore_gfp_highmem;
2343 bool ignore_gfp_reclaim;
2345 } fail_page_alloc = {
2346 .attr = FAULT_ATTR_INITIALIZER,
2347 .ignore_gfp_reclaim = true,
2348 .ignore_gfp_highmem = true,
2352 static int __init setup_fail_page_alloc(char *str)
2354 return setup_fault_attr(&fail_page_alloc.attr, str);
2356 __setup("fail_page_alloc=", setup_fail_page_alloc);
2358 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2360 if (order < fail_page_alloc.min_order)
2362 if (gfp_mask & __GFP_NOFAIL)
2364 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2366 if (fail_page_alloc.ignore_gfp_reclaim &&
2367 (gfp_mask & __GFP_DIRECT_RECLAIM))
2370 return should_fail(&fail_page_alloc.attr, 1 << order);
2373 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2375 static int __init fail_page_alloc_debugfs(void)
2377 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2380 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2381 &fail_page_alloc.attr);
2383 return PTR_ERR(dir);
2385 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2386 &fail_page_alloc.ignore_gfp_reclaim))
2388 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2389 &fail_page_alloc.ignore_gfp_highmem))
2391 if (!debugfs_create_u32("min-order", mode, dir,
2392 &fail_page_alloc.min_order))
2397 debugfs_remove_recursive(dir);
2402 late_initcall(fail_page_alloc_debugfs);
2404 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2406 #else /* CONFIG_FAIL_PAGE_ALLOC */
2408 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2413 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2416 * Return true if free base pages are above 'mark'. For high-order checks it
2417 * will return true of the order-0 watermark is reached and there is at least
2418 * one free page of a suitable size. Checking now avoids taking the zone lock
2419 * to check in the allocation paths if no pages are free.
2421 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2422 unsigned long mark, int classzone_idx, int alloc_flags,
2427 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2429 /* free_pages may go negative - that's OK */
2430 free_pages -= (1 << order) - 1;
2432 if (alloc_flags & ALLOC_HIGH)
2436 * If the caller does not have rights to ALLOC_HARDER then subtract
2437 * the high-atomic reserves. This will over-estimate the size of the
2438 * atomic reserve but it avoids a search.
2440 if (likely(!alloc_harder))
2441 free_pages -= z->nr_reserved_highatomic;
2446 /* If allocation can't use CMA areas don't use free CMA pages */
2447 if (!(alloc_flags & ALLOC_CMA))
2448 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2452 * Check watermarks for an order-0 allocation request. If these
2453 * are not met, then a high-order request also cannot go ahead
2454 * even if a suitable page happened to be free.
2456 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2459 /* If this is an order-0 request then the watermark is fine */
2463 /* For a high-order request, check at least one suitable page is free */
2464 for (o = order; o < MAX_ORDER; o++) {
2465 struct free_area *area = &z->free_area[o];
2471 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2472 if (!list_empty(&area->free_list[mt]))
2477 if ((alloc_flags & ALLOC_CMA) &&
2478 !list_empty(&area->free_list[MIGRATE_CMA])) {
2483 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
2489 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2490 int classzone_idx, int alloc_flags)
2492 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2493 zone_page_state(z, NR_FREE_PAGES));
2496 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2497 unsigned long mark, int classzone_idx)
2499 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2501 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2502 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2504 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2509 static bool zone_local(struct zone *local_zone, struct zone *zone)
2511 return local_zone->node == zone->node;
2514 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2516 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2519 #else /* CONFIG_NUMA */
2520 static bool zone_local(struct zone *local_zone, struct zone *zone)
2525 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2529 #endif /* CONFIG_NUMA */
2531 static void reset_alloc_batches(struct zone *preferred_zone)
2533 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2536 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2537 high_wmark_pages(zone) - low_wmark_pages(zone) -
2538 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2539 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2540 } while (zone++ != preferred_zone);
2544 * get_page_from_freelist goes through the zonelist trying to allocate
2547 static struct page *
2548 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2549 const struct alloc_context *ac)
2551 struct zonelist *zonelist = ac->zonelist;
2553 struct page *page = NULL;
2555 int nr_fair_skipped = 0;
2556 bool zonelist_rescan;
2559 zonelist_rescan = false;
2562 * Scan zonelist, looking for a zone with enough free.
2563 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2565 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2569 if (cpusets_enabled() &&
2570 (alloc_flags & ALLOC_CPUSET) &&
2571 !cpuset_zone_allowed(zone, gfp_mask))
2574 * Distribute pages in proportion to the individual
2575 * zone size to ensure fair page aging. The zone a
2576 * page was allocated in should have no effect on the
2577 * time the page has in memory before being reclaimed.
2579 if (alloc_flags & ALLOC_FAIR) {
2580 if (!zone_local(ac->preferred_zone, zone))
2582 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2588 * When allocating a page cache page for writing, we
2589 * want to get it from a zone that is within its dirty
2590 * limit, such that no single zone holds more than its
2591 * proportional share of globally allowed dirty pages.
2592 * The dirty limits take into account the zone's
2593 * lowmem reserves and high watermark so that kswapd
2594 * should be able to balance it without having to
2595 * write pages from its LRU list.
2597 * This may look like it could increase pressure on
2598 * lower zones by failing allocations in higher zones
2599 * before they are full. But the pages that do spill
2600 * over are limited as the lower zones are protected
2601 * by this very same mechanism. It should not become
2602 * a practical burden to them.
2604 * XXX: For now, allow allocations to potentially
2605 * exceed the per-zone dirty limit in the slowpath
2606 * (spread_dirty_pages unset) before going into reclaim,
2607 * which is important when on a NUMA setup the allowed
2608 * zones are together not big enough to reach the
2609 * global limit. The proper fix for these situations
2610 * will require awareness of zones in the
2611 * dirty-throttling and the flusher threads.
2613 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2616 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2617 if (!zone_watermark_ok(zone, order, mark,
2618 ac->classzone_idx, alloc_flags)) {
2621 /* Checked here to keep the fast path fast */
2622 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2623 if (alloc_flags & ALLOC_NO_WATERMARKS)
2626 if (zone_reclaim_mode == 0 ||
2627 !zone_allows_reclaim(ac->preferred_zone, zone))
2630 ret = zone_reclaim(zone, gfp_mask, order);
2632 case ZONE_RECLAIM_NOSCAN:
2635 case ZONE_RECLAIM_FULL:
2636 /* scanned but unreclaimable */
2639 /* did we reclaim enough */
2640 if (zone_watermark_ok(zone, order, mark,
2641 ac->classzone_idx, alloc_flags))
2649 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2650 gfp_mask, alloc_flags, ac->migratetype);
2652 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2656 * If this is a high-order atomic allocation then check
2657 * if the pageblock should be reserved for the future
2659 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2660 reserve_highatomic_pageblock(page, zone, order);
2667 * The first pass makes sure allocations are spread fairly within the
2668 * local node. However, the local node might have free pages left
2669 * after the fairness batches are exhausted, and remote zones haven't
2670 * even been considered yet. Try once more without fairness, and
2671 * include remote zones now, before entering the slowpath and waking
2672 * kswapd: prefer spilling to a remote zone over swapping locally.
2674 if (alloc_flags & ALLOC_FAIR) {
2675 alloc_flags &= ~ALLOC_FAIR;
2676 if (nr_fair_skipped) {
2677 zonelist_rescan = true;
2678 reset_alloc_batches(ac->preferred_zone);
2680 if (nr_online_nodes > 1)
2681 zonelist_rescan = true;
2684 if (zonelist_rescan)
2691 * Large machines with many possible nodes should not always dump per-node
2692 * meminfo in irq context.
2694 static inline bool should_suppress_show_mem(void)
2699 ret = in_interrupt();
2704 static DEFINE_RATELIMIT_STATE(nopage_rs,
2705 DEFAULT_RATELIMIT_INTERVAL,
2706 DEFAULT_RATELIMIT_BURST);
2708 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2710 unsigned int filter = SHOW_MEM_FILTER_NODES;
2712 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2713 debug_guardpage_minorder() > 0)
2717 * This documents exceptions given to allocations in certain
2718 * contexts that are allowed to allocate outside current's set
2721 if (!(gfp_mask & __GFP_NOMEMALLOC))
2722 if (test_thread_flag(TIF_MEMDIE) ||
2723 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2724 filter &= ~SHOW_MEM_FILTER_NODES;
2725 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2726 filter &= ~SHOW_MEM_FILTER_NODES;
2729 struct va_format vaf;
2732 va_start(args, fmt);
2737 pr_warn("%pV", &vaf);
2742 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2743 current->comm, order, gfp_mask);
2746 if (!should_suppress_show_mem())
2750 static inline struct page *
2751 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2752 const struct alloc_context *ac, unsigned long *did_some_progress)
2754 struct oom_control oc = {
2755 .zonelist = ac->zonelist,
2756 .nodemask = ac->nodemask,
2757 .gfp_mask = gfp_mask,
2762 *did_some_progress = 0;
2765 * Acquire the oom lock. If that fails, somebody else is
2766 * making progress for us.
2768 if (!mutex_trylock(&oom_lock)) {
2769 *did_some_progress = 1;
2770 schedule_timeout_uninterruptible(1);
2775 * Go through the zonelist yet one more time, keep very high watermark
2776 * here, this is only to catch a parallel oom killing, we must fail if
2777 * we're still under heavy pressure.
2779 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2780 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2784 if (!(gfp_mask & __GFP_NOFAIL)) {
2785 /* Coredumps can quickly deplete all memory reserves */
2786 if (current->flags & PF_DUMPCORE)
2788 /* The OOM killer will not help higher order allocs */
2789 if (order > PAGE_ALLOC_COSTLY_ORDER)
2791 /* The OOM killer does not needlessly kill tasks for lowmem */
2792 if (ac->high_zoneidx < ZONE_NORMAL)
2794 /* The OOM killer does not compensate for IO-less reclaim */
2795 if (!(gfp_mask & __GFP_FS)) {
2797 * XXX: Page reclaim didn't yield anything,
2798 * and the OOM killer can't be invoked, but
2799 * keep looping as per tradition.
2801 *did_some_progress = 1;
2804 if (pm_suspended_storage())
2806 /* The OOM killer may not free memory on a specific node */
2807 if (gfp_mask & __GFP_THISNODE)
2810 /* Exhausted what can be done so it's blamo time */
2811 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2812 *did_some_progress = 1;
2814 mutex_unlock(&oom_lock);
2818 #ifdef CONFIG_COMPACTION
2819 /* Try memory compaction for high-order allocations before reclaim */
2820 static struct page *
2821 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2822 int alloc_flags, const struct alloc_context *ac,
2823 enum migrate_mode mode, int *contended_compaction,
2824 bool *deferred_compaction)
2826 unsigned long compact_result;
2832 current->flags |= PF_MEMALLOC;
2833 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2834 mode, contended_compaction);
2835 current->flags &= ~PF_MEMALLOC;
2837 switch (compact_result) {
2838 case COMPACT_DEFERRED:
2839 *deferred_compaction = true;
2841 case COMPACT_SKIPPED:
2848 * At least in one zone compaction wasn't deferred or skipped, so let's
2849 * count a compaction stall
2851 count_vm_event(COMPACTSTALL);
2853 page = get_page_from_freelist(gfp_mask, order,
2854 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2857 struct zone *zone = page_zone(page);
2859 zone->compact_blockskip_flush = false;
2860 compaction_defer_reset(zone, order, true);
2861 count_vm_event(COMPACTSUCCESS);
2866 * It's bad if compaction run occurs and fails. The most likely reason
2867 * is that pages exist, but not enough to satisfy watermarks.
2869 count_vm_event(COMPACTFAIL);
2876 static inline struct page *
2877 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2878 int alloc_flags, const struct alloc_context *ac,
2879 enum migrate_mode mode, int *contended_compaction,
2880 bool *deferred_compaction)
2884 #endif /* CONFIG_COMPACTION */
2886 /* Perform direct synchronous page reclaim */
2888 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2889 const struct alloc_context *ac)
2891 struct reclaim_state reclaim_state;
2896 /* We now go into synchronous reclaim */
2897 cpuset_memory_pressure_bump();
2898 current->flags |= PF_MEMALLOC;
2899 lockdep_set_current_reclaim_state(gfp_mask);
2900 reclaim_state.reclaimed_slab = 0;
2901 current->reclaim_state = &reclaim_state;
2903 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2906 current->reclaim_state = NULL;
2907 lockdep_clear_current_reclaim_state();
2908 current->flags &= ~PF_MEMALLOC;
2915 /* The really slow allocator path where we enter direct reclaim */
2916 static inline struct page *
2917 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2918 int alloc_flags, const struct alloc_context *ac,
2919 unsigned long *did_some_progress)
2921 struct page *page = NULL;
2922 bool drained = false;
2924 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2925 if (unlikely(!(*did_some_progress)))
2929 page = get_page_from_freelist(gfp_mask, order,
2930 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2933 * If an allocation failed after direct reclaim, it could be because
2934 * pages are pinned on the per-cpu lists or in high alloc reserves.
2935 * Shrink them them and try again
2937 if (!page && !drained) {
2938 unreserve_highatomic_pageblock(ac);
2939 drain_all_pages(NULL);
2948 * This is called in the allocator slow-path if the allocation request is of
2949 * sufficient urgency to ignore watermarks and take other desperate measures
2951 static inline struct page *
2952 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2953 const struct alloc_context *ac)
2958 page = get_page_from_freelist(gfp_mask, order,
2959 ALLOC_NO_WATERMARKS, ac);
2961 if (!page && gfp_mask & __GFP_NOFAIL)
2962 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2964 } while (!page && (gfp_mask & __GFP_NOFAIL));
2969 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2974 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2975 ac->high_zoneidx, ac->nodemask)
2976 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2980 gfp_to_alloc_flags(gfp_t gfp_mask)
2982 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2984 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2985 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2988 * The caller may dip into page reserves a bit more if the caller
2989 * cannot run direct reclaim, or if the caller has realtime scheduling
2990 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2991 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2993 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2995 if (gfp_mask & __GFP_ATOMIC) {
2997 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2998 * if it can't schedule.
3000 if (!(gfp_mask & __GFP_NOMEMALLOC))
3001 alloc_flags |= ALLOC_HARDER;
3003 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3004 * comment for __cpuset_node_allowed().
3006 alloc_flags &= ~ALLOC_CPUSET;
3007 } else if (unlikely(rt_task(current)) && !in_interrupt())
3008 alloc_flags |= ALLOC_HARDER;
3010 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3011 if (gfp_mask & __GFP_MEMALLOC)
3012 alloc_flags |= ALLOC_NO_WATERMARKS;
3013 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3014 alloc_flags |= ALLOC_NO_WATERMARKS;
3015 else if (!in_interrupt() &&
3016 ((current->flags & PF_MEMALLOC) ||
3017 unlikely(test_thread_flag(TIF_MEMDIE))))
3018 alloc_flags |= ALLOC_NO_WATERMARKS;
3021 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3022 alloc_flags |= ALLOC_CMA;
3027 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3029 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3032 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3034 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3037 static inline struct page *
3038 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3039 struct alloc_context *ac)
3041 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3042 struct page *page = NULL;
3044 unsigned long pages_reclaimed = 0;
3045 unsigned long did_some_progress;
3046 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3047 bool deferred_compaction = false;
3048 int contended_compaction = COMPACT_CONTENDED_NONE;
3051 * In the slowpath, we sanity check order to avoid ever trying to
3052 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3053 * be using allocators in order of preference for an area that is
3056 if (order >= MAX_ORDER) {
3057 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3062 * We also sanity check to catch abuse of atomic reserves being used by
3063 * callers that are not in atomic context.
3065 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3066 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3067 gfp_mask &= ~__GFP_ATOMIC;
3070 * If this allocation cannot block and it is for a specific node, then
3071 * fail early. There's no need to wakeup kswapd or retry for a
3072 * speculative node-specific allocation.
3074 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3078 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3079 wake_all_kswapds(order, ac);
3082 * OK, we're below the kswapd watermark and have kicked background
3083 * reclaim. Now things get more complex, so set up alloc_flags according
3084 * to how we want to proceed.
3086 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3089 * Find the true preferred zone if the allocation is unconstrained by
3092 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3093 struct zoneref *preferred_zoneref;
3094 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3095 ac->high_zoneidx, NULL, &ac->preferred_zone);
3096 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3099 /* This is the last chance, in general, before the goto nopage. */
3100 page = get_page_from_freelist(gfp_mask, order,
3101 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3105 /* Allocate without watermarks if the context allows */
3106 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3108 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3109 * the allocation is high priority and these type of
3110 * allocations are system rather than user orientated
3112 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3119 /* Caller is not willing to reclaim, we can't balance anything */
3120 if (!can_direct_reclaim) {
3122 * All existing users of the deprecated __GFP_NOFAIL are
3123 * blockable, so warn of any new users that actually allow this
3124 * type of allocation to fail.
3126 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3130 /* Avoid recursion of direct reclaim */
3131 if (current->flags & PF_MEMALLOC)
3134 /* Avoid allocations with no watermarks from looping endlessly */
3135 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3139 * Try direct compaction. The first pass is asynchronous. Subsequent
3140 * attempts after direct reclaim are synchronous
3142 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3144 &contended_compaction,
3145 &deferred_compaction);
3149 /* Checks for THP-specific high-order allocations */
3150 if (is_thp_gfp_mask(gfp_mask)) {
3152 * If compaction is deferred for high-order allocations, it is
3153 * because sync compaction recently failed. If this is the case
3154 * and the caller requested a THP allocation, we do not want
3155 * to heavily disrupt the system, so we fail the allocation
3156 * instead of entering direct reclaim.
3158 if (deferred_compaction)
3162 * In all zones where compaction was attempted (and not
3163 * deferred or skipped), lock contention has been detected.
3164 * For THP allocation we do not want to disrupt the others
3165 * so we fallback to base pages instead.
3167 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3171 * If compaction was aborted due to need_resched(), we do not
3172 * want to further increase allocation latency, unless it is
3173 * khugepaged trying to collapse.
3175 if (contended_compaction == COMPACT_CONTENDED_SCHED
3176 && !(current->flags & PF_KTHREAD))
3181 * It can become very expensive to allocate transparent hugepages at
3182 * fault, so use asynchronous memory compaction for THP unless it is
3183 * khugepaged trying to collapse.
3185 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3186 migration_mode = MIGRATE_SYNC_LIGHT;
3188 /* Try direct reclaim and then allocating */
3189 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3190 &did_some_progress);
3194 /* Do not loop if specifically requested */
3195 if (gfp_mask & __GFP_NORETRY)
3198 /* Keep reclaiming pages as long as there is reasonable progress */
3199 pages_reclaimed += did_some_progress;
3200 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3201 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3202 /* Wait for some write requests to complete then retry */
3203 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3207 /* Reclaim has failed us, start killing things */
3208 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3212 /* Retry as long as the OOM killer is making progress */
3213 if (did_some_progress)
3218 * High-order allocations do not necessarily loop after
3219 * direct reclaim and reclaim/compaction depends on compaction
3220 * being called after reclaim so call directly if necessary
3222 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3224 &contended_compaction,
3225 &deferred_compaction);
3229 warn_alloc_failed(gfp_mask, order, NULL);
3235 * This is the 'heart' of the zoned buddy allocator.
3238 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3239 struct zonelist *zonelist, nodemask_t *nodemask)
3241 struct zoneref *preferred_zoneref;
3242 struct page *page = NULL;
3243 unsigned int cpuset_mems_cookie;
3244 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3245 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3246 struct alloc_context ac = {
3247 .high_zoneidx = gfp_zone(gfp_mask),
3248 .nodemask = nodemask,
3249 .migratetype = gfpflags_to_migratetype(gfp_mask),
3252 gfp_mask &= gfp_allowed_mask;
3254 lockdep_trace_alloc(gfp_mask);
3256 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3258 if (should_fail_alloc_page(gfp_mask, order))
3262 * Check the zones suitable for the gfp_mask contain at least one
3263 * valid zone. It's possible to have an empty zonelist as a result
3264 * of __GFP_THISNODE and a memoryless node
3266 if (unlikely(!zonelist->_zonerefs->zone))
3269 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3270 alloc_flags |= ALLOC_CMA;
3273 cpuset_mems_cookie = read_mems_allowed_begin();
3275 /* We set it here, as __alloc_pages_slowpath might have changed it */
3276 ac.zonelist = zonelist;
3278 /* Dirty zone balancing only done in the fast path */
3279 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3281 /* The preferred zone is used for statistics later */
3282 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3283 ac.nodemask ? : &cpuset_current_mems_allowed,
3284 &ac.preferred_zone);
3285 if (!ac.preferred_zone)
3287 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3289 /* First allocation attempt */
3290 alloc_mask = gfp_mask|__GFP_HARDWALL;
3291 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3292 if (unlikely(!page)) {
3294 * Runtime PM, block IO and its error handling path
3295 * can deadlock because I/O on the device might not
3298 alloc_mask = memalloc_noio_flags(gfp_mask);
3299 ac.spread_dirty_pages = false;
3301 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3304 if (kmemcheck_enabled && page)
3305 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3307 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3311 * When updating a task's mems_allowed, it is possible to race with
3312 * parallel threads in such a way that an allocation can fail while
3313 * the mask is being updated. If a page allocation is about to fail,
3314 * check if the cpuset changed during allocation and if so, retry.
3316 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3321 EXPORT_SYMBOL(__alloc_pages_nodemask);
3324 * Common helper functions.
3326 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3331 * __get_free_pages() returns a 32-bit address, which cannot represent
3334 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3336 page = alloc_pages(gfp_mask, order);
3339 return (unsigned long) page_address(page);
3341 EXPORT_SYMBOL(__get_free_pages);
3343 unsigned long get_zeroed_page(gfp_t gfp_mask)
3345 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3347 EXPORT_SYMBOL(get_zeroed_page);
3349 void __free_pages(struct page *page, unsigned int order)
3351 if (put_page_testzero(page)) {
3353 free_hot_cold_page(page, false);
3355 __free_pages_ok(page, order);
3359 EXPORT_SYMBOL(__free_pages);
3361 void free_pages(unsigned long addr, unsigned int order)
3364 VM_BUG_ON(!virt_addr_valid((void *)addr));
3365 __free_pages(virt_to_page((void *)addr), order);
3369 EXPORT_SYMBOL(free_pages);
3373 * An arbitrary-length arbitrary-offset area of memory which resides
3374 * within a 0 or higher order page. Multiple fragments within that page
3375 * are individually refcounted, in the page's reference counter.
3377 * The page_frag functions below provide a simple allocation framework for
3378 * page fragments. This is used by the network stack and network device
3379 * drivers to provide a backing region of memory for use as either an
3380 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3382 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3385 struct page *page = NULL;
3386 gfp_t gfp = gfp_mask;
3388 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3389 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3391 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3392 PAGE_FRAG_CACHE_MAX_ORDER);
3393 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3395 if (unlikely(!page))
3396 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3398 nc->va = page ? page_address(page) : NULL;
3403 void *__alloc_page_frag(struct page_frag_cache *nc,
3404 unsigned int fragsz, gfp_t gfp_mask)
3406 unsigned int size = PAGE_SIZE;
3410 if (unlikely(!nc->va)) {
3412 page = __page_frag_refill(nc, gfp_mask);
3416 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3417 /* if size can vary use size else just use PAGE_SIZE */
3420 /* Even if we own the page, we do not use atomic_set().
3421 * This would break get_page_unless_zero() users.
3423 atomic_add(size - 1, &page->_count);
3425 /* reset page count bias and offset to start of new frag */
3426 nc->pfmemalloc = page_is_pfmemalloc(page);
3427 nc->pagecnt_bias = size;
3431 offset = nc->offset - fragsz;
3432 if (unlikely(offset < 0)) {
3433 page = virt_to_page(nc->va);
3435 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3438 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3439 /* if size can vary use size else just use PAGE_SIZE */
3442 /* OK, page count is 0, we can safely set it */
3443 atomic_set(&page->_count, size);
3445 /* reset page count bias and offset to start of new frag */
3446 nc->pagecnt_bias = size;
3447 offset = size - fragsz;
3451 nc->offset = offset;
3453 return nc->va + offset;
3455 EXPORT_SYMBOL(__alloc_page_frag);
3458 * Frees a page fragment allocated out of either a compound or order 0 page.
3460 void __free_page_frag(void *addr)
3462 struct page *page = virt_to_head_page(addr);
3464 if (unlikely(put_page_testzero(page)))
3465 __free_pages_ok(page, compound_order(page));
3467 EXPORT_SYMBOL(__free_page_frag);
3470 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3471 * of the current memory cgroup.
3473 * It should be used when the caller would like to use kmalloc, but since the
3474 * allocation is large, it has to fall back to the page allocator.
3476 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3480 page = alloc_pages(gfp_mask, order);
3481 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3482 __free_pages(page, order);
3488 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3492 page = alloc_pages_node(nid, gfp_mask, order);
3493 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3494 __free_pages(page, order);
3501 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3504 void __free_kmem_pages(struct page *page, unsigned int order)
3506 memcg_kmem_uncharge(page, order);
3507 __free_pages(page, order);
3510 void free_kmem_pages(unsigned long addr, unsigned int order)
3513 VM_BUG_ON(!virt_addr_valid((void *)addr));
3514 __free_kmem_pages(virt_to_page((void *)addr), order);
3518 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3522 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3523 unsigned long used = addr + PAGE_ALIGN(size);
3525 split_page(virt_to_page((void *)addr), order);
3526 while (used < alloc_end) {
3531 return (void *)addr;
3535 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3536 * @size: the number of bytes to allocate
3537 * @gfp_mask: GFP flags for the allocation
3539 * This function is similar to alloc_pages(), except that it allocates the
3540 * minimum number of pages to satisfy the request. alloc_pages() can only
3541 * allocate memory in power-of-two pages.
3543 * This function is also limited by MAX_ORDER.
3545 * Memory allocated by this function must be released by free_pages_exact().
3547 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3549 unsigned int order = get_order(size);
3552 addr = __get_free_pages(gfp_mask, order);
3553 return make_alloc_exact(addr, order, size);
3555 EXPORT_SYMBOL(alloc_pages_exact);
3558 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3560 * @nid: the preferred node ID where memory should be allocated
3561 * @size: the number of bytes to allocate
3562 * @gfp_mask: GFP flags for the allocation
3564 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3567 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3569 unsigned int order = get_order(size);
3570 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3573 return make_alloc_exact((unsigned long)page_address(p), order, size);
3577 * free_pages_exact - release memory allocated via alloc_pages_exact()
3578 * @virt: the value returned by alloc_pages_exact.
3579 * @size: size of allocation, same value as passed to alloc_pages_exact().
3581 * Release the memory allocated by a previous call to alloc_pages_exact.
3583 void free_pages_exact(void *virt, size_t size)
3585 unsigned long addr = (unsigned long)virt;
3586 unsigned long end = addr + PAGE_ALIGN(size);
3588 while (addr < end) {
3593 EXPORT_SYMBOL(free_pages_exact);
3596 * nr_free_zone_pages - count number of pages beyond high watermark
3597 * @offset: The zone index of the highest zone
3599 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3600 * high watermark within all zones at or below a given zone index. For each
3601 * zone, the number of pages is calculated as:
3602 * managed_pages - high_pages
3604 static unsigned long nr_free_zone_pages(int offset)
3609 /* Just pick one node, since fallback list is circular */
3610 unsigned long sum = 0;
3612 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3614 for_each_zone_zonelist(zone, z, zonelist, offset) {
3615 unsigned long size = zone->managed_pages;
3616 unsigned long high = high_wmark_pages(zone);
3625 * nr_free_buffer_pages - count number of pages beyond high watermark
3627 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3628 * watermark within ZONE_DMA and ZONE_NORMAL.
3630 unsigned long nr_free_buffer_pages(void)
3632 return nr_free_zone_pages(gfp_zone(GFP_USER));
3634 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3637 * nr_free_pagecache_pages - count number of pages beyond high watermark
3639 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3640 * high watermark within all zones.
3642 unsigned long nr_free_pagecache_pages(void)
3644 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3647 static inline void show_node(struct zone *zone)
3649 if (IS_ENABLED(CONFIG_NUMA))
3650 printk("Node %d ", zone_to_nid(zone));
3653 long si_mem_available(void)
3656 unsigned long pagecache;
3657 unsigned long wmark_low = 0;
3658 unsigned long pages[NR_LRU_LISTS];
3662 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
3663 pages[lru] = global_page_state(NR_LRU_BASE + lru);
3666 wmark_low += zone->watermark[WMARK_LOW];
3669 * Estimate the amount of memory available for userspace allocations,
3670 * without causing swapping.
3672 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
3675 * Not all the page cache can be freed, otherwise the system will
3676 * start swapping. Assume at least half of the page cache, or the
3677 * low watermark worth of cache, needs to stay.
3679 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
3680 pagecache -= min(pagecache / 2, wmark_low);
3681 available += pagecache;
3684 * Part of the reclaimable slab consists of items that are in use,
3685 * and cannot be freed. Cap this estimate at the low watermark.
3687 available += global_page_state(NR_SLAB_RECLAIMABLE) -
3688 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
3694 EXPORT_SYMBOL_GPL(si_mem_available);
3696 void si_meminfo(struct sysinfo *val)
3698 val->totalram = totalram_pages;
3699 val->sharedram = global_page_state(NR_SHMEM);
3700 val->freeram = global_page_state(NR_FREE_PAGES);
3701 val->bufferram = nr_blockdev_pages();
3702 val->totalhigh = totalhigh_pages;
3703 val->freehigh = nr_free_highpages();
3704 val->mem_unit = PAGE_SIZE;
3707 EXPORT_SYMBOL(si_meminfo);
3710 void si_meminfo_node(struct sysinfo *val, int nid)
3712 int zone_type; /* needs to be signed */
3713 unsigned long managed_pages = 0;
3714 pg_data_t *pgdat = NODE_DATA(nid);
3716 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3717 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3718 val->totalram = managed_pages;
3719 val->sharedram = node_page_state(nid, NR_SHMEM);
3720 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3721 #ifdef CONFIG_HIGHMEM
3722 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3723 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3729 val->mem_unit = PAGE_SIZE;
3734 * Determine whether the node should be displayed or not, depending on whether
3735 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3737 bool skip_free_areas_node(unsigned int flags, int nid)
3740 unsigned int cpuset_mems_cookie;
3742 if (!(flags & SHOW_MEM_FILTER_NODES))
3746 cpuset_mems_cookie = read_mems_allowed_begin();
3747 ret = !node_isset(nid, cpuset_current_mems_allowed);
3748 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3753 #define K(x) ((x) << (PAGE_SHIFT-10))
3755 static void show_migration_types(unsigned char type)
3757 static const char types[MIGRATE_TYPES] = {
3758 [MIGRATE_UNMOVABLE] = 'U',
3759 [MIGRATE_MOVABLE] = 'M',
3760 [MIGRATE_RECLAIMABLE] = 'E',
3761 [MIGRATE_HIGHATOMIC] = 'H',
3763 [MIGRATE_CMA] = 'C',
3765 #ifdef CONFIG_MEMORY_ISOLATION
3766 [MIGRATE_ISOLATE] = 'I',
3769 char tmp[MIGRATE_TYPES + 1];
3773 for (i = 0; i < MIGRATE_TYPES; i++) {
3774 if (type & (1 << i))
3779 printk("(%s) ", tmp);
3783 * Show free area list (used inside shift_scroll-lock stuff)
3784 * We also calculate the percentage fragmentation. We do this by counting the
3785 * memory on each free list with the exception of the first item on the list.
3788 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3791 void show_free_areas(unsigned int filter)
3793 unsigned long free_pcp = 0;
3797 for_each_populated_zone(zone) {
3798 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3801 for_each_online_cpu(cpu)
3802 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3805 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3806 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3807 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3808 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3809 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3810 " free:%lu free_pcp:%lu free_cma:%lu\n",
3811 global_page_state(NR_ACTIVE_ANON),
3812 global_page_state(NR_INACTIVE_ANON),
3813 global_page_state(NR_ISOLATED_ANON),
3814 global_page_state(NR_ACTIVE_FILE),
3815 global_page_state(NR_INACTIVE_FILE),
3816 global_page_state(NR_ISOLATED_FILE),
3817 global_page_state(NR_UNEVICTABLE),
3818 global_page_state(NR_FILE_DIRTY),
3819 global_page_state(NR_WRITEBACK),
3820 global_page_state(NR_UNSTABLE_NFS),
3821 global_page_state(NR_SLAB_RECLAIMABLE),
3822 global_page_state(NR_SLAB_UNRECLAIMABLE),
3823 global_page_state(NR_FILE_MAPPED),
3824 global_page_state(NR_SHMEM),
3825 global_page_state(NR_PAGETABLE),
3826 global_page_state(NR_BOUNCE),
3827 global_page_state(NR_FREE_PAGES),
3829 global_page_state(NR_FREE_CMA_PAGES));
3831 for_each_populated_zone(zone) {
3834 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3838 for_each_online_cpu(cpu)
3839 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3847 " active_anon:%lukB"
3848 " inactive_anon:%lukB"
3849 " active_file:%lukB"
3850 " inactive_file:%lukB"
3851 " unevictable:%lukB"
3852 " isolated(anon):%lukB"
3853 " isolated(file):%lukB"
3861 " slab_reclaimable:%lukB"
3862 " slab_unreclaimable:%lukB"
3863 " kernel_stack:%lukB"
3870 " writeback_tmp:%lukB"
3871 " pages_scanned:%lu"
3872 " all_unreclaimable? %s"
3875 K(zone_page_state(zone, NR_FREE_PAGES)),
3876 K(min_wmark_pages(zone)),
3877 K(low_wmark_pages(zone)),
3878 K(high_wmark_pages(zone)),
3879 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3880 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3881 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3882 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3883 K(zone_page_state(zone, NR_UNEVICTABLE)),
3884 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3885 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3886 K(zone->present_pages),
3887 K(zone->managed_pages),
3888 K(zone_page_state(zone, NR_MLOCK)),
3889 K(zone_page_state(zone, NR_FILE_DIRTY)),
3890 K(zone_page_state(zone, NR_WRITEBACK)),
3891 K(zone_page_state(zone, NR_FILE_MAPPED)),
3892 K(zone_page_state(zone, NR_SHMEM)),
3893 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3894 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3895 zone_page_state(zone, NR_KERNEL_STACK) *
3897 K(zone_page_state(zone, NR_PAGETABLE)),
3898 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3899 K(zone_page_state(zone, NR_BOUNCE)),
3901 K(this_cpu_read(zone->pageset->pcp.count)),
3902 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3903 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3904 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3905 (!zone_reclaimable(zone) ? "yes" : "no")
3907 printk("lowmem_reserve[]:");
3908 for (i = 0; i < MAX_NR_ZONES; i++)
3909 printk(" %ld", zone->lowmem_reserve[i]);
3913 for_each_populated_zone(zone) {
3915 unsigned long nr[MAX_ORDER], flags, total = 0;
3916 unsigned char types[MAX_ORDER];
3918 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3921 printk("%s: ", zone->name);
3923 spin_lock_irqsave(&zone->lock, flags);
3924 for (order = 0; order < MAX_ORDER; order++) {
3925 struct free_area *area = &zone->free_area[order];
3928 nr[order] = area->nr_free;
3929 total += nr[order] << order;
3932 for (type = 0; type < MIGRATE_TYPES; type++) {
3933 if (!list_empty(&area->free_list[type]))
3934 types[order] |= 1 << type;
3937 spin_unlock_irqrestore(&zone->lock, flags);
3938 for (order = 0; order < MAX_ORDER; order++) {
3939 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3941 show_migration_types(types[order]);
3943 printk("= %lukB\n", K(total));
3946 hugetlb_show_meminfo();
3948 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3950 show_swap_cache_info();
3953 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3955 zoneref->zone = zone;
3956 zoneref->zone_idx = zone_idx(zone);
3960 * Builds allocation fallback zone lists.
3962 * Add all populated zones of a node to the zonelist.
3964 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3968 enum zone_type zone_type = MAX_NR_ZONES;
3972 zone = pgdat->node_zones + zone_type;
3973 if (populated_zone(zone)) {
3974 zoneref_set_zone(zone,
3975 &zonelist->_zonerefs[nr_zones++]);
3976 check_highest_zone(zone_type);
3978 } while (zone_type);
3986 * 0 = automatic detection of better ordering.
3987 * 1 = order by ([node] distance, -zonetype)
3988 * 2 = order by (-zonetype, [node] distance)
3990 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3991 * the same zonelist. So only NUMA can configure this param.
3993 #define ZONELIST_ORDER_DEFAULT 0
3994 #define ZONELIST_ORDER_NODE 1
3995 #define ZONELIST_ORDER_ZONE 2
3997 /* zonelist order in the kernel.
3998 * set_zonelist_order() will set this to NODE or ZONE.
4000 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4001 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4005 /* The value user specified ....changed by config */
4006 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4007 /* string for sysctl */
4008 #define NUMA_ZONELIST_ORDER_LEN 16
4009 char numa_zonelist_order[16] = "default";
4012 * interface for configure zonelist ordering.
4013 * command line option "numa_zonelist_order"
4014 * = "[dD]efault - default, automatic configuration.
4015 * = "[nN]ode - order by node locality, then by zone within node
4016 * = "[zZ]one - order by zone, then by locality within zone
4019 static int __parse_numa_zonelist_order(char *s)
4021 if (*s == 'd' || *s == 'D') {
4022 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4023 } else if (*s == 'n' || *s == 'N') {
4024 user_zonelist_order = ZONELIST_ORDER_NODE;
4025 } else if (*s == 'z' || *s == 'Z') {
4026 user_zonelist_order = ZONELIST_ORDER_ZONE;
4029 "Ignoring invalid numa_zonelist_order value: "
4036 static __init int setup_numa_zonelist_order(char *s)
4043 ret = __parse_numa_zonelist_order(s);
4045 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4049 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4052 * sysctl handler for numa_zonelist_order
4054 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4055 void __user *buffer, size_t *length,
4058 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4060 static DEFINE_MUTEX(zl_order_mutex);
4062 mutex_lock(&zl_order_mutex);
4064 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4068 strcpy(saved_string, (char *)table->data);
4070 ret = proc_dostring(table, write, buffer, length, ppos);
4074 int oldval = user_zonelist_order;
4076 ret = __parse_numa_zonelist_order((char *)table->data);
4079 * bogus value. restore saved string
4081 strncpy((char *)table->data, saved_string,
4082 NUMA_ZONELIST_ORDER_LEN);
4083 user_zonelist_order = oldval;
4084 } else if (oldval != user_zonelist_order) {
4085 mutex_lock(&zonelists_mutex);
4086 build_all_zonelists(NULL, NULL);
4087 mutex_unlock(&zonelists_mutex);
4091 mutex_unlock(&zl_order_mutex);
4096 #define MAX_NODE_LOAD (nr_online_nodes)
4097 static int node_load[MAX_NUMNODES];
4100 * find_next_best_node - find the next node that should appear in a given node's fallback list
4101 * @node: node whose fallback list we're appending
4102 * @used_node_mask: nodemask_t of already used nodes
4104 * We use a number of factors to determine which is the next node that should
4105 * appear on a given node's fallback list. The node should not have appeared
4106 * already in @node's fallback list, and it should be the next closest node
4107 * according to the distance array (which contains arbitrary distance values
4108 * from each node to each node in the system), and should also prefer nodes
4109 * with no CPUs, since presumably they'll have very little allocation pressure
4110 * on them otherwise.
4111 * It returns -1 if no node is found.
4113 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4116 int min_val = INT_MAX;
4117 int best_node = NUMA_NO_NODE;
4118 const struct cpumask *tmp = cpumask_of_node(0);
4120 /* Use the local node if we haven't already */
4121 if (!node_isset(node, *used_node_mask)) {
4122 node_set(node, *used_node_mask);
4126 for_each_node_state(n, N_MEMORY) {
4128 /* Don't want a node to appear more than once */
4129 if (node_isset(n, *used_node_mask))
4132 /* Use the distance array to find the distance */
4133 val = node_distance(node, n);
4135 /* Penalize nodes under us ("prefer the next node") */
4138 /* Give preference to headless and unused nodes */
4139 tmp = cpumask_of_node(n);
4140 if (!cpumask_empty(tmp))
4141 val += PENALTY_FOR_NODE_WITH_CPUS;
4143 /* Slight preference for less loaded node */
4144 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4145 val += node_load[n];
4147 if (val < min_val) {
4154 node_set(best_node, *used_node_mask);
4161 * Build zonelists ordered by node and zones within node.
4162 * This results in maximum locality--normal zone overflows into local
4163 * DMA zone, if any--but risks exhausting DMA zone.
4165 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4168 struct zonelist *zonelist;
4170 zonelist = &pgdat->node_zonelists[0];
4171 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4173 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4174 zonelist->_zonerefs[j].zone = NULL;
4175 zonelist->_zonerefs[j].zone_idx = 0;
4179 * Build gfp_thisnode zonelists
4181 static void build_thisnode_zonelists(pg_data_t *pgdat)
4184 struct zonelist *zonelist;
4186 zonelist = &pgdat->node_zonelists[1];
4187 j = build_zonelists_node(pgdat, zonelist, 0);
4188 zonelist->_zonerefs[j].zone = NULL;
4189 zonelist->_zonerefs[j].zone_idx = 0;
4193 * Build zonelists ordered by zone and nodes within zones.
4194 * This results in conserving DMA zone[s] until all Normal memory is
4195 * exhausted, but results in overflowing to remote node while memory
4196 * may still exist in local DMA zone.
4198 static int node_order[MAX_NUMNODES];
4200 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4203 int zone_type; /* needs to be signed */
4205 struct zonelist *zonelist;
4207 zonelist = &pgdat->node_zonelists[0];
4209 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4210 for (j = 0; j < nr_nodes; j++) {
4211 node = node_order[j];
4212 z = &NODE_DATA(node)->node_zones[zone_type];
4213 if (populated_zone(z)) {
4215 &zonelist->_zonerefs[pos++]);
4216 check_highest_zone(zone_type);
4220 zonelist->_zonerefs[pos].zone = NULL;
4221 zonelist->_zonerefs[pos].zone_idx = 0;
4224 #if defined(CONFIG_64BIT)
4226 * Devices that require DMA32/DMA are relatively rare and do not justify a
4227 * penalty to every machine in case the specialised case applies. Default
4228 * to Node-ordering on 64-bit NUMA machines
4230 static int default_zonelist_order(void)
4232 return ZONELIST_ORDER_NODE;
4236 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4237 * by the kernel. If processes running on node 0 deplete the low memory zone
4238 * then reclaim will occur more frequency increasing stalls and potentially
4239 * be easier to OOM if a large percentage of the zone is under writeback or
4240 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4241 * Hence, default to zone ordering on 32-bit.
4243 static int default_zonelist_order(void)
4245 return ZONELIST_ORDER_ZONE;
4247 #endif /* CONFIG_64BIT */
4249 static void set_zonelist_order(void)
4251 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4252 current_zonelist_order = default_zonelist_order();
4254 current_zonelist_order = user_zonelist_order;
4257 static void build_zonelists(pg_data_t *pgdat)
4261 nodemask_t used_mask;
4262 int local_node, prev_node;
4263 struct zonelist *zonelist;
4264 unsigned int order = current_zonelist_order;
4266 /* initialize zonelists */
4267 for (i = 0; i < MAX_ZONELISTS; i++) {
4268 zonelist = pgdat->node_zonelists + i;
4269 zonelist->_zonerefs[0].zone = NULL;
4270 zonelist->_zonerefs[0].zone_idx = 0;
4273 /* NUMA-aware ordering of nodes */
4274 local_node = pgdat->node_id;
4275 load = nr_online_nodes;
4276 prev_node = local_node;
4277 nodes_clear(used_mask);
4279 memset(node_order, 0, sizeof(node_order));
4282 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4284 * We don't want to pressure a particular node.
4285 * So adding penalty to the first node in same
4286 * distance group to make it round-robin.
4288 if (node_distance(local_node, node) !=
4289 node_distance(local_node, prev_node))
4290 node_load[node] = load;
4294 if (order == ZONELIST_ORDER_NODE)
4295 build_zonelists_in_node_order(pgdat, node);
4297 node_order[j++] = node; /* remember order */
4300 if (order == ZONELIST_ORDER_ZONE) {
4301 /* calculate node order -- i.e., DMA last! */
4302 build_zonelists_in_zone_order(pgdat, j);
4305 build_thisnode_zonelists(pgdat);
4308 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4310 * Return node id of node used for "local" allocations.
4311 * I.e., first node id of first zone in arg node's generic zonelist.
4312 * Used for initializing percpu 'numa_mem', which is used primarily
4313 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4315 int local_memory_node(int node)
4319 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4320 gfp_zone(GFP_KERNEL),
4327 #else /* CONFIG_NUMA */
4329 static void set_zonelist_order(void)
4331 current_zonelist_order = ZONELIST_ORDER_ZONE;
4334 static void build_zonelists(pg_data_t *pgdat)
4336 int node, local_node;
4338 struct zonelist *zonelist;
4340 local_node = pgdat->node_id;
4342 zonelist = &pgdat->node_zonelists[0];
4343 j = build_zonelists_node(pgdat, zonelist, 0);
4346 * Now we build the zonelist so that it contains the zones
4347 * of all the other nodes.
4348 * We don't want to pressure a particular node, so when
4349 * building the zones for node N, we make sure that the
4350 * zones coming right after the local ones are those from
4351 * node N+1 (modulo N)
4353 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4354 if (!node_online(node))
4356 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4358 for (node = 0; node < local_node; node++) {
4359 if (!node_online(node))
4361 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4364 zonelist->_zonerefs[j].zone = NULL;
4365 zonelist->_zonerefs[j].zone_idx = 0;
4368 #endif /* CONFIG_NUMA */
4371 * Boot pageset table. One per cpu which is going to be used for all
4372 * zones and all nodes. The parameters will be set in such a way
4373 * that an item put on a list will immediately be handed over to
4374 * the buddy list. This is safe since pageset manipulation is done
4375 * with interrupts disabled.
4377 * The boot_pagesets must be kept even after bootup is complete for
4378 * unused processors and/or zones. They do play a role for bootstrapping
4379 * hotplugged processors.
4381 * zoneinfo_show() and maybe other functions do
4382 * not check if the processor is online before following the pageset pointer.
4383 * Other parts of the kernel may not check if the zone is available.
4385 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4386 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4387 static void setup_zone_pageset(struct zone *zone);
4390 * Global mutex to protect against size modification of zonelists
4391 * as well as to serialize pageset setup for the new populated zone.
4393 DEFINE_MUTEX(zonelists_mutex);
4395 /* return values int ....just for stop_machine() */
4396 static int __build_all_zonelists(void *data)
4400 pg_data_t *self = data;
4403 memset(node_load, 0, sizeof(node_load));
4406 if (self && !node_online(self->node_id)) {
4407 build_zonelists(self);
4410 for_each_online_node(nid) {
4411 pg_data_t *pgdat = NODE_DATA(nid);
4413 build_zonelists(pgdat);
4417 * Initialize the boot_pagesets that are going to be used
4418 * for bootstrapping processors. The real pagesets for
4419 * each zone will be allocated later when the per cpu
4420 * allocator is available.
4422 * boot_pagesets are used also for bootstrapping offline
4423 * cpus if the system is already booted because the pagesets
4424 * are needed to initialize allocators on a specific cpu too.
4425 * F.e. the percpu allocator needs the page allocator which
4426 * needs the percpu allocator in order to allocate its pagesets
4427 * (a chicken-egg dilemma).
4429 for_each_possible_cpu(cpu) {
4430 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4432 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4434 * We now know the "local memory node" for each node--
4435 * i.e., the node of the first zone in the generic zonelist.
4436 * Set up numa_mem percpu variable for on-line cpus. During
4437 * boot, only the boot cpu should be on-line; we'll init the
4438 * secondary cpus' numa_mem as they come on-line. During
4439 * node/memory hotplug, we'll fixup all on-line cpus.
4441 if (cpu_online(cpu))
4442 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4449 static noinline void __init
4450 build_all_zonelists_init(void)
4452 __build_all_zonelists(NULL);
4453 mminit_verify_zonelist();
4454 cpuset_init_current_mems_allowed();
4458 * Called with zonelists_mutex held always
4459 * unless system_state == SYSTEM_BOOTING.
4461 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4462 * [we're only called with non-NULL zone through __meminit paths] and
4463 * (2) call of __init annotated helper build_all_zonelists_init
4464 * [protected by SYSTEM_BOOTING].
4466 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4468 set_zonelist_order();
4470 if (system_state == SYSTEM_BOOTING) {
4471 build_all_zonelists_init();
4473 #ifdef CONFIG_MEMORY_HOTPLUG
4475 setup_zone_pageset(zone);
4477 /* we have to stop all cpus to guarantee there is no user
4479 stop_machine(__build_all_zonelists, pgdat, NULL);
4480 /* cpuset refresh routine should be here */
4482 vm_total_pages = nr_free_pagecache_pages();
4484 * Disable grouping by mobility if the number of pages in the
4485 * system is too low to allow the mechanism to work. It would be
4486 * more accurate, but expensive to check per-zone. This check is
4487 * made on memory-hotadd so a system can start with mobility
4488 * disabled and enable it later
4490 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4491 page_group_by_mobility_disabled = 1;
4493 page_group_by_mobility_disabled = 0;
4495 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4496 "Total pages: %ld\n",
4498 zonelist_order_name[current_zonelist_order],
4499 page_group_by_mobility_disabled ? "off" : "on",
4502 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4507 * Helper functions to size the waitqueue hash table.
4508 * Essentially these want to choose hash table sizes sufficiently
4509 * large so that collisions trying to wait on pages are rare.
4510 * But in fact, the number of active page waitqueues on typical
4511 * systems is ridiculously low, less than 200. So this is even
4512 * conservative, even though it seems large.
4514 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4515 * waitqueues, i.e. the size of the waitq table given the number of pages.
4517 #define PAGES_PER_WAITQUEUE 256
4519 #ifndef CONFIG_MEMORY_HOTPLUG
4520 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4522 unsigned long size = 1;
4524 pages /= PAGES_PER_WAITQUEUE;
4526 while (size < pages)
4530 * Once we have dozens or even hundreds of threads sleeping
4531 * on IO we've got bigger problems than wait queue collision.
4532 * Limit the size of the wait table to a reasonable size.
4534 size = min(size, 4096UL);
4536 return max(size, 4UL);
4540 * A zone's size might be changed by hot-add, so it is not possible to determine
4541 * a suitable size for its wait_table. So we use the maximum size now.
4543 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4545 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4546 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4547 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4549 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4550 * or more by the traditional way. (See above). It equals:
4552 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4553 * ia64(16K page size) : = ( 8G + 4M)byte.
4554 * powerpc (64K page size) : = (32G +16M)byte.
4556 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4563 * This is an integer logarithm so that shifts can be used later
4564 * to extract the more random high bits from the multiplicative
4565 * hash function before the remainder is taken.
4567 static inline unsigned long wait_table_bits(unsigned long size)
4573 * Initially all pages are reserved - free ones are freed
4574 * up by free_all_bootmem() once the early boot process is
4575 * done. Non-atomic initialization, single-pass.
4577 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4578 unsigned long start_pfn, enum memmap_context context)
4580 pg_data_t *pgdat = NODE_DATA(nid);
4581 unsigned long end_pfn = start_pfn + size;
4584 unsigned long nr_initialised = 0;
4586 if (highest_memmap_pfn < end_pfn - 1)
4587 highest_memmap_pfn = end_pfn - 1;
4589 z = &pgdat->node_zones[zone];
4590 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4592 * There can be holes in boot-time mem_map[]s
4593 * handed to this function. They do not
4594 * exist on hotplugged memory.
4596 if (context == MEMMAP_EARLY) {
4597 if (!early_pfn_valid(pfn))
4599 if (!early_pfn_in_nid(pfn, nid))
4601 if (!update_defer_init(pgdat, pfn, end_pfn,
4607 * Mark the block movable so that blocks are reserved for
4608 * movable at startup. This will force kernel allocations
4609 * to reserve their blocks rather than leaking throughout
4610 * the address space during boot when many long-lived
4611 * kernel allocations are made.
4613 * bitmap is created for zone's valid pfn range. but memmap
4614 * can be created for invalid pages (for alignment)
4615 * check here not to call set_pageblock_migratetype() against
4618 if (!(pfn & (pageblock_nr_pages - 1))) {
4619 struct page *page = pfn_to_page(pfn);
4621 __init_single_page(page, pfn, zone, nid);
4622 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4624 __init_single_pfn(pfn, zone, nid);
4629 static void __meminit zone_init_free_lists(struct zone *zone)
4631 unsigned int order, t;
4632 for_each_migratetype_order(order, t) {
4633 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4634 zone->free_area[order].nr_free = 0;
4638 #ifndef __HAVE_ARCH_MEMMAP_INIT
4639 #define memmap_init(size, nid, zone, start_pfn) \
4640 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4643 static int zone_batchsize(struct zone *zone)
4649 * The per-cpu-pages pools are set to around 1000th of the
4650 * size of the zone. But no more than 1/2 of a meg.
4652 * OK, so we don't know how big the cache is. So guess.
4654 batch = zone->managed_pages / 1024;
4655 if (batch * PAGE_SIZE > 512 * 1024)
4656 batch = (512 * 1024) / PAGE_SIZE;
4657 batch /= 4; /* We effectively *= 4 below */
4662 * Clamp the batch to a 2^n - 1 value. Having a power
4663 * of 2 value was found to be more likely to have
4664 * suboptimal cache aliasing properties in some cases.
4666 * For example if 2 tasks are alternately allocating
4667 * batches of pages, one task can end up with a lot
4668 * of pages of one half of the possible page colors
4669 * and the other with pages of the other colors.
4671 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4676 /* The deferral and batching of frees should be suppressed under NOMMU
4679 * The problem is that NOMMU needs to be able to allocate large chunks
4680 * of contiguous memory as there's no hardware page translation to
4681 * assemble apparent contiguous memory from discontiguous pages.
4683 * Queueing large contiguous runs of pages for batching, however,
4684 * causes the pages to actually be freed in smaller chunks. As there
4685 * can be a significant delay between the individual batches being
4686 * recycled, this leads to the once large chunks of space being
4687 * fragmented and becoming unavailable for high-order allocations.
4694 * pcp->high and pcp->batch values are related and dependent on one another:
4695 * ->batch must never be higher then ->high.
4696 * The following function updates them in a safe manner without read side
4699 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4700 * those fields changing asynchronously (acording the the above rule).
4702 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4703 * outside of boot time (or some other assurance that no concurrent updaters
4706 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4707 unsigned long batch)
4709 /* start with a fail safe value for batch */
4713 /* Update high, then batch, in order */
4720 /* a companion to pageset_set_high() */
4721 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4723 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4726 static void pageset_init(struct per_cpu_pageset *p)
4728 struct per_cpu_pages *pcp;
4731 memset(p, 0, sizeof(*p));
4735 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4736 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4739 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4742 pageset_set_batch(p, batch);
4746 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4747 * to the value high for the pageset p.
4749 static void pageset_set_high(struct per_cpu_pageset *p,
4752 unsigned long batch = max(1UL, high / 4);
4753 if ((high / 4) > (PAGE_SHIFT * 8))
4754 batch = PAGE_SHIFT * 8;
4756 pageset_update(&p->pcp, high, batch);
4759 static void pageset_set_high_and_batch(struct zone *zone,
4760 struct per_cpu_pageset *pcp)
4762 if (percpu_pagelist_fraction)
4763 pageset_set_high(pcp,
4764 (zone->managed_pages /
4765 percpu_pagelist_fraction));
4767 pageset_set_batch(pcp, zone_batchsize(zone));
4770 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4772 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4775 pageset_set_high_and_batch(zone, pcp);
4778 static void __meminit setup_zone_pageset(struct zone *zone)
4781 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4782 for_each_possible_cpu(cpu)
4783 zone_pageset_init(zone, cpu);
4787 * Allocate per cpu pagesets and initialize them.
4788 * Before this call only boot pagesets were available.
4790 void __init setup_per_cpu_pageset(void)
4794 for_each_populated_zone(zone)
4795 setup_zone_pageset(zone);
4798 static noinline __init_refok
4799 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4805 * The per-page waitqueue mechanism uses hashed waitqueues
4808 zone->wait_table_hash_nr_entries =
4809 wait_table_hash_nr_entries(zone_size_pages);
4810 zone->wait_table_bits =
4811 wait_table_bits(zone->wait_table_hash_nr_entries);
4812 alloc_size = zone->wait_table_hash_nr_entries
4813 * sizeof(wait_queue_head_t);
4815 if (!slab_is_available()) {
4816 zone->wait_table = (wait_queue_head_t *)
4817 memblock_virt_alloc_node_nopanic(
4818 alloc_size, zone->zone_pgdat->node_id);
4821 * This case means that a zone whose size was 0 gets new memory
4822 * via memory hot-add.
4823 * But it may be the case that a new node was hot-added. In
4824 * this case vmalloc() will not be able to use this new node's
4825 * memory - this wait_table must be initialized to use this new
4826 * node itself as well.
4827 * To use this new node's memory, further consideration will be
4830 zone->wait_table = vmalloc(alloc_size);
4832 if (!zone->wait_table)
4835 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4836 init_waitqueue_head(zone->wait_table + i);
4841 static __meminit void zone_pcp_init(struct zone *zone)
4844 * per cpu subsystem is not up at this point. The following code
4845 * relies on the ability of the linker to provide the
4846 * offset of a (static) per cpu variable into the per cpu area.
4848 zone->pageset = &boot_pageset;
4850 if (populated_zone(zone))
4851 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4852 zone->name, zone->present_pages,
4853 zone_batchsize(zone));
4856 int __meminit init_currently_empty_zone(struct zone *zone,
4857 unsigned long zone_start_pfn,
4860 struct pglist_data *pgdat = zone->zone_pgdat;
4862 ret = zone_wait_table_init(zone, size);
4865 pgdat->nr_zones = zone_idx(zone) + 1;
4867 zone->zone_start_pfn = zone_start_pfn;
4869 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4870 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4872 (unsigned long)zone_idx(zone),
4873 zone_start_pfn, (zone_start_pfn + size));
4875 zone_init_free_lists(zone);
4880 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4881 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4884 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4886 int __meminit __early_pfn_to_nid(unsigned long pfn,
4887 struct mminit_pfnnid_cache *state)
4889 unsigned long start_pfn, end_pfn;
4892 if (state->last_start <= pfn && pfn < state->last_end)
4893 return state->last_nid;
4895 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4897 state->last_start = start_pfn;
4898 state->last_end = end_pfn;
4899 state->last_nid = nid;
4904 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4907 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4908 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4909 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4911 * If an architecture guarantees that all ranges registered contain no holes
4912 * and may be freed, this this function may be used instead of calling
4913 * memblock_free_early_nid() manually.
4915 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4917 unsigned long start_pfn, end_pfn;
4920 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4921 start_pfn = min(start_pfn, max_low_pfn);
4922 end_pfn = min(end_pfn, max_low_pfn);
4924 if (start_pfn < end_pfn)
4925 memblock_free_early_nid(PFN_PHYS(start_pfn),
4926 (end_pfn - start_pfn) << PAGE_SHIFT,
4932 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4933 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4935 * If an architecture guarantees that all ranges registered contain no holes and may
4936 * be freed, this function may be used instead of calling memory_present() manually.
4938 void __init sparse_memory_present_with_active_regions(int nid)
4940 unsigned long start_pfn, end_pfn;
4943 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4944 memory_present(this_nid, start_pfn, end_pfn);
4948 * get_pfn_range_for_nid - Return the start and end page frames for a node
4949 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4950 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4951 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4953 * It returns the start and end page frame of a node based on information
4954 * provided by memblock_set_node(). If called for a node
4955 * with no available memory, a warning is printed and the start and end
4958 void __meminit get_pfn_range_for_nid(unsigned int nid,
4959 unsigned long *start_pfn, unsigned long *end_pfn)
4961 unsigned long this_start_pfn, this_end_pfn;
4967 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4968 *start_pfn = min(*start_pfn, this_start_pfn);
4969 *end_pfn = max(*end_pfn, this_end_pfn);
4972 if (*start_pfn == -1UL)
4977 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4978 * assumption is made that zones within a node are ordered in monotonic
4979 * increasing memory addresses so that the "highest" populated zone is used
4981 static void __init find_usable_zone_for_movable(void)
4984 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4985 if (zone_index == ZONE_MOVABLE)
4988 if (arch_zone_highest_possible_pfn[zone_index] >
4989 arch_zone_lowest_possible_pfn[zone_index])
4993 VM_BUG_ON(zone_index == -1);
4994 movable_zone = zone_index;
4998 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4999 * because it is sized independent of architecture. Unlike the other zones,
5000 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5001 * in each node depending on the size of each node and how evenly kernelcore
5002 * is distributed. This helper function adjusts the zone ranges
5003 * provided by the architecture for a given node by using the end of the
5004 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5005 * zones within a node are in order of monotonic increases memory addresses
5007 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5008 unsigned long zone_type,
5009 unsigned long node_start_pfn,
5010 unsigned long node_end_pfn,
5011 unsigned long *zone_start_pfn,
5012 unsigned long *zone_end_pfn)
5014 /* Only adjust if ZONE_MOVABLE is on this node */
5015 if (zone_movable_pfn[nid]) {
5016 /* Size ZONE_MOVABLE */
5017 if (zone_type == ZONE_MOVABLE) {
5018 *zone_start_pfn = zone_movable_pfn[nid];
5019 *zone_end_pfn = min(node_end_pfn,
5020 arch_zone_highest_possible_pfn[movable_zone]);
5022 /* Adjust for ZONE_MOVABLE starting within this range */
5023 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
5024 *zone_end_pfn > zone_movable_pfn[nid]) {
5025 *zone_end_pfn = zone_movable_pfn[nid];
5027 /* Check if this whole range is within ZONE_MOVABLE */
5028 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5029 *zone_start_pfn = *zone_end_pfn;
5034 * Return the number of pages a zone spans in a node, including holes
5035 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5037 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5038 unsigned long zone_type,
5039 unsigned long node_start_pfn,
5040 unsigned long node_end_pfn,
5041 unsigned long *ignored)
5043 unsigned long zone_start_pfn, zone_end_pfn;
5045 /* When hotadd a new node from cpu_up(), the node should be empty */
5046 if (!node_start_pfn && !node_end_pfn)
5049 /* Get the start and end of the zone */
5050 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5051 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5052 adjust_zone_range_for_zone_movable(nid, zone_type,
5053 node_start_pfn, node_end_pfn,
5054 &zone_start_pfn, &zone_end_pfn);
5056 /* Check that this node has pages within the zone's required range */
5057 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
5060 /* Move the zone boundaries inside the node if necessary */
5061 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
5062 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5064 /* Return the spanned pages */
5065 return zone_end_pfn - zone_start_pfn;
5069 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5070 * then all holes in the requested range will be accounted for.
5072 unsigned long __meminit __absent_pages_in_range(int nid,
5073 unsigned long range_start_pfn,
5074 unsigned long range_end_pfn)
5076 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5077 unsigned long start_pfn, end_pfn;
5080 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5081 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5082 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5083 nr_absent -= end_pfn - start_pfn;
5089 * absent_pages_in_range - Return number of page frames in holes within a range
5090 * @start_pfn: The start PFN to start searching for holes
5091 * @end_pfn: The end PFN to stop searching for holes
5093 * It returns the number of pages frames in memory holes within a range.
5095 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5096 unsigned long end_pfn)
5098 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5101 /* Return the number of page frames in holes in a zone on a node */
5102 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5103 unsigned long zone_type,
5104 unsigned long node_start_pfn,
5105 unsigned long node_end_pfn,
5106 unsigned long *ignored)
5108 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5109 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5110 unsigned long zone_start_pfn, zone_end_pfn;
5112 /* When hotadd a new node from cpu_up(), the node should be empty */
5113 if (!node_start_pfn && !node_end_pfn)
5116 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5117 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5119 adjust_zone_range_for_zone_movable(nid, zone_type,
5120 node_start_pfn, node_end_pfn,
5121 &zone_start_pfn, &zone_end_pfn);
5122 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5125 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5126 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5127 unsigned long zone_type,
5128 unsigned long node_start_pfn,
5129 unsigned long node_end_pfn,
5130 unsigned long *zones_size)
5132 return zones_size[zone_type];
5135 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5136 unsigned long zone_type,
5137 unsigned long node_start_pfn,
5138 unsigned long node_end_pfn,
5139 unsigned long *zholes_size)
5144 return zholes_size[zone_type];
5147 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5149 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5150 unsigned long node_start_pfn,
5151 unsigned long node_end_pfn,
5152 unsigned long *zones_size,
5153 unsigned long *zholes_size)
5155 unsigned long realtotalpages = 0, totalpages = 0;
5158 for (i = 0; i < MAX_NR_ZONES; i++) {
5159 struct zone *zone = pgdat->node_zones + i;
5160 unsigned long size, real_size;
5162 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5166 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5167 node_start_pfn, node_end_pfn,
5169 zone->spanned_pages = size;
5170 zone->present_pages = real_size;
5173 realtotalpages += real_size;
5176 pgdat->node_spanned_pages = totalpages;
5177 pgdat->node_present_pages = realtotalpages;
5178 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5182 #ifndef CONFIG_SPARSEMEM
5184 * Calculate the size of the zone->blockflags rounded to an unsigned long
5185 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5186 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5187 * round what is now in bits to nearest long in bits, then return it in
5190 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5192 unsigned long usemapsize;
5194 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5195 usemapsize = roundup(zonesize, pageblock_nr_pages);
5196 usemapsize = usemapsize >> pageblock_order;
5197 usemapsize *= NR_PAGEBLOCK_BITS;
5198 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5200 return usemapsize / 8;
5203 static void __init setup_usemap(struct pglist_data *pgdat,
5205 unsigned long zone_start_pfn,
5206 unsigned long zonesize)
5208 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5209 zone->pageblock_flags = NULL;
5211 zone->pageblock_flags =
5212 memblock_virt_alloc_node_nopanic(usemapsize,
5216 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5217 unsigned long zone_start_pfn, unsigned long zonesize) {}
5218 #endif /* CONFIG_SPARSEMEM */
5220 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5222 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5223 void __paginginit set_pageblock_order(void)
5227 /* Check that pageblock_nr_pages has not already been setup */
5228 if (pageblock_order)
5231 if (HPAGE_SHIFT > PAGE_SHIFT)
5232 order = HUGETLB_PAGE_ORDER;
5234 order = MAX_ORDER - 1;
5237 * Assume the largest contiguous order of interest is a huge page.
5238 * This value may be variable depending on boot parameters on IA64 and
5241 pageblock_order = order;
5243 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5246 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5247 * is unused as pageblock_order is set at compile-time. See
5248 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5251 void __paginginit set_pageblock_order(void)
5255 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5257 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5258 unsigned long present_pages)
5260 unsigned long pages = spanned_pages;
5263 * Provide a more accurate estimation if there are holes within
5264 * the zone and SPARSEMEM is in use. If there are holes within the
5265 * zone, each populated memory region may cost us one or two extra
5266 * memmap pages due to alignment because memmap pages for each
5267 * populated regions may not naturally algined on page boundary.
5268 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5270 if (spanned_pages > present_pages + (present_pages >> 4) &&
5271 IS_ENABLED(CONFIG_SPARSEMEM))
5272 pages = present_pages;
5274 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5278 * Set up the zone data structures:
5279 * - mark all pages reserved
5280 * - mark all memory queues empty
5281 * - clear the memory bitmaps
5283 * NOTE: pgdat should get zeroed by caller.
5285 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5288 int nid = pgdat->node_id;
5289 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5292 pgdat_resize_init(pgdat);
5293 #ifdef CONFIG_NUMA_BALANCING
5294 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5295 pgdat->numabalancing_migrate_nr_pages = 0;
5296 pgdat->numabalancing_migrate_next_window = jiffies;
5298 init_waitqueue_head(&pgdat->kswapd_wait);
5299 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5300 pgdat_page_ext_init(pgdat);
5302 for (j = 0; j < MAX_NR_ZONES; j++) {
5303 struct zone *zone = pgdat->node_zones + j;
5304 unsigned long size, realsize, freesize, memmap_pages;
5306 size = zone->spanned_pages;
5307 realsize = freesize = zone->present_pages;
5310 * Adjust freesize so that it accounts for how much memory
5311 * is used by this zone for memmap. This affects the watermark
5312 * and per-cpu initialisations
5314 memmap_pages = calc_memmap_size(size, realsize);
5315 if (!is_highmem_idx(j)) {
5316 if (freesize >= memmap_pages) {
5317 freesize -= memmap_pages;
5320 " %s zone: %lu pages used for memmap\n",
5321 zone_names[j], memmap_pages);
5324 " %s zone: %lu pages exceeds freesize %lu\n",
5325 zone_names[j], memmap_pages, freesize);
5328 /* Account for reserved pages */
5329 if (j == 0 && freesize > dma_reserve) {
5330 freesize -= dma_reserve;
5331 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5332 zone_names[0], dma_reserve);
5335 if (!is_highmem_idx(j))
5336 nr_kernel_pages += freesize;
5337 /* Charge for highmem memmap if there are enough kernel pages */
5338 else if (nr_kernel_pages > memmap_pages * 2)
5339 nr_kernel_pages -= memmap_pages;
5340 nr_all_pages += freesize;
5343 * Set an approximate value for lowmem here, it will be adjusted
5344 * when the bootmem allocator frees pages into the buddy system.
5345 * And all highmem pages will be managed by the buddy system.
5347 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5350 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5352 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5354 zone->name = zone_names[j];
5355 spin_lock_init(&zone->lock);
5356 spin_lock_init(&zone->lru_lock);
5357 zone_seqlock_init(zone);
5358 zone->zone_pgdat = pgdat;
5359 zone_pcp_init(zone);
5361 /* For bootup, initialized properly in watermark setup */
5362 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5364 lruvec_init(&zone->lruvec);
5368 set_pageblock_order();
5369 setup_usemap(pgdat, zone, zone_start_pfn, size);
5370 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5372 memmap_init(size, nid, j, zone_start_pfn);
5373 zone_start_pfn += size;
5377 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5379 unsigned long __maybe_unused start = 0;
5380 unsigned long __maybe_unused offset = 0;
5382 /* Skip empty nodes */
5383 if (!pgdat->node_spanned_pages)
5386 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5387 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5388 offset = pgdat->node_start_pfn - start;
5389 /* ia64 gets its own node_mem_map, before this, without bootmem */
5390 if (!pgdat->node_mem_map) {
5391 unsigned long size, end;
5395 * The zone's endpoints aren't required to be MAX_ORDER
5396 * aligned but the node_mem_map endpoints must be in order
5397 * for the buddy allocator to function correctly.
5399 end = pgdat_end_pfn(pgdat);
5400 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5401 size = (end - start) * sizeof(struct page);
5402 map = alloc_remap(pgdat->node_id, size);
5404 map = memblock_virt_alloc_node_nopanic(size,
5406 pgdat->node_mem_map = map + offset;
5408 #ifndef CONFIG_NEED_MULTIPLE_NODES
5410 * With no DISCONTIG, the global mem_map is just set as node 0's
5412 if (pgdat == NODE_DATA(0)) {
5413 mem_map = NODE_DATA(0)->node_mem_map;
5414 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5415 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5417 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5420 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5423 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5424 unsigned long node_start_pfn, unsigned long *zholes_size)
5426 pg_data_t *pgdat = NODE_DATA(nid);
5427 unsigned long start_pfn = 0;
5428 unsigned long end_pfn = 0;
5430 /* pg_data_t should be reset to zero when it's allocated */
5431 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5433 pgdat->node_id = nid;
5434 pgdat->node_start_pfn = node_start_pfn;
5435 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5436 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5437 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5438 (u64)start_pfn << PAGE_SHIFT,
5439 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5441 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5442 zones_size, zholes_size);
5444 alloc_node_mem_map(pgdat);
5445 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5446 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5447 nid, (unsigned long)pgdat,
5448 (unsigned long)pgdat->node_mem_map);
5451 reset_deferred_meminit(pgdat);
5452 free_area_init_core(pgdat);
5455 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5457 #if MAX_NUMNODES > 1
5459 * Figure out the number of possible node ids.
5461 void __init setup_nr_node_ids(void)
5463 unsigned int highest;
5465 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5466 nr_node_ids = highest + 1;
5471 * node_map_pfn_alignment - determine the maximum internode alignment
5473 * This function should be called after node map is populated and sorted.
5474 * It calculates the maximum power of two alignment which can distinguish
5477 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5478 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5479 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5480 * shifted, 1GiB is enough and this function will indicate so.
5482 * This is used to test whether pfn -> nid mapping of the chosen memory
5483 * model has fine enough granularity to avoid incorrect mapping for the
5484 * populated node map.
5486 * Returns the determined alignment in pfn's. 0 if there is no alignment
5487 * requirement (single node).
5489 unsigned long __init node_map_pfn_alignment(void)
5491 unsigned long accl_mask = 0, last_end = 0;
5492 unsigned long start, end, mask;
5496 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5497 if (!start || last_nid < 0 || last_nid == nid) {
5504 * Start with a mask granular enough to pin-point to the
5505 * start pfn and tick off bits one-by-one until it becomes
5506 * too coarse to separate the current node from the last.
5508 mask = ~((1 << __ffs(start)) - 1);
5509 while (mask && last_end <= (start & (mask << 1)))
5512 /* accumulate all internode masks */
5516 /* convert mask to number of pages */
5517 return ~accl_mask + 1;
5520 /* Find the lowest pfn for a node */
5521 static unsigned long __init find_min_pfn_for_node(int nid)
5523 unsigned long min_pfn = ULONG_MAX;
5524 unsigned long start_pfn;
5527 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5528 min_pfn = min(min_pfn, start_pfn);
5530 if (min_pfn == ULONG_MAX) {
5532 "Could not find start_pfn for node %d\n", nid);
5540 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5542 * It returns the minimum PFN based on information provided via
5543 * memblock_set_node().
5545 unsigned long __init find_min_pfn_with_active_regions(void)
5547 return find_min_pfn_for_node(MAX_NUMNODES);
5551 * early_calculate_totalpages()
5552 * Sum pages in active regions for movable zone.
5553 * Populate N_MEMORY for calculating usable_nodes.
5555 static unsigned long __init early_calculate_totalpages(void)
5557 unsigned long totalpages = 0;
5558 unsigned long start_pfn, end_pfn;
5561 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5562 unsigned long pages = end_pfn - start_pfn;
5564 totalpages += pages;
5566 node_set_state(nid, N_MEMORY);
5572 * Find the PFN the Movable zone begins in each node. Kernel memory
5573 * is spread evenly between nodes as long as the nodes have enough
5574 * memory. When they don't, some nodes will have more kernelcore than
5577 static void __init find_zone_movable_pfns_for_nodes(void)
5580 unsigned long usable_startpfn;
5581 unsigned long kernelcore_node, kernelcore_remaining;
5582 /* save the state before borrow the nodemask */
5583 nodemask_t saved_node_state = node_states[N_MEMORY];
5584 unsigned long totalpages = early_calculate_totalpages();
5585 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5586 struct memblock_region *r;
5588 /* Need to find movable_zone earlier when movable_node is specified. */
5589 find_usable_zone_for_movable();
5592 * If movable_node is specified, ignore kernelcore and movablecore
5595 if (movable_node_is_enabled()) {
5596 for_each_memblock(memory, r) {
5597 if (!memblock_is_hotpluggable(r))
5602 usable_startpfn = PFN_DOWN(r->base);
5603 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5604 min(usable_startpfn, zone_movable_pfn[nid]) :
5612 * If movablecore=nn[KMG] was specified, calculate what size of
5613 * kernelcore that corresponds so that memory usable for
5614 * any allocation type is evenly spread. If both kernelcore
5615 * and movablecore are specified, then the value of kernelcore
5616 * will be used for required_kernelcore if it's greater than
5617 * what movablecore would have allowed.
5619 if (required_movablecore) {
5620 unsigned long corepages;
5623 * Round-up so that ZONE_MOVABLE is at least as large as what
5624 * was requested by the user
5626 required_movablecore =
5627 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5628 required_movablecore = min(totalpages, required_movablecore);
5629 corepages = totalpages - required_movablecore;
5631 required_kernelcore = max(required_kernelcore, corepages);
5635 * If kernelcore was not specified or kernelcore size is larger
5636 * than totalpages, there is no ZONE_MOVABLE.
5638 if (!required_kernelcore || required_kernelcore >= totalpages)
5641 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5642 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5645 /* Spread kernelcore memory as evenly as possible throughout nodes */
5646 kernelcore_node = required_kernelcore / usable_nodes;
5647 for_each_node_state(nid, N_MEMORY) {
5648 unsigned long start_pfn, end_pfn;
5651 * Recalculate kernelcore_node if the division per node
5652 * now exceeds what is necessary to satisfy the requested
5653 * amount of memory for the kernel
5655 if (required_kernelcore < kernelcore_node)
5656 kernelcore_node = required_kernelcore / usable_nodes;
5659 * As the map is walked, we track how much memory is usable
5660 * by the kernel using kernelcore_remaining. When it is
5661 * 0, the rest of the node is usable by ZONE_MOVABLE
5663 kernelcore_remaining = kernelcore_node;
5665 /* Go through each range of PFNs within this node */
5666 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5667 unsigned long size_pages;
5669 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5670 if (start_pfn >= end_pfn)
5673 /* Account for what is only usable for kernelcore */
5674 if (start_pfn < usable_startpfn) {
5675 unsigned long kernel_pages;
5676 kernel_pages = min(end_pfn, usable_startpfn)
5679 kernelcore_remaining -= min(kernel_pages,
5680 kernelcore_remaining);
5681 required_kernelcore -= min(kernel_pages,
5682 required_kernelcore);
5684 /* Continue if range is now fully accounted */
5685 if (end_pfn <= usable_startpfn) {
5688 * Push zone_movable_pfn to the end so
5689 * that if we have to rebalance
5690 * kernelcore across nodes, we will
5691 * not double account here
5693 zone_movable_pfn[nid] = end_pfn;
5696 start_pfn = usable_startpfn;
5700 * The usable PFN range for ZONE_MOVABLE is from
5701 * start_pfn->end_pfn. Calculate size_pages as the
5702 * number of pages used as kernelcore
5704 size_pages = end_pfn - start_pfn;
5705 if (size_pages > kernelcore_remaining)
5706 size_pages = kernelcore_remaining;
5707 zone_movable_pfn[nid] = start_pfn + size_pages;
5710 * Some kernelcore has been met, update counts and
5711 * break if the kernelcore for this node has been
5714 required_kernelcore -= min(required_kernelcore,
5716 kernelcore_remaining -= size_pages;
5717 if (!kernelcore_remaining)
5723 * If there is still required_kernelcore, we do another pass with one
5724 * less node in the count. This will push zone_movable_pfn[nid] further
5725 * along on the nodes that still have memory until kernelcore is
5729 if (usable_nodes && required_kernelcore > usable_nodes)
5733 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5734 for (nid = 0; nid < MAX_NUMNODES; nid++)
5735 zone_movable_pfn[nid] =
5736 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5739 /* restore the node_state */
5740 node_states[N_MEMORY] = saved_node_state;
5743 /* Any regular or high memory on that node ? */
5744 static void check_for_memory(pg_data_t *pgdat, int nid)
5746 enum zone_type zone_type;
5748 if (N_MEMORY == N_NORMAL_MEMORY)
5751 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5752 struct zone *zone = &pgdat->node_zones[zone_type];
5753 if (populated_zone(zone)) {
5754 node_set_state(nid, N_HIGH_MEMORY);
5755 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5756 zone_type <= ZONE_NORMAL)
5757 node_set_state(nid, N_NORMAL_MEMORY);
5764 * free_area_init_nodes - Initialise all pg_data_t and zone data
5765 * @max_zone_pfn: an array of max PFNs for each zone
5767 * This will call free_area_init_node() for each active node in the system.
5768 * Using the page ranges provided by memblock_set_node(), the size of each
5769 * zone in each node and their holes is calculated. If the maximum PFN
5770 * between two adjacent zones match, it is assumed that the zone is empty.
5771 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5772 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5773 * starts where the previous one ended. For example, ZONE_DMA32 starts
5774 * at arch_max_dma_pfn.
5776 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5778 unsigned long start_pfn, end_pfn;
5781 /* Record where the zone boundaries are */
5782 memset(arch_zone_lowest_possible_pfn, 0,
5783 sizeof(arch_zone_lowest_possible_pfn));
5784 memset(arch_zone_highest_possible_pfn, 0,
5785 sizeof(arch_zone_highest_possible_pfn));
5787 start_pfn = find_min_pfn_with_active_regions();
5789 for (i = 0; i < MAX_NR_ZONES; i++) {
5790 if (i == ZONE_MOVABLE)
5793 end_pfn = max(max_zone_pfn[i], start_pfn);
5794 arch_zone_lowest_possible_pfn[i] = start_pfn;
5795 arch_zone_highest_possible_pfn[i] = end_pfn;
5797 start_pfn = end_pfn;
5799 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5800 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5802 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5803 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5804 find_zone_movable_pfns_for_nodes();
5806 /* Print out the zone ranges */
5807 pr_info("Zone ranges:\n");
5808 for (i = 0; i < MAX_NR_ZONES; i++) {
5809 if (i == ZONE_MOVABLE)
5811 pr_info(" %-8s ", zone_names[i]);
5812 if (arch_zone_lowest_possible_pfn[i] ==
5813 arch_zone_highest_possible_pfn[i])
5816 pr_cont("[mem %#018Lx-%#018Lx]\n",
5817 (u64)arch_zone_lowest_possible_pfn[i]
5819 ((u64)arch_zone_highest_possible_pfn[i]
5820 << PAGE_SHIFT) - 1);
5823 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5824 pr_info("Movable zone start for each node\n");
5825 for (i = 0; i < MAX_NUMNODES; i++) {
5826 if (zone_movable_pfn[i])
5827 pr_info(" Node %d: %#018Lx\n", i,
5828 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5831 /* Print out the early node map */
5832 pr_info("Early memory node ranges\n");
5833 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5834 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5835 (u64)start_pfn << PAGE_SHIFT,
5836 ((u64)end_pfn << PAGE_SHIFT) - 1);
5838 /* Initialise every node */
5839 mminit_verify_pageflags_layout();
5840 setup_nr_node_ids();
5841 for_each_online_node(nid) {
5842 pg_data_t *pgdat = NODE_DATA(nid);
5843 free_area_init_node(nid, NULL,
5844 find_min_pfn_for_node(nid), NULL);
5846 /* Any memory on that node */
5847 if (pgdat->node_present_pages)
5848 node_set_state(nid, N_MEMORY);
5849 check_for_memory(pgdat, nid);
5853 static int __init cmdline_parse_core(char *p, unsigned long *core)
5855 unsigned long long coremem;
5859 coremem = memparse(p, &p);
5860 *core = coremem >> PAGE_SHIFT;
5862 /* Paranoid check that UL is enough for the coremem value */
5863 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5869 * kernelcore=size sets the amount of memory for use for allocations that
5870 * cannot be reclaimed or migrated.
5872 static int __init cmdline_parse_kernelcore(char *p)
5874 return cmdline_parse_core(p, &required_kernelcore);
5878 * movablecore=size sets the amount of memory for use for allocations that
5879 * can be reclaimed or migrated.
5881 static int __init cmdline_parse_movablecore(char *p)
5883 return cmdline_parse_core(p, &required_movablecore);
5886 early_param("kernelcore", cmdline_parse_kernelcore);
5887 early_param("movablecore", cmdline_parse_movablecore);
5889 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5891 void adjust_managed_page_count(struct page *page, long count)
5893 spin_lock(&managed_page_count_lock);
5894 page_zone(page)->managed_pages += count;
5895 totalram_pages += count;
5896 #ifdef CONFIG_HIGHMEM
5897 if (PageHighMem(page))
5898 totalhigh_pages += count;
5900 spin_unlock(&managed_page_count_lock);
5902 EXPORT_SYMBOL(adjust_managed_page_count);
5904 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5907 unsigned long pages = 0;
5909 start = (void *)PAGE_ALIGN((unsigned long)start);
5910 end = (void *)((unsigned long)end & PAGE_MASK);
5911 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5912 if ((unsigned int)poison <= 0xFF)
5913 memset(pos, poison, PAGE_SIZE);
5914 free_reserved_page(virt_to_page(pos));
5918 pr_info("Freeing %s memory: %ldK\n",
5919 s, pages << (PAGE_SHIFT - 10));
5923 EXPORT_SYMBOL(free_reserved_area);
5925 #ifdef CONFIG_HIGHMEM
5926 void free_highmem_page(struct page *page)
5928 __free_reserved_page(page);
5930 page_zone(page)->managed_pages++;
5936 void __init mem_init_print_info(const char *str)
5938 unsigned long physpages, codesize, datasize, rosize, bss_size;
5939 unsigned long init_code_size, init_data_size;
5941 physpages = get_num_physpages();
5942 codesize = _etext - _stext;
5943 datasize = _edata - _sdata;
5944 rosize = __end_rodata - __start_rodata;
5945 bss_size = __bss_stop - __bss_start;
5946 init_data_size = __init_end - __init_begin;
5947 init_code_size = _einittext - _sinittext;
5950 * Detect special cases and adjust section sizes accordingly:
5951 * 1) .init.* may be embedded into .data sections
5952 * 2) .init.text.* may be out of [__init_begin, __init_end],
5953 * please refer to arch/tile/kernel/vmlinux.lds.S.
5954 * 3) .rodata.* may be embedded into .text or .data sections.
5956 #define adj_init_size(start, end, size, pos, adj) \
5958 if (start <= pos && pos < end && size > adj) \
5962 adj_init_size(__init_begin, __init_end, init_data_size,
5963 _sinittext, init_code_size);
5964 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5965 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5966 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5967 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5969 #undef adj_init_size
5971 pr_info("Memory: %luK/%luK available "
5972 "(%luK kernel code, %luK rwdata, %luK rodata, "
5973 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5974 #ifdef CONFIG_HIGHMEM
5978 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5979 codesize >> 10, datasize >> 10, rosize >> 10,
5980 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5981 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5982 totalcma_pages << (PAGE_SHIFT-10),
5983 #ifdef CONFIG_HIGHMEM
5984 totalhigh_pages << (PAGE_SHIFT-10),
5986 str ? ", " : "", str ? str : "");
5990 * set_dma_reserve - set the specified number of pages reserved in the first zone
5991 * @new_dma_reserve: The number of pages to mark reserved
5993 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5994 * In the DMA zone, a significant percentage may be consumed by kernel image
5995 * and other unfreeable allocations which can skew the watermarks badly. This
5996 * function may optionally be used to account for unfreeable pages in the
5997 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5998 * smaller per-cpu batchsize.
6000 void __init set_dma_reserve(unsigned long new_dma_reserve)
6002 dma_reserve = new_dma_reserve;
6005 void __init free_area_init(unsigned long *zones_size)
6007 free_area_init_node(0, zones_size,
6008 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6011 static int page_alloc_cpu_notify(struct notifier_block *self,
6012 unsigned long action, void *hcpu)
6014 int cpu = (unsigned long)hcpu;
6016 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6017 lru_add_drain_cpu(cpu);
6021 * Spill the event counters of the dead processor
6022 * into the current processors event counters.
6023 * This artificially elevates the count of the current
6026 vm_events_fold_cpu(cpu);
6029 * Zero the differential counters of the dead processor
6030 * so that the vm statistics are consistent.
6032 * This is only okay since the processor is dead and cannot
6033 * race with what we are doing.
6035 cpu_vm_stats_fold(cpu);
6040 void __init page_alloc_init(void)
6042 hotcpu_notifier(page_alloc_cpu_notify, 0);
6046 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6047 * or min_free_kbytes changes.
6049 static void calculate_totalreserve_pages(void)
6051 struct pglist_data *pgdat;
6052 unsigned long reserve_pages = 0;
6053 enum zone_type i, j;
6055 for_each_online_pgdat(pgdat) {
6056 for (i = 0; i < MAX_NR_ZONES; i++) {
6057 struct zone *zone = pgdat->node_zones + i;
6060 /* Find valid and maximum lowmem_reserve in the zone */
6061 for (j = i; j < MAX_NR_ZONES; j++) {
6062 if (zone->lowmem_reserve[j] > max)
6063 max = zone->lowmem_reserve[j];
6066 /* we treat the high watermark as reserved pages. */
6067 max += high_wmark_pages(zone);
6069 if (max > zone->managed_pages)
6070 max = zone->managed_pages;
6071 reserve_pages += max;
6073 * Lowmem reserves are not available to
6074 * GFP_HIGHUSER page cache allocations and
6075 * kswapd tries to balance zones to their high
6076 * watermark. As a result, neither should be
6077 * regarded as dirtyable memory, to prevent a
6078 * situation where reclaim has to clean pages
6079 * in order to balance the zones.
6081 zone->dirty_balance_reserve = max;
6084 dirty_balance_reserve = reserve_pages;
6085 totalreserve_pages = reserve_pages;
6089 * setup_per_zone_lowmem_reserve - called whenever
6090 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6091 * has a correct pages reserved value, so an adequate number of
6092 * pages are left in the zone after a successful __alloc_pages().
6094 static void setup_per_zone_lowmem_reserve(void)
6096 struct pglist_data *pgdat;
6097 enum zone_type j, idx;
6099 for_each_online_pgdat(pgdat) {
6100 for (j = 0; j < MAX_NR_ZONES; j++) {
6101 struct zone *zone = pgdat->node_zones + j;
6102 unsigned long managed_pages = zone->managed_pages;
6104 zone->lowmem_reserve[j] = 0;
6108 struct zone *lower_zone;
6112 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6113 sysctl_lowmem_reserve_ratio[idx] = 1;
6115 lower_zone = pgdat->node_zones + idx;
6116 lower_zone->lowmem_reserve[j] = managed_pages /
6117 sysctl_lowmem_reserve_ratio[idx];
6118 managed_pages += lower_zone->managed_pages;
6123 /* update totalreserve_pages */
6124 calculate_totalreserve_pages();
6127 static void __setup_per_zone_wmarks(void)
6129 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6130 unsigned long lowmem_pages = 0;
6132 unsigned long flags;
6134 /* Calculate total number of !ZONE_HIGHMEM pages */
6135 for_each_zone(zone) {
6136 if (!is_highmem(zone))
6137 lowmem_pages += zone->managed_pages;
6140 for_each_zone(zone) {
6143 spin_lock_irqsave(&zone->lock, flags);
6144 tmp = (u64)pages_min * zone->managed_pages;
6145 do_div(tmp, lowmem_pages);
6146 if (is_highmem(zone)) {
6148 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6149 * need highmem pages, so cap pages_min to a small
6152 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6153 * deltas control asynch page reclaim, and so should
6154 * not be capped for highmem.
6156 unsigned long min_pages;
6158 min_pages = zone->managed_pages / 1024;
6159 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6160 zone->watermark[WMARK_MIN] = min_pages;
6163 * If it's a lowmem zone, reserve a number of pages
6164 * proportionate to the zone's size.
6166 zone->watermark[WMARK_MIN] = tmp;
6169 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6170 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6172 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6173 high_wmark_pages(zone) - low_wmark_pages(zone) -
6174 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6176 spin_unlock_irqrestore(&zone->lock, flags);
6179 /* update totalreserve_pages */
6180 calculate_totalreserve_pages();
6184 * setup_per_zone_wmarks - called when min_free_kbytes changes
6185 * or when memory is hot-{added|removed}
6187 * Ensures that the watermark[min,low,high] values for each zone are set
6188 * correctly with respect to min_free_kbytes.
6190 void setup_per_zone_wmarks(void)
6192 mutex_lock(&zonelists_mutex);
6193 __setup_per_zone_wmarks();
6194 mutex_unlock(&zonelists_mutex);
6198 * The inactive anon list should be small enough that the VM never has to
6199 * do too much work, but large enough that each inactive page has a chance
6200 * to be referenced again before it is swapped out.
6202 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6203 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6204 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6205 * the anonymous pages are kept on the inactive list.
6208 * memory ratio inactive anon
6209 * -------------------------------------
6218 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6220 unsigned int gb, ratio;
6222 /* Zone size in gigabytes */
6223 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6225 ratio = int_sqrt(10 * gb);
6229 zone->inactive_ratio = ratio;
6232 static void __meminit setup_per_zone_inactive_ratio(void)
6237 calculate_zone_inactive_ratio(zone);
6241 * Initialise min_free_kbytes.
6243 * For small machines we want it small (128k min). For large machines
6244 * we want it large (64MB max). But it is not linear, because network
6245 * bandwidth does not increase linearly with machine size. We use
6247 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6248 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6264 int __meminit init_per_zone_wmark_min(void)
6266 unsigned long lowmem_kbytes;
6267 int new_min_free_kbytes;
6269 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6270 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6272 if (new_min_free_kbytes > user_min_free_kbytes) {
6273 min_free_kbytes = new_min_free_kbytes;
6274 if (min_free_kbytes < 128)
6275 min_free_kbytes = 128;
6276 if (min_free_kbytes > 65536)
6277 min_free_kbytes = 65536;
6279 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6280 new_min_free_kbytes, user_min_free_kbytes);
6282 setup_per_zone_wmarks();
6283 refresh_zone_stat_thresholds();
6284 setup_per_zone_lowmem_reserve();
6285 setup_per_zone_inactive_ratio();
6288 core_initcall(init_per_zone_wmark_min)
6291 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6292 * that we can call two helper functions whenever min_free_kbytes
6295 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6296 void __user *buffer, size_t *length, loff_t *ppos)
6300 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6305 user_min_free_kbytes = min_free_kbytes;
6306 setup_per_zone_wmarks();
6312 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6313 void __user *buffer, size_t *length, loff_t *ppos)
6318 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6323 zone->min_unmapped_pages = (zone->managed_pages *
6324 sysctl_min_unmapped_ratio) / 100;
6328 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6329 void __user *buffer, size_t *length, loff_t *ppos)
6334 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6339 zone->min_slab_pages = (zone->managed_pages *
6340 sysctl_min_slab_ratio) / 100;
6346 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6347 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6348 * whenever sysctl_lowmem_reserve_ratio changes.
6350 * The reserve ratio obviously has absolutely no relation with the
6351 * minimum watermarks. The lowmem reserve ratio can only make sense
6352 * if in function of the boot time zone sizes.
6354 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6355 void __user *buffer, size_t *length, loff_t *ppos)
6357 proc_dointvec_minmax(table, write, buffer, length, ppos);
6358 setup_per_zone_lowmem_reserve();
6363 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6364 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6365 * pagelist can have before it gets flushed back to buddy allocator.
6367 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6368 void __user *buffer, size_t *length, loff_t *ppos)
6371 int old_percpu_pagelist_fraction;
6374 mutex_lock(&pcp_batch_high_lock);
6375 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6377 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6378 if (!write || ret < 0)
6381 /* Sanity checking to avoid pcp imbalance */
6382 if (percpu_pagelist_fraction &&
6383 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6384 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6390 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6393 for_each_populated_zone(zone) {
6396 for_each_possible_cpu(cpu)
6397 pageset_set_high_and_batch(zone,
6398 per_cpu_ptr(zone->pageset, cpu));
6401 mutex_unlock(&pcp_batch_high_lock);
6406 int hashdist = HASHDIST_DEFAULT;
6408 static int __init set_hashdist(char *str)
6412 hashdist = simple_strtoul(str, &str, 0);
6415 __setup("hashdist=", set_hashdist);
6419 * allocate a large system hash table from bootmem
6420 * - it is assumed that the hash table must contain an exact power-of-2
6421 * quantity of entries
6422 * - limit is the number of hash buckets, not the total allocation size
6424 void *__init alloc_large_system_hash(const char *tablename,
6425 unsigned long bucketsize,
6426 unsigned long numentries,
6429 unsigned int *_hash_shift,
6430 unsigned int *_hash_mask,
6431 unsigned long low_limit,
6432 unsigned long high_limit)
6434 unsigned long long max = high_limit;
6435 unsigned long log2qty, size;
6438 /* allow the kernel cmdline to have a say */
6440 /* round applicable memory size up to nearest megabyte */
6441 numentries = nr_kernel_pages;
6443 /* It isn't necessary when PAGE_SIZE >= 1MB */
6444 if (PAGE_SHIFT < 20)
6445 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6447 /* limit to 1 bucket per 2^scale bytes of low memory */
6448 if (scale > PAGE_SHIFT)
6449 numentries >>= (scale - PAGE_SHIFT);
6451 numentries <<= (PAGE_SHIFT - scale);
6453 /* Make sure we've got at least a 0-order allocation.. */
6454 if (unlikely(flags & HASH_SMALL)) {
6455 /* Makes no sense without HASH_EARLY */
6456 WARN_ON(!(flags & HASH_EARLY));
6457 if (!(numentries >> *_hash_shift)) {
6458 numentries = 1UL << *_hash_shift;
6459 BUG_ON(!numentries);
6461 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6462 numentries = PAGE_SIZE / bucketsize;
6464 numentries = roundup_pow_of_two(numentries);
6466 /* limit allocation size to 1/16 total memory by default */
6468 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6469 do_div(max, bucketsize);
6471 max = min(max, 0x80000000ULL);
6473 if (numentries < low_limit)
6474 numentries = low_limit;
6475 if (numentries > max)
6478 log2qty = ilog2(numentries);
6481 size = bucketsize << log2qty;
6482 if (flags & HASH_EARLY)
6483 table = memblock_virt_alloc_nopanic(size, 0);
6485 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6488 * If bucketsize is not a power-of-two, we may free
6489 * some pages at the end of hash table which
6490 * alloc_pages_exact() automatically does
6492 if (get_order(size) < MAX_ORDER) {
6493 table = alloc_pages_exact(size, GFP_ATOMIC);
6494 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6497 } while (!table && size > PAGE_SIZE && --log2qty);
6500 panic("Failed to allocate %s hash table\n", tablename);
6502 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6505 ilog2(size) - PAGE_SHIFT,
6509 *_hash_shift = log2qty;
6511 *_hash_mask = (1 << log2qty) - 1;
6516 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6517 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6520 #ifdef CONFIG_SPARSEMEM
6521 return __pfn_to_section(pfn)->pageblock_flags;
6523 return zone->pageblock_flags;
6524 #endif /* CONFIG_SPARSEMEM */
6527 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6529 #ifdef CONFIG_SPARSEMEM
6530 pfn &= (PAGES_PER_SECTION-1);
6531 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6533 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6534 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6535 #endif /* CONFIG_SPARSEMEM */
6539 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6540 * @page: The page within the block of interest
6541 * @pfn: The target page frame number
6542 * @end_bitidx: The last bit of interest to retrieve
6543 * @mask: mask of bits that the caller is interested in
6545 * Return: pageblock_bits flags
6547 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6548 unsigned long end_bitidx,
6552 unsigned long *bitmap;
6553 unsigned long bitidx, word_bitidx;
6556 zone = page_zone(page);
6557 bitmap = get_pageblock_bitmap(zone, pfn);
6558 bitidx = pfn_to_bitidx(zone, pfn);
6559 word_bitidx = bitidx / BITS_PER_LONG;
6560 bitidx &= (BITS_PER_LONG-1);
6562 word = bitmap[word_bitidx];
6563 bitidx += end_bitidx;
6564 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6568 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6569 * @page: The page within the block of interest
6570 * @flags: The flags to set
6571 * @pfn: The target page frame number
6572 * @end_bitidx: The last bit of interest
6573 * @mask: mask of bits that the caller is interested in
6575 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6577 unsigned long end_bitidx,
6581 unsigned long *bitmap;
6582 unsigned long bitidx, word_bitidx;
6583 unsigned long old_word, word;
6585 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6587 zone = page_zone(page);
6588 bitmap = get_pageblock_bitmap(zone, pfn);
6589 bitidx = pfn_to_bitidx(zone, pfn);
6590 word_bitidx = bitidx / BITS_PER_LONG;
6591 bitidx &= (BITS_PER_LONG-1);
6593 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6595 bitidx += end_bitidx;
6596 mask <<= (BITS_PER_LONG - bitidx - 1);
6597 flags <<= (BITS_PER_LONG - bitidx - 1);
6599 word = READ_ONCE(bitmap[word_bitidx]);
6601 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6602 if (word == old_word)
6609 * This function checks whether pageblock includes unmovable pages or not.
6610 * If @count is not zero, it is okay to include less @count unmovable pages
6612 * PageLRU check without isolation or lru_lock could race so that
6613 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6614 * expect this function should be exact.
6616 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6617 bool skip_hwpoisoned_pages)
6619 unsigned long pfn, iter, found;
6623 * For avoiding noise data, lru_add_drain_all() should be called
6624 * If ZONE_MOVABLE, the zone never contains unmovable pages
6626 if (zone_idx(zone) == ZONE_MOVABLE)
6628 mt = get_pageblock_migratetype(page);
6629 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6632 pfn = page_to_pfn(page);
6633 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6634 unsigned long check = pfn + iter;
6636 if (!pfn_valid_within(check))
6639 page = pfn_to_page(check);
6642 * Hugepages are not in LRU lists, but they're movable.
6643 * We need not scan over tail pages bacause we don't
6644 * handle each tail page individually in migration.
6646 if (PageHuge(page)) {
6647 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6652 * We can't use page_count without pin a page
6653 * because another CPU can free compound page.
6654 * This check already skips compound tails of THP
6655 * because their page->_count is zero at all time.
6657 if (!atomic_read(&page->_count)) {
6658 if (PageBuddy(page))
6659 iter += (1 << page_order(page)) - 1;
6664 * The HWPoisoned page may be not in buddy system, and
6665 * page_count() is not 0.
6667 if (skip_hwpoisoned_pages && PageHWPoison(page))
6673 * If there are RECLAIMABLE pages, we need to check
6674 * it. But now, memory offline itself doesn't call
6675 * shrink_node_slabs() and it still to be fixed.
6678 * If the page is not RAM, page_count()should be 0.
6679 * we don't need more check. This is an _used_ not-movable page.
6681 * The problematic thing here is PG_reserved pages. PG_reserved
6682 * is set to both of a memory hole page and a _used_ kernel
6691 bool is_pageblock_removable_nolock(struct page *page)
6697 * We have to be careful here because we are iterating over memory
6698 * sections which are not zone aware so we might end up outside of
6699 * the zone but still within the section.
6700 * We have to take care about the node as well. If the node is offline
6701 * its NODE_DATA will be NULL - see page_zone.
6703 if (!node_online(page_to_nid(page)))
6706 zone = page_zone(page);
6707 pfn = page_to_pfn(page);
6708 if (!zone_spans_pfn(zone, pfn))
6711 return !has_unmovable_pages(zone, page, 0, true);
6716 static unsigned long pfn_max_align_down(unsigned long pfn)
6718 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6719 pageblock_nr_pages) - 1);
6722 static unsigned long pfn_max_align_up(unsigned long pfn)
6724 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6725 pageblock_nr_pages));
6728 /* [start, end) must belong to a single zone. */
6729 static int __alloc_contig_migrate_range(struct compact_control *cc,
6730 unsigned long start, unsigned long end)
6732 /* This function is based on compact_zone() from compaction.c. */
6733 unsigned long nr_reclaimed;
6734 unsigned long pfn = start;
6735 unsigned int tries = 0;
6740 while (pfn < end || !list_empty(&cc->migratepages)) {
6741 if (fatal_signal_pending(current)) {
6746 if (list_empty(&cc->migratepages)) {
6747 cc->nr_migratepages = 0;
6748 pfn = isolate_migratepages_range(cc, pfn, end);
6754 } else if (++tries == 5) {
6755 ret = ret < 0 ? ret : -EBUSY;
6759 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6761 cc->nr_migratepages -= nr_reclaimed;
6763 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6764 NULL, 0, cc->mode, MR_CMA);
6767 putback_movable_pages(&cc->migratepages);
6774 * alloc_contig_range() -- tries to allocate given range of pages
6775 * @start: start PFN to allocate
6776 * @end: one-past-the-last PFN to allocate
6777 * @migratetype: migratetype of the underlaying pageblocks (either
6778 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6779 * in range must have the same migratetype and it must
6780 * be either of the two.
6782 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6783 * aligned, however it's the caller's responsibility to guarantee that
6784 * we are the only thread that changes migrate type of pageblocks the
6787 * The PFN range must belong to a single zone.
6789 * Returns zero on success or negative error code. On success all
6790 * pages which PFN is in [start, end) are allocated for the caller and
6791 * need to be freed with free_contig_range().
6793 int alloc_contig_range(unsigned long start, unsigned long end,
6794 unsigned migratetype)
6796 unsigned long outer_start, outer_end;
6800 struct compact_control cc = {
6801 .nr_migratepages = 0,
6803 .zone = page_zone(pfn_to_page(start)),
6804 .mode = MIGRATE_SYNC,
6805 .ignore_skip_hint = true,
6807 INIT_LIST_HEAD(&cc.migratepages);
6810 * What we do here is we mark all pageblocks in range as
6811 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6812 * have different sizes, and due to the way page allocator
6813 * work, we align the range to biggest of the two pages so
6814 * that page allocator won't try to merge buddies from
6815 * different pageblocks and change MIGRATE_ISOLATE to some
6816 * other migration type.
6818 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6819 * migrate the pages from an unaligned range (ie. pages that
6820 * we are interested in). This will put all the pages in
6821 * range back to page allocator as MIGRATE_ISOLATE.
6823 * When this is done, we take the pages in range from page
6824 * allocator removing them from the buddy system. This way
6825 * page allocator will never consider using them.
6827 * This lets us mark the pageblocks back as
6828 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6829 * aligned range but not in the unaligned, original range are
6830 * put back to page allocator so that buddy can use them.
6833 ret = start_isolate_page_range(pfn_max_align_down(start),
6834 pfn_max_align_up(end), migratetype,
6839 ret = __alloc_contig_migrate_range(&cc, start, end);
6844 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6845 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6846 * more, all pages in [start, end) are free in page allocator.
6847 * What we are going to do is to allocate all pages from
6848 * [start, end) (that is remove them from page allocator).
6850 * The only problem is that pages at the beginning and at the
6851 * end of interesting range may be not aligned with pages that
6852 * page allocator holds, ie. they can be part of higher order
6853 * pages. Because of this, we reserve the bigger range and
6854 * once this is done free the pages we are not interested in.
6856 * We don't have to hold zone->lock here because the pages are
6857 * isolated thus they won't get removed from buddy.
6860 lru_add_drain_all();
6861 drain_all_pages(cc.zone);
6864 outer_start = start;
6865 while (!PageBuddy(pfn_to_page(outer_start))) {
6866 if (++order >= MAX_ORDER) {
6870 outer_start &= ~0UL << order;
6873 /* Make sure the range is really isolated. */
6874 if (test_pages_isolated(outer_start, end, false)) {
6875 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
6876 __func__, outer_start, end);
6881 /* Grab isolated pages from freelists. */
6882 outer_end = isolate_freepages_range(&cc, outer_start, end);
6888 /* Free head and tail (if any) */
6889 if (start != outer_start)
6890 free_contig_range(outer_start, start - outer_start);
6891 if (end != outer_end)
6892 free_contig_range(end, outer_end - end);
6895 undo_isolate_page_range(pfn_max_align_down(start),
6896 pfn_max_align_up(end), migratetype);
6900 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6902 unsigned int count = 0;
6904 for (; nr_pages--; pfn++) {
6905 struct page *page = pfn_to_page(pfn);
6907 count += page_count(page) != 1;
6910 WARN(count != 0, "%d pages are still in use!\n", count);
6914 #ifdef CONFIG_MEMORY_HOTPLUG
6916 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6917 * page high values need to be recalulated.
6919 void __meminit zone_pcp_update(struct zone *zone)
6922 mutex_lock(&pcp_batch_high_lock);
6923 for_each_possible_cpu(cpu)
6924 pageset_set_high_and_batch(zone,
6925 per_cpu_ptr(zone->pageset, cpu));
6926 mutex_unlock(&pcp_batch_high_lock);
6930 void zone_pcp_reset(struct zone *zone)
6932 unsigned long flags;
6934 struct per_cpu_pageset *pset;
6936 /* avoid races with drain_pages() */
6937 local_irq_save(flags);
6938 if (zone->pageset != &boot_pageset) {
6939 for_each_online_cpu(cpu) {
6940 pset = per_cpu_ptr(zone->pageset, cpu);
6941 drain_zonestat(zone, pset);
6943 free_percpu(zone->pageset);
6944 zone->pageset = &boot_pageset;
6946 local_irq_restore(flags);
6949 #ifdef CONFIG_MEMORY_HOTREMOVE
6951 * All pages in the range must be isolated before calling this.
6954 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6958 unsigned int order, i;
6960 unsigned long flags;
6961 /* find the first valid pfn */
6962 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6967 zone = page_zone(pfn_to_page(pfn));
6968 spin_lock_irqsave(&zone->lock, flags);
6970 while (pfn < end_pfn) {
6971 if (!pfn_valid(pfn)) {
6975 page = pfn_to_page(pfn);
6977 * The HWPoisoned page may be not in buddy system, and
6978 * page_count() is not 0.
6980 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6982 SetPageReserved(page);
6986 BUG_ON(page_count(page));
6987 BUG_ON(!PageBuddy(page));
6988 order = page_order(page);
6989 #ifdef CONFIG_DEBUG_VM
6990 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6991 pfn, 1 << order, end_pfn);
6993 list_del(&page->lru);
6994 rmv_page_order(page);
6995 zone->free_area[order].nr_free--;
6996 for (i = 0; i < (1 << order); i++)
6997 SetPageReserved((page+i));
6998 pfn += (1 << order);
7000 spin_unlock_irqrestore(&zone->lock, flags);
7004 #ifdef CONFIG_MEMORY_FAILURE
7005 bool is_free_buddy_page(struct page *page)
7007 struct zone *zone = page_zone(page);
7008 unsigned long pfn = page_to_pfn(page);
7009 unsigned long flags;
7012 spin_lock_irqsave(&zone->lock, flags);
7013 for (order = 0; order < MAX_ORDER; order++) {
7014 struct page *page_head = page - (pfn & ((1 << order) - 1));
7016 if (PageBuddy(page_head) && page_order(page_head) >= order)
7019 spin_unlock_irqrestore(&zone->lock, flags);
7021 return order < MAX_ORDER;