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
242 * Try to keep at least this much lowmem free. Do not allow normal
243 * allocations below this point, only high priority ones. Automatically
244 * tuned according to the amount of memory in the system.
246 int min_free_kbytes = 1024;
247 int user_min_free_kbytes = -1;
250 * Extra memory for the system to try freeing. Used to temporarily
251 * free memory, to make space for new workloads. Anyone can allocate
252 * down to the min watermarks controlled by min_free_kbytes above.
254 int extra_free_kbytes = 0;
256 static unsigned long __meminitdata nr_kernel_pages;
257 static unsigned long __meminitdata nr_all_pages;
258 static unsigned long __meminitdata dma_reserve;
260 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
261 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
262 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
263 static unsigned long __initdata required_kernelcore;
264 static unsigned long __initdata required_movablecore;
265 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
267 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
269 EXPORT_SYMBOL(movable_zone);
270 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
273 int nr_node_ids __read_mostly = MAX_NUMNODES;
274 int nr_online_nodes __read_mostly = 1;
275 EXPORT_SYMBOL(nr_node_ids);
276 EXPORT_SYMBOL(nr_online_nodes);
279 int page_group_by_mobility_disabled __read_mostly;
281 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
284 * Determine how many pages need to be initialized durig early boot
285 * (non-deferred initialization).
286 * The value of first_deferred_pfn will be set later, once non-deferred pages
287 * are initialized, but for now set it ULONG_MAX.
289 static inline void reset_deferred_meminit(pg_data_t *pgdat)
291 phys_addr_t start_addr, end_addr;
292 unsigned long max_pgcnt;
293 unsigned long reserved;
296 * Initialise at least 2G of a node but also take into account that
297 * two large system hashes that can take up 1GB for 0.25TB/node.
299 max_pgcnt = max(2UL << (30 - PAGE_SHIFT),
300 (pgdat->node_spanned_pages >> 8));
303 * Compensate the all the memblock reservations (e.g. crash kernel)
304 * from the initial estimation to make sure we will initialize enough
307 start_addr = PFN_PHYS(pgdat->node_start_pfn);
308 end_addr = PFN_PHYS(pgdat->node_start_pfn + max_pgcnt);
309 reserved = memblock_reserved_memory_within(start_addr, end_addr);
310 max_pgcnt += PHYS_PFN(reserved);
312 pgdat->static_init_pgcnt = min(max_pgcnt, pgdat->node_spanned_pages);
313 pgdat->first_deferred_pfn = ULONG_MAX;
316 /* Returns true if the struct page for the pfn is uninitialised */
317 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
319 int nid = early_pfn_to_nid(pfn);
321 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
327 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
329 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
336 * Returns false when the remaining initialisation should be deferred until
337 * later in the boot cycle when it can be parallelised.
339 static inline bool update_defer_init(pg_data_t *pgdat,
340 unsigned long pfn, unsigned long zone_end,
341 unsigned long *nr_initialised)
343 /* Always populate low zones for address-contrained allocations */
344 if (zone_end < pgdat_end_pfn(pgdat))
346 /* Initialise at least 2G of the highest zone */
348 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
349 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
350 pgdat->first_deferred_pfn = pfn;
357 static inline void reset_deferred_meminit(pg_data_t *pgdat)
361 static inline bool early_page_uninitialised(unsigned long pfn)
366 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
371 static inline bool update_defer_init(pg_data_t *pgdat,
372 unsigned long pfn, unsigned long zone_end,
373 unsigned long *nr_initialised)
380 void set_pageblock_migratetype(struct page *page, int migratetype)
382 if (unlikely(page_group_by_mobility_disabled &&
383 migratetype < MIGRATE_PCPTYPES))
384 migratetype = MIGRATE_UNMOVABLE;
386 set_pageblock_flags_group(page, (unsigned long)migratetype,
387 PB_migrate, PB_migrate_end);
390 #ifdef CONFIG_DEBUG_VM
391 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
395 unsigned long pfn = page_to_pfn(page);
396 unsigned long sp, start_pfn;
399 seq = zone_span_seqbegin(zone);
400 start_pfn = zone->zone_start_pfn;
401 sp = zone->spanned_pages;
402 if (!zone_spans_pfn(zone, pfn))
404 } while (zone_span_seqretry(zone, seq));
407 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
408 pfn, zone_to_nid(zone), zone->name,
409 start_pfn, start_pfn + sp);
414 static int page_is_consistent(struct zone *zone, struct page *page)
416 if (!pfn_valid_within(page_to_pfn(page)))
418 if (zone != page_zone(page))
424 * Temporary debugging check for pages not lying within a given zone.
426 static int bad_range(struct zone *zone, struct page *page)
428 if (page_outside_zone_boundaries(zone, page))
430 if (!page_is_consistent(zone, page))
436 static inline int bad_range(struct zone *zone, struct page *page)
442 static void bad_page(struct page *page, const char *reason,
443 unsigned long bad_flags)
445 static unsigned long resume;
446 static unsigned long nr_shown;
447 static unsigned long nr_unshown;
449 /* Don't complain about poisoned pages */
450 if (PageHWPoison(page)) {
451 page_mapcount_reset(page); /* remove PageBuddy */
456 * Allow a burst of 60 reports, then keep quiet for that minute;
457 * or allow a steady drip of one report per second.
459 if (nr_shown == 60) {
460 if (time_before(jiffies, resume)) {
466 "BUG: Bad page state: %lu messages suppressed\n",
473 resume = jiffies + 60 * HZ;
475 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
476 current->comm, page_to_pfn(page));
477 dump_page_badflags(page, reason, bad_flags);
482 /* Leave bad fields for debug, except PageBuddy could make trouble */
483 page_mapcount_reset(page); /* remove PageBuddy */
484 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
488 * Higher-order pages are called "compound pages". They are structured thusly:
490 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
492 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
493 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
495 * The first tail page's ->compound_dtor holds the offset in array of compound
496 * page destructors. See compound_page_dtors.
498 * The first tail page's ->compound_order holds the order of allocation.
499 * This usage means that zero-order pages may not be compound.
502 static void free_compound_page(struct page *page)
504 __free_pages_ok(page, compound_order(page));
507 void prep_compound_page(struct page *page, unsigned int order)
510 int nr_pages = 1 << order;
512 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
513 set_compound_order(page, order);
515 for (i = 1; i < nr_pages; i++) {
516 struct page *p = page + i;
517 set_page_count(p, 0);
518 set_compound_head(p, page);
522 #ifdef CONFIG_DEBUG_PAGEALLOC
523 unsigned int _debug_guardpage_minorder;
524 bool _debug_pagealloc_enabled __read_mostly;
525 bool _debug_guardpage_enabled __read_mostly;
527 static int __init early_debug_pagealloc(char *buf)
532 if (strcmp(buf, "on") == 0)
533 _debug_pagealloc_enabled = true;
537 early_param("debug_pagealloc", early_debug_pagealloc);
539 static bool need_debug_guardpage(void)
541 /* If we don't use debug_pagealloc, we don't need guard page */
542 if (!debug_pagealloc_enabled())
548 static void init_debug_guardpage(void)
550 if (!debug_pagealloc_enabled())
553 _debug_guardpage_enabled = true;
556 struct page_ext_operations debug_guardpage_ops = {
557 .need = need_debug_guardpage,
558 .init = init_debug_guardpage,
561 static int __init debug_guardpage_minorder_setup(char *buf)
565 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
566 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
569 _debug_guardpage_minorder = res;
570 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
573 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
575 static inline void set_page_guard(struct zone *zone, struct page *page,
576 unsigned int order, int migratetype)
578 struct page_ext *page_ext;
580 if (!debug_guardpage_enabled())
583 page_ext = lookup_page_ext(page);
584 if (unlikely(!page_ext))
587 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
589 INIT_LIST_HEAD(&page->lru);
590 set_page_private(page, order);
591 /* Guard pages are not available for any usage */
592 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
595 static inline void clear_page_guard(struct zone *zone, struct page *page,
596 unsigned int order, int migratetype)
598 struct page_ext *page_ext;
600 if (!debug_guardpage_enabled())
603 page_ext = lookup_page_ext(page);
604 if (unlikely(!page_ext))
607 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
609 set_page_private(page, 0);
610 if (!is_migrate_isolate(migratetype))
611 __mod_zone_freepage_state(zone, (1 << order), migratetype);
614 struct page_ext_operations debug_guardpage_ops = { NULL, };
615 static inline void set_page_guard(struct zone *zone, struct page *page,
616 unsigned int order, int migratetype) {}
617 static inline void clear_page_guard(struct zone *zone, struct page *page,
618 unsigned int order, int migratetype) {}
621 static inline void set_page_order(struct page *page, unsigned int order)
623 set_page_private(page, order);
624 __SetPageBuddy(page);
627 static inline void rmv_page_order(struct page *page)
629 __ClearPageBuddy(page);
630 set_page_private(page, 0);
634 * This function checks whether a page is free && is the buddy
635 * we can do coalesce a page and its buddy if
636 * (a) the buddy is not in a hole &&
637 * (b) the buddy is in the buddy system &&
638 * (c) a page and its buddy have the same order &&
639 * (d) a page and its buddy are in the same zone.
641 * For recording whether a page is in the buddy system, we set ->_mapcount
642 * PAGE_BUDDY_MAPCOUNT_VALUE.
643 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
644 * serialized by zone->lock.
646 * For recording page's order, we use page_private(page).
648 static inline int page_is_buddy(struct page *page, struct page *buddy,
651 if (!pfn_valid_within(page_to_pfn(buddy)))
654 if (page_is_guard(buddy) && page_order(buddy) == order) {
655 if (page_zone_id(page) != page_zone_id(buddy))
658 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
663 if (PageBuddy(buddy) && page_order(buddy) == order) {
665 * zone check is done late to avoid uselessly
666 * calculating zone/node ids for pages that could
669 if (page_zone_id(page) != page_zone_id(buddy))
672 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
680 * Freeing function for a buddy system allocator.
682 * The concept of a buddy system is to maintain direct-mapped table
683 * (containing bit values) for memory blocks of various "orders".
684 * The bottom level table contains the map for the smallest allocatable
685 * units of memory (here, pages), and each level above it describes
686 * pairs of units from the levels below, hence, "buddies".
687 * At a high level, all that happens here is marking the table entry
688 * at the bottom level available, and propagating the changes upward
689 * as necessary, plus some accounting needed to play nicely with other
690 * parts of the VM system.
691 * At each level, we keep a list of pages, which are heads of continuous
692 * free pages of length of (1 << order) and marked with _mapcount
693 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
695 * So when we are allocating or freeing one, we can derive the state of the
696 * other. That is, if we allocate a small block, and both were
697 * free, the remainder of the region must be split into blocks.
698 * If a block is freed, and its buddy is also free, then this
699 * triggers coalescing into a block of larger size.
704 static inline void __free_one_page(struct page *page,
706 struct zone *zone, unsigned int order,
709 unsigned long page_idx;
710 unsigned long combined_idx;
711 unsigned long uninitialized_var(buddy_idx);
713 unsigned int max_order;
715 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
717 VM_BUG_ON(!zone_is_initialized(zone));
718 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
720 VM_BUG_ON(migratetype == -1);
721 if (likely(!is_migrate_isolate(migratetype)))
722 __mod_zone_freepage_state(zone, 1 << order, migratetype);
724 page_idx = pfn & ((1 << MAX_ORDER) - 1);
726 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
727 VM_BUG_ON_PAGE(bad_range(zone, page), page);
730 while (order < max_order - 1) {
731 buddy_idx = __find_buddy_index(page_idx, order);
732 buddy = page + (buddy_idx - page_idx);
733 if (!page_is_buddy(page, buddy, order))
736 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
737 * merge with it and move up one order.
739 if (page_is_guard(buddy)) {
740 clear_page_guard(zone, buddy, order, migratetype);
742 list_del(&buddy->lru);
743 zone->free_area[order].nr_free--;
744 rmv_page_order(buddy);
746 combined_idx = buddy_idx & page_idx;
747 page = page + (combined_idx - page_idx);
748 page_idx = combined_idx;
751 if (max_order < MAX_ORDER) {
752 /* If we are here, it means order is >= pageblock_order.
753 * We want to prevent merge between freepages on isolate
754 * pageblock and normal pageblock. Without this, pageblock
755 * isolation could cause incorrect freepage or CMA accounting.
757 * We don't want to hit this code for the more frequent
760 if (unlikely(has_isolate_pageblock(zone))) {
763 buddy_idx = __find_buddy_index(page_idx, order);
764 buddy = page + (buddy_idx - page_idx);
765 buddy_mt = get_pageblock_migratetype(buddy);
767 if (migratetype != buddy_mt
768 && (is_migrate_isolate(migratetype) ||
769 is_migrate_isolate(buddy_mt)))
773 goto continue_merging;
777 set_page_order(page, order);
780 * If this is not the largest possible page, check if the buddy
781 * of the next-highest order is free. If it is, it's possible
782 * that pages are being freed that will coalesce soon. In case,
783 * that is happening, add the free page to the tail of the list
784 * so it's less likely to be used soon and more likely to be merged
785 * as a higher order page
787 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
788 struct page *higher_page, *higher_buddy;
789 combined_idx = buddy_idx & page_idx;
790 higher_page = page + (combined_idx - page_idx);
791 buddy_idx = __find_buddy_index(combined_idx, order + 1);
792 higher_buddy = higher_page + (buddy_idx - combined_idx);
793 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
794 list_add_tail(&page->lru,
795 &zone->free_area[order].free_list[migratetype]);
800 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
802 zone->free_area[order].nr_free++;
805 static inline int free_pages_check(struct page *page)
807 const char *bad_reason = NULL;
808 unsigned long bad_flags = 0;
810 if (unlikely(page_mapcount(page)))
811 bad_reason = "nonzero mapcount";
812 if (unlikely(page->mapping != NULL))
813 bad_reason = "non-NULL mapping";
814 if (unlikely(atomic_read(&page->_count) != 0))
815 bad_reason = "nonzero _count";
816 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
817 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
818 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
821 if (unlikely(page->mem_cgroup))
822 bad_reason = "page still charged to cgroup";
824 if (unlikely(bad_reason)) {
825 bad_page(page, bad_reason, bad_flags);
828 page_cpupid_reset_last(page);
829 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
830 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
835 * Frees a number of pages from the PCP lists
836 * Assumes all pages on list are in same zone, and of same order.
837 * count is the number of pages to free.
839 * If the zone was previously in an "all pages pinned" state then look to
840 * see if this freeing clears that state.
842 * And clear the zone's pages_scanned counter, to hold off the "all pages are
843 * pinned" detection logic.
845 static void free_pcppages_bulk(struct zone *zone, int count,
846 struct per_cpu_pages *pcp)
851 unsigned long nr_scanned;
853 spin_lock(&zone->lock);
854 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
856 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
860 struct list_head *list;
863 * Remove pages from lists in a round-robin fashion. A
864 * batch_free count is maintained that is incremented when an
865 * empty list is encountered. This is so more pages are freed
866 * off fuller lists instead of spinning excessively around empty
871 if (++migratetype == MIGRATE_PCPTYPES)
873 list = &pcp->lists[migratetype];
874 } while (list_empty(list));
876 /* This is the only non-empty list. Free them all. */
877 if (batch_free == MIGRATE_PCPTYPES)
878 batch_free = to_free;
881 int mt; /* migratetype of the to-be-freed page */
883 page = list_entry(list->prev, struct page, lru);
884 /* must delete as __free_one_page list manipulates */
885 list_del(&page->lru);
887 mt = get_pcppage_migratetype(page);
888 /* MIGRATE_ISOLATE page should not go to pcplists */
889 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
890 /* Pageblock could have been isolated meanwhile */
891 if (unlikely(has_isolate_pageblock(zone)))
892 mt = get_pageblock_migratetype(page);
894 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
895 trace_mm_page_pcpu_drain(page, 0, mt);
896 } while (--to_free && --batch_free && !list_empty(list));
898 spin_unlock(&zone->lock);
901 static void free_one_page(struct zone *zone,
902 struct page *page, unsigned long pfn,
906 unsigned long nr_scanned;
907 spin_lock(&zone->lock);
908 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
910 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
912 if (unlikely(has_isolate_pageblock(zone) ||
913 is_migrate_isolate(migratetype))) {
914 migratetype = get_pfnblock_migratetype(page, pfn);
916 __free_one_page(page, pfn, zone, order, migratetype);
917 spin_unlock(&zone->lock);
920 static int free_tail_pages_check(struct page *head_page, struct page *page)
925 * We rely page->lru.next never has bit 0 set, unless the page
926 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
928 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
930 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
934 if (unlikely(!PageTail(page))) {
935 bad_page(page, "PageTail not set", 0);
938 if (unlikely(compound_head(page) != head_page)) {
939 bad_page(page, "compound_head not consistent", 0);
944 clear_compound_head(page);
948 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
949 unsigned long zone, int nid)
951 set_page_links(page, zone, nid, pfn);
952 init_page_count(page);
953 page_mapcount_reset(page);
954 page_cpupid_reset_last(page);
956 INIT_LIST_HEAD(&page->lru);
957 #ifdef WANT_PAGE_VIRTUAL
958 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
959 if (!is_highmem_idx(zone))
960 set_page_address(page, __va(pfn << PAGE_SHIFT));
964 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
967 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
970 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
971 static void init_reserved_page(unsigned long pfn)
976 if (!early_page_uninitialised(pfn))
979 nid = early_pfn_to_nid(pfn);
980 pgdat = NODE_DATA(nid);
982 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
983 struct zone *zone = &pgdat->node_zones[zid];
985 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
988 __init_single_pfn(pfn, zid, nid);
991 static inline void init_reserved_page(unsigned long pfn)
994 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
997 * Initialised pages do not have PageReserved set. This function is
998 * called for each range allocated by the bootmem allocator and
999 * marks the pages PageReserved. The remaining valid pages are later
1000 * sent to the buddy page allocator.
1002 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1004 unsigned long start_pfn = PFN_DOWN(start);
1005 unsigned long end_pfn = PFN_UP(end);
1007 for (; start_pfn < end_pfn; start_pfn++) {
1008 if (pfn_valid(start_pfn)) {
1009 struct page *page = pfn_to_page(start_pfn);
1011 init_reserved_page(start_pfn);
1013 /* Avoid false-positive PageTail() */
1014 INIT_LIST_HEAD(&page->lru);
1016 SetPageReserved(page);
1021 static bool free_pages_prepare(struct page *page, unsigned int order)
1023 bool compound = PageCompound(page);
1026 VM_BUG_ON_PAGE(PageTail(page), page);
1027 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1029 trace_mm_page_free(page, order);
1030 kmemcheck_free_shadow(page, order);
1031 kasan_free_pages(page, order);
1034 page->mapping = NULL;
1035 bad += free_pages_check(page);
1036 for (i = 1; i < (1 << order); i++) {
1038 bad += free_tail_pages_check(page, page + i);
1039 bad += free_pages_check(page + i);
1044 reset_page_owner(page, order);
1046 if (!PageHighMem(page)) {
1047 debug_check_no_locks_freed(page_address(page),
1048 PAGE_SIZE << order);
1049 debug_check_no_obj_freed(page_address(page),
1050 PAGE_SIZE << order);
1052 arch_free_page(page, order);
1053 kernel_map_pages(page, 1 << order, 0);
1058 static void __free_pages_ok(struct page *page, unsigned int order)
1060 unsigned long flags;
1062 unsigned long pfn = page_to_pfn(page);
1064 if (!free_pages_prepare(page, order))
1067 migratetype = get_pfnblock_migratetype(page, pfn);
1068 local_irq_save(flags);
1069 __count_vm_events(PGFREE, 1 << order);
1070 free_one_page(page_zone(page), page, pfn, order, migratetype);
1071 local_irq_restore(flags);
1074 static void __init __free_pages_boot_core(struct page *page,
1075 unsigned long pfn, unsigned int order)
1077 unsigned int nr_pages = 1 << order;
1078 struct page *p = page;
1082 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1084 __ClearPageReserved(p);
1085 set_page_count(p, 0);
1087 __ClearPageReserved(p);
1088 set_page_count(p, 0);
1090 page_zone(page)->managed_pages += nr_pages;
1091 set_page_refcounted(page);
1092 __free_pages(page, order);
1095 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1096 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1098 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1100 int __meminit early_pfn_to_nid(unsigned long pfn)
1102 static DEFINE_SPINLOCK(early_pfn_lock);
1105 spin_lock(&early_pfn_lock);
1106 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1108 nid = first_online_node;
1109 spin_unlock(&early_pfn_lock);
1115 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1116 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1117 struct mminit_pfnnid_cache *state)
1121 nid = __early_pfn_to_nid(pfn, state);
1122 if (nid >= 0 && nid != node)
1127 /* Only safe to use early in boot when initialisation is single-threaded */
1128 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1130 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1135 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1139 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1140 struct mminit_pfnnid_cache *state)
1147 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1150 if (early_page_uninitialised(pfn))
1152 return __free_pages_boot_core(page, pfn, order);
1155 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1156 static void __init deferred_free_range(struct page *page,
1157 unsigned long pfn, int nr_pages)
1164 /* Free a large naturally-aligned chunk if possible */
1165 if (nr_pages == MAX_ORDER_NR_PAGES &&
1166 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1167 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1168 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1172 for (i = 0; i < nr_pages; i++, page++, pfn++)
1173 __free_pages_boot_core(page, pfn, 0);
1176 /* Completion tracking for deferred_init_memmap() threads */
1177 static atomic_t pgdat_init_n_undone __initdata;
1178 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1180 static inline void __init pgdat_init_report_one_done(void)
1182 if (atomic_dec_and_test(&pgdat_init_n_undone))
1183 complete(&pgdat_init_all_done_comp);
1186 /* Initialise remaining memory on a node */
1187 static int __init deferred_init_memmap(void *data)
1189 pg_data_t *pgdat = data;
1190 int nid = pgdat->node_id;
1191 struct mminit_pfnnid_cache nid_init_state = { };
1192 unsigned long start = jiffies;
1193 unsigned long nr_pages = 0;
1194 unsigned long walk_start, walk_end;
1197 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1198 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1200 if (first_init_pfn == ULONG_MAX) {
1201 pgdat_init_report_one_done();
1205 /* Bind memory initialisation thread to a local node if possible */
1206 if (!cpumask_empty(cpumask))
1207 set_cpus_allowed_ptr(current, cpumask);
1209 /* Sanity check boundaries */
1210 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1211 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1212 pgdat->first_deferred_pfn = ULONG_MAX;
1214 /* Only the highest zone is deferred so find it */
1215 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1216 zone = pgdat->node_zones + zid;
1217 if (first_init_pfn < zone_end_pfn(zone))
1221 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1222 unsigned long pfn, end_pfn;
1223 struct page *page = NULL;
1224 struct page *free_base_page = NULL;
1225 unsigned long free_base_pfn = 0;
1228 end_pfn = min(walk_end, zone_end_pfn(zone));
1229 pfn = first_init_pfn;
1230 if (pfn < walk_start)
1232 if (pfn < zone->zone_start_pfn)
1233 pfn = zone->zone_start_pfn;
1235 for (; pfn < end_pfn; pfn++) {
1236 if (!pfn_valid_within(pfn))
1240 * Ensure pfn_valid is checked every
1241 * MAX_ORDER_NR_PAGES for memory holes
1243 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1244 if (!pfn_valid(pfn)) {
1250 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1255 /* Minimise pfn page lookups and scheduler checks */
1256 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1259 nr_pages += nr_to_free;
1260 deferred_free_range(free_base_page,
1261 free_base_pfn, nr_to_free);
1262 free_base_page = NULL;
1263 free_base_pfn = nr_to_free = 0;
1265 page = pfn_to_page(pfn);
1270 VM_BUG_ON(page_zone(page) != zone);
1274 __init_single_page(page, pfn, zid, nid);
1275 if (!free_base_page) {
1276 free_base_page = page;
1277 free_base_pfn = pfn;
1282 /* Where possible, batch up pages for a single free */
1285 /* Free the current block of pages to allocator */
1286 nr_pages += nr_to_free;
1287 deferred_free_range(free_base_page, free_base_pfn,
1289 free_base_page = NULL;
1290 free_base_pfn = nr_to_free = 0;
1293 first_init_pfn = max(end_pfn, first_init_pfn);
1296 /* Sanity check that the next zone really is unpopulated */
1297 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1299 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1300 jiffies_to_msecs(jiffies - start));
1302 pgdat_init_report_one_done();
1306 void __init page_alloc_init_late(void)
1310 /* There will be num_node_state(N_MEMORY) threads */
1311 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1312 for_each_node_state(nid, N_MEMORY) {
1313 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1316 /* Block until all are initialised */
1317 wait_for_completion(&pgdat_init_all_done_comp);
1319 /* Reinit limits that are based on free pages after the kernel is up */
1320 files_maxfiles_init();
1322 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1325 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1326 void __init init_cma_reserved_pageblock(struct page *page)
1328 unsigned i = pageblock_nr_pages;
1329 struct page *p = page;
1332 __ClearPageReserved(p);
1333 set_page_count(p, 0);
1336 set_pageblock_migratetype(page, MIGRATE_CMA);
1338 if (pageblock_order >= MAX_ORDER) {
1339 i = pageblock_nr_pages;
1342 set_page_refcounted(p);
1343 __free_pages(p, MAX_ORDER - 1);
1344 p += MAX_ORDER_NR_PAGES;
1345 } while (i -= MAX_ORDER_NR_PAGES);
1347 set_page_refcounted(page);
1348 __free_pages(page, pageblock_order);
1351 adjust_managed_page_count(page, pageblock_nr_pages);
1356 * The order of subdivision here is critical for the IO subsystem.
1357 * Please do not alter this order without good reasons and regression
1358 * testing. Specifically, as large blocks of memory are subdivided,
1359 * the order in which smaller blocks are delivered depends on the order
1360 * they're subdivided in this function. This is the primary factor
1361 * influencing the order in which pages are delivered to the IO
1362 * subsystem according to empirical testing, and this is also justified
1363 * by considering the behavior of a buddy system containing a single
1364 * large block of memory acted on by a series of small allocations.
1365 * This behavior is a critical factor in sglist merging's success.
1369 static inline void expand(struct zone *zone, struct page *page,
1370 int low, int high, struct free_area *area,
1373 unsigned long size = 1 << high;
1375 while (high > low) {
1379 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1381 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1382 debug_guardpage_enabled() &&
1383 high < debug_guardpage_minorder()) {
1385 * Mark as guard pages (or page), that will allow to
1386 * merge back to allocator when buddy will be freed.
1387 * Corresponding page table entries will not be touched,
1388 * pages will stay not present in virtual address space
1390 set_page_guard(zone, &page[size], high, migratetype);
1393 list_add(&page[size].lru, &area->free_list[migratetype]);
1395 set_page_order(&page[size], high);
1400 * This page is about to be returned from the page allocator
1402 static inline int check_new_page(struct page *page)
1404 const char *bad_reason = NULL;
1405 unsigned long bad_flags = 0;
1407 if (unlikely(page_mapcount(page)))
1408 bad_reason = "nonzero mapcount";
1409 if (unlikely(page->mapping != NULL))
1410 bad_reason = "non-NULL mapping";
1411 if (unlikely(atomic_read(&page->_count) != 0))
1412 bad_reason = "nonzero _count";
1413 if (unlikely(page->flags & __PG_HWPOISON)) {
1414 bad_reason = "HWPoisoned (hardware-corrupted)";
1415 bad_flags = __PG_HWPOISON;
1417 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1418 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1419 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1422 if (unlikely(page->mem_cgroup))
1423 bad_reason = "page still charged to cgroup";
1425 if (unlikely(bad_reason)) {
1426 bad_page(page, bad_reason, bad_flags);
1432 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1437 for (i = 0; i < (1 << order); i++) {
1438 struct page *p = page + i;
1439 if (unlikely(check_new_page(p)))
1443 set_page_private(page, 0);
1444 set_page_refcounted(page);
1446 arch_alloc_page(page, order);
1447 kernel_map_pages(page, 1 << order, 1);
1448 kasan_alloc_pages(page, order);
1450 if (gfp_flags & __GFP_ZERO)
1451 for (i = 0; i < (1 << order); i++)
1452 clear_highpage(page + i);
1454 if (order && (gfp_flags & __GFP_COMP))
1455 prep_compound_page(page, order);
1457 set_page_owner(page, order, gfp_flags);
1460 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1461 * allocate the page. The expectation is that the caller is taking
1462 * steps that will free more memory. The caller should avoid the page
1463 * being used for !PFMEMALLOC purposes.
1465 if (alloc_flags & ALLOC_NO_WATERMARKS)
1466 set_page_pfmemalloc(page);
1468 clear_page_pfmemalloc(page);
1474 * Go through the free lists for the given migratetype and remove
1475 * the smallest available page from the freelists
1478 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1481 unsigned int current_order;
1482 struct free_area *area;
1485 /* Find a page of the appropriate size in the preferred list */
1486 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1487 area = &(zone->free_area[current_order]);
1488 if (list_empty(&area->free_list[migratetype]))
1491 page = list_entry(area->free_list[migratetype].next,
1493 list_del(&page->lru);
1494 rmv_page_order(page);
1496 expand(zone, page, order, current_order, area, migratetype);
1497 set_pcppage_migratetype(page, migratetype);
1506 * This array describes the order lists are fallen back to when
1507 * the free lists for the desirable migrate type are depleted
1509 static int fallbacks[MIGRATE_TYPES][4] = {
1510 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1511 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1512 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1514 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1516 #ifdef CONFIG_MEMORY_ISOLATION
1517 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1522 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1525 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1528 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1529 unsigned int order) { return NULL; }
1533 * Move the free pages in a range to the free lists of the requested type.
1534 * Note that start_page and end_pages are not aligned on a pageblock
1535 * boundary. If alignment is required, use move_freepages_block()
1537 int move_freepages(struct zone *zone,
1538 struct page *start_page, struct page *end_page,
1543 int pages_moved = 0;
1545 #ifndef CONFIG_HOLES_IN_ZONE
1547 * page_zone is not safe to call in this context when
1548 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1549 * anyway as we check zone boundaries in move_freepages_block().
1550 * Remove at a later date when no bug reports exist related to
1551 * grouping pages by mobility
1553 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1556 for (page = start_page; page <= end_page;) {
1557 if (!pfn_valid_within(page_to_pfn(page))) {
1562 /* Make sure we are not inadvertently changing nodes */
1563 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1565 if (!PageBuddy(page)) {
1570 order = page_order(page);
1571 list_move(&page->lru,
1572 &zone->free_area[order].free_list[migratetype]);
1574 pages_moved += 1 << order;
1580 int move_freepages_block(struct zone *zone, struct page *page,
1583 unsigned long start_pfn, end_pfn;
1584 struct page *start_page, *end_page;
1586 start_pfn = page_to_pfn(page);
1587 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1588 start_page = pfn_to_page(start_pfn);
1589 end_page = start_page + pageblock_nr_pages - 1;
1590 end_pfn = start_pfn + pageblock_nr_pages - 1;
1592 /* Do not cross zone boundaries */
1593 if (!zone_spans_pfn(zone, start_pfn))
1595 if (!zone_spans_pfn(zone, end_pfn))
1598 return move_freepages(zone, start_page, end_page, migratetype);
1601 static void change_pageblock_range(struct page *pageblock_page,
1602 int start_order, int migratetype)
1604 int nr_pageblocks = 1 << (start_order - pageblock_order);
1606 while (nr_pageblocks--) {
1607 set_pageblock_migratetype(pageblock_page, migratetype);
1608 pageblock_page += pageblock_nr_pages;
1613 * When we are falling back to another migratetype during allocation, try to
1614 * steal extra free pages from the same pageblocks to satisfy further
1615 * allocations, instead of polluting multiple pageblocks.
1617 * If we are stealing a relatively large buddy page, it is likely there will
1618 * be more free pages in the pageblock, so try to steal them all. For
1619 * reclaimable and unmovable allocations, we steal regardless of page size,
1620 * as fragmentation caused by those allocations polluting movable pageblocks
1621 * is worse than movable allocations stealing from unmovable and reclaimable
1624 static bool can_steal_fallback(unsigned int order, int start_mt)
1627 * Leaving this order check is intended, although there is
1628 * relaxed order check in next check. The reason is that
1629 * we can actually steal whole pageblock if this condition met,
1630 * but, below check doesn't guarantee it and that is just heuristic
1631 * so could be changed anytime.
1633 if (order >= pageblock_order)
1636 if (order >= pageblock_order / 2 ||
1637 start_mt == MIGRATE_RECLAIMABLE ||
1638 start_mt == MIGRATE_UNMOVABLE ||
1639 page_group_by_mobility_disabled)
1646 * This function implements actual steal behaviour. If order is large enough,
1647 * we can steal whole pageblock. If not, we first move freepages in this
1648 * pageblock and check whether half of pages are moved or not. If half of
1649 * pages are moved, we can change migratetype of pageblock and permanently
1650 * use it's pages as requested migratetype in the future.
1652 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1655 unsigned int current_order = page_order(page);
1658 /* Take ownership for orders >= pageblock_order */
1659 if (current_order >= pageblock_order) {
1660 change_pageblock_range(page, current_order, start_type);
1664 pages = move_freepages_block(zone, page, start_type);
1666 /* Claim the whole block if over half of it is free */
1667 if (pages >= (1 << (pageblock_order-1)) ||
1668 page_group_by_mobility_disabled)
1669 set_pageblock_migratetype(page, start_type);
1673 * Check whether there is a suitable fallback freepage with requested order.
1674 * If only_stealable is true, this function returns fallback_mt only if
1675 * we can steal other freepages all together. This would help to reduce
1676 * fragmentation due to mixed migratetype pages in one pageblock.
1678 int find_suitable_fallback(struct free_area *area, unsigned int order,
1679 int migratetype, bool only_stealable, bool *can_steal)
1684 if (area->nr_free == 0)
1689 fallback_mt = fallbacks[migratetype][i];
1690 if (fallback_mt == MIGRATE_TYPES)
1693 if (list_empty(&area->free_list[fallback_mt]))
1696 if (can_steal_fallback(order, migratetype))
1699 if (!only_stealable)
1710 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1711 * there are no empty page blocks that contain a page with a suitable order
1713 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1714 unsigned int alloc_order)
1717 unsigned long max_managed, flags;
1720 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1721 * Check is race-prone but harmless.
1723 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1724 if (zone->nr_reserved_highatomic >= max_managed)
1727 spin_lock_irqsave(&zone->lock, flags);
1729 /* Recheck the nr_reserved_highatomic limit under the lock */
1730 if (zone->nr_reserved_highatomic >= max_managed)
1734 mt = get_pageblock_migratetype(page);
1735 if (mt != MIGRATE_HIGHATOMIC &&
1736 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1737 zone->nr_reserved_highatomic += pageblock_nr_pages;
1738 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1739 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1743 spin_unlock_irqrestore(&zone->lock, flags);
1747 * Used when an allocation is about to fail under memory pressure. This
1748 * potentially hurts the reliability of high-order allocations when under
1749 * intense memory pressure but failed atomic allocations should be easier
1750 * to recover from than an OOM.
1752 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1754 struct zonelist *zonelist = ac->zonelist;
1755 unsigned long flags;
1761 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1763 /* Preserve at least one pageblock */
1764 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1767 spin_lock_irqsave(&zone->lock, flags);
1768 for (order = 0; order < MAX_ORDER; order++) {
1769 struct free_area *area = &(zone->free_area[order]);
1771 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1774 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1778 * In page freeing path, migratetype change is racy so
1779 * we can counter several free pages in a pageblock
1780 * in this loop althoug we changed the pageblock type
1781 * from highatomic to ac->migratetype. So we should
1782 * adjust the count once.
1784 if (get_pageblock_migratetype(page) ==
1785 MIGRATE_HIGHATOMIC) {
1787 * It should never happen but changes to
1788 * locking could inadvertently allow a per-cpu
1789 * drain to add pages to MIGRATE_HIGHATOMIC
1790 * while unreserving so be safe and watch for
1793 zone->nr_reserved_highatomic -= min(
1795 zone->nr_reserved_highatomic);
1799 * Convert to ac->migratetype and avoid the normal
1800 * pageblock stealing heuristics. Minimally, the caller
1801 * is doing the work and needs the pages. More
1802 * importantly, if the block was always converted to
1803 * MIGRATE_UNMOVABLE or another type then the number
1804 * of pageblocks that cannot be completely freed
1807 set_pageblock_migratetype(page, ac->migratetype);
1808 move_freepages_block(zone, page, ac->migratetype);
1809 spin_unlock_irqrestore(&zone->lock, flags);
1812 spin_unlock_irqrestore(&zone->lock, flags);
1816 /* Remove an element from the buddy allocator from the fallback list */
1817 static inline struct page *
1818 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1820 struct free_area *area;
1821 unsigned int current_order;
1826 /* Find the largest possible block of pages in the other list */
1827 for (current_order = MAX_ORDER-1;
1828 current_order >= order && current_order <= MAX_ORDER-1;
1830 area = &(zone->free_area[current_order]);
1831 fallback_mt = find_suitable_fallback(area, current_order,
1832 start_migratetype, false, &can_steal);
1833 if (fallback_mt == -1)
1836 page = list_entry(area->free_list[fallback_mt].next,
1839 steal_suitable_fallback(zone, page, start_migratetype);
1841 /* Remove the page from the freelists */
1843 list_del(&page->lru);
1844 rmv_page_order(page);
1846 expand(zone, page, order, current_order, area,
1849 * The pcppage_migratetype may differ from pageblock's
1850 * migratetype depending on the decisions in
1851 * find_suitable_fallback(). This is OK as long as it does not
1852 * differ for MIGRATE_CMA pageblocks. Those can be used as
1853 * fallback only via special __rmqueue_cma_fallback() function
1855 set_pcppage_migratetype(page, start_migratetype);
1857 trace_mm_page_alloc_extfrag(page, order, current_order,
1858 start_migratetype, fallback_mt);
1867 * Do the hard work of removing an element from the buddy allocator.
1868 * Call me with the zone->lock already held.
1870 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1871 int migratetype, gfp_t gfp_flags)
1875 page = __rmqueue_smallest(zone, order, migratetype);
1876 if (unlikely(!page)) {
1877 if (migratetype == MIGRATE_MOVABLE)
1878 page = __rmqueue_cma_fallback(zone, order);
1881 page = __rmqueue_fallback(zone, order, migratetype);
1884 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1889 * Obtain a specified number of elements from the buddy allocator, all under
1890 * a single hold of the lock, for efficiency. Add them to the supplied list.
1891 * Returns the number of new pages which were placed at *list.
1893 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1894 unsigned long count, struct list_head *list,
1895 int migratetype, bool cold)
1899 spin_lock(&zone->lock);
1900 for (i = 0; i < count; ++i) {
1901 struct page *page = __rmqueue(zone, order, migratetype, 0);
1902 if (unlikely(page == NULL))
1906 * Split buddy pages returned by expand() are received here
1907 * in physical page order. The page is added to the callers and
1908 * list and the list head then moves forward. From the callers
1909 * perspective, the linked list is ordered by page number in
1910 * some conditions. This is useful for IO devices that can
1911 * merge IO requests if the physical pages are ordered
1915 list_add(&page->lru, list);
1917 list_add_tail(&page->lru, list);
1919 if (is_migrate_cma(get_pcppage_migratetype(page)))
1920 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1923 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1924 spin_unlock(&zone->lock);
1930 * Called from the vmstat counter updater to drain pagesets of this
1931 * currently executing processor on remote nodes after they have
1934 * Note that this function must be called with the thread pinned to
1935 * a single processor.
1937 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1939 unsigned long flags;
1940 int to_drain, batch;
1942 local_irq_save(flags);
1943 batch = READ_ONCE(pcp->batch);
1944 to_drain = min(pcp->count, batch);
1946 free_pcppages_bulk(zone, to_drain, pcp);
1947 pcp->count -= to_drain;
1949 local_irq_restore(flags);
1954 * Drain pcplists of the indicated processor and zone.
1956 * The processor must either be the current processor and the
1957 * thread pinned to the current processor or a processor that
1960 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1962 unsigned long flags;
1963 struct per_cpu_pageset *pset;
1964 struct per_cpu_pages *pcp;
1966 local_irq_save(flags);
1967 pset = per_cpu_ptr(zone->pageset, cpu);
1971 free_pcppages_bulk(zone, pcp->count, pcp);
1974 local_irq_restore(flags);
1978 * Drain pcplists of all zones on the indicated processor.
1980 * The processor must either be the current processor and the
1981 * thread pinned to the current processor or a processor that
1984 static void drain_pages(unsigned int cpu)
1988 for_each_populated_zone(zone) {
1989 drain_pages_zone(cpu, zone);
1994 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1996 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1997 * the single zone's pages.
1999 void drain_local_pages(struct zone *zone)
2001 int cpu = smp_processor_id();
2004 drain_pages_zone(cpu, zone);
2010 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2012 * When zone parameter is non-NULL, spill just the single zone's pages.
2014 * Note that this code is protected against sending an IPI to an offline
2015 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2016 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2017 * nothing keeps CPUs from showing up after we populated the cpumask and
2018 * before the call to on_each_cpu_mask().
2020 void drain_all_pages(struct zone *zone)
2025 * Allocate in the BSS so we wont require allocation in
2026 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2028 static cpumask_t cpus_with_pcps;
2031 * We don't care about racing with CPU hotplug event
2032 * as offline notification will cause the notified
2033 * cpu to drain that CPU pcps and on_each_cpu_mask
2034 * disables preemption as part of its processing
2036 for_each_online_cpu(cpu) {
2037 struct per_cpu_pageset *pcp;
2039 bool has_pcps = false;
2042 pcp = per_cpu_ptr(zone->pageset, cpu);
2046 for_each_populated_zone(z) {
2047 pcp = per_cpu_ptr(z->pageset, cpu);
2048 if (pcp->pcp.count) {
2056 cpumask_set_cpu(cpu, &cpus_with_pcps);
2058 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2060 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2064 #ifdef CONFIG_HIBERNATION
2066 void mark_free_pages(struct zone *zone)
2068 unsigned long pfn, max_zone_pfn;
2069 unsigned long flags;
2070 unsigned int order, t;
2071 struct list_head *curr;
2073 if (zone_is_empty(zone))
2076 spin_lock_irqsave(&zone->lock, flags);
2078 max_zone_pfn = zone_end_pfn(zone);
2079 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2080 if (pfn_valid(pfn)) {
2081 struct page *page = pfn_to_page(pfn);
2083 if (!swsusp_page_is_forbidden(page))
2084 swsusp_unset_page_free(page);
2087 for_each_migratetype_order(order, t) {
2088 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2091 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2092 for (i = 0; i < (1UL << order); i++)
2093 swsusp_set_page_free(pfn_to_page(pfn + i));
2096 spin_unlock_irqrestore(&zone->lock, flags);
2098 #endif /* CONFIG_PM */
2101 * Free a 0-order page
2102 * cold == true ? free a cold page : free a hot page
2104 void free_hot_cold_page(struct page *page, bool cold)
2106 struct zone *zone = page_zone(page);
2107 struct per_cpu_pages *pcp;
2108 unsigned long flags;
2109 unsigned long pfn = page_to_pfn(page);
2112 if (!free_pages_prepare(page, 0))
2115 migratetype = get_pfnblock_migratetype(page, pfn);
2116 set_pcppage_migratetype(page, migratetype);
2117 local_irq_save(flags);
2118 __count_vm_event(PGFREE);
2121 * We only track unmovable, reclaimable and movable on pcp lists.
2122 * Free ISOLATE pages back to the allocator because they are being
2123 * offlined but treat RESERVE as movable pages so we can get those
2124 * areas back if necessary. Otherwise, we may have to free
2125 * excessively into the page allocator
2127 if (migratetype >= MIGRATE_PCPTYPES) {
2128 if (unlikely(is_migrate_isolate(migratetype))) {
2129 free_one_page(zone, page, pfn, 0, migratetype);
2132 migratetype = MIGRATE_MOVABLE;
2135 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2137 list_add(&page->lru, &pcp->lists[migratetype]);
2139 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2141 if (pcp->count >= pcp->high) {
2142 unsigned long batch = READ_ONCE(pcp->batch);
2143 free_pcppages_bulk(zone, batch, pcp);
2144 pcp->count -= batch;
2148 local_irq_restore(flags);
2152 * Free a list of 0-order pages
2154 void free_hot_cold_page_list(struct list_head *list, bool cold)
2156 struct page *page, *next;
2158 list_for_each_entry_safe(page, next, list, lru) {
2159 trace_mm_page_free_batched(page, cold);
2160 free_hot_cold_page(page, cold);
2165 * split_page takes a non-compound higher-order page, and splits it into
2166 * n (1<<order) sub-pages: page[0..n]
2167 * Each sub-page must be freed individually.
2169 * Note: this is probably too low level an operation for use in drivers.
2170 * Please consult with lkml before using this in your driver.
2172 void split_page(struct page *page, unsigned int order)
2177 VM_BUG_ON_PAGE(PageCompound(page), page);
2178 VM_BUG_ON_PAGE(!page_count(page), page);
2180 #ifdef CONFIG_KMEMCHECK
2182 * Split shadow pages too, because free(page[0]) would
2183 * otherwise free the whole shadow.
2185 if (kmemcheck_page_is_tracked(page))
2186 split_page(virt_to_page(page[0].shadow), order);
2189 gfp_mask = get_page_owner_gfp(page);
2190 set_page_owner(page, 0, gfp_mask);
2191 for (i = 1; i < (1 << order); i++) {
2192 set_page_refcounted(page + i);
2193 set_page_owner(page + i, 0, gfp_mask);
2196 EXPORT_SYMBOL_GPL(split_page);
2198 int __isolate_free_page(struct page *page, unsigned int order)
2200 unsigned long watermark;
2204 BUG_ON(!PageBuddy(page));
2206 zone = page_zone(page);
2207 mt = get_pageblock_migratetype(page);
2209 if (!is_migrate_isolate(mt)) {
2210 /* Obey watermarks as if the page was being allocated */
2211 watermark = low_wmark_pages(zone) + (1 << order);
2212 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2215 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2218 /* Remove page from free list */
2219 list_del(&page->lru);
2220 zone->free_area[order].nr_free--;
2221 rmv_page_order(page);
2223 set_page_owner(page, order, __GFP_MOVABLE);
2225 /* Set the pageblock if the isolated page is at least a pageblock */
2226 if (order >= pageblock_order - 1) {
2227 struct page *endpage = page + (1 << order) - 1;
2228 for (; page < endpage; page += pageblock_nr_pages) {
2229 int mt = get_pageblock_migratetype(page);
2230 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2231 set_pageblock_migratetype(page,
2237 return 1UL << order;
2241 * Similar to split_page except the page is already free. As this is only
2242 * being used for migration, the migratetype of the block also changes.
2243 * As this is called with interrupts disabled, the caller is responsible
2244 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2247 * Note: this is probably too low level an operation for use in drivers.
2248 * Please consult with lkml before using this in your driver.
2250 int split_free_page(struct page *page)
2255 order = page_order(page);
2257 nr_pages = __isolate_free_page(page, order);
2261 /* Split into individual pages */
2262 set_page_refcounted(page);
2263 split_page(page, order);
2268 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2271 struct page *buffered_rmqueue(struct zone *preferred_zone,
2272 struct zone *zone, unsigned int order,
2273 gfp_t gfp_flags, int alloc_flags, int migratetype)
2275 unsigned long flags;
2277 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2279 if (likely(order == 0)) {
2280 struct per_cpu_pages *pcp;
2281 struct list_head *list;
2283 local_irq_save(flags);
2284 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2285 list = &pcp->lists[migratetype];
2286 if (list_empty(list)) {
2287 pcp->count += rmqueue_bulk(zone, 0,
2290 if (unlikely(list_empty(list)))
2295 page = list_entry(list->prev, struct page, lru);
2297 page = list_entry(list->next, struct page, lru);
2299 list_del(&page->lru);
2302 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2304 * __GFP_NOFAIL is not to be used in new code.
2306 * All __GFP_NOFAIL callers should be fixed so that they
2307 * properly detect and handle allocation failures.
2309 * We most definitely don't want callers attempting to
2310 * allocate greater than order-1 page units with
2313 WARN_ON_ONCE(order > 1);
2315 spin_lock_irqsave(&zone->lock, flags);
2318 if (alloc_flags & ALLOC_HARDER) {
2319 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2321 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2324 page = __rmqueue(zone, order, migratetype, gfp_flags);
2325 spin_unlock(&zone->lock);
2328 __mod_zone_freepage_state(zone, -(1 << order),
2329 get_pcppage_migratetype(page));
2332 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2333 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2334 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2335 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2337 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2338 zone_statistics(preferred_zone, zone, gfp_flags);
2339 local_irq_restore(flags);
2341 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2345 local_irq_restore(flags);
2349 #ifdef CONFIG_FAIL_PAGE_ALLOC
2352 struct fault_attr attr;
2354 bool ignore_gfp_highmem;
2355 bool ignore_gfp_reclaim;
2357 } fail_page_alloc = {
2358 .attr = FAULT_ATTR_INITIALIZER,
2359 .ignore_gfp_reclaim = true,
2360 .ignore_gfp_highmem = true,
2364 static int __init setup_fail_page_alloc(char *str)
2366 return setup_fault_attr(&fail_page_alloc.attr, str);
2368 __setup("fail_page_alloc=", setup_fail_page_alloc);
2370 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2372 if (order < fail_page_alloc.min_order)
2374 if (gfp_mask & __GFP_NOFAIL)
2376 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2378 if (fail_page_alloc.ignore_gfp_reclaim &&
2379 (gfp_mask & __GFP_DIRECT_RECLAIM))
2382 return should_fail(&fail_page_alloc.attr, 1 << order);
2385 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2387 static int __init fail_page_alloc_debugfs(void)
2389 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2392 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2393 &fail_page_alloc.attr);
2395 return PTR_ERR(dir);
2397 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2398 &fail_page_alloc.ignore_gfp_reclaim))
2400 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2401 &fail_page_alloc.ignore_gfp_highmem))
2403 if (!debugfs_create_u32("min-order", mode, dir,
2404 &fail_page_alloc.min_order))
2409 debugfs_remove_recursive(dir);
2414 late_initcall(fail_page_alloc_debugfs);
2416 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2418 #else /* CONFIG_FAIL_PAGE_ALLOC */
2420 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2425 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2428 * Return true if free base pages are above 'mark'. For high-order checks it
2429 * will return true of the order-0 watermark is reached and there is at least
2430 * one free page of a suitable size. Checking now avoids taking the zone lock
2431 * to check in the allocation paths if no pages are free.
2433 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2434 unsigned long mark, int classzone_idx, int alloc_flags,
2439 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2441 /* free_pages may go negative - that's OK */
2442 free_pages -= (1 << order) - 1;
2444 if (alloc_flags & ALLOC_HIGH)
2448 * If the caller does not have rights to ALLOC_HARDER then subtract
2449 * the high-atomic reserves. This will over-estimate the size of the
2450 * atomic reserve but it avoids a search.
2452 if (likely(!alloc_harder))
2453 free_pages -= z->nr_reserved_highatomic;
2458 /* If allocation can't use CMA areas don't use free CMA pages */
2459 if (!(alloc_flags & ALLOC_CMA))
2460 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2464 * Check watermarks for an order-0 allocation request. If these
2465 * are not met, then a high-order request also cannot go ahead
2466 * even if a suitable page happened to be free.
2468 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2471 /* If this is an order-0 request then the watermark is fine */
2475 /* For a high-order request, check at least one suitable page is free */
2476 for (o = order; o < MAX_ORDER; o++) {
2477 struct free_area *area = &z->free_area[o];
2486 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2487 if (!list_empty(&area->free_list[mt]))
2492 if ((alloc_flags & ALLOC_CMA) &&
2493 !list_empty(&area->free_list[MIGRATE_CMA])) {
2501 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2502 int classzone_idx, int alloc_flags)
2504 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2505 zone_page_state(z, NR_FREE_PAGES));
2508 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2509 unsigned long mark, int classzone_idx)
2511 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2513 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2514 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2516 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2521 static bool zone_local(struct zone *local_zone, struct zone *zone)
2523 return local_zone->node == zone->node;
2526 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2528 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2531 #else /* CONFIG_NUMA */
2532 static bool zone_local(struct zone *local_zone, struct zone *zone)
2537 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2541 #endif /* CONFIG_NUMA */
2543 static void reset_alloc_batches(struct zone *preferred_zone)
2545 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2548 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2549 high_wmark_pages(zone) - low_wmark_pages(zone) -
2550 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2551 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2552 } while (zone++ != preferred_zone);
2556 * get_page_from_freelist goes through the zonelist trying to allocate
2559 static struct page *
2560 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2561 const struct alloc_context *ac)
2563 struct zonelist *zonelist = ac->zonelist;
2565 struct page *page = NULL;
2567 int nr_fair_skipped = 0;
2568 bool zonelist_rescan;
2571 zonelist_rescan = false;
2574 * Scan zonelist, looking for a zone with enough free.
2575 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2577 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2581 if (cpusets_enabled() &&
2582 (alloc_flags & ALLOC_CPUSET) &&
2583 !cpuset_zone_allowed(zone, gfp_mask))
2586 * Distribute pages in proportion to the individual
2587 * zone size to ensure fair page aging. The zone a
2588 * page was allocated in should have no effect on the
2589 * time the page has in memory before being reclaimed.
2591 if (alloc_flags & ALLOC_FAIR) {
2592 if (!zone_local(ac->preferred_zone, zone))
2594 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2600 * When allocating a page cache page for writing, we
2601 * want to get it from a zone that is within its dirty
2602 * limit, such that no single zone holds more than its
2603 * proportional share of globally allowed dirty pages.
2604 * The dirty limits take into account the zone's
2605 * lowmem reserves and high watermark so that kswapd
2606 * should be able to balance it without having to
2607 * write pages from its LRU list.
2609 * This may look like it could increase pressure on
2610 * lower zones by failing allocations in higher zones
2611 * before they are full. But the pages that do spill
2612 * over are limited as the lower zones are protected
2613 * by this very same mechanism. It should not become
2614 * a practical burden to them.
2616 * XXX: For now, allow allocations to potentially
2617 * exceed the per-zone dirty limit in the slowpath
2618 * (spread_dirty_pages unset) before going into reclaim,
2619 * which is important when on a NUMA setup the allowed
2620 * zones are together not big enough to reach the
2621 * global limit. The proper fix for these situations
2622 * will require awareness of zones in the
2623 * dirty-throttling and the flusher threads.
2625 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2628 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2629 if (!zone_watermark_ok(zone, order, mark,
2630 ac->classzone_idx, alloc_flags)) {
2633 /* Checked here to keep the fast path fast */
2634 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2635 if (alloc_flags & ALLOC_NO_WATERMARKS)
2638 if (zone_reclaim_mode == 0 ||
2639 !zone_allows_reclaim(ac->preferred_zone, zone))
2642 ret = zone_reclaim(zone, gfp_mask, order);
2644 case ZONE_RECLAIM_NOSCAN:
2647 case ZONE_RECLAIM_FULL:
2648 /* scanned but unreclaimable */
2651 /* did we reclaim enough */
2652 if (zone_watermark_ok(zone, order, mark,
2653 ac->classzone_idx, alloc_flags))
2661 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2662 gfp_mask, alloc_flags, ac->migratetype);
2664 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2668 * If this is a high-order atomic allocation then check
2669 * if the pageblock should be reserved for the future
2671 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2672 reserve_highatomic_pageblock(page, zone, order);
2679 * The first pass makes sure allocations are spread fairly within the
2680 * local node. However, the local node might have free pages left
2681 * after the fairness batches are exhausted, and remote zones haven't
2682 * even been considered yet. Try once more without fairness, and
2683 * include remote zones now, before entering the slowpath and waking
2684 * kswapd: prefer spilling to a remote zone over swapping locally.
2686 if (alloc_flags & ALLOC_FAIR) {
2687 alloc_flags &= ~ALLOC_FAIR;
2688 if (nr_fair_skipped) {
2689 zonelist_rescan = true;
2690 reset_alloc_batches(ac->preferred_zone);
2692 if (nr_online_nodes > 1)
2693 zonelist_rescan = true;
2696 if (zonelist_rescan)
2703 * Large machines with many possible nodes should not always dump per-node
2704 * meminfo in irq context.
2706 static inline bool should_suppress_show_mem(void)
2711 ret = in_interrupt();
2716 static DEFINE_RATELIMIT_STATE(nopage_rs,
2717 DEFAULT_RATELIMIT_INTERVAL,
2718 DEFAULT_RATELIMIT_BURST);
2720 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2722 unsigned int filter = SHOW_MEM_FILTER_NODES;
2724 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2725 debug_guardpage_minorder() > 0)
2729 * This documents exceptions given to allocations in certain
2730 * contexts that are allowed to allocate outside current's set
2733 if (!(gfp_mask & __GFP_NOMEMALLOC))
2734 if (test_thread_flag(TIF_MEMDIE) ||
2735 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2736 filter &= ~SHOW_MEM_FILTER_NODES;
2737 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2738 filter &= ~SHOW_MEM_FILTER_NODES;
2741 struct va_format vaf;
2744 va_start(args, fmt);
2749 pr_warn("%pV", &vaf);
2754 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2755 current->comm, order, gfp_mask);
2758 if (!should_suppress_show_mem())
2762 static inline struct page *
2763 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2764 const struct alloc_context *ac, unsigned long *did_some_progress)
2766 struct oom_control oc = {
2767 .zonelist = ac->zonelist,
2768 .nodemask = ac->nodemask,
2769 .gfp_mask = gfp_mask,
2774 *did_some_progress = 0;
2777 * Acquire the oom lock. If that fails, somebody else is
2778 * making progress for us.
2780 if (!mutex_trylock(&oom_lock)) {
2781 *did_some_progress = 1;
2782 schedule_timeout_uninterruptible(1);
2787 * Go through the zonelist yet one more time, keep very high watermark
2788 * here, this is only to catch a parallel oom killing, we must fail if
2789 * we're still under heavy pressure.
2791 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2792 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2796 if (!(gfp_mask & __GFP_NOFAIL)) {
2797 /* Coredumps can quickly deplete all memory reserves */
2798 if (current->flags & PF_DUMPCORE)
2800 /* The OOM killer will not help higher order allocs */
2801 if (order > PAGE_ALLOC_COSTLY_ORDER)
2803 /* The OOM killer does not needlessly kill tasks for lowmem */
2804 if (ac->high_zoneidx < ZONE_NORMAL)
2806 /* The OOM killer does not compensate for IO-less reclaim */
2807 if (!(gfp_mask & __GFP_FS)) {
2809 * XXX: Page reclaim didn't yield anything,
2810 * and the OOM killer can't be invoked, but
2811 * keep looping as per tradition.
2813 *did_some_progress = 1;
2816 if (pm_suspended_storage())
2818 /* The OOM killer may not free memory on a specific node */
2819 if (gfp_mask & __GFP_THISNODE)
2822 /* Exhausted what can be done so it's blamo time */
2823 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2824 *did_some_progress = 1;
2826 mutex_unlock(&oom_lock);
2830 #ifdef CONFIG_COMPACTION
2831 /* Try memory compaction for high-order allocations before reclaim */
2832 static struct page *
2833 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2834 int alloc_flags, const struct alloc_context *ac,
2835 enum migrate_mode mode, int *contended_compaction,
2836 bool *deferred_compaction)
2838 unsigned long compact_result;
2844 current->flags |= PF_MEMALLOC;
2845 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2846 mode, contended_compaction);
2847 current->flags &= ~PF_MEMALLOC;
2849 switch (compact_result) {
2850 case COMPACT_DEFERRED:
2851 *deferred_compaction = true;
2853 case COMPACT_SKIPPED:
2860 * At least in one zone compaction wasn't deferred or skipped, so let's
2861 * count a compaction stall
2863 count_vm_event(COMPACTSTALL);
2865 page = get_page_from_freelist(gfp_mask, order,
2866 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2869 struct zone *zone = page_zone(page);
2871 zone->compact_blockskip_flush = false;
2872 compaction_defer_reset(zone, order, true);
2873 count_vm_event(COMPACTSUCCESS);
2878 * It's bad if compaction run occurs and fails. The most likely reason
2879 * is that pages exist, but not enough to satisfy watermarks.
2881 count_vm_event(COMPACTFAIL);
2888 static inline struct page *
2889 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2890 int alloc_flags, const struct alloc_context *ac,
2891 enum migrate_mode mode, int *contended_compaction,
2892 bool *deferred_compaction)
2896 #endif /* CONFIG_COMPACTION */
2898 /* Perform direct synchronous page reclaim */
2900 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2901 const struct alloc_context *ac)
2903 struct reclaim_state reclaim_state;
2908 /* We now go into synchronous reclaim */
2909 cpuset_memory_pressure_bump();
2910 current->flags |= PF_MEMALLOC;
2911 lockdep_set_current_reclaim_state(gfp_mask);
2912 reclaim_state.reclaimed_slab = 0;
2913 current->reclaim_state = &reclaim_state;
2915 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2918 current->reclaim_state = NULL;
2919 lockdep_clear_current_reclaim_state();
2920 current->flags &= ~PF_MEMALLOC;
2927 /* The really slow allocator path where we enter direct reclaim */
2928 static inline struct page *
2929 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2930 int alloc_flags, const struct alloc_context *ac,
2931 unsigned long *did_some_progress)
2933 struct page *page = NULL;
2934 bool drained = false;
2936 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2937 if (unlikely(!(*did_some_progress)))
2941 page = get_page_from_freelist(gfp_mask, order,
2942 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2945 * If an allocation failed after direct reclaim, it could be because
2946 * pages are pinned on the per-cpu lists or in high alloc reserves.
2947 * Shrink them them and try again
2949 if (!page && !drained) {
2950 unreserve_highatomic_pageblock(ac);
2951 drain_all_pages(NULL);
2960 * This is called in the allocator slow-path if the allocation request is of
2961 * sufficient urgency to ignore watermarks and take other desperate measures
2963 static inline struct page *
2964 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2965 const struct alloc_context *ac)
2970 page = get_page_from_freelist(gfp_mask, order,
2971 ALLOC_NO_WATERMARKS, ac);
2973 if (!page && gfp_mask & __GFP_NOFAIL)
2974 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2976 } while (!page && (gfp_mask & __GFP_NOFAIL));
2981 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2986 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2987 ac->high_zoneidx, ac->nodemask)
2988 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2992 gfp_to_alloc_flags(gfp_t gfp_mask)
2994 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2996 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2997 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3000 * The caller may dip into page reserves a bit more if the caller
3001 * cannot run direct reclaim, or if the caller has realtime scheduling
3002 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3003 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3005 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3007 if (gfp_mask & __GFP_ATOMIC) {
3009 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3010 * if it can't schedule.
3012 if (!(gfp_mask & __GFP_NOMEMALLOC))
3013 alloc_flags |= ALLOC_HARDER;
3015 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3016 * comment for __cpuset_node_allowed().
3018 alloc_flags &= ~ALLOC_CPUSET;
3019 } else if (unlikely(rt_task(current)) && !in_interrupt())
3020 alloc_flags |= ALLOC_HARDER;
3022 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3023 if (gfp_mask & __GFP_MEMALLOC)
3024 alloc_flags |= ALLOC_NO_WATERMARKS;
3025 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3026 alloc_flags |= ALLOC_NO_WATERMARKS;
3027 else if (!in_interrupt() &&
3028 ((current->flags & PF_MEMALLOC) ||
3029 unlikely(test_thread_flag(TIF_MEMDIE))))
3030 alloc_flags |= ALLOC_NO_WATERMARKS;
3033 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3034 alloc_flags |= ALLOC_CMA;
3039 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3041 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3044 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3046 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3049 static inline struct page *
3050 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3051 struct alloc_context *ac)
3053 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3054 struct page *page = NULL;
3056 unsigned long pages_reclaimed = 0;
3057 unsigned long did_some_progress;
3058 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3059 bool deferred_compaction = false;
3060 int contended_compaction = COMPACT_CONTENDED_NONE;
3063 * In the slowpath, we sanity check order to avoid ever trying to
3064 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3065 * be using allocators in order of preference for an area that is
3068 if (order >= MAX_ORDER) {
3069 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3074 * We also sanity check to catch abuse of atomic reserves being used by
3075 * callers that are not in atomic context.
3077 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3078 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3079 gfp_mask &= ~__GFP_ATOMIC;
3082 * If this allocation cannot block and it is for a specific node, then
3083 * fail early. There's no need to wakeup kswapd or retry for a
3084 * speculative node-specific allocation.
3086 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3090 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3091 wake_all_kswapds(order, ac);
3094 * OK, we're below the kswapd watermark and have kicked background
3095 * reclaim. Now things get more complex, so set up alloc_flags according
3096 * to how we want to proceed.
3098 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3101 * Find the true preferred zone if the allocation is unconstrained by
3104 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3105 struct zoneref *preferred_zoneref;
3106 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3107 ac->high_zoneidx, NULL, &ac->preferred_zone);
3108 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3111 /* This is the last chance, in general, before the goto nopage. */
3112 page = get_page_from_freelist(gfp_mask, order,
3113 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3117 /* Allocate without watermarks if the context allows */
3118 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3120 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3121 * the allocation is high priority and these type of
3122 * allocations are system rather than user orientated
3124 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3126 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3133 /* Caller is not willing to reclaim, we can't balance anything */
3134 if (!can_direct_reclaim) {
3136 * All existing users of the deprecated __GFP_NOFAIL are
3137 * blockable, so warn of any new users that actually allow this
3138 * type of allocation to fail.
3140 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3144 /* Avoid recursion of direct reclaim */
3145 if (current->flags & PF_MEMALLOC)
3148 /* Avoid allocations with no watermarks from looping endlessly */
3149 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3153 * Try direct compaction. The first pass is asynchronous. Subsequent
3154 * attempts after direct reclaim are synchronous
3156 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3158 &contended_compaction,
3159 &deferred_compaction);
3163 /* Checks for THP-specific high-order allocations */
3164 if (is_thp_gfp_mask(gfp_mask)) {
3166 * If compaction is deferred for high-order allocations, it is
3167 * because sync compaction recently failed. If this is the case
3168 * and the caller requested a THP allocation, we do not want
3169 * to heavily disrupt the system, so we fail the allocation
3170 * instead of entering direct reclaim.
3172 if (deferred_compaction)
3176 * In all zones where compaction was attempted (and not
3177 * deferred or skipped), lock contention has been detected.
3178 * For THP allocation we do not want to disrupt the others
3179 * so we fallback to base pages instead.
3181 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3185 * If compaction was aborted due to need_resched(), we do not
3186 * want to further increase allocation latency, unless it is
3187 * khugepaged trying to collapse.
3189 if (contended_compaction == COMPACT_CONTENDED_SCHED
3190 && !(current->flags & PF_KTHREAD))
3195 * It can become very expensive to allocate transparent hugepages at
3196 * fault, so use asynchronous memory compaction for THP unless it is
3197 * khugepaged trying to collapse.
3199 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3200 migration_mode = MIGRATE_SYNC_LIGHT;
3202 /* Try direct reclaim and then allocating */
3203 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3204 &did_some_progress);
3208 /* Do not loop if specifically requested */
3209 if (gfp_mask & __GFP_NORETRY)
3212 /* Keep reclaiming pages as long as there is reasonable progress */
3213 pages_reclaimed += did_some_progress;
3214 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3215 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3216 /* Wait for some write requests to complete then retry */
3217 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3221 /* Reclaim has failed us, start killing things */
3222 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3226 /* Retry as long as the OOM killer is making progress */
3227 if (did_some_progress)
3232 * High-order allocations do not necessarily loop after
3233 * direct reclaim and reclaim/compaction depends on compaction
3234 * being called after reclaim so call directly if necessary
3236 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3238 &contended_compaction,
3239 &deferred_compaction);
3243 warn_alloc_failed(gfp_mask, order, NULL);
3249 * This is the 'heart' of the zoned buddy allocator.
3252 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3253 struct zonelist *zonelist, nodemask_t *nodemask)
3255 struct zoneref *preferred_zoneref;
3256 struct page *page = NULL;
3257 unsigned int cpuset_mems_cookie;
3258 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3259 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3260 struct alloc_context ac = {
3261 .high_zoneidx = gfp_zone(gfp_mask),
3262 .nodemask = nodemask,
3263 .migratetype = gfpflags_to_migratetype(gfp_mask),
3266 gfp_mask &= gfp_allowed_mask;
3268 lockdep_trace_alloc(gfp_mask);
3270 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3272 if (should_fail_alloc_page(gfp_mask, order))
3276 * Check the zones suitable for the gfp_mask contain at least one
3277 * valid zone. It's possible to have an empty zonelist as a result
3278 * of __GFP_THISNODE and a memoryless node
3280 if (unlikely(!zonelist->_zonerefs->zone))
3283 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3284 alloc_flags |= ALLOC_CMA;
3287 cpuset_mems_cookie = read_mems_allowed_begin();
3289 /* We set it here, as __alloc_pages_slowpath might have changed it */
3290 ac.zonelist = zonelist;
3292 /* Dirty zone balancing only done in the fast path */
3293 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3295 /* The preferred zone is used for statistics later */
3296 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3297 ac.nodemask ? : &cpuset_current_mems_allowed,
3298 &ac.preferred_zone);
3299 if (!ac.preferred_zone)
3301 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3303 /* First allocation attempt */
3304 alloc_mask = gfp_mask|__GFP_HARDWALL;
3305 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3306 if (unlikely(!page)) {
3308 * Runtime PM, block IO and its error handling path
3309 * can deadlock because I/O on the device might not
3312 alloc_mask = memalloc_noio_flags(gfp_mask);
3313 ac.spread_dirty_pages = false;
3315 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3318 if (kmemcheck_enabled && page)
3319 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3321 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3325 * When updating a task's mems_allowed, it is possible to race with
3326 * parallel threads in such a way that an allocation can fail while
3327 * the mask is being updated. If a page allocation is about to fail,
3328 * check if the cpuset changed during allocation and if so, retry.
3330 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3335 EXPORT_SYMBOL(__alloc_pages_nodemask);
3338 * Common helper functions.
3340 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3345 * __get_free_pages() returns a 32-bit address, which cannot represent
3348 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3350 page = alloc_pages(gfp_mask, order);
3353 return (unsigned long) page_address(page);
3355 EXPORT_SYMBOL(__get_free_pages);
3357 unsigned long get_zeroed_page(gfp_t gfp_mask)
3359 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3361 EXPORT_SYMBOL(get_zeroed_page);
3363 void __free_pages(struct page *page, unsigned int order)
3365 if (put_page_testzero(page)) {
3367 free_hot_cold_page(page, false);
3369 __free_pages_ok(page, order);
3373 EXPORT_SYMBOL(__free_pages);
3375 void free_pages(unsigned long addr, unsigned int order)
3378 VM_BUG_ON(!virt_addr_valid((void *)addr));
3379 __free_pages(virt_to_page((void *)addr), order);
3383 EXPORT_SYMBOL(free_pages);
3387 * An arbitrary-length arbitrary-offset area of memory which resides
3388 * within a 0 or higher order page. Multiple fragments within that page
3389 * are individually refcounted, in the page's reference counter.
3391 * The page_frag functions below provide a simple allocation framework for
3392 * page fragments. This is used by the network stack and network device
3393 * drivers to provide a backing region of memory for use as either an
3394 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3396 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3399 struct page *page = NULL;
3400 gfp_t gfp = gfp_mask;
3402 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3403 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3405 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3406 PAGE_FRAG_CACHE_MAX_ORDER);
3407 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3409 if (unlikely(!page))
3410 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3412 nc->va = page ? page_address(page) : NULL;
3417 void *__alloc_page_frag(struct page_frag_cache *nc,
3418 unsigned int fragsz, gfp_t gfp_mask)
3420 unsigned int size = PAGE_SIZE;
3424 if (unlikely(!nc->va)) {
3426 page = __page_frag_refill(nc, gfp_mask);
3430 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3431 /* if size can vary use size else just use PAGE_SIZE */
3434 /* Even if we own the page, we do not use atomic_set().
3435 * This would break get_page_unless_zero() users.
3437 atomic_add(size - 1, &page->_count);
3439 /* reset page count bias and offset to start of new frag */
3440 nc->pfmemalloc = page_is_pfmemalloc(page);
3441 nc->pagecnt_bias = size;
3445 offset = nc->offset - fragsz;
3446 if (unlikely(offset < 0)) {
3447 page = virt_to_page(nc->va);
3449 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3452 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3453 /* if size can vary use size else just use PAGE_SIZE */
3456 /* OK, page count is 0, we can safely set it */
3457 atomic_set(&page->_count, size);
3459 /* reset page count bias and offset to start of new frag */
3460 nc->pagecnt_bias = size;
3461 offset = size - fragsz;
3465 nc->offset = offset;
3467 return nc->va + offset;
3469 EXPORT_SYMBOL(__alloc_page_frag);
3472 * Frees a page fragment allocated out of either a compound or order 0 page.
3474 void __free_page_frag(void *addr)
3476 struct page *page = virt_to_head_page(addr);
3478 if (unlikely(put_page_testzero(page)))
3479 __free_pages_ok(page, compound_order(page));
3481 EXPORT_SYMBOL(__free_page_frag);
3484 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3485 * of the current memory cgroup.
3487 * It should be used when the caller would like to use kmalloc, but since the
3488 * allocation is large, it has to fall back to the page allocator.
3490 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3494 page = alloc_pages(gfp_mask, order);
3495 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3496 __free_pages(page, order);
3502 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3506 page = alloc_pages_node(nid, gfp_mask, order);
3507 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3508 __free_pages(page, order);
3515 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3518 void __free_kmem_pages(struct page *page, unsigned int order)
3520 memcg_kmem_uncharge(page, order);
3521 __free_pages(page, order);
3524 void free_kmem_pages(unsigned long addr, unsigned int order)
3527 VM_BUG_ON(!virt_addr_valid((void *)addr));
3528 __free_kmem_pages(virt_to_page((void *)addr), order);
3532 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3536 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3537 unsigned long used = addr + PAGE_ALIGN(size);
3539 split_page(virt_to_page((void *)addr), order);
3540 while (used < alloc_end) {
3545 return (void *)addr;
3549 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3550 * @size: the number of bytes to allocate
3551 * @gfp_mask: GFP flags for the allocation
3553 * This function is similar to alloc_pages(), except that it allocates the
3554 * minimum number of pages to satisfy the request. alloc_pages() can only
3555 * allocate memory in power-of-two pages.
3557 * This function is also limited by MAX_ORDER.
3559 * Memory allocated by this function must be released by free_pages_exact().
3561 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3563 unsigned int order = get_order(size);
3566 addr = __get_free_pages(gfp_mask, order);
3567 return make_alloc_exact(addr, order, size);
3569 EXPORT_SYMBOL(alloc_pages_exact);
3572 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3574 * @nid: the preferred node ID where memory should be allocated
3575 * @size: the number of bytes to allocate
3576 * @gfp_mask: GFP flags for the allocation
3578 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3581 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3583 unsigned int order = get_order(size);
3584 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3587 return make_alloc_exact((unsigned long)page_address(p), order, size);
3591 * free_pages_exact - release memory allocated via alloc_pages_exact()
3592 * @virt: the value returned by alloc_pages_exact.
3593 * @size: size of allocation, same value as passed to alloc_pages_exact().
3595 * Release the memory allocated by a previous call to alloc_pages_exact.
3597 void free_pages_exact(void *virt, size_t size)
3599 unsigned long addr = (unsigned long)virt;
3600 unsigned long end = addr + PAGE_ALIGN(size);
3602 while (addr < end) {
3607 EXPORT_SYMBOL(free_pages_exact);
3610 * nr_free_zone_pages - count number of pages beyond high watermark
3611 * @offset: The zone index of the highest zone
3613 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3614 * high watermark within all zones at or below a given zone index. For each
3615 * zone, the number of pages is calculated as:
3616 * managed_pages - high_pages
3618 static unsigned long nr_free_zone_pages(int offset)
3623 /* Just pick one node, since fallback list is circular */
3624 unsigned long sum = 0;
3626 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3628 for_each_zone_zonelist(zone, z, zonelist, offset) {
3629 unsigned long size = zone->managed_pages;
3630 unsigned long high = high_wmark_pages(zone);
3639 * nr_free_buffer_pages - count number of pages beyond high watermark
3641 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3642 * watermark within ZONE_DMA and ZONE_NORMAL.
3644 unsigned long nr_free_buffer_pages(void)
3646 return nr_free_zone_pages(gfp_zone(GFP_USER));
3648 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3651 * nr_free_pagecache_pages - count number of pages beyond high watermark
3653 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3654 * high watermark within all zones.
3656 unsigned long nr_free_pagecache_pages(void)
3658 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3661 static inline void show_node(struct zone *zone)
3663 if (IS_ENABLED(CONFIG_NUMA))
3664 printk("Node %d ", zone_to_nid(zone));
3667 void si_meminfo(struct sysinfo *val)
3669 val->totalram = totalram_pages;
3670 val->sharedram = global_page_state(NR_SHMEM);
3671 val->freeram = global_page_state(NR_FREE_PAGES);
3672 val->bufferram = nr_blockdev_pages();
3673 val->totalhigh = totalhigh_pages;
3674 val->freehigh = nr_free_highpages();
3675 val->mem_unit = PAGE_SIZE;
3678 EXPORT_SYMBOL(si_meminfo);
3681 void si_meminfo_node(struct sysinfo *val, int nid)
3683 int zone_type; /* needs to be signed */
3684 unsigned long managed_pages = 0;
3685 pg_data_t *pgdat = NODE_DATA(nid);
3687 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3688 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3689 val->totalram = managed_pages;
3690 val->sharedram = node_page_state(nid, NR_SHMEM);
3691 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3692 #ifdef CONFIG_HIGHMEM
3693 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3694 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3700 val->mem_unit = PAGE_SIZE;
3705 * Determine whether the node should be displayed or not, depending on whether
3706 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3708 bool skip_free_areas_node(unsigned int flags, int nid)
3711 unsigned int cpuset_mems_cookie;
3713 if (!(flags & SHOW_MEM_FILTER_NODES))
3717 cpuset_mems_cookie = read_mems_allowed_begin();
3718 ret = !node_isset(nid, cpuset_current_mems_allowed);
3719 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3724 #define K(x) ((x) << (PAGE_SHIFT-10))
3726 static void show_migration_types(unsigned char type)
3728 static const char types[MIGRATE_TYPES] = {
3729 [MIGRATE_UNMOVABLE] = 'U',
3730 [MIGRATE_MOVABLE] = 'M',
3731 [MIGRATE_RECLAIMABLE] = 'E',
3732 [MIGRATE_HIGHATOMIC] = 'H',
3734 [MIGRATE_CMA] = 'C',
3736 #ifdef CONFIG_MEMORY_ISOLATION
3737 [MIGRATE_ISOLATE] = 'I',
3740 char tmp[MIGRATE_TYPES + 1];
3744 for (i = 0; i < MIGRATE_TYPES; i++) {
3745 if (type & (1 << i))
3750 printk("(%s) ", tmp);
3754 * Show free area list (used inside shift_scroll-lock stuff)
3755 * We also calculate the percentage fragmentation. We do this by counting the
3756 * memory on each free list with the exception of the first item on the list.
3759 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3762 void show_free_areas(unsigned int filter)
3764 unsigned long free_pcp = 0;
3768 for_each_populated_zone(zone) {
3769 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3772 for_each_online_cpu(cpu)
3773 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3776 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3777 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3778 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3779 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3780 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3781 " free:%lu free_pcp:%lu free_cma:%lu\n",
3782 global_page_state(NR_ACTIVE_ANON),
3783 global_page_state(NR_INACTIVE_ANON),
3784 global_page_state(NR_ISOLATED_ANON),
3785 global_page_state(NR_ACTIVE_FILE),
3786 global_page_state(NR_INACTIVE_FILE),
3787 global_page_state(NR_ISOLATED_FILE),
3788 global_page_state(NR_UNEVICTABLE),
3789 global_page_state(NR_FILE_DIRTY),
3790 global_page_state(NR_WRITEBACK),
3791 global_page_state(NR_UNSTABLE_NFS),
3792 global_page_state(NR_SLAB_RECLAIMABLE),
3793 global_page_state(NR_SLAB_UNRECLAIMABLE),
3794 global_page_state(NR_FILE_MAPPED),
3795 global_page_state(NR_SHMEM),
3796 global_page_state(NR_PAGETABLE),
3797 global_page_state(NR_BOUNCE),
3798 global_page_state(NR_FREE_PAGES),
3800 global_page_state(NR_FREE_CMA_PAGES));
3802 for_each_populated_zone(zone) {
3805 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3809 for_each_online_cpu(cpu)
3810 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3818 " active_anon:%lukB"
3819 " inactive_anon:%lukB"
3820 " active_file:%lukB"
3821 " inactive_file:%lukB"
3822 " unevictable:%lukB"
3823 " isolated(anon):%lukB"
3824 " isolated(file):%lukB"
3832 " slab_reclaimable:%lukB"
3833 " slab_unreclaimable:%lukB"
3834 " kernel_stack:%lukB"
3841 " writeback_tmp:%lukB"
3842 " pages_scanned:%lu"
3843 " all_unreclaimable? %s"
3846 K(zone_page_state(zone, NR_FREE_PAGES)),
3847 K(min_wmark_pages(zone)),
3848 K(low_wmark_pages(zone)),
3849 K(high_wmark_pages(zone)),
3850 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3851 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3852 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3853 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3854 K(zone_page_state(zone, NR_UNEVICTABLE)),
3855 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3856 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3857 K(zone->present_pages),
3858 K(zone->managed_pages),
3859 K(zone_page_state(zone, NR_MLOCK)),
3860 K(zone_page_state(zone, NR_FILE_DIRTY)),
3861 K(zone_page_state(zone, NR_WRITEBACK)),
3862 K(zone_page_state(zone, NR_FILE_MAPPED)),
3863 K(zone_page_state(zone, NR_SHMEM)),
3864 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3865 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3866 zone_page_state(zone, NR_KERNEL_STACK) *
3868 K(zone_page_state(zone, NR_PAGETABLE)),
3869 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3870 K(zone_page_state(zone, NR_BOUNCE)),
3872 K(this_cpu_read(zone->pageset->pcp.count)),
3873 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3874 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3875 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3876 (!zone_reclaimable(zone) ? "yes" : "no")
3878 printk("lowmem_reserve[]:");
3879 for (i = 0; i < MAX_NR_ZONES; i++)
3880 printk(" %ld", zone->lowmem_reserve[i]);
3884 for_each_populated_zone(zone) {
3886 unsigned long nr[MAX_ORDER], flags, total = 0;
3887 unsigned char types[MAX_ORDER];
3889 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3892 printk("%s: ", zone->name);
3894 spin_lock_irqsave(&zone->lock, flags);
3895 for (order = 0; order < MAX_ORDER; order++) {
3896 struct free_area *area = &zone->free_area[order];
3899 nr[order] = area->nr_free;
3900 total += nr[order] << order;
3903 for (type = 0; type < MIGRATE_TYPES; type++) {
3904 if (!list_empty(&area->free_list[type]))
3905 types[order] |= 1 << type;
3908 spin_unlock_irqrestore(&zone->lock, flags);
3909 for (order = 0; order < MAX_ORDER; order++) {
3910 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3912 show_migration_types(types[order]);
3914 printk("= %lukB\n", K(total));
3917 hugetlb_show_meminfo();
3919 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3921 show_swap_cache_info();
3924 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3926 zoneref->zone = zone;
3927 zoneref->zone_idx = zone_idx(zone);
3931 * Builds allocation fallback zone lists.
3933 * Add all populated zones of a node to the zonelist.
3935 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3939 enum zone_type zone_type = MAX_NR_ZONES;
3943 zone = pgdat->node_zones + zone_type;
3944 if (populated_zone(zone)) {
3945 zoneref_set_zone(zone,
3946 &zonelist->_zonerefs[nr_zones++]);
3947 check_highest_zone(zone_type);
3949 } while (zone_type);
3957 * 0 = automatic detection of better ordering.
3958 * 1 = order by ([node] distance, -zonetype)
3959 * 2 = order by (-zonetype, [node] distance)
3961 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3962 * the same zonelist. So only NUMA can configure this param.
3964 #define ZONELIST_ORDER_DEFAULT 0
3965 #define ZONELIST_ORDER_NODE 1
3966 #define ZONELIST_ORDER_ZONE 2
3968 /* zonelist order in the kernel.
3969 * set_zonelist_order() will set this to NODE or ZONE.
3971 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3972 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3976 /* The value user specified ....changed by config */
3977 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3978 /* string for sysctl */
3979 #define NUMA_ZONELIST_ORDER_LEN 16
3980 char numa_zonelist_order[16] = "default";
3983 * interface for configure zonelist ordering.
3984 * command line option "numa_zonelist_order"
3985 * = "[dD]efault - default, automatic configuration.
3986 * = "[nN]ode - order by node locality, then by zone within node
3987 * = "[zZ]one - order by zone, then by locality within zone
3990 static int __parse_numa_zonelist_order(char *s)
3992 if (*s == 'd' || *s == 'D') {
3993 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3994 } else if (*s == 'n' || *s == 'N') {
3995 user_zonelist_order = ZONELIST_ORDER_NODE;
3996 } else if (*s == 'z' || *s == 'Z') {
3997 user_zonelist_order = ZONELIST_ORDER_ZONE;
4000 "Ignoring invalid numa_zonelist_order value: %s\n", s);
4006 static __init int setup_numa_zonelist_order(char *s)
4013 ret = __parse_numa_zonelist_order(s);
4015 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4019 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4022 * sysctl handler for numa_zonelist_order
4024 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4025 void __user *buffer, size_t *length,
4028 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4030 static DEFINE_MUTEX(zl_order_mutex);
4032 mutex_lock(&zl_order_mutex);
4034 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4038 strcpy(saved_string, (char *)table->data);
4040 ret = proc_dostring(table, write, buffer, length, ppos);
4044 int oldval = user_zonelist_order;
4046 ret = __parse_numa_zonelist_order((char *)table->data);
4049 * bogus value. restore saved string
4051 strncpy((char *)table->data, saved_string,
4052 NUMA_ZONELIST_ORDER_LEN);
4053 user_zonelist_order = oldval;
4054 } else if (oldval != user_zonelist_order) {
4055 mutex_lock(&zonelists_mutex);
4056 build_all_zonelists(NULL, NULL);
4057 mutex_unlock(&zonelists_mutex);
4061 mutex_unlock(&zl_order_mutex);
4066 #define MAX_NODE_LOAD (nr_online_nodes)
4067 static int node_load[MAX_NUMNODES];
4070 * find_next_best_node - find the next node that should appear in a given node's fallback list
4071 * @node: node whose fallback list we're appending
4072 * @used_node_mask: nodemask_t of already used nodes
4074 * We use a number of factors to determine which is the next node that should
4075 * appear on a given node's fallback list. The node should not have appeared
4076 * already in @node's fallback list, and it should be the next closest node
4077 * according to the distance array (which contains arbitrary distance values
4078 * from each node to each node in the system), and should also prefer nodes
4079 * with no CPUs, since presumably they'll have very little allocation pressure
4080 * on them otherwise.
4081 * It returns -1 if no node is found.
4083 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4086 int min_val = INT_MAX;
4087 int best_node = NUMA_NO_NODE;
4088 const struct cpumask *tmp = cpumask_of_node(0);
4090 /* Use the local node if we haven't already */
4091 if (!node_isset(node, *used_node_mask)) {
4092 node_set(node, *used_node_mask);
4096 for_each_node_state(n, N_MEMORY) {
4098 /* Don't want a node to appear more than once */
4099 if (node_isset(n, *used_node_mask))
4102 /* Use the distance array to find the distance */
4103 val = node_distance(node, n);
4105 /* Penalize nodes under us ("prefer the next node") */
4108 /* Give preference to headless and unused nodes */
4109 tmp = cpumask_of_node(n);
4110 if (!cpumask_empty(tmp))
4111 val += PENALTY_FOR_NODE_WITH_CPUS;
4113 /* Slight preference for less loaded node */
4114 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4115 val += node_load[n];
4117 if (val < min_val) {
4124 node_set(best_node, *used_node_mask);
4131 * Build zonelists ordered by node and zones within node.
4132 * This results in maximum locality--normal zone overflows into local
4133 * DMA zone, if any--but risks exhausting DMA zone.
4135 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4138 struct zonelist *zonelist;
4140 zonelist = &pgdat->node_zonelists[0];
4141 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4143 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4144 zonelist->_zonerefs[j].zone = NULL;
4145 zonelist->_zonerefs[j].zone_idx = 0;
4149 * Build gfp_thisnode zonelists
4151 static void build_thisnode_zonelists(pg_data_t *pgdat)
4154 struct zonelist *zonelist;
4156 zonelist = &pgdat->node_zonelists[1];
4157 j = build_zonelists_node(pgdat, zonelist, 0);
4158 zonelist->_zonerefs[j].zone = NULL;
4159 zonelist->_zonerefs[j].zone_idx = 0;
4163 * Build zonelists ordered by zone and nodes within zones.
4164 * This results in conserving DMA zone[s] until all Normal memory is
4165 * exhausted, but results in overflowing to remote node while memory
4166 * may still exist in local DMA zone.
4168 static int node_order[MAX_NUMNODES];
4170 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4173 int zone_type; /* needs to be signed */
4175 struct zonelist *zonelist;
4177 zonelist = &pgdat->node_zonelists[0];
4179 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4180 for (j = 0; j < nr_nodes; j++) {
4181 node = node_order[j];
4182 z = &NODE_DATA(node)->node_zones[zone_type];
4183 if (populated_zone(z)) {
4185 &zonelist->_zonerefs[pos++]);
4186 check_highest_zone(zone_type);
4190 zonelist->_zonerefs[pos].zone = NULL;
4191 zonelist->_zonerefs[pos].zone_idx = 0;
4194 #if defined(CONFIG_64BIT)
4196 * Devices that require DMA32/DMA are relatively rare and do not justify a
4197 * penalty to every machine in case the specialised case applies. Default
4198 * to Node-ordering on 64-bit NUMA machines
4200 static int default_zonelist_order(void)
4202 return ZONELIST_ORDER_NODE;
4206 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4207 * by the kernel. If processes running on node 0 deplete the low memory zone
4208 * then reclaim will occur more frequency increasing stalls and potentially
4209 * be easier to OOM if a large percentage of the zone is under writeback or
4210 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4211 * Hence, default to zone ordering on 32-bit.
4213 static int default_zonelist_order(void)
4215 return ZONELIST_ORDER_ZONE;
4217 #endif /* CONFIG_64BIT */
4219 static void set_zonelist_order(void)
4221 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4222 current_zonelist_order = default_zonelist_order();
4224 current_zonelist_order = user_zonelist_order;
4227 static void build_zonelists(pg_data_t *pgdat)
4231 nodemask_t used_mask;
4232 int local_node, prev_node;
4233 struct zonelist *zonelist;
4234 unsigned int order = current_zonelist_order;
4236 /* initialize zonelists */
4237 for (i = 0; i < MAX_ZONELISTS; i++) {
4238 zonelist = pgdat->node_zonelists + i;
4239 zonelist->_zonerefs[0].zone = NULL;
4240 zonelist->_zonerefs[0].zone_idx = 0;
4243 /* NUMA-aware ordering of nodes */
4244 local_node = pgdat->node_id;
4245 load = nr_online_nodes;
4246 prev_node = local_node;
4247 nodes_clear(used_mask);
4249 memset(node_order, 0, sizeof(node_order));
4252 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4254 * We don't want to pressure a particular node.
4255 * So adding penalty to the first node in same
4256 * distance group to make it round-robin.
4258 if (node_distance(local_node, node) !=
4259 node_distance(local_node, prev_node))
4260 node_load[node] = load;
4264 if (order == ZONELIST_ORDER_NODE)
4265 build_zonelists_in_node_order(pgdat, node);
4267 node_order[j++] = node; /* remember order */
4270 if (order == ZONELIST_ORDER_ZONE) {
4271 /* calculate node order -- i.e., DMA last! */
4272 build_zonelists_in_zone_order(pgdat, j);
4275 build_thisnode_zonelists(pgdat);
4278 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4280 * Return node id of node used for "local" allocations.
4281 * I.e., first node id of first zone in arg node's generic zonelist.
4282 * Used for initializing percpu 'numa_mem', which is used primarily
4283 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4285 int local_memory_node(int node)
4289 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4290 gfp_zone(GFP_KERNEL),
4297 #else /* CONFIG_NUMA */
4299 static void set_zonelist_order(void)
4301 current_zonelist_order = ZONELIST_ORDER_ZONE;
4304 static void build_zonelists(pg_data_t *pgdat)
4306 int node, local_node;
4308 struct zonelist *zonelist;
4310 local_node = pgdat->node_id;
4312 zonelist = &pgdat->node_zonelists[0];
4313 j = build_zonelists_node(pgdat, zonelist, 0);
4316 * Now we build the zonelist so that it contains the zones
4317 * of all the other nodes.
4318 * We don't want to pressure a particular node, so when
4319 * building the zones for node N, we make sure that the
4320 * zones coming right after the local ones are those from
4321 * node N+1 (modulo N)
4323 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4324 if (!node_online(node))
4326 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4328 for (node = 0; node < local_node; node++) {
4329 if (!node_online(node))
4331 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4334 zonelist->_zonerefs[j].zone = NULL;
4335 zonelist->_zonerefs[j].zone_idx = 0;
4338 #endif /* CONFIG_NUMA */
4341 * Boot pageset table. One per cpu which is going to be used for all
4342 * zones and all nodes. The parameters will be set in such a way
4343 * that an item put on a list will immediately be handed over to
4344 * the buddy list. This is safe since pageset manipulation is done
4345 * with interrupts disabled.
4347 * The boot_pagesets must be kept even after bootup is complete for
4348 * unused processors and/or zones. They do play a role for bootstrapping
4349 * hotplugged processors.
4351 * zoneinfo_show() and maybe other functions do
4352 * not check if the processor is online before following the pageset pointer.
4353 * Other parts of the kernel may not check if the zone is available.
4355 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4356 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4357 static void setup_zone_pageset(struct zone *zone);
4360 * Global mutex to protect against size modification of zonelists
4361 * as well as to serialize pageset setup for the new populated zone.
4363 DEFINE_MUTEX(zonelists_mutex);
4365 /* return values int ....just for stop_machine() */
4366 static int __build_all_zonelists(void *data)
4370 pg_data_t *self = data;
4373 memset(node_load, 0, sizeof(node_load));
4376 if (self && !node_online(self->node_id)) {
4377 build_zonelists(self);
4380 for_each_online_node(nid) {
4381 pg_data_t *pgdat = NODE_DATA(nid);
4383 build_zonelists(pgdat);
4387 * Initialize the boot_pagesets that are going to be used
4388 * for bootstrapping processors. The real pagesets for
4389 * each zone will be allocated later when the per cpu
4390 * allocator is available.
4392 * boot_pagesets are used also for bootstrapping offline
4393 * cpus if the system is already booted because the pagesets
4394 * are needed to initialize allocators on a specific cpu too.
4395 * F.e. the percpu allocator needs the page allocator which
4396 * needs the percpu allocator in order to allocate its pagesets
4397 * (a chicken-egg dilemma).
4399 for_each_possible_cpu(cpu) {
4400 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4402 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4404 * We now know the "local memory node" for each node--
4405 * i.e., the node of the first zone in the generic zonelist.
4406 * Set up numa_mem percpu variable for on-line cpus. During
4407 * boot, only the boot cpu should be on-line; we'll init the
4408 * secondary cpus' numa_mem as they come on-line. During
4409 * node/memory hotplug, we'll fixup all on-line cpus.
4411 if (cpu_online(cpu))
4412 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4419 static noinline void __init
4420 build_all_zonelists_init(void)
4422 __build_all_zonelists(NULL);
4423 mminit_verify_zonelist();
4424 cpuset_init_current_mems_allowed();
4428 * Called with zonelists_mutex held always
4429 * unless system_state == SYSTEM_BOOTING.
4431 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4432 * [we're only called with non-NULL zone through __meminit paths] and
4433 * (2) call of __init annotated helper build_all_zonelists_init
4434 * [protected by SYSTEM_BOOTING].
4436 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4438 set_zonelist_order();
4440 if (system_state == SYSTEM_BOOTING) {
4441 build_all_zonelists_init();
4443 #ifdef CONFIG_MEMORY_HOTPLUG
4445 setup_zone_pageset(zone);
4447 /* we have to stop all cpus to guarantee there is no user
4449 stop_machine(__build_all_zonelists, pgdat, NULL);
4450 /* cpuset refresh routine should be here */
4452 vm_total_pages = nr_free_pagecache_pages();
4454 * Disable grouping by mobility if the number of pages in the
4455 * system is too low to allow the mechanism to work. It would be
4456 * more accurate, but expensive to check per-zone. This check is
4457 * made on memory-hotadd so a system can start with mobility
4458 * disabled and enable it later
4460 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4461 page_group_by_mobility_disabled = 1;
4463 page_group_by_mobility_disabled = 0;
4465 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4467 zonelist_order_name[current_zonelist_order],
4468 page_group_by_mobility_disabled ? "off" : "on",
4471 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4476 * Helper functions to size the waitqueue hash table.
4477 * Essentially these want to choose hash table sizes sufficiently
4478 * large so that collisions trying to wait on pages are rare.
4479 * But in fact, the number of active page waitqueues on typical
4480 * systems is ridiculously low, less than 200. So this is even
4481 * conservative, even though it seems large.
4483 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4484 * waitqueues, i.e. the size of the waitq table given the number of pages.
4486 #define PAGES_PER_WAITQUEUE 256
4488 #ifndef CONFIG_MEMORY_HOTPLUG
4489 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4491 unsigned long size = 1;
4493 pages /= PAGES_PER_WAITQUEUE;
4495 while (size < pages)
4499 * Once we have dozens or even hundreds of threads sleeping
4500 * on IO we've got bigger problems than wait queue collision.
4501 * Limit the size of the wait table to a reasonable size.
4503 size = min(size, 4096UL);
4505 return max(size, 4UL);
4509 * A zone's size might be changed by hot-add, so it is not possible to determine
4510 * a suitable size for its wait_table. So we use the maximum size now.
4512 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4514 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4515 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4516 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4518 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4519 * or more by the traditional way. (See above). It equals:
4521 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4522 * ia64(16K page size) : = ( 8G + 4M)byte.
4523 * powerpc (64K page size) : = (32G +16M)byte.
4525 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4532 * This is an integer logarithm so that shifts can be used later
4533 * to extract the more random high bits from the multiplicative
4534 * hash function before the remainder is taken.
4536 static inline unsigned long wait_table_bits(unsigned long size)
4542 * Initially all pages are reserved - free ones are freed
4543 * up by free_all_bootmem() once the early boot process is
4544 * done. Non-atomic initialization, single-pass.
4546 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4547 unsigned long start_pfn, enum memmap_context context)
4549 pg_data_t *pgdat = NODE_DATA(nid);
4550 unsigned long end_pfn = start_pfn + size;
4553 unsigned long nr_initialised = 0;
4555 if (highest_memmap_pfn < end_pfn - 1)
4556 highest_memmap_pfn = end_pfn - 1;
4558 z = &pgdat->node_zones[zone];
4559 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4561 * There can be holes in boot-time mem_map[]s
4562 * handed to this function. They do not
4563 * exist on hotplugged memory.
4565 if (context == MEMMAP_EARLY) {
4566 if (!early_pfn_valid(pfn))
4568 if (!early_pfn_in_nid(pfn, nid))
4570 if (!update_defer_init(pgdat, pfn, end_pfn,
4576 * Mark the block movable so that blocks are reserved for
4577 * movable at startup. This will force kernel allocations
4578 * to reserve their blocks rather than leaking throughout
4579 * the address space during boot when many long-lived
4580 * kernel allocations are made.
4582 * bitmap is created for zone's valid pfn range. but memmap
4583 * can be created for invalid pages (for alignment)
4584 * check here not to call set_pageblock_migratetype() against
4587 if (!(pfn & (pageblock_nr_pages - 1))) {
4588 struct page *page = pfn_to_page(pfn);
4590 __init_single_page(page, pfn, zone, nid);
4591 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4593 __init_single_pfn(pfn, zone, nid);
4598 static void __meminit zone_init_free_lists(struct zone *zone)
4600 unsigned int order, t;
4601 for_each_migratetype_order(order, t) {
4602 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4603 zone->free_area[order].nr_free = 0;
4607 #ifndef __HAVE_ARCH_MEMMAP_INIT
4608 #define memmap_init(size, nid, zone, start_pfn) \
4609 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4612 static int zone_batchsize(struct zone *zone)
4618 * The per-cpu-pages pools are set to around 1000th of the
4619 * size of the zone. But no more than 1/2 of a meg.
4621 * OK, so we don't know how big the cache is. So guess.
4623 batch = zone->managed_pages / 1024;
4624 if (batch * PAGE_SIZE > 512 * 1024)
4625 batch = (512 * 1024) / PAGE_SIZE;
4626 batch /= 4; /* We effectively *= 4 below */
4631 * Clamp the batch to a 2^n - 1 value. Having a power
4632 * of 2 value was found to be more likely to have
4633 * suboptimal cache aliasing properties in some cases.
4635 * For example if 2 tasks are alternately allocating
4636 * batches of pages, one task can end up with a lot
4637 * of pages of one half of the possible page colors
4638 * and the other with pages of the other colors.
4640 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4645 /* The deferral and batching of frees should be suppressed under NOMMU
4648 * The problem is that NOMMU needs to be able to allocate large chunks
4649 * of contiguous memory as there's no hardware page translation to
4650 * assemble apparent contiguous memory from discontiguous pages.
4652 * Queueing large contiguous runs of pages for batching, however,
4653 * causes the pages to actually be freed in smaller chunks. As there
4654 * can be a significant delay between the individual batches being
4655 * recycled, this leads to the once large chunks of space being
4656 * fragmented and becoming unavailable for high-order allocations.
4663 * pcp->high and pcp->batch values are related and dependent on one another:
4664 * ->batch must never be higher then ->high.
4665 * The following function updates them in a safe manner without read side
4668 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4669 * those fields changing asynchronously (acording the the above rule).
4671 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4672 * outside of boot time (or some other assurance that no concurrent updaters
4675 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4676 unsigned long batch)
4678 /* start with a fail safe value for batch */
4682 /* Update high, then batch, in order */
4689 /* a companion to pageset_set_high() */
4690 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4692 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4695 static void pageset_init(struct per_cpu_pageset *p)
4697 struct per_cpu_pages *pcp;
4700 memset(p, 0, sizeof(*p));
4704 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4705 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4708 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4711 pageset_set_batch(p, batch);
4715 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4716 * to the value high for the pageset p.
4718 static void pageset_set_high(struct per_cpu_pageset *p,
4721 unsigned long batch = max(1UL, high / 4);
4722 if ((high / 4) > (PAGE_SHIFT * 8))
4723 batch = PAGE_SHIFT * 8;
4725 pageset_update(&p->pcp, high, batch);
4728 static void pageset_set_high_and_batch(struct zone *zone,
4729 struct per_cpu_pageset *pcp)
4731 if (percpu_pagelist_fraction)
4732 pageset_set_high(pcp,
4733 (zone->managed_pages /
4734 percpu_pagelist_fraction));
4736 pageset_set_batch(pcp, zone_batchsize(zone));
4739 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4741 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4744 pageset_set_high_and_batch(zone, pcp);
4747 static void __meminit setup_zone_pageset(struct zone *zone)
4750 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4751 for_each_possible_cpu(cpu)
4752 zone_pageset_init(zone, cpu);
4756 * Allocate per cpu pagesets and initialize them.
4757 * Before this call only boot pagesets were available.
4759 void __init setup_per_cpu_pageset(void)
4763 for_each_populated_zone(zone)
4764 setup_zone_pageset(zone);
4767 static noinline __init_refok
4768 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4774 * The per-page waitqueue mechanism uses hashed waitqueues
4777 zone->wait_table_hash_nr_entries =
4778 wait_table_hash_nr_entries(zone_size_pages);
4779 zone->wait_table_bits =
4780 wait_table_bits(zone->wait_table_hash_nr_entries);
4781 alloc_size = zone->wait_table_hash_nr_entries
4782 * sizeof(wait_queue_head_t);
4784 if (!slab_is_available()) {
4785 zone->wait_table = (wait_queue_head_t *)
4786 memblock_virt_alloc_node_nopanic(
4787 alloc_size, zone->zone_pgdat->node_id);
4790 * This case means that a zone whose size was 0 gets new memory
4791 * via memory hot-add.
4792 * But it may be the case that a new node was hot-added. In
4793 * this case vmalloc() will not be able to use this new node's
4794 * memory - this wait_table must be initialized to use this new
4795 * node itself as well.
4796 * To use this new node's memory, further consideration will be
4799 zone->wait_table = vmalloc(alloc_size);
4801 if (!zone->wait_table)
4804 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4805 init_waitqueue_head(zone->wait_table + i);
4810 static __meminit void zone_pcp_init(struct zone *zone)
4813 * per cpu subsystem is not up at this point. The following code
4814 * relies on the ability of the linker to provide the
4815 * offset of a (static) per cpu variable into the per cpu area.
4817 zone->pageset = &boot_pageset;
4819 if (populated_zone(zone))
4820 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4821 zone->name, zone->present_pages,
4822 zone_batchsize(zone));
4825 int __meminit init_currently_empty_zone(struct zone *zone,
4826 unsigned long zone_start_pfn,
4829 struct pglist_data *pgdat = zone->zone_pgdat;
4831 ret = zone_wait_table_init(zone, size);
4834 pgdat->nr_zones = zone_idx(zone) + 1;
4836 zone->zone_start_pfn = zone_start_pfn;
4838 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4839 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4841 (unsigned long)zone_idx(zone),
4842 zone_start_pfn, (zone_start_pfn + size));
4844 zone_init_free_lists(zone);
4849 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4850 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4853 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4855 int __meminit __early_pfn_to_nid(unsigned long pfn,
4856 struct mminit_pfnnid_cache *state)
4858 unsigned long start_pfn, end_pfn;
4861 if (state->last_start <= pfn && pfn < state->last_end)
4862 return state->last_nid;
4864 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4866 state->last_start = start_pfn;
4867 state->last_end = end_pfn;
4868 state->last_nid = nid;
4873 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4876 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4877 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4878 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4880 * If an architecture guarantees that all ranges registered contain no holes
4881 * and may be freed, this this function may be used instead of calling
4882 * memblock_free_early_nid() manually.
4884 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4886 unsigned long start_pfn, end_pfn;
4889 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4890 start_pfn = min(start_pfn, max_low_pfn);
4891 end_pfn = min(end_pfn, max_low_pfn);
4893 if (start_pfn < end_pfn)
4894 memblock_free_early_nid(PFN_PHYS(start_pfn),
4895 (end_pfn - start_pfn) << PAGE_SHIFT,
4901 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4902 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4904 * If an architecture guarantees that all ranges registered contain no holes and may
4905 * be freed, this function may be used instead of calling memory_present() manually.
4907 void __init sparse_memory_present_with_active_regions(int nid)
4909 unsigned long start_pfn, end_pfn;
4912 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4913 memory_present(this_nid, start_pfn, end_pfn);
4917 * get_pfn_range_for_nid - Return the start and end page frames for a node
4918 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4919 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4920 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4922 * It returns the start and end page frame of a node based on information
4923 * provided by memblock_set_node(). If called for a node
4924 * with no available memory, a warning is printed and the start and end
4927 void __meminit get_pfn_range_for_nid(unsigned int nid,
4928 unsigned long *start_pfn, unsigned long *end_pfn)
4930 unsigned long this_start_pfn, this_end_pfn;
4936 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4937 *start_pfn = min(*start_pfn, this_start_pfn);
4938 *end_pfn = max(*end_pfn, this_end_pfn);
4941 if (*start_pfn == -1UL)
4946 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4947 * assumption is made that zones within a node are ordered in monotonic
4948 * increasing memory addresses so that the "highest" populated zone is used
4950 static void __init find_usable_zone_for_movable(void)
4953 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4954 if (zone_index == ZONE_MOVABLE)
4957 if (arch_zone_highest_possible_pfn[zone_index] >
4958 arch_zone_lowest_possible_pfn[zone_index])
4962 VM_BUG_ON(zone_index == -1);
4963 movable_zone = zone_index;
4967 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4968 * because it is sized independent of architecture. Unlike the other zones,
4969 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4970 * in each node depending on the size of each node and how evenly kernelcore
4971 * is distributed. This helper function adjusts the zone ranges
4972 * provided by the architecture for a given node by using the end of the
4973 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4974 * zones within a node are in order of monotonic increases memory addresses
4976 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4977 unsigned long zone_type,
4978 unsigned long node_start_pfn,
4979 unsigned long node_end_pfn,
4980 unsigned long *zone_start_pfn,
4981 unsigned long *zone_end_pfn)
4983 /* Only adjust if ZONE_MOVABLE is on this node */
4984 if (zone_movable_pfn[nid]) {
4985 /* Size ZONE_MOVABLE */
4986 if (zone_type == ZONE_MOVABLE) {
4987 *zone_start_pfn = zone_movable_pfn[nid];
4988 *zone_end_pfn = min(node_end_pfn,
4989 arch_zone_highest_possible_pfn[movable_zone]);
4991 /* Adjust for ZONE_MOVABLE starting within this range */
4992 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4993 *zone_end_pfn > zone_movable_pfn[nid]) {
4994 *zone_end_pfn = zone_movable_pfn[nid];
4996 /* Check if this whole range is within ZONE_MOVABLE */
4997 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4998 *zone_start_pfn = *zone_end_pfn;
5003 * Return the number of pages a zone spans in a node, including holes
5004 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5006 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5007 unsigned long zone_type,
5008 unsigned long node_start_pfn,
5009 unsigned long node_end_pfn,
5010 unsigned long *ignored)
5012 unsigned long zone_start_pfn, zone_end_pfn;
5014 /* When hotadd a new node from cpu_up(), the node should be empty */
5015 if (!node_start_pfn && !node_end_pfn)
5018 /* Get the start and end of the zone */
5019 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5020 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5021 adjust_zone_range_for_zone_movable(nid, zone_type,
5022 node_start_pfn, node_end_pfn,
5023 &zone_start_pfn, &zone_end_pfn);
5025 /* Check that this node has pages within the zone's required range */
5026 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
5029 /* Move the zone boundaries inside the node if necessary */
5030 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
5031 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5033 /* Return the spanned pages */
5034 return zone_end_pfn - zone_start_pfn;
5038 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5039 * then all holes in the requested range will be accounted for.
5041 unsigned long __meminit __absent_pages_in_range(int nid,
5042 unsigned long range_start_pfn,
5043 unsigned long range_end_pfn)
5045 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5046 unsigned long start_pfn, end_pfn;
5049 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5050 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5051 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5052 nr_absent -= end_pfn - start_pfn;
5058 * absent_pages_in_range - Return number of page frames in holes within a range
5059 * @start_pfn: The start PFN to start searching for holes
5060 * @end_pfn: The end PFN to stop searching for holes
5062 * It returns the number of pages frames in memory holes within a range.
5064 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5065 unsigned long end_pfn)
5067 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5070 /* Return the number of page frames in holes in a zone on a node */
5071 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5072 unsigned long zone_type,
5073 unsigned long node_start_pfn,
5074 unsigned long node_end_pfn,
5075 unsigned long *ignored)
5077 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5078 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5079 unsigned long zone_start_pfn, zone_end_pfn;
5081 /* When hotadd a new node from cpu_up(), the node should be empty */
5082 if (!node_start_pfn && !node_end_pfn)
5085 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5086 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5088 adjust_zone_range_for_zone_movable(nid, zone_type,
5089 node_start_pfn, node_end_pfn,
5090 &zone_start_pfn, &zone_end_pfn);
5091 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5094 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5095 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5096 unsigned long zone_type,
5097 unsigned long node_start_pfn,
5098 unsigned long node_end_pfn,
5099 unsigned long *zones_size)
5101 return zones_size[zone_type];
5104 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5105 unsigned long zone_type,
5106 unsigned long node_start_pfn,
5107 unsigned long node_end_pfn,
5108 unsigned long *zholes_size)
5113 return zholes_size[zone_type];
5116 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5118 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5119 unsigned long node_start_pfn,
5120 unsigned long node_end_pfn,
5121 unsigned long *zones_size,
5122 unsigned long *zholes_size)
5124 unsigned long realtotalpages = 0, totalpages = 0;
5127 for (i = 0; i < MAX_NR_ZONES; i++) {
5128 struct zone *zone = pgdat->node_zones + i;
5129 unsigned long size, real_size;
5131 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5135 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5136 node_start_pfn, node_end_pfn,
5138 zone->spanned_pages = size;
5139 zone->present_pages = real_size;
5142 realtotalpages += real_size;
5145 pgdat->node_spanned_pages = totalpages;
5146 pgdat->node_present_pages = realtotalpages;
5147 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5151 #ifndef CONFIG_SPARSEMEM
5153 * Calculate the size of the zone->blockflags rounded to an unsigned long
5154 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5155 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5156 * round what is now in bits to nearest long in bits, then return it in
5159 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5161 unsigned long usemapsize;
5163 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5164 usemapsize = roundup(zonesize, pageblock_nr_pages);
5165 usemapsize = usemapsize >> pageblock_order;
5166 usemapsize *= NR_PAGEBLOCK_BITS;
5167 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5169 return usemapsize / 8;
5172 static void __init setup_usemap(struct pglist_data *pgdat,
5174 unsigned long zone_start_pfn,
5175 unsigned long zonesize)
5177 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5178 zone->pageblock_flags = NULL;
5180 zone->pageblock_flags =
5181 memblock_virt_alloc_node_nopanic(usemapsize,
5185 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5186 unsigned long zone_start_pfn, unsigned long zonesize) {}
5187 #endif /* CONFIG_SPARSEMEM */
5189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5191 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5192 void __paginginit set_pageblock_order(void)
5196 /* Check that pageblock_nr_pages has not already been setup */
5197 if (pageblock_order)
5200 if (HPAGE_SHIFT > PAGE_SHIFT)
5201 order = HUGETLB_PAGE_ORDER;
5203 order = MAX_ORDER - 1;
5206 * Assume the largest contiguous order of interest is a huge page.
5207 * This value may be variable depending on boot parameters on IA64 and
5210 pageblock_order = order;
5212 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5215 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5216 * is unused as pageblock_order is set at compile-time. See
5217 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5220 void __paginginit set_pageblock_order(void)
5224 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5226 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5227 unsigned long present_pages)
5229 unsigned long pages = spanned_pages;
5232 * Provide a more accurate estimation if there are holes within
5233 * the zone and SPARSEMEM is in use. If there are holes within the
5234 * zone, each populated memory region may cost us one or two extra
5235 * memmap pages due to alignment because memmap pages for each
5236 * populated regions may not naturally algined on page boundary.
5237 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5239 if (spanned_pages > present_pages + (present_pages >> 4) &&
5240 IS_ENABLED(CONFIG_SPARSEMEM))
5241 pages = present_pages;
5243 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5247 * Set up the zone data structures:
5248 * - mark all pages reserved
5249 * - mark all memory queues empty
5250 * - clear the memory bitmaps
5252 * NOTE: pgdat should get zeroed by caller.
5254 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5257 int nid = pgdat->node_id;
5258 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5261 pgdat_resize_init(pgdat);
5262 #ifdef CONFIG_NUMA_BALANCING
5263 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5264 pgdat->numabalancing_migrate_nr_pages = 0;
5265 pgdat->numabalancing_migrate_next_window = jiffies;
5267 init_waitqueue_head(&pgdat->kswapd_wait);
5268 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5269 pgdat_page_ext_init(pgdat);
5271 for (j = 0; j < MAX_NR_ZONES; j++) {
5272 struct zone *zone = pgdat->node_zones + j;
5273 unsigned long size, realsize, freesize, memmap_pages;
5275 size = zone->spanned_pages;
5276 realsize = freesize = zone->present_pages;
5279 * Adjust freesize so that it accounts for how much memory
5280 * is used by this zone for memmap. This affects the watermark
5281 * and per-cpu initialisations
5283 memmap_pages = calc_memmap_size(size, realsize);
5284 if (!is_highmem_idx(j)) {
5285 if (freesize >= memmap_pages) {
5286 freesize -= memmap_pages;
5289 " %s zone: %lu pages used for memmap\n",
5290 zone_names[j], memmap_pages);
5293 " %s zone: %lu pages exceeds freesize %lu\n",
5294 zone_names[j], memmap_pages, freesize);
5297 /* Account for reserved pages */
5298 if (j == 0 && freesize > dma_reserve) {
5299 freesize -= dma_reserve;
5300 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5301 zone_names[0], dma_reserve);
5304 if (!is_highmem_idx(j))
5305 nr_kernel_pages += freesize;
5306 /* Charge for highmem memmap if there are enough kernel pages */
5307 else if (nr_kernel_pages > memmap_pages * 2)
5308 nr_kernel_pages -= memmap_pages;
5309 nr_all_pages += freesize;
5312 * Set an approximate value for lowmem here, it will be adjusted
5313 * when the bootmem allocator frees pages into the buddy system.
5314 * And all highmem pages will be managed by the buddy system.
5316 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5319 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5321 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5323 zone->name = zone_names[j];
5324 spin_lock_init(&zone->lock);
5325 spin_lock_init(&zone->lru_lock);
5326 zone_seqlock_init(zone);
5327 zone->zone_pgdat = pgdat;
5328 zone_pcp_init(zone);
5330 /* For bootup, initialized properly in watermark setup */
5331 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5333 lruvec_init(&zone->lruvec);
5337 set_pageblock_order();
5338 setup_usemap(pgdat, zone, zone_start_pfn, size);
5339 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5341 memmap_init(size, nid, j, zone_start_pfn);
5342 zone_start_pfn += size;
5346 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5348 unsigned long __maybe_unused start = 0;
5349 unsigned long __maybe_unused offset = 0;
5351 /* Skip empty nodes */
5352 if (!pgdat->node_spanned_pages)
5355 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5356 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5357 offset = pgdat->node_start_pfn - start;
5358 /* ia64 gets its own node_mem_map, before this, without bootmem */
5359 if (!pgdat->node_mem_map) {
5360 unsigned long size, end;
5364 * The zone's endpoints aren't required to be MAX_ORDER
5365 * aligned but the node_mem_map endpoints must be in order
5366 * for the buddy allocator to function correctly.
5368 end = pgdat_end_pfn(pgdat);
5369 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5370 size = (end - start) * sizeof(struct page);
5371 map = alloc_remap(pgdat->node_id, size);
5373 map = memblock_virt_alloc_node_nopanic(size,
5375 pgdat->node_mem_map = map + offset;
5377 #ifndef CONFIG_NEED_MULTIPLE_NODES
5379 * With no DISCONTIG, the global mem_map is just set as node 0's
5381 if (pgdat == NODE_DATA(0)) {
5382 mem_map = NODE_DATA(0)->node_mem_map;
5383 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5384 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5386 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5389 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5392 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5393 unsigned long node_start_pfn, unsigned long *zholes_size)
5395 pg_data_t *pgdat = NODE_DATA(nid);
5396 unsigned long start_pfn = 0;
5397 unsigned long end_pfn = 0;
5399 /* pg_data_t should be reset to zero when it's allocated */
5400 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5402 pgdat->node_id = nid;
5403 pgdat->node_start_pfn = node_start_pfn;
5404 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5405 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5406 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5407 (u64)start_pfn << PAGE_SHIFT,
5408 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5410 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5411 zones_size, zholes_size);
5413 alloc_node_mem_map(pgdat);
5414 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5415 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5416 nid, (unsigned long)pgdat,
5417 (unsigned long)pgdat->node_mem_map);
5420 reset_deferred_meminit(pgdat);
5421 free_area_init_core(pgdat);
5424 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5426 #if MAX_NUMNODES > 1
5428 * Figure out the number of possible node ids.
5430 void __init setup_nr_node_ids(void)
5432 unsigned int highest;
5434 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5435 nr_node_ids = highest + 1;
5440 * node_map_pfn_alignment - determine the maximum internode alignment
5442 * This function should be called after node map is populated and sorted.
5443 * It calculates the maximum power of two alignment which can distinguish
5446 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5447 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5448 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5449 * shifted, 1GiB is enough and this function will indicate so.
5451 * This is used to test whether pfn -> nid mapping of the chosen memory
5452 * model has fine enough granularity to avoid incorrect mapping for the
5453 * populated node map.
5455 * Returns the determined alignment in pfn's. 0 if there is no alignment
5456 * requirement (single node).
5458 unsigned long __init node_map_pfn_alignment(void)
5460 unsigned long accl_mask = 0, last_end = 0;
5461 unsigned long start, end, mask;
5465 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5466 if (!start || last_nid < 0 || last_nid == nid) {
5473 * Start with a mask granular enough to pin-point to the
5474 * start pfn and tick off bits one-by-one until it becomes
5475 * too coarse to separate the current node from the last.
5477 mask = ~((1 << __ffs(start)) - 1);
5478 while (mask && last_end <= (start & (mask << 1)))
5481 /* accumulate all internode masks */
5485 /* convert mask to number of pages */
5486 return ~accl_mask + 1;
5489 /* Find the lowest pfn for a node */
5490 static unsigned long __init find_min_pfn_for_node(int nid)
5492 unsigned long min_pfn = ULONG_MAX;
5493 unsigned long start_pfn;
5496 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5497 min_pfn = min(min_pfn, start_pfn);
5499 if (min_pfn == ULONG_MAX) {
5501 "Could not find start_pfn for node %d\n", nid);
5509 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5511 * It returns the minimum PFN based on information provided via
5512 * memblock_set_node().
5514 unsigned long __init find_min_pfn_with_active_regions(void)
5516 return find_min_pfn_for_node(MAX_NUMNODES);
5520 * early_calculate_totalpages()
5521 * Sum pages in active regions for movable zone.
5522 * Populate N_MEMORY for calculating usable_nodes.
5524 static unsigned long __init early_calculate_totalpages(void)
5526 unsigned long totalpages = 0;
5527 unsigned long start_pfn, end_pfn;
5530 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5531 unsigned long pages = end_pfn - start_pfn;
5533 totalpages += pages;
5535 node_set_state(nid, N_MEMORY);
5541 * Find the PFN the Movable zone begins in each node. Kernel memory
5542 * is spread evenly between nodes as long as the nodes have enough
5543 * memory. When they don't, some nodes will have more kernelcore than
5546 static void __init find_zone_movable_pfns_for_nodes(void)
5549 unsigned long usable_startpfn;
5550 unsigned long kernelcore_node, kernelcore_remaining;
5551 /* save the state before borrow the nodemask */
5552 nodemask_t saved_node_state = node_states[N_MEMORY];
5553 unsigned long totalpages = early_calculate_totalpages();
5554 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5555 struct memblock_region *r;
5557 /* Need to find movable_zone earlier when movable_node is specified. */
5558 find_usable_zone_for_movable();
5561 * If movable_node is specified, ignore kernelcore and movablecore
5564 if (movable_node_is_enabled()) {
5565 for_each_memblock(memory, r) {
5566 if (!memblock_is_hotpluggable(r))
5571 usable_startpfn = PFN_DOWN(r->base);
5572 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5573 min(usable_startpfn, zone_movable_pfn[nid]) :
5581 * If movablecore=nn[KMG] was specified, calculate what size of
5582 * kernelcore that corresponds so that memory usable for
5583 * any allocation type is evenly spread. If both kernelcore
5584 * and movablecore are specified, then the value of kernelcore
5585 * will be used for required_kernelcore if it's greater than
5586 * what movablecore would have allowed.
5588 if (required_movablecore) {
5589 unsigned long corepages;
5592 * Round-up so that ZONE_MOVABLE is at least as large as what
5593 * was requested by the user
5595 required_movablecore =
5596 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5597 required_movablecore = min(totalpages, required_movablecore);
5598 corepages = totalpages - required_movablecore;
5600 required_kernelcore = max(required_kernelcore, corepages);
5604 * If kernelcore was not specified or kernelcore size is larger
5605 * than totalpages, there is no ZONE_MOVABLE.
5607 if (!required_kernelcore || required_kernelcore >= totalpages)
5610 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5611 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5614 /* Spread kernelcore memory as evenly as possible throughout nodes */
5615 kernelcore_node = required_kernelcore / usable_nodes;
5616 for_each_node_state(nid, N_MEMORY) {
5617 unsigned long start_pfn, end_pfn;
5620 * Recalculate kernelcore_node if the division per node
5621 * now exceeds what is necessary to satisfy the requested
5622 * amount of memory for the kernel
5624 if (required_kernelcore < kernelcore_node)
5625 kernelcore_node = required_kernelcore / usable_nodes;
5628 * As the map is walked, we track how much memory is usable
5629 * by the kernel using kernelcore_remaining. When it is
5630 * 0, the rest of the node is usable by ZONE_MOVABLE
5632 kernelcore_remaining = kernelcore_node;
5634 /* Go through each range of PFNs within this node */
5635 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5636 unsigned long size_pages;
5638 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5639 if (start_pfn >= end_pfn)
5642 /* Account for what is only usable for kernelcore */
5643 if (start_pfn < usable_startpfn) {
5644 unsigned long kernel_pages;
5645 kernel_pages = min(end_pfn, usable_startpfn)
5648 kernelcore_remaining -= min(kernel_pages,
5649 kernelcore_remaining);
5650 required_kernelcore -= min(kernel_pages,
5651 required_kernelcore);
5653 /* Continue if range is now fully accounted */
5654 if (end_pfn <= usable_startpfn) {
5657 * Push zone_movable_pfn to the end so
5658 * that if we have to rebalance
5659 * kernelcore across nodes, we will
5660 * not double account here
5662 zone_movable_pfn[nid] = end_pfn;
5665 start_pfn = usable_startpfn;
5669 * The usable PFN range for ZONE_MOVABLE is from
5670 * start_pfn->end_pfn. Calculate size_pages as the
5671 * number of pages used as kernelcore
5673 size_pages = end_pfn - start_pfn;
5674 if (size_pages > kernelcore_remaining)
5675 size_pages = kernelcore_remaining;
5676 zone_movable_pfn[nid] = start_pfn + size_pages;
5679 * Some kernelcore has been met, update counts and
5680 * break if the kernelcore for this node has been
5683 required_kernelcore -= min(required_kernelcore,
5685 kernelcore_remaining -= size_pages;
5686 if (!kernelcore_remaining)
5692 * If there is still required_kernelcore, we do another pass with one
5693 * less node in the count. This will push zone_movable_pfn[nid] further
5694 * along on the nodes that still have memory until kernelcore is
5698 if (usable_nodes && required_kernelcore > usable_nodes)
5702 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5703 for (nid = 0; nid < MAX_NUMNODES; nid++)
5704 zone_movable_pfn[nid] =
5705 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5708 /* restore the node_state */
5709 node_states[N_MEMORY] = saved_node_state;
5712 /* Any regular or high memory on that node ? */
5713 static void check_for_memory(pg_data_t *pgdat, int nid)
5715 enum zone_type zone_type;
5717 if (N_MEMORY == N_NORMAL_MEMORY)
5720 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5721 struct zone *zone = &pgdat->node_zones[zone_type];
5722 if (populated_zone(zone)) {
5723 node_set_state(nid, N_HIGH_MEMORY);
5724 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5725 zone_type <= ZONE_NORMAL)
5726 node_set_state(nid, N_NORMAL_MEMORY);
5733 * free_area_init_nodes - Initialise all pg_data_t and zone data
5734 * @max_zone_pfn: an array of max PFNs for each zone
5736 * This will call free_area_init_node() for each active node in the system.
5737 * Using the page ranges provided by memblock_set_node(), the size of each
5738 * zone in each node and their holes is calculated. If the maximum PFN
5739 * between two adjacent zones match, it is assumed that the zone is empty.
5740 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5741 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5742 * starts where the previous one ended. For example, ZONE_DMA32 starts
5743 * at arch_max_dma_pfn.
5745 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5747 unsigned long start_pfn, end_pfn;
5750 /* Record where the zone boundaries are */
5751 memset(arch_zone_lowest_possible_pfn, 0,
5752 sizeof(arch_zone_lowest_possible_pfn));
5753 memset(arch_zone_highest_possible_pfn, 0,
5754 sizeof(arch_zone_highest_possible_pfn));
5756 start_pfn = find_min_pfn_with_active_regions();
5758 for (i = 0; i < MAX_NR_ZONES; i++) {
5759 if (i == ZONE_MOVABLE)
5762 end_pfn = max(max_zone_pfn[i], start_pfn);
5763 arch_zone_lowest_possible_pfn[i] = start_pfn;
5764 arch_zone_highest_possible_pfn[i] = end_pfn;
5766 start_pfn = end_pfn;
5768 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5769 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5771 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5772 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5773 find_zone_movable_pfns_for_nodes();
5775 /* Print out the zone ranges */
5776 pr_info("Zone ranges:\n");
5777 for (i = 0; i < MAX_NR_ZONES; i++) {
5778 if (i == ZONE_MOVABLE)
5780 pr_info(" %-8s ", zone_names[i]);
5781 if (arch_zone_lowest_possible_pfn[i] ==
5782 arch_zone_highest_possible_pfn[i])
5785 pr_cont("[mem %#018Lx-%#018Lx]\n",
5786 (u64)arch_zone_lowest_possible_pfn[i]
5788 ((u64)arch_zone_highest_possible_pfn[i]
5789 << PAGE_SHIFT) - 1);
5792 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5793 pr_info("Movable zone start for each node\n");
5794 for (i = 0; i < MAX_NUMNODES; i++) {
5795 if (zone_movable_pfn[i])
5796 pr_info(" Node %d: %#018Lx\n", i,
5797 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5800 /* Print out the early node map */
5801 pr_info("Early memory node ranges\n");
5802 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5803 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5804 (u64)start_pfn << PAGE_SHIFT,
5805 ((u64)end_pfn << PAGE_SHIFT) - 1);
5807 /* Initialise every node */
5808 mminit_verify_pageflags_layout();
5809 setup_nr_node_ids();
5810 for_each_online_node(nid) {
5811 pg_data_t *pgdat = NODE_DATA(nid);
5812 free_area_init_node(nid, NULL,
5813 find_min_pfn_for_node(nid), NULL);
5815 /* Any memory on that node */
5816 if (pgdat->node_present_pages)
5817 node_set_state(nid, N_MEMORY);
5818 check_for_memory(pgdat, nid);
5822 static int __init cmdline_parse_core(char *p, unsigned long *core)
5824 unsigned long long coremem;
5828 coremem = memparse(p, &p);
5829 *core = coremem >> PAGE_SHIFT;
5831 /* Paranoid check that UL is enough for the coremem value */
5832 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5838 * kernelcore=size sets the amount of memory for use for allocations that
5839 * cannot be reclaimed or migrated.
5841 static int __init cmdline_parse_kernelcore(char *p)
5843 return cmdline_parse_core(p, &required_kernelcore);
5847 * movablecore=size sets the amount of memory for use for allocations that
5848 * can be reclaimed or migrated.
5850 static int __init cmdline_parse_movablecore(char *p)
5852 return cmdline_parse_core(p, &required_movablecore);
5855 early_param("kernelcore", cmdline_parse_kernelcore);
5856 early_param("movablecore", cmdline_parse_movablecore);
5858 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5860 void adjust_managed_page_count(struct page *page, long count)
5862 spin_lock(&managed_page_count_lock);
5863 page_zone(page)->managed_pages += count;
5864 totalram_pages += count;
5865 #ifdef CONFIG_HIGHMEM
5866 if (PageHighMem(page))
5867 totalhigh_pages += count;
5869 spin_unlock(&managed_page_count_lock);
5871 EXPORT_SYMBOL(adjust_managed_page_count);
5873 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5876 unsigned long pages = 0;
5878 start = (void *)PAGE_ALIGN((unsigned long)start);
5879 end = (void *)((unsigned long)end & PAGE_MASK);
5880 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5881 if ((unsigned int)poison <= 0xFF)
5882 memset(pos, poison, PAGE_SIZE);
5883 free_reserved_page(virt_to_page(pos));
5887 pr_info("Freeing %s memory: %ldK\n",
5888 s, pages << (PAGE_SHIFT - 10));
5892 EXPORT_SYMBOL(free_reserved_area);
5894 #ifdef CONFIG_HIGHMEM
5895 void free_highmem_page(struct page *page)
5897 __free_reserved_page(page);
5899 page_zone(page)->managed_pages++;
5905 void __init mem_init_print_info(const char *str)
5907 unsigned long physpages, codesize, datasize, rosize, bss_size;
5908 unsigned long init_code_size, init_data_size;
5910 physpages = get_num_physpages();
5911 codesize = _etext - _stext;
5912 datasize = _edata - _sdata;
5913 rosize = __end_rodata - __start_rodata;
5914 bss_size = __bss_stop - __bss_start;
5915 init_data_size = __init_end - __init_begin;
5916 init_code_size = _einittext - _sinittext;
5919 * Detect special cases and adjust section sizes accordingly:
5920 * 1) .init.* may be embedded into .data sections
5921 * 2) .init.text.* may be out of [__init_begin, __init_end],
5922 * please refer to arch/tile/kernel/vmlinux.lds.S.
5923 * 3) .rodata.* may be embedded into .text or .data sections.
5925 #define adj_init_size(start, end, size, pos, adj) \
5927 if (start <= pos && pos < end && size > adj) \
5931 adj_init_size(__init_begin, __init_end, init_data_size,
5932 _sinittext, init_code_size);
5933 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5934 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5935 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5936 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5938 #undef adj_init_size
5940 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
5941 #ifdef CONFIG_HIGHMEM
5945 nr_free_pages() << (PAGE_SHIFT - 10),
5946 physpages << (PAGE_SHIFT - 10),
5947 codesize >> 10, datasize >> 10, rosize >> 10,
5948 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5949 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
5950 totalcma_pages << (PAGE_SHIFT - 10),
5951 #ifdef CONFIG_HIGHMEM
5952 totalhigh_pages << (PAGE_SHIFT - 10),
5954 str ? ", " : "", str ? str : "");
5958 * set_dma_reserve - set the specified number of pages reserved in the first zone
5959 * @new_dma_reserve: The number of pages to mark reserved
5961 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5962 * In the DMA zone, a significant percentage may be consumed by kernel image
5963 * and other unfreeable allocations which can skew the watermarks badly. This
5964 * function may optionally be used to account for unfreeable pages in the
5965 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5966 * smaller per-cpu batchsize.
5968 void __init set_dma_reserve(unsigned long new_dma_reserve)
5970 dma_reserve = new_dma_reserve;
5973 void __init free_area_init(unsigned long *zones_size)
5975 free_area_init_node(0, zones_size,
5976 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5979 static int page_alloc_cpu_notify(struct notifier_block *self,
5980 unsigned long action, void *hcpu)
5982 int cpu = (unsigned long)hcpu;
5984 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5985 lru_add_drain_cpu(cpu);
5989 * Spill the event counters of the dead processor
5990 * into the current processors event counters.
5991 * This artificially elevates the count of the current
5994 vm_events_fold_cpu(cpu);
5997 * Zero the differential counters of the dead processor
5998 * so that the vm statistics are consistent.
6000 * This is only okay since the processor is dead and cannot
6001 * race with what we are doing.
6003 cpu_vm_stats_fold(cpu);
6008 void __init page_alloc_init(void)
6010 hotcpu_notifier(page_alloc_cpu_notify, 0);
6014 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6015 * or min_free_kbytes changes.
6017 static void calculate_totalreserve_pages(void)
6019 struct pglist_data *pgdat;
6020 unsigned long reserve_pages = 0;
6021 enum zone_type i, j;
6023 for_each_online_pgdat(pgdat) {
6024 for (i = 0; i < MAX_NR_ZONES; i++) {
6025 struct zone *zone = pgdat->node_zones + i;
6028 /* Find valid and maximum lowmem_reserve in the zone */
6029 for (j = i; j < MAX_NR_ZONES; j++) {
6030 if (zone->lowmem_reserve[j] > max)
6031 max = zone->lowmem_reserve[j];
6034 /* we treat the high watermark as reserved pages. */
6035 max += high_wmark_pages(zone);
6037 if (max > zone->managed_pages)
6038 max = zone->managed_pages;
6039 reserve_pages += max;
6041 * Lowmem reserves are not available to
6042 * GFP_HIGHUSER page cache allocations and
6043 * kswapd tries to balance zones to their high
6044 * watermark. As a result, neither should be
6045 * regarded as dirtyable memory, to prevent a
6046 * situation where reclaim has to clean pages
6047 * in order to balance the zones.
6049 zone->dirty_balance_reserve = max;
6052 dirty_balance_reserve = reserve_pages;
6053 totalreserve_pages = reserve_pages;
6057 * setup_per_zone_lowmem_reserve - called whenever
6058 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6059 * has a correct pages reserved value, so an adequate number of
6060 * pages are left in the zone after a successful __alloc_pages().
6062 static void setup_per_zone_lowmem_reserve(void)
6064 struct pglist_data *pgdat;
6065 enum zone_type j, idx;
6067 for_each_online_pgdat(pgdat) {
6068 for (j = 0; j < MAX_NR_ZONES; j++) {
6069 struct zone *zone = pgdat->node_zones + j;
6070 unsigned long managed_pages = zone->managed_pages;
6072 zone->lowmem_reserve[j] = 0;
6076 struct zone *lower_zone;
6080 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6081 sysctl_lowmem_reserve_ratio[idx] = 1;
6083 lower_zone = pgdat->node_zones + idx;
6084 lower_zone->lowmem_reserve[j] = managed_pages /
6085 sysctl_lowmem_reserve_ratio[idx];
6086 managed_pages += lower_zone->managed_pages;
6091 /* update totalreserve_pages */
6092 calculate_totalreserve_pages();
6095 static void __setup_per_zone_wmarks(void)
6097 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6098 unsigned long pages_low = extra_free_kbytes >> (PAGE_SHIFT - 10);
6099 unsigned long lowmem_pages = 0;
6101 unsigned long flags;
6103 /* Calculate total number of !ZONE_HIGHMEM pages */
6104 for_each_zone(zone) {
6105 if (!is_highmem(zone))
6106 lowmem_pages += zone->managed_pages;
6109 for_each_zone(zone) {
6112 spin_lock_irqsave(&zone->lock, flags);
6113 min = (u64)pages_min * zone->managed_pages;
6114 do_div(min, lowmem_pages);
6115 low = (u64)pages_low * zone->managed_pages;
6116 do_div(low, vm_total_pages);
6118 if (is_highmem(zone)) {
6120 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6121 * need highmem pages, so cap pages_min to a small
6124 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6125 * deltas control asynch page reclaim, and so should
6126 * not be capped for highmem.
6128 unsigned long min_pages;
6130 min_pages = zone->managed_pages / 1024;
6131 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6132 zone->watermark[WMARK_MIN] = min_pages;
6135 * If it's a lowmem zone, reserve a number of pages
6136 * proportionate to the zone's size.
6138 zone->watermark[WMARK_MIN] = min;
6141 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) +
6143 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) +
6146 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6147 high_wmark_pages(zone) - low_wmark_pages(zone) -
6148 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6150 spin_unlock_irqrestore(&zone->lock, flags);
6153 /* update totalreserve_pages */
6154 calculate_totalreserve_pages();
6158 * setup_per_zone_wmarks - called when min_free_kbytes changes
6159 * or when memory is hot-{added|removed}
6161 * Ensures that the watermark[min,low,high] values for each zone are set
6162 * correctly with respect to min_free_kbytes.
6164 void setup_per_zone_wmarks(void)
6166 mutex_lock(&zonelists_mutex);
6167 __setup_per_zone_wmarks();
6168 mutex_unlock(&zonelists_mutex);
6172 * The inactive anon list should be small enough that the VM never has to
6173 * do too much work, but large enough that each inactive page has a chance
6174 * to be referenced again before it is swapped out.
6176 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6177 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6178 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6179 * the anonymous pages are kept on the inactive list.
6182 * memory ratio inactive anon
6183 * -------------------------------------
6192 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6194 unsigned int gb, ratio;
6196 /* Zone size in gigabytes */
6197 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6199 ratio = int_sqrt(10 * gb);
6203 zone->inactive_ratio = ratio;
6206 static void __meminit setup_per_zone_inactive_ratio(void)
6211 calculate_zone_inactive_ratio(zone);
6215 * Initialise min_free_kbytes.
6217 * For small machines we want it small (128k min). For large machines
6218 * we want it large (64MB max). But it is not linear, because network
6219 * bandwidth does not increase linearly with machine size. We use
6221 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6222 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6238 int __meminit init_per_zone_wmark_min(void)
6240 unsigned long lowmem_kbytes;
6241 int new_min_free_kbytes;
6243 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6244 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6246 if (new_min_free_kbytes > user_min_free_kbytes) {
6247 min_free_kbytes = new_min_free_kbytes;
6248 if (min_free_kbytes < 128)
6249 min_free_kbytes = 128;
6250 if (min_free_kbytes > 65536)
6251 min_free_kbytes = 65536;
6253 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6254 new_min_free_kbytes, user_min_free_kbytes);
6256 setup_per_zone_wmarks();
6257 refresh_zone_stat_thresholds();
6258 setup_per_zone_lowmem_reserve();
6259 setup_per_zone_inactive_ratio();
6262 core_initcall(init_per_zone_wmark_min)
6265 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6266 * that we can call two helper functions whenever min_free_kbytes
6267 * or extra_free_kbytes changes.
6269 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6270 void __user *buffer, size_t *length, loff_t *ppos)
6274 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6279 user_min_free_kbytes = min_free_kbytes;
6280 setup_per_zone_wmarks();
6286 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6287 void __user *buffer, size_t *length, loff_t *ppos)
6292 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6297 zone->min_unmapped_pages = (zone->managed_pages *
6298 sysctl_min_unmapped_ratio) / 100;
6302 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6303 void __user *buffer, size_t *length, loff_t *ppos)
6308 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6313 zone->min_slab_pages = (zone->managed_pages *
6314 sysctl_min_slab_ratio) / 100;
6320 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6321 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6322 * whenever sysctl_lowmem_reserve_ratio changes.
6324 * The reserve ratio obviously has absolutely no relation with the
6325 * minimum watermarks. The lowmem reserve ratio can only make sense
6326 * if in function of the boot time zone sizes.
6328 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6329 void __user *buffer, size_t *length, loff_t *ppos)
6331 proc_dointvec_minmax(table, write, buffer, length, ppos);
6332 setup_per_zone_lowmem_reserve();
6337 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6338 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6339 * pagelist can have before it gets flushed back to buddy allocator.
6341 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6342 void __user *buffer, size_t *length, loff_t *ppos)
6345 int old_percpu_pagelist_fraction;
6348 mutex_lock(&pcp_batch_high_lock);
6349 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6351 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6352 if (!write || ret < 0)
6355 /* Sanity checking to avoid pcp imbalance */
6356 if (percpu_pagelist_fraction &&
6357 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6358 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6364 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6367 for_each_populated_zone(zone) {
6370 for_each_possible_cpu(cpu)
6371 pageset_set_high_and_batch(zone,
6372 per_cpu_ptr(zone->pageset, cpu));
6375 mutex_unlock(&pcp_batch_high_lock);
6380 int hashdist = HASHDIST_DEFAULT;
6382 static int __init set_hashdist(char *str)
6386 hashdist = simple_strtoul(str, &str, 0);
6389 __setup("hashdist=", set_hashdist);
6393 * allocate a large system hash table from bootmem
6394 * - it is assumed that the hash table must contain an exact power-of-2
6395 * quantity of entries
6396 * - limit is the number of hash buckets, not the total allocation size
6398 void *__init alloc_large_system_hash(const char *tablename,
6399 unsigned long bucketsize,
6400 unsigned long numentries,
6403 unsigned int *_hash_shift,
6404 unsigned int *_hash_mask,
6405 unsigned long low_limit,
6406 unsigned long high_limit)
6408 unsigned long long max = high_limit;
6409 unsigned long log2qty, size;
6412 /* allow the kernel cmdline to have a say */
6414 /* round applicable memory size up to nearest megabyte */
6415 numentries = nr_kernel_pages;
6417 /* It isn't necessary when PAGE_SIZE >= 1MB */
6418 if (PAGE_SHIFT < 20)
6419 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6421 /* limit to 1 bucket per 2^scale bytes of low memory */
6422 if (scale > PAGE_SHIFT)
6423 numentries >>= (scale - PAGE_SHIFT);
6425 numentries <<= (PAGE_SHIFT - scale);
6427 /* Make sure we've got at least a 0-order allocation.. */
6428 if (unlikely(flags & HASH_SMALL)) {
6429 /* Makes no sense without HASH_EARLY */
6430 WARN_ON(!(flags & HASH_EARLY));
6431 if (!(numentries >> *_hash_shift)) {
6432 numentries = 1UL << *_hash_shift;
6433 BUG_ON(!numentries);
6435 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6436 numentries = PAGE_SIZE / bucketsize;
6438 numentries = roundup_pow_of_two(numentries);
6440 /* limit allocation size to 1/16 total memory by default */
6442 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6443 do_div(max, bucketsize);
6445 max = min(max, 0x80000000ULL);
6447 if (numentries < low_limit)
6448 numentries = low_limit;
6449 if (numentries > max)
6452 log2qty = ilog2(numentries);
6455 size = bucketsize << log2qty;
6456 if (flags & HASH_EARLY)
6457 table = memblock_virt_alloc_nopanic(size, 0);
6459 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6462 * If bucketsize is not a power-of-two, we may free
6463 * some pages at the end of hash table which
6464 * alloc_pages_exact() automatically does
6466 if (get_order(size) < MAX_ORDER) {
6467 table = alloc_pages_exact(size, GFP_ATOMIC);
6468 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6471 } while (!table && size > PAGE_SIZE && --log2qty);
6474 panic("Failed to allocate %s hash table\n", tablename);
6476 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6479 ilog2(size) - PAGE_SHIFT,
6483 *_hash_shift = log2qty;
6485 *_hash_mask = (1 << log2qty) - 1;
6490 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6491 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6494 #ifdef CONFIG_SPARSEMEM
6495 return __pfn_to_section(pfn)->pageblock_flags;
6497 return zone->pageblock_flags;
6498 #endif /* CONFIG_SPARSEMEM */
6501 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6503 #ifdef CONFIG_SPARSEMEM
6504 pfn &= (PAGES_PER_SECTION-1);
6505 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6507 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6508 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6509 #endif /* CONFIG_SPARSEMEM */
6513 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6514 * @page: The page within the block of interest
6515 * @pfn: The target page frame number
6516 * @end_bitidx: The last bit of interest to retrieve
6517 * @mask: mask of bits that the caller is interested in
6519 * Return: pageblock_bits flags
6521 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6522 unsigned long end_bitidx,
6526 unsigned long *bitmap;
6527 unsigned long bitidx, word_bitidx;
6530 zone = page_zone(page);
6531 bitmap = get_pageblock_bitmap(zone, pfn);
6532 bitidx = pfn_to_bitidx(zone, pfn);
6533 word_bitidx = bitidx / BITS_PER_LONG;
6534 bitidx &= (BITS_PER_LONG-1);
6536 word = bitmap[word_bitidx];
6537 bitidx += end_bitidx;
6538 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6542 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6543 * @page: The page within the block of interest
6544 * @flags: The flags to set
6545 * @pfn: The target page frame number
6546 * @end_bitidx: The last bit of interest
6547 * @mask: mask of bits that the caller is interested in
6549 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6551 unsigned long end_bitidx,
6555 unsigned long *bitmap;
6556 unsigned long bitidx, word_bitidx;
6557 unsigned long old_word, word;
6559 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6561 zone = page_zone(page);
6562 bitmap = get_pageblock_bitmap(zone, pfn);
6563 bitidx = pfn_to_bitidx(zone, pfn);
6564 word_bitidx = bitidx / BITS_PER_LONG;
6565 bitidx &= (BITS_PER_LONG-1);
6567 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6569 bitidx += end_bitidx;
6570 mask <<= (BITS_PER_LONG - bitidx - 1);
6571 flags <<= (BITS_PER_LONG - bitidx - 1);
6573 word = READ_ONCE(bitmap[word_bitidx]);
6575 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6576 if (word == old_word)
6583 * This function checks whether pageblock includes unmovable pages or not.
6584 * If @count is not zero, it is okay to include less @count unmovable pages
6586 * PageLRU check without isolation or lru_lock could race so that
6587 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6588 * expect this function should be exact.
6590 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6591 bool skip_hwpoisoned_pages)
6593 unsigned long pfn, iter, found;
6597 * For avoiding noise data, lru_add_drain_all() should be called
6598 * If ZONE_MOVABLE, the zone never contains unmovable pages
6600 if (zone_idx(zone) == ZONE_MOVABLE)
6602 mt = get_pageblock_migratetype(page);
6603 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6606 pfn = page_to_pfn(page);
6607 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6608 unsigned long check = pfn + iter;
6610 if (!pfn_valid_within(check))
6613 page = pfn_to_page(check);
6616 * Hugepages are not in LRU lists, but they're movable.
6617 * We need not scan over tail pages bacause we don't
6618 * handle each tail page individually in migration.
6620 if (PageHuge(page)) {
6621 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6626 * We can't use page_count without pin a page
6627 * because another CPU can free compound page.
6628 * This check already skips compound tails of THP
6629 * because their page->_count is zero at all time.
6631 if (!atomic_read(&page->_count)) {
6632 if (PageBuddy(page))
6633 iter += (1 << page_order(page)) - 1;
6638 * The HWPoisoned page may be not in buddy system, and
6639 * page_count() is not 0.
6641 if (skip_hwpoisoned_pages && PageHWPoison(page))
6647 * If there are RECLAIMABLE pages, we need to check
6648 * it. But now, memory offline itself doesn't call
6649 * shrink_node_slabs() and it still to be fixed.
6652 * If the page is not RAM, page_count()should be 0.
6653 * we don't need more check. This is an _used_ not-movable page.
6655 * The problematic thing here is PG_reserved pages. PG_reserved
6656 * is set to both of a memory hole page and a _used_ kernel
6665 bool is_pageblock_removable_nolock(struct page *page)
6671 * We have to be careful here because we are iterating over memory
6672 * sections which are not zone aware so we might end up outside of
6673 * the zone but still within the section.
6674 * We have to take care about the node as well. If the node is offline
6675 * its NODE_DATA will be NULL - see page_zone.
6677 if (!node_online(page_to_nid(page)))
6680 zone = page_zone(page);
6681 pfn = page_to_pfn(page);
6682 if (!zone_spans_pfn(zone, pfn))
6685 return !has_unmovable_pages(zone, page, 0, true);
6690 static unsigned long pfn_max_align_down(unsigned long pfn)
6692 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6693 pageblock_nr_pages) - 1);
6696 static unsigned long pfn_max_align_up(unsigned long pfn)
6698 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6699 pageblock_nr_pages));
6702 /* [start, end) must belong to a single zone. */
6703 static int __alloc_contig_migrate_range(struct compact_control *cc,
6704 unsigned long start, unsigned long end)
6706 /* This function is based on compact_zone() from compaction.c. */
6707 unsigned long nr_reclaimed;
6708 unsigned long pfn = start;
6709 unsigned int tries = 0;
6714 while (pfn < end || !list_empty(&cc->migratepages)) {
6715 if (fatal_signal_pending(current)) {
6720 if (list_empty(&cc->migratepages)) {
6721 cc->nr_migratepages = 0;
6722 pfn = isolate_migratepages_range(cc, pfn, end);
6728 } else if (++tries == 5) {
6729 ret = ret < 0 ? ret : -EBUSY;
6733 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6735 cc->nr_migratepages -= nr_reclaimed;
6737 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6738 NULL, 0, cc->mode, MR_CMA);
6741 putback_movable_pages(&cc->migratepages);
6748 * alloc_contig_range() -- tries to allocate given range of pages
6749 * @start: start PFN to allocate
6750 * @end: one-past-the-last PFN to allocate
6751 * @migratetype: migratetype of the underlaying pageblocks (either
6752 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6753 * in range must have the same migratetype and it must
6754 * be either of the two.
6756 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6757 * aligned, however it's the caller's responsibility to guarantee that
6758 * we are the only thread that changes migrate type of pageblocks the
6761 * The PFN range must belong to a single zone.
6763 * Returns zero on success or negative error code. On success all
6764 * pages which PFN is in [start, end) are allocated for the caller and
6765 * need to be freed with free_contig_range().
6767 int alloc_contig_range(unsigned long start, unsigned long end,
6768 unsigned migratetype)
6770 unsigned long outer_start, outer_end;
6774 struct compact_control cc = {
6775 .nr_migratepages = 0,
6777 .zone = page_zone(pfn_to_page(start)),
6778 .mode = MIGRATE_SYNC,
6779 .ignore_skip_hint = true,
6781 INIT_LIST_HEAD(&cc.migratepages);
6784 * What we do here is we mark all pageblocks in range as
6785 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6786 * have different sizes, and due to the way page allocator
6787 * work, we align the range to biggest of the two pages so
6788 * that page allocator won't try to merge buddies from
6789 * different pageblocks and change MIGRATE_ISOLATE to some
6790 * other migration type.
6792 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6793 * migrate the pages from an unaligned range (ie. pages that
6794 * we are interested in). This will put all the pages in
6795 * range back to page allocator as MIGRATE_ISOLATE.
6797 * When this is done, we take the pages in range from page
6798 * allocator removing them from the buddy system. This way
6799 * page allocator will never consider using them.
6801 * This lets us mark the pageblocks back as
6802 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6803 * aligned range but not in the unaligned, original range are
6804 * put back to page allocator so that buddy can use them.
6807 ret = start_isolate_page_range(pfn_max_align_down(start),
6808 pfn_max_align_up(end), migratetype,
6813 ret = __alloc_contig_migrate_range(&cc, start, end);
6818 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6819 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6820 * more, all pages in [start, end) are free in page allocator.
6821 * What we are going to do is to allocate all pages from
6822 * [start, end) (that is remove them from page allocator).
6824 * The only problem is that pages at the beginning and at the
6825 * end of interesting range may be not aligned with pages that
6826 * page allocator holds, ie. they can be part of higher order
6827 * pages. Because of this, we reserve the bigger range and
6828 * once this is done free the pages we are not interested in.
6830 * We don't have to hold zone->lock here because the pages are
6831 * isolated thus they won't get removed from buddy.
6834 lru_add_drain_all();
6835 drain_all_pages(cc.zone);
6838 outer_start = start;
6839 while (!PageBuddy(pfn_to_page(outer_start))) {
6840 if (++order >= MAX_ORDER) {
6844 outer_start &= ~0UL << order;
6847 /* Make sure the range is really isolated. */
6848 if (test_pages_isolated(outer_start, end, false)) {
6849 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
6850 __func__, outer_start, end);
6855 /* Grab isolated pages from freelists. */
6856 outer_end = isolate_freepages_range(&cc, outer_start, end);
6862 /* Free head and tail (if any) */
6863 if (start != outer_start)
6864 free_contig_range(outer_start, start - outer_start);
6865 if (end != outer_end)
6866 free_contig_range(end, outer_end - end);
6869 undo_isolate_page_range(pfn_max_align_down(start),
6870 pfn_max_align_up(end), migratetype);
6874 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6876 unsigned int count = 0;
6878 for (; nr_pages--; pfn++) {
6879 struct page *page = pfn_to_page(pfn);
6881 count += page_count(page) != 1;
6884 WARN(count != 0, "%d pages are still in use!\n", count);
6888 #ifdef CONFIG_MEMORY_HOTPLUG
6890 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6891 * page high values need to be recalulated.
6893 void __meminit zone_pcp_update(struct zone *zone)
6896 mutex_lock(&pcp_batch_high_lock);
6897 for_each_possible_cpu(cpu)
6898 pageset_set_high_and_batch(zone,
6899 per_cpu_ptr(zone->pageset, cpu));
6900 mutex_unlock(&pcp_batch_high_lock);
6904 void zone_pcp_reset(struct zone *zone)
6906 unsigned long flags;
6908 struct per_cpu_pageset *pset;
6910 /* avoid races with drain_pages() */
6911 local_irq_save(flags);
6912 if (zone->pageset != &boot_pageset) {
6913 for_each_online_cpu(cpu) {
6914 pset = per_cpu_ptr(zone->pageset, cpu);
6915 drain_zonestat(zone, pset);
6917 free_percpu(zone->pageset);
6918 zone->pageset = &boot_pageset;
6920 local_irq_restore(flags);
6923 #ifdef CONFIG_MEMORY_HOTREMOVE
6925 * All pages in the range must be isolated before calling this.
6928 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6932 unsigned int order, i;
6934 unsigned long flags;
6935 /* find the first valid pfn */
6936 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6941 zone = page_zone(pfn_to_page(pfn));
6942 spin_lock_irqsave(&zone->lock, flags);
6944 while (pfn < end_pfn) {
6945 if (!pfn_valid(pfn)) {
6949 page = pfn_to_page(pfn);
6951 * The HWPoisoned page may be not in buddy system, and
6952 * page_count() is not 0.
6954 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6956 SetPageReserved(page);
6960 BUG_ON(page_count(page));
6961 BUG_ON(!PageBuddy(page));
6962 order = page_order(page);
6963 #ifdef CONFIG_DEBUG_VM
6964 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6965 pfn, 1 << order, end_pfn);
6967 list_del(&page->lru);
6968 rmv_page_order(page);
6969 zone->free_area[order].nr_free--;
6970 for (i = 0; i < (1 << order); i++)
6971 SetPageReserved((page+i));
6972 pfn += (1 << order);
6974 spin_unlock_irqrestore(&zone->lock, flags);
6978 #ifdef CONFIG_MEMORY_FAILURE
6979 bool is_free_buddy_page(struct page *page)
6981 struct zone *zone = page_zone(page);
6982 unsigned long pfn = page_to_pfn(page);
6983 unsigned long flags;
6986 spin_lock_irqsave(&zone->lock, flags);
6987 for (order = 0; order < MAX_ORDER; order++) {
6988 struct page *page_head = page - (pfn & ((1 << order) - 1));
6990 if (PageBuddy(page_head) && page_order(page_head) >= order)
6993 spin_unlock_irqrestore(&zone->lock, flags);
6995 return order < MAX_ORDER;