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 compound_page_dtor * const compound_page_dtors[] = {
235 #ifdef CONFIG_HUGETLB_PAGE
238 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
243 int min_free_kbytes = 1024;
244 int user_min_free_kbytes = -1;
246 static unsigned long __meminitdata nr_kernel_pages;
247 static unsigned long __meminitdata nr_all_pages;
248 static unsigned long __meminitdata dma_reserve;
250 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
251 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
252 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
253 static unsigned long __initdata required_kernelcore;
254 static unsigned long __initdata required_movablecore;
255 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
257 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
259 EXPORT_SYMBOL(movable_zone);
260 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
263 int nr_node_ids __read_mostly = MAX_NUMNODES;
264 int nr_online_nodes __read_mostly = 1;
265 EXPORT_SYMBOL(nr_node_ids);
266 EXPORT_SYMBOL(nr_online_nodes);
269 int page_group_by_mobility_disabled __read_mostly;
271 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
272 static inline void reset_deferred_meminit(pg_data_t *pgdat)
274 pgdat->first_deferred_pfn = ULONG_MAX;
277 /* Returns true if the struct page for the pfn is uninitialised */
278 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
280 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
286 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
288 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
295 * Returns false when the remaining initialisation should be deferred until
296 * later in the boot cycle when it can be parallelised.
298 static inline bool update_defer_init(pg_data_t *pgdat,
299 unsigned long pfn, unsigned long zone_end,
300 unsigned long *nr_initialised)
302 /* Always populate low zones for address-contrained allocations */
303 if (zone_end < pgdat_end_pfn(pgdat))
306 /* Initialise at least 2G of the highest zone */
308 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
309 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
310 pgdat->first_deferred_pfn = pfn;
317 static inline void reset_deferred_meminit(pg_data_t *pgdat)
321 static inline bool early_page_uninitialised(unsigned long pfn)
326 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
331 static inline bool update_defer_init(pg_data_t *pgdat,
332 unsigned long pfn, unsigned long zone_end,
333 unsigned long *nr_initialised)
340 void set_pageblock_migratetype(struct page *page, int migratetype)
342 if (unlikely(page_group_by_mobility_disabled &&
343 migratetype < MIGRATE_PCPTYPES))
344 migratetype = MIGRATE_UNMOVABLE;
346 set_pageblock_flags_group(page, (unsigned long)migratetype,
347 PB_migrate, PB_migrate_end);
350 #ifdef CONFIG_DEBUG_VM
351 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
355 unsigned long pfn = page_to_pfn(page);
356 unsigned long sp, start_pfn;
359 seq = zone_span_seqbegin(zone);
360 start_pfn = zone->zone_start_pfn;
361 sp = zone->spanned_pages;
362 if (!zone_spans_pfn(zone, pfn))
364 } while (zone_span_seqretry(zone, seq));
367 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
368 pfn, zone_to_nid(zone), zone->name,
369 start_pfn, start_pfn + sp);
374 static int page_is_consistent(struct zone *zone, struct page *page)
376 if (!pfn_valid_within(page_to_pfn(page)))
378 if (zone != page_zone(page))
384 * Temporary debugging check for pages not lying within a given zone.
386 static int bad_range(struct zone *zone, struct page *page)
388 if (page_outside_zone_boundaries(zone, page))
390 if (!page_is_consistent(zone, page))
396 static inline int bad_range(struct zone *zone, struct page *page)
402 static void bad_page(struct page *page, const char *reason,
403 unsigned long bad_flags)
405 static unsigned long resume;
406 static unsigned long nr_shown;
407 static unsigned long nr_unshown;
409 /* Don't complain about poisoned pages */
410 if (PageHWPoison(page)) {
411 page_mapcount_reset(page); /* remove PageBuddy */
416 * Allow a burst of 60 reports, then keep quiet for that minute;
417 * or allow a steady drip of one report per second.
419 if (nr_shown == 60) {
420 if (time_before(jiffies, resume)) {
426 "BUG: Bad page state: %lu messages suppressed\n",
433 resume = jiffies + 60 * HZ;
435 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
436 current->comm, page_to_pfn(page));
437 dump_page_badflags(page, reason, bad_flags);
442 /* Leave bad fields for debug, except PageBuddy could make trouble */
443 page_mapcount_reset(page); /* remove PageBuddy */
444 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
448 * Higher-order pages are called "compound pages". They are structured thusly:
450 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
452 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
453 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
455 * The first tail page's ->compound_dtor holds the offset in array of compound
456 * page destructors. See compound_page_dtors.
458 * The first tail page's ->compound_order holds the order of allocation.
459 * This usage means that zero-order pages may not be compound.
462 void free_compound_page(struct page *page)
464 __free_pages_ok(page, compound_order(page));
467 void prep_compound_page(struct page *page, unsigned int order)
470 int nr_pages = 1 << order;
472 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
473 set_compound_order(page, order);
475 for (i = 1; i < nr_pages; i++) {
476 struct page *p = page + i;
477 set_page_count(p, 0);
478 p->mapping = TAIL_MAPPING;
479 set_compound_head(p, page);
481 atomic_set(compound_mapcount_ptr(page), -1);
484 #ifdef CONFIG_DEBUG_PAGEALLOC
485 unsigned int _debug_guardpage_minorder;
486 bool _debug_pagealloc_enabled __read_mostly;
487 bool _debug_guardpage_enabled __read_mostly;
489 static int __init early_debug_pagealloc(char *buf)
494 if (strcmp(buf, "on") == 0)
495 _debug_pagealloc_enabled = true;
499 early_param("debug_pagealloc", early_debug_pagealloc);
501 static bool need_debug_guardpage(void)
503 /* If we don't use debug_pagealloc, we don't need guard page */
504 if (!debug_pagealloc_enabled())
510 static void init_debug_guardpage(void)
512 if (!debug_pagealloc_enabled())
515 _debug_guardpage_enabled = true;
518 struct page_ext_operations debug_guardpage_ops = {
519 .need = need_debug_guardpage,
520 .init = init_debug_guardpage,
523 static int __init debug_guardpage_minorder_setup(char *buf)
527 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
528 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
531 _debug_guardpage_minorder = res;
532 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
535 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
537 static inline void set_page_guard(struct zone *zone, struct page *page,
538 unsigned int order, int migratetype)
540 struct page_ext *page_ext;
542 if (!debug_guardpage_enabled())
545 page_ext = lookup_page_ext(page);
546 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
548 INIT_LIST_HEAD(&page->lru);
549 set_page_private(page, order);
550 /* Guard pages are not available for any usage */
551 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
554 static inline void clear_page_guard(struct zone *zone, struct page *page,
555 unsigned int order, int migratetype)
557 struct page_ext *page_ext;
559 if (!debug_guardpage_enabled())
562 page_ext = lookup_page_ext(page);
563 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
565 set_page_private(page, 0);
566 if (!is_migrate_isolate(migratetype))
567 __mod_zone_freepage_state(zone, (1 << order), migratetype);
570 struct page_ext_operations debug_guardpage_ops = { NULL, };
571 static inline void set_page_guard(struct zone *zone, struct page *page,
572 unsigned int order, int migratetype) {}
573 static inline void clear_page_guard(struct zone *zone, struct page *page,
574 unsigned int order, int migratetype) {}
577 static inline void set_page_order(struct page *page, unsigned int order)
579 set_page_private(page, order);
580 __SetPageBuddy(page);
583 static inline void rmv_page_order(struct page *page)
585 __ClearPageBuddy(page);
586 set_page_private(page, 0);
590 * This function checks whether a page is free && is the buddy
591 * we can do coalesce a page and its buddy if
592 * (a) the buddy is not in a hole &&
593 * (b) the buddy is in the buddy system &&
594 * (c) a page and its buddy have the same order &&
595 * (d) a page and its buddy are in the same zone.
597 * For recording whether a page is in the buddy system, we set ->_mapcount
598 * PAGE_BUDDY_MAPCOUNT_VALUE.
599 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
600 * serialized by zone->lock.
602 * For recording page's order, we use page_private(page).
604 static inline int page_is_buddy(struct page *page, struct page *buddy,
607 if (!pfn_valid_within(page_to_pfn(buddy)))
610 if (page_is_guard(buddy) && page_order(buddy) == order) {
611 if (page_zone_id(page) != page_zone_id(buddy))
614 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
619 if (PageBuddy(buddy) && page_order(buddy) == order) {
621 * zone check is done late to avoid uselessly
622 * calculating zone/node ids for pages that could
625 if (page_zone_id(page) != page_zone_id(buddy))
628 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
636 * Freeing function for a buddy system allocator.
638 * The concept of a buddy system is to maintain direct-mapped table
639 * (containing bit values) for memory blocks of various "orders".
640 * The bottom level table contains the map for the smallest allocatable
641 * units of memory (here, pages), and each level above it describes
642 * pairs of units from the levels below, hence, "buddies".
643 * At a high level, all that happens here is marking the table entry
644 * at the bottom level available, and propagating the changes upward
645 * as necessary, plus some accounting needed to play nicely with other
646 * parts of the VM system.
647 * At each level, we keep a list of pages, which are heads of continuous
648 * free pages of length of (1 << order) and marked with _mapcount
649 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
651 * So when we are allocating or freeing one, we can derive the state of the
652 * other. That is, if we allocate a small block, and both were
653 * free, the remainder of the region must be split into blocks.
654 * If a block is freed, and its buddy is also free, then this
655 * triggers coalescing into a block of larger size.
660 static inline void __free_one_page(struct page *page,
662 struct zone *zone, unsigned int order,
665 unsigned long page_idx;
666 unsigned long combined_idx;
667 unsigned long uninitialized_var(buddy_idx);
669 unsigned int max_order = MAX_ORDER;
671 VM_BUG_ON(!zone_is_initialized(zone));
672 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
674 VM_BUG_ON(migratetype == -1);
675 if (is_migrate_isolate(migratetype)) {
677 * We restrict max order of merging to prevent merge
678 * between freepages on isolate pageblock and normal
679 * pageblock. Without this, pageblock isolation
680 * could cause incorrect freepage accounting.
682 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
684 __mod_zone_freepage_state(zone, 1 << order, migratetype);
687 page_idx = pfn & ((1 << max_order) - 1);
689 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
690 VM_BUG_ON_PAGE(bad_range(zone, page), page);
692 while (order < max_order - 1) {
693 buddy_idx = __find_buddy_index(page_idx, order);
694 buddy = page + (buddy_idx - page_idx);
695 if (!page_is_buddy(page, buddy, order))
698 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
699 * merge with it and move up one order.
701 if (page_is_guard(buddy)) {
702 clear_page_guard(zone, buddy, order, migratetype);
704 list_del(&buddy->lru);
705 zone->free_area[order].nr_free--;
706 rmv_page_order(buddy);
708 combined_idx = buddy_idx & page_idx;
709 page = page + (combined_idx - page_idx);
710 page_idx = combined_idx;
713 set_page_order(page, order);
716 * If this is not the largest possible page, check if the buddy
717 * of the next-highest order is free. If it is, it's possible
718 * that pages are being freed that will coalesce soon. In case,
719 * that is happening, add the free page to the tail of the list
720 * so it's less likely to be used soon and more likely to be merged
721 * as a higher order page
723 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
724 struct page *higher_page, *higher_buddy;
725 combined_idx = buddy_idx & page_idx;
726 higher_page = page + (combined_idx - page_idx);
727 buddy_idx = __find_buddy_index(combined_idx, order + 1);
728 higher_buddy = higher_page + (buddy_idx - combined_idx);
729 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
730 list_add_tail(&page->lru,
731 &zone->free_area[order].free_list[migratetype]);
736 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
738 zone->free_area[order].nr_free++;
741 static inline int free_pages_check(struct page *page)
743 const char *bad_reason = NULL;
744 unsigned long bad_flags = 0;
746 if (unlikely(atomic_read(&page->_mapcount) != -1))
747 bad_reason = "nonzero mapcount";
748 if (unlikely(page->mapping != NULL))
749 bad_reason = "non-NULL mapping";
750 if (unlikely(atomic_read(&page->_count) != 0))
751 bad_reason = "nonzero _count";
752 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
753 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
754 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
757 if (unlikely(page->mem_cgroup))
758 bad_reason = "page still charged to cgroup";
760 if (unlikely(bad_reason)) {
761 bad_page(page, bad_reason, bad_flags);
764 page_cpupid_reset_last(page);
765 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
766 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
771 * Frees a number of pages from the PCP lists
772 * Assumes all pages on list are in same zone, and of same order.
773 * count is the number of pages to free.
775 * If the zone was previously in an "all pages pinned" state then look to
776 * see if this freeing clears that state.
778 * And clear the zone's pages_scanned counter, to hold off the "all pages are
779 * pinned" detection logic.
781 static void free_pcppages_bulk(struct zone *zone, int count,
782 struct per_cpu_pages *pcp)
787 unsigned long nr_scanned;
789 spin_lock(&zone->lock);
790 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
792 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
796 struct list_head *list;
799 * Remove pages from lists in a round-robin fashion. A
800 * batch_free count is maintained that is incremented when an
801 * empty list is encountered. This is so more pages are freed
802 * off fuller lists instead of spinning excessively around empty
807 if (++migratetype == MIGRATE_PCPTYPES)
809 list = &pcp->lists[migratetype];
810 } while (list_empty(list));
812 /* This is the only non-empty list. Free them all. */
813 if (batch_free == MIGRATE_PCPTYPES)
814 batch_free = to_free;
817 int mt; /* migratetype of the to-be-freed page */
819 page = list_entry(list->prev, struct page, lru);
820 /* must delete as __free_one_page list manipulates */
821 list_del(&page->lru);
823 mt = get_pcppage_migratetype(page);
824 /* MIGRATE_ISOLATE page should not go to pcplists */
825 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
826 /* Pageblock could have been isolated meanwhile */
827 if (unlikely(has_isolate_pageblock(zone)))
828 mt = get_pageblock_migratetype(page);
830 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
831 trace_mm_page_pcpu_drain(page, 0, mt);
832 } while (--to_free && --batch_free && !list_empty(list));
834 spin_unlock(&zone->lock);
837 static void free_one_page(struct zone *zone,
838 struct page *page, unsigned long pfn,
842 unsigned long nr_scanned;
843 spin_lock(&zone->lock);
844 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
846 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
848 if (unlikely(has_isolate_pageblock(zone) ||
849 is_migrate_isolate(migratetype))) {
850 migratetype = get_pfnblock_migratetype(page, pfn);
852 __free_one_page(page, pfn, zone, order, migratetype);
853 spin_unlock(&zone->lock);
856 static int free_tail_pages_check(struct page *head_page, struct page *page)
861 * We rely page->lru.next never has bit 0 set, unless the page
862 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
864 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
866 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
870 switch (page - head_page) {
872 /* the first tail page: ->mapping is compound_mapcount() */
873 if (unlikely(compound_mapcount(page))) {
874 bad_page(page, "nonzero compound_mapcount", 0);
880 * the second tail page: ->mapping is
881 * page_deferred_list().next -- ignore value.
885 if (page->mapping != TAIL_MAPPING) {
886 bad_page(page, "corrupted mapping in tail page", 0);
891 if (unlikely(!PageTail(page))) {
892 bad_page(page, "PageTail not set", 0);
895 if (unlikely(compound_head(page) != head_page)) {
896 bad_page(page, "compound_head not consistent", 0);
901 page->mapping = NULL;
902 clear_compound_head(page);
906 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
907 unsigned long zone, int nid)
909 set_page_links(page, zone, nid, pfn);
910 init_page_count(page);
911 page_mapcount_reset(page);
912 page_cpupid_reset_last(page);
914 INIT_LIST_HEAD(&page->lru);
915 #ifdef WANT_PAGE_VIRTUAL
916 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
917 if (!is_highmem_idx(zone))
918 set_page_address(page, __va(pfn << PAGE_SHIFT));
922 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
925 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
928 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
929 static void init_reserved_page(unsigned long pfn)
934 if (!early_page_uninitialised(pfn))
937 nid = early_pfn_to_nid(pfn);
938 pgdat = NODE_DATA(nid);
940 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
941 struct zone *zone = &pgdat->node_zones[zid];
943 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
946 __init_single_pfn(pfn, zid, nid);
949 static inline void init_reserved_page(unsigned long pfn)
952 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
955 * Initialised pages do not have PageReserved set. This function is
956 * called for each range allocated by the bootmem allocator and
957 * marks the pages PageReserved. The remaining valid pages are later
958 * sent to the buddy page allocator.
960 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
962 unsigned long start_pfn = PFN_DOWN(start);
963 unsigned long end_pfn = PFN_UP(end);
965 for (; start_pfn < end_pfn; start_pfn++) {
966 if (pfn_valid(start_pfn)) {
967 struct page *page = pfn_to_page(start_pfn);
969 init_reserved_page(start_pfn);
971 /* Avoid false-positive PageTail() */
972 INIT_LIST_HEAD(&page->lru);
974 SetPageReserved(page);
979 static bool free_pages_prepare(struct page *page, unsigned int order)
981 bool compound = PageCompound(page);
984 VM_BUG_ON_PAGE(PageTail(page), page);
985 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
987 trace_mm_page_free(page, order);
988 kmemcheck_free_shadow(page, order);
989 kasan_free_pages(page, order);
992 page->mapping = NULL;
993 bad += free_pages_check(page);
994 for (i = 1; i < (1 << order); i++) {
996 bad += free_tail_pages_check(page, page + i);
997 bad += free_pages_check(page + i);
1002 reset_page_owner(page, order);
1004 if (!PageHighMem(page)) {
1005 debug_check_no_locks_freed(page_address(page),
1006 PAGE_SIZE << order);
1007 debug_check_no_obj_freed(page_address(page),
1008 PAGE_SIZE << order);
1010 arch_free_page(page, order);
1011 kernel_map_pages(page, 1 << order, 0);
1016 static void __free_pages_ok(struct page *page, unsigned int order)
1018 unsigned long flags;
1020 unsigned long pfn = page_to_pfn(page);
1022 if (!free_pages_prepare(page, order))
1025 migratetype = get_pfnblock_migratetype(page, pfn);
1026 local_irq_save(flags);
1027 __count_vm_events(PGFREE, 1 << order);
1028 free_one_page(page_zone(page), page, pfn, order, migratetype);
1029 local_irq_restore(flags);
1032 static void __init __free_pages_boot_core(struct page *page,
1033 unsigned long pfn, unsigned int order)
1035 unsigned int nr_pages = 1 << order;
1036 struct page *p = page;
1040 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1042 __ClearPageReserved(p);
1043 set_page_count(p, 0);
1045 __ClearPageReserved(p);
1046 set_page_count(p, 0);
1048 page_zone(page)->managed_pages += nr_pages;
1049 set_page_refcounted(page);
1050 __free_pages(page, order);
1053 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1054 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1056 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1058 int __meminit early_pfn_to_nid(unsigned long pfn)
1060 static DEFINE_SPINLOCK(early_pfn_lock);
1063 spin_lock(&early_pfn_lock);
1064 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1067 spin_unlock(&early_pfn_lock);
1073 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1074 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1075 struct mminit_pfnnid_cache *state)
1079 nid = __early_pfn_to_nid(pfn, state);
1080 if (nid >= 0 && nid != node)
1085 /* Only safe to use early in boot when initialisation is single-threaded */
1086 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1088 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1093 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1097 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1098 struct mminit_pfnnid_cache *state)
1105 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1108 if (early_page_uninitialised(pfn))
1110 return __free_pages_boot_core(page, pfn, order);
1113 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1114 static void __init deferred_free_range(struct page *page,
1115 unsigned long pfn, int nr_pages)
1122 /* Free a large naturally-aligned chunk if possible */
1123 if (nr_pages == MAX_ORDER_NR_PAGES &&
1124 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1125 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1126 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1130 for (i = 0; i < nr_pages; i++, page++, pfn++)
1131 __free_pages_boot_core(page, pfn, 0);
1134 /* Completion tracking for deferred_init_memmap() threads */
1135 static atomic_t pgdat_init_n_undone __initdata;
1136 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1138 static inline void __init pgdat_init_report_one_done(void)
1140 if (atomic_dec_and_test(&pgdat_init_n_undone))
1141 complete(&pgdat_init_all_done_comp);
1144 /* Initialise remaining memory on a node */
1145 static int __init deferred_init_memmap(void *data)
1147 pg_data_t *pgdat = data;
1148 int nid = pgdat->node_id;
1149 struct mminit_pfnnid_cache nid_init_state = { };
1150 unsigned long start = jiffies;
1151 unsigned long nr_pages = 0;
1152 unsigned long walk_start, walk_end;
1155 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1156 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1158 if (first_init_pfn == ULONG_MAX) {
1159 pgdat_init_report_one_done();
1163 /* Bind memory initialisation thread to a local node if possible */
1164 if (!cpumask_empty(cpumask))
1165 set_cpus_allowed_ptr(current, cpumask);
1167 /* Sanity check boundaries */
1168 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1169 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1170 pgdat->first_deferred_pfn = ULONG_MAX;
1172 /* Only the highest zone is deferred so find it */
1173 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1174 zone = pgdat->node_zones + zid;
1175 if (first_init_pfn < zone_end_pfn(zone))
1179 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1180 unsigned long pfn, end_pfn;
1181 struct page *page = NULL;
1182 struct page *free_base_page = NULL;
1183 unsigned long free_base_pfn = 0;
1186 end_pfn = min(walk_end, zone_end_pfn(zone));
1187 pfn = first_init_pfn;
1188 if (pfn < walk_start)
1190 if (pfn < zone->zone_start_pfn)
1191 pfn = zone->zone_start_pfn;
1193 for (; pfn < end_pfn; pfn++) {
1194 if (!pfn_valid_within(pfn))
1198 * Ensure pfn_valid is checked every
1199 * MAX_ORDER_NR_PAGES for memory holes
1201 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1202 if (!pfn_valid(pfn)) {
1208 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1213 /* Minimise pfn page lookups and scheduler checks */
1214 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1217 nr_pages += nr_to_free;
1218 deferred_free_range(free_base_page,
1219 free_base_pfn, nr_to_free);
1220 free_base_page = NULL;
1221 free_base_pfn = nr_to_free = 0;
1223 page = pfn_to_page(pfn);
1228 VM_BUG_ON(page_zone(page) != zone);
1232 __init_single_page(page, pfn, zid, nid);
1233 if (!free_base_page) {
1234 free_base_page = page;
1235 free_base_pfn = pfn;
1240 /* Where possible, batch up pages for a single free */
1243 /* Free the current block of pages to allocator */
1244 nr_pages += nr_to_free;
1245 deferred_free_range(free_base_page, free_base_pfn,
1247 free_base_page = NULL;
1248 free_base_pfn = nr_to_free = 0;
1251 first_init_pfn = max(end_pfn, first_init_pfn);
1254 /* Sanity check that the next zone really is unpopulated */
1255 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1257 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1258 jiffies_to_msecs(jiffies - start));
1260 pgdat_init_report_one_done();
1264 void __init page_alloc_init_late(void)
1268 /* There will be num_node_state(N_MEMORY) threads */
1269 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1270 for_each_node_state(nid, N_MEMORY) {
1271 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1274 /* Block until all are initialised */
1275 wait_for_completion(&pgdat_init_all_done_comp);
1277 /* Reinit limits that are based on free pages after the kernel is up */
1278 files_maxfiles_init();
1280 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1283 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1284 void __init init_cma_reserved_pageblock(struct page *page)
1286 unsigned i = pageblock_nr_pages;
1287 struct page *p = page;
1290 __ClearPageReserved(p);
1291 set_page_count(p, 0);
1294 set_pageblock_migratetype(page, MIGRATE_CMA);
1296 if (pageblock_order >= MAX_ORDER) {
1297 i = pageblock_nr_pages;
1300 set_page_refcounted(p);
1301 __free_pages(p, MAX_ORDER - 1);
1302 p += MAX_ORDER_NR_PAGES;
1303 } while (i -= MAX_ORDER_NR_PAGES);
1305 set_page_refcounted(page);
1306 __free_pages(page, pageblock_order);
1309 adjust_managed_page_count(page, pageblock_nr_pages);
1314 * The order of subdivision here is critical for the IO subsystem.
1315 * Please do not alter this order without good reasons and regression
1316 * testing. Specifically, as large blocks of memory are subdivided,
1317 * the order in which smaller blocks are delivered depends on the order
1318 * they're subdivided in this function. This is the primary factor
1319 * influencing the order in which pages are delivered to the IO
1320 * subsystem according to empirical testing, and this is also justified
1321 * by considering the behavior of a buddy system containing a single
1322 * large block of memory acted on by a series of small allocations.
1323 * This behavior is a critical factor in sglist merging's success.
1327 static inline void expand(struct zone *zone, struct page *page,
1328 int low, int high, struct free_area *area,
1331 unsigned long size = 1 << high;
1333 while (high > low) {
1337 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1339 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1340 debug_guardpage_enabled() &&
1341 high < debug_guardpage_minorder()) {
1343 * Mark as guard pages (or page), that will allow to
1344 * merge back to allocator when buddy will be freed.
1345 * Corresponding page table entries will not be touched,
1346 * pages will stay not present in virtual address space
1348 set_page_guard(zone, &page[size], high, migratetype);
1351 list_add(&page[size].lru, &area->free_list[migratetype]);
1353 set_page_order(&page[size], high);
1358 * This page is about to be returned from the page allocator
1360 static inline int check_new_page(struct page *page)
1362 const char *bad_reason = NULL;
1363 unsigned long bad_flags = 0;
1365 if (unlikely(atomic_read(&page->_mapcount) != -1))
1366 bad_reason = "nonzero mapcount";
1367 if (unlikely(page->mapping != NULL))
1368 bad_reason = "non-NULL mapping";
1369 if (unlikely(atomic_read(&page->_count) != 0))
1370 bad_reason = "nonzero _count";
1371 if (unlikely(page->flags & __PG_HWPOISON)) {
1372 bad_reason = "HWPoisoned (hardware-corrupted)";
1373 bad_flags = __PG_HWPOISON;
1375 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1376 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1377 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1380 if (unlikely(page->mem_cgroup))
1381 bad_reason = "page still charged to cgroup";
1383 if (unlikely(bad_reason)) {
1384 bad_page(page, bad_reason, bad_flags);
1390 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1395 for (i = 0; i < (1 << order); i++) {
1396 struct page *p = page + i;
1397 if (unlikely(check_new_page(p)))
1401 set_page_private(page, 0);
1402 set_page_refcounted(page);
1404 arch_alloc_page(page, order);
1405 kernel_map_pages(page, 1 << order, 1);
1406 kasan_alloc_pages(page, order);
1408 if (gfp_flags & __GFP_ZERO)
1409 for (i = 0; i < (1 << order); i++)
1410 clear_highpage(page + i);
1412 if (order && (gfp_flags & __GFP_COMP))
1413 prep_compound_page(page, order);
1415 set_page_owner(page, order, gfp_flags);
1418 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1419 * allocate the page. The expectation is that the caller is taking
1420 * steps that will free more memory. The caller should avoid the page
1421 * being used for !PFMEMALLOC purposes.
1423 if (alloc_flags & ALLOC_NO_WATERMARKS)
1424 set_page_pfmemalloc(page);
1426 clear_page_pfmemalloc(page);
1432 * Go through the free lists for the given migratetype and remove
1433 * the smallest available page from the freelists
1436 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1439 unsigned int current_order;
1440 struct free_area *area;
1443 /* Find a page of the appropriate size in the preferred list */
1444 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1445 area = &(zone->free_area[current_order]);
1446 if (list_empty(&area->free_list[migratetype]))
1449 page = list_entry(area->free_list[migratetype].next,
1451 list_del(&page->lru);
1452 rmv_page_order(page);
1454 expand(zone, page, order, current_order, area, migratetype);
1455 set_pcppage_migratetype(page, migratetype);
1464 * This array describes the order lists are fallen back to when
1465 * the free lists for the desirable migrate type are depleted
1467 static int fallbacks[MIGRATE_TYPES][4] = {
1468 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1469 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1470 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1472 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1474 #ifdef CONFIG_MEMORY_ISOLATION
1475 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1480 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1483 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1486 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1487 unsigned int order) { return NULL; }
1491 * Move the free pages in a range to the free lists of the requested type.
1492 * Note that start_page and end_pages are not aligned on a pageblock
1493 * boundary. If alignment is required, use move_freepages_block()
1495 int move_freepages(struct zone *zone,
1496 struct page *start_page, struct page *end_page,
1501 int pages_moved = 0;
1503 #ifndef CONFIG_HOLES_IN_ZONE
1505 * page_zone is not safe to call in this context when
1506 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1507 * anyway as we check zone boundaries in move_freepages_block().
1508 * Remove at a later date when no bug reports exist related to
1509 * grouping pages by mobility
1511 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1514 for (page = start_page; page <= end_page;) {
1515 /* Make sure we are not inadvertently changing nodes */
1516 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1518 if (!pfn_valid_within(page_to_pfn(page))) {
1523 if (!PageBuddy(page)) {
1528 order = page_order(page);
1529 list_move(&page->lru,
1530 &zone->free_area[order].free_list[migratetype]);
1532 pages_moved += 1 << order;
1538 int move_freepages_block(struct zone *zone, struct page *page,
1541 unsigned long start_pfn, end_pfn;
1542 struct page *start_page, *end_page;
1544 start_pfn = page_to_pfn(page);
1545 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1546 start_page = pfn_to_page(start_pfn);
1547 end_page = start_page + pageblock_nr_pages - 1;
1548 end_pfn = start_pfn + pageblock_nr_pages - 1;
1550 /* Do not cross zone boundaries */
1551 if (!zone_spans_pfn(zone, start_pfn))
1553 if (!zone_spans_pfn(zone, end_pfn))
1556 return move_freepages(zone, start_page, end_page, migratetype);
1559 static void change_pageblock_range(struct page *pageblock_page,
1560 int start_order, int migratetype)
1562 int nr_pageblocks = 1 << (start_order - pageblock_order);
1564 while (nr_pageblocks--) {
1565 set_pageblock_migratetype(pageblock_page, migratetype);
1566 pageblock_page += pageblock_nr_pages;
1571 * When we are falling back to another migratetype during allocation, try to
1572 * steal extra free pages from the same pageblocks to satisfy further
1573 * allocations, instead of polluting multiple pageblocks.
1575 * If we are stealing a relatively large buddy page, it is likely there will
1576 * be more free pages in the pageblock, so try to steal them all. For
1577 * reclaimable and unmovable allocations, we steal regardless of page size,
1578 * as fragmentation caused by those allocations polluting movable pageblocks
1579 * is worse than movable allocations stealing from unmovable and reclaimable
1582 static bool can_steal_fallback(unsigned int order, int start_mt)
1585 * Leaving this order check is intended, although there is
1586 * relaxed order check in next check. The reason is that
1587 * we can actually steal whole pageblock if this condition met,
1588 * but, below check doesn't guarantee it and that is just heuristic
1589 * so could be changed anytime.
1591 if (order >= pageblock_order)
1594 if (order >= pageblock_order / 2 ||
1595 start_mt == MIGRATE_RECLAIMABLE ||
1596 start_mt == MIGRATE_UNMOVABLE ||
1597 page_group_by_mobility_disabled)
1604 * This function implements actual steal behaviour. If order is large enough,
1605 * we can steal whole pageblock. If not, we first move freepages in this
1606 * pageblock and check whether half of pages are moved or not. If half of
1607 * pages are moved, we can change migratetype of pageblock and permanently
1608 * use it's pages as requested migratetype in the future.
1610 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1613 unsigned int current_order = page_order(page);
1616 /* Take ownership for orders >= pageblock_order */
1617 if (current_order >= pageblock_order) {
1618 change_pageblock_range(page, current_order, start_type);
1622 pages = move_freepages_block(zone, page, start_type);
1624 /* Claim the whole block if over half of it is free */
1625 if (pages >= (1 << (pageblock_order-1)) ||
1626 page_group_by_mobility_disabled)
1627 set_pageblock_migratetype(page, start_type);
1631 * Check whether there is a suitable fallback freepage with requested order.
1632 * If only_stealable is true, this function returns fallback_mt only if
1633 * we can steal other freepages all together. This would help to reduce
1634 * fragmentation due to mixed migratetype pages in one pageblock.
1636 int find_suitable_fallback(struct free_area *area, unsigned int order,
1637 int migratetype, bool only_stealable, bool *can_steal)
1642 if (area->nr_free == 0)
1647 fallback_mt = fallbacks[migratetype][i];
1648 if (fallback_mt == MIGRATE_TYPES)
1651 if (list_empty(&area->free_list[fallback_mt]))
1654 if (can_steal_fallback(order, migratetype))
1657 if (!only_stealable)
1668 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1669 * there are no empty page blocks that contain a page with a suitable order
1671 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1672 unsigned int alloc_order)
1675 unsigned long max_managed, flags;
1678 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1679 * Check is race-prone but harmless.
1681 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1682 if (zone->nr_reserved_highatomic >= max_managed)
1685 spin_lock_irqsave(&zone->lock, flags);
1687 /* Recheck the nr_reserved_highatomic limit under the lock */
1688 if (zone->nr_reserved_highatomic >= max_managed)
1692 mt = get_pageblock_migratetype(page);
1693 if (mt != MIGRATE_HIGHATOMIC &&
1694 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1695 zone->nr_reserved_highatomic += pageblock_nr_pages;
1696 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1697 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1701 spin_unlock_irqrestore(&zone->lock, flags);
1705 * Used when an allocation is about to fail under memory pressure. This
1706 * potentially hurts the reliability of high-order allocations when under
1707 * intense memory pressure but failed atomic allocations should be easier
1708 * to recover from than an OOM.
1710 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1712 struct zonelist *zonelist = ac->zonelist;
1713 unsigned long flags;
1719 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1721 /* Preserve at least one pageblock */
1722 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1725 spin_lock_irqsave(&zone->lock, flags);
1726 for (order = 0; order < MAX_ORDER; order++) {
1727 struct free_area *area = &(zone->free_area[order]);
1729 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1732 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1736 * It should never happen but changes to locking could
1737 * inadvertently allow a per-cpu drain to add pages
1738 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1739 * and watch for underflows.
1741 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1742 zone->nr_reserved_highatomic);
1745 * Convert to ac->migratetype and avoid the normal
1746 * pageblock stealing heuristics. Minimally, the caller
1747 * is doing the work and needs the pages. More
1748 * importantly, if the block was always converted to
1749 * MIGRATE_UNMOVABLE or another type then the number
1750 * of pageblocks that cannot be completely freed
1753 set_pageblock_migratetype(page, ac->migratetype);
1754 move_freepages_block(zone, page, ac->migratetype);
1755 spin_unlock_irqrestore(&zone->lock, flags);
1758 spin_unlock_irqrestore(&zone->lock, flags);
1762 /* Remove an element from the buddy allocator from the fallback list */
1763 static inline struct page *
1764 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1766 struct free_area *area;
1767 unsigned int current_order;
1772 /* Find the largest possible block of pages in the other list */
1773 for (current_order = MAX_ORDER-1;
1774 current_order >= order && current_order <= MAX_ORDER-1;
1776 area = &(zone->free_area[current_order]);
1777 fallback_mt = find_suitable_fallback(area, current_order,
1778 start_migratetype, false, &can_steal);
1779 if (fallback_mt == -1)
1782 page = list_entry(area->free_list[fallback_mt].next,
1785 steal_suitable_fallback(zone, page, start_migratetype);
1787 /* Remove the page from the freelists */
1789 list_del(&page->lru);
1790 rmv_page_order(page);
1792 expand(zone, page, order, current_order, area,
1795 * The pcppage_migratetype may differ from pageblock's
1796 * migratetype depending on the decisions in
1797 * find_suitable_fallback(). This is OK as long as it does not
1798 * differ for MIGRATE_CMA pageblocks. Those can be used as
1799 * fallback only via special __rmqueue_cma_fallback() function
1801 set_pcppage_migratetype(page, start_migratetype);
1803 trace_mm_page_alloc_extfrag(page, order, current_order,
1804 start_migratetype, fallback_mt);
1813 * Do the hard work of removing an element from the buddy allocator.
1814 * Call me with the zone->lock already held.
1816 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1817 int migratetype, gfp_t gfp_flags)
1821 page = __rmqueue_smallest(zone, order, migratetype);
1822 if (unlikely(!page)) {
1823 if (migratetype == MIGRATE_MOVABLE)
1824 page = __rmqueue_cma_fallback(zone, order);
1827 page = __rmqueue_fallback(zone, order, migratetype);
1830 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1835 * Obtain a specified number of elements from the buddy allocator, all under
1836 * a single hold of the lock, for efficiency. Add them to the supplied list.
1837 * Returns the number of new pages which were placed at *list.
1839 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1840 unsigned long count, struct list_head *list,
1841 int migratetype, bool cold)
1845 spin_lock(&zone->lock);
1846 for (i = 0; i < count; ++i) {
1847 struct page *page = __rmqueue(zone, order, migratetype, 0);
1848 if (unlikely(page == NULL))
1852 * Split buddy pages returned by expand() are received here
1853 * in physical page order. The page is added to the callers and
1854 * list and the list head then moves forward. From the callers
1855 * perspective, the linked list is ordered by page number in
1856 * some conditions. This is useful for IO devices that can
1857 * merge IO requests if the physical pages are ordered
1861 list_add(&page->lru, list);
1863 list_add_tail(&page->lru, list);
1865 if (is_migrate_cma(get_pcppage_migratetype(page)))
1866 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1869 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1870 spin_unlock(&zone->lock);
1876 * Called from the vmstat counter updater to drain pagesets of this
1877 * currently executing processor on remote nodes after they have
1880 * Note that this function must be called with the thread pinned to
1881 * a single processor.
1883 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1885 unsigned long flags;
1886 int to_drain, batch;
1888 local_irq_save(flags);
1889 batch = READ_ONCE(pcp->batch);
1890 to_drain = min(pcp->count, batch);
1892 free_pcppages_bulk(zone, to_drain, pcp);
1893 pcp->count -= to_drain;
1895 local_irq_restore(flags);
1900 * Drain pcplists of the indicated processor and zone.
1902 * The processor must either be the current processor and the
1903 * thread pinned to the current processor or a processor that
1906 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1908 unsigned long flags;
1909 struct per_cpu_pageset *pset;
1910 struct per_cpu_pages *pcp;
1912 local_irq_save(flags);
1913 pset = per_cpu_ptr(zone->pageset, cpu);
1917 free_pcppages_bulk(zone, pcp->count, pcp);
1920 local_irq_restore(flags);
1924 * Drain pcplists of all zones on the indicated processor.
1926 * The processor must either be the current processor and the
1927 * thread pinned to the current processor or a processor that
1930 static void drain_pages(unsigned int cpu)
1934 for_each_populated_zone(zone) {
1935 drain_pages_zone(cpu, zone);
1940 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1942 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1943 * the single zone's pages.
1945 void drain_local_pages(struct zone *zone)
1947 int cpu = smp_processor_id();
1950 drain_pages_zone(cpu, zone);
1956 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1958 * When zone parameter is non-NULL, spill just the single zone's pages.
1960 * Note that this code is protected against sending an IPI to an offline
1961 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1962 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1963 * nothing keeps CPUs from showing up after we populated the cpumask and
1964 * before the call to on_each_cpu_mask().
1966 void drain_all_pages(struct zone *zone)
1971 * Allocate in the BSS so we wont require allocation in
1972 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1974 static cpumask_t cpus_with_pcps;
1977 * We don't care about racing with CPU hotplug event
1978 * as offline notification will cause the notified
1979 * cpu to drain that CPU pcps and on_each_cpu_mask
1980 * disables preemption as part of its processing
1982 for_each_online_cpu(cpu) {
1983 struct per_cpu_pageset *pcp;
1985 bool has_pcps = false;
1988 pcp = per_cpu_ptr(zone->pageset, cpu);
1992 for_each_populated_zone(z) {
1993 pcp = per_cpu_ptr(z->pageset, cpu);
1994 if (pcp->pcp.count) {
2002 cpumask_set_cpu(cpu, &cpus_with_pcps);
2004 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2006 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2010 #ifdef CONFIG_HIBERNATION
2012 void mark_free_pages(struct zone *zone)
2014 unsigned long pfn, max_zone_pfn;
2015 unsigned long flags;
2016 unsigned int order, t;
2017 struct list_head *curr;
2019 if (zone_is_empty(zone))
2022 spin_lock_irqsave(&zone->lock, flags);
2024 max_zone_pfn = zone_end_pfn(zone);
2025 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2026 if (pfn_valid(pfn)) {
2027 struct page *page = pfn_to_page(pfn);
2029 if (!swsusp_page_is_forbidden(page))
2030 swsusp_unset_page_free(page);
2033 for_each_migratetype_order(order, t) {
2034 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2037 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2038 for (i = 0; i < (1UL << order); i++)
2039 swsusp_set_page_free(pfn_to_page(pfn + i));
2042 spin_unlock_irqrestore(&zone->lock, flags);
2044 #endif /* CONFIG_PM */
2047 * Free a 0-order page
2048 * cold == true ? free a cold page : free a hot page
2050 void free_hot_cold_page(struct page *page, bool cold)
2052 struct zone *zone = page_zone(page);
2053 struct per_cpu_pages *pcp;
2054 unsigned long flags;
2055 unsigned long pfn = page_to_pfn(page);
2058 if (!free_pages_prepare(page, 0))
2061 migratetype = get_pfnblock_migratetype(page, pfn);
2062 set_pcppage_migratetype(page, migratetype);
2063 local_irq_save(flags);
2064 __count_vm_event(PGFREE);
2067 * We only track unmovable, reclaimable and movable on pcp lists.
2068 * Free ISOLATE pages back to the allocator because they are being
2069 * offlined but treat RESERVE as movable pages so we can get those
2070 * areas back if necessary. Otherwise, we may have to free
2071 * excessively into the page allocator
2073 if (migratetype >= MIGRATE_PCPTYPES) {
2074 if (unlikely(is_migrate_isolate(migratetype))) {
2075 free_one_page(zone, page, pfn, 0, migratetype);
2078 migratetype = MIGRATE_MOVABLE;
2081 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2083 list_add(&page->lru, &pcp->lists[migratetype]);
2085 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2087 if (pcp->count >= pcp->high) {
2088 unsigned long batch = READ_ONCE(pcp->batch);
2089 free_pcppages_bulk(zone, batch, pcp);
2090 pcp->count -= batch;
2094 local_irq_restore(flags);
2098 * Free a list of 0-order pages
2100 void free_hot_cold_page_list(struct list_head *list, bool cold)
2102 struct page *page, *next;
2104 list_for_each_entry_safe(page, next, list, lru) {
2105 trace_mm_page_free_batched(page, cold);
2106 free_hot_cold_page(page, cold);
2111 * split_page takes a non-compound higher-order page, and splits it into
2112 * n (1<<order) sub-pages: page[0..n]
2113 * Each sub-page must be freed individually.
2115 * Note: this is probably too low level an operation for use in drivers.
2116 * Please consult with lkml before using this in your driver.
2118 void split_page(struct page *page, unsigned int order)
2123 VM_BUG_ON_PAGE(PageCompound(page), page);
2124 VM_BUG_ON_PAGE(!page_count(page), page);
2126 #ifdef CONFIG_KMEMCHECK
2128 * Split shadow pages too, because free(page[0]) would
2129 * otherwise free the whole shadow.
2131 if (kmemcheck_page_is_tracked(page))
2132 split_page(virt_to_page(page[0].shadow), order);
2135 gfp_mask = get_page_owner_gfp(page);
2136 set_page_owner(page, 0, gfp_mask);
2137 for (i = 1; i < (1 << order); i++) {
2138 set_page_refcounted(page + i);
2139 set_page_owner(page + i, 0, gfp_mask);
2142 EXPORT_SYMBOL_GPL(split_page);
2144 int __isolate_free_page(struct page *page, unsigned int order)
2146 unsigned long watermark;
2150 BUG_ON(!PageBuddy(page));
2152 zone = page_zone(page);
2153 mt = get_pageblock_migratetype(page);
2155 if (!is_migrate_isolate(mt)) {
2156 /* Obey watermarks as if the page was being allocated */
2157 watermark = low_wmark_pages(zone) + (1 << order);
2158 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2161 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2164 /* Remove page from free list */
2165 list_del(&page->lru);
2166 zone->free_area[order].nr_free--;
2167 rmv_page_order(page);
2169 set_page_owner(page, order, __GFP_MOVABLE);
2171 /* Set the pageblock if the isolated page is at least a pageblock */
2172 if (order >= pageblock_order - 1) {
2173 struct page *endpage = page + (1 << order) - 1;
2174 for (; page < endpage; page += pageblock_nr_pages) {
2175 int mt = get_pageblock_migratetype(page);
2176 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2177 set_pageblock_migratetype(page,
2183 return 1UL << order;
2187 * Similar to split_page except the page is already free. As this is only
2188 * being used for migration, the migratetype of the block also changes.
2189 * As this is called with interrupts disabled, the caller is responsible
2190 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2193 * Note: this is probably too low level an operation for use in drivers.
2194 * Please consult with lkml before using this in your driver.
2196 int split_free_page(struct page *page)
2201 order = page_order(page);
2203 nr_pages = __isolate_free_page(page, order);
2207 /* Split into individual pages */
2208 set_page_refcounted(page);
2209 split_page(page, order);
2214 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2217 struct page *buffered_rmqueue(struct zone *preferred_zone,
2218 struct zone *zone, unsigned int order,
2219 gfp_t gfp_flags, int alloc_flags, int migratetype)
2221 unsigned long flags;
2223 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2225 if (likely(order == 0)) {
2226 struct per_cpu_pages *pcp;
2227 struct list_head *list;
2229 local_irq_save(flags);
2230 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2231 list = &pcp->lists[migratetype];
2232 if (list_empty(list)) {
2233 pcp->count += rmqueue_bulk(zone, 0,
2236 if (unlikely(list_empty(list)))
2241 page = list_entry(list->prev, struct page, lru);
2243 page = list_entry(list->next, struct page, lru);
2245 list_del(&page->lru);
2248 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2250 * __GFP_NOFAIL is not to be used in new code.
2252 * All __GFP_NOFAIL callers should be fixed so that they
2253 * properly detect and handle allocation failures.
2255 * We most definitely don't want callers attempting to
2256 * allocate greater than order-1 page units with
2259 WARN_ON_ONCE(order > 1);
2261 spin_lock_irqsave(&zone->lock, flags);
2264 if (alloc_flags & ALLOC_HARDER) {
2265 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2267 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2270 page = __rmqueue(zone, order, migratetype, gfp_flags);
2271 spin_unlock(&zone->lock);
2274 __mod_zone_freepage_state(zone, -(1 << order),
2275 get_pcppage_migratetype(page));
2278 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2279 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2280 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2281 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2283 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2284 zone_statistics(preferred_zone, zone, gfp_flags);
2285 local_irq_restore(flags);
2287 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2291 local_irq_restore(flags);
2295 #ifdef CONFIG_FAIL_PAGE_ALLOC
2298 struct fault_attr attr;
2300 bool ignore_gfp_highmem;
2301 bool ignore_gfp_reclaim;
2303 } fail_page_alloc = {
2304 .attr = FAULT_ATTR_INITIALIZER,
2305 .ignore_gfp_reclaim = true,
2306 .ignore_gfp_highmem = true,
2310 static int __init setup_fail_page_alloc(char *str)
2312 return setup_fault_attr(&fail_page_alloc.attr, str);
2314 __setup("fail_page_alloc=", setup_fail_page_alloc);
2316 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2318 if (order < fail_page_alloc.min_order)
2320 if (gfp_mask & __GFP_NOFAIL)
2322 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2324 if (fail_page_alloc.ignore_gfp_reclaim &&
2325 (gfp_mask & __GFP_DIRECT_RECLAIM))
2328 return should_fail(&fail_page_alloc.attr, 1 << order);
2331 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2333 static int __init fail_page_alloc_debugfs(void)
2335 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2338 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2339 &fail_page_alloc.attr);
2341 return PTR_ERR(dir);
2343 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2344 &fail_page_alloc.ignore_gfp_reclaim))
2346 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2347 &fail_page_alloc.ignore_gfp_highmem))
2349 if (!debugfs_create_u32("min-order", mode, dir,
2350 &fail_page_alloc.min_order))
2355 debugfs_remove_recursive(dir);
2360 late_initcall(fail_page_alloc_debugfs);
2362 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2364 #else /* CONFIG_FAIL_PAGE_ALLOC */
2366 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2371 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2374 * Return true if free base pages are above 'mark'. For high-order checks it
2375 * will return true of the order-0 watermark is reached and there is at least
2376 * one free page of a suitable size. Checking now avoids taking the zone lock
2377 * to check in the allocation paths if no pages are free.
2379 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2380 unsigned long mark, int classzone_idx, int alloc_flags,
2385 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2387 /* free_pages may go negative - that's OK */
2388 free_pages -= (1 << order) - 1;
2390 if (alloc_flags & ALLOC_HIGH)
2394 * If the caller does not have rights to ALLOC_HARDER then subtract
2395 * the high-atomic reserves. This will over-estimate the size of the
2396 * atomic reserve but it avoids a search.
2398 if (likely(!alloc_harder))
2399 free_pages -= z->nr_reserved_highatomic;
2404 /* If allocation can't use CMA areas don't use free CMA pages */
2405 if (!(alloc_flags & ALLOC_CMA))
2406 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2410 * Check watermarks for an order-0 allocation request. If these
2411 * are not met, then a high-order request also cannot go ahead
2412 * even if a suitable page happened to be free.
2414 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2417 /* If this is an order-0 request then the watermark is fine */
2421 /* For a high-order request, check at least one suitable page is free */
2422 for (o = order; o < MAX_ORDER; o++) {
2423 struct free_area *area = &z->free_area[o];
2432 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2433 if (!list_empty(&area->free_list[mt]))
2438 if ((alloc_flags & ALLOC_CMA) &&
2439 !list_empty(&area->free_list[MIGRATE_CMA])) {
2447 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2448 int classzone_idx, int alloc_flags)
2450 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2451 zone_page_state(z, NR_FREE_PAGES));
2454 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2455 unsigned long mark, int classzone_idx)
2457 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2459 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2460 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2462 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2467 static bool zone_local(struct zone *local_zone, struct zone *zone)
2469 return local_zone->node == zone->node;
2472 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2474 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2477 #else /* CONFIG_NUMA */
2478 static bool zone_local(struct zone *local_zone, struct zone *zone)
2483 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2487 #endif /* CONFIG_NUMA */
2489 static void reset_alloc_batches(struct zone *preferred_zone)
2491 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2494 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2495 high_wmark_pages(zone) - low_wmark_pages(zone) -
2496 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2497 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2498 } while (zone++ != preferred_zone);
2502 * get_page_from_freelist goes through the zonelist trying to allocate
2505 static struct page *
2506 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2507 const struct alloc_context *ac)
2509 struct zonelist *zonelist = ac->zonelist;
2511 struct page *page = NULL;
2513 int nr_fair_skipped = 0;
2514 bool zonelist_rescan;
2517 zonelist_rescan = false;
2520 * Scan zonelist, looking for a zone with enough free.
2521 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2523 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2527 if (cpusets_enabled() &&
2528 (alloc_flags & ALLOC_CPUSET) &&
2529 !cpuset_zone_allowed(zone, gfp_mask))
2532 * Distribute pages in proportion to the individual
2533 * zone size to ensure fair page aging. The zone a
2534 * page was allocated in should have no effect on the
2535 * time the page has in memory before being reclaimed.
2537 if (alloc_flags & ALLOC_FAIR) {
2538 if (!zone_local(ac->preferred_zone, zone))
2540 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2546 * When allocating a page cache page for writing, we
2547 * want to get it from a zone that is within its dirty
2548 * limit, such that no single zone holds more than its
2549 * proportional share of globally allowed dirty pages.
2550 * The dirty limits take into account the zone's
2551 * lowmem reserves and high watermark so that kswapd
2552 * should be able to balance it without having to
2553 * write pages from its LRU list.
2555 * This may look like it could increase pressure on
2556 * lower zones by failing allocations in higher zones
2557 * before they are full. But the pages that do spill
2558 * over are limited as the lower zones are protected
2559 * by this very same mechanism. It should not become
2560 * a practical burden to them.
2562 * XXX: For now, allow allocations to potentially
2563 * exceed the per-zone dirty limit in the slowpath
2564 * (spread_dirty_pages unset) before going into reclaim,
2565 * which is important when on a NUMA setup the allowed
2566 * zones are together not big enough to reach the
2567 * global limit. The proper fix for these situations
2568 * will require awareness of zones in the
2569 * dirty-throttling and the flusher threads.
2571 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2574 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2575 if (!zone_watermark_ok(zone, order, mark,
2576 ac->classzone_idx, alloc_flags)) {
2579 /* Checked here to keep the fast path fast */
2580 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2581 if (alloc_flags & ALLOC_NO_WATERMARKS)
2584 if (zone_reclaim_mode == 0 ||
2585 !zone_allows_reclaim(ac->preferred_zone, zone))
2588 ret = zone_reclaim(zone, gfp_mask, order);
2590 case ZONE_RECLAIM_NOSCAN:
2593 case ZONE_RECLAIM_FULL:
2594 /* scanned but unreclaimable */
2597 /* did we reclaim enough */
2598 if (zone_watermark_ok(zone, order, mark,
2599 ac->classzone_idx, alloc_flags))
2607 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2608 gfp_mask, alloc_flags, ac->migratetype);
2610 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2614 * If this is a high-order atomic allocation then check
2615 * if the pageblock should be reserved for the future
2617 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2618 reserve_highatomic_pageblock(page, zone, order);
2625 * The first pass makes sure allocations are spread fairly within the
2626 * local node. However, the local node might have free pages left
2627 * after the fairness batches are exhausted, and remote zones haven't
2628 * even been considered yet. Try once more without fairness, and
2629 * include remote zones now, before entering the slowpath and waking
2630 * kswapd: prefer spilling to a remote zone over swapping locally.
2632 if (alloc_flags & ALLOC_FAIR) {
2633 alloc_flags &= ~ALLOC_FAIR;
2634 if (nr_fair_skipped) {
2635 zonelist_rescan = true;
2636 reset_alloc_batches(ac->preferred_zone);
2638 if (nr_online_nodes > 1)
2639 zonelist_rescan = true;
2642 if (zonelist_rescan)
2649 * Large machines with many possible nodes should not always dump per-node
2650 * meminfo in irq context.
2652 static inline bool should_suppress_show_mem(void)
2657 ret = in_interrupt();
2662 static DEFINE_RATELIMIT_STATE(nopage_rs,
2663 DEFAULT_RATELIMIT_INTERVAL,
2664 DEFAULT_RATELIMIT_BURST);
2666 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2668 unsigned int filter = SHOW_MEM_FILTER_NODES;
2670 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2671 debug_guardpage_minorder() > 0)
2675 * This documents exceptions given to allocations in certain
2676 * contexts that are allowed to allocate outside current's set
2679 if (!(gfp_mask & __GFP_NOMEMALLOC))
2680 if (test_thread_flag(TIF_MEMDIE) ||
2681 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2682 filter &= ~SHOW_MEM_FILTER_NODES;
2683 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2684 filter &= ~SHOW_MEM_FILTER_NODES;
2687 struct va_format vaf;
2690 va_start(args, fmt);
2695 pr_warn("%pV", &vaf);
2700 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2701 current->comm, order, gfp_mask);
2704 if (!should_suppress_show_mem())
2708 static inline struct page *
2709 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2710 const struct alloc_context *ac, unsigned long *did_some_progress)
2712 struct oom_control oc = {
2713 .zonelist = ac->zonelist,
2714 .nodemask = ac->nodemask,
2715 .gfp_mask = gfp_mask,
2720 *did_some_progress = 0;
2723 * Acquire the oom lock. If that fails, somebody else is
2724 * making progress for us.
2726 if (!mutex_trylock(&oom_lock)) {
2727 *did_some_progress = 1;
2728 schedule_timeout_uninterruptible(1);
2733 * Go through the zonelist yet one more time, keep very high watermark
2734 * here, this is only to catch a parallel oom killing, we must fail if
2735 * we're still under heavy pressure.
2737 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2738 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2742 if (!(gfp_mask & __GFP_NOFAIL)) {
2743 /* Coredumps can quickly deplete all memory reserves */
2744 if (current->flags & PF_DUMPCORE)
2746 /* The OOM killer will not help higher order allocs */
2747 if (order > PAGE_ALLOC_COSTLY_ORDER)
2749 /* The OOM killer does not needlessly kill tasks for lowmem */
2750 if (ac->high_zoneidx < ZONE_NORMAL)
2752 /* The OOM killer does not compensate for IO-less reclaim */
2753 if (!(gfp_mask & __GFP_FS)) {
2755 * XXX: Page reclaim didn't yield anything,
2756 * and the OOM killer can't be invoked, but
2757 * keep looping as per tradition.
2759 *did_some_progress = 1;
2762 if (pm_suspended_storage())
2764 /* The OOM killer may not free memory on a specific node */
2765 if (gfp_mask & __GFP_THISNODE)
2768 /* Exhausted what can be done so it's blamo time */
2769 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2770 *did_some_progress = 1;
2772 mutex_unlock(&oom_lock);
2776 #ifdef CONFIG_COMPACTION
2777 /* Try memory compaction for high-order allocations before reclaim */
2778 static struct page *
2779 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2780 int alloc_flags, const struct alloc_context *ac,
2781 enum migrate_mode mode, int *contended_compaction,
2782 bool *deferred_compaction)
2784 unsigned long compact_result;
2790 current->flags |= PF_MEMALLOC;
2791 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2792 mode, contended_compaction);
2793 current->flags &= ~PF_MEMALLOC;
2795 switch (compact_result) {
2796 case COMPACT_DEFERRED:
2797 *deferred_compaction = true;
2799 case COMPACT_SKIPPED:
2806 * At least in one zone compaction wasn't deferred or skipped, so let's
2807 * count a compaction stall
2809 count_vm_event(COMPACTSTALL);
2811 page = get_page_from_freelist(gfp_mask, order,
2812 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2815 struct zone *zone = page_zone(page);
2817 zone->compact_blockskip_flush = false;
2818 compaction_defer_reset(zone, order, true);
2819 count_vm_event(COMPACTSUCCESS);
2824 * It's bad if compaction run occurs and fails. The most likely reason
2825 * is that pages exist, but not enough to satisfy watermarks.
2827 count_vm_event(COMPACTFAIL);
2834 static inline struct page *
2835 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2836 int alloc_flags, const struct alloc_context *ac,
2837 enum migrate_mode mode, int *contended_compaction,
2838 bool *deferred_compaction)
2842 #endif /* CONFIG_COMPACTION */
2844 /* Perform direct synchronous page reclaim */
2846 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2847 const struct alloc_context *ac)
2849 struct reclaim_state reclaim_state;
2854 /* We now go into synchronous reclaim */
2855 cpuset_memory_pressure_bump();
2856 current->flags |= PF_MEMALLOC;
2857 lockdep_set_current_reclaim_state(gfp_mask);
2858 reclaim_state.reclaimed_slab = 0;
2859 current->reclaim_state = &reclaim_state;
2861 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2864 current->reclaim_state = NULL;
2865 lockdep_clear_current_reclaim_state();
2866 current->flags &= ~PF_MEMALLOC;
2873 /* The really slow allocator path where we enter direct reclaim */
2874 static inline struct page *
2875 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2876 int alloc_flags, const struct alloc_context *ac,
2877 unsigned long *did_some_progress)
2879 struct page *page = NULL;
2880 bool drained = false;
2882 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2883 if (unlikely(!(*did_some_progress)))
2887 page = get_page_from_freelist(gfp_mask, order,
2888 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2891 * If an allocation failed after direct reclaim, it could be because
2892 * pages are pinned on the per-cpu lists or in high alloc reserves.
2893 * Shrink them them and try again
2895 if (!page && !drained) {
2896 unreserve_highatomic_pageblock(ac);
2897 drain_all_pages(NULL);
2906 * This is called in the allocator slow-path if the allocation request is of
2907 * sufficient urgency to ignore watermarks and take other desperate measures
2909 static inline struct page *
2910 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2911 const struct alloc_context *ac)
2916 page = get_page_from_freelist(gfp_mask, order,
2917 ALLOC_NO_WATERMARKS, ac);
2919 if (!page && gfp_mask & __GFP_NOFAIL)
2920 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2922 } while (!page && (gfp_mask & __GFP_NOFAIL));
2927 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2932 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2933 ac->high_zoneidx, ac->nodemask)
2934 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2938 gfp_to_alloc_flags(gfp_t gfp_mask)
2940 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2942 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2943 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2946 * The caller may dip into page reserves a bit more if the caller
2947 * cannot run direct reclaim, or if the caller has realtime scheduling
2948 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2949 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2951 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2953 if (gfp_mask & __GFP_ATOMIC) {
2955 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2956 * if it can't schedule.
2958 if (!(gfp_mask & __GFP_NOMEMALLOC))
2959 alloc_flags |= ALLOC_HARDER;
2961 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2962 * comment for __cpuset_node_allowed().
2964 alloc_flags &= ~ALLOC_CPUSET;
2965 } else if (unlikely(rt_task(current)) && !in_interrupt())
2966 alloc_flags |= ALLOC_HARDER;
2968 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2969 if (gfp_mask & __GFP_MEMALLOC)
2970 alloc_flags |= ALLOC_NO_WATERMARKS;
2971 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2972 alloc_flags |= ALLOC_NO_WATERMARKS;
2973 else if (!in_interrupt() &&
2974 ((current->flags & PF_MEMALLOC) ||
2975 unlikely(test_thread_flag(TIF_MEMDIE))))
2976 alloc_flags |= ALLOC_NO_WATERMARKS;
2979 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2980 alloc_flags |= ALLOC_CMA;
2985 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2987 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2990 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2992 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2995 static inline struct page *
2996 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2997 struct alloc_context *ac)
2999 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3000 struct page *page = NULL;
3002 unsigned long pages_reclaimed = 0;
3003 unsigned long did_some_progress;
3004 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3005 bool deferred_compaction = false;
3006 int contended_compaction = COMPACT_CONTENDED_NONE;
3009 * In the slowpath, we sanity check order to avoid ever trying to
3010 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3011 * be using allocators in order of preference for an area that is
3014 if (order >= MAX_ORDER) {
3015 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3020 * We also sanity check to catch abuse of atomic reserves being used by
3021 * callers that are not in atomic context.
3023 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3024 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3025 gfp_mask &= ~__GFP_ATOMIC;
3028 * If this allocation cannot block and it is for a specific node, then
3029 * fail early. There's no need to wakeup kswapd or retry for a
3030 * speculative node-specific allocation.
3032 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3036 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3037 wake_all_kswapds(order, ac);
3040 * OK, we're below the kswapd watermark and have kicked background
3041 * reclaim. Now things get more complex, so set up alloc_flags according
3042 * to how we want to proceed.
3044 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3047 * Find the true preferred zone if the allocation is unconstrained by
3050 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3051 struct zoneref *preferred_zoneref;
3052 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3053 ac->high_zoneidx, NULL, &ac->preferred_zone);
3054 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3057 /* This is the last chance, in general, before the goto nopage. */
3058 page = get_page_from_freelist(gfp_mask, order,
3059 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3063 /* Allocate without watermarks if the context allows */
3064 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3066 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3067 * the allocation is high priority and these type of
3068 * allocations are system rather than user orientated
3070 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3072 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3079 /* Caller is not willing to reclaim, we can't balance anything */
3080 if (!can_direct_reclaim) {
3082 * All existing users of the deprecated __GFP_NOFAIL are
3083 * blockable, so warn of any new users that actually allow this
3084 * type of allocation to fail.
3086 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3090 /* Avoid recursion of direct reclaim */
3091 if (current->flags & PF_MEMALLOC)
3094 /* Avoid allocations with no watermarks from looping endlessly */
3095 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3099 * Try direct compaction. The first pass is asynchronous. Subsequent
3100 * attempts after direct reclaim are synchronous
3102 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3104 &contended_compaction,
3105 &deferred_compaction);
3109 /* Checks for THP-specific high-order allocations */
3110 if (is_thp_gfp_mask(gfp_mask)) {
3112 * If compaction is deferred for high-order allocations, it is
3113 * because sync compaction recently failed. If this is the case
3114 * and the caller requested a THP allocation, we do not want
3115 * to heavily disrupt the system, so we fail the allocation
3116 * instead of entering direct reclaim.
3118 if (deferred_compaction)
3122 * In all zones where compaction was attempted (and not
3123 * deferred or skipped), lock contention has been detected.
3124 * For THP allocation we do not want to disrupt the others
3125 * so we fallback to base pages instead.
3127 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3131 * If compaction was aborted due to need_resched(), we do not
3132 * want to further increase allocation latency, unless it is
3133 * khugepaged trying to collapse.
3135 if (contended_compaction == COMPACT_CONTENDED_SCHED
3136 && !(current->flags & PF_KTHREAD))
3141 * It can become very expensive to allocate transparent hugepages at
3142 * fault, so use asynchronous memory compaction for THP unless it is
3143 * khugepaged trying to collapse.
3145 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3146 migration_mode = MIGRATE_SYNC_LIGHT;
3148 /* Try direct reclaim and then allocating */
3149 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3150 &did_some_progress);
3154 /* Do not loop if specifically requested */
3155 if (gfp_mask & __GFP_NORETRY)
3158 /* Keep reclaiming pages as long as there is reasonable progress */
3159 pages_reclaimed += did_some_progress;
3160 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3161 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3162 /* Wait for some write requests to complete then retry */
3163 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3167 /* Reclaim has failed us, start killing things */
3168 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3172 /* Retry as long as the OOM killer is making progress */
3173 if (did_some_progress)
3178 * High-order allocations do not necessarily loop after
3179 * direct reclaim and reclaim/compaction depends on compaction
3180 * being called after reclaim so call directly if necessary
3182 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3184 &contended_compaction,
3185 &deferred_compaction);
3189 warn_alloc_failed(gfp_mask, order, NULL);
3195 * This is the 'heart' of the zoned buddy allocator.
3198 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3199 struct zonelist *zonelist, nodemask_t *nodemask)
3201 struct zoneref *preferred_zoneref;
3202 struct page *page = NULL;
3203 unsigned int cpuset_mems_cookie;
3204 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3205 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3206 struct alloc_context ac = {
3207 .high_zoneidx = gfp_zone(gfp_mask),
3208 .nodemask = nodemask,
3209 .migratetype = gfpflags_to_migratetype(gfp_mask),
3212 gfp_mask &= gfp_allowed_mask;
3214 lockdep_trace_alloc(gfp_mask);
3216 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3218 if (should_fail_alloc_page(gfp_mask, order))
3222 * Check the zones suitable for the gfp_mask contain at least one
3223 * valid zone. It's possible to have an empty zonelist as a result
3224 * of __GFP_THISNODE and a memoryless node
3226 if (unlikely(!zonelist->_zonerefs->zone))
3229 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3230 alloc_flags |= ALLOC_CMA;
3233 cpuset_mems_cookie = read_mems_allowed_begin();
3235 /* We set it here, as __alloc_pages_slowpath might have changed it */
3236 ac.zonelist = zonelist;
3238 /* Dirty zone balancing only done in the fast path */
3239 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3241 /* The preferred zone is used for statistics later */
3242 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3243 ac.nodemask ? : &cpuset_current_mems_allowed,
3244 &ac.preferred_zone);
3245 if (!ac.preferred_zone)
3247 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3249 /* First allocation attempt */
3250 alloc_mask = gfp_mask|__GFP_HARDWALL;
3251 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3252 if (unlikely(!page)) {
3254 * Runtime PM, block IO and its error handling path
3255 * can deadlock because I/O on the device might not
3258 alloc_mask = memalloc_noio_flags(gfp_mask);
3259 ac.spread_dirty_pages = false;
3261 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3264 if (kmemcheck_enabled && page)
3265 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3267 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3271 * When updating a task's mems_allowed, it is possible to race with
3272 * parallel threads in such a way that an allocation can fail while
3273 * the mask is being updated. If a page allocation is about to fail,
3274 * check if the cpuset changed during allocation and if so, retry.
3276 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3281 EXPORT_SYMBOL(__alloc_pages_nodemask);
3284 * Common helper functions.
3286 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3291 * __get_free_pages() returns a 32-bit address, which cannot represent
3294 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3296 page = alloc_pages(gfp_mask, order);
3299 return (unsigned long) page_address(page);
3301 EXPORT_SYMBOL(__get_free_pages);
3303 unsigned long get_zeroed_page(gfp_t gfp_mask)
3305 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3307 EXPORT_SYMBOL(get_zeroed_page);
3309 void __free_pages(struct page *page, unsigned int order)
3311 if (put_page_testzero(page)) {
3313 free_hot_cold_page(page, false);
3315 __free_pages_ok(page, order);
3319 EXPORT_SYMBOL(__free_pages);
3321 void free_pages(unsigned long addr, unsigned int order)
3324 VM_BUG_ON(!virt_addr_valid((void *)addr));
3325 __free_pages(virt_to_page((void *)addr), order);
3329 EXPORT_SYMBOL(free_pages);
3333 * An arbitrary-length arbitrary-offset area of memory which resides
3334 * within a 0 or higher order page. Multiple fragments within that page
3335 * are individually refcounted, in the page's reference counter.
3337 * The page_frag functions below provide a simple allocation framework for
3338 * page fragments. This is used by the network stack and network device
3339 * drivers to provide a backing region of memory for use as either an
3340 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3342 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3345 struct page *page = NULL;
3346 gfp_t gfp = gfp_mask;
3348 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3349 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3351 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3352 PAGE_FRAG_CACHE_MAX_ORDER);
3353 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3355 if (unlikely(!page))
3356 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3358 nc->va = page ? page_address(page) : NULL;
3363 void *__alloc_page_frag(struct page_frag_cache *nc,
3364 unsigned int fragsz, gfp_t gfp_mask)
3366 unsigned int size = PAGE_SIZE;
3370 if (unlikely(!nc->va)) {
3372 page = __page_frag_refill(nc, gfp_mask);
3376 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3377 /* if size can vary use size else just use PAGE_SIZE */
3380 /* Even if we own the page, we do not use atomic_set().
3381 * This would break get_page_unless_zero() users.
3383 atomic_add(size - 1, &page->_count);
3385 /* reset page count bias and offset to start of new frag */
3386 nc->pfmemalloc = page_is_pfmemalloc(page);
3387 nc->pagecnt_bias = size;
3391 offset = nc->offset - fragsz;
3392 if (unlikely(offset < 0)) {
3393 page = virt_to_page(nc->va);
3395 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3398 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3399 /* if size can vary use size else just use PAGE_SIZE */
3402 /* OK, page count is 0, we can safely set it */
3403 atomic_set(&page->_count, size);
3405 /* reset page count bias and offset to start of new frag */
3406 nc->pagecnt_bias = size;
3407 offset = size - fragsz;
3411 nc->offset = offset;
3413 return nc->va + offset;
3415 EXPORT_SYMBOL(__alloc_page_frag);
3418 * Frees a page fragment allocated out of either a compound or order 0 page.
3420 void __free_page_frag(void *addr)
3422 struct page *page = virt_to_head_page(addr);
3424 if (unlikely(put_page_testzero(page)))
3425 __free_pages_ok(page, compound_order(page));
3427 EXPORT_SYMBOL(__free_page_frag);
3430 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3431 * of the current memory cgroup.
3433 * It should be used when the caller would like to use kmalloc, but since the
3434 * allocation is large, it has to fall back to the page allocator.
3436 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3440 page = alloc_pages(gfp_mask, order);
3441 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3442 __free_pages(page, order);
3448 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3452 page = alloc_pages_node(nid, gfp_mask, order);
3453 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3454 __free_pages(page, order);
3461 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3464 void __free_kmem_pages(struct page *page, unsigned int order)
3466 memcg_kmem_uncharge(page, order);
3467 __free_pages(page, order);
3470 void free_kmem_pages(unsigned long addr, unsigned int order)
3473 VM_BUG_ON(!virt_addr_valid((void *)addr));
3474 __free_kmem_pages(virt_to_page((void *)addr), order);
3478 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3482 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3483 unsigned long used = addr + PAGE_ALIGN(size);
3485 split_page(virt_to_page((void *)addr), order);
3486 while (used < alloc_end) {
3491 return (void *)addr;
3495 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3496 * @size: the number of bytes to allocate
3497 * @gfp_mask: GFP flags for the allocation
3499 * This function is similar to alloc_pages(), except that it allocates the
3500 * minimum number of pages to satisfy the request. alloc_pages() can only
3501 * allocate memory in power-of-two pages.
3503 * This function is also limited by MAX_ORDER.
3505 * Memory allocated by this function must be released by free_pages_exact().
3507 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3509 unsigned int order = get_order(size);
3512 addr = __get_free_pages(gfp_mask, order);
3513 return make_alloc_exact(addr, order, size);
3515 EXPORT_SYMBOL(alloc_pages_exact);
3518 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3520 * @nid: the preferred node ID where memory should be allocated
3521 * @size: the number of bytes to allocate
3522 * @gfp_mask: GFP flags for the allocation
3524 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3527 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3529 unsigned int order = get_order(size);
3530 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3533 return make_alloc_exact((unsigned long)page_address(p), order, size);
3537 * free_pages_exact - release memory allocated via alloc_pages_exact()
3538 * @virt: the value returned by alloc_pages_exact.
3539 * @size: size of allocation, same value as passed to alloc_pages_exact().
3541 * Release the memory allocated by a previous call to alloc_pages_exact.
3543 void free_pages_exact(void *virt, size_t size)
3545 unsigned long addr = (unsigned long)virt;
3546 unsigned long end = addr + PAGE_ALIGN(size);
3548 while (addr < end) {
3553 EXPORT_SYMBOL(free_pages_exact);
3556 * nr_free_zone_pages - count number of pages beyond high watermark
3557 * @offset: The zone index of the highest zone
3559 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3560 * high watermark within all zones at or below a given zone index. For each
3561 * zone, the number of pages is calculated as:
3562 * managed_pages - high_pages
3564 static unsigned long nr_free_zone_pages(int offset)
3569 /* Just pick one node, since fallback list is circular */
3570 unsigned long sum = 0;
3572 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3574 for_each_zone_zonelist(zone, z, zonelist, offset) {
3575 unsigned long size = zone->managed_pages;
3576 unsigned long high = high_wmark_pages(zone);
3585 * nr_free_buffer_pages - count number of pages beyond high watermark
3587 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3588 * watermark within ZONE_DMA and ZONE_NORMAL.
3590 unsigned long nr_free_buffer_pages(void)
3592 return nr_free_zone_pages(gfp_zone(GFP_USER));
3594 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3597 * nr_free_pagecache_pages - count number of pages beyond high watermark
3599 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3600 * high watermark within all zones.
3602 unsigned long nr_free_pagecache_pages(void)
3604 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3607 static inline void show_node(struct zone *zone)
3609 if (IS_ENABLED(CONFIG_NUMA))
3610 printk("Node %d ", zone_to_nid(zone));
3613 void si_meminfo(struct sysinfo *val)
3615 val->totalram = totalram_pages;
3616 val->sharedram = global_page_state(NR_SHMEM);
3617 val->freeram = global_page_state(NR_FREE_PAGES);
3618 val->bufferram = nr_blockdev_pages();
3619 val->totalhigh = totalhigh_pages;
3620 val->freehigh = nr_free_highpages();
3621 val->mem_unit = PAGE_SIZE;
3624 EXPORT_SYMBOL(si_meminfo);
3627 void si_meminfo_node(struct sysinfo *val, int nid)
3629 int zone_type; /* needs to be signed */
3630 unsigned long managed_pages = 0;
3631 pg_data_t *pgdat = NODE_DATA(nid);
3633 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3634 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3635 val->totalram = managed_pages;
3636 val->sharedram = node_page_state(nid, NR_SHMEM);
3637 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3638 #ifdef CONFIG_HIGHMEM
3639 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3640 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3646 val->mem_unit = PAGE_SIZE;
3651 * Determine whether the node should be displayed or not, depending on whether
3652 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3654 bool skip_free_areas_node(unsigned int flags, int nid)
3657 unsigned int cpuset_mems_cookie;
3659 if (!(flags & SHOW_MEM_FILTER_NODES))
3663 cpuset_mems_cookie = read_mems_allowed_begin();
3664 ret = !node_isset(nid, cpuset_current_mems_allowed);
3665 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3670 #define K(x) ((x) << (PAGE_SHIFT-10))
3672 static void show_migration_types(unsigned char type)
3674 static const char types[MIGRATE_TYPES] = {
3675 [MIGRATE_UNMOVABLE] = 'U',
3676 [MIGRATE_RECLAIMABLE] = 'E',
3677 [MIGRATE_MOVABLE] = 'M',
3679 [MIGRATE_CMA] = 'C',
3681 #ifdef CONFIG_MEMORY_ISOLATION
3682 [MIGRATE_ISOLATE] = 'I',
3685 char tmp[MIGRATE_TYPES + 1];
3689 for (i = 0; i < MIGRATE_TYPES; i++) {
3690 if (type & (1 << i))
3695 printk("(%s) ", tmp);
3699 * Show free area list (used inside shift_scroll-lock stuff)
3700 * We also calculate the percentage fragmentation. We do this by counting the
3701 * memory on each free list with the exception of the first item on the list.
3704 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3707 void show_free_areas(unsigned int filter)
3709 unsigned long free_pcp = 0;
3713 for_each_populated_zone(zone) {
3714 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3717 for_each_online_cpu(cpu)
3718 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3721 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3722 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3723 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3724 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3725 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3726 " free:%lu free_pcp:%lu free_cma:%lu\n",
3727 global_page_state(NR_ACTIVE_ANON),
3728 global_page_state(NR_INACTIVE_ANON),
3729 global_page_state(NR_ISOLATED_ANON),
3730 global_page_state(NR_ACTIVE_FILE),
3731 global_page_state(NR_INACTIVE_FILE),
3732 global_page_state(NR_ISOLATED_FILE),
3733 global_page_state(NR_UNEVICTABLE),
3734 global_page_state(NR_FILE_DIRTY),
3735 global_page_state(NR_WRITEBACK),
3736 global_page_state(NR_UNSTABLE_NFS),
3737 global_page_state(NR_SLAB_RECLAIMABLE),
3738 global_page_state(NR_SLAB_UNRECLAIMABLE),
3739 global_page_state(NR_FILE_MAPPED),
3740 global_page_state(NR_SHMEM),
3741 global_page_state(NR_PAGETABLE),
3742 global_page_state(NR_BOUNCE),
3743 global_page_state(NR_FREE_PAGES),
3745 global_page_state(NR_FREE_CMA_PAGES));
3747 for_each_populated_zone(zone) {
3750 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3754 for_each_online_cpu(cpu)
3755 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3763 " active_anon:%lukB"
3764 " inactive_anon:%lukB"
3765 " active_file:%lukB"
3766 " inactive_file:%lukB"
3767 " unevictable:%lukB"
3768 " isolated(anon):%lukB"
3769 " isolated(file):%lukB"
3777 " slab_reclaimable:%lukB"
3778 " slab_unreclaimable:%lukB"
3779 " kernel_stack:%lukB"
3786 " writeback_tmp:%lukB"
3787 " pages_scanned:%lu"
3788 " all_unreclaimable? %s"
3791 K(zone_page_state(zone, NR_FREE_PAGES)),
3792 K(min_wmark_pages(zone)),
3793 K(low_wmark_pages(zone)),
3794 K(high_wmark_pages(zone)),
3795 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3796 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3797 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3798 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3799 K(zone_page_state(zone, NR_UNEVICTABLE)),
3800 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3801 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3802 K(zone->present_pages),
3803 K(zone->managed_pages),
3804 K(zone_page_state(zone, NR_MLOCK)),
3805 K(zone_page_state(zone, NR_FILE_DIRTY)),
3806 K(zone_page_state(zone, NR_WRITEBACK)),
3807 K(zone_page_state(zone, NR_FILE_MAPPED)),
3808 K(zone_page_state(zone, NR_SHMEM)),
3809 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3810 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3811 zone_page_state(zone, NR_KERNEL_STACK) *
3813 K(zone_page_state(zone, NR_PAGETABLE)),
3814 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3815 K(zone_page_state(zone, NR_BOUNCE)),
3817 K(this_cpu_read(zone->pageset->pcp.count)),
3818 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3819 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3820 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3821 (!zone_reclaimable(zone) ? "yes" : "no")
3823 printk("lowmem_reserve[]:");
3824 for (i = 0; i < MAX_NR_ZONES; i++)
3825 printk(" %ld", zone->lowmem_reserve[i]);
3829 for_each_populated_zone(zone) {
3831 unsigned long nr[MAX_ORDER], flags, total = 0;
3832 unsigned char types[MAX_ORDER];
3834 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3837 printk("%s: ", zone->name);
3839 spin_lock_irqsave(&zone->lock, flags);
3840 for (order = 0; order < MAX_ORDER; order++) {
3841 struct free_area *area = &zone->free_area[order];
3844 nr[order] = area->nr_free;
3845 total += nr[order] << order;
3848 for (type = 0; type < MIGRATE_TYPES; type++) {
3849 if (!list_empty(&area->free_list[type]))
3850 types[order] |= 1 << type;
3853 spin_unlock_irqrestore(&zone->lock, flags);
3854 for (order = 0; order < MAX_ORDER; order++) {
3855 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3857 show_migration_types(types[order]);
3859 printk("= %lukB\n", K(total));
3862 hugetlb_show_meminfo();
3864 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3866 show_swap_cache_info();
3869 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3871 zoneref->zone = zone;
3872 zoneref->zone_idx = zone_idx(zone);
3876 * Builds allocation fallback zone lists.
3878 * Add all populated zones of a node to the zonelist.
3880 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3884 enum zone_type zone_type = MAX_NR_ZONES;
3888 zone = pgdat->node_zones + zone_type;
3889 if (populated_zone(zone)) {
3890 zoneref_set_zone(zone,
3891 &zonelist->_zonerefs[nr_zones++]);
3892 check_highest_zone(zone_type);
3894 } while (zone_type);
3902 * 0 = automatic detection of better ordering.
3903 * 1 = order by ([node] distance, -zonetype)
3904 * 2 = order by (-zonetype, [node] distance)
3906 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3907 * the same zonelist. So only NUMA can configure this param.
3909 #define ZONELIST_ORDER_DEFAULT 0
3910 #define ZONELIST_ORDER_NODE 1
3911 #define ZONELIST_ORDER_ZONE 2
3913 /* zonelist order in the kernel.
3914 * set_zonelist_order() will set this to NODE or ZONE.
3916 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3917 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3921 /* The value user specified ....changed by config */
3922 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3923 /* string for sysctl */
3924 #define NUMA_ZONELIST_ORDER_LEN 16
3925 char numa_zonelist_order[16] = "default";
3928 * interface for configure zonelist ordering.
3929 * command line option "numa_zonelist_order"
3930 * = "[dD]efault - default, automatic configuration.
3931 * = "[nN]ode - order by node locality, then by zone within node
3932 * = "[zZ]one - order by zone, then by locality within zone
3935 static int __parse_numa_zonelist_order(char *s)
3937 if (*s == 'd' || *s == 'D') {
3938 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3939 } else if (*s == 'n' || *s == 'N') {
3940 user_zonelist_order = ZONELIST_ORDER_NODE;
3941 } else if (*s == 'z' || *s == 'Z') {
3942 user_zonelist_order = ZONELIST_ORDER_ZONE;
3945 "Ignoring invalid numa_zonelist_order value: "
3952 static __init int setup_numa_zonelist_order(char *s)
3959 ret = __parse_numa_zonelist_order(s);
3961 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3965 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3968 * sysctl handler for numa_zonelist_order
3970 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3971 void __user *buffer, size_t *length,
3974 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3976 static DEFINE_MUTEX(zl_order_mutex);
3978 mutex_lock(&zl_order_mutex);
3980 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3984 strcpy(saved_string, (char *)table->data);
3986 ret = proc_dostring(table, write, buffer, length, ppos);
3990 int oldval = user_zonelist_order;
3992 ret = __parse_numa_zonelist_order((char *)table->data);
3995 * bogus value. restore saved string
3997 strncpy((char *)table->data, saved_string,
3998 NUMA_ZONELIST_ORDER_LEN);
3999 user_zonelist_order = oldval;
4000 } else if (oldval != user_zonelist_order) {
4001 mutex_lock(&zonelists_mutex);
4002 build_all_zonelists(NULL, NULL);
4003 mutex_unlock(&zonelists_mutex);
4007 mutex_unlock(&zl_order_mutex);
4012 #define MAX_NODE_LOAD (nr_online_nodes)
4013 static int node_load[MAX_NUMNODES];
4016 * find_next_best_node - find the next node that should appear in a given node's fallback list
4017 * @node: node whose fallback list we're appending
4018 * @used_node_mask: nodemask_t of already used nodes
4020 * We use a number of factors to determine which is the next node that should
4021 * appear on a given node's fallback list. The node should not have appeared
4022 * already in @node's fallback list, and it should be the next closest node
4023 * according to the distance array (which contains arbitrary distance values
4024 * from each node to each node in the system), and should also prefer nodes
4025 * with no CPUs, since presumably they'll have very little allocation pressure
4026 * on them otherwise.
4027 * It returns -1 if no node is found.
4029 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4032 int min_val = INT_MAX;
4033 int best_node = NUMA_NO_NODE;
4034 const struct cpumask *tmp = cpumask_of_node(0);
4036 /* Use the local node if we haven't already */
4037 if (!node_isset(node, *used_node_mask)) {
4038 node_set(node, *used_node_mask);
4042 for_each_node_state(n, N_MEMORY) {
4044 /* Don't want a node to appear more than once */
4045 if (node_isset(n, *used_node_mask))
4048 /* Use the distance array to find the distance */
4049 val = node_distance(node, n);
4051 /* Penalize nodes under us ("prefer the next node") */
4054 /* Give preference to headless and unused nodes */
4055 tmp = cpumask_of_node(n);
4056 if (!cpumask_empty(tmp))
4057 val += PENALTY_FOR_NODE_WITH_CPUS;
4059 /* Slight preference for less loaded node */
4060 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4061 val += node_load[n];
4063 if (val < min_val) {
4070 node_set(best_node, *used_node_mask);
4077 * Build zonelists ordered by node and zones within node.
4078 * This results in maximum locality--normal zone overflows into local
4079 * DMA zone, if any--but risks exhausting DMA zone.
4081 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4084 struct zonelist *zonelist;
4086 zonelist = &pgdat->node_zonelists[0];
4087 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4089 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4090 zonelist->_zonerefs[j].zone = NULL;
4091 zonelist->_zonerefs[j].zone_idx = 0;
4095 * Build gfp_thisnode zonelists
4097 static void build_thisnode_zonelists(pg_data_t *pgdat)
4100 struct zonelist *zonelist;
4102 zonelist = &pgdat->node_zonelists[1];
4103 j = build_zonelists_node(pgdat, zonelist, 0);
4104 zonelist->_zonerefs[j].zone = NULL;
4105 zonelist->_zonerefs[j].zone_idx = 0;
4109 * Build zonelists ordered by zone and nodes within zones.
4110 * This results in conserving DMA zone[s] until all Normal memory is
4111 * exhausted, but results in overflowing to remote node while memory
4112 * may still exist in local DMA zone.
4114 static int node_order[MAX_NUMNODES];
4116 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4119 int zone_type; /* needs to be signed */
4121 struct zonelist *zonelist;
4123 zonelist = &pgdat->node_zonelists[0];
4125 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4126 for (j = 0; j < nr_nodes; j++) {
4127 node = node_order[j];
4128 z = &NODE_DATA(node)->node_zones[zone_type];
4129 if (populated_zone(z)) {
4131 &zonelist->_zonerefs[pos++]);
4132 check_highest_zone(zone_type);
4136 zonelist->_zonerefs[pos].zone = NULL;
4137 zonelist->_zonerefs[pos].zone_idx = 0;
4140 #if defined(CONFIG_64BIT)
4142 * Devices that require DMA32/DMA are relatively rare and do not justify a
4143 * penalty to every machine in case the specialised case applies. Default
4144 * to Node-ordering on 64-bit NUMA machines
4146 static int default_zonelist_order(void)
4148 return ZONELIST_ORDER_NODE;
4152 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4153 * by the kernel. If processes running on node 0 deplete the low memory zone
4154 * then reclaim will occur more frequency increasing stalls and potentially
4155 * be easier to OOM if a large percentage of the zone is under writeback or
4156 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4157 * Hence, default to zone ordering on 32-bit.
4159 static int default_zonelist_order(void)
4161 return ZONELIST_ORDER_ZONE;
4163 #endif /* CONFIG_64BIT */
4165 static void set_zonelist_order(void)
4167 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4168 current_zonelist_order = default_zonelist_order();
4170 current_zonelist_order = user_zonelist_order;
4173 static void build_zonelists(pg_data_t *pgdat)
4177 nodemask_t used_mask;
4178 int local_node, prev_node;
4179 struct zonelist *zonelist;
4180 unsigned int order = current_zonelist_order;
4182 /* initialize zonelists */
4183 for (i = 0; i < MAX_ZONELISTS; i++) {
4184 zonelist = pgdat->node_zonelists + i;
4185 zonelist->_zonerefs[0].zone = NULL;
4186 zonelist->_zonerefs[0].zone_idx = 0;
4189 /* NUMA-aware ordering of nodes */
4190 local_node = pgdat->node_id;
4191 load = nr_online_nodes;
4192 prev_node = local_node;
4193 nodes_clear(used_mask);
4195 memset(node_order, 0, sizeof(node_order));
4198 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4200 * We don't want to pressure a particular node.
4201 * So adding penalty to the first node in same
4202 * distance group to make it round-robin.
4204 if (node_distance(local_node, node) !=
4205 node_distance(local_node, prev_node))
4206 node_load[node] = load;
4210 if (order == ZONELIST_ORDER_NODE)
4211 build_zonelists_in_node_order(pgdat, node);
4213 node_order[j++] = node; /* remember order */
4216 if (order == ZONELIST_ORDER_ZONE) {
4217 /* calculate node order -- i.e., DMA last! */
4218 build_zonelists_in_zone_order(pgdat, j);
4221 build_thisnode_zonelists(pgdat);
4224 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4226 * Return node id of node used for "local" allocations.
4227 * I.e., first node id of first zone in arg node's generic zonelist.
4228 * Used for initializing percpu 'numa_mem', which is used primarily
4229 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4231 int local_memory_node(int node)
4235 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4236 gfp_zone(GFP_KERNEL),
4243 #else /* CONFIG_NUMA */
4245 static void set_zonelist_order(void)
4247 current_zonelist_order = ZONELIST_ORDER_ZONE;
4250 static void build_zonelists(pg_data_t *pgdat)
4252 int node, local_node;
4254 struct zonelist *zonelist;
4256 local_node = pgdat->node_id;
4258 zonelist = &pgdat->node_zonelists[0];
4259 j = build_zonelists_node(pgdat, zonelist, 0);
4262 * Now we build the zonelist so that it contains the zones
4263 * of all the other nodes.
4264 * We don't want to pressure a particular node, so when
4265 * building the zones for node N, we make sure that the
4266 * zones coming right after the local ones are those from
4267 * node N+1 (modulo N)
4269 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4270 if (!node_online(node))
4272 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4274 for (node = 0; node < local_node; node++) {
4275 if (!node_online(node))
4277 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4280 zonelist->_zonerefs[j].zone = NULL;
4281 zonelist->_zonerefs[j].zone_idx = 0;
4284 #endif /* CONFIG_NUMA */
4287 * Boot pageset table. One per cpu which is going to be used for all
4288 * zones and all nodes. The parameters will be set in such a way
4289 * that an item put on a list will immediately be handed over to
4290 * the buddy list. This is safe since pageset manipulation is done
4291 * with interrupts disabled.
4293 * The boot_pagesets must be kept even after bootup is complete for
4294 * unused processors and/or zones. They do play a role for bootstrapping
4295 * hotplugged processors.
4297 * zoneinfo_show() and maybe other functions do
4298 * not check if the processor is online before following the pageset pointer.
4299 * Other parts of the kernel may not check if the zone is available.
4301 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4302 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4303 static void setup_zone_pageset(struct zone *zone);
4306 * Global mutex to protect against size modification of zonelists
4307 * as well as to serialize pageset setup for the new populated zone.
4309 DEFINE_MUTEX(zonelists_mutex);
4311 /* return values int ....just for stop_machine() */
4312 static int __build_all_zonelists(void *data)
4316 pg_data_t *self = data;
4319 memset(node_load, 0, sizeof(node_load));
4322 if (self && !node_online(self->node_id)) {
4323 build_zonelists(self);
4326 for_each_online_node(nid) {
4327 pg_data_t *pgdat = NODE_DATA(nid);
4329 build_zonelists(pgdat);
4333 * Initialize the boot_pagesets that are going to be used
4334 * for bootstrapping processors. The real pagesets for
4335 * each zone will be allocated later when the per cpu
4336 * allocator is available.
4338 * boot_pagesets are used also for bootstrapping offline
4339 * cpus if the system is already booted because the pagesets
4340 * are needed to initialize allocators on a specific cpu too.
4341 * F.e. the percpu allocator needs the page allocator which
4342 * needs the percpu allocator in order to allocate its pagesets
4343 * (a chicken-egg dilemma).
4345 for_each_possible_cpu(cpu) {
4346 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4348 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4350 * We now know the "local memory node" for each node--
4351 * i.e., the node of the first zone in the generic zonelist.
4352 * Set up numa_mem percpu variable for all possible cpus
4353 * if associated node has been onlined.
4355 if (node_online(cpu_to_node(cpu)))
4356 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4358 set_cpu_numa_mem(cpu, NUMA_NO_NODE);
4365 static noinline void __init
4366 build_all_zonelists_init(void)
4368 __build_all_zonelists(NULL);
4369 mminit_verify_zonelist();
4370 cpuset_init_current_mems_allowed();
4374 * Called with zonelists_mutex held always
4375 * unless system_state == SYSTEM_BOOTING.
4377 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4378 * [we're only called with non-NULL zone through __meminit paths] and
4379 * (2) call of __init annotated helper build_all_zonelists_init
4380 * [protected by SYSTEM_BOOTING].
4382 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4384 set_zonelist_order();
4386 if (system_state == SYSTEM_BOOTING) {
4387 build_all_zonelists_init();
4389 #ifdef CONFIG_MEMORY_HOTPLUG
4391 setup_zone_pageset(zone);
4393 /* we have to stop all cpus to guarantee there is no user
4395 stop_machine(__build_all_zonelists, pgdat, NULL);
4396 /* cpuset refresh routine should be here */
4398 vm_total_pages = nr_free_pagecache_pages();
4400 * Disable grouping by mobility if the number of pages in the
4401 * system is too low to allow the mechanism to work. It would be
4402 * more accurate, but expensive to check per-zone. This check is
4403 * made on memory-hotadd so a system can start with mobility
4404 * disabled and enable it later
4406 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4407 page_group_by_mobility_disabled = 1;
4409 page_group_by_mobility_disabled = 0;
4411 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4412 "Total pages: %ld\n",
4414 zonelist_order_name[current_zonelist_order],
4415 page_group_by_mobility_disabled ? "off" : "on",
4418 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4423 * Helper functions to size the waitqueue hash table.
4424 * Essentially these want to choose hash table sizes sufficiently
4425 * large so that collisions trying to wait on pages are rare.
4426 * But in fact, the number of active page waitqueues on typical
4427 * systems is ridiculously low, less than 200. So this is even
4428 * conservative, even though it seems large.
4430 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4431 * waitqueues, i.e. the size of the waitq table given the number of pages.
4433 #define PAGES_PER_WAITQUEUE 256
4435 #ifndef CONFIG_MEMORY_HOTPLUG
4436 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4438 unsigned long size = 1;
4440 pages /= PAGES_PER_WAITQUEUE;
4442 while (size < pages)
4446 * Once we have dozens or even hundreds of threads sleeping
4447 * on IO we've got bigger problems than wait queue collision.
4448 * Limit the size of the wait table to a reasonable size.
4450 size = min(size, 4096UL);
4452 return max(size, 4UL);
4456 * A zone's size might be changed by hot-add, so it is not possible to determine
4457 * a suitable size for its wait_table. So we use the maximum size now.
4459 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4461 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4462 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4463 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4465 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4466 * or more by the traditional way. (See above). It equals:
4468 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4469 * ia64(16K page size) : = ( 8G + 4M)byte.
4470 * powerpc (64K page size) : = (32G +16M)byte.
4472 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4479 * This is an integer logarithm so that shifts can be used later
4480 * to extract the more random high bits from the multiplicative
4481 * hash function before the remainder is taken.
4483 static inline unsigned long wait_table_bits(unsigned long size)
4489 * Initially all pages are reserved - free ones are freed
4490 * up by free_all_bootmem() once the early boot process is
4491 * done. Non-atomic initialization, single-pass.
4493 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4494 unsigned long start_pfn, enum memmap_context context)
4496 pg_data_t *pgdat = NODE_DATA(nid);
4497 unsigned long end_pfn = start_pfn + size;
4500 unsigned long nr_initialised = 0;
4502 if (highest_memmap_pfn < end_pfn - 1)
4503 highest_memmap_pfn = end_pfn - 1;
4505 z = &pgdat->node_zones[zone];
4506 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4508 * There can be holes in boot-time mem_map[]s
4509 * handed to this function. They do not
4510 * exist on hotplugged memory.
4512 if (context == MEMMAP_EARLY) {
4513 if (!early_pfn_valid(pfn))
4515 if (!early_pfn_in_nid(pfn, nid))
4517 if (!update_defer_init(pgdat, pfn, end_pfn,
4523 * Mark the block movable so that blocks are reserved for
4524 * movable at startup. This will force kernel allocations
4525 * to reserve their blocks rather than leaking throughout
4526 * the address space during boot when many long-lived
4527 * kernel allocations are made.
4529 * bitmap is created for zone's valid pfn range. but memmap
4530 * can be created for invalid pages (for alignment)
4531 * check here not to call set_pageblock_migratetype() against
4534 if (!(pfn & (pageblock_nr_pages - 1))) {
4535 struct page *page = pfn_to_page(pfn);
4537 __init_single_page(page, pfn, zone, nid);
4538 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4540 __init_single_pfn(pfn, zone, nid);
4545 static void __meminit zone_init_free_lists(struct zone *zone)
4547 unsigned int order, t;
4548 for_each_migratetype_order(order, t) {
4549 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4550 zone->free_area[order].nr_free = 0;
4554 #ifndef __HAVE_ARCH_MEMMAP_INIT
4555 #define memmap_init(size, nid, zone, start_pfn) \
4556 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4559 static int zone_batchsize(struct zone *zone)
4565 * The per-cpu-pages pools are set to around 1000th of the
4566 * size of the zone. But no more than 1/2 of a meg.
4568 * OK, so we don't know how big the cache is. So guess.
4570 batch = zone->managed_pages / 1024;
4571 if (batch * PAGE_SIZE > 512 * 1024)
4572 batch = (512 * 1024) / PAGE_SIZE;
4573 batch /= 4; /* We effectively *= 4 below */
4578 * Clamp the batch to a 2^n - 1 value. Having a power
4579 * of 2 value was found to be more likely to have
4580 * suboptimal cache aliasing properties in some cases.
4582 * For example if 2 tasks are alternately allocating
4583 * batches of pages, one task can end up with a lot
4584 * of pages of one half of the possible page colors
4585 * and the other with pages of the other colors.
4587 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4592 /* The deferral and batching of frees should be suppressed under NOMMU
4595 * The problem is that NOMMU needs to be able to allocate large chunks
4596 * of contiguous memory as there's no hardware page translation to
4597 * assemble apparent contiguous memory from discontiguous pages.
4599 * Queueing large contiguous runs of pages for batching, however,
4600 * causes the pages to actually be freed in smaller chunks. As there
4601 * can be a significant delay between the individual batches being
4602 * recycled, this leads to the once large chunks of space being
4603 * fragmented and becoming unavailable for high-order allocations.
4610 * pcp->high and pcp->batch values are related and dependent on one another:
4611 * ->batch must never be higher then ->high.
4612 * The following function updates them in a safe manner without read side
4615 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4616 * those fields changing asynchronously (acording the the above rule).
4618 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4619 * outside of boot time (or some other assurance that no concurrent updaters
4622 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4623 unsigned long batch)
4625 /* start with a fail safe value for batch */
4629 /* Update high, then batch, in order */
4636 /* a companion to pageset_set_high() */
4637 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4639 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4642 static void pageset_init(struct per_cpu_pageset *p)
4644 struct per_cpu_pages *pcp;
4647 memset(p, 0, sizeof(*p));
4651 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4652 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4655 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4658 pageset_set_batch(p, batch);
4662 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4663 * to the value high for the pageset p.
4665 static void pageset_set_high(struct per_cpu_pageset *p,
4668 unsigned long batch = max(1UL, high / 4);
4669 if ((high / 4) > (PAGE_SHIFT * 8))
4670 batch = PAGE_SHIFT * 8;
4672 pageset_update(&p->pcp, high, batch);
4675 static void pageset_set_high_and_batch(struct zone *zone,
4676 struct per_cpu_pageset *pcp)
4678 if (percpu_pagelist_fraction)
4679 pageset_set_high(pcp,
4680 (zone->managed_pages /
4681 percpu_pagelist_fraction));
4683 pageset_set_batch(pcp, zone_batchsize(zone));
4686 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4688 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4691 pageset_set_high_and_batch(zone, pcp);
4694 static void __meminit setup_zone_pageset(struct zone *zone)
4697 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4698 for_each_possible_cpu(cpu)
4699 zone_pageset_init(zone, cpu);
4703 * Allocate per cpu pagesets and initialize them.
4704 * Before this call only boot pagesets were available.
4706 void __init setup_per_cpu_pageset(void)
4710 for_each_populated_zone(zone)
4711 setup_zone_pageset(zone);
4714 static noinline __init_refok
4715 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4721 * The per-page waitqueue mechanism uses hashed waitqueues
4724 zone->wait_table_hash_nr_entries =
4725 wait_table_hash_nr_entries(zone_size_pages);
4726 zone->wait_table_bits =
4727 wait_table_bits(zone->wait_table_hash_nr_entries);
4728 alloc_size = zone->wait_table_hash_nr_entries
4729 * sizeof(wait_queue_head_t);
4731 if (!slab_is_available()) {
4732 zone->wait_table = (wait_queue_head_t *)
4733 memblock_virt_alloc_node_nopanic(
4734 alloc_size, zone->zone_pgdat->node_id);
4737 * This case means that a zone whose size was 0 gets new memory
4738 * via memory hot-add.
4739 * But it may be the case that a new node was hot-added. In
4740 * this case vmalloc() will not be able to use this new node's
4741 * memory - this wait_table must be initialized to use this new
4742 * node itself as well.
4743 * To use this new node's memory, further consideration will be
4746 zone->wait_table = vmalloc(alloc_size);
4748 if (!zone->wait_table)
4751 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4752 init_waitqueue_head(zone->wait_table + i);
4757 static __meminit void zone_pcp_init(struct zone *zone)
4760 * per cpu subsystem is not up at this point. The following code
4761 * relies on the ability of the linker to provide the
4762 * offset of a (static) per cpu variable into the per cpu area.
4764 zone->pageset = &boot_pageset;
4766 if (populated_zone(zone))
4767 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4768 zone->name, zone->present_pages,
4769 zone_batchsize(zone));
4772 int __meminit init_currently_empty_zone(struct zone *zone,
4773 unsigned long zone_start_pfn,
4776 struct pglist_data *pgdat = zone->zone_pgdat;
4778 ret = zone_wait_table_init(zone, size);
4781 pgdat->nr_zones = zone_idx(zone) + 1;
4783 zone->zone_start_pfn = zone_start_pfn;
4785 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4786 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4788 (unsigned long)zone_idx(zone),
4789 zone_start_pfn, (zone_start_pfn + size));
4791 zone_init_free_lists(zone);
4796 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4797 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4800 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4802 int __meminit __early_pfn_to_nid(unsigned long pfn,
4803 struct mminit_pfnnid_cache *state)
4805 unsigned long start_pfn, end_pfn;
4808 if (state->last_start <= pfn && pfn < state->last_end)
4809 return state->last_nid;
4811 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4813 state->last_start = start_pfn;
4814 state->last_end = end_pfn;
4815 state->last_nid = nid;
4820 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4823 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4824 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4825 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4827 * If an architecture guarantees that all ranges registered contain no holes
4828 * and may be freed, this this function may be used instead of calling
4829 * memblock_free_early_nid() manually.
4831 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4833 unsigned long start_pfn, end_pfn;
4836 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4837 start_pfn = min(start_pfn, max_low_pfn);
4838 end_pfn = min(end_pfn, max_low_pfn);
4840 if (start_pfn < end_pfn)
4841 memblock_free_early_nid(PFN_PHYS(start_pfn),
4842 (end_pfn - start_pfn) << PAGE_SHIFT,
4848 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4849 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4851 * If an architecture guarantees that all ranges registered contain no holes and may
4852 * be freed, this function may be used instead of calling memory_present() manually.
4854 void __init sparse_memory_present_with_active_regions(int nid)
4856 unsigned long start_pfn, end_pfn;
4859 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4860 memory_present(this_nid, start_pfn, end_pfn);
4864 * get_pfn_range_for_nid - Return the start and end page frames for a node
4865 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4866 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4867 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4869 * It returns the start and end page frame of a node based on information
4870 * provided by memblock_set_node(). If called for a node
4871 * with no available memory, a warning is printed and the start and end
4874 void __meminit get_pfn_range_for_nid(unsigned int nid,
4875 unsigned long *start_pfn, unsigned long *end_pfn)
4877 unsigned long this_start_pfn, this_end_pfn;
4883 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4884 *start_pfn = min(*start_pfn, this_start_pfn);
4885 *end_pfn = max(*end_pfn, this_end_pfn);
4888 if (*start_pfn == -1UL)
4893 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4894 * assumption is made that zones within a node are ordered in monotonic
4895 * increasing memory addresses so that the "highest" populated zone is used
4897 static void __init find_usable_zone_for_movable(void)
4900 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4901 if (zone_index == ZONE_MOVABLE)
4904 if (arch_zone_highest_possible_pfn[zone_index] >
4905 arch_zone_lowest_possible_pfn[zone_index])
4909 VM_BUG_ON(zone_index == -1);
4910 movable_zone = zone_index;
4914 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4915 * because it is sized independent of architecture. Unlike the other zones,
4916 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4917 * in each node depending on the size of each node and how evenly kernelcore
4918 * is distributed. This helper function adjusts the zone ranges
4919 * provided by the architecture for a given node by using the end of the
4920 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4921 * zones within a node are in order of monotonic increases memory addresses
4923 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4924 unsigned long zone_type,
4925 unsigned long node_start_pfn,
4926 unsigned long node_end_pfn,
4927 unsigned long *zone_start_pfn,
4928 unsigned long *zone_end_pfn)
4930 /* Only adjust if ZONE_MOVABLE is on this node */
4931 if (zone_movable_pfn[nid]) {
4932 /* Size ZONE_MOVABLE */
4933 if (zone_type == ZONE_MOVABLE) {
4934 *zone_start_pfn = zone_movable_pfn[nid];
4935 *zone_end_pfn = min(node_end_pfn,
4936 arch_zone_highest_possible_pfn[movable_zone]);
4938 /* Adjust for ZONE_MOVABLE starting within this range */
4939 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4940 *zone_end_pfn > zone_movable_pfn[nid]) {
4941 *zone_end_pfn = zone_movable_pfn[nid];
4943 /* Check if this whole range is within ZONE_MOVABLE */
4944 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4945 *zone_start_pfn = *zone_end_pfn;
4950 * Return the number of pages a zone spans in a node, including holes
4951 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4953 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4954 unsigned long zone_type,
4955 unsigned long node_start_pfn,
4956 unsigned long node_end_pfn,
4957 unsigned long *ignored)
4959 unsigned long zone_start_pfn, zone_end_pfn;
4961 /* When hotadd a new node from cpu_up(), the node should be empty */
4962 if (!node_start_pfn && !node_end_pfn)
4965 /* Get the start and end of the zone */
4966 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4967 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4968 adjust_zone_range_for_zone_movable(nid, zone_type,
4969 node_start_pfn, node_end_pfn,
4970 &zone_start_pfn, &zone_end_pfn);
4972 /* Check that this node has pages within the zone's required range */
4973 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4976 /* Move the zone boundaries inside the node if necessary */
4977 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4978 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4980 /* Return the spanned pages */
4981 return zone_end_pfn - zone_start_pfn;
4985 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4986 * then all holes in the requested range will be accounted for.
4988 unsigned long __meminit __absent_pages_in_range(int nid,
4989 unsigned long range_start_pfn,
4990 unsigned long range_end_pfn)
4992 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4993 unsigned long start_pfn, end_pfn;
4996 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4997 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4998 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4999 nr_absent -= end_pfn - start_pfn;
5005 * absent_pages_in_range - Return number of page frames in holes within a range
5006 * @start_pfn: The start PFN to start searching for holes
5007 * @end_pfn: The end PFN to stop searching for holes
5009 * It returns the number of pages frames in memory holes within a range.
5011 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5012 unsigned long end_pfn)
5014 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5017 /* Return the number of page frames in holes in a zone on a node */
5018 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5019 unsigned long zone_type,
5020 unsigned long node_start_pfn,
5021 unsigned long node_end_pfn,
5022 unsigned long *ignored)
5024 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5025 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5026 unsigned long zone_start_pfn, zone_end_pfn;
5028 /* When hotadd a new node from cpu_up(), the node should be empty */
5029 if (!node_start_pfn && !node_end_pfn)
5032 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5033 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5035 adjust_zone_range_for_zone_movable(nid, zone_type,
5036 node_start_pfn, node_end_pfn,
5037 &zone_start_pfn, &zone_end_pfn);
5038 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5041 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5042 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5043 unsigned long zone_type,
5044 unsigned long node_start_pfn,
5045 unsigned long node_end_pfn,
5046 unsigned long *zones_size)
5048 return zones_size[zone_type];
5051 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5052 unsigned long zone_type,
5053 unsigned long node_start_pfn,
5054 unsigned long node_end_pfn,
5055 unsigned long *zholes_size)
5060 return zholes_size[zone_type];
5063 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5065 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5066 unsigned long node_start_pfn,
5067 unsigned long node_end_pfn,
5068 unsigned long *zones_size,
5069 unsigned long *zholes_size)
5071 unsigned long realtotalpages = 0, totalpages = 0;
5074 for (i = 0; i < MAX_NR_ZONES; i++) {
5075 struct zone *zone = pgdat->node_zones + i;
5076 unsigned long size, real_size;
5078 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5082 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5083 node_start_pfn, node_end_pfn,
5085 zone->spanned_pages = size;
5086 zone->present_pages = real_size;
5089 realtotalpages += real_size;
5092 pgdat->node_spanned_pages = totalpages;
5093 pgdat->node_present_pages = realtotalpages;
5094 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5098 #ifndef CONFIG_SPARSEMEM
5100 * Calculate the size of the zone->blockflags rounded to an unsigned long
5101 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5102 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5103 * round what is now in bits to nearest long in bits, then return it in
5106 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5108 unsigned long usemapsize;
5110 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5111 usemapsize = roundup(zonesize, pageblock_nr_pages);
5112 usemapsize = usemapsize >> pageblock_order;
5113 usemapsize *= NR_PAGEBLOCK_BITS;
5114 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5116 return usemapsize / 8;
5119 static void __init setup_usemap(struct pglist_data *pgdat,
5121 unsigned long zone_start_pfn,
5122 unsigned long zonesize)
5124 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5125 zone->pageblock_flags = NULL;
5127 zone->pageblock_flags =
5128 memblock_virt_alloc_node_nopanic(usemapsize,
5132 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5133 unsigned long zone_start_pfn, unsigned long zonesize) {}
5134 #endif /* CONFIG_SPARSEMEM */
5136 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5138 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5139 void __paginginit set_pageblock_order(void)
5143 /* Check that pageblock_nr_pages has not already been setup */
5144 if (pageblock_order)
5147 if (HPAGE_SHIFT > PAGE_SHIFT)
5148 order = HUGETLB_PAGE_ORDER;
5150 order = MAX_ORDER - 1;
5153 * Assume the largest contiguous order of interest is a huge page.
5154 * This value may be variable depending on boot parameters on IA64 and
5157 pageblock_order = order;
5159 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5162 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5163 * is unused as pageblock_order is set at compile-time. See
5164 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5167 void __paginginit set_pageblock_order(void)
5171 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5173 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5174 unsigned long present_pages)
5176 unsigned long pages = spanned_pages;
5179 * Provide a more accurate estimation if there are holes within
5180 * the zone and SPARSEMEM is in use. If there are holes within the
5181 * zone, each populated memory region may cost us one or two extra
5182 * memmap pages due to alignment because memmap pages for each
5183 * populated regions may not naturally algined on page boundary.
5184 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5186 if (spanned_pages > present_pages + (present_pages >> 4) &&
5187 IS_ENABLED(CONFIG_SPARSEMEM))
5188 pages = present_pages;
5190 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5194 * Set up the zone data structures:
5195 * - mark all pages reserved
5196 * - mark all memory queues empty
5197 * - clear the memory bitmaps
5199 * NOTE: pgdat should get zeroed by caller.
5201 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5204 int nid = pgdat->node_id;
5205 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5208 pgdat_resize_init(pgdat);
5209 #ifdef CONFIG_NUMA_BALANCING
5210 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5211 pgdat->numabalancing_migrate_nr_pages = 0;
5212 pgdat->numabalancing_migrate_next_window = jiffies;
5214 init_waitqueue_head(&pgdat->kswapd_wait);
5215 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5216 pgdat_page_ext_init(pgdat);
5218 for (j = 0; j < MAX_NR_ZONES; j++) {
5219 struct zone *zone = pgdat->node_zones + j;
5220 unsigned long size, realsize, freesize, memmap_pages;
5222 size = zone->spanned_pages;
5223 realsize = freesize = zone->present_pages;
5226 * Adjust freesize so that it accounts for how much memory
5227 * is used by this zone for memmap. This affects the watermark
5228 * and per-cpu initialisations
5230 memmap_pages = calc_memmap_size(size, realsize);
5231 if (!is_highmem_idx(j)) {
5232 if (freesize >= memmap_pages) {
5233 freesize -= memmap_pages;
5236 " %s zone: %lu pages used for memmap\n",
5237 zone_names[j], memmap_pages);
5240 " %s zone: %lu pages exceeds freesize %lu\n",
5241 zone_names[j], memmap_pages, freesize);
5244 /* Account for reserved pages */
5245 if (j == 0 && freesize > dma_reserve) {
5246 freesize -= dma_reserve;
5247 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5248 zone_names[0], dma_reserve);
5251 if (!is_highmem_idx(j))
5252 nr_kernel_pages += freesize;
5253 /* Charge for highmem memmap if there are enough kernel pages */
5254 else if (nr_kernel_pages > memmap_pages * 2)
5255 nr_kernel_pages -= memmap_pages;
5256 nr_all_pages += freesize;
5259 * Set an approximate value for lowmem here, it will be adjusted
5260 * when the bootmem allocator frees pages into the buddy system.
5261 * And all highmem pages will be managed by the buddy system.
5263 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5266 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5268 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5270 zone->name = zone_names[j];
5271 spin_lock_init(&zone->lock);
5272 spin_lock_init(&zone->lru_lock);
5273 zone_seqlock_init(zone);
5274 zone->zone_pgdat = pgdat;
5275 zone_pcp_init(zone);
5277 /* For bootup, initialized properly in watermark setup */
5278 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5280 lruvec_init(&zone->lruvec);
5284 set_pageblock_order();
5285 setup_usemap(pgdat, zone, zone_start_pfn, size);
5286 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5288 memmap_init(size, nid, j, zone_start_pfn);
5289 zone_start_pfn += size;
5293 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5295 unsigned long __maybe_unused offset = 0;
5297 /* Skip empty nodes */
5298 if (!pgdat->node_spanned_pages)
5301 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5302 /* ia64 gets its own node_mem_map, before this, without bootmem */
5303 if (!pgdat->node_mem_map) {
5304 unsigned long size, start, end;
5308 * The zone's endpoints aren't required to be MAX_ORDER
5309 * aligned but the node_mem_map endpoints must be in order
5310 * for the buddy allocator to function correctly.
5312 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5313 offset = pgdat->node_start_pfn - start;
5314 end = pgdat_end_pfn(pgdat);
5315 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5316 size = (end - start) * sizeof(struct page);
5317 map = alloc_remap(pgdat->node_id, size);
5319 map = memblock_virt_alloc_node_nopanic(size,
5321 pgdat->node_mem_map = map + offset;
5323 #ifndef CONFIG_NEED_MULTIPLE_NODES
5325 * With no DISCONTIG, the global mem_map is just set as node 0's
5327 if (pgdat == NODE_DATA(0)) {
5328 mem_map = NODE_DATA(0)->node_mem_map;
5329 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5330 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5332 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5335 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5338 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5339 unsigned long node_start_pfn, unsigned long *zholes_size)
5341 pg_data_t *pgdat = NODE_DATA(nid);
5342 unsigned long start_pfn = 0;
5343 unsigned long end_pfn = 0;
5345 /* pg_data_t should be reset to zero when it's allocated */
5346 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5348 reset_deferred_meminit(pgdat);
5349 pgdat->node_id = nid;
5350 pgdat->node_start_pfn = node_start_pfn;
5351 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5352 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5353 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5354 (u64)start_pfn << PAGE_SHIFT,
5355 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5357 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5358 zones_size, zholes_size);
5360 alloc_node_mem_map(pgdat);
5361 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5362 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5363 nid, (unsigned long)pgdat,
5364 (unsigned long)pgdat->node_mem_map);
5367 free_area_init_core(pgdat);
5370 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5372 #if MAX_NUMNODES > 1
5374 * Figure out the number of possible node ids.
5376 void __init setup_nr_node_ids(void)
5378 unsigned int highest;
5380 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5381 nr_node_ids = highest + 1;
5386 * node_map_pfn_alignment - determine the maximum internode alignment
5388 * This function should be called after node map is populated and sorted.
5389 * It calculates the maximum power of two alignment which can distinguish
5392 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5393 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5394 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5395 * shifted, 1GiB is enough and this function will indicate so.
5397 * This is used to test whether pfn -> nid mapping of the chosen memory
5398 * model has fine enough granularity to avoid incorrect mapping for the
5399 * populated node map.
5401 * Returns the determined alignment in pfn's. 0 if there is no alignment
5402 * requirement (single node).
5404 unsigned long __init node_map_pfn_alignment(void)
5406 unsigned long accl_mask = 0, last_end = 0;
5407 unsigned long start, end, mask;
5411 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5412 if (!start || last_nid < 0 || last_nid == nid) {
5419 * Start with a mask granular enough to pin-point to the
5420 * start pfn and tick off bits one-by-one until it becomes
5421 * too coarse to separate the current node from the last.
5423 mask = ~((1 << __ffs(start)) - 1);
5424 while (mask && last_end <= (start & (mask << 1)))
5427 /* accumulate all internode masks */
5431 /* convert mask to number of pages */
5432 return ~accl_mask + 1;
5435 /* Find the lowest pfn for a node */
5436 static unsigned long __init find_min_pfn_for_node(int nid)
5438 unsigned long min_pfn = ULONG_MAX;
5439 unsigned long start_pfn;
5442 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5443 min_pfn = min(min_pfn, start_pfn);
5445 if (min_pfn == ULONG_MAX) {
5447 "Could not find start_pfn for node %d\n", nid);
5455 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5457 * It returns the minimum PFN based on information provided via
5458 * memblock_set_node().
5460 unsigned long __init find_min_pfn_with_active_regions(void)
5462 return find_min_pfn_for_node(MAX_NUMNODES);
5466 * early_calculate_totalpages()
5467 * Sum pages in active regions for movable zone.
5468 * Populate N_MEMORY for calculating usable_nodes.
5470 static unsigned long __init early_calculate_totalpages(void)
5472 unsigned long totalpages = 0;
5473 unsigned long start_pfn, end_pfn;
5476 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5477 unsigned long pages = end_pfn - start_pfn;
5479 totalpages += pages;
5481 node_set_state(nid, N_MEMORY);
5487 * Find the PFN the Movable zone begins in each node. Kernel memory
5488 * is spread evenly between nodes as long as the nodes have enough
5489 * memory. When they don't, some nodes will have more kernelcore than
5492 static void __init find_zone_movable_pfns_for_nodes(void)
5495 unsigned long usable_startpfn;
5496 unsigned long kernelcore_node, kernelcore_remaining;
5497 /* save the state before borrow the nodemask */
5498 nodemask_t saved_node_state = node_states[N_MEMORY];
5499 unsigned long totalpages = early_calculate_totalpages();
5500 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5501 struct memblock_region *r;
5503 /* Need to find movable_zone earlier when movable_node is specified. */
5504 find_usable_zone_for_movable();
5507 * If movable_node is specified, ignore kernelcore and movablecore
5510 if (movable_node_is_enabled()) {
5511 for_each_memblock(memory, r) {
5512 if (!memblock_is_hotpluggable(r))
5517 usable_startpfn = PFN_DOWN(r->base);
5518 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5519 min(usable_startpfn, zone_movable_pfn[nid]) :
5527 * If movablecore=nn[KMG] was specified, calculate what size of
5528 * kernelcore that corresponds so that memory usable for
5529 * any allocation type is evenly spread. If both kernelcore
5530 * and movablecore are specified, then the value of kernelcore
5531 * will be used for required_kernelcore if it's greater than
5532 * what movablecore would have allowed.
5534 if (required_movablecore) {
5535 unsigned long corepages;
5538 * Round-up so that ZONE_MOVABLE is at least as large as what
5539 * was requested by the user
5541 required_movablecore =
5542 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5543 required_movablecore = min(totalpages, required_movablecore);
5544 corepages = totalpages - required_movablecore;
5546 required_kernelcore = max(required_kernelcore, corepages);
5550 * If kernelcore was not specified or kernelcore size is larger
5551 * than totalpages, there is no ZONE_MOVABLE.
5553 if (!required_kernelcore || required_kernelcore >= totalpages)
5556 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5557 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5560 /* Spread kernelcore memory as evenly as possible throughout nodes */
5561 kernelcore_node = required_kernelcore / usable_nodes;
5562 for_each_node_state(nid, N_MEMORY) {
5563 unsigned long start_pfn, end_pfn;
5566 * Recalculate kernelcore_node if the division per node
5567 * now exceeds what is necessary to satisfy the requested
5568 * amount of memory for the kernel
5570 if (required_kernelcore < kernelcore_node)
5571 kernelcore_node = required_kernelcore / usable_nodes;
5574 * As the map is walked, we track how much memory is usable
5575 * by the kernel using kernelcore_remaining. When it is
5576 * 0, the rest of the node is usable by ZONE_MOVABLE
5578 kernelcore_remaining = kernelcore_node;
5580 /* Go through each range of PFNs within this node */
5581 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5582 unsigned long size_pages;
5584 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5585 if (start_pfn >= end_pfn)
5588 /* Account for what is only usable for kernelcore */
5589 if (start_pfn < usable_startpfn) {
5590 unsigned long kernel_pages;
5591 kernel_pages = min(end_pfn, usable_startpfn)
5594 kernelcore_remaining -= min(kernel_pages,
5595 kernelcore_remaining);
5596 required_kernelcore -= min(kernel_pages,
5597 required_kernelcore);
5599 /* Continue if range is now fully accounted */
5600 if (end_pfn <= usable_startpfn) {
5603 * Push zone_movable_pfn to the end so
5604 * that if we have to rebalance
5605 * kernelcore across nodes, we will
5606 * not double account here
5608 zone_movable_pfn[nid] = end_pfn;
5611 start_pfn = usable_startpfn;
5615 * The usable PFN range for ZONE_MOVABLE is from
5616 * start_pfn->end_pfn. Calculate size_pages as the
5617 * number of pages used as kernelcore
5619 size_pages = end_pfn - start_pfn;
5620 if (size_pages > kernelcore_remaining)
5621 size_pages = kernelcore_remaining;
5622 zone_movable_pfn[nid] = start_pfn + size_pages;
5625 * Some kernelcore has been met, update counts and
5626 * break if the kernelcore for this node has been
5629 required_kernelcore -= min(required_kernelcore,
5631 kernelcore_remaining -= size_pages;
5632 if (!kernelcore_remaining)
5638 * If there is still required_kernelcore, we do another pass with one
5639 * less node in the count. This will push zone_movable_pfn[nid] further
5640 * along on the nodes that still have memory until kernelcore is
5644 if (usable_nodes && required_kernelcore > usable_nodes)
5648 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5649 for (nid = 0; nid < MAX_NUMNODES; nid++)
5650 zone_movable_pfn[nid] =
5651 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5654 /* restore the node_state */
5655 node_states[N_MEMORY] = saved_node_state;
5658 /* Any regular or high memory on that node ? */
5659 static void check_for_memory(pg_data_t *pgdat, int nid)
5661 enum zone_type zone_type;
5663 if (N_MEMORY == N_NORMAL_MEMORY)
5666 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5667 struct zone *zone = &pgdat->node_zones[zone_type];
5668 if (populated_zone(zone)) {
5669 node_set_state(nid, N_HIGH_MEMORY);
5670 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5671 zone_type <= ZONE_NORMAL)
5672 node_set_state(nid, N_NORMAL_MEMORY);
5679 * free_area_init_nodes - Initialise all pg_data_t and zone data
5680 * @max_zone_pfn: an array of max PFNs for each zone
5682 * This will call free_area_init_node() for each active node in the system.
5683 * Using the page ranges provided by memblock_set_node(), the size of each
5684 * zone in each node and their holes is calculated. If the maximum PFN
5685 * between two adjacent zones match, it is assumed that the zone is empty.
5686 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5687 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5688 * starts where the previous one ended. For example, ZONE_DMA32 starts
5689 * at arch_max_dma_pfn.
5691 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5693 unsigned long start_pfn, end_pfn;
5696 /* Record where the zone boundaries are */
5697 memset(arch_zone_lowest_possible_pfn, 0,
5698 sizeof(arch_zone_lowest_possible_pfn));
5699 memset(arch_zone_highest_possible_pfn, 0,
5700 sizeof(arch_zone_highest_possible_pfn));
5701 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5702 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5703 for (i = 1; i < MAX_NR_ZONES; i++) {
5704 if (i == ZONE_MOVABLE)
5706 arch_zone_lowest_possible_pfn[i] =
5707 arch_zone_highest_possible_pfn[i-1];
5708 arch_zone_highest_possible_pfn[i] =
5709 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5711 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5712 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5714 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5715 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5716 find_zone_movable_pfns_for_nodes();
5718 /* Print out the zone ranges */
5719 pr_info("Zone ranges:\n");
5720 for (i = 0; i < MAX_NR_ZONES; i++) {
5721 if (i == ZONE_MOVABLE)
5723 pr_info(" %-8s ", zone_names[i]);
5724 if (arch_zone_lowest_possible_pfn[i] ==
5725 arch_zone_highest_possible_pfn[i])
5728 pr_cont("[mem %#018Lx-%#018Lx]\n",
5729 (u64)arch_zone_lowest_possible_pfn[i]
5731 ((u64)arch_zone_highest_possible_pfn[i]
5732 << PAGE_SHIFT) - 1);
5735 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5736 pr_info("Movable zone start for each node\n");
5737 for (i = 0; i < MAX_NUMNODES; i++) {
5738 if (zone_movable_pfn[i])
5739 pr_info(" Node %d: %#018Lx\n", i,
5740 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5743 /* Print out the early node map */
5744 pr_info("Early memory node ranges\n");
5745 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5746 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5747 (u64)start_pfn << PAGE_SHIFT,
5748 ((u64)end_pfn << PAGE_SHIFT) - 1);
5750 /* Initialise every node */
5751 mminit_verify_pageflags_layout();
5752 setup_nr_node_ids();
5753 for_each_online_node(nid) {
5754 pg_data_t *pgdat = NODE_DATA(nid);
5755 free_area_init_node(nid, NULL,
5756 find_min_pfn_for_node(nid), NULL);
5758 /* Any memory on that node */
5759 if (pgdat->node_present_pages)
5760 node_set_state(nid, N_MEMORY);
5761 check_for_memory(pgdat, nid);
5765 static int __init cmdline_parse_core(char *p, unsigned long *core)
5767 unsigned long long coremem;
5771 coremem = memparse(p, &p);
5772 *core = coremem >> PAGE_SHIFT;
5774 /* Paranoid check that UL is enough for the coremem value */
5775 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5781 * kernelcore=size sets the amount of memory for use for allocations that
5782 * cannot be reclaimed or migrated.
5784 static int __init cmdline_parse_kernelcore(char *p)
5786 return cmdline_parse_core(p, &required_kernelcore);
5790 * movablecore=size sets the amount of memory for use for allocations that
5791 * can be reclaimed or migrated.
5793 static int __init cmdline_parse_movablecore(char *p)
5795 return cmdline_parse_core(p, &required_movablecore);
5798 early_param("kernelcore", cmdline_parse_kernelcore);
5799 early_param("movablecore", cmdline_parse_movablecore);
5801 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5803 void adjust_managed_page_count(struct page *page, long count)
5805 spin_lock(&managed_page_count_lock);
5806 page_zone(page)->managed_pages += count;
5807 totalram_pages += count;
5808 #ifdef CONFIG_HIGHMEM
5809 if (PageHighMem(page))
5810 totalhigh_pages += count;
5812 spin_unlock(&managed_page_count_lock);
5814 EXPORT_SYMBOL(adjust_managed_page_count);
5816 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5819 unsigned long pages = 0;
5821 start = (void *)PAGE_ALIGN((unsigned long)start);
5822 end = (void *)((unsigned long)end & PAGE_MASK);
5823 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5824 if ((unsigned int)poison <= 0xFF)
5825 memset(pos, poison, PAGE_SIZE);
5826 free_reserved_page(virt_to_page(pos));
5830 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5831 s, pages << (PAGE_SHIFT - 10), start, end);
5835 EXPORT_SYMBOL(free_reserved_area);
5837 #ifdef CONFIG_HIGHMEM
5838 void free_highmem_page(struct page *page)
5840 __free_reserved_page(page);
5842 page_zone(page)->managed_pages++;
5848 void __init mem_init_print_info(const char *str)
5850 unsigned long physpages, codesize, datasize, rosize, bss_size;
5851 unsigned long init_code_size, init_data_size;
5853 physpages = get_num_physpages();
5854 codesize = _etext - _stext;
5855 datasize = _edata - _sdata;
5856 rosize = __end_rodata - __start_rodata;
5857 bss_size = __bss_stop - __bss_start;
5858 init_data_size = __init_end - __init_begin;
5859 init_code_size = _einittext - _sinittext;
5862 * Detect special cases and adjust section sizes accordingly:
5863 * 1) .init.* may be embedded into .data sections
5864 * 2) .init.text.* may be out of [__init_begin, __init_end],
5865 * please refer to arch/tile/kernel/vmlinux.lds.S.
5866 * 3) .rodata.* may be embedded into .text or .data sections.
5868 #define adj_init_size(start, end, size, pos, adj) \
5870 if (start <= pos && pos < end && size > adj) \
5874 adj_init_size(__init_begin, __init_end, init_data_size,
5875 _sinittext, init_code_size);
5876 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5877 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5878 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5879 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5881 #undef adj_init_size
5883 pr_info("Memory: %luK/%luK available "
5884 "(%luK kernel code, %luK rwdata, %luK rodata, "
5885 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5886 #ifdef CONFIG_HIGHMEM
5890 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5891 codesize >> 10, datasize >> 10, rosize >> 10,
5892 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5893 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5894 totalcma_pages << (PAGE_SHIFT-10),
5895 #ifdef CONFIG_HIGHMEM
5896 totalhigh_pages << (PAGE_SHIFT-10),
5898 str ? ", " : "", str ? str : "");
5902 * set_dma_reserve - set the specified number of pages reserved in the first zone
5903 * @new_dma_reserve: The number of pages to mark reserved
5905 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5906 * In the DMA zone, a significant percentage may be consumed by kernel image
5907 * and other unfreeable allocations which can skew the watermarks badly. This
5908 * function may optionally be used to account for unfreeable pages in the
5909 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5910 * smaller per-cpu batchsize.
5912 void __init set_dma_reserve(unsigned long new_dma_reserve)
5914 dma_reserve = new_dma_reserve;
5917 void __init free_area_init(unsigned long *zones_size)
5919 free_area_init_node(0, zones_size,
5920 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5923 static int page_alloc_cpu_notify(struct notifier_block *self,
5924 unsigned long action, void *hcpu)
5926 int cpu = (unsigned long)hcpu;
5928 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5929 lru_add_drain_cpu(cpu);
5933 * Spill the event counters of the dead processor
5934 * into the current processors event counters.
5935 * This artificially elevates the count of the current
5938 vm_events_fold_cpu(cpu);
5941 * Zero the differential counters of the dead processor
5942 * so that the vm statistics are consistent.
5944 * This is only okay since the processor is dead and cannot
5945 * race with what we are doing.
5947 cpu_vm_stats_fold(cpu);
5952 void __init page_alloc_init(void)
5954 hotcpu_notifier(page_alloc_cpu_notify, 0);
5958 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5959 * or min_free_kbytes changes.
5961 static void calculate_totalreserve_pages(void)
5963 struct pglist_data *pgdat;
5964 unsigned long reserve_pages = 0;
5965 enum zone_type i, j;
5967 for_each_online_pgdat(pgdat) {
5968 for (i = 0; i < MAX_NR_ZONES; i++) {
5969 struct zone *zone = pgdat->node_zones + i;
5972 /* Find valid and maximum lowmem_reserve in the zone */
5973 for (j = i; j < MAX_NR_ZONES; j++) {
5974 if (zone->lowmem_reserve[j] > max)
5975 max = zone->lowmem_reserve[j];
5978 /* we treat the high watermark as reserved pages. */
5979 max += high_wmark_pages(zone);
5981 if (max > zone->managed_pages)
5982 max = zone->managed_pages;
5983 reserve_pages += max;
5985 * Lowmem reserves are not available to
5986 * GFP_HIGHUSER page cache allocations and
5987 * kswapd tries to balance zones to their high
5988 * watermark. As a result, neither should be
5989 * regarded as dirtyable memory, to prevent a
5990 * situation where reclaim has to clean pages
5991 * in order to balance the zones.
5993 zone->dirty_balance_reserve = max;
5996 dirty_balance_reserve = reserve_pages;
5997 totalreserve_pages = reserve_pages;
6001 * setup_per_zone_lowmem_reserve - called whenever
6002 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6003 * has a correct pages reserved value, so an adequate number of
6004 * pages are left in the zone after a successful __alloc_pages().
6006 static void setup_per_zone_lowmem_reserve(void)
6008 struct pglist_data *pgdat;
6009 enum zone_type j, idx;
6011 for_each_online_pgdat(pgdat) {
6012 for (j = 0; j < MAX_NR_ZONES; j++) {
6013 struct zone *zone = pgdat->node_zones + j;
6014 unsigned long managed_pages = zone->managed_pages;
6016 zone->lowmem_reserve[j] = 0;
6020 struct zone *lower_zone;
6024 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6025 sysctl_lowmem_reserve_ratio[idx] = 1;
6027 lower_zone = pgdat->node_zones + idx;
6028 lower_zone->lowmem_reserve[j] = managed_pages /
6029 sysctl_lowmem_reserve_ratio[idx];
6030 managed_pages += lower_zone->managed_pages;
6035 /* update totalreserve_pages */
6036 calculate_totalreserve_pages();
6039 static void __setup_per_zone_wmarks(void)
6041 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6042 unsigned long lowmem_pages = 0;
6044 unsigned long flags;
6046 /* Calculate total number of !ZONE_HIGHMEM pages */
6047 for_each_zone(zone) {
6048 if (!is_highmem(zone))
6049 lowmem_pages += zone->managed_pages;
6052 for_each_zone(zone) {
6055 spin_lock_irqsave(&zone->lock, flags);
6056 tmp = (u64)pages_min * zone->managed_pages;
6057 do_div(tmp, lowmem_pages);
6058 if (is_highmem(zone)) {
6060 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6061 * need highmem pages, so cap pages_min to a small
6064 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6065 * deltas control asynch page reclaim, and so should
6066 * not be capped for highmem.
6068 unsigned long min_pages;
6070 min_pages = zone->managed_pages / 1024;
6071 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6072 zone->watermark[WMARK_MIN] = min_pages;
6075 * If it's a lowmem zone, reserve a number of pages
6076 * proportionate to the zone's size.
6078 zone->watermark[WMARK_MIN] = tmp;
6081 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6082 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6084 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6085 high_wmark_pages(zone) - low_wmark_pages(zone) -
6086 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6088 spin_unlock_irqrestore(&zone->lock, flags);
6091 /* update totalreserve_pages */
6092 calculate_totalreserve_pages();
6096 * setup_per_zone_wmarks - called when min_free_kbytes changes
6097 * or when memory is hot-{added|removed}
6099 * Ensures that the watermark[min,low,high] values for each zone are set
6100 * correctly with respect to min_free_kbytes.
6102 void setup_per_zone_wmarks(void)
6104 mutex_lock(&zonelists_mutex);
6105 __setup_per_zone_wmarks();
6106 mutex_unlock(&zonelists_mutex);
6110 * The inactive anon list should be small enough that the VM never has to
6111 * do too much work, but large enough that each inactive page has a chance
6112 * to be referenced again before it is swapped out.
6114 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6115 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6116 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6117 * the anonymous pages are kept on the inactive list.
6120 * memory ratio inactive anon
6121 * -------------------------------------
6130 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6132 unsigned int gb, ratio;
6134 /* Zone size in gigabytes */
6135 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6137 ratio = int_sqrt(10 * gb);
6141 zone->inactive_ratio = ratio;
6144 static void __meminit setup_per_zone_inactive_ratio(void)
6149 calculate_zone_inactive_ratio(zone);
6153 * Initialise min_free_kbytes.
6155 * For small machines we want it small (128k min). For large machines
6156 * we want it large (64MB max). But it is not linear, because network
6157 * bandwidth does not increase linearly with machine size. We use
6159 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6160 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6176 int __meminit init_per_zone_wmark_min(void)
6178 unsigned long lowmem_kbytes;
6179 int new_min_free_kbytes;
6181 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6182 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6184 if (new_min_free_kbytes > user_min_free_kbytes) {
6185 min_free_kbytes = new_min_free_kbytes;
6186 if (min_free_kbytes < 128)
6187 min_free_kbytes = 128;
6188 if (min_free_kbytes > 65536)
6189 min_free_kbytes = 65536;
6191 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6192 new_min_free_kbytes, user_min_free_kbytes);
6194 setup_per_zone_wmarks();
6195 refresh_zone_stat_thresholds();
6196 setup_per_zone_lowmem_reserve();
6197 setup_per_zone_inactive_ratio();
6200 module_init(init_per_zone_wmark_min)
6203 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6204 * that we can call two helper functions whenever min_free_kbytes
6207 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6208 void __user *buffer, size_t *length, loff_t *ppos)
6212 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6217 user_min_free_kbytes = min_free_kbytes;
6218 setup_per_zone_wmarks();
6224 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6225 void __user *buffer, size_t *length, loff_t *ppos)
6230 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6235 zone->min_unmapped_pages = (zone->managed_pages *
6236 sysctl_min_unmapped_ratio) / 100;
6240 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6241 void __user *buffer, size_t *length, loff_t *ppos)
6246 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6251 zone->min_slab_pages = (zone->managed_pages *
6252 sysctl_min_slab_ratio) / 100;
6258 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6259 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6260 * whenever sysctl_lowmem_reserve_ratio changes.
6262 * The reserve ratio obviously has absolutely no relation with the
6263 * minimum watermarks. The lowmem reserve ratio can only make sense
6264 * if in function of the boot time zone sizes.
6266 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6267 void __user *buffer, size_t *length, loff_t *ppos)
6269 proc_dointvec_minmax(table, write, buffer, length, ppos);
6270 setup_per_zone_lowmem_reserve();
6275 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6276 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6277 * pagelist can have before it gets flushed back to buddy allocator.
6279 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6280 void __user *buffer, size_t *length, loff_t *ppos)
6283 int old_percpu_pagelist_fraction;
6286 mutex_lock(&pcp_batch_high_lock);
6287 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6289 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6290 if (!write || ret < 0)
6293 /* Sanity checking to avoid pcp imbalance */
6294 if (percpu_pagelist_fraction &&
6295 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6296 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6302 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6305 for_each_populated_zone(zone) {
6308 for_each_possible_cpu(cpu)
6309 pageset_set_high_and_batch(zone,
6310 per_cpu_ptr(zone->pageset, cpu));
6313 mutex_unlock(&pcp_batch_high_lock);
6318 int hashdist = HASHDIST_DEFAULT;
6320 static int __init set_hashdist(char *str)
6324 hashdist = simple_strtoul(str, &str, 0);
6327 __setup("hashdist=", set_hashdist);
6331 * allocate a large system hash table from bootmem
6332 * - it is assumed that the hash table must contain an exact power-of-2
6333 * quantity of entries
6334 * - limit is the number of hash buckets, not the total allocation size
6336 void *__init alloc_large_system_hash(const char *tablename,
6337 unsigned long bucketsize,
6338 unsigned long numentries,
6341 unsigned int *_hash_shift,
6342 unsigned int *_hash_mask,
6343 unsigned long low_limit,
6344 unsigned long high_limit)
6346 unsigned long long max = high_limit;
6347 unsigned long log2qty, size;
6350 /* allow the kernel cmdline to have a say */
6352 /* round applicable memory size up to nearest megabyte */
6353 numentries = nr_kernel_pages;
6355 /* It isn't necessary when PAGE_SIZE >= 1MB */
6356 if (PAGE_SHIFT < 20)
6357 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6359 /* limit to 1 bucket per 2^scale bytes of low memory */
6360 if (scale > PAGE_SHIFT)
6361 numentries >>= (scale - PAGE_SHIFT);
6363 numentries <<= (PAGE_SHIFT - scale);
6365 /* Make sure we've got at least a 0-order allocation.. */
6366 if (unlikely(flags & HASH_SMALL)) {
6367 /* Makes no sense without HASH_EARLY */
6368 WARN_ON(!(flags & HASH_EARLY));
6369 if (!(numentries >> *_hash_shift)) {
6370 numentries = 1UL << *_hash_shift;
6371 BUG_ON(!numentries);
6373 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6374 numentries = PAGE_SIZE / bucketsize;
6376 numentries = roundup_pow_of_two(numentries);
6378 /* limit allocation size to 1/16 total memory by default */
6380 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6381 do_div(max, bucketsize);
6383 max = min(max, 0x80000000ULL);
6385 if (numentries < low_limit)
6386 numentries = low_limit;
6387 if (numentries > max)
6390 log2qty = ilog2(numentries);
6393 size = bucketsize << log2qty;
6394 if (flags & HASH_EARLY)
6395 table = memblock_virt_alloc_nopanic(size, 0);
6397 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6400 * If bucketsize is not a power-of-two, we may free
6401 * some pages at the end of hash table which
6402 * alloc_pages_exact() automatically does
6404 if (get_order(size) < MAX_ORDER) {
6405 table = alloc_pages_exact(size, GFP_ATOMIC);
6406 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6409 } while (!table && size > PAGE_SIZE && --log2qty);
6412 panic("Failed to allocate %s hash table\n", tablename);
6414 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6417 ilog2(size) - PAGE_SHIFT,
6421 *_hash_shift = log2qty;
6423 *_hash_mask = (1 << log2qty) - 1;
6428 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6429 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6432 #ifdef CONFIG_SPARSEMEM
6433 return __pfn_to_section(pfn)->pageblock_flags;
6435 return zone->pageblock_flags;
6436 #endif /* CONFIG_SPARSEMEM */
6439 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6441 #ifdef CONFIG_SPARSEMEM
6442 pfn &= (PAGES_PER_SECTION-1);
6443 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6445 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6446 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6447 #endif /* CONFIG_SPARSEMEM */
6451 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6452 * @page: The page within the block of interest
6453 * @pfn: The target page frame number
6454 * @end_bitidx: The last bit of interest to retrieve
6455 * @mask: mask of bits that the caller is interested in
6457 * Return: pageblock_bits flags
6459 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6460 unsigned long end_bitidx,
6464 unsigned long *bitmap;
6465 unsigned long bitidx, word_bitidx;
6468 zone = page_zone(page);
6469 bitmap = get_pageblock_bitmap(zone, pfn);
6470 bitidx = pfn_to_bitidx(zone, pfn);
6471 word_bitidx = bitidx / BITS_PER_LONG;
6472 bitidx &= (BITS_PER_LONG-1);
6474 word = bitmap[word_bitidx];
6475 bitidx += end_bitidx;
6476 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6480 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6481 * @page: The page within the block of interest
6482 * @flags: The flags to set
6483 * @pfn: The target page frame number
6484 * @end_bitidx: The last bit of interest
6485 * @mask: mask of bits that the caller is interested in
6487 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6489 unsigned long end_bitidx,
6493 unsigned long *bitmap;
6494 unsigned long bitidx, word_bitidx;
6495 unsigned long old_word, word;
6497 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6499 zone = page_zone(page);
6500 bitmap = get_pageblock_bitmap(zone, pfn);
6501 bitidx = pfn_to_bitidx(zone, pfn);
6502 word_bitidx = bitidx / BITS_PER_LONG;
6503 bitidx &= (BITS_PER_LONG-1);
6505 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6507 bitidx += end_bitidx;
6508 mask <<= (BITS_PER_LONG - bitidx - 1);
6509 flags <<= (BITS_PER_LONG - bitidx - 1);
6511 word = READ_ONCE(bitmap[word_bitidx]);
6513 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6514 if (word == old_word)
6521 * This function checks whether pageblock includes unmovable pages or not.
6522 * If @count is not zero, it is okay to include less @count unmovable pages
6524 * PageLRU check without isolation or lru_lock could race so that
6525 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6526 * expect this function should be exact.
6528 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6529 bool skip_hwpoisoned_pages)
6531 unsigned long pfn, iter, found;
6535 * For avoiding noise data, lru_add_drain_all() should be called
6536 * If ZONE_MOVABLE, the zone never contains unmovable pages
6538 if (zone_idx(zone) == ZONE_MOVABLE)
6540 mt = get_pageblock_migratetype(page);
6541 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6544 pfn = page_to_pfn(page);
6545 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6546 unsigned long check = pfn + iter;
6548 if (!pfn_valid_within(check))
6551 page = pfn_to_page(check);
6554 * Hugepages are not in LRU lists, but they're movable.
6555 * We need not scan over tail pages bacause we don't
6556 * handle each tail page individually in migration.
6558 if (PageHuge(page)) {
6559 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6564 * We can't use page_count without pin a page
6565 * because another CPU can free compound page.
6566 * This check already skips compound tails of THP
6567 * because their page->_count is zero at all time.
6569 if (!atomic_read(&page->_count)) {
6570 if (PageBuddy(page))
6571 iter += (1 << page_order(page)) - 1;
6576 * The HWPoisoned page may be not in buddy system, and
6577 * page_count() is not 0.
6579 if (skip_hwpoisoned_pages && PageHWPoison(page))
6585 * If there are RECLAIMABLE pages, we need to check
6586 * it. But now, memory offline itself doesn't call
6587 * shrink_node_slabs() and it still to be fixed.
6590 * If the page is not RAM, page_count()should be 0.
6591 * we don't need more check. This is an _used_ not-movable page.
6593 * The problematic thing here is PG_reserved pages. PG_reserved
6594 * is set to both of a memory hole page and a _used_ kernel
6603 bool is_pageblock_removable_nolock(struct page *page)
6609 * We have to be careful here because we are iterating over memory
6610 * sections which are not zone aware so we might end up outside of
6611 * the zone but still within the section.
6612 * We have to take care about the node as well. If the node is offline
6613 * its NODE_DATA will be NULL - see page_zone.
6615 if (!node_online(page_to_nid(page)))
6618 zone = page_zone(page);
6619 pfn = page_to_pfn(page);
6620 if (!zone_spans_pfn(zone, pfn))
6623 return !has_unmovable_pages(zone, page, 0, true);
6628 static unsigned long pfn_max_align_down(unsigned long pfn)
6630 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6631 pageblock_nr_pages) - 1);
6634 static unsigned long pfn_max_align_up(unsigned long pfn)
6636 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6637 pageblock_nr_pages));
6640 /* [start, end) must belong to a single zone. */
6641 static int __alloc_contig_migrate_range(struct compact_control *cc,
6642 unsigned long start, unsigned long end)
6644 /* This function is based on compact_zone() from compaction.c. */
6645 unsigned long nr_reclaimed;
6646 unsigned long pfn = start;
6647 unsigned int tries = 0;
6652 while (pfn < end || !list_empty(&cc->migratepages)) {
6653 if (fatal_signal_pending(current)) {
6658 if (list_empty(&cc->migratepages)) {
6659 cc->nr_migratepages = 0;
6660 pfn = isolate_migratepages_range(cc, pfn, end);
6666 } else if (++tries == 5) {
6667 ret = ret < 0 ? ret : -EBUSY;
6671 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6673 cc->nr_migratepages -= nr_reclaimed;
6675 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6676 NULL, 0, cc->mode, MR_CMA);
6679 putback_movable_pages(&cc->migratepages);
6686 * alloc_contig_range() -- tries to allocate given range of pages
6687 * @start: start PFN to allocate
6688 * @end: one-past-the-last PFN to allocate
6689 * @migratetype: migratetype of the underlaying pageblocks (either
6690 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6691 * in range must have the same migratetype and it must
6692 * be either of the two.
6694 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6695 * aligned, however it's the caller's responsibility to guarantee that
6696 * we are the only thread that changes migrate type of pageblocks the
6699 * The PFN range must belong to a single zone.
6701 * Returns zero on success or negative error code. On success all
6702 * pages which PFN is in [start, end) are allocated for the caller and
6703 * need to be freed with free_contig_range().
6705 int alloc_contig_range(unsigned long start, unsigned long end,
6706 unsigned migratetype)
6708 unsigned long outer_start, outer_end;
6712 struct compact_control cc = {
6713 .nr_migratepages = 0,
6715 .zone = page_zone(pfn_to_page(start)),
6716 .mode = MIGRATE_SYNC,
6717 .ignore_skip_hint = true,
6719 INIT_LIST_HEAD(&cc.migratepages);
6722 * What we do here is we mark all pageblocks in range as
6723 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6724 * have different sizes, and due to the way page allocator
6725 * work, we align the range to biggest of the two pages so
6726 * that page allocator won't try to merge buddies from
6727 * different pageblocks and change MIGRATE_ISOLATE to some
6728 * other migration type.
6730 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6731 * migrate the pages from an unaligned range (ie. pages that
6732 * we are interested in). This will put all the pages in
6733 * range back to page allocator as MIGRATE_ISOLATE.
6735 * When this is done, we take the pages in range from page
6736 * allocator removing them from the buddy system. This way
6737 * page allocator will never consider using them.
6739 * This lets us mark the pageblocks back as
6740 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6741 * aligned range but not in the unaligned, original range are
6742 * put back to page allocator so that buddy can use them.
6745 ret = start_isolate_page_range(pfn_max_align_down(start),
6746 pfn_max_align_up(end), migratetype,
6751 ret = __alloc_contig_migrate_range(&cc, start, end);
6756 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6757 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6758 * more, all pages in [start, end) are free in page allocator.
6759 * What we are going to do is to allocate all pages from
6760 * [start, end) (that is remove them from page allocator).
6762 * The only problem is that pages at the beginning and at the
6763 * end of interesting range may be not aligned with pages that
6764 * page allocator holds, ie. they can be part of higher order
6765 * pages. Because of this, we reserve the bigger range and
6766 * once this is done free the pages we are not interested in.
6768 * We don't have to hold zone->lock here because the pages are
6769 * isolated thus they won't get removed from buddy.
6772 lru_add_drain_all();
6773 drain_all_pages(cc.zone);
6776 outer_start = start;
6777 while (!PageBuddy(pfn_to_page(outer_start))) {
6778 if (++order >= MAX_ORDER) {
6782 outer_start &= ~0UL << order;
6785 /* Make sure the range is really isolated. */
6786 if (test_pages_isolated(outer_start, end, false)) {
6787 pr_info("%s: [%lx, %lx) PFNs busy\n",
6788 __func__, outer_start, end);
6793 /* Grab isolated pages from freelists. */
6794 outer_end = isolate_freepages_range(&cc, outer_start, end);
6800 /* Free head and tail (if any) */
6801 if (start != outer_start)
6802 free_contig_range(outer_start, start - outer_start);
6803 if (end != outer_end)
6804 free_contig_range(end, outer_end - end);
6807 undo_isolate_page_range(pfn_max_align_down(start),
6808 pfn_max_align_up(end), migratetype);
6812 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6814 unsigned int count = 0;
6816 for (; nr_pages--; pfn++) {
6817 struct page *page = pfn_to_page(pfn);
6819 count += page_count(page) != 1;
6822 WARN(count != 0, "%d pages are still in use!\n", count);
6826 #ifdef CONFIG_MEMORY_HOTPLUG
6828 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6829 * page high values need to be recalulated.
6831 void __meminit zone_pcp_update(struct zone *zone)
6834 mutex_lock(&pcp_batch_high_lock);
6835 for_each_possible_cpu(cpu)
6836 pageset_set_high_and_batch(zone,
6837 per_cpu_ptr(zone->pageset, cpu));
6838 mutex_unlock(&pcp_batch_high_lock);
6842 void zone_pcp_reset(struct zone *zone)
6844 unsigned long flags;
6846 struct per_cpu_pageset *pset;
6848 /* avoid races with drain_pages() */
6849 local_irq_save(flags);
6850 if (zone->pageset != &boot_pageset) {
6851 for_each_online_cpu(cpu) {
6852 pset = per_cpu_ptr(zone->pageset, cpu);
6853 drain_zonestat(zone, pset);
6855 free_percpu(zone->pageset);
6856 zone->pageset = &boot_pageset;
6858 local_irq_restore(flags);
6861 #ifdef CONFIG_MEMORY_HOTREMOVE
6863 * All pages in the range must be isolated before calling this.
6866 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6870 unsigned int order, i;
6872 unsigned long flags;
6873 /* find the first valid pfn */
6874 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6879 zone = page_zone(pfn_to_page(pfn));
6880 spin_lock_irqsave(&zone->lock, flags);
6882 while (pfn < end_pfn) {
6883 if (!pfn_valid(pfn)) {
6887 page = pfn_to_page(pfn);
6889 * The HWPoisoned page may be not in buddy system, and
6890 * page_count() is not 0.
6892 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6894 SetPageReserved(page);
6898 BUG_ON(page_count(page));
6899 BUG_ON(!PageBuddy(page));
6900 order = page_order(page);
6901 #ifdef CONFIG_DEBUG_VM
6902 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6903 pfn, 1 << order, end_pfn);
6905 list_del(&page->lru);
6906 rmv_page_order(page);
6907 zone->free_area[order].nr_free--;
6908 for (i = 0; i < (1 << order); i++)
6909 SetPageReserved((page+i));
6910 pfn += (1 << order);
6912 spin_unlock_irqrestore(&zone->lock, flags);
6916 #ifdef CONFIG_MEMORY_FAILURE
6917 bool is_free_buddy_page(struct page *page)
6919 struct zone *zone = page_zone(page);
6920 unsigned long pfn = page_to_pfn(page);
6921 unsigned long flags;
6924 spin_lock_irqsave(&zone->lock, flags);
6925 for (order = 0; order < MAX_ORDER; order++) {
6926 struct page *page_head = page - (pfn & ((1 << order) - 1));
6928 if (PageBuddy(page_head) && page_order(page_head) >= order)
6931 spin_unlock_irqrestore(&zone->lock, flags);
6933 return order < MAX_ORDER;