2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page *page)
141 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
143 page->index = migratetype;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly;
187 static void __free_pages_ok(struct page *page, unsigned int order);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages);
215 static char * const zone_names[MAX_NR_ZONES] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 static void free_compound_page(struct page *page);
233 compound_page_dtor * const compound_page_dtors[] = {
236 #ifdef CONFIG_HUGETLB_PAGE
242 * Try to keep at least this much lowmem free. Do not allow normal
243 * allocations below this point, only high priority ones. Automatically
244 * tuned according to the amount of memory in the system.
246 int min_free_kbytes = 1024;
247 int user_min_free_kbytes = -1;
250 * Extra memory for the system to try freeing. Used to temporarily
251 * free memory, to make space for new workloads. Anyone can allocate
252 * down to the min watermarks controlled by min_free_kbytes above.
254 int extra_free_kbytes = 0;
256 static unsigned long __meminitdata nr_kernel_pages;
257 static unsigned long __meminitdata nr_all_pages;
258 static unsigned long __meminitdata dma_reserve;
260 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
261 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
262 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
263 static unsigned long __initdata required_kernelcore;
264 static unsigned long __initdata required_movablecore;
265 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
267 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
269 EXPORT_SYMBOL(movable_zone);
270 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
273 int nr_node_ids __read_mostly = MAX_NUMNODES;
274 int nr_online_nodes __read_mostly = 1;
275 EXPORT_SYMBOL(nr_node_ids);
276 EXPORT_SYMBOL(nr_online_nodes);
279 int page_group_by_mobility_disabled __read_mostly;
281 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
282 static inline void reset_deferred_meminit(pg_data_t *pgdat)
284 pgdat->first_deferred_pfn = ULONG_MAX;
287 /* Returns true if the struct page for the pfn is uninitialised */
288 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
290 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
296 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
298 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
305 * Returns false when the remaining initialisation should be deferred until
306 * later in the boot cycle when it can be parallelised.
308 static inline bool update_defer_init(pg_data_t *pgdat,
309 unsigned long pfn, unsigned long zone_end,
310 unsigned long *nr_initialised)
312 /* Always populate low zones for address-contrained allocations */
313 if (zone_end < pgdat_end_pfn(pgdat))
316 /* Initialise at least 2G of the highest zone */
318 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
319 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
320 pgdat->first_deferred_pfn = pfn;
327 static inline void reset_deferred_meminit(pg_data_t *pgdat)
331 static inline bool early_page_uninitialised(unsigned long pfn)
336 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
341 static inline bool update_defer_init(pg_data_t *pgdat,
342 unsigned long pfn, unsigned long zone_end,
343 unsigned long *nr_initialised)
350 void set_pageblock_migratetype(struct page *page, int migratetype)
352 if (unlikely(page_group_by_mobility_disabled &&
353 migratetype < MIGRATE_PCPTYPES))
354 migratetype = MIGRATE_UNMOVABLE;
356 set_pageblock_flags_group(page, (unsigned long)migratetype,
357 PB_migrate, PB_migrate_end);
360 #ifdef CONFIG_DEBUG_VM
361 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
365 unsigned long pfn = page_to_pfn(page);
366 unsigned long sp, start_pfn;
369 seq = zone_span_seqbegin(zone);
370 start_pfn = zone->zone_start_pfn;
371 sp = zone->spanned_pages;
372 if (!zone_spans_pfn(zone, pfn))
374 } while (zone_span_seqretry(zone, seq));
377 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
378 pfn, zone_to_nid(zone), zone->name,
379 start_pfn, start_pfn + sp);
384 static int page_is_consistent(struct zone *zone, struct page *page)
386 if (!pfn_valid_within(page_to_pfn(page)))
388 if (zone != page_zone(page))
394 * Temporary debugging check for pages not lying within a given zone.
396 static int bad_range(struct zone *zone, struct page *page)
398 if (page_outside_zone_boundaries(zone, page))
400 if (!page_is_consistent(zone, page))
406 static inline int bad_range(struct zone *zone, struct page *page)
412 static void bad_page(struct page *page, const char *reason,
413 unsigned long bad_flags)
415 static unsigned long resume;
416 static unsigned long nr_shown;
417 static unsigned long nr_unshown;
419 /* Don't complain about poisoned pages */
420 if (PageHWPoison(page)) {
421 page_mapcount_reset(page); /* remove PageBuddy */
426 * Allow a burst of 60 reports, then keep quiet for that minute;
427 * or allow a steady drip of one report per second.
429 if (nr_shown == 60) {
430 if (time_before(jiffies, resume)) {
436 "BUG: Bad page state: %lu messages suppressed\n",
443 resume = jiffies + 60 * HZ;
445 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
446 current->comm, page_to_pfn(page));
447 dump_page_badflags(page, reason, bad_flags);
452 /* Leave bad fields for debug, except PageBuddy could make trouble */
453 page_mapcount_reset(page); /* remove PageBuddy */
454 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
458 * Higher-order pages are called "compound pages". They are structured thusly:
460 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
462 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
463 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
465 * The first tail page's ->compound_dtor holds the offset in array of compound
466 * page destructors. See compound_page_dtors.
468 * The first tail page's ->compound_order holds the order of allocation.
469 * This usage means that zero-order pages may not be compound.
472 static void free_compound_page(struct page *page)
474 __free_pages_ok(page, compound_order(page));
477 void prep_compound_page(struct page *page, unsigned int order)
480 int nr_pages = 1 << order;
482 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
483 set_compound_order(page, order);
485 for (i = 1; i < nr_pages; i++) {
486 struct page *p = page + i;
487 set_page_count(p, 0);
488 set_compound_head(p, page);
492 #ifdef CONFIG_DEBUG_PAGEALLOC
493 unsigned int _debug_guardpage_minorder;
494 bool _debug_pagealloc_enabled __read_mostly;
495 bool _debug_guardpage_enabled __read_mostly;
497 static int __init early_debug_pagealloc(char *buf)
502 if (strcmp(buf, "on") == 0)
503 _debug_pagealloc_enabled = true;
507 early_param("debug_pagealloc", early_debug_pagealloc);
509 static bool need_debug_guardpage(void)
511 /* If we don't use debug_pagealloc, we don't need guard page */
512 if (!debug_pagealloc_enabled())
518 static void init_debug_guardpage(void)
520 if (!debug_pagealloc_enabled())
523 _debug_guardpage_enabled = true;
526 struct page_ext_operations debug_guardpage_ops = {
527 .need = need_debug_guardpage,
528 .init = init_debug_guardpage,
531 static int __init debug_guardpage_minorder_setup(char *buf)
535 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
536 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
539 _debug_guardpage_minorder = res;
540 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
543 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
545 static inline void set_page_guard(struct zone *zone, struct page *page,
546 unsigned int order, int migratetype)
548 struct page_ext *page_ext;
550 if (!debug_guardpage_enabled())
553 page_ext = lookup_page_ext(page);
554 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
556 INIT_LIST_HEAD(&page->lru);
557 set_page_private(page, order);
558 /* Guard pages are not available for any usage */
559 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
562 static inline void clear_page_guard(struct zone *zone, struct page *page,
563 unsigned int order, int migratetype)
565 struct page_ext *page_ext;
567 if (!debug_guardpage_enabled())
570 page_ext = lookup_page_ext(page);
571 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
573 set_page_private(page, 0);
574 if (!is_migrate_isolate(migratetype))
575 __mod_zone_freepage_state(zone, (1 << order), migratetype);
578 struct page_ext_operations debug_guardpage_ops = { NULL, };
579 static inline void set_page_guard(struct zone *zone, struct page *page,
580 unsigned int order, int migratetype) {}
581 static inline void clear_page_guard(struct zone *zone, struct page *page,
582 unsigned int order, int migratetype) {}
585 static inline void set_page_order(struct page *page, unsigned int order)
587 set_page_private(page, order);
588 __SetPageBuddy(page);
591 static inline void rmv_page_order(struct page *page)
593 __ClearPageBuddy(page);
594 set_page_private(page, 0);
598 * This function checks whether a page is free && is the buddy
599 * we can do coalesce a page and its buddy if
600 * (a) the buddy is not in a hole &&
601 * (b) the buddy is in the buddy system &&
602 * (c) a page and its buddy have the same order &&
603 * (d) a page and its buddy are in the same zone.
605 * For recording whether a page is in the buddy system, we set ->_mapcount
606 * PAGE_BUDDY_MAPCOUNT_VALUE.
607 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
608 * serialized by zone->lock.
610 * For recording page's order, we use page_private(page).
612 static inline int page_is_buddy(struct page *page, struct page *buddy,
615 if (!pfn_valid_within(page_to_pfn(buddy)))
618 if (page_is_guard(buddy) && page_order(buddy) == order) {
619 if (page_zone_id(page) != page_zone_id(buddy))
622 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
627 if (PageBuddy(buddy) && page_order(buddy) == order) {
629 * zone check is done late to avoid uselessly
630 * calculating zone/node ids for pages that could
633 if (page_zone_id(page) != page_zone_id(buddy))
636 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
644 * Freeing function for a buddy system allocator.
646 * The concept of a buddy system is to maintain direct-mapped table
647 * (containing bit values) for memory blocks of various "orders".
648 * The bottom level table contains the map for the smallest allocatable
649 * units of memory (here, pages), and each level above it describes
650 * pairs of units from the levels below, hence, "buddies".
651 * At a high level, all that happens here is marking the table entry
652 * at the bottom level available, and propagating the changes upward
653 * as necessary, plus some accounting needed to play nicely with other
654 * parts of the VM system.
655 * At each level, we keep a list of pages, which are heads of continuous
656 * free pages of length of (1 << order) and marked with _mapcount
657 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
659 * So when we are allocating or freeing one, we can derive the state of the
660 * other. That is, if we allocate a small block, and both were
661 * free, the remainder of the region must be split into blocks.
662 * If a block is freed, and its buddy is also free, then this
663 * triggers coalescing into a block of larger size.
668 static inline void __free_one_page(struct page *page,
670 struct zone *zone, unsigned int order,
673 unsigned long page_idx;
674 unsigned long combined_idx;
675 unsigned long uninitialized_var(buddy_idx);
677 unsigned int max_order;
679 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
681 VM_BUG_ON(!zone_is_initialized(zone));
682 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
684 VM_BUG_ON(migratetype == -1);
685 if (likely(!is_migrate_isolate(migratetype)))
686 __mod_zone_freepage_state(zone, 1 << order, migratetype);
688 page_idx = pfn & ((1 << MAX_ORDER) - 1);
690 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
691 VM_BUG_ON_PAGE(bad_range(zone, page), page);
694 while (order < max_order - 1) {
695 buddy_idx = __find_buddy_index(page_idx, order);
696 buddy = page + (buddy_idx - page_idx);
697 if (!page_is_buddy(page, buddy, order))
700 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
701 * merge with it and move up one order.
703 if (page_is_guard(buddy)) {
704 clear_page_guard(zone, buddy, order, migratetype);
706 list_del(&buddy->lru);
707 zone->free_area[order].nr_free--;
708 rmv_page_order(buddy);
710 combined_idx = buddy_idx & page_idx;
711 page = page + (combined_idx - page_idx);
712 page_idx = combined_idx;
715 if (max_order < MAX_ORDER) {
716 /* If we are here, it means order is >= pageblock_order.
717 * We want to prevent merge between freepages on isolate
718 * pageblock and normal pageblock. Without this, pageblock
719 * isolation could cause incorrect freepage or CMA accounting.
721 * We don't want to hit this code for the more frequent
724 if (unlikely(has_isolate_pageblock(zone))) {
727 buddy_idx = __find_buddy_index(page_idx, order);
728 buddy = page + (buddy_idx - page_idx);
729 buddy_mt = get_pageblock_migratetype(buddy);
731 if (migratetype != buddy_mt
732 && (is_migrate_isolate(migratetype) ||
733 is_migrate_isolate(buddy_mt)))
737 goto continue_merging;
741 set_page_order(page, order);
744 * If this is not the largest possible page, check if the buddy
745 * of the next-highest order is free. If it is, it's possible
746 * that pages are being freed that will coalesce soon. In case,
747 * that is happening, add the free page to the tail of the list
748 * so it's less likely to be used soon and more likely to be merged
749 * as a higher order page
751 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
752 struct page *higher_page, *higher_buddy;
753 combined_idx = buddy_idx & page_idx;
754 higher_page = page + (combined_idx - page_idx);
755 buddy_idx = __find_buddy_index(combined_idx, order + 1);
756 higher_buddy = higher_page + (buddy_idx - combined_idx);
757 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
758 list_add_tail(&page->lru,
759 &zone->free_area[order].free_list[migratetype]);
764 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
766 zone->free_area[order].nr_free++;
769 static inline int free_pages_check(struct page *page)
771 const char *bad_reason = NULL;
772 unsigned long bad_flags = 0;
774 if (unlikely(page_mapcount(page)))
775 bad_reason = "nonzero mapcount";
776 if (unlikely(page->mapping != NULL))
777 bad_reason = "non-NULL mapping";
778 if (unlikely(atomic_read(&page->_count) != 0))
779 bad_reason = "nonzero _count";
780 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
781 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
782 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
785 if (unlikely(page->mem_cgroup))
786 bad_reason = "page still charged to cgroup";
788 if (unlikely(bad_reason)) {
789 bad_page(page, bad_reason, bad_flags);
792 page_cpupid_reset_last(page);
793 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
794 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
799 * Frees a number of pages from the PCP lists
800 * Assumes all pages on list are in same zone, and of same order.
801 * count is the number of pages to free.
803 * If the zone was previously in an "all pages pinned" state then look to
804 * see if this freeing clears that state.
806 * And clear the zone's pages_scanned counter, to hold off the "all pages are
807 * pinned" detection logic.
809 static void free_pcppages_bulk(struct zone *zone, int count,
810 struct per_cpu_pages *pcp)
815 unsigned long nr_scanned;
817 spin_lock(&zone->lock);
818 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
820 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
824 struct list_head *list;
827 * Remove pages from lists in a round-robin fashion. A
828 * batch_free count is maintained that is incremented when an
829 * empty list is encountered. This is so more pages are freed
830 * off fuller lists instead of spinning excessively around empty
835 if (++migratetype == MIGRATE_PCPTYPES)
837 list = &pcp->lists[migratetype];
838 } while (list_empty(list));
840 /* This is the only non-empty list. Free them all. */
841 if (batch_free == MIGRATE_PCPTYPES)
842 batch_free = to_free;
845 int mt; /* migratetype of the to-be-freed page */
847 page = list_entry(list->prev, struct page, lru);
848 /* must delete as __free_one_page list manipulates */
849 list_del(&page->lru);
851 mt = get_pcppage_migratetype(page);
852 /* MIGRATE_ISOLATE page should not go to pcplists */
853 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
854 /* Pageblock could have been isolated meanwhile */
855 if (unlikely(has_isolate_pageblock(zone)))
856 mt = get_pageblock_migratetype(page);
858 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
859 trace_mm_page_pcpu_drain(page, 0, mt);
860 } while (--to_free && --batch_free && !list_empty(list));
862 spin_unlock(&zone->lock);
865 static void free_one_page(struct zone *zone,
866 struct page *page, unsigned long pfn,
870 unsigned long nr_scanned;
871 spin_lock(&zone->lock);
872 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
874 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
876 if (unlikely(has_isolate_pageblock(zone) ||
877 is_migrate_isolate(migratetype))) {
878 migratetype = get_pfnblock_migratetype(page, pfn);
880 __free_one_page(page, pfn, zone, order, migratetype);
881 spin_unlock(&zone->lock);
884 static int free_tail_pages_check(struct page *head_page, struct page *page)
889 * We rely page->lru.next never has bit 0 set, unless the page
890 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
892 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
894 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
898 if (unlikely(!PageTail(page))) {
899 bad_page(page, "PageTail not set", 0);
902 if (unlikely(compound_head(page) != head_page)) {
903 bad_page(page, "compound_head not consistent", 0);
908 clear_compound_head(page);
912 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
913 unsigned long zone, int nid)
915 set_page_links(page, zone, nid, pfn);
916 init_page_count(page);
917 page_mapcount_reset(page);
918 page_cpupid_reset_last(page);
920 INIT_LIST_HEAD(&page->lru);
921 #ifdef WANT_PAGE_VIRTUAL
922 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
923 if (!is_highmem_idx(zone))
924 set_page_address(page, __va(pfn << PAGE_SHIFT));
928 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
931 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
934 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
935 static void init_reserved_page(unsigned long pfn)
940 if (!early_page_uninitialised(pfn))
943 nid = early_pfn_to_nid(pfn);
944 pgdat = NODE_DATA(nid);
946 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
947 struct zone *zone = &pgdat->node_zones[zid];
949 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
952 __init_single_pfn(pfn, zid, nid);
955 static inline void init_reserved_page(unsigned long pfn)
958 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
961 * Initialised pages do not have PageReserved set. This function is
962 * called for each range allocated by the bootmem allocator and
963 * marks the pages PageReserved. The remaining valid pages are later
964 * sent to the buddy page allocator.
966 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
968 unsigned long start_pfn = PFN_DOWN(start);
969 unsigned long end_pfn = PFN_UP(end);
971 for (; start_pfn < end_pfn; start_pfn++) {
972 if (pfn_valid(start_pfn)) {
973 struct page *page = pfn_to_page(start_pfn);
975 init_reserved_page(start_pfn);
977 /* Avoid false-positive PageTail() */
978 INIT_LIST_HEAD(&page->lru);
980 SetPageReserved(page);
985 static bool free_pages_prepare(struct page *page, unsigned int order)
987 bool compound = PageCompound(page);
990 VM_BUG_ON_PAGE(PageTail(page), page);
991 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
993 trace_mm_page_free(page, order);
994 kmemcheck_free_shadow(page, order);
995 kasan_free_pages(page, order);
998 page->mapping = NULL;
999 bad += free_pages_check(page);
1000 for (i = 1; i < (1 << order); i++) {
1002 bad += free_tail_pages_check(page, page + i);
1003 bad += free_pages_check(page + i);
1008 reset_page_owner(page, order);
1010 if (!PageHighMem(page)) {
1011 debug_check_no_locks_freed(page_address(page),
1012 PAGE_SIZE << order);
1013 debug_check_no_obj_freed(page_address(page),
1014 PAGE_SIZE << order);
1016 arch_free_page(page, order);
1017 kernel_map_pages(page, 1 << order, 0);
1022 static void __free_pages_ok(struct page *page, unsigned int order)
1024 unsigned long flags;
1026 unsigned long pfn = page_to_pfn(page);
1028 if (!free_pages_prepare(page, order))
1031 migratetype = get_pfnblock_migratetype(page, pfn);
1032 local_irq_save(flags);
1033 __count_vm_events(PGFREE, 1 << order);
1034 free_one_page(page_zone(page), page, pfn, order, migratetype);
1035 local_irq_restore(flags);
1038 static void __init __free_pages_boot_core(struct page *page,
1039 unsigned long pfn, unsigned int order)
1041 unsigned int nr_pages = 1 << order;
1042 struct page *p = page;
1046 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1048 __ClearPageReserved(p);
1049 set_page_count(p, 0);
1051 __ClearPageReserved(p);
1052 set_page_count(p, 0);
1054 page_zone(page)->managed_pages += nr_pages;
1055 set_page_refcounted(page);
1056 __free_pages(page, order);
1059 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1060 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1062 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1064 int __meminit early_pfn_to_nid(unsigned long pfn)
1066 static DEFINE_SPINLOCK(early_pfn_lock);
1069 spin_lock(&early_pfn_lock);
1070 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1073 spin_unlock(&early_pfn_lock);
1079 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1080 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1081 struct mminit_pfnnid_cache *state)
1085 nid = __early_pfn_to_nid(pfn, state);
1086 if (nid >= 0 && nid != node)
1091 /* Only safe to use early in boot when initialisation is single-threaded */
1092 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1094 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1099 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1103 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1104 struct mminit_pfnnid_cache *state)
1111 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1114 if (early_page_uninitialised(pfn))
1116 return __free_pages_boot_core(page, pfn, order);
1119 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1120 static void __init deferred_free_range(struct page *page,
1121 unsigned long pfn, int nr_pages)
1128 /* Free a large naturally-aligned chunk if possible */
1129 if (nr_pages == MAX_ORDER_NR_PAGES &&
1130 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1131 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1132 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1136 for (i = 0; i < nr_pages; i++, page++, pfn++)
1137 __free_pages_boot_core(page, pfn, 0);
1140 /* Completion tracking for deferred_init_memmap() threads */
1141 static atomic_t pgdat_init_n_undone __initdata;
1142 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1144 static inline void __init pgdat_init_report_one_done(void)
1146 if (atomic_dec_and_test(&pgdat_init_n_undone))
1147 complete(&pgdat_init_all_done_comp);
1150 /* Initialise remaining memory on a node */
1151 static int __init deferred_init_memmap(void *data)
1153 pg_data_t *pgdat = data;
1154 int nid = pgdat->node_id;
1155 struct mminit_pfnnid_cache nid_init_state = { };
1156 unsigned long start = jiffies;
1157 unsigned long nr_pages = 0;
1158 unsigned long walk_start, walk_end;
1161 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1162 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1164 if (first_init_pfn == ULONG_MAX) {
1165 pgdat_init_report_one_done();
1169 /* Bind memory initialisation thread to a local node if possible */
1170 if (!cpumask_empty(cpumask))
1171 set_cpus_allowed_ptr(current, cpumask);
1173 /* Sanity check boundaries */
1174 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1175 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1176 pgdat->first_deferred_pfn = ULONG_MAX;
1178 /* Only the highest zone is deferred so find it */
1179 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1180 zone = pgdat->node_zones + zid;
1181 if (first_init_pfn < zone_end_pfn(zone))
1185 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1186 unsigned long pfn, end_pfn;
1187 struct page *page = NULL;
1188 struct page *free_base_page = NULL;
1189 unsigned long free_base_pfn = 0;
1192 end_pfn = min(walk_end, zone_end_pfn(zone));
1193 pfn = first_init_pfn;
1194 if (pfn < walk_start)
1196 if (pfn < zone->zone_start_pfn)
1197 pfn = zone->zone_start_pfn;
1199 for (; pfn < end_pfn; pfn++) {
1200 if (!pfn_valid_within(pfn))
1204 * Ensure pfn_valid is checked every
1205 * MAX_ORDER_NR_PAGES for memory holes
1207 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1208 if (!pfn_valid(pfn)) {
1214 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1219 /* Minimise pfn page lookups and scheduler checks */
1220 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1223 nr_pages += nr_to_free;
1224 deferred_free_range(free_base_page,
1225 free_base_pfn, nr_to_free);
1226 free_base_page = NULL;
1227 free_base_pfn = nr_to_free = 0;
1229 page = pfn_to_page(pfn);
1234 VM_BUG_ON(page_zone(page) != zone);
1238 __init_single_page(page, pfn, zid, nid);
1239 if (!free_base_page) {
1240 free_base_page = page;
1241 free_base_pfn = pfn;
1246 /* Where possible, batch up pages for a single free */
1249 /* Free the current block of pages to allocator */
1250 nr_pages += nr_to_free;
1251 deferred_free_range(free_base_page, free_base_pfn,
1253 free_base_page = NULL;
1254 free_base_pfn = nr_to_free = 0;
1257 first_init_pfn = max(end_pfn, first_init_pfn);
1260 /* Sanity check that the next zone really is unpopulated */
1261 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1263 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1264 jiffies_to_msecs(jiffies - start));
1266 pgdat_init_report_one_done();
1270 void __init page_alloc_init_late(void)
1274 /* There will be num_node_state(N_MEMORY) threads */
1275 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1276 for_each_node_state(nid, N_MEMORY) {
1277 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1280 /* Block until all are initialised */
1281 wait_for_completion(&pgdat_init_all_done_comp);
1283 /* Reinit limits that are based on free pages after the kernel is up */
1284 files_maxfiles_init();
1286 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1289 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1290 void __init init_cma_reserved_pageblock(struct page *page)
1292 unsigned i = pageblock_nr_pages;
1293 struct page *p = page;
1296 __ClearPageReserved(p);
1297 set_page_count(p, 0);
1300 set_pageblock_migratetype(page, MIGRATE_CMA);
1302 if (pageblock_order >= MAX_ORDER) {
1303 i = pageblock_nr_pages;
1306 set_page_refcounted(p);
1307 __free_pages(p, MAX_ORDER - 1);
1308 p += MAX_ORDER_NR_PAGES;
1309 } while (i -= MAX_ORDER_NR_PAGES);
1311 set_page_refcounted(page);
1312 __free_pages(page, pageblock_order);
1315 adjust_managed_page_count(page, pageblock_nr_pages);
1320 * The order of subdivision here is critical for the IO subsystem.
1321 * Please do not alter this order without good reasons and regression
1322 * testing. Specifically, as large blocks of memory are subdivided,
1323 * the order in which smaller blocks are delivered depends on the order
1324 * they're subdivided in this function. This is the primary factor
1325 * influencing the order in which pages are delivered to the IO
1326 * subsystem according to empirical testing, and this is also justified
1327 * by considering the behavior of a buddy system containing a single
1328 * large block of memory acted on by a series of small allocations.
1329 * This behavior is a critical factor in sglist merging's success.
1333 static inline void expand(struct zone *zone, struct page *page,
1334 int low, int high, struct free_area *area,
1337 unsigned long size = 1 << high;
1339 while (high > low) {
1343 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1345 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1346 debug_guardpage_enabled() &&
1347 high < debug_guardpage_minorder()) {
1349 * Mark as guard pages (or page), that will allow to
1350 * merge back to allocator when buddy will be freed.
1351 * Corresponding page table entries will not be touched,
1352 * pages will stay not present in virtual address space
1354 set_page_guard(zone, &page[size], high, migratetype);
1357 list_add(&page[size].lru, &area->free_list[migratetype]);
1359 set_page_order(&page[size], high);
1364 * This page is about to be returned from the page allocator
1366 static inline int check_new_page(struct page *page)
1368 const char *bad_reason = NULL;
1369 unsigned long bad_flags = 0;
1371 if (unlikely(page_mapcount(page)))
1372 bad_reason = "nonzero mapcount";
1373 if (unlikely(page->mapping != NULL))
1374 bad_reason = "non-NULL mapping";
1375 if (unlikely(atomic_read(&page->_count) != 0))
1376 bad_reason = "nonzero _count";
1377 if (unlikely(page->flags & __PG_HWPOISON)) {
1378 bad_reason = "HWPoisoned (hardware-corrupted)";
1379 bad_flags = __PG_HWPOISON;
1381 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1382 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1383 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1386 if (unlikely(page->mem_cgroup))
1387 bad_reason = "page still charged to cgroup";
1389 if (unlikely(bad_reason)) {
1390 bad_page(page, bad_reason, bad_flags);
1396 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1401 for (i = 0; i < (1 << order); i++) {
1402 struct page *p = page + i;
1403 if (unlikely(check_new_page(p)))
1407 set_page_private(page, 0);
1408 set_page_refcounted(page);
1410 arch_alloc_page(page, order);
1411 kernel_map_pages(page, 1 << order, 1);
1412 kasan_alloc_pages(page, order);
1414 if (gfp_flags & __GFP_ZERO)
1415 for (i = 0; i < (1 << order); i++)
1416 clear_highpage(page + i);
1418 if (order && (gfp_flags & __GFP_COMP))
1419 prep_compound_page(page, order);
1421 set_page_owner(page, order, gfp_flags);
1424 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1425 * allocate the page. The expectation is that the caller is taking
1426 * steps that will free more memory. The caller should avoid the page
1427 * being used for !PFMEMALLOC purposes.
1429 if (alloc_flags & ALLOC_NO_WATERMARKS)
1430 set_page_pfmemalloc(page);
1432 clear_page_pfmemalloc(page);
1438 * Go through the free lists for the given migratetype and remove
1439 * the smallest available page from the freelists
1442 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1445 unsigned int current_order;
1446 struct free_area *area;
1449 /* Find a page of the appropriate size in the preferred list */
1450 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1451 area = &(zone->free_area[current_order]);
1452 if (list_empty(&area->free_list[migratetype]))
1455 page = list_entry(area->free_list[migratetype].next,
1457 list_del(&page->lru);
1458 rmv_page_order(page);
1460 expand(zone, page, order, current_order, area, migratetype);
1461 set_pcppage_migratetype(page, migratetype);
1470 * This array describes the order lists are fallen back to when
1471 * the free lists for the desirable migrate type are depleted
1473 static int fallbacks[MIGRATE_TYPES][4] = {
1474 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1475 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1476 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1478 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1480 #ifdef CONFIG_MEMORY_ISOLATION
1481 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1486 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1489 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1492 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1493 unsigned int order) { return NULL; }
1497 * Move the free pages in a range to the free lists of the requested type.
1498 * Note that start_page and end_pages are not aligned on a pageblock
1499 * boundary. If alignment is required, use move_freepages_block()
1501 int move_freepages(struct zone *zone,
1502 struct page *start_page, struct page *end_page,
1507 int pages_moved = 0;
1509 #ifndef CONFIG_HOLES_IN_ZONE
1511 * page_zone is not safe to call in this context when
1512 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1513 * anyway as we check zone boundaries in move_freepages_block().
1514 * Remove at a later date when no bug reports exist related to
1515 * grouping pages by mobility
1517 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1520 for (page = start_page; page <= end_page;) {
1521 /* Make sure we are not inadvertently changing nodes */
1522 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1524 if (!pfn_valid_within(page_to_pfn(page))) {
1529 if (!PageBuddy(page)) {
1534 order = page_order(page);
1535 list_move(&page->lru,
1536 &zone->free_area[order].free_list[migratetype]);
1538 pages_moved += 1 << order;
1544 int move_freepages_block(struct zone *zone, struct page *page,
1547 unsigned long start_pfn, end_pfn;
1548 struct page *start_page, *end_page;
1550 start_pfn = page_to_pfn(page);
1551 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1552 start_page = pfn_to_page(start_pfn);
1553 end_page = start_page + pageblock_nr_pages - 1;
1554 end_pfn = start_pfn + pageblock_nr_pages - 1;
1556 /* Do not cross zone boundaries */
1557 if (!zone_spans_pfn(zone, start_pfn))
1559 if (!zone_spans_pfn(zone, end_pfn))
1562 return move_freepages(zone, start_page, end_page, migratetype);
1565 static void change_pageblock_range(struct page *pageblock_page,
1566 int start_order, int migratetype)
1568 int nr_pageblocks = 1 << (start_order - pageblock_order);
1570 while (nr_pageblocks--) {
1571 set_pageblock_migratetype(pageblock_page, migratetype);
1572 pageblock_page += pageblock_nr_pages;
1577 * When we are falling back to another migratetype during allocation, try to
1578 * steal extra free pages from the same pageblocks to satisfy further
1579 * allocations, instead of polluting multiple pageblocks.
1581 * If we are stealing a relatively large buddy page, it is likely there will
1582 * be more free pages in the pageblock, so try to steal them all. For
1583 * reclaimable and unmovable allocations, we steal regardless of page size,
1584 * as fragmentation caused by those allocations polluting movable pageblocks
1585 * is worse than movable allocations stealing from unmovable and reclaimable
1588 static bool can_steal_fallback(unsigned int order, int start_mt)
1591 * Leaving this order check is intended, although there is
1592 * relaxed order check in next check. The reason is that
1593 * we can actually steal whole pageblock if this condition met,
1594 * but, below check doesn't guarantee it and that is just heuristic
1595 * so could be changed anytime.
1597 if (order >= pageblock_order)
1600 if (order >= pageblock_order / 2 ||
1601 start_mt == MIGRATE_RECLAIMABLE ||
1602 start_mt == MIGRATE_UNMOVABLE ||
1603 page_group_by_mobility_disabled)
1610 * This function implements actual steal behaviour. If order is large enough,
1611 * we can steal whole pageblock. If not, we first move freepages in this
1612 * pageblock and check whether half of pages are moved or not. If half of
1613 * pages are moved, we can change migratetype of pageblock and permanently
1614 * use it's pages as requested migratetype in the future.
1616 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1619 unsigned int current_order = page_order(page);
1622 /* Take ownership for orders >= pageblock_order */
1623 if (current_order >= pageblock_order) {
1624 change_pageblock_range(page, current_order, start_type);
1628 pages = move_freepages_block(zone, page, start_type);
1630 /* Claim the whole block if over half of it is free */
1631 if (pages >= (1 << (pageblock_order-1)) ||
1632 page_group_by_mobility_disabled)
1633 set_pageblock_migratetype(page, start_type);
1637 * Check whether there is a suitable fallback freepage with requested order.
1638 * If only_stealable is true, this function returns fallback_mt only if
1639 * we can steal other freepages all together. This would help to reduce
1640 * fragmentation due to mixed migratetype pages in one pageblock.
1642 int find_suitable_fallback(struct free_area *area, unsigned int order,
1643 int migratetype, bool only_stealable, bool *can_steal)
1648 if (area->nr_free == 0)
1653 fallback_mt = fallbacks[migratetype][i];
1654 if (fallback_mt == MIGRATE_TYPES)
1657 if (list_empty(&area->free_list[fallback_mt]))
1660 if (can_steal_fallback(order, migratetype))
1663 if (!only_stealable)
1674 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1675 * there are no empty page blocks that contain a page with a suitable order
1677 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1678 unsigned int alloc_order)
1681 unsigned long max_managed, flags;
1684 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1685 * Check is race-prone but harmless.
1687 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1688 if (zone->nr_reserved_highatomic >= max_managed)
1691 spin_lock_irqsave(&zone->lock, flags);
1693 /* Recheck the nr_reserved_highatomic limit under the lock */
1694 if (zone->nr_reserved_highatomic >= max_managed)
1698 mt = get_pageblock_migratetype(page);
1699 if (mt != MIGRATE_HIGHATOMIC &&
1700 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1701 zone->nr_reserved_highatomic += pageblock_nr_pages;
1702 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1703 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1707 spin_unlock_irqrestore(&zone->lock, flags);
1711 * Used when an allocation is about to fail under memory pressure. This
1712 * potentially hurts the reliability of high-order allocations when under
1713 * intense memory pressure but failed atomic allocations should be easier
1714 * to recover from than an OOM.
1716 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1718 struct zonelist *zonelist = ac->zonelist;
1719 unsigned long flags;
1725 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1727 /* Preserve at least one pageblock */
1728 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1731 spin_lock_irqsave(&zone->lock, flags);
1732 for (order = 0; order < MAX_ORDER; order++) {
1733 struct free_area *area = &(zone->free_area[order]);
1735 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1738 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1742 * It should never happen but changes to locking could
1743 * inadvertently allow a per-cpu drain to add pages
1744 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1745 * and watch for underflows.
1747 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1748 zone->nr_reserved_highatomic);
1751 * Convert to ac->migratetype and avoid the normal
1752 * pageblock stealing heuristics. Minimally, the caller
1753 * is doing the work and needs the pages. More
1754 * importantly, if the block was always converted to
1755 * MIGRATE_UNMOVABLE or another type then the number
1756 * of pageblocks that cannot be completely freed
1759 set_pageblock_migratetype(page, ac->migratetype);
1760 move_freepages_block(zone, page, ac->migratetype);
1761 spin_unlock_irqrestore(&zone->lock, flags);
1764 spin_unlock_irqrestore(&zone->lock, flags);
1768 /* Remove an element from the buddy allocator from the fallback list */
1769 static inline struct page *
1770 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1772 struct free_area *area;
1773 unsigned int current_order;
1778 /* Find the largest possible block of pages in the other list */
1779 for (current_order = MAX_ORDER-1;
1780 current_order >= order && current_order <= MAX_ORDER-1;
1782 area = &(zone->free_area[current_order]);
1783 fallback_mt = find_suitable_fallback(area, current_order,
1784 start_migratetype, false, &can_steal);
1785 if (fallback_mt == -1)
1788 page = list_entry(area->free_list[fallback_mt].next,
1791 steal_suitable_fallback(zone, page, start_migratetype);
1793 /* Remove the page from the freelists */
1795 list_del(&page->lru);
1796 rmv_page_order(page);
1798 expand(zone, page, order, current_order, area,
1801 * The pcppage_migratetype may differ from pageblock's
1802 * migratetype depending on the decisions in
1803 * find_suitable_fallback(). This is OK as long as it does not
1804 * differ for MIGRATE_CMA pageblocks. Those can be used as
1805 * fallback only via special __rmqueue_cma_fallback() function
1807 set_pcppage_migratetype(page, start_migratetype);
1809 trace_mm_page_alloc_extfrag(page, order, current_order,
1810 start_migratetype, fallback_mt);
1819 * Do the hard work of removing an element from the buddy allocator.
1820 * Call me with the zone->lock already held.
1822 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1823 int migratetype, gfp_t gfp_flags)
1827 page = __rmqueue_smallest(zone, order, migratetype);
1828 if (unlikely(!page)) {
1829 if (migratetype == MIGRATE_MOVABLE)
1830 page = __rmqueue_cma_fallback(zone, order);
1833 page = __rmqueue_fallback(zone, order, migratetype);
1836 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1841 * Obtain a specified number of elements from the buddy allocator, all under
1842 * a single hold of the lock, for efficiency. Add them to the supplied list.
1843 * Returns the number of new pages which were placed at *list.
1845 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1846 unsigned long count, struct list_head *list,
1847 int migratetype, bool cold)
1851 spin_lock(&zone->lock);
1852 for (i = 0; i < count; ++i) {
1853 struct page *page = __rmqueue(zone, order, migratetype, 0);
1854 if (unlikely(page == NULL))
1858 * Split buddy pages returned by expand() are received here
1859 * in physical page order. The page is added to the callers and
1860 * list and the list head then moves forward. From the callers
1861 * perspective, the linked list is ordered by page number in
1862 * some conditions. This is useful for IO devices that can
1863 * merge IO requests if the physical pages are ordered
1867 list_add(&page->lru, list);
1869 list_add_tail(&page->lru, list);
1871 if (is_migrate_cma(get_pcppage_migratetype(page)))
1872 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1875 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1876 spin_unlock(&zone->lock);
1882 * Called from the vmstat counter updater to drain pagesets of this
1883 * currently executing processor on remote nodes after they have
1886 * Note that this function must be called with the thread pinned to
1887 * a single processor.
1889 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1891 unsigned long flags;
1892 int to_drain, batch;
1894 local_irq_save(flags);
1895 batch = READ_ONCE(pcp->batch);
1896 to_drain = min(pcp->count, batch);
1898 free_pcppages_bulk(zone, to_drain, pcp);
1899 pcp->count -= to_drain;
1901 local_irq_restore(flags);
1906 * Drain pcplists of the indicated processor and zone.
1908 * The processor must either be the current processor and the
1909 * thread pinned to the current processor or a processor that
1912 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1914 unsigned long flags;
1915 struct per_cpu_pageset *pset;
1916 struct per_cpu_pages *pcp;
1918 local_irq_save(flags);
1919 pset = per_cpu_ptr(zone->pageset, cpu);
1923 free_pcppages_bulk(zone, pcp->count, pcp);
1926 local_irq_restore(flags);
1930 * Drain pcplists of all zones on the indicated processor.
1932 * The processor must either be the current processor and the
1933 * thread pinned to the current processor or a processor that
1936 static void drain_pages(unsigned int cpu)
1940 for_each_populated_zone(zone) {
1941 drain_pages_zone(cpu, zone);
1946 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1948 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1949 * the single zone's pages.
1951 void drain_local_pages(struct zone *zone)
1953 int cpu = smp_processor_id();
1956 drain_pages_zone(cpu, zone);
1962 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1964 * When zone parameter is non-NULL, spill just the single zone's pages.
1966 * Note that this code is protected against sending an IPI to an offline
1967 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1968 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1969 * nothing keeps CPUs from showing up after we populated the cpumask and
1970 * before the call to on_each_cpu_mask().
1972 void drain_all_pages(struct zone *zone)
1977 * Allocate in the BSS so we wont require allocation in
1978 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1980 static cpumask_t cpus_with_pcps;
1983 * We don't care about racing with CPU hotplug event
1984 * as offline notification will cause the notified
1985 * cpu to drain that CPU pcps and on_each_cpu_mask
1986 * disables preemption as part of its processing
1988 for_each_online_cpu(cpu) {
1989 struct per_cpu_pageset *pcp;
1991 bool has_pcps = false;
1994 pcp = per_cpu_ptr(zone->pageset, cpu);
1998 for_each_populated_zone(z) {
1999 pcp = per_cpu_ptr(z->pageset, cpu);
2000 if (pcp->pcp.count) {
2008 cpumask_set_cpu(cpu, &cpus_with_pcps);
2010 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2012 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2016 #ifdef CONFIG_HIBERNATION
2018 void mark_free_pages(struct zone *zone)
2020 unsigned long pfn, max_zone_pfn;
2021 unsigned long flags;
2022 unsigned int order, t;
2023 struct list_head *curr;
2025 if (zone_is_empty(zone))
2028 spin_lock_irqsave(&zone->lock, flags);
2030 max_zone_pfn = zone_end_pfn(zone);
2031 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2032 if (pfn_valid(pfn)) {
2033 struct page *page = pfn_to_page(pfn);
2035 if (!swsusp_page_is_forbidden(page))
2036 swsusp_unset_page_free(page);
2039 for_each_migratetype_order(order, t) {
2040 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2043 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2044 for (i = 0; i < (1UL << order); i++)
2045 swsusp_set_page_free(pfn_to_page(pfn + i));
2048 spin_unlock_irqrestore(&zone->lock, flags);
2050 #endif /* CONFIG_PM */
2053 * Free a 0-order page
2054 * cold == true ? free a cold page : free a hot page
2056 void free_hot_cold_page(struct page *page, bool cold)
2058 struct zone *zone = page_zone(page);
2059 struct per_cpu_pages *pcp;
2060 unsigned long flags;
2061 unsigned long pfn = page_to_pfn(page);
2064 if (!free_pages_prepare(page, 0))
2067 migratetype = get_pfnblock_migratetype(page, pfn);
2068 set_pcppage_migratetype(page, migratetype);
2069 local_irq_save(flags);
2070 __count_vm_event(PGFREE);
2073 * We only track unmovable, reclaimable and movable on pcp lists.
2074 * Free ISOLATE pages back to the allocator because they are being
2075 * offlined but treat RESERVE as movable pages so we can get those
2076 * areas back if necessary. Otherwise, we may have to free
2077 * excessively into the page allocator
2079 if (migratetype >= MIGRATE_PCPTYPES) {
2080 if (unlikely(is_migrate_isolate(migratetype))) {
2081 free_one_page(zone, page, pfn, 0, migratetype);
2084 migratetype = MIGRATE_MOVABLE;
2087 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2089 list_add(&page->lru, &pcp->lists[migratetype]);
2091 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2093 if (pcp->count >= pcp->high) {
2094 unsigned long batch = READ_ONCE(pcp->batch);
2095 free_pcppages_bulk(zone, batch, pcp);
2096 pcp->count -= batch;
2100 local_irq_restore(flags);
2104 * Free a list of 0-order pages
2106 void free_hot_cold_page_list(struct list_head *list, bool cold)
2108 struct page *page, *next;
2110 list_for_each_entry_safe(page, next, list, lru) {
2111 trace_mm_page_free_batched(page, cold);
2112 free_hot_cold_page(page, cold);
2117 * split_page takes a non-compound higher-order page, and splits it into
2118 * n (1<<order) sub-pages: page[0..n]
2119 * Each sub-page must be freed individually.
2121 * Note: this is probably too low level an operation for use in drivers.
2122 * Please consult with lkml before using this in your driver.
2124 void split_page(struct page *page, unsigned int order)
2129 VM_BUG_ON_PAGE(PageCompound(page), page);
2130 VM_BUG_ON_PAGE(!page_count(page), page);
2132 #ifdef CONFIG_KMEMCHECK
2134 * Split shadow pages too, because free(page[0]) would
2135 * otherwise free the whole shadow.
2137 if (kmemcheck_page_is_tracked(page))
2138 split_page(virt_to_page(page[0].shadow), order);
2141 gfp_mask = get_page_owner_gfp(page);
2142 set_page_owner(page, 0, gfp_mask);
2143 for (i = 1; i < (1 << order); i++) {
2144 set_page_refcounted(page + i);
2145 set_page_owner(page + i, 0, gfp_mask);
2148 EXPORT_SYMBOL_GPL(split_page);
2150 int __isolate_free_page(struct page *page, unsigned int order)
2152 unsigned long watermark;
2156 BUG_ON(!PageBuddy(page));
2158 zone = page_zone(page);
2159 mt = get_pageblock_migratetype(page);
2161 if (!is_migrate_isolate(mt)) {
2162 /* Obey watermarks as if the page was being allocated */
2163 watermark = low_wmark_pages(zone) + (1 << order);
2164 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2167 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2170 /* Remove page from free list */
2171 list_del(&page->lru);
2172 zone->free_area[order].nr_free--;
2173 rmv_page_order(page);
2175 set_page_owner(page, order, __GFP_MOVABLE);
2177 /* Set the pageblock if the isolated page is at least a pageblock */
2178 if (order >= pageblock_order - 1) {
2179 struct page *endpage = page + (1 << order) - 1;
2180 for (; page < endpage; page += pageblock_nr_pages) {
2181 int mt = get_pageblock_migratetype(page);
2182 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2183 set_pageblock_migratetype(page,
2189 return 1UL << order;
2193 * Similar to split_page except the page is already free. As this is only
2194 * being used for migration, the migratetype of the block also changes.
2195 * As this is called with interrupts disabled, the caller is responsible
2196 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2199 * Note: this is probably too low level an operation for use in drivers.
2200 * Please consult with lkml before using this in your driver.
2202 int split_free_page(struct page *page)
2207 order = page_order(page);
2209 nr_pages = __isolate_free_page(page, order);
2213 /* Split into individual pages */
2214 set_page_refcounted(page);
2215 split_page(page, order);
2220 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2223 struct page *buffered_rmqueue(struct zone *preferred_zone,
2224 struct zone *zone, unsigned int order,
2225 gfp_t gfp_flags, int alloc_flags, int migratetype)
2227 unsigned long flags;
2229 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2231 if (likely(order == 0)) {
2232 struct per_cpu_pages *pcp;
2233 struct list_head *list;
2235 local_irq_save(flags);
2236 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2237 list = &pcp->lists[migratetype];
2238 if (list_empty(list)) {
2239 pcp->count += rmqueue_bulk(zone, 0,
2242 if (unlikely(list_empty(list)))
2247 page = list_entry(list->prev, struct page, lru);
2249 page = list_entry(list->next, struct page, lru);
2251 list_del(&page->lru);
2254 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2256 * __GFP_NOFAIL is not to be used in new code.
2258 * All __GFP_NOFAIL callers should be fixed so that they
2259 * properly detect and handle allocation failures.
2261 * We most definitely don't want callers attempting to
2262 * allocate greater than order-1 page units with
2265 WARN_ON_ONCE(order > 1);
2267 spin_lock_irqsave(&zone->lock, flags);
2270 if (alloc_flags & ALLOC_HARDER) {
2271 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2273 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2276 page = __rmqueue(zone, order, migratetype, gfp_flags);
2277 spin_unlock(&zone->lock);
2280 __mod_zone_freepage_state(zone, -(1 << order),
2281 get_pcppage_migratetype(page));
2284 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2285 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2286 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2287 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2289 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2290 zone_statistics(preferred_zone, zone, gfp_flags);
2291 local_irq_restore(flags);
2293 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2297 local_irq_restore(flags);
2301 #ifdef CONFIG_FAIL_PAGE_ALLOC
2304 struct fault_attr attr;
2306 bool ignore_gfp_highmem;
2307 bool ignore_gfp_reclaim;
2309 } fail_page_alloc = {
2310 .attr = FAULT_ATTR_INITIALIZER,
2311 .ignore_gfp_reclaim = true,
2312 .ignore_gfp_highmem = true,
2316 static int __init setup_fail_page_alloc(char *str)
2318 return setup_fault_attr(&fail_page_alloc.attr, str);
2320 __setup("fail_page_alloc=", setup_fail_page_alloc);
2322 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2324 if (order < fail_page_alloc.min_order)
2326 if (gfp_mask & __GFP_NOFAIL)
2328 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2330 if (fail_page_alloc.ignore_gfp_reclaim &&
2331 (gfp_mask & __GFP_DIRECT_RECLAIM))
2334 return should_fail(&fail_page_alloc.attr, 1 << order);
2337 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2339 static int __init fail_page_alloc_debugfs(void)
2341 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2344 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2345 &fail_page_alloc.attr);
2347 return PTR_ERR(dir);
2349 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2350 &fail_page_alloc.ignore_gfp_reclaim))
2352 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2353 &fail_page_alloc.ignore_gfp_highmem))
2355 if (!debugfs_create_u32("min-order", mode, dir,
2356 &fail_page_alloc.min_order))
2361 debugfs_remove_recursive(dir);
2366 late_initcall(fail_page_alloc_debugfs);
2368 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2370 #else /* CONFIG_FAIL_PAGE_ALLOC */
2372 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2377 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2380 * Return true if free base pages are above 'mark'. For high-order checks it
2381 * will return true of the order-0 watermark is reached and there is at least
2382 * one free page of a suitable size. Checking now avoids taking the zone lock
2383 * to check in the allocation paths if no pages are free.
2385 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2386 unsigned long mark, int classzone_idx, int alloc_flags,
2391 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2393 /* free_pages may go negative - that's OK */
2394 free_pages -= (1 << order) - 1;
2396 if (alloc_flags & ALLOC_HIGH)
2400 * If the caller does not have rights to ALLOC_HARDER then subtract
2401 * the high-atomic reserves. This will over-estimate the size of the
2402 * atomic reserve but it avoids a search.
2404 if (likely(!alloc_harder))
2405 free_pages -= z->nr_reserved_highatomic;
2410 /* If allocation can't use CMA areas don't use free CMA pages */
2411 if (!(alloc_flags & ALLOC_CMA))
2412 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2416 * Check watermarks for an order-0 allocation request. If these
2417 * are not met, then a high-order request also cannot go ahead
2418 * even if a suitable page happened to be free.
2420 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2423 /* If this is an order-0 request then the watermark is fine */
2427 /* For a high-order request, check at least one suitable page is free */
2428 for (o = order; o < MAX_ORDER; o++) {
2429 struct free_area *area = &z->free_area[o];
2438 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2439 if (!list_empty(&area->free_list[mt]))
2444 if ((alloc_flags & ALLOC_CMA) &&
2445 !list_empty(&area->free_list[MIGRATE_CMA])) {
2453 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2454 int classzone_idx, int alloc_flags)
2456 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2457 zone_page_state(z, NR_FREE_PAGES));
2460 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2461 unsigned long mark, int classzone_idx)
2463 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2465 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2466 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2468 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2473 static bool zone_local(struct zone *local_zone, struct zone *zone)
2475 return local_zone->node == zone->node;
2478 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2480 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2483 #else /* CONFIG_NUMA */
2484 static bool zone_local(struct zone *local_zone, struct zone *zone)
2489 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2493 #endif /* CONFIG_NUMA */
2495 static void reset_alloc_batches(struct zone *preferred_zone)
2497 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2500 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2501 high_wmark_pages(zone) - low_wmark_pages(zone) -
2502 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2503 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2504 } while (zone++ != preferred_zone);
2508 * get_page_from_freelist goes through the zonelist trying to allocate
2511 static struct page *
2512 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2513 const struct alloc_context *ac)
2515 struct zonelist *zonelist = ac->zonelist;
2517 struct page *page = NULL;
2519 int nr_fair_skipped = 0;
2520 bool zonelist_rescan;
2523 zonelist_rescan = false;
2526 * Scan zonelist, looking for a zone with enough free.
2527 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2529 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2533 if (cpusets_enabled() &&
2534 (alloc_flags & ALLOC_CPUSET) &&
2535 !cpuset_zone_allowed(zone, gfp_mask))
2538 * Distribute pages in proportion to the individual
2539 * zone size to ensure fair page aging. The zone a
2540 * page was allocated in should have no effect on the
2541 * time the page has in memory before being reclaimed.
2543 if (alloc_flags & ALLOC_FAIR) {
2544 if (!zone_local(ac->preferred_zone, zone))
2546 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2552 * When allocating a page cache page for writing, we
2553 * want to get it from a zone that is within its dirty
2554 * limit, such that no single zone holds more than its
2555 * proportional share of globally allowed dirty pages.
2556 * The dirty limits take into account the zone's
2557 * lowmem reserves and high watermark so that kswapd
2558 * should be able to balance it without having to
2559 * write pages from its LRU list.
2561 * This may look like it could increase pressure on
2562 * lower zones by failing allocations in higher zones
2563 * before they are full. But the pages that do spill
2564 * over are limited as the lower zones are protected
2565 * by this very same mechanism. It should not become
2566 * a practical burden to them.
2568 * XXX: For now, allow allocations to potentially
2569 * exceed the per-zone dirty limit in the slowpath
2570 * (spread_dirty_pages unset) before going into reclaim,
2571 * which is important when on a NUMA setup the allowed
2572 * zones are together not big enough to reach the
2573 * global limit. The proper fix for these situations
2574 * will require awareness of zones in the
2575 * dirty-throttling and the flusher threads.
2577 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2580 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2581 if (!zone_watermark_ok(zone, order, mark,
2582 ac->classzone_idx, alloc_flags)) {
2585 /* Checked here to keep the fast path fast */
2586 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2587 if (alloc_flags & ALLOC_NO_WATERMARKS)
2590 if (zone_reclaim_mode == 0 ||
2591 !zone_allows_reclaim(ac->preferred_zone, zone))
2594 ret = zone_reclaim(zone, gfp_mask, order);
2596 case ZONE_RECLAIM_NOSCAN:
2599 case ZONE_RECLAIM_FULL:
2600 /* scanned but unreclaimable */
2603 /* did we reclaim enough */
2604 if (zone_watermark_ok(zone, order, mark,
2605 ac->classzone_idx, alloc_flags))
2613 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2614 gfp_mask, alloc_flags, ac->migratetype);
2616 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2620 * If this is a high-order atomic allocation then check
2621 * if the pageblock should be reserved for the future
2623 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2624 reserve_highatomic_pageblock(page, zone, order);
2631 * The first pass makes sure allocations are spread fairly within the
2632 * local node. However, the local node might have free pages left
2633 * after the fairness batches are exhausted, and remote zones haven't
2634 * even been considered yet. Try once more without fairness, and
2635 * include remote zones now, before entering the slowpath and waking
2636 * kswapd: prefer spilling to a remote zone over swapping locally.
2638 if (alloc_flags & ALLOC_FAIR) {
2639 alloc_flags &= ~ALLOC_FAIR;
2640 if (nr_fair_skipped) {
2641 zonelist_rescan = true;
2642 reset_alloc_batches(ac->preferred_zone);
2644 if (nr_online_nodes > 1)
2645 zonelist_rescan = true;
2648 if (zonelist_rescan)
2655 * Large machines with many possible nodes should not always dump per-node
2656 * meminfo in irq context.
2658 static inline bool should_suppress_show_mem(void)
2663 ret = in_interrupt();
2668 static DEFINE_RATELIMIT_STATE(nopage_rs,
2669 DEFAULT_RATELIMIT_INTERVAL,
2670 DEFAULT_RATELIMIT_BURST);
2672 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2674 unsigned int filter = SHOW_MEM_FILTER_NODES;
2676 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2677 debug_guardpage_minorder() > 0)
2681 * This documents exceptions given to allocations in certain
2682 * contexts that are allowed to allocate outside current's set
2685 if (!(gfp_mask & __GFP_NOMEMALLOC))
2686 if (test_thread_flag(TIF_MEMDIE) ||
2687 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2688 filter &= ~SHOW_MEM_FILTER_NODES;
2689 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2690 filter &= ~SHOW_MEM_FILTER_NODES;
2693 struct va_format vaf;
2696 va_start(args, fmt);
2701 pr_warn("%pV", &vaf);
2706 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2707 current->comm, order, gfp_mask);
2710 if (!should_suppress_show_mem())
2714 static inline struct page *
2715 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2716 const struct alloc_context *ac, unsigned long *did_some_progress)
2718 struct oom_control oc = {
2719 .zonelist = ac->zonelist,
2720 .nodemask = ac->nodemask,
2721 .gfp_mask = gfp_mask,
2726 *did_some_progress = 0;
2729 * Acquire the oom lock. If that fails, somebody else is
2730 * making progress for us.
2732 if (!mutex_trylock(&oom_lock)) {
2733 *did_some_progress = 1;
2734 schedule_timeout_uninterruptible(1);
2739 * Go through the zonelist yet one more time, keep very high watermark
2740 * here, this is only to catch a parallel oom killing, we must fail if
2741 * we're still under heavy pressure.
2743 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2744 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2748 if (!(gfp_mask & __GFP_NOFAIL)) {
2749 /* Coredumps can quickly deplete all memory reserves */
2750 if (current->flags & PF_DUMPCORE)
2752 /* The OOM killer will not help higher order allocs */
2753 if (order > PAGE_ALLOC_COSTLY_ORDER)
2755 /* The OOM killer does not needlessly kill tasks for lowmem */
2756 if (ac->high_zoneidx < ZONE_NORMAL)
2758 /* The OOM killer does not compensate for IO-less reclaim */
2759 if (!(gfp_mask & __GFP_FS)) {
2761 * XXX: Page reclaim didn't yield anything,
2762 * and the OOM killer can't be invoked, but
2763 * keep looping as per tradition.
2765 *did_some_progress = 1;
2768 if (pm_suspended_storage())
2770 /* The OOM killer may not free memory on a specific node */
2771 if (gfp_mask & __GFP_THISNODE)
2774 /* Exhausted what can be done so it's blamo time */
2775 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2776 *did_some_progress = 1;
2778 mutex_unlock(&oom_lock);
2782 #ifdef CONFIG_COMPACTION
2783 /* Try memory compaction for high-order allocations before reclaim */
2784 static struct page *
2785 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2786 int alloc_flags, const struct alloc_context *ac,
2787 enum migrate_mode mode, int *contended_compaction,
2788 bool *deferred_compaction)
2790 unsigned long compact_result;
2796 current->flags |= PF_MEMALLOC;
2797 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2798 mode, contended_compaction);
2799 current->flags &= ~PF_MEMALLOC;
2801 switch (compact_result) {
2802 case COMPACT_DEFERRED:
2803 *deferred_compaction = true;
2805 case COMPACT_SKIPPED:
2812 * At least in one zone compaction wasn't deferred or skipped, so let's
2813 * count a compaction stall
2815 count_vm_event(COMPACTSTALL);
2817 page = get_page_from_freelist(gfp_mask, order,
2818 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2821 struct zone *zone = page_zone(page);
2823 zone->compact_blockskip_flush = false;
2824 compaction_defer_reset(zone, order, true);
2825 count_vm_event(COMPACTSUCCESS);
2830 * It's bad if compaction run occurs and fails. The most likely reason
2831 * is that pages exist, but not enough to satisfy watermarks.
2833 count_vm_event(COMPACTFAIL);
2840 static inline struct page *
2841 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2842 int alloc_flags, const struct alloc_context *ac,
2843 enum migrate_mode mode, int *contended_compaction,
2844 bool *deferred_compaction)
2848 #endif /* CONFIG_COMPACTION */
2850 /* Perform direct synchronous page reclaim */
2852 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2853 const struct alloc_context *ac)
2855 struct reclaim_state reclaim_state;
2860 /* We now go into synchronous reclaim */
2861 cpuset_memory_pressure_bump();
2862 current->flags |= PF_MEMALLOC;
2863 lockdep_set_current_reclaim_state(gfp_mask);
2864 reclaim_state.reclaimed_slab = 0;
2865 current->reclaim_state = &reclaim_state;
2867 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2870 current->reclaim_state = NULL;
2871 lockdep_clear_current_reclaim_state();
2872 current->flags &= ~PF_MEMALLOC;
2879 /* The really slow allocator path where we enter direct reclaim */
2880 static inline struct page *
2881 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2882 int alloc_flags, const struct alloc_context *ac,
2883 unsigned long *did_some_progress)
2885 struct page *page = NULL;
2886 bool drained = false;
2888 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2889 if (unlikely(!(*did_some_progress)))
2893 page = get_page_from_freelist(gfp_mask, order,
2894 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2897 * If an allocation failed after direct reclaim, it could be because
2898 * pages are pinned on the per-cpu lists or in high alloc reserves.
2899 * Shrink them them and try again
2901 if (!page && !drained) {
2902 unreserve_highatomic_pageblock(ac);
2903 drain_all_pages(NULL);
2912 * This is called in the allocator slow-path if the allocation request is of
2913 * sufficient urgency to ignore watermarks and take other desperate measures
2915 static inline struct page *
2916 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2917 const struct alloc_context *ac)
2922 page = get_page_from_freelist(gfp_mask, order,
2923 ALLOC_NO_WATERMARKS, ac);
2925 if (!page && gfp_mask & __GFP_NOFAIL)
2926 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2928 } while (!page && (gfp_mask & __GFP_NOFAIL));
2933 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2938 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2939 ac->high_zoneidx, ac->nodemask)
2940 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2944 gfp_to_alloc_flags(gfp_t gfp_mask)
2946 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2948 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2949 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2952 * The caller may dip into page reserves a bit more if the caller
2953 * cannot run direct reclaim, or if the caller has realtime scheduling
2954 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2955 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2957 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2959 if (gfp_mask & __GFP_ATOMIC) {
2961 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2962 * if it can't schedule.
2964 if (!(gfp_mask & __GFP_NOMEMALLOC))
2965 alloc_flags |= ALLOC_HARDER;
2967 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2968 * comment for __cpuset_node_allowed().
2970 alloc_flags &= ~ALLOC_CPUSET;
2971 } else if (unlikely(rt_task(current)) && !in_interrupt())
2972 alloc_flags |= ALLOC_HARDER;
2974 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2975 if (gfp_mask & __GFP_MEMALLOC)
2976 alloc_flags |= ALLOC_NO_WATERMARKS;
2977 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2978 alloc_flags |= ALLOC_NO_WATERMARKS;
2979 else if (!in_interrupt() &&
2980 ((current->flags & PF_MEMALLOC) ||
2981 unlikely(test_thread_flag(TIF_MEMDIE))))
2982 alloc_flags |= ALLOC_NO_WATERMARKS;
2985 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2986 alloc_flags |= ALLOC_CMA;
2991 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2993 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2996 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2998 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3001 static inline struct page *
3002 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3003 struct alloc_context *ac)
3005 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3006 struct page *page = NULL;
3008 unsigned long pages_reclaimed = 0;
3009 unsigned long did_some_progress;
3010 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3011 bool deferred_compaction = false;
3012 int contended_compaction = COMPACT_CONTENDED_NONE;
3015 * In the slowpath, we sanity check order to avoid ever trying to
3016 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3017 * be using allocators in order of preference for an area that is
3020 if (order >= MAX_ORDER) {
3021 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3026 * We also sanity check to catch abuse of atomic reserves being used by
3027 * callers that are not in atomic context.
3029 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3030 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3031 gfp_mask &= ~__GFP_ATOMIC;
3034 * If this allocation cannot block and it is for a specific node, then
3035 * fail early. There's no need to wakeup kswapd or retry for a
3036 * speculative node-specific allocation.
3038 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3042 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3043 wake_all_kswapds(order, ac);
3046 * OK, we're below the kswapd watermark and have kicked background
3047 * reclaim. Now things get more complex, so set up alloc_flags according
3048 * to how we want to proceed.
3050 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3053 * Find the true preferred zone if the allocation is unconstrained by
3056 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3057 struct zoneref *preferred_zoneref;
3058 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3059 ac->high_zoneidx, NULL, &ac->preferred_zone);
3060 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3063 /* This is the last chance, in general, before the goto nopage. */
3064 page = get_page_from_freelist(gfp_mask, order,
3065 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3069 /* Allocate without watermarks if the context allows */
3070 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3072 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3073 * the allocation is high priority and these type of
3074 * allocations are system rather than user orientated
3076 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3078 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3085 /* Caller is not willing to reclaim, we can't balance anything */
3086 if (!can_direct_reclaim) {
3088 * All existing users of the deprecated __GFP_NOFAIL are
3089 * blockable, so warn of any new users that actually allow this
3090 * type of allocation to fail.
3092 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3096 /* Avoid recursion of direct reclaim */
3097 if (current->flags & PF_MEMALLOC)
3100 /* Avoid allocations with no watermarks from looping endlessly */
3101 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3105 * Try direct compaction. The first pass is asynchronous. Subsequent
3106 * attempts after direct reclaim are synchronous
3108 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3110 &contended_compaction,
3111 &deferred_compaction);
3115 /* Checks for THP-specific high-order allocations */
3116 if (is_thp_gfp_mask(gfp_mask)) {
3118 * If compaction is deferred for high-order allocations, it is
3119 * because sync compaction recently failed. If this is the case
3120 * and the caller requested a THP allocation, we do not want
3121 * to heavily disrupt the system, so we fail the allocation
3122 * instead of entering direct reclaim.
3124 if (deferred_compaction)
3128 * In all zones where compaction was attempted (and not
3129 * deferred or skipped), lock contention has been detected.
3130 * For THP allocation we do not want to disrupt the others
3131 * so we fallback to base pages instead.
3133 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3137 * If compaction was aborted due to need_resched(), we do not
3138 * want to further increase allocation latency, unless it is
3139 * khugepaged trying to collapse.
3141 if (contended_compaction == COMPACT_CONTENDED_SCHED
3142 && !(current->flags & PF_KTHREAD))
3147 * It can become very expensive to allocate transparent hugepages at
3148 * fault, so use asynchronous memory compaction for THP unless it is
3149 * khugepaged trying to collapse.
3151 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3152 migration_mode = MIGRATE_SYNC_LIGHT;
3154 /* Try direct reclaim and then allocating */
3155 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3156 &did_some_progress);
3160 /* Do not loop if specifically requested */
3161 if (gfp_mask & __GFP_NORETRY)
3164 /* Keep reclaiming pages as long as there is reasonable progress */
3165 pages_reclaimed += did_some_progress;
3166 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3167 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3168 /* Wait for some write requests to complete then retry */
3169 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3173 /* Reclaim has failed us, start killing things */
3174 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3178 /* Retry as long as the OOM killer is making progress */
3179 if (did_some_progress)
3184 * High-order allocations do not necessarily loop after
3185 * direct reclaim and reclaim/compaction depends on compaction
3186 * being called after reclaim so call directly if necessary
3188 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3190 &contended_compaction,
3191 &deferred_compaction);
3195 warn_alloc_failed(gfp_mask, order, NULL);
3201 * This is the 'heart' of the zoned buddy allocator.
3204 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3205 struct zonelist *zonelist, nodemask_t *nodemask)
3207 struct zoneref *preferred_zoneref;
3208 struct page *page = NULL;
3209 unsigned int cpuset_mems_cookie;
3210 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3211 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3212 struct alloc_context ac = {
3213 .high_zoneidx = gfp_zone(gfp_mask),
3214 .nodemask = nodemask,
3215 .migratetype = gfpflags_to_migratetype(gfp_mask),
3218 gfp_mask &= gfp_allowed_mask;
3220 lockdep_trace_alloc(gfp_mask);
3222 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3224 if (should_fail_alloc_page(gfp_mask, order))
3228 * Check the zones suitable for the gfp_mask contain at least one
3229 * valid zone. It's possible to have an empty zonelist as a result
3230 * of __GFP_THISNODE and a memoryless node
3232 if (unlikely(!zonelist->_zonerefs->zone))
3235 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3236 alloc_flags |= ALLOC_CMA;
3239 cpuset_mems_cookie = read_mems_allowed_begin();
3241 /* We set it here, as __alloc_pages_slowpath might have changed it */
3242 ac.zonelist = zonelist;
3244 /* Dirty zone balancing only done in the fast path */
3245 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3247 /* The preferred zone is used for statistics later */
3248 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3249 ac.nodemask ? : &cpuset_current_mems_allowed,
3250 &ac.preferred_zone);
3251 if (!ac.preferred_zone)
3253 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3255 /* First allocation attempt */
3256 alloc_mask = gfp_mask|__GFP_HARDWALL;
3257 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3258 if (unlikely(!page)) {
3260 * Runtime PM, block IO and its error handling path
3261 * can deadlock because I/O on the device might not
3264 alloc_mask = memalloc_noio_flags(gfp_mask);
3265 ac.spread_dirty_pages = false;
3267 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3270 if (kmemcheck_enabled && page)
3271 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3273 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3277 * When updating a task's mems_allowed, it is possible to race with
3278 * parallel threads in such a way that an allocation can fail while
3279 * the mask is being updated. If a page allocation is about to fail,
3280 * check if the cpuset changed during allocation and if so, retry.
3282 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3287 EXPORT_SYMBOL(__alloc_pages_nodemask);
3290 * Common helper functions.
3292 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3297 * __get_free_pages() returns a 32-bit address, which cannot represent
3300 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3302 page = alloc_pages(gfp_mask, order);
3305 return (unsigned long) page_address(page);
3307 EXPORT_SYMBOL(__get_free_pages);
3309 unsigned long get_zeroed_page(gfp_t gfp_mask)
3311 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3313 EXPORT_SYMBOL(get_zeroed_page);
3315 void __free_pages(struct page *page, unsigned int order)
3317 if (put_page_testzero(page)) {
3319 free_hot_cold_page(page, false);
3321 __free_pages_ok(page, order);
3325 EXPORT_SYMBOL(__free_pages);
3327 void free_pages(unsigned long addr, unsigned int order)
3330 VM_BUG_ON(!virt_addr_valid((void *)addr));
3331 __free_pages(virt_to_page((void *)addr), order);
3335 EXPORT_SYMBOL(free_pages);
3339 * An arbitrary-length arbitrary-offset area of memory which resides
3340 * within a 0 or higher order page. Multiple fragments within that page
3341 * are individually refcounted, in the page's reference counter.
3343 * The page_frag functions below provide a simple allocation framework for
3344 * page fragments. This is used by the network stack and network device
3345 * drivers to provide a backing region of memory for use as either an
3346 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3348 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3351 struct page *page = NULL;
3352 gfp_t gfp = gfp_mask;
3354 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3355 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3357 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3358 PAGE_FRAG_CACHE_MAX_ORDER);
3359 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3361 if (unlikely(!page))
3362 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3364 nc->va = page ? page_address(page) : NULL;
3369 void *__alloc_page_frag(struct page_frag_cache *nc,
3370 unsigned int fragsz, gfp_t gfp_mask)
3372 unsigned int size = PAGE_SIZE;
3376 if (unlikely(!nc->va)) {
3378 page = __page_frag_refill(nc, gfp_mask);
3382 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3383 /* if size can vary use size else just use PAGE_SIZE */
3386 /* Even if we own the page, we do not use atomic_set().
3387 * This would break get_page_unless_zero() users.
3389 atomic_add(size - 1, &page->_count);
3391 /* reset page count bias and offset to start of new frag */
3392 nc->pfmemalloc = page_is_pfmemalloc(page);
3393 nc->pagecnt_bias = size;
3397 offset = nc->offset - fragsz;
3398 if (unlikely(offset < 0)) {
3399 page = virt_to_page(nc->va);
3401 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3404 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3405 /* if size can vary use size else just use PAGE_SIZE */
3408 /* OK, page count is 0, we can safely set it */
3409 atomic_set(&page->_count, size);
3411 /* reset page count bias and offset to start of new frag */
3412 nc->pagecnt_bias = size;
3413 offset = size - fragsz;
3417 nc->offset = offset;
3419 return nc->va + offset;
3421 EXPORT_SYMBOL(__alloc_page_frag);
3424 * Frees a page fragment allocated out of either a compound or order 0 page.
3426 void __free_page_frag(void *addr)
3428 struct page *page = virt_to_head_page(addr);
3430 if (unlikely(put_page_testzero(page)))
3431 __free_pages_ok(page, compound_order(page));
3433 EXPORT_SYMBOL(__free_page_frag);
3436 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3437 * of the current memory cgroup.
3439 * It should be used when the caller would like to use kmalloc, but since the
3440 * allocation is large, it has to fall back to the page allocator.
3442 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3446 page = alloc_pages(gfp_mask, order);
3447 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3448 __free_pages(page, order);
3454 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3458 page = alloc_pages_node(nid, gfp_mask, order);
3459 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3460 __free_pages(page, order);
3467 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3470 void __free_kmem_pages(struct page *page, unsigned int order)
3472 memcg_kmem_uncharge(page, order);
3473 __free_pages(page, order);
3476 void free_kmem_pages(unsigned long addr, unsigned int order)
3479 VM_BUG_ON(!virt_addr_valid((void *)addr));
3480 __free_kmem_pages(virt_to_page((void *)addr), order);
3484 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3488 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3489 unsigned long used = addr + PAGE_ALIGN(size);
3491 split_page(virt_to_page((void *)addr), order);
3492 while (used < alloc_end) {
3497 return (void *)addr;
3501 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3502 * @size: the number of bytes to allocate
3503 * @gfp_mask: GFP flags for the allocation
3505 * This function is similar to alloc_pages(), except that it allocates the
3506 * minimum number of pages to satisfy the request. alloc_pages() can only
3507 * allocate memory in power-of-two pages.
3509 * This function is also limited by MAX_ORDER.
3511 * Memory allocated by this function must be released by free_pages_exact().
3513 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3515 unsigned int order = get_order(size);
3518 addr = __get_free_pages(gfp_mask, order);
3519 return make_alloc_exact(addr, order, size);
3521 EXPORT_SYMBOL(alloc_pages_exact);
3524 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3526 * @nid: the preferred node ID where memory should be allocated
3527 * @size: the number of bytes to allocate
3528 * @gfp_mask: GFP flags for the allocation
3530 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3533 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3535 unsigned int order = get_order(size);
3536 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3539 return make_alloc_exact((unsigned long)page_address(p), order, size);
3543 * free_pages_exact - release memory allocated via alloc_pages_exact()
3544 * @virt: the value returned by alloc_pages_exact.
3545 * @size: size of allocation, same value as passed to alloc_pages_exact().
3547 * Release the memory allocated by a previous call to alloc_pages_exact.
3549 void free_pages_exact(void *virt, size_t size)
3551 unsigned long addr = (unsigned long)virt;
3552 unsigned long end = addr + PAGE_ALIGN(size);
3554 while (addr < end) {
3559 EXPORT_SYMBOL(free_pages_exact);
3562 * nr_free_zone_pages - count number of pages beyond high watermark
3563 * @offset: The zone index of the highest zone
3565 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3566 * high watermark within all zones at or below a given zone index. For each
3567 * zone, the number of pages is calculated as:
3568 * managed_pages - high_pages
3570 static unsigned long nr_free_zone_pages(int offset)
3575 /* Just pick one node, since fallback list is circular */
3576 unsigned long sum = 0;
3578 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3580 for_each_zone_zonelist(zone, z, zonelist, offset) {
3581 unsigned long size = zone->managed_pages;
3582 unsigned long high = high_wmark_pages(zone);
3591 * nr_free_buffer_pages - count number of pages beyond high watermark
3593 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3594 * watermark within ZONE_DMA and ZONE_NORMAL.
3596 unsigned long nr_free_buffer_pages(void)
3598 return nr_free_zone_pages(gfp_zone(GFP_USER));
3600 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3603 * nr_free_pagecache_pages - count number of pages beyond high watermark
3605 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3606 * high watermark within all zones.
3608 unsigned long nr_free_pagecache_pages(void)
3610 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3613 static inline void show_node(struct zone *zone)
3615 if (IS_ENABLED(CONFIG_NUMA))
3616 printk("Node %d ", zone_to_nid(zone));
3619 void si_meminfo(struct sysinfo *val)
3621 val->totalram = totalram_pages;
3622 val->sharedram = global_page_state(NR_SHMEM);
3623 val->freeram = global_page_state(NR_FREE_PAGES);
3624 val->bufferram = nr_blockdev_pages();
3625 val->totalhigh = totalhigh_pages;
3626 val->freehigh = nr_free_highpages();
3627 val->mem_unit = PAGE_SIZE;
3630 EXPORT_SYMBOL(si_meminfo);
3633 void si_meminfo_node(struct sysinfo *val, int nid)
3635 int zone_type; /* needs to be signed */
3636 unsigned long managed_pages = 0;
3637 pg_data_t *pgdat = NODE_DATA(nid);
3639 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3640 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3641 val->totalram = managed_pages;
3642 val->sharedram = node_page_state(nid, NR_SHMEM);
3643 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3644 #ifdef CONFIG_HIGHMEM
3645 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3646 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3652 val->mem_unit = PAGE_SIZE;
3657 * Determine whether the node should be displayed or not, depending on whether
3658 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3660 bool skip_free_areas_node(unsigned int flags, int nid)
3663 unsigned int cpuset_mems_cookie;
3665 if (!(flags & SHOW_MEM_FILTER_NODES))
3669 cpuset_mems_cookie = read_mems_allowed_begin();
3670 ret = !node_isset(nid, cpuset_current_mems_allowed);
3671 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3676 #define K(x) ((x) << (PAGE_SHIFT-10))
3678 static void show_migration_types(unsigned char type)
3680 static const char types[MIGRATE_TYPES] = {
3681 [MIGRATE_UNMOVABLE] = 'U',
3682 [MIGRATE_MOVABLE] = 'M',
3683 [MIGRATE_RECLAIMABLE] = 'E',
3684 [MIGRATE_HIGHATOMIC] = 'H',
3686 [MIGRATE_CMA] = 'C',
3688 #ifdef CONFIG_MEMORY_ISOLATION
3689 [MIGRATE_ISOLATE] = 'I',
3692 char tmp[MIGRATE_TYPES + 1];
3696 for (i = 0; i < MIGRATE_TYPES; i++) {
3697 if (type & (1 << i))
3702 printk("(%s) ", tmp);
3706 * Show free area list (used inside shift_scroll-lock stuff)
3707 * We also calculate the percentage fragmentation. We do this by counting the
3708 * memory on each free list with the exception of the first item on the list.
3711 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3714 void show_free_areas(unsigned int filter)
3716 unsigned long free_pcp = 0;
3720 for_each_populated_zone(zone) {
3721 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3724 for_each_online_cpu(cpu)
3725 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3728 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3729 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3730 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3731 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3732 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3733 " free:%lu free_pcp:%lu free_cma:%lu\n",
3734 global_page_state(NR_ACTIVE_ANON),
3735 global_page_state(NR_INACTIVE_ANON),
3736 global_page_state(NR_ISOLATED_ANON),
3737 global_page_state(NR_ACTIVE_FILE),
3738 global_page_state(NR_INACTIVE_FILE),
3739 global_page_state(NR_ISOLATED_FILE),
3740 global_page_state(NR_UNEVICTABLE),
3741 global_page_state(NR_FILE_DIRTY),
3742 global_page_state(NR_WRITEBACK),
3743 global_page_state(NR_UNSTABLE_NFS),
3744 global_page_state(NR_SLAB_RECLAIMABLE),
3745 global_page_state(NR_SLAB_UNRECLAIMABLE),
3746 global_page_state(NR_FILE_MAPPED),
3747 global_page_state(NR_SHMEM),
3748 global_page_state(NR_PAGETABLE),
3749 global_page_state(NR_BOUNCE),
3750 global_page_state(NR_FREE_PAGES),
3752 global_page_state(NR_FREE_CMA_PAGES));
3754 for_each_populated_zone(zone) {
3757 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3761 for_each_online_cpu(cpu)
3762 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3770 " active_anon:%lukB"
3771 " inactive_anon:%lukB"
3772 " active_file:%lukB"
3773 " inactive_file:%lukB"
3774 " unevictable:%lukB"
3775 " isolated(anon):%lukB"
3776 " isolated(file):%lukB"
3784 " slab_reclaimable:%lukB"
3785 " slab_unreclaimable:%lukB"
3786 " kernel_stack:%lukB"
3793 " writeback_tmp:%lukB"
3794 " pages_scanned:%lu"
3795 " all_unreclaimable? %s"
3798 K(zone_page_state(zone, NR_FREE_PAGES)),
3799 K(min_wmark_pages(zone)),
3800 K(low_wmark_pages(zone)),
3801 K(high_wmark_pages(zone)),
3802 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3803 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3804 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3805 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3806 K(zone_page_state(zone, NR_UNEVICTABLE)),
3807 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3808 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3809 K(zone->present_pages),
3810 K(zone->managed_pages),
3811 K(zone_page_state(zone, NR_MLOCK)),
3812 K(zone_page_state(zone, NR_FILE_DIRTY)),
3813 K(zone_page_state(zone, NR_WRITEBACK)),
3814 K(zone_page_state(zone, NR_FILE_MAPPED)),
3815 K(zone_page_state(zone, NR_SHMEM)),
3816 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3817 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3818 zone_page_state(zone, NR_KERNEL_STACK) *
3820 K(zone_page_state(zone, NR_PAGETABLE)),
3821 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3822 K(zone_page_state(zone, NR_BOUNCE)),
3824 K(this_cpu_read(zone->pageset->pcp.count)),
3825 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3826 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3827 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3828 (!zone_reclaimable(zone) ? "yes" : "no")
3830 printk("lowmem_reserve[]:");
3831 for (i = 0; i < MAX_NR_ZONES; i++)
3832 printk(" %ld", zone->lowmem_reserve[i]);
3836 for_each_populated_zone(zone) {
3838 unsigned long nr[MAX_ORDER], flags, total = 0;
3839 unsigned char types[MAX_ORDER];
3841 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3844 printk("%s: ", zone->name);
3846 spin_lock_irqsave(&zone->lock, flags);
3847 for (order = 0; order < MAX_ORDER; order++) {
3848 struct free_area *area = &zone->free_area[order];
3851 nr[order] = area->nr_free;
3852 total += nr[order] << order;
3855 for (type = 0; type < MIGRATE_TYPES; type++) {
3856 if (!list_empty(&area->free_list[type]))
3857 types[order] |= 1 << type;
3860 spin_unlock_irqrestore(&zone->lock, flags);
3861 for (order = 0; order < MAX_ORDER; order++) {
3862 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3864 show_migration_types(types[order]);
3866 printk("= %lukB\n", K(total));
3869 hugetlb_show_meminfo();
3871 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3873 show_swap_cache_info();
3876 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3878 zoneref->zone = zone;
3879 zoneref->zone_idx = zone_idx(zone);
3883 * Builds allocation fallback zone lists.
3885 * Add all populated zones of a node to the zonelist.
3887 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3891 enum zone_type zone_type = MAX_NR_ZONES;
3895 zone = pgdat->node_zones + zone_type;
3896 if (populated_zone(zone)) {
3897 zoneref_set_zone(zone,
3898 &zonelist->_zonerefs[nr_zones++]);
3899 check_highest_zone(zone_type);
3901 } while (zone_type);
3909 * 0 = automatic detection of better ordering.
3910 * 1 = order by ([node] distance, -zonetype)
3911 * 2 = order by (-zonetype, [node] distance)
3913 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3914 * the same zonelist. So only NUMA can configure this param.
3916 #define ZONELIST_ORDER_DEFAULT 0
3917 #define ZONELIST_ORDER_NODE 1
3918 #define ZONELIST_ORDER_ZONE 2
3920 /* zonelist order in the kernel.
3921 * set_zonelist_order() will set this to NODE or ZONE.
3923 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3924 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3928 /* The value user specified ....changed by config */
3929 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3930 /* string for sysctl */
3931 #define NUMA_ZONELIST_ORDER_LEN 16
3932 char numa_zonelist_order[16] = "default";
3935 * interface for configure zonelist ordering.
3936 * command line option "numa_zonelist_order"
3937 * = "[dD]efault - default, automatic configuration.
3938 * = "[nN]ode - order by node locality, then by zone within node
3939 * = "[zZ]one - order by zone, then by locality within zone
3942 static int __parse_numa_zonelist_order(char *s)
3944 if (*s == 'd' || *s == 'D') {
3945 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3946 } else if (*s == 'n' || *s == 'N') {
3947 user_zonelist_order = ZONELIST_ORDER_NODE;
3948 } else if (*s == 'z' || *s == 'Z') {
3949 user_zonelist_order = ZONELIST_ORDER_ZONE;
3952 "Ignoring invalid numa_zonelist_order value: "
3959 static __init int setup_numa_zonelist_order(char *s)
3966 ret = __parse_numa_zonelist_order(s);
3968 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3972 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3975 * sysctl handler for numa_zonelist_order
3977 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3978 void __user *buffer, size_t *length,
3981 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3983 static DEFINE_MUTEX(zl_order_mutex);
3985 mutex_lock(&zl_order_mutex);
3987 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3991 strcpy(saved_string, (char *)table->data);
3993 ret = proc_dostring(table, write, buffer, length, ppos);
3997 int oldval = user_zonelist_order;
3999 ret = __parse_numa_zonelist_order((char *)table->data);
4002 * bogus value. restore saved string
4004 strncpy((char *)table->data, saved_string,
4005 NUMA_ZONELIST_ORDER_LEN);
4006 user_zonelist_order = oldval;
4007 } else if (oldval != user_zonelist_order) {
4008 mutex_lock(&zonelists_mutex);
4009 build_all_zonelists(NULL, NULL);
4010 mutex_unlock(&zonelists_mutex);
4014 mutex_unlock(&zl_order_mutex);
4019 #define MAX_NODE_LOAD (nr_online_nodes)
4020 static int node_load[MAX_NUMNODES];
4023 * find_next_best_node - find the next node that should appear in a given node's fallback list
4024 * @node: node whose fallback list we're appending
4025 * @used_node_mask: nodemask_t of already used nodes
4027 * We use a number of factors to determine which is the next node that should
4028 * appear on a given node's fallback list. The node should not have appeared
4029 * already in @node's fallback list, and it should be the next closest node
4030 * according to the distance array (which contains arbitrary distance values
4031 * from each node to each node in the system), and should also prefer nodes
4032 * with no CPUs, since presumably they'll have very little allocation pressure
4033 * on them otherwise.
4034 * It returns -1 if no node is found.
4036 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4039 int min_val = INT_MAX;
4040 int best_node = NUMA_NO_NODE;
4041 const struct cpumask *tmp = cpumask_of_node(0);
4043 /* Use the local node if we haven't already */
4044 if (!node_isset(node, *used_node_mask)) {
4045 node_set(node, *used_node_mask);
4049 for_each_node_state(n, N_MEMORY) {
4051 /* Don't want a node to appear more than once */
4052 if (node_isset(n, *used_node_mask))
4055 /* Use the distance array to find the distance */
4056 val = node_distance(node, n);
4058 /* Penalize nodes under us ("prefer the next node") */
4061 /* Give preference to headless and unused nodes */
4062 tmp = cpumask_of_node(n);
4063 if (!cpumask_empty(tmp))
4064 val += PENALTY_FOR_NODE_WITH_CPUS;
4066 /* Slight preference for less loaded node */
4067 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4068 val += node_load[n];
4070 if (val < min_val) {
4077 node_set(best_node, *used_node_mask);
4084 * Build zonelists ordered by node and zones within node.
4085 * This results in maximum locality--normal zone overflows into local
4086 * DMA zone, if any--but risks exhausting DMA zone.
4088 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4091 struct zonelist *zonelist;
4093 zonelist = &pgdat->node_zonelists[0];
4094 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4096 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4097 zonelist->_zonerefs[j].zone = NULL;
4098 zonelist->_zonerefs[j].zone_idx = 0;
4102 * Build gfp_thisnode zonelists
4104 static void build_thisnode_zonelists(pg_data_t *pgdat)
4107 struct zonelist *zonelist;
4109 zonelist = &pgdat->node_zonelists[1];
4110 j = build_zonelists_node(pgdat, zonelist, 0);
4111 zonelist->_zonerefs[j].zone = NULL;
4112 zonelist->_zonerefs[j].zone_idx = 0;
4116 * Build zonelists ordered by zone and nodes within zones.
4117 * This results in conserving DMA zone[s] until all Normal memory is
4118 * exhausted, but results in overflowing to remote node while memory
4119 * may still exist in local DMA zone.
4121 static int node_order[MAX_NUMNODES];
4123 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4126 int zone_type; /* needs to be signed */
4128 struct zonelist *zonelist;
4130 zonelist = &pgdat->node_zonelists[0];
4132 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4133 for (j = 0; j < nr_nodes; j++) {
4134 node = node_order[j];
4135 z = &NODE_DATA(node)->node_zones[zone_type];
4136 if (populated_zone(z)) {
4138 &zonelist->_zonerefs[pos++]);
4139 check_highest_zone(zone_type);
4143 zonelist->_zonerefs[pos].zone = NULL;
4144 zonelist->_zonerefs[pos].zone_idx = 0;
4147 #if defined(CONFIG_64BIT)
4149 * Devices that require DMA32/DMA are relatively rare and do not justify a
4150 * penalty to every machine in case the specialised case applies. Default
4151 * to Node-ordering on 64-bit NUMA machines
4153 static int default_zonelist_order(void)
4155 return ZONELIST_ORDER_NODE;
4159 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4160 * by the kernel. If processes running on node 0 deplete the low memory zone
4161 * then reclaim will occur more frequency increasing stalls and potentially
4162 * be easier to OOM if a large percentage of the zone is under writeback or
4163 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4164 * Hence, default to zone ordering on 32-bit.
4166 static int default_zonelist_order(void)
4168 return ZONELIST_ORDER_ZONE;
4170 #endif /* CONFIG_64BIT */
4172 static void set_zonelist_order(void)
4174 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4175 current_zonelist_order = default_zonelist_order();
4177 current_zonelist_order = user_zonelist_order;
4180 static void build_zonelists(pg_data_t *pgdat)
4184 nodemask_t used_mask;
4185 int local_node, prev_node;
4186 struct zonelist *zonelist;
4187 unsigned int order = current_zonelist_order;
4189 /* initialize zonelists */
4190 for (i = 0; i < MAX_ZONELISTS; i++) {
4191 zonelist = pgdat->node_zonelists + i;
4192 zonelist->_zonerefs[0].zone = NULL;
4193 zonelist->_zonerefs[0].zone_idx = 0;
4196 /* NUMA-aware ordering of nodes */
4197 local_node = pgdat->node_id;
4198 load = nr_online_nodes;
4199 prev_node = local_node;
4200 nodes_clear(used_mask);
4202 memset(node_order, 0, sizeof(node_order));
4205 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4207 * We don't want to pressure a particular node.
4208 * So adding penalty to the first node in same
4209 * distance group to make it round-robin.
4211 if (node_distance(local_node, node) !=
4212 node_distance(local_node, prev_node))
4213 node_load[node] = load;
4217 if (order == ZONELIST_ORDER_NODE)
4218 build_zonelists_in_node_order(pgdat, node);
4220 node_order[j++] = node; /* remember order */
4223 if (order == ZONELIST_ORDER_ZONE) {
4224 /* calculate node order -- i.e., DMA last! */
4225 build_zonelists_in_zone_order(pgdat, j);
4228 build_thisnode_zonelists(pgdat);
4231 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4233 * Return node id of node used for "local" allocations.
4234 * I.e., first node id of first zone in arg node's generic zonelist.
4235 * Used for initializing percpu 'numa_mem', which is used primarily
4236 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4238 int local_memory_node(int node)
4242 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4243 gfp_zone(GFP_KERNEL),
4250 #else /* CONFIG_NUMA */
4252 static void set_zonelist_order(void)
4254 current_zonelist_order = ZONELIST_ORDER_ZONE;
4257 static void build_zonelists(pg_data_t *pgdat)
4259 int node, local_node;
4261 struct zonelist *zonelist;
4263 local_node = pgdat->node_id;
4265 zonelist = &pgdat->node_zonelists[0];
4266 j = build_zonelists_node(pgdat, zonelist, 0);
4269 * Now we build the zonelist so that it contains the zones
4270 * of all the other nodes.
4271 * We don't want to pressure a particular node, so when
4272 * building the zones for node N, we make sure that the
4273 * zones coming right after the local ones are those from
4274 * node N+1 (modulo N)
4276 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4277 if (!node_online(node))
4279 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4281 for (node = 0; node < local_node; node++) {
4282 if (!node_online(node))
4284 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4287 zonelist->_zonerefs[j].zone = NULL;
4288 zonelist->_zonerefs[j].zone_idx = 0;
4291 #endif /* CONFIG_NUMA */
4294 * Boot pageset table. One per cpu which is going to be used for all
4295 * zones and all nodes. The parameters will be set in such a way
4296 * that an item put on a list will immediately be handed over to
4297 * the buddy list. This is safe since pageset manipulation is done
4298 * with interrupts disabled.
4300 * The boot_pagesets must be kept even after bootup is complete for
4301 * unused processors and/or zones. They do play a role for bootstrapping
4302 * hotplugged processors.
4304 * zoneinfo_show() and maybe other functions do
4305 * not check if the processor is online before following the pageset pointer.
4306 * Other parts of the kernel may not check if the zone is available.
4308 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4309 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4310 static void setup_zone_pageset(struct zone *zone);
4313 * Global mutex to protect against size modification of zonelists
4314 * as well as to serialize pageset setup for the new populated zone.
4316 DEFINE_MUTEX(zonelists_mutex);
4318 /* return values int ....just for stop_machine() */
4319 static int __build_all_zonelists(void *data)
4323 pg_data_t *self = data;
4326 memset(node_load, 0, sizeof(node_load));
4329 if (self && !node_online(self->node_id)) {
4330 build_zonelists(self);
4333 for_each_online_node(nid) {
4334 pg_data_t *pgdat = NODE_DATA(nid);
4336 build_zonelists(pgdat);
4340 * Initialize the boot_pagesets that are going to be used
4341 * for bootstrapping processors. The real pagesets for
4342 * each zone will be allocated later when the per cpu
4343 * allocator is available.
4345 * boot_pagesets are used also for bootstrapping offline
4346 * cpus if the system is already booted because the pagesets
4347 * are needed to initialize allocators on a specific cpu too.
4348 * F.e. the percpu allocator needs the page allocator which
4349 * needs the percpu allocator in order to allocate its pagesets
4350 * (a chicken-egg dilemma).
4352 for_each_possible_cpu(cpu) {
4353 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4355 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4357 * We now know the "local memory node" for each node--
4358 * i.e., the node of the first zone in the generic zonelist.
4359 * Set up numa_mem percpu variable for on-line cpus. During
4360 * boot, only the boot cpu should be on-line; we'll init the
4361 * secondary cpus' numa_mem as they come on-line. During
4362 * node/memory hotplug, we'll fixup all on-line cpus.
4364 if (cpu_online(cpu))
4365 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4372 static noinline void __init
4373 build_all_zonelists_init(void)
4375 __build_all_zonelists(NULL);
4376 mminit_verify_zonelist();
4377 cpuset_init_current_mems_allowed();
4381 * Called with zonelists_mutex held always
4382 * unless system_state == SYSTEM_BOOTING.
4384 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4385 * [we're only called with non-NULL zone through __meminit paths] and
4386 * (2) call of __init annotated helper build_all_zonelists_init
4387 * [protected by SYSTEM_BOOTING].
4389 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4391 set_zonelist_order();
4393 if (system_state == SYSTEM_BOOTING) {
4394 build_all_zonelists_init();
4396 #ifdef CONFIG_MEMORY_HOTPLUG
4398 setup_zone_pageset(zone);
4400 /* we have to stop all cpus to guarantee there is no user
4402 stop_machine(__build_all_zonelists, pgdat, NULL);
4403 /* cpuset refresh routine should be here */
4405 vm_total_pages = nr_free_pagecache_pages();
4407 * Disable grouping by mobility if the number of pages in the
4408 * system is too low to allow the mechanism to work. It would be
4409 * more accurate, but expensive to check per-zone. This check is
4410 * made on memory-hotadd so a system can start with mobility
4411 * disabled and enable it later
4413 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4414 page_group_by_mobility_disabled = 1;
4416 page_group_by_mobility_disabled = 0;
4418 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4419 "Total pages: %ld\n",
4421 zonelist_order_name[current_zonelist_order],
4422 page_group_by_mobility_disabled ? "off" : "on",
4425 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4430 * Helper functions to size the waitqueue hash table.
4431 * Essentially these want to choose hash table sizes sufficiently
4432 * large so that collisions trying to wait on pages are rare.
4433 * But in fact, the number of active page waitqueues on typical
4434 * systems is ridiculously low, less than 200. So this is even
4435 * conservative, even though it seems large.
4437 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4438 * waitqueues, i.e. the size of the waitq table given the number of pages.
4440 #define PAGES_PER_WAITQUEUE 256
4442 #ifndef CONFIG_MEMORY_HOTPLUG
4443 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4445 unsigned long size = 1;
4447 pages /= PAGES_PER_WAITQUEUE;
4449 while (size < pages)
4453 * Once we have dozens or even hundreds of threads sleeping
4454 * on IO we've got bigger problems than wait queue collision.
4455 * Limit the size of the wait table to a reasonable size.
4457 size = min(size, 4096UL);
4459 return max(size, 4UL);
4463 * A zone's size might be changed by hot-add, so it is not possible to determine
4464 * a suitable size for its wait_table. So we use the maximum size now.
4466 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4468 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4469 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4470 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4472 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4473 * or more by the traditional way. (See above). It equals:
4475 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4476 * ia64(16K page size) : = ( 8G + 4M)byte.
4477 * powerpc (64K page size) : = (32G +16M)byte.
4479 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4486 * This is an integer logarithm so that shifts can be used later
4487 * to extract the more random high bits from the multiplicative
4488 * hash function before the remainder is taken.
4490 static inline unsigned long wait_table_bits(unsigned long size)
4496 * Initially all pages are reserved - free ones are freed
4497 * up by free_all_bootmem() once the early boot process is
4498 * done. Non-atomic initialization, single-pass.
4500 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4501 unsigned long start_pfn, enum memmap_context context)
4503 pg_data_t *pgdat = NODE_DATA(nid);
4504 unsigned long end_pfn = start_pfn + size;
4507 unsigned long nr_initialised = 0;
4509 if (highest_memmap_pfn < end_pfn - 1)
4510 highest_memmap_pfn = end_pfn - 1;
4512 z = &pgdat->node_zones[zone];
4513 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4515 * There can be holes in boot-time mem_map[]s
4516 * handed to this function. They do not
4517 * exist on hotplugged memory.
4519 if (context == MEMMAP_EARLY) {
4520 if (!early_pfn_valid(pfn))
4522 if (!early_pfn_in_nid(pfn, nid))
4524 if (!update_defer_init(pgdat, pfn, end_pfn,
4530 * Mark the block movable so that blocks are reserved for
4531 * movable at startup. This will force kernel allocations
4532 * to reserve their blocks rather than leaking throughout
4533 * the address space during boot when many long-lived
4534 * kernel allocations are made.
4536 * bitmap is created for zone's valid pfn range. but memmap
4537 * can be created for invalid pages (for alignment)
4538 * check here not to call set_pageblock_migratetype() against
4541 if (!(pfn & (pageblock_nr_pages - 1))) {
4542 struct page *page = pfn_to_page(pfn);
4544 __init_single_page(page, pfn, zone, nid);
4545 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4547 __init_single_pfn(pfn, zone, nid);
4552 static void __meminit zone_init_free_lists(struct zone *zone)
4554 unsigned int order, t;
4555 for_each_migratetype_order(order, t) {
4556 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4557 zone->free_area[order].nr_free = 0;
4561 #ifndef __HAVE_ARCH_MEMMAP_INIT
4562 #define memmap_init(size, nid, zone, start_pfn) \
4563 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4566 static int zone_batchsize(struct zone *zone)
4572 * The per-cpu-pages pools are set to around 1000th of the
4573 * size of the zone. But no more than 1/2 of a meg.
4575 * OK, so we don't know how big the cache is. So guess.
4577 batch = zone->managed_pages / 1024;
4578 if (batch * PAGE_SIZE > 512 * 1024)
4579 batch = (512 * 1024) / PAGE_SIZE;
4580 batch /= 4; /* We effectively *= 4 below */
4585 * Clamp the batch to a 2^n - 1 value. Having a power
4586 * of 2 value was found to be more likely to have
4587 * suboptimal cache aliasing properties in some cases.
4589 * For example if 2 tasks are alternately allocating
4590 * batches of pages, one task can end up with a lot
4591 * of pages of one half of the possible page colors
4592 * and the other with pages of the other colors.
4594 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4599 /* The deferral and batching of frees should be suppressed under NOMMU
4602 * The problem is that NOMMU needs to be able to allocate large chunks
4603 * of contiguous memory as there's no hardware page translation to
4604 * assemble apparent contiguous memory from discontiguous pages.
4606 * Queueing large contiguous runs of pages for batching, however,
4607 * causes the pages to actually be freed in smaller chunks. As there
4608 * can be a significant delay between the individual batches being
4609 * recycled, this leads to the once large chunks of space being
4610 * fragmented and becoming unavailable for high-order allocations.
4617 * pcp->high and pcp->batch values are related and dependent on one another:
4618 * ->batch must never be higher then ->high.
4619 * The following function updates them in a safe manner without read side
4622 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4623 * those fields changing asynchronously (acording the the above rule).
4625 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4626 * outside of boot time (or some other assurance that no concurrent updaters
4629 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4630 unsigned long batch)
4632 /* start with a fail safe value for batch */
4636 /* Update high, then batch, in order */
4643 /* a companion to pageset_set_high() */
4644 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4646 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4649 static void pageset_init(struct per_cpu_pageset *p)
4651 struct per_cpu_pages *pcp;
4654 memset(p, 0, sizeof(*p));
4658 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4659 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4662 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4665 pageset_set_batch(p, batch);
4669 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4670 * to the value high for the pageset p.
4672 static void pageset_set_high(struct per_cpu_pageset *p,
4675 unsigned long batch = max(1UL, high / 4);
4676 if ((high / 4) > (PAGE_SHIFT * 8))
4677 batch = PAGE_SHIFT * 8;
4679 pageset_update(&p->pcp, high, batch);
4682 static void pageset_set_high_and_batch(struct zone *zone,
4683 struct per_cpu_pageset *pcp)
4685 if (percpu_pagelist_fraction)
4686 pageset_set_high(pcp,
4687 (zone->managed_pages /
4688 percpu_pagelist_fraction));
4690 pageset_set_batch(pcp, zone_batchsize(zone));
4693 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4695 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4698 pageset_set_high_and_batch(zone, pcp);
4701 static void __meminit setup_zone_pageset(struct zone *zone)
4704 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4705 for_each_possible_cpu(cpu)
4706 zone_pageset_init(zone, cpu);
4710 * Allocate per cpu pagesets and initialize them.
4711 * Before this call only boot pagesets were available.
4713 void __init setup_per_cpu_pageset(void)
4717 for_each_populated_zone(zone)
4718 setup_zone_pageset(zone);
4721 static noinline __init_refok
4722 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4728 * The per-page waitqueue mechanism uses hashed waitqueues
4731 zone->wait_table_hash_nr_entries =
4732 wait_table_hash_nr_entries(zone_size_pages);
4733 zone->wait_table_bits =
4734 wait_table_bits(zone->wait_table_hash_nr_entries);
4735 alloc_size = zone->wait_table_hash_nr_entries
4736 * sizeof(wait_queue_head_t);
4738 if (!slab_is_available()) {
4739 zone->wait_table = (wait_queue_head_t *)
4740 memblock_virt_alloc_node_nopanic(
4741 alloc_size, zone->zone_pgdat->node_id);
4744 * This case means that a zone whose size was 0 gets new memory
4745 * via memory hot-add.
4746 * But it may be the case that a new node was hot-added. In
4747 * this case vmalloc() will not be able to use this new node's
4748 * memory - this wait_table must be initialized to use this new
4749 * node itself as well.
4750 * To use this new node's memory, further consideration will be
4753 zone->wait_table = vmalloc(alloc_size);
4755 if (!zone->wait_table)
4758 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4759 init_waitqueue_head(zone->wait_table + i);
4764 static __meminit void zone_pcp_init(struct zone *zone)
4767 * per cpu subsystem is not up at this point. The following code
4768 * relies on the ability of the linker to provide the
4769 * offset of a (static) per cpu variable into the per cpu area.
4771 zone->pageset = &boot_pageset;
4773 if (populated_zone(zone))
4774 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4775 zone->name, zone->present_pages,
4776 zone_batchsize(zone));
4779 int __meminit init_currently_empty_zone(struct zone *zone,
4780 unsigned long zone_start_pfn,
4783 struct pglist_data *pgdat = zone->zone_pgdat;
4785 ret = zone_wait_table_init(zone, size);
4788 pgdat->nr_zones = zone_idx(zone) + 1;
4790 zone->zone_start_pfn = zone_start_pfn;
4792 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4793 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4795 (unsigned long)zone_idx(zone),
4796 zone_start_pfn, (zone_start_pfn + size));
4798 zone_init_free_lists(zone);
4803 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4804 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4807 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4809 int __meminit __early_pfn_to_nid(unsigned long pfn,
4810 struct mminit_pfnnid_cache *state)
4812 unsigned long start_pfn, end_pfn;
4815 if (state->last_start <= pfn && pfn < state->last_end)
4816 return state->last_nid;
4818 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4820 state->last_start = start_pfn;
4821 state->last_end = end_pfn;
4822 state->last_nid = nid;
4827 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4830 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4831 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4832 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4834 * If an architecture guarantees that all ranges registered contain no holes
4835 * and may be freed, this this function may be used instead of calling
4836 * memblock_free_early_nid() manually.
4838 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4840 unsigned long start_pfn, end_pfn;
4843 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4844 start_pfn = min(start_pfn, max_low_pfn);
4845 end_pfn = min(end_pfn, max_low_pfn);
4847 if (start_pfn < end_pfn)
4848 memblock_free_early_nid(PFN_PHYS(start_pfn),
4849 (end_pfn - start_pfn) << PAGE_SHIFT,
4855 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4856 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4858 * If an architecture guarantees that all ranges registered contain no holes and may
4859 * be freed, this function may be used instead of calling memory_present() manually.
4861 void __init sparse_memory_present_with_active_regions(int nid)
4863 unsigned long start_pfn, end_pfn;
4866 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4867 memory_present(this_nid, start_pfn, end_pfn);
4871 * get_pfn_range_for_nid - Return the start and end page frames for a node
4872 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4873 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4874 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4876 * It returns the start and end page frame of a node based on information
4877 * provided by memblock_set_node(). If called for a node
4878 * with no available memory, a warning is printed and the start and end
4881 void __meminit get_pfn_range_for_nid(unsigned int nid,
4882 unsigned long *start_pfn, unsigned long *end_pfn)
4884 unsigned long this_start_pfn, this_end_pfn;
4890 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4891 *start_pfn = min(*start_pfn, this_start_pfn);
4892 *end_pfn = max(*end_pfn, this_end_pfn);
4895 if (*start_pfn == -1UL)
4900 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4901 * assumption is made that zones within a node are ordered in monotonic
4902 * increasing memory addresses so that the "highest" populated zone is used
4904 static void __init find_usable_zone_for_movable(void)
4907 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4908 if (zone_index == ZONE_MOVABLE)
4911 if (arch_zone_highest_possible_pfn[zone_index] >
4912 arch_zone_lowest_possible_pfn[zone_index])
4916 VM_BUG_ON(zone_index == -1);
4917 movable_zone = zone_index;
4921 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4922 * because it is sized independent of architecture. Unlike the other zones,
4923 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4924 * in each node depending on the size of each node and how evenly kernelcore
4925 * is distributed. This helper function adjusts the zone ranges
4926 * provided by the architecture for a given node by using the end of the
4927 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4928 * zones within a node are in order of monotonic increases memory addresses
4930 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4931 unsigned long zone_type,
4932 unsigned long node_start_pfn,
4933 unsigned long node_end_pfn,
4934 unsigned long *zone_start_pfn,
4935 unsigned long *zone_end_pfn)
4937 /* Only adjust if ZONE_MOVABLE is on this node */
4938 if (zone_movable_pfn[nid]) {
4939 /* Size ZONE_MOVABLE */
4940 if (zone_type == ZONE_MOVABLE) {
4941 *zone_start_pfn = zone_movable_pfn[nid];
4942 *zone_end_pfn = min(node_end_pfn,
4943 arch_zone_highest_possible_pfn[movable_zone]);
4945 /* Adjust for ZONE_MOVABLE starting within this range */
4946 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4947 *zone_end_pfn > zone_movable_pfn[nid]) {
4948 *zone_end_pfn = zone_movable_pfn[nid];
4950 /* Check if this whole range is within ZONE_MOVABLE */
4951 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4952 *zone_start_pfn = *zone_end_pfn;
4957 * Return the number of pages a zone spans in a node, including holes
4958 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4960 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4961 unsigned long zone_type,
4962 unsigned long node_start_pfn,
4963 unsigned long node_end_pfn,
4964 unsigned long *ignored)
4966 unsigned long zone_start_pfn, zone_end_pfn;
4968 /* When hotadd a new node from cpu_up(), the node should be empty */
4969 if (!node_start_pfn && !node_end_pfn)
4972 /* Get the start and end of the zone */
4973 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4974 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4975 adjust_zone_range_for_zone_movable(nid, zone_type,
4976 node_start_pfn, node_end_pfn,
4977 &zone_start_pfn, &zone_end_pfn);
4979 /* Check that this node has pages within the zone's required range */
4980 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4983 /* Move the zone boundaries inside the node if necessary */
4984 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4985 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4987 /* Return the spanned pages */
4988 return zone_end_pfn - zone_start_pfn;
4992 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4993 * then all holes in the requested range will be accounted for.
4995 unsigned long __meminit __absent_pages_in_range(int nid,
4996 unsigned long range_start_pfn,
4997 unsigned long range_end_pfn)
4999 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5000 unsigned long start_pfn, end_pfn;
5003 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5004 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5005 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5006 nr_absent -= end_pfn - start_pfn;
5012 * absent_pages_in_range - Return number of page frames in holes within a range
5013 * @start_pfn: The start PFN to start searching for holes
5014 * @end_pfn: The end PFN to stop searching for holes
5016 * It returns the number of pages frames in memory holes within a range.
5018 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5019 unsigned long end_pfn)
5021 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5024 /* Return the number of page frames in holes in a zone on a node */
5025 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5026 unsigned long zone_type,
5027 unsigned long node_start_pfn,
5028 unsigned long node_end_pfn,
5029 unsigned long *ignored)
5031 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5032 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5033 unsigned long zone_start_pfn, zone_end_pfn;
5035 /* When hotadd a new node from cpu_up(), the node should be empty */
5036 if (!node_start_pfn && !node_end_pfn)
5039 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5040 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5042 adjust_zone_range_for_zone_movable(nid, zone_type,
5043 node_start_pfn, node_end_pfn,
5044 &zone_start_pfn, &zone_end_pfn);
5045 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5048 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5049 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5050 unsigned long zone_type,
5051 unsigned long node_start_pfn,
5052 unsigned long node_end_pfn,
5053 unsigned long *zones_size)
5055 return zones_size[zone_type];
5058 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5059 unsigned long zone_type,
5060 unsigned long node_start_pfn,
5061 unsigned long node_end_pfn,
5062 unsigned long *zholes_size)
5067 return zholes_size[zone_type];
5070 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5072 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5073 unsigned long node_start_pfn,
5074 unsigned long node_end_pfn,
5075 unsigned long *zones_size,
5076 unsigned long *zholes_size)
5078 unsigned long realtotalpages = 0, totalpages = 0;
5081 for (i = 0; i < MAX_NR_ZONES; i++) {
5082 struct zone *zone = pgdat->node_zones + i;
5083 unsigned long size, real_size;
5085 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5089 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5090 node_start_pfn, node_end_pfn,
5092 zone->spanned_pages = size;
5093 zone->present_pages = real_size;
5096 realtotalpages += real_size;
5099 pgdat->node_spanned_pages = totalpages;
5100 pgdat->node_present_pages = realtotalpages;
5101 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5105 #ifndef CONFIG_SPARSEMEM
5107 * Calculate the size of the zone->blockflags rounded to an unsigned long
5108 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5109 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5110 * round what is now in bits to nearest long in bits, then return it in
5113 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5115 unsigned long usemapsize;
5117 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5118 usemapsize = roundup(zonesize, pageblock_nr_pages);
5119 usemapsize = usemapsize >> pageblock_order;
5120 usemapsize *= NR_PAGEBLOCK_BITS;
5121 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5123 return usemapsize / 8;
5126 static void __init setup_usemap(struct pglist_data *pgdat,
5128 unsigned long zone_start_pfn,
5129 unsigned long zonesize)
5131 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5132 zone->pageblock_flags = NULL;
5134 zone->pageblock_flags =
5135 memblock_virt_alloc_node_nopanic(usemapsize,
5139 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5140 unsigned long zone_start_pfn, unsigned long zonesize) {}
5141 #endif /* CONFIG_SPARSEMEM */
5143 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5145 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5146 void __paginginit set_pageblock_order(void)
5150 /* Check that pageblock_nr_pages has not already been setup */
5151 if (pageblock_order)
5154 if (HPAGE_SHIFT > PAGE_SHIFT)
5155 order = HUGETLB_PAGE_ORDER;
5157 order = MAX_ORDER - 1;
5160 * Assume the largest contiguous order of interest is a huge page.
5161 * This value may be variable depending on boot parameters on IA64 and
5164 pageblock_order = order;
5166 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5169 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5170 * is unused as pageblock_order is set at compile-time. See
5171 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5174 void __paginginit set_pageblock_order(void)
5178 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5180 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5181 unsigned long present_pages)
5183 unsigned long pages = spanned_pages;
5186 * Provide a more accurate estimation if there are holes within
5187 * the zone and SPARSEMEM is in use. If there are holes within the
5188 * zone, each populated memory region may cost us one or two extra
5189 * memmap pages due to alignment because memmap pages for each
5190 * populated regions may not naturally algined on page boundary.
5191 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5193 if (spanned_pages > present_pages + (present_pages >> 4) &&
5194 IS_ENABLED(CONFIG_SPARSEMEM))
5195 pages = present_pages;
5197 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5201 * Set up the zone data structures:
5202 * - mark all pages reserved
5203 * - mark all memory queues empty
5204 * - clear the memory bitmaps
5206 * NOTE: pgdat should get zeroed by caller.
5208 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5211 int nid = pgdat->node_id;
5212 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5215 pgdat_resize_init(pgdat);
5216 #ifdef CONFIG_NUMA_BALANCING
5217 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5218 pgdat->numabalancing_migrate_nr_pages = 0;
5219 pgdat->numabalancing_migrate_next_window = jiffies;
5221 init_waitqueue_head(&pgdat->kswapd_wait);
5222 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5223 pgdat_page_ext_init(pgdat);
5225 for (j = 0; j < MAX_NR_ZONES; j++) {
5226 struct zone *zone = pgdat->node_zones + j;
5227 unsigned long size, realsize, freesize, memmap_pages;
5229 size = zone->spanned_pages;
5230 realsize = freesize = zone->present_pages;
5233 * Adjust freesize so that it accounts for how much memory
5234 * is used by this zone for memmap. This affects the watermark
5235 * and per-cpu initialisations
5237 memmap_pages = calc_memmap_size(size, realsize);
5238 if (!is_highmem_idx(j)) {
5239 if (freesize >= memmap_pages) {
5240 freesize -= memmap_pages;
5243 " %s zone: %lu pages used for memmap\n",
5244 zone_names[j], memmap_pages);
5247 " %s zone: %lu pages exceeds freesize %lu\n",
5248 zone_names[j], memmap_pages, freesize);
5251 /* Account for reserved pages */
5252 if (j == 0 && freesize > dma_reserve) {
5253 freesize -= dma_reserve;
5254 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5255 zone_names[0], dma_reserve);
5258 if (!is_highmem_idx(j))
5259 nr_kernel_pages += freesize;
5260 /* Charge for highmem memmap if there are enough kernel pages */
5261 else if (nr_kernel_pages > memmap_pages * 2)
5262 nr_kernel_pages -= memmap_pages;
5263 nr_all_pages += freesize;
5266 * Set an approximate value for lowmem here, it will be adjusted
5267 * when the bootmem allocator frees pages into the buddy system.
5268 * And all highmem pages will be managed by the buddy system.
5270 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5273 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5275 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5277 zone->name = zone_names[j];
5278 spin_lock_init(&zone->lock);
5279 spin_lock_init(&zone->lru_lock);
5280 zone_seqlock_init(zone);
5281 zone->zone_pgdat = pgdat;
5282 zone_pcp_init(zone);
5284 /* For bootup, initialized properly in watermark setup */
5285 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5287 lruvec_init(&zone->lruvec);
5291 set_pageblock_order();
5292 setup_usemap(pgdat, zone, zone_start_pfn, size);
5293 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5295 memmap_init(size, nid, j, zone_start_pfn);
5296 zone_start_pfn += size;
5300 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5302 unsigned long __maybe_unused start = 0;
5303 unsigned long __maybe_unused offset = 0;
5305 /* Skip empty nodes */
5306 if (!pgdat->node_spanned_pages)
5309 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5310 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5311 offset = pgdat->node_start_pfn - start;
5312 /* ia64 gets its own node_mem_map, before this, without bootmem */
5313 if (!pgdat->node_mem_map) {
5314 unsigned long size, end;
5318 * The zone's endpoints aren't required to be MAX_ORDER
5319 * aligned but the node_mem_map endpoints must be in order
5320 * for the buddy allocator to function correctly.
5322 end = pgdat_end_pfn(pgdat);
5323 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5324 size = (end - start) * sizeof(struct page);
5325 map = alloc_remap(pgdat->node_id, size);
5327 map = memblock_virt_alloc_node_nopanic(size,
5329 pgdat->node_mem_map = map + offset;
5331 #ifndef CONFIG_NEED_MULTIPLE_NODES
5333 * With no DISCONTIG, the global mem_map is just set as node 0's
5335 if (pgdat == NODE_DATA(0)) {
5336 mem_map = NODE_DATA(0)->node_mem_map;
5337 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5338 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5340 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5343 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5346 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5347 unsigned long node_start_pfn, unsigned long *zholes_size)
5349 pg_data_t *pgdat = NODE_DATA(nid);
5350 unsigned long start_pfn = 0;
5351 unsigned long end_pfn = 0;
5353 /* pg_data_t should be reset to zero when it's allocated */
5354 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5356 reset_deferred_meminit(pgdat);
5357 pgdat->node_id = nid;
5358 pgdat->node_start_pfn = node_start_pfn;
5359 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5360 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5361 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5362 (u64)start_pfn << PAGE_SHIFT,
5363 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5365 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5366 zones_size, zholes_size);
5368 alloc_node_mem_map(pgdat);
5369 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5370 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5371 nid, (unsigned long)pgdat,
5372 (unsigned long)pgdat->node_mem_map);
5375 free_area_init_core(pgdat);
5378 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5380 #if MAX_NUMNODES > 1
5382 * Figure out the number of possible node ids.
5384 void __init setup_nr_node_ids(void)
5386 unsigned int highest;
5388 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5389 nr_node_ids = highest + 1;
5394 * node_map_pfn_alignment - determine the maximum internode alignment
5396 * This function should be called after node map is populated and sorted.
5397 * It calculates the maximum power of two alignment which can distinguish
5400 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5401 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5402 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5403 * shifted, 1GiB is enough and this function will indicate so.
5405 * This is used to test whether pfn -> nid mapping of the chosen memory
5406 * model has fine enough granularity to avoid incorrect mapping for the
5407 * populated node map.
5409 * Returns the determined alignment in pfn's. 0 if there is no alignment
5410 * requirement (single node).
5412 unsigned long __init node_map_pfn_alignment(void)
5414 unsigned long accl_mask = 0, last_end = 0;
5415 unsigned long start, end, mask;
5419 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5420 if (!start || last_nid < 0 || last_nid == nid) {
5427 * Start with a mask granular enough to pin-point to the
5428 * start pfn and tick off bits one-by-one until it becomes
5429 * too coarse to separate the current node from the last.
5431 mask = ~((1 << __ffs(start)) - 1);
5432 while (mask && last_end <= (start & (mask << 1)))
5435 /* accumulate all internode masks */
5439 /* convert mask to number of pages */
5440 return ~accl_mask + 1;
5443 /* Find the lowest pfn for a node */
5444 static unsigned long __init find_min_pfn_for_node(int nid)
5446 unsigned long min_pfn = ULONG_MAX;
5447 unsigned long start_pfn;
5450 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5451 min_pfn = min(min_pfn, start_pfn);
5453 if (min_pfn == ULONG_MAX) {
5455 "Could not find start_pfn for node %d\n", nid);
5463 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5465 * It returns the minimum PFN based on information provided via
5466 * memblock_set_node().
5468 unsigned long __init find_min_pfn_with_active_regions(void)
5470 return find_min_pfn_for_node(MAX_NUMNODES);
5474 * early_calculate_totalpages()
5475 * Sum pages in active regions for movable zone.
5476 * Populate N_MEMORY for calculating usable_nodes.
5478 static unsigned long __init early_calculate_totalpages(void)
5480 unsigned long totalpages = 0;
5481 unsigned long start_pfn, end_pfn;
5484 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5485 unsigned long pages = end_pfn - start_pfn;
5487 totalpages += pages;
5489 node_set_state(nid, N_MEMORY);
5495 * Find the PFN the Movable zone begins in each node. Kernel memory
5496 * is spread evenly between nodes as long as the nodes have enough
5497 * memory. When they don't, some nodes will have more kernelcore than
5500 static void __init find_zone_movable_pfns_for_nodes(void)
5503 unsigned long usable_startpfn;
5504 unsigned long kernelcore_node, kernelcore_remaining;
5505 /* save the state before borrow the nodemask */
5506 nodemask_t saved_node_state = node_states[N_MEMORY];
5507 unsigned long totalpages = early_calculate_totalpages();
5508 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5509 struct memblock_region *r;
5511 /* Need to find movable_zone earlier when movable_node is specified. */
5512 find_usable_zone_for_movable();
5515 * If movable_node is specified, ignore kernelcore and movablecore
5518 if (movable_node_is_enabled()) {
5519 for_each_memblock(memory, r) {
5520 if (!memblock_is_hotpluggable(r))
5525 usable_startpfn = PFN_DOWN(r->base);
5526 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5527 min(usable_startpfn, zone_movable_pfn[nid]) :
5535 * If movablecore=nn[KMG] was specified, calculate what size of
5536 * kernelcore that corresponds so that memory usable for
5537 * any allocation type is evenly spread. If both kernelcore
5538 * and movablecore are specified, then the value of kernelcore
5539 * will be used for required_kernelcore if it's greater than
5540 * what movablecore would have allowed.
5542 if (required_movablecore) {
5543 unsigned long corepages;
5546 * Round-up so that ZONE_MOVABLE is at least as large as what
5547 * was requested by the user
5549 required_movablecore =
5550 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5551 required_movablecore = min(totalpages, required_movablecore);
5552 corepages = totalpages - required_movablecore;
5554 required_kernelcore = max(required_kernelcore, corepages);
5558 * If kernelcore was not specified or kernelcore size is larger
5559 * than totalpages, there is no ZONE_MOVABLE.
5561 if (!required_kernelcore || required_kernelcore >= totalpages)
5564 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5565 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5568 /* Spread kernelcore memory as evenly as possible throughout nodes */
5569 kernelcore_node = required_kernelcore / usable_nodes;
5570 for_each_node_state(nid, N_MEMORY) {
5571 unsigned long start_pfn, end_pfn;
5574 * Recalculate kernelcore_node if the division per node
5575 * now exceeds what is necessary to satisfy the requested
5576 * amount of memory for the kernel
5578 if (required_kernelcore < kernelcore_node)
5579 kernelcore_node = required_kernelcore / usable_nodes;
5582 * As the map is walked, we track how much memory is usable
5583 * by the kernel using kernelcore_remaining. When it is
5584 * 0, the rest of the node is usable by ZONE_MOVABLE
5586 kernelcore_remaining = kernelcore_node;
5588 /* Go through each range of PFNs within this node */
5589 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5590 unsigned long size_pages;
5592 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5593 if (start_pfn >= end_pfn)
5596 /* Account for what is only usable for kernelcore */
5597 if (start_pfn < usable_startpfn) {
5598 unsigned long kernel_pages;
5599 kernel_pages = min(end_pfn, usable_startpfn)
5602 kernelcore_remaining -= min(kernel_pages,
5603 kernelcore_remaining);
5604 required_kernelcore -= min(kernel_pages,
5605 required_kernelcore);
5607 /* Continue if range is now fully accounted */
5608 if (end_pfn <= usable_startpfn) {
5611 * Push zone_movable_pfn to the end so
5612 * that if we have to rebalance
5613 * kernelcore across nodes, we will
5614 * not double account here
5616 zone_movable_pfn[nid] = end_pfn;
5619 start_pfn = usable_startpfn;
5623 * The usable PFN range for ZONE_MOVABLE is from
5624 * start_pfn->end_pfn. Calculate size_pages as the
5625 * number of pages used as kernelcore
5627 size_pages = end_pfn - start_pfn;
5628 if (size_pages > kernelcore_remaining)
5629 size_pages = kernelcore_remaining;
5630 zone_movable_pfn[nid] = start_pfn + size_pages;
5633 * Some kernelcore has been met, update counts and
5634 * break if the kernelcore for this node has been
5637 required_kernelcore -= min(required_kernelcore,
5639 kernelcore_remaining -= size_pages;
5640 if (!kernelcore_remaining)
5646 * If there is still required_kernelcore, we do another pass with one
5647 * less node in the count. This will push zone_movable_pfn[nid] further
5648 * along on the nodes that still have memory until kernelcore is
5652 if (usable_nodes && required_kernelcore > usable_nodes)
5656 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5657 for (nid = 0; nid < MAX_NUMNODES; nid++)
5658 zone_movable_pfn[nid] =
5659 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5662 /* restore the node_state */
5663 node_states[N_MEMORY] = saved_node_state;
5666 /* Any regular or high memory on that node ? */
5667 static void check_for_memory(pg_data_t *pgdat, int nid)
5669 enum zone_type zone_type;
5671 if (N_MEMORY == N_NORMAL_MEMORY)
5674 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5675 struct zone *zone = &pgdat->node_zones[zone_type];
5676 if (populated_zone(zone)) {
5677 node_set_state(nid, N_HIGH_MEMORY);
5678 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5679 zone_type <= ZONE_NORMAL)
5680 node_set_state(nid, N_NORMAL_MEMORY);
5687 * free_area_init_nodes - Initialise all pg_data_t and zone data
5688 * @max_zone_pfn: an array of max PFNs for each zone
5690 * This will call free_area_init_node() for each active node in the system.
5691 * Using the page ranges provided by memblock_set_node(), the size of each
5692 * zone in each node and their holes is calculated. If the maximum PFN
5693 * between two adjacent zones match, it is assumed that the zone is empty.
5694 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5695 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5696 * starts where the previous one ended. For example, ZONE_DMA32 starts
5697 * at arch_max_dma_pfn.
5699 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5701 unsigned long start_pfn, end_pfn;
5704 /* Record where the zone boundaries are */
5705 memset(arch_zone_lowest_possible_pfn, 0,
5706 sizeof(arch_zone_lowest_possible_pfn));
5707 memset(arch_zone_highest_possible_pfn, 0,
5708 sizeof(arch_zone_highest_possible_pfn));
5709 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5710 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5711 for (i = 1; i < MAX_NR_ZONES; i++) {
5712 if (i == ZONE_MOVABLE)
5714 arch_zone_lowest_possible_pfn[i] =
5715 arch_zone_highest_possible_pfn[i-1];
5716 arch_zone_highest_possible_pfn[i] =
5717 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5719 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5720 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5722 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5723 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5724 find_zone_movable_pfns_for_nodes();
5726 /* Print out the zone ranges */
5727 pr_info("Zone ranges:\n");
5728 for (i = 0; i < MAX_NR_ZONES; i++) {
5729 if (i == ZONE_MOVABLE)
5731 pr_info(" %-8s ", zone_names[i]);
5732 if (arch_zone_lowest_possible_pfn[i] ==
5733 arch_zone_highest_possible_pfn[i])
5736 pr_cont("[mem %#018Lx-%#018Lx]\n",
5737 (u64)arch_zone_lowest_possible_pfn[i]
5739 ((u64)arch_zone_highest_possible_pfn[i]
5740 << PAGE_SHIFT) - 1);
5743 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5744 pr_info("Movable zone start for each node\n");
5745 for (i = 0; i < MAX_NUMNODES; i++) {
5746 if (zone_movable_pfn[i])
5747 pr_info(" Node %d: %#018Lx\n", i,
5748 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5751 /* Print out the early node map */
5752 pr_info("Early memory node ranges\n");
5753 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5754 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5755 (u64)start_pfn << PAGE_SHIFT,
5756 ((u64)end_pfn << PAGE_SHIFT) - 1);
5758 /* Initialise every node */
5759 mminit_verify_pageflags_layout();
5760 setup_nr_node_ids();
5761 for_each_online_node(nid) {
5762 pg_data_t *pgdat = NODE_DATA(nid);
5763 free_area_init_node(nid, NULL,
5764 find_min_pfn_for_node(nid), NULL);
5766 /* Any memory on that node */
5767 if (pgdat->node_present_pages)
5768 node_set_state(nid, N_MEMORY);
5769 check_for_memory(pgdat, nid);
5773 static int __init cmdline_parse_core(char *p, unsigned long *core)
5775 unsigned long long coremem;
5779 coremem = memparse(p, &p);
5780 *core = coremem >> PAGE_SHIFT;
5782 /* Paranoid check that UL is enough for the coremem value */
5783 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5789 * kernelcore=size sets the amount of memory for use for allocations that
5790 * cannot be reclaimed or migrated.
5792 static int __init cmdline_parse_kernelcore(char *p)
5794 return cmdline_parse_core(p, &required_kernelcore);
5798 * movablecore=size sets the amount of memory for use for allocations that
5799 * can be reclaimed or migrated.
5801 static int __init cmdline_parse_movablecore(char *p)
5803 return cmdline_parse_core(p, &required_movablecore);
5806 early_param("kernelcore", cmdline_parse_kernelcore);
5807 early_param("movablecore", cmdline_parse_movablecore);
5809 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5811 void adjust_managed_page_count(struct page *page, long count)
5813 spin_lock(&managed_page_count_lock);
5814 page_zone(page)->managed_pages += count;
5815 totalram_pages += count;
5816 #ifdef CONFIG_HIGHMEM
5817 if (PageHighMem(page))
5818 totalhigh_pages += count;
5820 spin_unlock(&managed_page_count_lock);
5822 EXPORT_SYMBOL(adjust_managed_page_count);
5824 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5827 unsigned long pages = 0;
5829 start = (void *)PAGE_ALIGN((unsigned long)start);
5830 end = (void *)((unsigned long)end & PAGE_MASK);
5831 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5832 if ((unsigned int)poison <= 0xFF)
5833 memset(pos, poison, PAGE_SIZE);
5834 free_reserved_page(virt_to_page(pos));
5838 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5839 s, pages << (PAGE_SHIFT - 10), start, end);
5843 EXPORT_SYMBOL(free_reserved_area);
5845 #ifdef CONFIG_HIGHMEM
5846 void free_highmem_page(struct page *page)
5848 __free_reserved_page(page);
5850 page_zone(page)->managed_pages++;
5856 void __init mem_init_print_info(const char *str)
5858 unsigned long physpages, codesize, datasize, rosize, bss_size;
5859 unsigned long init_code_size, init_data_size;
5861 physpages = get_num_physpages();
5862 codesize = _etext - _stext;
5863 datasize = _edata - _sdata;
5864 rosize = __end_rodata - __start_rodata;
5865 bss_size = __bss_stop - __bss_start;
5866 init_data_size = __init_end - __init_begin;
5867 init_code_size = _einittext - _sinittext;
5870 * Detect special cases and adjust section sizes accordingly:
5871 * 1) .init.* may be embedded into .data sections
5872 * 2) .init.text.* may be out of [__init_begin, __init_end],
5873 * please refer to arch/tile/kernel/vmlinux.lds.S.
5874 * 3) .rodata.* may be embedded into .text or .data sections.
5876 #define adj_init_size(start, end, size, pos, adj) \
5878 if (start <= pos && pos < end && size > adj) \
5882 adj_init_size(__init_begin, __init_end, init_data_size,
5883 _sinittext, init_code_size);
5884 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5885 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5886 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5887 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5889 #undef adj_init_size
5891 pr_info("Memory: %luK/%luK available "
5892 "(%luK kernel code, %luK rwdata, %luK rodata, "
5893 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5894 #ifdef CONFIG_HIGHMEM
5898 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5899 codesize >> 10, datasize >> 10, rosize >> 10,
5900 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5901 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5902 totalcma_pages << (PAGE_SHIFT-10),
5903 #ifdef CONFIG_HIGHMEM
5904 totalhigh_pages << (PAGE_SHIFT-10),
5906 str ? ", " : "", str ? str : "");
5910 * set_dma_reserve - set the specified number of pages reserved in the first zone
5911 * @new_dma_reserve: The number of pages to mark reserved
5913 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5914 * In the DMA zone, a significant percentage may be consumed by kernel image
5915 * and other unfreeable allocations which can skew the watermarks badly. This
5916 * function may optionally be used to account for unfreeable pages in the
5917 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5918 * smaller per-cpu batchsize.
5920 void __init set_dma_reserve(unsigned long new_dma_reserve)
5922 dma_reserve = new_dma_reserve;
5925 void __init free_area_init(unsigned long *zones_size)
5927 free_area_init_node(0, zones_size,
5928 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5931 static int page_alloc_cpu_notify(struct notifier_block *self,
5932 unsigned long action, void *hcpu)
5934 int cpu = (unsigned long)hcpu;
5936 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5937 lru_add_drain_cpu(cpu);
5941 * Spill the event counters of the dead processor
5942 * into the current processors event counters.
5943 * This artificially elevates the count of the current
5946 vm_events_fold_cpu(cpu);
5949 * Zero the differential counters of the dead processor
5950 * so that the vm statistics are consistent.
5952 * This is only okay since the processor is dead and cannot
5953 * race with what we are doing.
5955 cpu_vm_stats_fold(cpu);
5960 void __init page_alloc_init(void)
5962 hotcpu_notifier(page_alloc_cpu_notify, 0);
5966 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5967 * or min_free_kbytes changes.
5969 static void calculate_totalreserve_pages(void)
5971 struct pglist_data *pgdat;
5972 unsigned long reserve_pages = 0;
5973 enum zone_type i, j;
5975 for_each_online_pgdat(pgdat) {
5976 for (i = 0; i < MAX_NR_ZONES; i++) {
5977 struct zone *zone = pgdat->node_zones + i;
5980 /* Find valid and maximum lowmem_reserve in the zone */
5981 for (j = i; j < MAX_NR_ZONES; j++) {
5982 if (zone->lowmem_reserve[j] > max)
5983 max = zone->lowmem_reserve[j];
5986 /* we treat the high watermark as reserved pages. */
5987 max += high_wmark_pages(zone);
5989 if (max > zone->managed_pages)
5990 max = zone->managed_pages;
5991 reserve_pages += max;
5993 * Lowmem reserves are not available to
5994 * GFP_HIGHUSER page cache allocations and
5995 * kswapd tries to balance zones to their high
5996 * watermark. As a result, neither should be
5997 * regarded as dirtyable memory, to prevent a
5998 * situation where reclaim has to clean pages
5999 * in order to balance the zones.
6001 zone->dirty_balance_reserve = max;
6004 dirty_balance_reserve = reserve_pages;
6005 totalreserve_pages = reserve_pages;
6009 * setup_per_zone_lowmem_reserve - called whenever
6010 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6011 * has a correct pages reserved value, so an adequate number of
6012 * pages are left in the zone after a successful __alloc_pages().
6014 static void setup_per_zone_lowmem_reserve(void)
6016 struct pglist_data *pgdat;
6017 enum zone_type j, idx;
6019 for_each_online_pgdat(pgdat) {
6020 for (j = 0; j < MAX_NR_ZONES; j++) {
6021 struct zone *zone = pgdat->node_zones + j;
6022 unsigned long managed_pages = zone->managed_pages;
6024 zone->lowmem_reserve[j] = 0;
6028 struct zone *lower_zone;
6032 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6033 sysctl_lowmem_reserve_ratio[idx] = 1;
6035 lower_zone = pgdat->node_zones + idx;
6036 lower_zone->lowmem_reserve[j] = managed_pages /
6037 sysctl_lowmem_reserve_ratio[idx];
6038 managed_pages += lower_zone->managed_pages;
6043 /* update totalreserve_pages */
6044 calculate_totalreserve_pages();
6047 static void __setup_per_zone_wmarks(void)
6049 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6050 unsigned long pages_low = extra_free_kbytes >> (PAGE_SHIFT - 10);
6051 unsigned long lowmem_pages = 0;
6053 unsigned long flags;
6055 /* Calculate total number of !ZONE_HIGHMEM pages */
6056 for_each_zone(zone) {
6057 if (!is_highmem(zone))
6058 lowmem_pages += zone->managed_pages;
6061 for_each_zone(zone) {
6064 spin_lock_irqsave(&zone->lock, flags);
6065 min = (u64)pages_min * zone->managed_pages;
6066 do_div(min, lowmem_pages);
6067 low = (u64)pages_low * zone->managed_pages;
6068 do_div(low, vm_total_pages);
6070 if (is_highmem(zone)) {
6072 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6073 * need highmem pages, so cap pages_min to a small
6076 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6077 * deltas control asynch page reclaim, and so should
6078 * not be capped for highmem.
6080 unsigned long min_pages;
6082 min_pages = zone->managed_pages / 1024;
6083 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6084 zone->watermark[WMARK_MIN] = min_pages;
6087 * If it's a lowmem zone, reserve a number of pages
6088 * proportionate to the zone's size.
6090 zone->watermark[WMARK_MIN] = min;
6093 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) +
6095 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) +
6098 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6099 high_wmark_pages(zone) - low_wmark_pages(zone) -
6100 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6102 spin_unlock_irqrestore(&zone->lock, flags);
6105 /* update totalreserve_pages */
6106 calculate_totalreserve_pages();
6110 * setup_per_zone_wmarks - called when min_free_kbytes changes
6111 * or when memory is hot-{added|removed}
6113 * Ensures that the watermark[min,low,high] values for each zone are set
6114 * correctly with respect to min_free_kbytes.
6116 void setup_per_zone_wmarks(void)
6118 mutex_lock(&zonelists_mutex);
6119 __setup_per_zone_wmarks();
6120 mutex_unlock(&zonelists_mutex);
6124 * The inactive anon list should be small enough that the VM never has to
6125 * do too much work, but large enough that each inactive page has a chance
6126 * to be referenced again before it is swapped out.
6128 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6129 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6130 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6131 * the anonymous pages are kept on the inactive list.
6134 * memory ratio inactive anon
6135 * -------------------------------------
6144 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6146 unsigned int gb, ratio;
6148 /* Zone size in gigabytes */
6149 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6151 ratio = int_sqrt(10 * gb);
6155 zone->inactive_ratio = ratio;
6158 static void __meminit setup_per_zone_inactive_ratio(void)
6163 calculate_zone_inactive_ratio(zone);
6167 * Initialise min_free_kbytes.
6169 * For small machines we want it small (128k min). For large machines
6170 * we want it large (64MB max). But it is not linear, because network
6171 * bandwidth does not increase linearly with machine size. We use
6173 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6174 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6190 int __meminit init_per_zone_wmark_min(void)
6192 unsigned long lowmem_kbytes;
6193 int new_min_free_kbytes;
6195 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6196 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6198 if (new_min_free_kbytes > user_min_free_kbytes) {
6199 min_free_kbytes = new_min_free_kbytes;
6200 if (min_free_kbytes < 128)
6201 min_free_kbytes = 128;
6202 if (min_free_kbytes > 65536)
6203 min_free_kbytes = 65536;
6205 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6206 new_min_free_kbytes, user_min_free_kbytes);
6208 setup_per_zone_wmarks();
6209 refresh_zone_stat_thresholds();
6210 setup_per_zone_lowmem_reserve();
6211 setup_per_zone_inactive_ratio();
6214 core_initcall(init_per_zone_wmark_min)
6217 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6218 * that we can call two helper functions whenever min_free_kbytes
6219 * or extra_free_kbytes changes.
6221 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6222 void __user *buffer, size_t *length, loff_t *ppos)
6226 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6231 user_min_free_kbytes = min_free_kbytes;
6232 setup_per_zone_wmarks();
6238 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6239 void __user *buffer, size_t *length, loff_t *ppos)
6244 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6249 zone->min_unmapped_pages = (zone->managed_pages *
6250 sysctl_min_unmapped_ratio) / 100;
6254 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6255 void __user *buffer, size_t *length, loff_t *ppos)
6260 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6265 zone->min_slab_pages = (zone->managed_pages *
6266 sysctl_min_slab_ratio) / 100;
6272 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6273 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6274 * whenever sysctl_lowmem_reserve_ratio changes.
6276 * The reserve ratio obviously has absolutely no relation with the
6277 * minimum watermarks. The lowmem reserve ratio can only make sense
6278 * if in function of the boot time zone sizes.
6280 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6281 void __user *buffer, size_t *length, loff_t *ppos)
6283 proc_dointvec_minmax(table, write, buffer, length, ppos);
6284 setup_per_zone_lowmem_reserve();
6289 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6290 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6291 * pagelist can have before it gets flushed back to buddy allocator.
6293 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6294 void __user *buffer, size_t *length, loff_t *ppos)
6297 int old_percpu_pagelist_fraction;
6300 mutex_lock(&pcp_batch_high_lock);
6301 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6303 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6304 if (!write || ret < 0)
6307 /* Sanity checking to avoid pcp imbalance */
6308 if (percpu_pagelist_fraction &&
6309 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6310 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6316 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6319 for_each_populated_zone(zone) {
6322 for_each_possible_cpu(cpu)
6323 pageset_set_high_and_batch(zone,
6324 per_cpu_ptr(zone->pageset, cpu));
6327 mutex_unlock(&pcp_batch_high_lock);
6332 int hashdist = HASHDIST_DEFAULT;
6334 static int __init set_hashdist(char *str)
6338 hashdist = simple_strtoul(str, &str, 0);
6341 __setup("hashdist=", set_hashdist);
6345 * allocate a large system hash table from bootmem
6346 * - it is assumed that the hash table must contain an exact power-of-2
6347 * quantity of entries
6348 * - limit is the number of hash buckets, not the total allocation size
6350 void *__init alloc_large_system_hash(const char *tablename,
6351 unsigned long bucketsize,
6352 unsigned long numentries,
6355 unsigned int *_hash_shift,
6356 unsigned int *_hash_mask,
6357 unsigned long low_limit,
6358 unsigned long high_limit)
6360 unsigned long long max = high_limit;
6361 unsigned long log2qty, size;
6364 /* allow the kernel cmdline to have a say */
6366 /* round applicable memory size up to nearest megabyte */
6367 numentries = nr_kernel_pages;
6369 /* It isn't necessary when PAGE_SIZE >= 1MB */
6370 if (PAGE_SHIFT < 20)
6371 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6373 /* limit to 1 bucket per 2^scale bytes of low memory */
6374 if (scale > PAGE_SHIFT)
6375 numentries >>= (scale - PAGE_SHIFT);
6377 numentries <<= (PAGE_SHIFT - scale);
6379 /* Make sure we've got at least a 0-order allocation.. */
6380 if (unlikely(flags & HASH_SMALL)) {
6381 /* Makes no sense without HASH_EARLY */
6382 WARN_ON(!(flags & HASH_EARLY));
6383 if (!(numentries >> *_hash_shift)) {
6384 numentries = 1UL << *_hash_shift;
6385 BUG_ON(!numentries);
6387 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6388 numentries = PAGE_SIZE / bucketsize;
6390 numentries = roundup_pow_of_two(numentries);
6392 /* limit allocation size to 1/16 total memory by default */
6394 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6395 do_div(max, bucketsize);
6397 max = min(max, 0x80000000ULL);
6399 if (numentries < low_limit)
6400 numentries = low_limit;
6401 if (numentries > max)
6404 log2qty = ilog2(numentries);
6407 size = bucketsize << log2qty;
6408 if (flags & HASH_EARLY)
6409 table = memblock_virt_alloc_nopanic(size, 0);
6411 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6414 * If bucketsize is not a power-of-two, we may free
6415 * some pages at the end of hash table which
6416 * alloc_pages_exact() automatically does
6418 if (get_order(size) < MAX_ORDER) {
6419 table = alloc_pages_exact(size, GFP_ATOMIC);
6420 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6423 } while (!table && size > PAGE_SIZE && --log2qty);
6426 panic("Failed to allocate %s hash table\n", tablename);
6428 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6431 ilog2(size) - PAGE_SHIFT,
6435 *_hash_shift = log2qty;
6437 *_hash_mask = (1 << log2qty) - 1;
6442 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6443 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6446 #ifdef CONFIG_SPARSEMEM
6447 return __pfn_to_section(pfn)->pageblock_flags;
6449 return zone->pageblock_flags;
6450 #endif /* CONFIG_SPARSEMEM */
6453 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6455 #ifdef CONFIG_SPARSEMEM
6456 pfn &= (PAGES_PER_SECTION-1);
6457 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6459 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6460 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6461 #endif /* CONFIG_SPARSEMEM */
6465 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6466 * @page: The page within the block of interest
6467 * @pfn: The target page frame number
6468 * @end_bitidx: The last bit of interest to retrieve
6469 * @mask: mask of bits that the caller is interested in
6471 * Return: pageblock_bits flags
6473 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6474 unsigned long end_bitidx,
6478 unsigned long *bitmap;
6479 unsigned long bitidx, word_bitidx;
6482 zone = page_zone(page);
6483 bitmap = get_pageblock_bitmap(zone, pfn);
6484 bitidx = pfn_to_bitidx(zone, pfn);
6485 word_bitidx = bitidx / BITS_PER_LONG;
6486 bitidx &= (BITS_PER_LONG-1);
6488 word = bitmap[word_bitidx];
6489 bitidx += end_bitidx;
6490 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6494 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6495 * @page: The page within the block of interest
6496 * @flags: The flags to set
6497 * @pfn: The target page frame number
6498 * @end_bitidx: The last bit of interest
6499 * @mask: mask of bits that the caller is interested in
6501 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6503 unsigned long end_bitidx,
6507 unsigned long *bitmap;
6508 unsigned long bitidx, word_bitidx;
6509 unsigned long old_word, word;
6511 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6513 zone = page_zone(page);
6514 bitmap = get_pageblock_bitmap(zone, pfn);
6515 bitidx = pfn_to_bitidx(zone, pfn);
6516 word_bitidx = bitidx / BITS_PER_LONG;
6517 bitidx &= (BITS_PER_LONG-1);
6519 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6521 bitidx += end_bitidx;
6522 mask <<= (BITS_PER_LONG - bitidx - 1);
6523 flags <<= (BITS_PER_LONG - bitidx - 1);
6525 word = READ_ONCE(bitmap[word_bitidx]);
6527 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6528 if (word == old_word)
6535 * This function checks whether pageblock includes unmovable pages or not.
6536 * If @count is not zero, it is okay to include less @count unmovable pages
6538 * PageLRU check without isolation or lru_lock could race so that
6539 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6540 * expect this function should be exact.
6542 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6543 bool skip_hwpoisoned_pages)
6545 unsigned long pfn, iter, found;
6549 * For avoiding noise data, lru_add_drain_all() should be called
6550 * If ZONE_MOVABLE, the zone never contains unmovable pages
6552 if (zone_idx(zone) == ZONE_MOVABLE)
6554 mt = get_pageblock_migratetype(page);
6555 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6558 pfn = page_to_pfn(page);
6559 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6560 unsigned long check = pfn + iter;
6562 if (!pfn_valid_within(check))
6565 page = pfn_to_page(check);
6568 * Hugepages are not in LRU lists, but they're movable.
6569 * We need not scan over tail pages bacause we don't
6570 * handle each tail page individually in migration.
6572 if (PageHuge(page)) {
6573 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6578 * We can't use page_count without pin a page
6579 * because another CPU can free compound page.
6580 * This check already skips compound tails of THP
6581 * because their page->_count is zero at all time.
6583 if (!atomic_read(&page->_count)) {
6584 if (PageBuddy(page))
6585 iter += (1 << page_order(page)) - 1;
6590 * The HWPoisoned page may be not in buddy system, and
6591 * page_count() is not 0.
6593 if (skip_hwpoisoned_pages && PageHWPoison(page))
6599 * If there are RECLAIMABLE pages, we need to check
6600 * it. But now, memory offline itself doesn't call
6601 * shrink_node_slabs() and it still to be fixed.
6604 * If the page is not RAM, page_count()should be 0.
6605 * we don't need more check. This is an _used_ not-movable page.
6607 * The problematic thing here is PG_reserved pages. PG_reserved
6608 * is set to both of a memory hole page and a _used_ kernel
6617 bool is_pageblock_removable_nolock(struct page *page)
6623 * We have to be careful here because we are iterating over memory
6624 * sections which are not zone aware so we might end up outside of
6625 * the zone but still within the section.
6626 * We have to take care about the node as well. If the node is offline
6627 * its NODE_DATA will be NULL - see page_zone.
6629 if (!node_online(page_to_nid(page)))
6632 zone = page_zone(page);
6633 pfn = page_to_pfn(page);
6634 if (!zone_spans_pfn(zone, pfn))
6637 return !has_unmovable_pages(zone, page, 0, true);
6642 static unsigned long pfn_max_align_down(unsigned long pfn)
6644 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6645 pageblock_nr_pages) - 1);
6648 static unsigned long pfn_max_align_up(unsigned long pfn)
6650 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6651 pageblock_nr_pages));
6654 /* [start, end) must belong to a single zone. */
6655 static int __alloc_contig_migrate_range(struct compact_control *cc,
6656 unsigned long start, unsigned long end)
6658 /* This function is based on compact_zone() from compaction.c. */
6659 unsigned long nr_reclaimed;
6660 unsigned long pfn = start;
6661 unsigned int tries = 0;
6666 while (pfn < end || !list_empty(&cc->migratepages)) {
6667 if (fatal_signal_pending(current)) {
6672 if (list_empty(&cc->migratepages)) {
6673 cc->nr_migratepages = 0;
6674 pfn = isolate_migratepages_range(cc, pfn, end);
6680 } else if (++tries == 5) {
6681 ret = ret < 0 ? ret : -EBUSY;
6685 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6687 cc->nr_migratepages -= nr_reclaimed;
6689 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6690 NULL, 0, cc->mode, MR_CMA);
6693 putback_movable_pages(&cc->migratepages);
6700 * alloc_contig_range() -- tries to allocate given range of pages
6701 * @start: start PFN to allocate
6702 * @end: one-past-the-last PFN to allocate
6703 * @migratetype: migratetype of the underlaying pageblocks (either
6704 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6705 * in range must have the same migratetype and it must
6706 * be either of the two.
6708 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6709 * aligned, however it's the caller's responsibility to guarantee that
6710 * we are the only thread that changes migrate type of pageblocks the
6713 * The PFN range must belong to a single zone.
6715 * Returns zero on success or negative error code. On success all
6716 * pages which PFN is in [start, end) are allocated for the caller and
6717 * need to be freed with free_contig_range().
6719 int alloc_contig_range(unsigned long start, unsigned long end,
6720 unsigned migratetype)
6722 unsigned long outer_start, outer_end;
6726 struct compact_control cc = {
6727 .nr_migratepages = 0,
6729 .zone = page_zone(pfn_to_page(start)),
6730 .mode = MIGRATE_SYNC,
6731 .ignore_skip_hint = true,
6733 INIT_LIST_HEAD(&cc.migratepages);
6736 * What we do here is we mark all pageblocks in range as
6737 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6738 * have different sizes, and due to the way page allocator
6739 * work, we align the range to biggest of the two pages so
6740 * that page allocator won't try to merge buddies from
6741 * different pageblocks and change MIGRATE_ISOLATE to some
6742 * other migration type.
6744 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6745 * migrate the pages from an unaligned range (ie. pages that
6746 * we are interested in). This will put all the pages in
6747 * range back to page allocator as MIGRATE_ISOLATE.
6749 * When this is done, we take the pages in range from page
6750 * allocator removing them from the buddy system. This way
6751 * page allocator will never consider using them.
6753 * This lets us mark the pageblocks back as
6754 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6755 * aligned range but not in the unaligned, original range are
6756 * put back to page allocator so that buddy can use them.
6759 ret = start_isolate_page_range(pfn_max_align_down(start),
6760 pfn_max_align_up(end), migratetype,
6765 ret = __alloc_contig_migrate_range(&cc, start, end);
6770 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6771 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6772 * more, all pages in [start, end) are free in page allocator.
6773 * What we are going to do is to allocate all pages from
6774 * [start, end) (that is remove them from page allocator).
6776 * The only problem is that pages at the beginning and at the
6777 * end of interesting range may be not aligned with pages that
6778 * page allocator holds, ie. they can be part of higher order
6779 * pages. Because of this, we reserve the bigger range and
6780 * once this is done free the pages we are not interested in.
6782 * We don't have to hold zone->lock here because the pages are
6783 * isolated thus they won't get removed from buddy.
6786 lru_add_drain_all();
6787 drain_all_pages(cc.zone);
6790 outer_start = start;
6791 while (!PageBuddy(pfn_to_page(outer_start))) {
6792 if (++order >= MAX_ORDER) {
6796 outer_start &= ~0UL << order;
6799 /* Make sure the range is really isolated. */
6800 if (test_pages_isolated(outer_start, end, false)) {
6801 pr_info("%s: [%lx, %lx) PFNs busy\n",
6802 __func__, outer_start, end);
6807 /* Grab isolated pages from freelists. */
6808 outer_end = isolate_freepages_range(&cc, outer_start, end);
6814 /* Free head and tail (if any) */
6815 if (start != outer_start)
6816 free_contig_range(outer_start, start - outer_start);
6817 if (end != outer_end)
6818 free_contig_range(end, outer_end - end);
6821 undo_isolate_page_range(pfn_max_align_down(start),
6822 pfn_max_align_up(end), migratetype);
6826 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6828 unsigned int count = 0;
6830 for (; nr_pages--; pfn++) {
6831 struct page *page = pfn_to_page(pfn);
6833 count += page_count(page) != 1;
6836 WARN(count != 0, "%d pages are still in use!\n", count);
6840 #ifdef CONFIG_MEMORY_HOTPLUG
6842 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6843 * page high values need to be recalulated.
6845 void __meminit zone_pcp_update(struct zone *zone)
6848 mutex_lock(&pcp_batch_high_lock);
6849 for_each_possible_cpu(cpu)
6850 pageset_set_high_and_batch(zone,
6851 per_cpu_ptr(zone->pageset, cpu));
6852 mutex_unlock(&pcp_batch_high_lock);
6856 void zone_pcp_reset(struct zone *zone)
6858 unsigned long flags;
6860 struct per_cpu_pageset *pset;
6862 /* avoid races with drain_pages() */
6863 local_irq_save(flags);
6864 if (zone->pageset != &boot_pageset) {
6865 for_each_online_cpu(cpu) {
6866 pset = per_cpu_ptr(zone->pageset, cpu);
6867 drain_zonestat(zone, pset);
6869 free_percpu(zone->pageset);
6870 zone->pageset = &boot_pageset;
6872 local_irq_restore(flags);
6875 #ifdef CONFIG_MEMORY_HOTREMOVE
6877 * All pages in the range must be isolated before calling this.
6880 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6884 unsigned int order, i;
6886 unsigned long flags;
6887 /* find the first valid pfn */
6888 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6893 zone = page_zone(pfn_to_page(pfn));
6894 spin_lock_irqsave(&zone->lock, flags);
6896 while (pfn < end_pfn) {
6897 if (!pfn_valid(pfn)) {
6901 page = pfn_to_page(pfn);
6903 * The HWPoisoned page may be not in buddy system, and
6904 * page_count() is not 0.
6906 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6908 SetPageReserved(page);
6912 BUG_ON(page_count(page));
6913 BUG_ON(!PageBuddy(page));
6914 order = page_order(page);
6915 #ifdef CONFIG_DEBUG_VM
6916 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6917 pfn, 1 << order, end_pfn);
6919 list_del(&page->lru);
6920 rmv_page_order(page);
6921 zone->free_area[order].nr_free--;
6922 for (i = 0; i < (1 << order); i++)
6923 SetPageReserved((page+i));
6924 pfn += (1 << order);
6926 spin_unlock_irqrestore(&zone->lock, flags);
6930 #ifdef CONFIG_MEMORY_FAILURE
6931 bool is_free_buddy_page(struct page *page)
6933 struct zone *zone = page_zone(page);
6934 unsigned long pfn = page_to_pfn(page);
6935 unsigned long flags;
6938 spin_lock_irqsave(&zone->lock, flags);
6939 for (order = 0; order < MAX_ORDER; order++) {
6940 struct page *page_head = page - (pfn & ((1 << order) - 1));
6942 if (PageBuddy(page_head) && page_order(page_head) >= order)
6945 spin_unlock_irqrestore(&zone->lock, flags);
6947 return order < MAX_ORDER;