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 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 int min_free_kbytes = 1024;
233 int user_min_free_kbytes = -1;
235 static unsigned long __meminitdata nr_kernel_pages;
236 static unsigned long __meminitdata nr_all_pages;
237 static unsigned long __meminitdata dma_reserve;
239 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
240 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
241 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
242 static unsigned long __initdata required_kernelcore;
243 static unsigned long __initdata required_movablecore;
244 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
246 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
248 EXPORT_SYMBOL(movable_zone);
249 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
252 int nr_node_ids __read_mostly = MAX_NUMNODES;
253 int nr_online_nodes __read_mostly = 1;
254 EXPORT_SYMBOL(nr_node_ids);
255 EXPORT_SYMBOL(nr_online_nodes);
258 int page_group_by_mobility_disabled __read_mostly;
260 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
261 static inline void reset_deferred_meminit(pg_data_t *pgdat)
263 pgdat->first_deferred_pfn = ULONG_MAX;
266 /* Returns true if the struct page for the pfn is uninitialised */
267 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
269 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
275 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
277 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
284 * Returns false when the remaining initialisation should be deferred until
285 * later in the boot cycle when it can be parallelised.
287 static inline bool update_defer_init(pg_data_t *pgdat,
288 unsigned long pfn, unsigned long zone_end,
289 unsigned long *nr_initialised)
291 /* Always populate low zones for address-contrained allocations */
292 if (zone_end < pgdat_end_pfn(pgdat))
295 /* Initialise at least 2G of the highest zone */
297 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
298 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
299 pgdat->first_deferred_pfn = pfn;
306 static inline void reset_deferred_meminit(pg_data_t *pgdat)
310 static inline bool early_page_uninitialised(unsigned long pfn)
315 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
320 static inline bool update_defer_init(pg_data_t *pgdat,
321 unsigned long pfn, unsigned long zone_end,
322 unsigned long *nr_initialised)
329 void set_pageblock_migratetype(struct page *page, int migratetype)
331 if (unlikely(page_group_by_mobility_disabled &&
332 migratetype < MIGRATE_PCPTYPES))
333 migratetype = MIGRATE_UNMOVABLE;
335 set_pageblock_flags_group(page, (unsigned long)migratetype,
336 PB_migrate, PB_migrate_end);
339 #ifdef CONFIG_DEBUG_VM
340 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
344 unsigned long pfn = page_to_pfn(page);
345 unsigned long sp, start_pfn;
348 seq = zone_span_seqbegin(zone);
349 start_pfn = zone->zone_start_pfn;
350 sp = zone->spanned_pages;
351 if (!zone_spans_pfn(zone, pfn))
353 } while (zone_span_seqretry(zone, seq));
356 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
357 pfn, zone_to_nid(zone), zone->name,
358 start_pfn, start_pfn + sp);
363 static int page_is_consistent(struct zone *zone, struct page *page)
365 if (!pfn_valid_within(page_to_pfn(page)))
367 if (zone != page_zone(page))
373 * Temporary debugging check for pages not lying within a given zone.
375 static int bad_range(struct zone *zone, struct page *page)
377 if (page_outside_zone_boundaries(zone, page))
379 if (!page_is_consistent(zone, page))
385 static inline int bad_range(struct zone *zone, struct page *page)
391 static void bad_page(struct page *page, const char *reason,
392 unsigned long bad_flags)
394 static unsigned long resume;
395 static unsigned long nr_shown;
396 static unsigned long nr_unshown;
398 /* Don't complain about poisoned pages */
399 if (PageHWPoison(page)) {
400 page_mapcount_reset(page); /* remove PageBuddy */
405 * Allow a burst of 60 reports, then keep quiet for that minute;
406 * or allow a steady drip of one report per second.
408 if (nr_shown == 60) {
409 if (time_before(jiffies, resume)) {
415 "BUG: Bad page state: %lu messages suppressed\n",
422 resume = jiffies + 60 * HZ;
424 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
425 current->comm, page_to_pfn(page));
426 dump_page_badflags(page, reason, bad_flags);
431 /* Leave bad fields for debug, except PageBuddy could make trouble */
432 page_mapcount_reset(page); /* remove PageBuddy */
433 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
437 * Higher-order pages are called "compound pages". They are structured thusly:
439 * The first PAGE_SIZE page is called the "head page".
441 * The remaining PAGE_SIZE pages are called "tail pages".
443 * All pages have PG_compound set. All tail pages have their ->first_page
444 * pointing at the head page.
446 * The first tail page's ->lru.next holds the address of the compound page's
447 * put_page() function. Its ->lru.prev holds the order of allocation.
448 * This usage means that zero-order pages may not be compound.
451 static void free_compound_page(struct page *page)
453 __free_pages_ok(page, compound_order(page));
456 void prep_compound_page(struct page *page, unsigned long order)
459 int nr_pages = 1 << order;
461 set_compound_page_dtor(page, free_compound_page);
462 set_compound_order(page, order);
464 for (i = 1; i < nr_pages; i++) {
465 struct page *p = page + i;
466 set_page_count(p, 0);
467 p->first_page = page;
468 /* Make sure p->first_page is always valid for PageTail() */
474 #ifdef CONFIG_DEBUG_PAGEALLOC
475 unsigned int _debug_guardpage_minorder;
476 bool _debug_pagealloc_enabled __read_mostly;
477 bool _debug_guardpage_enabled __read_mostly;
479 static int __init early_debug_pagealloc(char *buf)
484 if (strcmp(buf, "on") == 0)
485 _debug_pagealloc_enabled = true;
489 early_param("debug_pagealloc", early_debug_pagealloc);
491 static bool need_debug_guardpage(void)
493 /* If we don't use debug_pagealloc, we don't need guard page */
494 if (!debug_pagealloc_enabled())
500 static void init_debug_guardpage(void)
502 if (!debug_pagealloc_enabled())
505 _debug_guardpage_enabled = true;
508 struct page_ext_operations debug_guardpage_ops = {
509 .need = need_debug_guardpage,
510 .init = init_debug_guardpage,
513 static int __init debug_guardpage_minorder_setup(char *buf)
517 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
518 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
521 _debug_guardpage_minorder = res;
522 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
525 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
527 static inline void set_page_guard(struct zone *zone, struct page *page,
528 unsigned int order, int migratetype)
530 struct page_ext *page_ext;
532 if (!debug_guardpage_enabled())
535 page_ext = lookup_page_ext(page);
536 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
538 INIT_LIST_HEAD(&page->lru);
539 set_page_private(page, order);
540 /* Guard pages are not available for any usage */
541 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
544 static inline void clear_page_guard(struct zone *zone, struct page *page,
545 unsigned int order, int migratetype)
547 struct page_ext *page_ext;
549 if (!debug_guardpage_enabled())
552 page_ext = lookup_page_ext(page);
553 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
555 set_page_private(page, 0);
556 if (!is_migrate_isolate(migratetype))
557 __mod_zone_freepage_state(zone, (1 << order), migratetype);
560 struct page_ext_operations debug_guardpage_ops = { NULL, };
561 static inline void set_page_guard(struct zone *zone, struct page *page,
562 unsigned int order, int migratetype) {}
563 static inline void clear_page_guard(struct zone *zone, struct page *page,
564 unsigned int order, int migratetype) {}
567 static inline void set_page_order(struct page *page, unsigned int order)
569 set_page_private(page, order);
570 __SetPageBuddy(page);
573 static inline void rmv_page_order(struct page *page)
575 __ClearPageBuddy(page);
576 set_page_private(page, 0);
580 * This function checks whether a page is free && is the buddy
581 * we can do coalesce a page and its buddy if
582 * (a) the buddy is not in a hole &&
583 * (b) the buddy is in the buddy system &&
584 * (c) a page and its buddy have the same order &&
585 * (d) a page and its buddy are in the same zone.
587 * For recording whether a page is in the buddy system, we set ->_mapcount
588 * PAGE_BUDDY_MAPCOUNT_VALUE.
589 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
590 * serialized by zone->lock.
592 * For recording page's order, we use page_private(page).
594 static inline int page_is_buddy(struct page *page, struct page *buddy,
597 if (!pfn_valid_within(page_to_pfn(buddy)))
600 if (page_is_guard(buddy) && page_order(buddy) == order) {
601 if (page_zone_id(page) != page_zone_id(buddy))
604 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
609 if (PageBuddy(buddy) && page_order(buddy) == order) {
611 * zone check is done late to avoid uselessly
612 * calculating zone/node ids for pages that could
615 if (page_zone_id(page) != page_zone_id(buddy))
618 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
626 * Freeing function for a buddy system allocator.
628 * The concept of a buddy system is to maintain direct-mapped table
629 * (containing bit values) for memory blocks of various "orders".
630 * The bottom level table contains the map for the smallest allocatable
631 * units of memory (here, pages), and each level above it describes
632 * pairs of units from the levels below, hence, "buddies".
633 * At a high level, all that happens here is marking the table entry
634 * at the bottom level available, and propagating the changes upward
635 * as necessary, plus some accounting needed to play nicely with other
636 * parts of the VM system.
637 * At each level, we keep a list of pages, which are heads of continuous
638 * free pages of length of (1 << order) and marked with _mapcount
639 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
641 * So when we are allocating or freeing one, we can derive the state of the
642 * other. That is, if we allocate a small block, and both were
643 * free, the remainder of the region must be split into blocks.
644 * If a block is freed, and its buddy is also free, then this
645 * triggers coalescing into a block of larger size.
650 static inline void __free_one_page(struct page *page,
652 struct zone *zone, unsigned int order,
655 unsigned long page_idx;
656 unsigned long combined_idx;
657 unsigned long uninitialized_var(buddy_idx);
659 int max_order = MAX_ORDER;
661 VM_BUG_ON(!zone_is_initialized(zone));
662 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
664 VM_BUG_ON(migratetype == -1);
665 if (is_migrate_isolate(migratetype)) {
667 * We restrict max order of merging to prevent merge
668 * between freepages on isolate pageblock and normal
669 * pageblock. Without this, pageblock isolation
670 * could cause incorrect freepage accounting.
672 max_order = min(MAX_ORDER, pageblock_order + 1);
674 __mod_zone_freepage_state(zone, 1 << order, migratetype);
677 page_idx = pfn & ((1 << max_order) - 1);
679 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
680 VM_BUG_ON_PAGE(bad_range(zone, page), page);
682 while (order < max_order - 1) {
683 buddy_idx = __find_buddy_index(page_idx, order);
684 buddy = page + (buddy_idx - page_idx);
685 if (!page_is_buddy(page, buddy, order))
688 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
689 * merge with it and move up one order.
691 if (page_is_guard(buddy)) {
692 clear_page_guard(zone, buddy, order, migratetype);
694 list_del(&buddy->lru);
695 zone->free_area[order].nr_free--;
696 rmv_page_order(buddy);
698 combined_idx = buddy_idx & page_idx;
699 page = page + (combined_idx - page_idx);
700 page_idx = combined_idx;
703 set_page_order(page, order);
706 * If this is not the largest possible page, check if the buddy
707 * of the next-highest order is free. If it is, it's possible
708 * that pages are being freed that will coalesce soon. In case,
709 * that is happening, add the free page to the tail of the list
710 * so it's less likely to be used soon and more likely to be merged
711 * as a higher order page
713 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
714 struct page *higher_page, *higher_buddy;
715 combined_idx = buddy_idx & page_idx;
716 higher_page = page + (combined_idx - page_idx);
717 buddy_idx = __find_buddy_index(combined_idx, order + 1);
718 higher_buddy = higher_page + (buddy_idx - combined_idx);
719 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
720 list_add_tail(&page->lru,
721 &zone->free_area[order].free_list[migratetype]);
726 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
728 zone->free_area[order].nr_free++;
731 static inline int free_pages_check(struct page *page)
733 const char *bad_reason = NULL;
734 unsigned long bad_flags = 0;
736 if (unlikely(page_mapcount(page)))
737 bad_reason = "nonzero mapcount";
738 if (unlikely(page->mapping != NULL))
739 bad_reason = "non-NULL mapping";
740 if (unlikely(atomic_read(&page->_count) != 0))
741 bad_reason = "nonzero _count";
742 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
743 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
744 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
747 if (unlikely(page->mem_cgroup))
748 bad_reason = "page still charged to cgroup";
750 if (unlikely(bad_reason)) {
751 bad_page(page, bad_reason, bad_flags);
754 page_cpupid_reset_last(page);
755 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
756 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
761 * Frees a number of pages from the PCP lists
762 * Assumes all pages on list are in same zone, and of same order.
763 * count is the number of pages to free.
765 * If the zone was previously in an "all pages pinned" state then look to
766 * see if this freeing clears that state.
768 * And clear the zone's pages_scanned counter, to hold off the "all pages are
769 * pinned" detection logic.
771 static void free_pcppages_bulk(struct zone *zone, int count,
772 struct per_cpu_pages *pcp)
777 unsigned long nr_scanned;
779 spin_lock(&zone->lock);
780 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
782 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
786 struct list_head *list;
789 * Remove pages from lists in a round-robin fashion. A
790 * batch_free count is maintained that is incremented when an
791 * empty list is encountered. This is so more pages are freed
792 * off fuller lists instead of spinning excessively around empty
797 if (++migratetype == MIGRATE_PCPTYPES)
799 list = &pcp->lists[migratetype];
800 } while (list_empty(list));
802 /* This is the only non-empty list. Free them all. */
803 if (batch_free == MIGRATE_PCPTYPES)
804 batch_free = to_free;
807 int mt; /* migratetype of the to-be-freed page */
809 page = list_entry(list->prev, struct page, lru);
810 /* must delete as __free_one_page list manipulates */
811 list_del(&page->lru);
813 mt = get_pcppage_migratetype(page);
814 /* MIGRATE_ISOLATE page should not go to pcplists */
815 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
816 /* Pageblock could have been isolated meanwhile */
817 if (unlikely(has_isolate_pageblock(zone)))
818 mt = get_pageblock_migratetype(page);
820 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
821 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
822 trace_mm_page_pcpu_drain(page, 0, mt);
823 } while (--to_free && --batch_free && !list_empty(list));
825 spin_unlock(&zone->lock);
828 static void free_one_page(struct zone *zone,
829 struct page *page, unsigned long pfn,
833 unsigned long nr_scanned;
834 spin_lock(&zone->lock);
835 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
837 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
839 if (unlikely(has_isolate_pageblock(zone) ||
840 is_migrate_isolate(migratetype))) {
841 migratetype = get_pfnblock_migratetype(page, pfn);
843 __free_one_page(page, pfn, zone, order, migratetype);
844 spin_unlock(&zone->lock);
847 static int free_tail_pages_check(struct page *head_page, struct page *page)
849 if (!IS_ENABLED(CONFIG_DEBUG_VM))
851 if (unlikely(!PageTail(page))) {
852 bad_page(page, "PageTail not set", 0);
855 if (unlikely(page->first_page != head_page)) {
856 bad_page(page, "first_page not consistent", 0);
862 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
863 unsigned long zone, int nid)
865 set_page_links(page, zone, nid, pfn);
866 init_page_count(page);
867 page_mapcount_reset(page);
868 page_cpupid_reset_last(page);
870 INIT_LIST_HEAD(&page->lru);
871 #ifdef WANT_PAGE_VIRTUAL
872 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
873 if (!is_highmem_idx(zone))
874 set_page_address(page, __va(pfn << PAGE_SHIFT));
878 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
881 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
884 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
885 static void init_reserved_page(unsigned long pfn)
890 if (!early_page_uninitialised(pfn))
893 nid = early_pfn_to_nid(pfn);
894 pgdat = NODE_DATA(nid);
896 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
897 struct zone *zone = &pgdat->node_zones[zid];
899 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
902 __init_single_pfn(pfn, zid, nid);
905 static inline void init_reserved_page(unsigned long pfn)
908 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
911 * Initialised pages do not have PageReserved set. This function is
912 * called for each range allocated by the bootmem allocator and
913 * marks the pages PageReserved. The remaining valid pages are later
914 * sent to the buddy page allocator.
916 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
918 unsigned long start_pfn = PFN_DOWN(start);
919 unsigned long end_pfn = PFN_UP(end);
921 for (; start_pfn < end_pfn; start_pfn++) {
922 if (pfn_valid(start_pfn)) {
923 struct page *page = pfn_to_page(start_pfn);
925 init_reserved_page(start_pfn);
926 SetPageReserved(page);
931 static bool free_pages_prepare(struct page *page, unsigned int order)
933 bool compound = PageCompound(page);
936 VM_BUG_ON_PAGE(PageTail(page), page);
937 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
939 trace_mm_page_free(page, order);
940 kmemcheck_free_shadow(page, order);
941 kasan_free_pages(page, order);
944 page->mapping = NULL;
945 bad += free_pages_check(page);
946 for (i = 1; i < (1 << order); i++) {
948 bad += free_tail_pages_check(page, page + i);
949 bad += free_pages_check(page + i);
954 reset_page_owner(page, order);
956 if (!PageHighMem(page)) {
957 debug_check_no_locks_freed(page_address(page),
959 debug_check_no_obj_freed(page_address(page),
962 arch_free_page(page, order);
963 kernel_map_pages(page, 1 << order, 0);
968 static void __free_pages_ok(struct page *page, unsigned int order)
972 unsigned long pfn = page_to_pfn(page);
974 if (!free_pages_prepare(page, order))
977 migratetype = get_pfnblock_migratetype(page, pfn);
978 local_irq_save(flags);
979 __count_vm_events(PGFREE, 1 << order);
980 free_one_page(page_zone(page), page, pfn, order, migratetype);
981 local_irq_restore(flags);
984 static void __init __free_pages_boot_core(struct page *page,
985 unsigned long pfn, unsigned int order)
987 unsigned int nr_pages = 1 << order;
988 struct page *p = page;
992 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
994 __ClearPageReserved(p);
995 set_page_count(p, 0);
997 __ClearPageReserved(p);
998 set_page_count(p, 0);
1000 page_zone(page)->managed_pages += nr_pages;
1001 set_page_refcounted(page);
1002 __free_pages(page, order);
1005 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1006 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1008 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1010 int __meminit early_pfn_to_nid(unsigned long pfn)
1012 static DEFINE_SPINLOCK(early_pfn_lock);
1015 spin_lock(&early_pfn_lock);
1016 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1019 spin_unlock(&early_pfn_lock);
1025 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1026 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1027 struct mminit_pfnnid_cache *state)
1031 nid = __early_pfn_to_nid(pfn, state);
1032 if (nid >= 0 && nid != node)
1037 /* Only safe to use early in boot when initialisation is single-threaded */
1038 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1040 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1045 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1049 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1050 struct mminit_pfnnid_cache *state)
1057 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1060 if (early_page_uninitialised(pfn))
1062 return __free_pages_boot_core(page, pfn, order);
1065 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1066 static void __init deferred_free_range(struct page *page,
1067 unsigned long pfn, int nr_pages)
1074 /* Free a large naturally-aligned chunk if possible */
1075 if (nr_pages == MAX_ORDER_NR_PAGES &&
1076 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1077 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1078 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1082 for (i = 0; i < nr_pages; i++, page++, pfn++)
1083 __free_pages_boot_core(page, pfn, 0);
1086 /* Completion tracking for deferred_init_memmap() threads */
1087 static atomic_t pgdat_init_n_undone __initdata;
1088 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1090 static inline void __init pgdat_init_report_one_done(void)
1092 if (atomic_dec_and_test(&pgdat_init_n_undone))
1093 complete(&pgdat_init_all_done_comp);
1096 /* Initialise remaining memory on a node */
1097 static int __init deferred_init_memmap(void *data)
1099 pg_data_t *pgdat = data;
1100 int nid = pgdat->node_id;
1101 struct mminit_pfnnid_cache nid_init_state = { };
1102 unsigned long start = jiffies;
1103 unsigned long nr_pages = 0;
1104 unsigned long walk_start, walk_end;
1107 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1108 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1110 if (first_init_pfn == ULONG_MAX) {
1111 pgdat_init_report_one_done();
1115 /* Bind memory initialisation thread to a local node if possible */
1116 if (!cpumask_empty(cpumask))
1117 set_cpus_allowed_ptr(current, cpumask);
1119 /* Sanity check boundaries */
1120 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1121 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1122 pgdat->first_deferred_pfn = ULONG_MAX;
1124 /* Only the highest zone is deferred so find it */
1125 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1126 zone = pgdat->node_zones + zid;
1127 if (first_init_pfn < zone_end_pfn(zone))
1131 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1132 unsigned long pfn, end_pfn;
1133 struct page *page = NULL;
1134 struct page *free_base_page = NULL;
1135 unsigned long free_base_pfn = 0;
1138 end_pfn = min(walk_end, zone_end_pfn(zone));
1139 pfn = first_init_pfn;
1140 if (pfn < walk_start)
1142 if (pfn < zone->zone_start_pfn)
1143 pfn = zone->zone_start_pfn;
1145 for (; pfn < end_pfn; pfn++) {
1146 if (!pfn_valid_within(pfn))
1150 * Ensure pfn_valid is checked every
1151 * MAX_ORDER_NR_PAGES for memory holes
1153 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1154 if (!pfn_valid(pfn)) {
1160 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1165 /* Minimise pfn page lookups and scheduler checks */
1166 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1169 nr_pages += nr_to_free;
1170 deferred_free_range(free_base_page,
1171 free_base_pfn, nr_to_free);
1172 free_base_page = NULL;
1173 free_base_pfn = nr_to_free = 0;
1175 page = pfn_to_page(pfn);
1180 VM_BUG_ON(page_zone(page) != zone);
1184 __init_single_page(page, pfn, zid, nid);
1185 if (!free_base_page) {
1186 free_base_page = page;
1187 free_base_pfn = pfn;
1192 /* Where possible, batch up pages for a single free */
1195 /* Free the current block of pages to allocator */
1196 nr_pages += nr_to_free;
1197 deferred_free_range(free_base_page, free_base_pfn,
1199 free_base_page = NULL;
1200 free_base_pfn = nr_to_free = 0;
1203 first_init_pfn = max(end_pfn, first_init_pfn);
1206 /* Sanity check that the next zone really is unpopulated */
1207 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1209 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1210 jiffies_to_msecs(jiffies - start));
1212 pgdat_init_report_one_done();
1216 void __init page_alloc_init_late(void)
1220 /* There will be num_node_state(N_MEMORY) threads */
1221 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1222 for_each_node_state(nid, N_MEMORY) {
1223 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1226 /* Block until all are initialised */
1227 wait_for_completion(&pgdat_init_all_done_comp);
1229 /* Reinit limits that are based on free pages after the kernel is up */
1230 files_maxfiles_init();
1232 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1235 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1236 void __init init_cma_reserved_pageblock(struct page *page)
1238 unsigned i = pageblock_nr_pages;
1239 struct page *p = page;
1242 __ClearPageReserved(p);
1243 set_page_count(p, 0);
1246 set_pageblock_migratetype(page, MIGRATE_CMA);
1248 if (pageblock_order >= MAX_ORDER) {
1249 i = pageblock_nr_pages;
1252 set_page_refcounted(p);
1253 __free_pages(p, MAX_ORDER - 1);
1254 p += MAX_ORDER_NR_PAGES;
1255 } while (i -= MAX_ORDER_NR_PAGES);
1257 set_page_refcounted(page);
1258 __free_pages(page, pageblock_order);
1261 adjust_managed_page_count(page, pageblock_nr_pages);
1266 * The order of subdivision here is critical for the IO subsystem.
1267 * Please do not alter this order without good reasons and regression
1268 * testing. Specifically, as large blocks of memory are subdivided,
1269 * the order in which smaller blocks are delivered depends on the order
1270 * they're subdivided in this function. This is the primary factor
1271 * influencing the order in which pages are delivered to the IO
1272 * subsystem according to empirical testing, and this is also justified
1273 * by considering the behavior of a buddy system containing a single
1274 * large block of memory acted on by a series of small allocations.
1275 * This behavior is a critical factor in sglist merging's success.
1279 static inline void expand(struct zone *zone, struct page *page,
1280 int low, int high, struct free_area *area,
1283 unsigned long size = 1 << high;
1285 while (high > low) {
1289 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1291 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1292 debug_guardpage_enabled() &&
1293 high < debug_guardpage_minorder()) {
1295 * Mark as guard pages (or page), that will allow to
1296 * merge back to allocator when buddy will be freed.
1297 * Corresponding page table entries will not be touched,
1298 * pages will stay not present in virtual address space
1300 set_page_guard(zone, &page[size], high, migratetype);
1303 list_add(&page[size].lru, &area->free_list[migratetype]);
1305 set_page_order(&page[size], high);
1310 * This page is about to be returned from the page allocator
1312 static inline int check_new_page(struct page *page)
1314 const char *bad_reason = NULL;
1315 unsigned long bad_flags = 0;
1317 if (unlikely(page_mapcount(page)))
1318 bad_reason = "nonzero mapcount";
1319 if (unlikely(page->mapping != NULL))
1320 bad_reason = "non-NULL mapping";
1321 if (unlikely(atomic_read(&page->_count) != 0))
1322 bad_reason = "nonzero _count";
1323 if (unlikely(page->flags & __PG_HWPOISON)) {
1324 bad_reason = "HWPoisoned (hardware-corrupted)";
1325 bad_flags = __PG_HWPOISON;
1327 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1328 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1329 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1332 if (unlikely(page->mem_cgroup))
1333 bad_reason = "page still charged to cgroup";
1335 if (unlikely(bad_reason)) {
1336 bad_page(page, bad_reason, bad_flags);
1342 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1347 for (i = 0; i < (1 << order); i++) {
1348 struct page *p = page + i;
1349 if (unlikely(check_new_page(p)))
1353 set_page_private(page, 0);
1354 set_page_refcounted(page);
1356 arch_alloc_page(page, order);
1357 kernel_map_pages(page, 1 << order, 1);
1358 kasan_alloc_pages(page, order);
1360 if (gfp_flags & __GFP_ZERO)
1361 for (i = 0; i < (1 << order); i++)
1362 clear_highpage(page + i);
1364 if (order && (gfp_flags & __GFP_COMP))
1365 prep_compound_page(page, order);
1367 set_page_owner(page, order, gfp_flags);
1370 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1371 * allocate the page. The expectation is that the caller is taking
1372 * steps that will free more memory. The caller should avoid the page
1373 * being used for !PFMEMALLOC purposes.
1375 if (alloc_flags & ALLOC_NO_WATERMARKS)
1376 set_page_pfmemalloc(page);
1378 clear_page_pfmemalloc(page);
1384 * Go through the free lists for the given migratetype and remove
1385 * the smallest available page from the freelists
1388 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1391 unsigned int current_order;
1392 struct free_area *area;
1395 /* Find a page of the appropriate size in the preferred list */
1396 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1397 area = &(zone->free_area[current_order]);
1398 if (list_empty(&area->free_list[migratetype]))
1401 page = list_entry(area->free_list[migratetype].next,
1403 list_del(&page->lru);
1404 rmv_page_order(page);
1406 expand(zone, page, order, current_order, area, migratetype);
1407 set_pcppage_migratetype(page, migratetype);
1416 * This array describes the order lists are fallen back to when
1417 * the free lists for the desirable migrate type are depleted
1419 static int fallbacks[MIGRATE_TYPES][4] = {
1420 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1421 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1422 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1424 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1426 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1427 #ifdef CONFIG_MEMORY_ISOLATION
1428 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1433 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1436 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1439 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1440 unsigned int order) { return NULL; }
1444 * Move the free pages in a range to the free lists of the requested type.
1445 * Note that start_page and end_pages are not aligned on a pageblock
1446 * boundary. If alignment is required, use move_freepages_block()
1448 int move_freepages(struct zone *zone,
1449 struct page *start_page, struct page *end_page,
1453 unsigned long order;
1454 int pages_moved = 0;
1456 #ifndef CONFIG_HOLES_IN_ZONE
1458 * page_zone is not safe to call in this context when
1459 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1460 * anyway as we check zone boundaries in move_freepages_block().
1461 * Remove at a later date when no bug reports exist related to
1462 * grouping pages by mobility
1464 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1467 for (page = start_page; page <= end_page;) {
1468 /* Make sure we are not inadvertently changing nodes */
1469 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1471 if (!pfn_valid_within(page_to_pfn(page))) {
1476 if (!PageBuddy(page)) {
1481 order = page_order(page);
1482 list_move(&page->lru,
1483 &zone->free_area[order].free_list[migratetype]);
1485 pages_moved += 1 << order;
1491 int move_freepages_block(struct zone *zone, struct page *page,
1494 unsigned long start_pfn, end_pfn;
1495 struct page *start_page, *end_page;
1497 start_pfn = page_to_pfn(page);
1498 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1499 start_page = pfn_to_page(start_pfn);
1500 end_page = start_page + pageblock_nr_pages - 1;
1501 end_pfn = start_pfn + pageblock_nr_pages - 1;
1503 /* Do not cross zone boundaries */
1504 if (!zone_spans_pfn(zone, start_pfn))
1506 if (!zone_spans_pfn(zone, end_pfn))
1509 return move_freepages(zone, start_page, end_page, migratetype);
1512 static void change_pageblock_range(struct page *pageblock_page,
1513 int start_order, int migratetype)
1515 int nr_pageblocks = 1 << (start_order - pageblock_order);
1517 while (nr_pageblocks--) {
1518 set_pageblock_migratetype(pageblock_page, migratetype);
1519 pageblock_page += pageblock_nr_pages;
1524 * When we are falling back to another migratetype during allocation, try to
1525 * steal extra free pages from the same pageblocks to satisfy further
1526 * allocations, instead of polluting multiple pageblocks.
1528 * If we are stealing a relatively large buddy page, it is likely there will
1529 * be more free pages in the pageblock, so try to steal them all. For
1530 * reclaimable and unmovable allocations, we steal regardless of page size,
1531 * as fragmentation caused by those allocations polluting movable pageblocks
1532 * is worse than movable allocations stealing from unmovable and reclaimable
1535 static bool can_steal_fallback(unsigned int order, int start_mt)
1538 * Leaving this order check is intended, although there is
1539 * relaxed order check in next check. The reason is that
1540 * we can actually steal whole pageblock if this condition met,
1541 * but, below check doesn't guarantee it and that is just heuristic
1542 * so could be changed anytime.
1544 if (order >= pageblock_order)
1547 if (order >= pageblock_order / 2 ||
1548 start_mt == MIGRATE_RECLAIMABLE ||
1549 start_mt == MIGRATE_UNMOVABLE ||
1550 page_group_by_mobility_disabled)
1557 * This function implements actual steal behaviour. If order is large enough,
1558 * we can steal whole pageblock. If not, we first move freepages in this
1559 * pageblock and check whether half of pages are moved or not. If half of
1560 * pages are moved, we can change migratetype of pageblock and permanently
1561 * use it's pages as requested migratetype in the future.
1563 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1566 int current_order = page_order(page);
1569 /* Take ownership for orders >= pageblock_order */
1570 if (current_order >= pageblock_order) {
1571 change_pageblock_range(page, current_order, start_type);
1575 pages = move_freepages_block(zone, page, start_type);
1577 /* Claim the whole block if over half of it is free */
1578 if (pages >= (1 << (pageblock_order-1)) ||
1579 page_group_by_mobility_disabled)
1580 set_pageblock_migratetype(page, start_type);
1584 * Check whether there is a suitable fallback freepage with requested order.
1585 * If only_stealable is true, this function returns fallback_mt only if
1586 * we can steal other freepages all together. This would help to reduce
1587 * fragmentation due to mixed migratetype pages in one pageblock.
1589 int find_suitable_fallback(struct free_area *area, unsigned int order,
1590 int migratetype, bool only_stealable, bool *can_steal)
1595 if (area->nr_free == 0)
1600 fallback_mt = fallbacks[migratetype][i];
1601 if (fallback_mt == MIGRATE_RESERVE)
1604 if (list_empty(&area->free_list[fallback_mt]))
1607 if (can_steal_fallback(order, migratetype))
1610 if (!only_stealable)
1620 /* Remove an element from the buddy allocator from the fallback list */
1621 static inline struct page *
1622 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1624 struct free_area *area;
1625 unsigned int current_order;
1630 /* Find the largest possible block of pages in the other list */
1631 for (current_order = MAX_ORDER-1;
1632 current_order >= order && current_order <= MAX_ORDER-1;
1634 area = &(zone->free_area[current_order]);
1635 fallback_mt = find_suitable_fallback(area, current_order,
1636 start_migratetype, false, &can_steal);
1637 if (fallback_mt == -1)
1640 page = list_entry(area->free_list[fallback_mt].next,
1643 steal_suitable_fallback(zone, page, start_migratetype);
1645 /* Remove the page from the freelists */
1647 list_del(&page->lru);
1648 rmv_page_order(page);
1650 expand(zone, page, order, current_order, area,
1653 * The pcppage_migratetype may differ from pageblock's
1654 * migratetype depending on the decisions in
1655 * find_suitable_fallback(). This is OK as long as it does not
1656 * differ for MIGRATE_CMA pageblocks. Those can be used as
1657 * fallback only via special __rmqueue_cma_fallback() function
1659 set_pcppage_migratetype(page, start_migratetype);
1661 trace_mm_page_alloc_extfrag(page, order, current_order,
1662 start_migratetype, fallback_mt);
1671 * Do the hard work of removing an element from the buddy allocator.
1672 * Call me with the zone->lock already held.
1674 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1680 page = __rmqueue_smallest(zone, order, migratetype);
1682 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1683 if (migratetype == MIGRATE_MOVABLE)
1684 page = __rmqueue_cma_fallback(zone, order);
1687 page = __rmqueue_fallback(zone, order, migratetype);
1690 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1691 * is used because __rmqueue_smallest is an inline function
1692 * and we want just one call site
1695 migratetype = MIGRATE_RESERVE;
1700 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1705 * Obtain a specified number of elements from the buddy allocator, all under
1706 * a single hold of the lock, for efficiency. Add them to the supplied list.
1707 * Returns the number of new pages which were placed at *list.
1709 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1710 unsigned long count, struct list_head *list,
1711 int migratetype, bool cold)
1715 spin_lock(&zone->lock);
1716 for (i = 0; i < count; ++i) {
1717 struct page *page = __rmqueue(zone, order, migratetype);
1718 if (unlikely(page == NULL))
1722 * Split buddy pages returned by expand() are received here
1723 * in physical page order. The page is added to the callers and
1724 * list and the list head then moves forward. From the callers
1725 * perspective, the linked list is ordered by page number in
1726 * some conditions. This is useful for IO devices that can
1727 * merge IO requests if the physical pages are ordered
1731 list_add(&page->lru, list);
1733 list_add_tail(&page->lru, list);
1735 if (is_migrate_cma(get_pcppage_migratetype(page)))
1736 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1739 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1740 spin_unlock(&zone->lock);
1746 * Called from the vmstat counter updater to drain pagesets of this
1747 * currently executing processor on remote nodes after they have
1750 * Note that this function must be called with the thread pinned to
1751 * a single processor.
1753 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1755 unsigned long flags;
1756 int to_drain, batch;
1758 local_irq_save(flags);
1759 batch = READ_ONCE(pcp->batch);
1760 to_drain = min(pcp->count, batch);
1762 free_pcppages_bulk(zone, to_drain, pcp);
1763 pcp->count -= to_drain;
1765 local_irq_restore(flags);
1770 * Drain pcplists of the indicated processor and zone.
1772 * The processor must either be the current processor and the
1773 * thread pinned to the current processor or a processor that
1776 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1778 unsigned long flags;
1779 struct per_cpu_pageset *pset;
1780 struct per_cpu_pages *pcp;
1782 local_irq_save(flags);
1783 pset = per_cpu_ptr(zone->pageset, cpu);
1787 free_pcppages_bulk(zone, pcp->count, pcp);
1790 local_irq_restore(flags);
1794 * Drain pcplists of all zones on the indicated processor.
1796 * The processor must either be the current processor and the
1797 * thread pinned to the current processor or a processor that
1800 static void drain_pages(unsigned int cpu)
1804 for_each_populated_zone(zone) {
1805 drain_pages_zone(cpu, zone);
1810 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1812 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1813 * the single zone's pages.
1815 void drain_local_pages(struct zone *zone)
1817 int cpu = smp_processor_id();
1820 drain_pages_zone(cpu, zone);
1826 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1828 * When zone parameter is non-NULL, spill just the single zone's pages.
1830 * Note that this code is protected against sending an IPI to an offline
1831 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1832 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1833 * nothing keeps CPUs from showing up after we populated the cpumask and
1834 * before the call to on_each_cpu_mask().
1836 void drain_all_pages(struct zone *zone)
1841 * Allocate in the BSS so we wont require allocation in
1842 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1844 static cpumask_t cpus_with_pcps;
1847 * We don't care about racing with CPU hotplug event
1848 * as offline notification will cause the notified
1849 * cpu to drain that CPU pcps and on_each_cpu_mask
1850 * disables preemption as part of its processing
1852 for_each_online_cpu(cpu) {
1853 struct per_cpu_pageset *pcp;
1855 bool has_pcps = false;
1858 pcp = per_cpu_ptr(zone->pageset, cpu);
1862 for_each_populated_zone(z) {
1863 pcp = per_cpu_ptr(z->pageset, cpu);
1864 if (pcp->pcp.count) {
1872 cpumask_set_cpu(cpu, &cpus_with_pcps);
1874 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1876 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1880 #ifdef CONFIG_HIBERNATION
1882 void mark_free_pages(struct zone *zone)
1884 unsigned long pfn, max_zone_pfn;
1885 unsigned long flags;
1886 unsigned int order, t;
1887 struct list_head *curr;
1889 if (zone_is_empty(zone))
1892 spin_lock_irqsave(&zone->lock, flags);
1894 max_zone_pfn = zone_end_pfn(zone);
1895 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1896 if (pfn_valid(pfn)) {
1897 struct page *page = pfn_to_page(pfn);
1899 if (!swsusp_page_is_forbidden(page))
1900 swsusp_unset_page_free(page);
1903 for_each_migratetype_order(order, t) {
1904 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1907 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1908 for (i = 0; i < (1UL << order); i++)
1909 swsusp_set_page_free(pfn_to_page(pfn + i));
1912 spin_unlock_irqrestore(&zone->lock, flags);
1914 #endif /* CONFIG_PM */
1917 * Free a 0-order page
1918 * cold == true ? free a cold page : free a hot page
1920 void free_hot_cold_page(struct page *page, bool cold)
1922 struct zone *zone = page_zone(page);
1923 struct per_cpu_pages *pcp;
1924 unsigned long flags;
1925 unsigned long pfn = page_to_pfn(page);
1928 if (!free_pages_prepare(page, 0))
1931 migratetype = get_pfnblock_migratetype(page, pfn);
1932 set_pcppage_migratetype(page, migratetype);
1933 local_irq_save(flags);
1934 __count_vm_event(PGFREE);
1937 * We only track unmovable, reclaimable and movable on pcp lists.
1938 * Free ISOLATE pages back to the allocator because they are being
1939 * offlined but treat RESERVE as movable pages so we can get those
1940 * areas back if necessary. Otherwise, we may have to free
1941 * excessively into the page allocator
1943 if (migratetype >= MIGRATE_PCPTYPES) {
1944 if (unlikely(is_migrate_isolate(migratetype))) {
1945 free_one_page(zone, page, pfn, 0, migratetype);
1948 migratetype = MIGRATE_MOVABLE;
1951 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1953 list_add(&page->lru, &pcp->lists[migratetype]);
1955 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1957 if (pcp->count >= pcp->high) {
1958 unsigned long batch = READ_ONCE(pcp->batch);
1959 free_pcppages_bulk(zone, batch, pcp);
1960 pcp->count -= batch;
1964 local_irq_restore(flags);
1968 * Free a list of 0-order pages
1970 void free_hot_cold_page_list(struct list_head *list, bool cold)
1972 struct page *page, *next;
1974 list_for_each_entry_safe(page, next, list, lru) {
1975 trace_mm_page_free_batched(page, cold);
1976 free_hot_cold_page(page, cold);
1981 * split_page takes a non-compound higher-order page, and splits it into
1982 * n (1<<order) sub-pages: page[0..n]
1983 * Each sub-page must be freed individually.
1985 * Note: this is probably too low level an operation for use in drivers.
1986 * Please consult with lkml before using this in your driver.
1988 void split_page(struct page *page, unsigned int order)
1993 VM_BUG_ON_PAGE(PageCompound(page), page);
1994 VM_BUG_ON_PAGE(!page_count(page), page);
1996 #ifdef CONFIG_KMEMCHECK
1998 * Split shadow pages too, because free(page[0]) would
1999 * otherwise free the whole shadow.
2001 if (kmemcheck_page_is_tracked(page))
2002 split_page(virt_to_page(page[0].shadow), order);
2005 gfp_mask = get_page_owner_gfp(page);
2006 set_page_owner(page, 0, gfp_mask);
2007 for (i = 1; i < (1 << order); i++) {
2008 set_page_refcounted(page + i);
2009 set_page_owner(page + i, 0, gfp_mask);
2012 EXPORT_SYMBOL_GPL(split_page);
2014 int __isolate_free_page(struct page *page, unsigned int order)
2016 unsigned long watermark;
2020 BUG_ON(!PageBuddy(page));
2022 zone = page_zone(page);
2023 mt = get_pageblock_migratetype(page);
2025 if (!is_migrate_isolate(mt)) {
2026 /* Obey watermarks as if the page was being allocated */
2027 watermark = low_wmark_pages(zone) + (1 << order);
2028 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2031 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2034 /* Remove page from free list */
2035 list_del(&page->lru);
2036 zone->free_area[order].nr_free--;
2037 rmv_page_order(page);
2039 set_page_owner(page, order, __GFP_MOVABLE);
2041 /* Set the pageblock if the isolated page is at least a pageblock */
2042 if (order >= pageblock_order - 1) {
2043 struct page *endpage = page + (1 << order) - 1;
2044 for (; page < endpage; page += pageblock_nr_pages) {
2045 int mt = get_pageblock_migratetype(page);
2046 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2047 set_pageblock_migratetype(page,
2053 return 1UL << order;
2057 * Similar to split_page except the page is already free. As this is only
2058 * being used for migration, the migratetype of the block also changes.
2059 * As this is called with interrupts disabled, the caller is responsible
2060 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2063 * Note: this is probably too low level an operation for use in drivers.
2064 * Please consult with lkml before using this in your driver.
2066 int split_free_page(struct page *page)
2071 order = page_order(page);
2073 nr_pages = __isolate_free_page(page, order);
2077 /* Split into individual pages */
2078 set_page_refcounted(page);
2079 split_page(page, order);
2084 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2087 struct page *buffered_rmqueue(struct zone *preferred_zone,
2088 struct zone *zone, unsigned int order,
2089 gfp_t gfp_flags, int migratetype)
2091 unsigned long flags;
2093 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2095 if (likely(order == 0)) {
2096 struct per_cpu_pages *pcp;
2097 struct list_head *list;
2099 local_irq_save(flags);
2100 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2101 list = &pcp->lists[migratetype];
2102 if (list_empty(list)) {
2103 pcp->count += rmqueue_bulk(zone, 0,
2106 if (unlikely(list_empty(list)))
2111 page = list_entry(list->prev, struct page, lru);
2113 page = list_entry(list->next, struct page, lru);
2115 list_del(&page->lru);
2118 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2120 * __GFP_NOFAIL is not to be used in new code.
2122 * All __GFP_NOFAIL callers should be fixed so that they
2123 * properly detect and handle allocation failures.
2125 * We most definitely don't want callers attempting to
2126 * allocate greater than order-1 page units with
2129 WARN_ON_ONCE(order > 1);
2131 spin_lock_irqsave(&zone->lock, flags);
2132 page = __rmqueue(zone, order, migratetype);
2133 spin_unlock(&zone->lock);
2136 __mod_zone_freepage_state(zone, -(1 << order),
2137 get_pcppage_migratetype(page));
2140 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2141 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2142 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2143 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2145 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2146 zone_statistics(preferred_zone, zone, gfp_flags);
2147 local_irq_restore(flags);
2149 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2153 local_irq_restore(flags);
2157 #ifdef CONFIG_FAIL_PAGE_ALLOC
2160 struct fault_attr attr;
2162 bool ignore_gfp_highmem;
2163 bool ignore_gfp_reclaim;
2165 } fail_page_alloc = {
2166 .attr = FAULT_ATTR_INITIALIZER,
2167 .ignore_gfp_reclaim = true,
2168 .ignore_gfp_highmem = true,
2172 static int __init setup_fail_page_alloc(char *str)
2174 return setup_fault_attr(&fail_page_alloc.attr, str);
2176 __setup("fail_page_alloc=", setup_fail_page_alloc);
2178 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2180 if (order < fail_page_alloc.min_order)
2182 if (gfp_mask & __GFP_NOFAIL)
2184 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2186 if (fail_page_alloc.ignore_gfp_reclaim &&
2187 (gfp_mask & __GFP_DIRECT_RECLAIM))
2190 return should_fail(&fail_page_alloc.attr, 1 << order);
2193 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2195 static int __init fail_page_alloc_debugfs(void)
2197 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2200 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2201 &fail_page_alloc.attr);
2203 return PTR_ERR(dir);
2205 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2206 &fail_page_alloc.ignore_gfp_reclaim))
2208 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2209 &fail_page_alloc.ignore_gfp_highmem))
2211 if (!debugfs_create_u32("min-order", mode, dir,
2212 &fail_page_alloc.min_order))
2217 debugfs_remove_recursive(dir);
2222 late_initcall(fail_page_alloc_debugfs);
2224 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2226 #else /* CONFIG_FAIL_PAGE_ALLOC */
2228 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2233 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2236 * Return true if free pages are above 'mark'. This takes into account the order
2237 * of the allocation.
2239 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2240 unsigned long mark, int classzone_idx, int alloc_flags,
2243 /* free_pages may go negative - that's OK */
2248 free_pages -= (1 << order) - 1;
2249 if (alloc_flags & ALLOC_HIGH)
2251 if (alloc_flags & ALLOC_HARDER)
2255 /* If allocation can't use CMA areas don't use free CMA pages */
2256 if (!(alloc_flags & ALLOC_CMA))
2257 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
2260 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
2262 for (o = 0; o < order; o++) {
2263 /* At the next order, this order's pages become unavailable */
2264 free_pages -= z->free_area[o].nr_free << o;
2266 /* Require fewer higher order pages to be free */
2269 if (free_pages <= min)
2275 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2276 int classzone_idx, int alloc_flags)
2278 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2279 zone_page_state(z, NR_FREE_PAGES));
2282 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2283 unsigned long mark, int classzone_idx)
2285 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2287 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2288 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2290 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2296 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2297 * skip over zones that are not allowed by the cpuset, or that have
2298 * been recently (in last second) found to be nearly full. See further
2299 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2300 * that have to skip over a lot of full or unallowed zones.
2302 * If the zonelist cache is present in the passed zonelist, then
2303 * returns a pointer to the allowed node mask (either the current
2304 * tasks mems_allowed, or node_states[N_MEMORY].)
2306 * If the zonelist cache is not available for this zonelist, does
2307 * nothing and returns NULL.
2309 * If the fullzones BITMAP in the zonelist cache is stale (more than
2310 * a second since last zap'd) then we zap it out (clear its bits.)
2312 * We hold off even calling zlc_setup, until after we've checked the
2313 * first zone in the zonelist, on the theory that most allocations will
2314 * be satisfied from that first zone, so best to examine that zone as
2315 * quickly as we can.
2317 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2319 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2320 nodemask_t *allowednodes; /* zonelist_cache approximation */
2322 zlc = zonelist->zlcache_ptr;
2326 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
2327 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2328 zlc->last_full_zap = jiffies;
2331 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2332 &cpuset_current_mems_allowed :
2333 &node_states[N_MEMORY];
2334 return allowednodes;
2338 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2339 * if it is worth looking at further for free memory:
2340 * 1) Check that the zone isn't thought to be full (doesn't have its
2341 * bit set in the zonelist_cache fullzones BITMAP).
2342 * 2) Check that the zones node (obtained from the zonelist_cache
2343 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2344 * Return true (non-zero) if zone is worth looking at further, or
2345 * else return false (zero) if it is not.
2347 * This check -ignores- the distinction between various watermarks,
2348 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2349 * found to be full for any variation of these watermarks, it will
2350 * be considered full for up to one second by all requests, unless
2351 * we are so low on memory on all allowed nodes that we are forced
2352 * into the second scan of the zonelist.
2354 * In the second scan we ignore this zonelist cache and exactly
2355 * apply the watermarks to all zones, even it is slower to do so.
2356 * We are low on memory in the second scan, and should leave no stone
2357 * unturned looking for a free page.
2359 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2360 nodemask_t *allowednodes)
2362 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2363 int i; /* index of *z in zonelist zones */
2364 int n; /* node that zone *z is on */
2366 zlc = zonelist->zlcache_ptr;
2370 i = z - zonelist->_zonerefs;
2373 /* This zone is worth trying if it is allowed but not full */
2374 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2378 * Given 'z' scanning a zonelist, set the corresponding bit in
2379 * zlc->fullzones, so that subsequent attempts to allocate a page
2380 * from that zone don't waste time re-examining it.
2382 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2384 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2385 int i; /* index of *z in zonelist zones */
2387 zlc = zonelist->zlcache_ptr;
2391 i = z - zonelist->_zonerefs;
2393 set_bit(i, zlc->fullzones);
2397 * clear all zones full, called after direct reclaim makes progress so that
2398 * a zone that was recently full is not skipped over for up to a second
2400 static void zlc_clear_zones_full(struct zonelist *zonelist)
2402 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2404 zlc = zonelist->zlcache_ptr;
2408 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2411 static bool zone_local(struct zone *local_zone, struct zone *zone)
2413 return local_zone->node == zone->node;
2416 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2418 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2422 #else /* CONFIG_NUMA */
2424 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2429 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2430 nodemask_t *allowednodes)
2435 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2439 static void zlc_clear_zones_full(struct zonelist *zonelist)
2443 static bool zone_local(struct zone *local_zone, struct zone *zone)
2448 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2453 #endif /* CONFIG_NUMA */
2455 static void reset_alloc_batches(struct zone *preferred_zone)
2457 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2460 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2461 high_wmark_pages(zone) - low_wmark_pages(zone) -
2462 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2463 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2464 } while (zone++ != preferred_zone);
2468 * get_page_from_freelist goes through the zonelist trying to allocate
2471 static struct page *
2472 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2473 const struct alloc_context *ac)
2475 struct zonelist *zonelist = ac->zonelist;
2477 struct page *page = NULL;
2479 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2480 int zlc_active = 0; /* set if using zonelist_cache */
2481 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2482 int nr_fair_skipped = 0;
2483 bool zonelist_rescan;
2486 zonelist_rescan = false;
2489 * Scan zonelist, looking for a zone with enough free.
2490 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2492 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2496 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2497 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2499 if (cpusets_enabled() &&
2500 (alloc_flags & ALLOC_CPUSET) &&
2501 !cpuset_zone_allowed(zone, gfp_mask))
2504 * Distribute pages in proportion to the individual
2505 * zone size to ensure fair page aging. The zone a
2506 * page was allocated in should have no effect on the
2507 * time the page has in memory before being reclaimed.
2509 if (alloc_flags & ALLOC_FAIR) {
2510 if (!zone_local(ac->preferred_zone, zone))
2512 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2518 * When allocating a page cache page for writing, we
2519 * want to get it from a zone that is within its dirty
2520 * limit, such that no single zone holds more than its
2521 * proportional share of globally allowed dirty pages.
2522 * The dirty limits take into account the zone's
2523 * lowmem reserves and high watermark so that kswapd
2524 * should be able to balance it without having to
2525 * write pages from its LRU list.
2527 * This may look like it could increase pressure on
2528 * lower zones by failing allocations in higher zones
2529 * before they are full. But the pages that do spill
2530 * over are limited as the lower zones are protected
2531 * by this very same mechanism. It should not become
2532 * a practical burden to them.
2534 * XXX: For now, allow allocations to potentially
2535 * exceed the per-zone dirty limit in the slowpath
2536 * (spread_dirty_pages unset) before going into reclaim,
2537 * which is important when on a NUMA setup the allowed
2538 * zones are together not big enough to reach the
2539 * global limit. The proper fix for these situations
2540 * will require awareness of zones in the
2541 * dirty-throttling and the flusher threads.
2543 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2546 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2547 if (!zone_watermark_ok(zone, order, mark,
2548 ac->classzone_idx, alloc_flags)) {
2551 /* Checked here to keep the fast path fast */
2552 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2553 if (alloc_flags & ALLOC_NO_WATERMARKS)
2556 if (IS_ENABLED(CONFIG_NUMA) &&
2557 !did_zlc_setup && nr_online_nodes > 1) {
2559 * we do zlc_setup if there are multiple nodes
2560 * and before considering the first zone allowed
2563 allowednodes = zlc_setup(zonelist, alloc_flags);
2568 if (zone_reclaim_mode == 0 ||
2569 !zone_allows_reclaim(ac->preferred_zone, zone))
2570 goto this_zone_full;
2573 * As we may have just activated ZLC, check if the first
2574 * eligible zone has failed zone_reclaim recently.
2576 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2577 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2580 ret = zone_reclaim(zone, gfp_mask, order);
2582 case ZONE_RECLAIM_NOSCAN:
2585 case ZONE_RECLAIM_FULL:
2586 /* scanned but unreclaimable */
2589 /* did we reclaim enough */
2590 if (zone_watermark_ok(zone, order, mark,
2591 ac->classzone_idx, alloc_flags))
2595 * Failed to reclaim enough to meet watermark.
2596 * Only mark the zone full if checking the min
2597 * watermark or if we failed to reclaim just
2598 * 1<<order pages or else the page allocator
2599 * fastpath will prematurely mark zones full
2600 * when the watermark is between the low and
2603 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2604 ret == ZONE_RECLAIM_SOME)
2605 goto this_zone_full;
2612 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2613 gfp_mask, ac->migratetype);
2615 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2620 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2621 zlc_mark_zone_full(zonelist, z);
2625 * The first pass makes sure allocations are spread fairly within the
2626 * local node. However, the local node might have free pages left
2627 * after the fairness batches are exhausted, and remote zones haven't
2628 * even been considered yet. Try once more without fairness, and
2629 * include remote zones now, before entering the slowpath and waking
2630 * kswapd: prefer spilling to a remote zone over swapping locally.
2632 if (alloc_flags & ALLOC_FAIR) {
2633 alloc_flags &= ~ALLOC_FAIR;
2634 if (nr_fair_skipped) {
2635 zonelist_rescan = true;
2636 reset_alloc_batches(ac->preferred_zone);
2638 if (nr_online_nodes > 1)
2639 zonelist_rescan = true;
2642 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2643 /* Disable zlc cache for second zonelist scan */
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, 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:%d, 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)))
2892 /* After successful reclaim, reconsider all zones for allocation */
2893 if (IS_ENABLED(CONFIG_NUMA))
2894 zlc_clear_zones_full(ac->zonelist);
2897 page = get_page_from_freelist(gfp_mask, order,
2898 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2901 * If an allocation failed after direct reclaim, it could be because
2902 * pages are pinned on the per-cpu lists. Drain them and try again
2904 if (!page && !drained) {
2905 drain_all_pages(NULL);
2914 * This is called in the allocator slow-path if the allocation request is of
2915 * sufficient urgency to ignore watermarks and take other desperate measures
2917 static inline struct page *
2918 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2919 const struct alloc_context *ac)
2924 page = get_page_from_freelist(gfp_mask, order,
2925 ALLOC_NO_WATERMARKS, ac);
2927 if (!page && gfp_mask & __GFP_NOFAIL)
2928 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2930 } while (!page && (gfp_mask & __GFP_NOFAIL));
2935 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2940 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2941 ac->high_zoneidx, ac->nodemask)
2942 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2946 gfp_to_alloc_flags(gfp_t gfp_mask)
2948 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2950 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2951 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2954 * The caller may dip into page reserves a bit more if the caller
2955 * cannot run direct reclaim, or if the caller has realtime scheduling
2956 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2957 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2959 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2961 if (gfp_mask & __GFP_ATOMIC) {
2963 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2964 * if it can't schedule.
2966 if (!(gfp_mask & __GFP_NOMEMALLOC))
2967 alloc_flags |= ALLOC_HARDER;
2969 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2970 * comment for __cpuset_node_allowed().
2972 alloc_flags &= ~ALLOC_CPUSET;
2973 } else if (unlikely(rt_task(current)) && !in_interrupt())
2974 alloc_flags |= ALLOC_HARDER;
2976 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2977 if (gfp_mask & __GFP_MEMALLOC)
2978 alloc_flags |= ALLOC_NO_WATERMARKS;
2979 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2980 alloc_flags |= ALLOC_NO_WATERMARKS;
2981 else if (!in_interrupt() &&
2982 ((current->flags & PF_MEMALLOC) ||
2983 unlikely(test_thread_flag(TIF_MEMDIE))))
2984 alloc_flags |= ALLOC_NO_WATERMARKS;
2987 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2988 alloc_flags |= ALLOC_CMA;
2993 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2995 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2998 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3000 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3003 static inline struct page *
3004 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3005 struct alloc_context *ac)
3007 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3008 struct page *page = NULL;
3010 unsigned long pages_reclaimed = 0;
3011 unsigned long did_some_progress;
3012 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3013 bool deferred_compaction = false;
3014 int contended_compaction = COMPACT_CONTENDED_NONE;
3017 * In the slowpath, we sanity check order to avoid ever trying to
3018 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3019 * be using allocators in order of preference for an area that is
3022 if (order >= MAX_ORDER) {
3023 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3028 * We also sanity check to catch abuse of atomic reserves being used by
3029 * callers that are not in atomic context.
3031 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3032 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3033 gfp_mask &= ~__GFP_ATOMIC;
3036 * If this allocation cannot block and it is for a specific node, then
3037 * fail early. There's no need to wakeup kswapd or retry for a
3038 * speculative node-specific allocation.
3040 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3044 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3045 wake_all_kswapds(order, ac);
3048 * OK, we're below the kswapd watermark and have kicked background
3049 * reclaim. Now things get more complex, so set up alloc_flags according
3050 * to how we want to proceed.
3052 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3055 * Find the true preferred zone if the allocation is unconstrained by
3058 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3059 struct zoneref *preferred_zoneref;
3060 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3061 ac->high_zoneidx, NULL, &ac->preferred_zone);
3062 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3065 /* This is the last chance, in general, before the goto nopage. */
3066 page = get_page_from_freelist(gfp_mask, order,
3067 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3071 /* Allocate without watermarks if the context allows */
3072 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3074 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3075 * the allocation is high priority and these type of
3076 * allocations are system rather than user orientated
3078 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3080 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3087 /* Caller is not willing to reclaim, we can't balance anything */
3088 if (!can_direct_reclaim) {
3090 * All existing users of the deprecated __GFP_NOFAIL are
3091 * blockable, so warn of any new users that actually allow this
3092 * type of allocation to fail.
3094 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3098 /* Avoid recursion of direct reclaim */
3099 if (current->flags & PF_MEMALLOC)
3102 /* Avoid allocations with no watermarks from looping endlessly */
3103 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3107 * Try direct compaction. The first pass is asynchronous. Subsequent
3108 * attempts after direct reclaim are synchronous
3110 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3112 &contended_compaction,
3113 &deferred_compaction);
3117 /* Checks for THP-specific high-order allocations */
3118 if (is_thp_gfp_mask(gfp_mask)) {
3120 * If compaction is deferred for high-order allocations, it is
3121 * because sync compaction recently failed. If this is the case
3122 * and the caller requested a THP allocation, we do not want
3123 * to heavily disrupt the system, so we fail the allocation
3124 * instead of entering direct reclaim.
3126 if (deferred_compaction)
3130 * In all zones where compaction was attempted (and not
3131 * deferred or skipped), lock contention has been detected.
3132 * For THP allocation we do not want to disrupt the others
3133 * so we fallback to base pages instead.
3135 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3139 * If compaction was aborted due to need_resched(), we do not
3140 * want to further increase allocation latency, unless it is
3141 * khugepaged trying to collapse.
3143 if (contended_compaction == COMPACT_CONTENDED_SCHED
3144 && !(current->flags & PF_KTHREAD))
3149 * It can become very expensive to allocate transparent hugepages at
3150 * fault, so use asynchronous memory compaction for THP unless it is
3151 * khugepaged trying to collapse.
3153 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3154 migration_mode = MIGRATE_SYNC_LIGHT;
3156 /* Try direct reclaim and then allocating */
3157 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3158 &did_some_progress);
3162 /* Do not loop if specifically requested */
3163 if (gfp_mask & __GFP_NORETRY)
3166 /* Keep reclaiming pages as long as there is reasonable progress */
3167 pages_reclaimed += did_some_progress;
3168 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3169 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3170 /* Wait for some write requests to complete then retry */
3171 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3175 /* Reclaim has failed us, start killing things */
3176 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3180 /* Retry as long as the OOM killer is making progress */
3181 if (did_some_progress)
3186 * High-order allocations do not necessarily loop after
3187 * direct reclaim and reclaim/compaction depends on compaction
3188 * being called after reclaim so call directly if necessary
3190 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3192 &contended_compaction,
3193 &deferred_compaction);
3197 warn_alloc_failed(gfp_mask, order, NULL);
3203 * This is the 'heart' of the zoned buddy allocator.
3206 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3207 struct zonelist *zonelist, nodemask_t *nodemask)
3209 struct zoneref *preferred_zoneref;
3210 struct page *page = NULL;
3211 unsigned int cpuset_mems_cookie;
3212 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3213 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3214 struct alloc_context ac = {
3215 .high_zoneidx = gfp_zone(gfp_mask),
3216 .nodemask = nodemask,
3217 .migratetype = gfpflags_to_migratetype(gfp_mask),
3220 gfp_mask &= gfp_allowed_mask;
3222 lockdep_trace_alloc(gfp_mask);
3224 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3226 if (should_fail_alloc_page(gfp_mask, order))
3230 * Check the zones suitable for the gfp_mask contain at least one
3231 * valid zone. It's possible to have an empty zonelist as a result
3232 * of __GFP_THISNODE and a memoryless node
3234 if (unlikely(!zonelist->_zonerefs->zone))
3237 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3238 alloc_flags |= ALLOC_CMA;
3241 cpuset_mems_cookie = read_mems_allowed_begin();
3243 /* We set it here, as __alloc_pages_slowpath might have changed it */
3244 ac.zonelist = zonelist;
3246 /* Dirty zone balancing only done in the fast path */
3247 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3249 /* The preferred zone is used for statistics later */
3250 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3251 ac.nodemask ? : &cpuset_current_mems_allowed,
3252 &ac.preferred_zone);
3253 if (!ac.preferred_zone)
3255 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3257 /* First allocation attempt */
3258 alloc_mask = gfp_mask|__GFP_HARDWALL;
3259 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3260 if (unlikely(!page)) {
3262 * Runtime PM, block IO and its error handling path
3263 * can deadlock because I/O on the device might not
3266 alloc_mask = memalloc_noio_flags(gfp_mask);
3267 ac.spread_dirty_pages = false;
3269 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3272 if (kmemcheck_enabled && page)
3273 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3275 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3279 * When updating a task's mems_allowed, it is possible to race with
3280 * parallel threads in such a way that an allocation can fail while
3281 * the mask is being updated. If a page allocation is about to fail,
3282 * check if the cpuset changed during allocation and if so, retry.
3284 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3289 EXPORT_SYMBOL(__alloc_pages_nodemask);
3292 * Common helper functions.
3294 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3299 * __get_free_pages() returns a 32-bit address, which cannot represent
3302 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3304 page = alloc_pages(gfp_mask, order);
3307 return (unsigned long) page_address(page);
3309 EXPORT_SYMBOL(__get_free_pages);
3311 unsigned long get_zeroed_page(gfp_t gfp_mask)
3313 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3315 EXPORT_SYMBOL(get_zeroed_page);
3317 void __free_pages(struct page *page, unsigned int order)
3319 if (put_page_testzero(page)) {
3321 free_hot_cold_page(page, false);
3323 __free_pages_ok(page, order);
3327 EXPORT_SYMBOL(__free_pages);
3329 void free_pages(unsigned long addr, unsigned int order)
3332 VM_BUG_ON(!virt_addr_valid((void *)addr));
3333 __free_pages(virt_to_page((void *)addr), order);
3337 EXPORT_SYMBOL(free_pages);
3341 * An arbitrary-length arbitrary-offset area of memory which resides
3342 * within a 0 or higher order page. Multiple fragments within that page
3343 * are individually refcounted, in the page's reference counter.
3345 * The page_frag functions below provide a simple allocation framework for
3346 * page fragments. This is used by the network stack and network device
3347 * drivers to provide a backing region of memory for use as either an
3348 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3350 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3353 struct page *page = NULL;
3354 gfp_t gfp = gfp_mask;
3356 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3357 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3359 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3360 PAGE_FRAG_CACHE_MAX_ORDER);
3361 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3363 if (unlikely(!page))
3364 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3366 nc->va = page ? page_address(page) : NULL;
3371 void *__alloc_page_frag(struct page_frag_cache *nc,
3372 unsigned int fragsz, gfp_t gfp_mask)
3374 unsigned int size = PAGE_SIZE;
3378 if (unlikely(!nc->va)) {
3380 page = __page_frag_refill(nc, gfp_mask);
3384 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3385 /* if size can vary use size else just use PAGE_SIZE */
3388 /* Even if we own the page, we do not use atomic_set().
3389 * This would break get_page_unless_zero() users.
3391 atomic_add(size - 1, &page->_count);
3393 /* reset page count bias and offset to start of new frag */
3394 nc->pfmemalloc = page_is_pfmemalloc(page);
3395 nc->pagecnt_bias = size;
3399 offset = nc->offset - fragsz;
3400 if (unlikely(offset < 0)) {
3401 page = virt_to_page(nc->va);
3403 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3406 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3407 /* if size can vary use size else just use PAGE_SIZE */
3410 /* OK, page count is 0, we can safely set it */
3411 atomic_set(&page->_count, size);
3413 /* reset page count bias and offset to start of new frag */
3414 nc->pagecnt_bias = size;
3415 offset = size - fragsz;
3419 nc->offset = offset;
3421 return nc->va + offset;
3423 EXPORT_SYMBOL(__alloc_page_frag);
3426 * Frees a page fragment allocated out of either a compound or order 0 page.
3428 void __free_page_frag(void *addr)
3430 struct page *page = virt_to_head_page(addr);
3432 if (unlikely(put_page_testzero(page)))
3433 __free_pages_ok(page, compound_order(page));
3435 EXPORT_SYMBOL(__free_page_frag);
3438 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3439 * of the current memory cgroup.
3441 * It should be used when the caller would like to use kmalloc, but since the
3442 * allocation is large, it has to fall back to the page allocator.
3444 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3448 page = alloc_pages(gfp_mask, order);
3449 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3450 __free_pages(page, order);
3456 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3460 page = alloc_pages_node(nid, gfp_mask, order);
3461 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3462 __free_pages(page, order);
3469 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3472 void __free_kmem_pages(struct page *page, unsigned int order)
3474 memcg_kmem_uncharge(page, order);
3475 __free_pages(page, order);
3478 void free_kmem_pages(unsigned long addr, unsigned int order)
3481 VM_BUG_ON(!virt_addr_valid((void *)addr));
3482 __free_kmem_pages(virt_to_page((void *)addr), order);
3486 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3489 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3490 unsigned long used = addr + PAGE_ALIGN(size);
3492 split_page(virt_to_page((void *)addr), order);
3493 while (used < alloc_end) {
3498 return (void *)addr;
3502 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3503 * @size: the number of bytes to allocate
3504 * @gfp_mask: GFP flags for the allocation
3506 * This function is similar to alloc_pages(), except that it allocates the
3507 * minimum number of pages to satisfy the request. alloc_pages() can only
3508 * allocate memory in power-of-two pages.
3510 * This function is also limited by MAX_ORDER.
3512 * Memory allocated by this function must be released by free_pages_exact().
3514 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3516 unsigned int order = get_order(size);
3519 addr = __get_free_pages(gfp_mask, order);
3520 return make_alloc_exact(addr, order, size);
3522 EXPORT_SYMBOL(alloc_pages_exact);
3525 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3527 * @nid: the preferred node ID where memory should be allocated
3528 * @size: the number of bytes to allocate
3529 * @gfp_mask: GFP flags for the allocation
3531 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3534 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3536 unsigned order = get_order(size);
3537 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3540 return make_alloc_exact((unsigned long)page_address(p), order, size);
3544 * free_pages_exact - release memory allocated via alloc_pages_exact()
3545 * @virt: the value returned by alloc_pages_exact.
3546 * @size: size of allocation, same value as passed to alloc_pages_exact().
3548 * Release the memory allocated by a previous call to alloc_pages_exact.
3550 void free_pages_exact(void *virt, size_t size)
3552 unsigned long addr = (unsigned long)virt;
3553 unsigned long end = addr + PAGE_ALIGN(size);
3555 while (addr < end) {
3560 EXPORT_SYMBOL(free_pages_exact);
3563 * nr_free_zone_pages - count number of pages beyond high watermark
3564 * @offset: The zone index of the highest zone
3566 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3567 * high watermark within all zones at or below a given zone index. For each
3568 * zone, the number of pages is calculated as:
3569 * managed_pages - high_pages
3571 static unsigned long nr_free_zone_pages(int offset)
3576 /* Just pick one node, since fallback list is circular */
3577 unsigned long sum = 0;
3579 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3581 for_each_zone_zonelist(zone, z, zonelist, offset) {
3582 unsigned long size = zone->managed_pages;
3583 unsigned long high = high_wmark_pages(zone);
3592 * nr_free_buffer_pages - count number of pages beyond high watermark
3594 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3595 * watermark within ZONE_DMA and ZONE_NORMAL.
3597 unsigned long nr_free_buffer_pages(void)
3599 return nr_free_zone_pages(gfp_zone(GFP_USER));
3601 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3604 * nr_free_pagecache_pages - count number of pages beyond high watermark
3606 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3607 * high watermark within all zones.
3609 unsigned long nr_free_pagecache_pages(void)
3611 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3614 static inline void show_node(struct zone *zone)
3616 if (IS_ENABLED(CONFIG_NUMA))
3617 printk("Node %d ", zone_to_nid(zone));
3620 void si_meminfo(struct sysinfo *val)
3622 val->totalram = totalram_pages;
3623 val->sharedram = global_page_state(NR_SHMEM);
3624 val->freeram = global_page_state(NR_FREE_PAGES);
3625 val->bufferram = nr_blockdev_pages();
3626 val->totalhigh = totalhigh_pages;
3627 val->freehigh = nr_free_highpages();
3628 val->mem_unit = PAGE_SIZE;
3631 EXPORT_SYMBOL(si_meminfo);
3634 void si_meminfo_node(struct sysinfo *val, int nid)
3636 int zone_type; /* needs to be signed */
3637 unsigned long managed_pages = 0;
3638 pg_data_t *pgdat = NODE_DATA(nid);
3640 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3641 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3642 val->totalram = managed_pages;
3643 val->sharedram = node_page_state(nid, NR_SHMEM);
3644 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3645 #ifdef CONFIG_HIGHMEM
3646 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3647 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3653 val->mem_unit = PAGE_SIZE;
3658 * Determine whether the node should be displayed or not, depending on whether
3659 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3661 bool skip_free_areas_node(unsigned int flags, int nid)
3664 unsigned int cpuset_mems_cookie;
3666 if (!(flags & SHOW_MEM_FILTER_NODES))
3670 cpuset_mems_cookie = read_mems_allowed_begin();
3671 ret = !node_isset(nid, cpuset_current_mems_allowed);
3672 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3677 #define K(x) ((x) << (PAGE_SHIFT-10))
3679 static void show_migration_types(unsigned char type)
3681 static const char types[MIGRATE_TYPES] = {
3682 [MIGRATE_UNMOVABLE] = 'U',
3683 [MIGRATE_RECLAIMABLE] = 'E',
3684 [MIGRATE_MOVABLE] = 'M',
3685 [MIGRATE_RESERVE] = 'R',
3687 [MIGRATE_CMA] = 'C',
3689 #ifdef CONFIG_MEMORY_ISOLATION
3690 [MIGRATE_ISOLATE] = 'I',
3693 char tmp[MIGRATE_TYPES + 1];
3697 for (i = 0; i < MIGRATE_TYPES; i++) {
3698 if (type & (1 << i))
3703 printk("(%s) ", tmp);
3707 * Show free area list (used inside shift_scroll-lock stuff)
3708 * We also calculate the percentage fragmentation. We do this by counting the
3709 * memory on each free list with the exception of the first item on the list.
3712 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3715 void show_free_areas(unsigned int filter)
3717 unsigned long free_pcp = 0;
3721 for_each_populated_zone(zone) {
3722 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3725 for_each_online_cpu(cpu)
3726 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3729 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3730 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3731 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3732 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3733 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3734 " free:%lu free_pcp:%lu free_cma:%lu\n",
3735 global_page_state(NR_ACTIVE_ANON),
3736 global_page_state(NR_INACTIVE_ANON),
3737 global_page_state(NR_ISOLATED_ANON),
3738 global_page_state(NR_ACTIVE_FILE),
3739 global_page_state(NR_INACTIVE_FILE),
3740 global_page_state(NR_ISOLATED_FILE),
3741 global_page_state(NR_UNEVICTABLE),
3742 global_page_state(NR_FILE_DIRTY),
3743 global_page_state(NR_WRITEBACK),
3744 global_page_state(NR_UNSTABLE_NFS),
3745 global_page_state(NR_SLAB_RECLAIMABLE),
3746 global_page_state(NR_SLAB_UNRECLAIMABLE),
3747 global_page_state(NR_FILE_MAPPED),
3748 global_page_state(NR_SHMEM),
3749 global_page_state(NR_PAGETABLE),
3750 global_page_state(NR_BOUNCE),
3751 global_page_state(NR_FREE_PAGES),
3753 global_page_state(NR_FREE_CMA_PAGES));
3755 for_each_populated_zone(zone) {
3758 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3762 for_each_online_cpu(cpu)
3763 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3771 " active_anon:%lukB"
3772 " inactive_anon:%lukB"
3773 " active_file:%lukB"
3774 " inactive_file:%lukB"
3775 " unevictable:%lukB"
3776 " isolated(anon):%lukB"
3777 " isolated(file):%lukB"
3785 " slab_reclaimable:%lukB"
3786 " slab_unreclaimable:%lukB"
3787 " kernel_stack:%lukB"
3794 " writeback_tmp:%lukB"
3795 " pages_scanned:%lu"
3796 " all_unreclaimable? %s"
3799 K(zone_page_state(zone, NR_FREE_PAGES)),
3800 K(min_wmark_pages(zone)),
3801 K(low_wmark_pages(zone)),
3802 K(high_wmark_pages(zone)),
3803 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3804 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3805 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3806 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3807 K(zone_page_state(zone, NR_UNEVICTABLE)),
3808 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3809 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3810 K(zone->present_pages),
3811 K(zone->managed_pages),
3812 K(zone_page_state(zone, NR_MLOCK)),
3813 K(zone_page_state(zone, NR_FILE_DIRTY)),
3814 K(zone_page_state(zone, NR_WRITEBACK)),
3815 K(zone_page_state(zone, NR_FILE_MAPPED)),
3816 K(zone_page_state(zone, NR_SHMEM)),
3817 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3818 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3819 zone_page_state(zone, NR_KERNEL_STACK) *
3821 K(zone_page_state(zone, NR_PAGETABLE)),
3822 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3823 K(zone_page_state(zone, NR_BOUNCE)),
3825 K(this_cpu_read(zone->pageset->pcp.count)),
3826 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3827 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3828 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3829 (!zone_reclaimable(zone) ? "yes" : "no")
3831 printk("lowmem_reserve[]:");
3832 for (i = 0; i < MAX_NR_ZONES; i++)
3833 printk(" %ld", zone->lowmem_reserve[i]);
3837 for_each_populated_zone(zone) {
3838 unsigned long nr[MAX_ORDER], flags, order, 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 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 /* Construct the zonelist performance cache - see further mmzone.h */
4232 static void build_zonelist_cache(pg_data_t *pgdat)
4234 struct zonelist *zonelist;
4235 struct zonelist_cache *zlc;
4238 zonelist = &pgdat->node_zonelists[0];
4239 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
4240 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
4241 for (z = zonelist->_zonerefs; z->zone; z++)
4242 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
4245 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4247 * Return node id of node used for "local" allocations.
4248 * I.e., first node id of first zone in arg node's generic zonelist.
4249 * Used for initializing percpu 'numa_mem', which is used primarily
4250 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4252 int local_memory_node(int node)
4256 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4257 gfp_zone(GFP_KERNEL),
4264 #else /* CONFIG_NUMA */
4266 static void set_zonelist_order(void)
4268 current_zonelist_order = ZONELIST_ORDER_ZONE;
4271 static void build_zonelists(pg_data_t *pgdat)
4273 int node, local_node;
4275 struct zonelist *zonelist;
4277 local_node = pgdat->node_id;
4279 zonelist = &pgdat->node_zonelists[0];
4280 j = build_zonelists_node(pgdat, zonelist, 0);
4283 * Now we build the zonelist so that it contains the zones
4284 * of all the other nodes.
4285 * We don't want to pressure a particular node, so when
4286 * building the zones for node N, we make sure that the
4287 * zones coming right after the local ones are those from
4288 * node N+1 (modulo N)
4290 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4291 if (!node_online(node))
4293 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4295 for (node = 0; node < local_node; node++) {
4296 if (!node_online(node))
4298 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4301 zonelist->_zonerefs[j].zone = NULL;
4302 zonelist->_zonerefs[j].zone_idx = 0;
4305 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4306 static void build_zonelist_cache(pg_data_t *pgdat)
4308 pgdat->node_zonelists[0].zlcache_ptr = NULL;
4311 #endif /* CONFIG_NUMA */
4314 * Boot pageset table. One per cpu which is going to be used for all
4315 * zones and all nodes. The parameters will be set in such a way
4316 * that an item put on a list will immediately be handed over to
4317 * the buddy list. This is safe since pageset manipulation is done
4318 * with interrupts disabled.
4320 * The boot_pagesets must be kept even after bootup is complete for
4321 * unused processors and/or zones. They do play a role for bootstrapping
4322 * hotplugged processors.
4324 * zoneinfo_show() and maybe other functions do
4325 * not check if the processor is online before following the pageset pointer.
4326 * Other parts of the kernel may not check if the zone is available.
4328 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4329 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4330 static void setup_zone_pageset(struct zone *zone);
4333 * Global mutex to protect against size modification of zonelists
4334 * as well as to serialize pageset setup for the new populated zone.
4336 DEFINE_MUTEX(zonelists_mutex);
4338 /* return values int ....just for stop_machine() */
4339 static int __build_all_zonelists(void *data)
4343 pg_data_t *self = data;
4346 memset(node_load, 0, sizeof(node_load));
4349 if (self && !node_online(self->node_id)) {
4350 build_zonelists(self);
4351 build_zonelist_cache(self);
4354 for_each_online_node(nid) {
4355 pg_data_t *pgdat = NODE_DATA(nid);
4357 build_zonelists(pgdat);
4358 build_zonelist_cache(pgdat);
4362 * Initialize the boot_pagesets that are going to be used
4363 * for bootstrapping processors. The real pagesets for
4364 * each zone will be allocated later when the per cpu
4365 * allocator is available.
4367 * boot_pagesets are used also for bootstrapping offline
4368 * cpus if the system is already booted because the pagesets
4369 * are needed to initialize allocators on a specific cpu too.
4370 * F.e. the percpu allocator needs the page allocator which
4371 * needs the percpu allocator in order to allocate its pagesets
4372 * (a chicken-egg dilemma).
4374 for_each_possible_cpu(cpu) {
4375 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4377 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4379 * We now know the "local memory node" for each node--
4380 * i.e., the node of the first zone in the generic zonelist.
4381 * Set up numa_mem percpu variable for on-line cpus. During
4382 * boot, only the boot cpu should be on-line; we'll init the
4383 * secondary cpus' numa_mem as they come on-line. During
4384 * node/memory hotplug, we'll fixup all on-line cpus.
4386 if (cpu_online(cpu))
4387 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4394 static noinline void __init
4395 build_all_zonelists_init(void)
4397 __build_all_zonelists(NULL);
4398 mminit_verify_zonelist();
4399 cpuset_init_current_mems_allowed();
4403 * Called with zonelists_mutex held always
4404 * unless system_state == SYSTEM_BOOTING.
4406 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4407 * [we're only called with non-NULL zone through __meminit paths] and
4408 * (2) call of __init annotated helper build_all_zonelists_init
4409 * [protected by SYSTEM_BOOTING].
4411 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4413 set_zonelist_order();
4415 if (system_state == SYSTEM_BOOTING) {
4416 build_all_zonelists_init();
4418 #ifdef CONFIG_MEMORY_HOTPLUG
4420 setup_zone_pageset(zone);
4422 /* we have to stop all cpus to guarantee there is no user
4424 stop_machine(__build_all_zonelists, pgdat, NULL);
4425 /* cpuset refresh routine should be here */
4427 vm_total_pages = nr_free_pagecache_pages();
4429 * Disable grouping by mobility if the number of pages in the
4430 * system is too low to allow the mechanism to work. It would be
4431 * more accurate, but expensive to check per-zone. This check is
4432 * made on memory-hotadd so a system can start with mobility
4433 * disabled and enable it later
4435 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4436 page_group_by_mobility_disabled = 1;
4438 page_group_by_mobility_disabled = 0;
4440 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4441 "Total pages: %ld\n",
4443 zonelist_order_name[current_zonelist_order],
4444 page_group_by_mobility_disabled ? "off" : "on",
4447 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4452 * Helper functions to size the waitqueue hash table.
4453 * Essentially these want to choose hash table sizes sufficiently
4454 * large so that collisions trying to wait on pages are rare.
4455 * But in fact, the number of active page waitqueues on typical
4456 * systems is ridiculously low, less than 200. So this is even
4457 * conservative, even though it seems large.
4459 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4460 * waitqueues, i.e. the size of the waitq table given the number of pages.
4462 #define PAGES_PER_WAITQUEUE 256
4464 #ifndef CONFIG_MEMORY_HOTPLUG
4465 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4467 unsigned long size = 1;
4469 pages /= PAGES_PER_WAITQUEUE;
4471 while (size < pages)
4475 * Once we have dozens or even hundreds of threads sleeping
4476 * on IO we've got bigger problems than wait queue collision.
4477 * Limit the size of the wait table to a reasonable size.
4479 size = min(size, 4096UL);
4481 return max(size, 4UL);
4485 * A zone's size might be changed by hot-add, so it is not possible to determine
4486 * a suitable size for its wait_table. So we use the maximum size now.
4488 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4490 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4491 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4492 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4494 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4495 * or more by the traditional way. (See above). It equals:
4497 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4498 * ia64(16K page size) : = ( 8G + 4M)byte.
4499 * powerpc (64K page size) : = (32G +16M)byte.
4501 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4508 * This is an integer logarithm so that shifts can be used later
4509 * to extract the more random high bits from the multiplicative
4510 * hash function before the remainder is taken.
4512 static inline unsigned long wait_table_bits(unsigned long size)
4518 * Check if a pageblock contains reserved pages
4520 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4524 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4525 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4532 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4533 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4534 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4535 * higher will lead to a bigger reserve which will get freed as contiguous
4536 * blocks as reclaim kicks in
4538 static void setup_zone_migrate_reserve(struct zone *zone)
4540 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4542 unsigned long block_migratetype;
4547 * Get the start pfn, end pfn and the number of blocks to reserve
4548 * We have to be careful to be aligned to pageblock_nr_pages to
4549 * make sure that we always check pfn_valid for the first page in
4552 start_pfn = zone->zone_start_pfn;
4553 end_pfn = zone_end_pfn(zone);
4554 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4555 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4559 * Reserve blocks are generally in place to help high-order atomic
4560 * allocations that are short-lived. A min_free_kbytes value that
4561 * would result in more than 2 reserve blocks for atomic allocations
4562 * is assumed to be in place to help anti-fragmentation for the
4563 * future allocation of hugepages at runtime.
4565 reserve = min(2, reserve);
4566 old_reserve = zone->nr_migrate_reserve_block;
4568 /* When memory hot-add, we almost always need to do nothing */
4569 if (reserve == old_reserve)
4571 zone->nr_migrate_reserve_block = reserve;
4573 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4574 if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone)))
4577 if (!pfn_valid(pfn))
4579 page = pfn_to_page(pfn);
4581 /* Watch out for overlapping nodes */
4582 if (page_to_nid(page) != zone_to_nid(zone))
4585 block_migratetype = get_pageblock_migratetype(page);
4587 /* Only test what is necessary when the reserves are not met */
4590 * Blocks with reserved pages will never free, skip
4593 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4594 if (pageblock_is_reserved(pfn, block_end_pfn))
4597 /* If this block is reserved, account for it */
4598 if (block_migratetype == MIGRATE_RESERVE) {
4603 /* Suitable for reserving if this block is movable */
4604 if (block_migratetype == MIGRATE_MOVABLE) {
4605 set_pageblock_migratetype(page,
4607 move_freepages_block(zone, page,
4612 } else if (!old_reserve) {
4614 * At boot time we don't need to scan the whole zone
4615 * for turning off MIGRATE_RESERVE.
4621 * If the reserve is met and this is a previous reserved block,
4624 if (block_migratetype == MIGRATE_RESERVE) {
4625 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4626 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4632 * Initially all pages are reserved - free ones are freed
4633 * up by free_all_bootmem() once the early boot process is
4634 * done. Non-atomic initialization, single-pass.
4636 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4637 unsigned long start_pfn, enum memmap_context context)
4639 pg_data_t *pgdat = NODE_DATA(nid);
4640 unsigned long end_pfn = start_pfn + size;
4643 unsigned long nr_initialised = 0;
4645 if (highest_memmap_pfn < end_pfn - 1)
4646 highest_memmap_pfn = end_pfn - 1;
4648 z = &pgdat->node_zones[zone];
4649 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4651 * There can be holes in boot-time mem_map[]s
4652 * handed to this function. They do not
4653 * exist on hotplugged memory.
4655 if (context == MEMMAP_EARLY) {
4656 if (!early_pfn_valid(pfn))
4658 if (!early_pfn_in_nid(pfn, nid))
4660 if (!update_defer_init(pgdat, pfn, end_pfn,
4666 * Mark the block movable so that blocks are reserved for
4667 * movable at startup. This will force kernel allocations
4668 * to reserve their blocks rather than leaking throughout
4669 * the address space during boot when many long-lived
4670 * kernel allocations are made. Later some blocks near
4671 * the start are marked MIGRATE_RESERVE by
4672 * setup_zone_migrate_reserve()
4674 * bitmap is created for zone's valid pfn range. but memmap
4675 * can be created for invalid pages (for alignment)
4676 * check here not to call set_pageblock_migratetype() against
4679 if (!(pfn & (pageblock_nr_pages - 1))) {
4680 struct page *page = pfn_to_page(pfn);
4682 __init_single_page(page, pfn, zone, nid);
4683 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4685 __init_single_pfn(pfn, zone, nid);
4690 static void __meminit zone_init_free_lists(struct zone *zone)
4692 unsigned int order, t;
4693 for_each_migratetype_order(order, t) {
4694 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4695 zone->free_area[order].nr_free = 0;
4699 #ifndef __HAVE_ARCH_MEMMAP_INIT
4700 #define memmap_init(size, nid, zone, start_pfn) \
4701 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4704 static int zone_batchsize(struct zone *zone)
4710 * The per-cpu-pages pools are set to around 1000th of the
4711 * size of the zone. But no more than 1/2 of a meg.
4713 * OK, so we don't know how big the cache is. So guess.
4715 batch = zone->managed_pages / 1024;
4716 if (batch * PAGE_SIZE > 512 * 1024)
4717 batch = (512 * 1024) / PAGE_SIZE;
4718 batch /= 4; /* We effectively *= 4 below */
4723 * Clamp the batch to a 2^n - 1 value. Having a power
4724 * of 2 value was found to be more likely to have
4725 * suboptimal cache aliasing properties in some cases.
4727 * For example if 2 tasks are alternately allocating
4728 * batches of pages, one task can end up with a lot
4729 * of pages of one half of the possible page colors
4730 * and the other with pages of the other colors.
4732 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4737 /* The deferral and batching of frees should be suppressed under NOMMU
4740 * The problem is that NOMMU needs to be able to allocate large chunks
4741 * of contiguous memory as there's no hardware page translation to
4742 * assemble apparent contiguous memory from discontiguous pages.
4744 * Queueing large contiguous runs of pages for batching, however,
4745 * causes the pages to actually be freed in smaller chunks. As there
4746 * can be a significant delay between the individual batches being
4747 * recycled, this leads to the once large chunks of space being
4748 * fragmented and becoming unavailable for high-order allocations.
4755 * pcp->high and pcp->batch values are related and dependent on one another:
4756 * ->batch must never be higher then ->high.
4757 * The following function updates them in a safe manner without read side
4760 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4761 * those fields changing asynchronously (acording the the above rule).
4763 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4764 * outside of boot time (or some other assurance that no concurrent updaters
4767 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4768 unsigned long batch)
4770 /* start with a fail safe value for batch */
4774 /* Update high, then batch, in order */
4781 /* a companion to pageset_set_high() */
4782 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4784 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4787 static void pageset_init(struct per_cpu_pageset *p)
4789 struct per_cpu_pages *pcp;
4792 memset(p, 0, sizeof(*p));
4796 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4797 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4800 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4803 pageset_set_batch(p, batch);
4807 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4808 * to the value high for the pageset p.
4810 static void pageset_set_high(struct per_cpu_pageset *p,
4813 unsigned long batch = max(1UL, high / 4);
4814 if ((high / 4) > (PAGE_SHIFT * 8))
4815 batch = PAGE_SHIFT * 8;
4817 pageset_update(&p->pcp, high, batch);
4820 static void pageset_set_high_and_batch(struct zone *zone,
4821 struct per_cpu_pageset *pcp)
4823 if (percpu_pagelist_fraction)
4824 pageset_set_high(pcp,
4825 (zone->managed_pages /
4826 percpu_pagelist_fraction));
4828 pageset_set_batch(pcp, zone_batchsize(zone));
4831 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4833 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4836 pageset_set_high_and_batch(zone, pcp);
4839 static void __meminit setup_zone_pageset(struct zone *zone)
4842 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4843 for_each_possible_cpu(cpu)
4844 zone_pageset_init(zone, cpu);
4848 * Allocate per cpu pagesets and initialize them.
4849 * Before this call only boot pagesets were available.
4851 void __init setup_per_cpu_pageset(void)
4855 for_each_populated_zone(zone)
4856 setup_zone_pageset(zone);
4859 static noinline __init_refok
4860 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4866 * The per-page waitqueue mechanism uses hashed waitqueues
4869 zone->wait_table_hash_nr_entries =
4870 wait_table_hash_nr_entries(zone_size_pages);
4871 zone->wait_table_bits =
4872 wait_table_bits(zone->wait_table_hash_nr_entries);
4873 alloc_size = zone->wait_table_hash_nr_entries
4874 * sizeof(wait_queue_head_t);
4876 if (!slab_is_available()) {
4877 zone->wait_table = (wait_queue_head_t *)
4878 memblock_virt_alloc_node_nopanic(
4879 alloc_size, zone->zone_pgdat->node_id);
4882 * This case means that a zone whose size was 0 gets new memory
4883 * via memory hot-add.
4884 * But it may be the case that a new node was hot-added. In
4885 * this case vmalloc() will not be able to use this new node's
4886 * memory - this wait_table must be initialized to use this new
4887 * node itself as well.
4888 * To use this new node's memory, further consideration will be
4891 zone->wait_table = vmalloc(alloc_size);
4893 if (!zone->wait_table)
4896 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4897 init_waitqueue_head(zone->wait_table + i);
4902 static __meminit void zone_pcp_init(struct zone *zone)
4905 * per cpu subsystem is not up at this point. The following code
4906 * relies on the ability of the linker to provide the
4907 * offset of a (static) per cpu variable into the per cpu area.
4909 zone->pageset = &boot_pageset;
4911 if (populated_zone(zone))
4912 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4913 zone->name, zone->present_pages,
4914 zone_batchsize(zone));
4917 int __meminit init_currently_empty_zone(struct zone *zone,
4918 unsigned long zone_start_pfn,
4921 struct pglist_data *pgdat = zone->zone_pgdat;
4923 ret = zone_wait_table_init(zone, size);
4926 pgdat->nr_zones = zone_idx(zone) + 1;
4928 zone->zone_start_pfn = zone_start_pfn;
4930 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4931 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4933 (unsigned long)zone_idx(zone),
4934 zone_start_pfn, (zone_start_pfn + size));
4936 zone_init_free_lists(zone);
4941 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4942 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4945 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4947 int __meminit __early_pfn_to_nid(unsigned long pfn,
4948 struct mminit_pfnnid_cache *state)
4950 unsigned long start_pfn, end_pfn;
4953 if (state->last_start <= pfn && pfn < state->last_end)
4954 return state->last_nid;
4956 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4958 state->last_start = start_pfn;
4959 state->last_end = end_pfn;
4960 state->last_nid = nid;
4965 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4968 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4969 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4970 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4972 * If an architecture guarantees that all ranges registered contain no holes
4973 * and may be freed, this this function may be used instead of calling
4974 * memblock_free_early_nid() manually.
4976 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4978 unsigned long start_pfn, end_pfn;
4981 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4982 start_pfn = min(start_pfn, max_low_pfn);
4983 end_pfn = min(end_pfn, max_low_pfn);
4985 if (start_pfn < end_pfn)
4986 memblock_free_early_nid(PFN_PHYS(start_pfn),
4987 (end_pfn - start_pfn) << PAGE_SHIFT,
4993 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4994 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4996 * If an architecture guarantees that all ranges registered contain no holes and may
4997 * be freed, this function may be used instead of calling memory_present() manually.
4999 void __init sparse_memory_present_with_active_regions(int nid)
5001 unsigned long start_pfn, end_pfn;
5004 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5005 memory_present(this_nid, start_pfn, end_pfn);
5009 * get_pfn_range_for_nid - Return the start and end page frames for a node
5010 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5011 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5012 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5014 * It returns the start and end page frame of a node based on information
5015 * provided by memblock_set_node(). If called for a node
5016 * with no available memory, a warning is printed and the start and end
5019 void __meminit get_pfn_range_for_nid(unsigned int nid,
5020 unsigned long *start_pfn, unsigned long *end_pfn)
5022 unsigned long this_start_pfn, this_end_pfn;
5028 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5029 *start_pfn = min(*start_pfn, this_start_pfn);
5030 *end_pfn = max(*end_pfn, this_end_pfn);
5033 if (*start_pfn == -1UL)
5038 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5039 * assumption is made that zones within a node are ordered in monotonic
5040 * increasing memory addresses so that the "highest" populated zone is used
5042 static void __init find_usable_zone_for_movable(void)
5045 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5046 if (zone_index == ZONE_MOVABLE)
5049 if (arch_zone_highest_possible_pfn[zone_index] >
5050 arch_zone_lowest_possible_pfn[zone_index])
5054 VM_BUG_ON(zone_index == -1);
5055 movable_zone = zone_index;
5059 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5060 * because it is sized independent of architecture. Unlike the other zones,
5061 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5062 * in each node depending on the size of each node and how evenly kernelcore
5063 * is distributed. This helper function adjusts the zone ranges
5064 * provided by the architecture for a given node by using the end of the
5065 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5066 * zones within a node are in order of monotonic increases memory addresses
5068 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5069 unsigned long zone_type,
5070 unsigned long node_start_pfn,
5071 unsigned long node_end_pfn,
5072 unsigned long *zone_start_pfn,
5073 unsigned long *zone_end_pfn)
5075 /* Only adjust if ZONE_MOVABLE is on this node */
5076 if (zone_movable_pfn[nid]) {
5077 /* Size ZONE_MOVABLE */
5078 if (zone_type == ZONE_MOVABLE) {
5079 *zone_start_pfn = zone_movable_pfn[nid];
5080 *zone_end_pfn = min(node_end_pfn,
5081 arch_zone_highest_possible_pfn[movable_zone]);
5083 /* Adjust for ZONE_MOVABLE starting within this range */
5084 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
5085 *zone_end_pfn > zone_movable_pfn[nid]) {
5086 *zone_end_pfn = zone_movable_pfn[nid];
5088 /* Check if this whole range is within ZONE_MOVABLE */
5089 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5090 *zone_start_pfn = *zone_end_pfn;
5095 * Return the number of pages a zone spans in a node, including holes
5096 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5098 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5099 unsigned long zone_type,
5100 unsigned long node_start_pfn,
5101 unsigned long node_end_pfn,
5102 unsigned long *ignored)
5104 unsigned long zone_start_pfn, zone_end_pfn;
5106 /* When hotadd a new node from cpu_up(), the node should be empty */
5107 if (!node_start_pfn && !node_end_pfn)
5110 /* Get the start and end of the zone */
5111 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5112 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5113 adjust_zone_range_for_zone_movable(nid, zone_type,
5114 node_start_pfn, node_end_pfn,
5115 &zone_start_pfn, &zone_end_pfn);
5117 /* Check that this node has pages within the zone's required range */
5118 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
5121 /* Move the zone boundaries inside the node if necessary */
5122 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
5123 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5125 /* Return the spanned pages */
5126 return zone_end_pfn - zone_start_pfn;
5130 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5131 * then all holes in the requested range will be accounted for.
5133 unsigned long __meminit __absent_pages_in_range(int nid,
5134 unsigned long range_start_pfn,
5135 unsigned long range_end_pfn)
5137 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5138 unsigned long start_pfn, end_pfn;
5141 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5142 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5143 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5144 nr_absent -= end_pfn - start_pfn;
5150 * absent_pages_in_range - Return number of page frames in holes within a range
5151 * @start_pfn: The start PFN to start searching for holes
5152 * @end_pfn: The end PFN to stop searching for holes
5154 * It returns the number of pages frames in memory holes within a range.
5156 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5157 unsigned long end_pfn)
5159 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5162 /* Return the number of page frames in holes in a zone on a node */
5163 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5164 unsigned long zone_type,
5165 unsigned long node_start_pfn,
5166 unsigned long node_end_pfn,
5167 unsigned long *ignored)
5169 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5170 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5171 unsigned long zone_start_pfn, zone_end_pfn;
5173 /* When hotadd a new node from cpu_up(), the node should be empty */
5174 if (!node_start_pfn && !node_end_pfn)
5177 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5178 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5180 adjust_zone_range_for_zone_movable(nid, zone_type,
5181 node_start_pfn, node_end_pfn,
5182 &zone_start_pfn, &zone_end_pfn);
5183 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5186 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5187 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5188 unsigned long zone_type,
5189 unsigned long node_start_pfn,
5190 unsigned long node_end_pfn,
5191 unsigned long *zones_size)
5193 return zones_size[zone_type];
5196 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5197 unsigned long zone_type,
5198 unsigned long node_start_pfn,
5199 unsigned long node_end_pfn,
5200 unsigned long *zholes_size)
5205 return zholes_size[zone_type];
5208 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5210 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5211 unsigned long node_start_pfn,
5212 unsigned long node_end_pfn,
5213 unsigned long *zones_size,
5214 unsigned long *zholes_size)
5216 unsigned long realtotalpages = 0, totalpages = 0;
5219 for (i = 0; i < MAX_NR_ZONES; i++) {
5220 struct zone *zone = pgdat->node_zones + i;
5221 unsigned long size, real_size;
5223 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5227 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5228 node_start_pfn, node_end_pfn,
5230 zone->spanned_pages = size;
5231 zone->present_pages = real_size;
5234 realtotalpages += real_size;
5237 pgdat->node_spanned_pages = totalpages;
5238 pgdat->node_present_pages = realtotalpages;
5239 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5243 #ifndef CONFIG_SPARSEMEM
5245 * Calculate the size of the zone->blockflags rounded to an unsigned long
5246 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5247 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5248 * round what is now in bits to nearest long in bits, then return it in
5251 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5253 unsigned long usemapsize;
5255 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5256 usemapsize = roundup(zonesize, pageblock_nr_pages);
5257 usemapsize = usemapsize >> pageblock_order;
5258 usemapsize *= NR_PAGEBLOCK_BITS;
5259 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5261 return usemapsize / 8;
5264 static void __init setup_usemap(struct pglist_data *pgdat,
5266 unsigned long zone_start_pfn,
5267 unsigned long zonesize)
5269 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5270 zone->pageblock_flags = NULL;
5272 zone->pageblock_flags =
5273 memblock_virt_alloc_node_nopanic(usemapsize,
5277 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5278 unsigned long zone_start_pfn, unsigned long zonesize) {}
5279 #endif /* CONFIG_SPARSEMEM */
5281 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5283 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5284 void __paginginit set_pageblock_order(void)
5288 /* Check that pageblock_nr_pages has not already been setup */
5289 if (pageblock_order)
5292 if (HPAGE_SHIFT > PAGE_SHIFT)
5293 order = HUGETLB_PAGE_ORDER;
5295 order = MAX_ORDER - 1;
5298 * Assume the largest contiguous order of interest is a huge page.
5299 * This value may be variable depending on boot parameters on IA64 and
5302 pageblock_order = order;
5304 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5307 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5308 * is unused as pageblock_order is set at compile-time. See
5309 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5312 void __paginginit set_pageblock_order(void)
5316 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5318 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5319 unsigned long present_pages)
5321 unsigned long pages = spanned_pages;
5324 * Provide a more accurate estimation if there are holes within
5325 * the zone and SPARSEMEM is in use. If there are holes within the
5326 * zone, each populated memory region may cost us one or two extra
5327 * memmap pages due to alignment because memmap pages for each
5328 * populated regions may not naturally algined on page boundary.
5329 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5331 if (spanned_pages > present_pages + (present_pages >> 4) &&
5332 IS_ENABLED(CONFIG_SPARSEMEM))
5333 pages = present_pages;
5335 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5339 * Set up the zone data structures:
5340 * - mark all pages reserved
5341 * - mark all memory queues empty
5342 * - clear the memory bitmaps
5344 * NOTE: pgdat should get zeroed by caller.
5346 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5349 int nid = pgdat->node_id;
5350 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5353 pgdat_resize_init(pgdat);
5354 #ifdef CONFIG_NUMA_BALANCING
5355 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5356 pgdat->numabalancing_migrate_nr_pages = 0;
5357 pgdat->numabalancing_migrate_next_window = jiffies;
5359 init_waitqueue_head(&pgdat->kswapd_wait);
5360 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5361 pgdat_page_ext_init(pgdat);
5363 for (j = 0; j < MAX_NR_ZONES; j++) {
5364 struct zone *zone = pgdat->node_zones + j;
5365 unsigned long size, realsize, freesize, memmap_pages;
5367 size = zone->spanned_pages;
5368 realsize = freesize = zone->present_pages;
5371 * Adjust freesize so that it accounts for how much memory
5372 * is used by this zone for memmap. This affects the watermark
5373 * and per-cpu initialisations
5375 memmap_pages = calc_memmap_size(size, realsize);
5376 if (!is_highmem_idx(j)) {
5377 if (freesize >= memmap_pages) {
5378 freesize -= memmap_pages;
5381 " %s zone: %lu pages used for memmap\n",
5382 zone_names[j], memmap_pages);
5385 " %s zone: %lu pages exceeds freesize %lu\n",
5386 zone_names[j], memmap_pages, freesize);
5389 /* Account for reserved pages */
5390 if (j == 0 && freesize > dma_reserve) {
5391 freesize -= dma_reserve;
5392 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5393 zone_names[0], dma_reserve);
5396 if (!is_highmem_idx(j))
5397 nr_kernel_pages += freesize;
5398 /* Charge for highmem memmap if there are enough kernel pages */
5399 else if (nr_kernel_pages > memmap_pages * 2)
5400 nr_kernel_pages -= memmap_pages;
5401 nr_all_pages += freesize;
5404 * Set an approximate value for lowmem here, it will be adjusted
5405 * when the bootmem allocator frees pages into the buddy system.
5406 * And all highmem pages will be managed by the buddy system.
5408 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5411 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5413 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5415 zone->name = zone_names[j];
5416 spin_lock_init(&zone->lock);
5417 spin_lock_init(&zone->lru_lock);
5418 zone_seqlock_init(zone);
5419 zone->zone_pgdat = pgdat;
5420 zone_pcp_init(zone);
5422 /* For bootup, initialized properly in watermark setup */
5423 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5425 lruvec_init(&zone->lruvec);
5429 set_pageblock_order();
5430 setup_usemap(pgdat, zone, zone_start_pfn, size);
5431 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5433 memmap_init(size, nid, j, zone_start_pfn);
5434 zone_start_pfn += size;
5438 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5440 unsigned long __maybe_unused offset = 0;
5442 /* Skip empty nodes */
5443 if (!pgdat->node_spanned_pages)
5446 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5447 /* ia64 gets its own node_mem_map, before this, without bootmem */
5448 if (!pgdat->node_mem_map) {
5449 unsigned long size, start, end;
5453 * The zone's endpoints aren't required to be MAX_ORDER
5454 * aligned but the node_mem_map endpoints must be in order
5455 * for the buddy allocator to function correctly.
5457 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5458 offset = pgdat->node_start_pfn - start;
5459 end = pgdat_end_pfn(pgdat);
5460 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5461 size = (end - start) * sizeof(struct page);
5462 map = alloc_remap(pgdat->node_id, size);
5464 map = memblock_virt_alloc_node_nopanic(size,
5466 pgdat->node_mem_map = map + offset;
5468 #ifndef CONFIG_NEED_MULTIPLE_NODES
5470 * With no DISCONTIG, the global mem_map is just set as node 0's
5472 if (pgdat == NODE_DATA(0)) {
5473 mem_map = NODE_DATA(0)->node_mem_map;
5474 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5475 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5477 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5480 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5483 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5484 unsigned long node_start_pfn, unsigned long *zholes_size)
5486 pg_data_t *pgdat = NODE_DATA(nid);
5487 unsigned long start_pfn = 0;
5488 unsigned long end_pfn = 0;
5490 /* pg_data_t should be reset to zero when it's allocated */
5491 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5493 reset_deferred_meminit(pgdat);
5494 pgdat->node_id = nid;
5495 pgdat->node_start_pfn = node_start_pfn;
5496 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5497 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5498 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5499 (u64)start_pfn << PAGE_SHIFT,
5500 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5502 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5503 zones_size, zholes_size);
5505 alloc_node_mem_map(pgdat);
5506 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5507 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5508 nid, (unsigned long)pgdat,
5509 (unsigned long)pgdat->node_mem_map);
5512 free_area_init_core(pgdat);
5515 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5517 #if MAX_NUMNODES > 1
5519 * Figure out the number of possible node ids.
5521 void __init setup_nr_node_ids(void)
5523 unsigned int highest;
5525 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5526 nr_node_ids = highest + 1;
5531 * node_map_pfn_alignment - determine the maximum internode alignment
5533 * This function should be called after node map is populated and sorted.
5534 * It calculates the maximum power of two alignment which can distinguish
5537 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5538 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5539 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5540 * shifted, 1GiB is enough and this function will indicate so.
5542 * This is used to test whether pfn -> nid mapping of the chosen memory
5543 * model has fine enough granularity to avoid incorrect mapping for the
5544 * populated node map.
5546 * Returns the determined alignment in pfn's. 0 if there is no alignment
5547 * requirement (single node).
5549 unsigned long __init node_map_pfn_alignment(void)
5551 unsigned long accl_mask = 0, last_end = 0;
5552 unsigned long start, end, mask;
5556 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5557 if (!start || last_nid < 0 || last_nid == nid) {
5564 * Start with a mask granular enough to pin-point to the
5565 * start pfn and tick off bits one-by-one until it becomes
5566 * too coarse to separate the current node from the last.
5568 mask = ~((1 << __ffs(start)) - 1);
5569 while (mask && last_end <= (start & (mask << 1)))
5572 /* accumulate all internode masks */
5576 /* convert mask to number of pages */
5577 return ~accl_mask + 1;
5580 /* Find the lowest pfn for a node */
5581 static unsigned long __init find_min_pfn_for_node(int nid)
5583 unsigned long min_pfn = ULONG_MAX;
5584 unsigned long start_pfn;
5587 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5588 min_pfn = min(min_pfn, start_pfn);
5590 if (min_pfn == ULONG_MAX) {
5592 "Could not find start_pfn for node %d\n", nid);
5600 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5602 * It returns the minimum PFN based on information provided via
5603 * memblock_set_node().
5605 unsigned long __init find_min_pfn_with_active_regions(void)
5607 return find_min_pfn_for_node(MAX_NUMNODES);
5611 * early_calculate_totalpages()
5612 * Sum pages in active regions for movable zone.
5613 * Populate N_MEMORY for calculating usable_nodes.
5615 static unsigned long __init early_calculate_totalpages(void)
5617 unsigned long totalpages = 0;
5618 unsigned long start_pfn, end_pfn;
5621 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5622 unsigned long pages = end_pfn - start_pfn;
5624 totalpages += pages;
5626 node_set_state(nid, N_MEMORY);
5632 * Find the PFN the Movable zone begins in each node. Kernel memory
5633 * is spread evenly between nodes as long as the nodes have enough
5634 * memory. When they don't, some nodes will have more kernelcore than
5637 static void __init find_zone_movable_pfns_for_nodes(void)
5640 unsigned long usable_startpfn;
5641 unsigned long kernelcore_node, kernelcore_remaining;
5642 /* save the state before borrow the nodemask */
5643 nodemask_t saved_node_state = node_states[N_MEMORY];
5644 unsigned long totalpages = early_calculate_totalpages();
5645 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5646 struct memblock_region *r;
5648 /* Need to find movable_zone earlier when movable_node is specified. */
5649 find_usable_zone_for_movable();
5652 * If movable_node is specified, ignore kernelcore and movablecore
5655 if (movable_node_is_enabled()) {
5656 for_each_memblock(memory, r) {
5657 if (!memblock_is_hotpluggable(r))
5662 usable_startpfn = PFN_DOWN(r->base);
5663 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5664 min(usable_startpfn, zone_movable_pfn[nid]) :
5672 * If movablecore=nn[KMG] was specified, calculate what size of
5673 * kernelcore that corresponds so that memory usable for
5674 * any allocation type is evenly spread. If both kernelcore
5675 * and movablecore are specified, then the value of kernelcore
5676 * will be used for required_kernelcore if it's greater than
5677 * what movablecore would have allowed.
5679 if (required_movablecore) {
5680 unsigned long corepages;
5683 * Round-up so that ZONE_MOVABLE is at least as large as what
5684 * was requested by the user
5686 required_movablecore =
5687 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5688 required_movablecore = min(totalpages, required_movablecore);
5689 corepages = totalpages - required_movablecore;
5691 required_kernelcore = max(required_kernelcore, corepages);
5695 * If kernelcore was not specified or kernelcore size is larger
5696 * than totalpages, there is no ZONE_MOVABLE.
5698 if (!required_kernelcore || required_kernelcore >= totalpages)
5701 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5702 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5705 /* Spread kernelcore memory as evenly as possible throughout nodes */
5706 kernelcore_node = required_kernelcore / usable_nodes;
5707 for_each_node_state(nid, N_MEMORY) {
5708 unsigned long start_pfn, end_pfn;
5711 * Recalculate kernelcore_node if the division per node
5712 * now exceeds what is necessary to satisfy the requested
5713 * amount of memory for the kernel
5715 if (required_kernelcore < kernelcore_node)
5716 kernelcore_node = required_kernelcore / usable_nodes;
5719 * As the map is walked, we track how much memory is usable
5720 * by the kernel using kernelcore_remaining. When it is
5721 * 0, the rest of the node is usable by ZONE_MOVABLE
5723 kernelcore_remaining = kernelcore_node;
5725 /* Go through each range of PFNs within this node */
5726 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5727 unsigned long size_pages;
5729 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5730 if (start_pfn >= end_pfn)
5733 /* Account for what is only usable for kernelcore */
5734 if (start_pfn < usable_startpfn) {
5735 unsigned long kernel_pages;
5736 kernel_pages = min(end_pfn, usable_startpfn)
5739 kernelcore_remaining -= min(kernel_pages,
5740 kernelcore_remaining);
5741 required_kernelcore -= min(kernel_pages,
5742 required_kernelcore);
5744 /* Continue if range is now fully accounted */
5745 if (end_pfn <= usable_startpfn) {
5748 * Push zone_movable_pfn to the end so
5749 * that if we have to rebalance
5750 * kernelcore across nodes, we will
5751 * not double account here
5753 zone_movable_pfn[nid] = end_pfn;
5756 start_pfn = usable_startpfn;
5760 * The usable PFN range for ZONE_MOVABLE is from
5761 * start_pfn->end_pfn. Calculate size_pages as the
5762 * number of pages used as kernelcore
5764 size_pages = end_pfn - start_pfn;
5765 if (size_pages > kernelcore_remaining)
5766 size_pages = kernelcore_remaining;
5767 zone_movable_pfn[nid] = start_pfn + size_pages;
5770 * Some kernelcore has been met, update counts and
5771 * break if the kernelcore for this node has been
5774 required_kernelcore -= min(required_kernelcore,
5776 kernelcore_remaining -= size_pages;
5777 if (!kernelcore_remaining)
5783 * If there is still required_kernelcore, we do another pass with one
5784 * less node in the count. This will push zone_movable_pfn[nid] further
5785 * along on the nodes that still have memory until kernelcore is
5789 if (usable_nodes && required_kernelcore > usable_nodes)
5793 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5794 for (nid = 0; nid < MAX_NUMNODES; nid++)
5795 zone_movable_pfn[nid] =
5796 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5799 /* restore the node_state */
5800 node_states[N_MEMORY] = saved_node_state;
5803 /* Any regular or high memory on that node ? */
5804 static void check_for_memory(pg_data_t *pgdat, int nid)
5806 enum zone_type zone_type;
5808 if (N_MEMORY == N_NORMAL_MEMORY)
5811 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5812 struct zone *zone = &pgdat->node_zones[zone_type];
5813 if (populated_zone(zone)) {
5814 node_set_state(nid, N_HIGH_MEMORY);
5815 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5816 zone_type <= ZONE_NORMAL)
5817 node_set_state(nid, N_NORMAL_MEMORY);
5824 * free_area_init_nodes - Initialise all pg_data_t and zone data
5825 * @max_zone_pfn: an array of max PFNs for each zone
5827 * This will call free_area_init_node() for each active node in the system.
5828 * Using the page ranges provided by memblock_set_node(), the size of each
5829 * zone in each node and their holes is calculated. If the maximum PFN
5830 * between two adjacent zones match, it is assumed that the zone is empty.
5831 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5832 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5833 * starts where the previous one ended. For example, ZONE_DMA32 starts
5834 * at arch_max_dma_pfn.
5836 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5838 unsigned long start_pfn, end_pfn;
5841 /* Record where the zone boundaries are */
5842 memset(arch_zone_lowest_possible_pfn, 0,
5843 sizeof(arch_zone_lowest_possible_pfn));
5844 memset(arch_zone_highest_possible_pfn, 0,
5845 sizeof(arch_zone_highest_possible_pfn));
5846 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5847 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5848 for (i = 1; i < MAX_NR_ZONES; i++) {
5849 if (i == ZONE_MOVABLE)
5851 arch_zone_lowest_possible_pfn[i] =
5852 arch_zone_highest_possible_pfn[i-1];
5853 arch_zone_highest_possible_pfn[i] =
5854 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5856 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5857 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5859 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5860 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5861 find_zone_movable_pfns_for_nodes();
5863 /* Print out the zone ranges */
5864 pr_info("Zone ranges:\n");
5865 for (i = 0; i < MAX_NR_ZONES; i++) {
5866 if (i == ZONE_MOVABLE)
5868 pr_info(" %-8s ", zone_names[i]);
5869 if (arch_zone_lowest_possible_pfn[i] ==
5870 arch_zone_highest_possible_pfn[i])
5873 pr_cont("[mem %#018Lx-%#018Lx]\n",
5874 (u64)arch_zone_lowest_possible_pfn[i]
5876 ((u64)arch_zone_highest_possible_pfn[i]
5877 << PAGE_SHIFT) - 1);
5880 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5881 pr_info("Movable zone start for each node\n");
5882 for (i = 0; i < MAX_NUMNODES; i++) {
5883 if (zone_movable_pfn[i])
5884 pr_info(" Node %d: %#018Lx\n", i,
5885 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5888 /* Print out the early node map */
5889 pr_info("Early memory node ranges\n");
5890 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5891 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5892 (u64)start_pfn << PAGE_SHIFT,
5893 ((u64)end_pfn << PAGE_SHIFT) - 1);
5895 /* Initialise every node */
5896 mminit_verify_pageflags_layout();
5897 setup_nr_node_ids();
5898 for_each_online_node(nid) {
5899 pg_data_t *pgdat = NODE_DATA(nid);
5900 free_area_init_node(nid, NULL,
5901 find_min_pfn_for_node(nid), NULL);
5903 /* Any memory on that node */
5904 if (pgdat->node_present_pages)
5905 node_set_state(nid, N_MEMORY);
5906 check_for_memory(pgdat, nid);
5910 static int __init cmdline_parse_core(char *p, unsigned long *core)
5912 unsigned long long coremem;
5916 coremem = memparse(p, &p);
5917 *core = coremem >> PAGE_SHIFT;
5919 /* Paranoid check that UL is enough for the coremem value */
5920 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5926 * kernelcore=size sets the amount of memory for use for allocations that
5927 * cannot be reclaimed or migrated.
5929 static int __init cmdline_parse_kernelcore(char *p)
5931 return cmdline_parse_core(p, &required_kernelcore);
5935 * movablecore=size sets the amount of memory for use for allocations that
5936 * can be reclaimed or migrated.
5938 static int __init cmdline_parse_movablecore(char *p)
5940 return cmdline_parse_core(p, &required_movablecore);
5943 early_param("kernelcore", cmdline_parse_kernelcore);
5944 early_param("movablecore", cmdline_parse_movablecore);
5946 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5948 void adjust_managed_page_count(struct page *page, long count)
5950 spin_lock(&managed_page_count_lock);
5951 page_zone(page)->managed_pages += count;
5952 totalram_pages += count;
5953 #ifdef CONFIG_HIGHMEM
5954 if (PageHighMem(page))
5955 totalhigh_pages += count;
5957 spin_unlock(&managed_page_count_lock);
5959 EXPORT_SYMBOL(adjust_managed_page_count);
5961 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5964 unsigned long pages = 0;
5966 start = (void *)PAGE_ALIGN((unsigned long)start);
5967 end = (void *)((unsigned long)end & PAGE_MASK);
5968 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5969 if ((unsigned int)poison <= 0xFF)
5970 memset(pos, poison, PAGE_SIZE);
5971 free_reserved_page(virt_to_page(pos));
5975 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5976 s, pages << (PAGE_SHIFT - 10), start, end);
5980 EXPORT_SYMBOL(free_reserved_area);
5982 #ifdef CONFIG_HIGHMEM
5983 void free_highmem_page(struct page *page)
5985 __free_reserved_page(page);
5987 page_zone(page)->managed_pages++;
5993 void __init mem_init_print_info(const char *str)
5995 unsigned long physpages, codesize, datasize, rosize, bss_size;
5996 unsigned long init_code_size, init_data_size;
5998 physpages = get_num_physpages();
5999 codesize = _etext - _stext;
6000 datasize = _edata - _sdata;
6001 rosize = __end_rodata - __start_rodata;
6002 bss_size = __bss_stop - __bss_start;
6003 init_data_size = __init_end - __init_begin;
6004 init_code_size = _einittext - _sinittext;
6007 * Detect special cases and adjust section sizes accordingly:
6008 * 1) .init.* may be embedded into .data sections
6009 * 2) .init.text.* may be out of [__init_begin, __init_end],
6010 * please refer to arch/tile/kernel/vmlinux.lds.S.
6011 * 3) .rodata.* may be embedded into .text or .data sections.
6013 #define adj_init_size(start, end, size, pos, adj) \
6015 if (start <= pos && pos < end && size > adj) \
6019 adj_init_size(__init_begin, __init_end, init_data_size,
6020 _sinittext, init_code_size);
6021 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6022 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6023 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6024 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6026 #undef adj_init_size
6028 pr_info("Memory: %luK/%luK available "
6029 "(%luK kernel code, %luK rwdata, %luK rodata, "
6030 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
6031 #ifdef CONFIG_HIGHMEM
6035 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
6036 codesize >> 10, datasize >> 10, rosize >> 10,
6037 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6038 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
6039 totalcma_pages << (PAGE_SHIFT-10),
6040 #ifdef CONFIG_HIGHMEM
6041 totalhigh_pages << (PAGE_SHIFT-10),
6043 str ? ", " : "", str ? str : "");
6047 * set_dma_reserve - set the specified number of pages reserved in the first zone
6048 * @new_dma_reserve: The number of pages to mark reserved
6050 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6051 * In the DMA zone, a significant percentage may be consumed by kernel image
6052 * and other unfreeable allocations which can skew the watermarks badly. This
6053 * function may optionally be used to account for unfreeable pages in the
6054 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6055 * smaller per-cpu batchsize.
6057 void __init set_dma_reserve(unsigned long new_dma_reserve)
6059 dma_reserve = new_dma_reserve;
6062 void __init free_area_init(unsigned long *zones_size)
6064 free_area_init_node(0, zones_size,
6065 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6068 static int page_alloc_cpu_notify(struct notifier_block *self,
6069 unsigned long action, void *hcpu)
6071 int cpu = (unsigned long)hcpu;
6073 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6074 lru_add_drain_cpu(cpu);
6078 * Spill the event counters of the dead processor
6079 * into the current processors event counters.
6080 * This artificially elevates the count of the current
6083 vm_events_fold_cpu(cpu);
6086 * Zero the differential counters of the dead processor
6087 * so that the vm statistics are consistent.
6089 * This is only okay since the processor is dead and cannot
6090 * race with what we are doing.
6092 cpu_vm_stats_fold(cpu);
6097 void __init page_alloc_init(void)
6099 hotcpu_notifier(page_alloc_cpu_notify, 0);
6103 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6104 * or min_free_kbytes changes.
6106 static void calculate_totalreserve_pages(void)
6108 struct pglist_data *pgdat;
6109 unsigned long reserve_pages = 0;
6110 enum zone_type i, j;
6112 for_each_online_pgdat(pgdat) {
6113 for (i = 0; i < MAX_NR_ZONES; i++) {
6114 struct zone *zone = pgdat->node_zones + i;
6117 /* Find valid and maximum lowmem_reserve in the zone */
6118 for (j = i; j < MAX_NR_ZONES; j++) {
6119 if (zone->lowmem_reserve[j] > max)
6120 max = zone->lowmem_reserve[j];
6123 /* we treat the high watermark as reserved pages. */
6124 max += high_wmark_pages(zone);
6126 if (max > zone->managed_pages)
6127 max = zone->managed_pages;
6128 reserve_pages += max;
6130 * Lowmem reserves are not available to
6131 * GFP_HIGHUSER page cache allocations and
6132 * kswapd tries to balance zones to their high
6133 * watermark. As a result, neither should be
6134 * regarded as dirtyable memory, to prevent a
6135 * situation where reclaim has to clean pages
6136 * in order to balance the zones.
6138 zone->dirty_balance_reserve = max;
6141 dirty_balance_reserve = reserve_pages;
6142 totalreserve_pages = reserve_pages;
6146 * setup_per_zone_lowmem_reserve - called whenever
6147 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6148 * has a correct pages reserved value, so an adequate number of
6149 * pages are left in the zone after a successful __alloc_pages().
6151 static void setup_per_zone_lowmem_reserve(void)
6153 struct pglist_data *pgdat;
6154 enum zone_type j, idx;
6156 for_each_online_pgdat(pgdat) {
6157 for (j = 0; j < MAX_NR_ZONES; j++) {
6158 struct zone *zone = pgdat->node_zones + j;
6159 unsigned long managed_pages = zone->managed_pages;
6161 zone->lowmem_reserve[j] = 0;
6165 struct zone *lower_zone;
6169 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6170 sysctl_lowmem_reserve_ratio[idx] = 1;
6172 lower_zone = pgdat->node_zones + idx;
6173 lower_zone->lowmem_reserve[j] = managed_pages /
6174 sysctl_lowmem_reserve_ratio[idx];
6175 managed_pages += lower_zone->managed_pages;
6180 /* update totalreserve_pages */
6181 calculate_totalreserve_pages();
6184 static void __setup_per_zone_wmarks(void)
6186 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6187 unsigned long lowmem_pages = 0;
6189 unsigned long flags;
6191 /* Calculate total number of !ZONE_HIGHMEM pages */
6192 for_each_zone(zone) {
6193 if (!is_highmem(zone))
6194 lowmem_pages += zone->managed_pages;
6197 for_each_zone(zone) {
6200 spin_lock_irqsave(&zone->lock, flags);
6201 tmp = (u64)pages_min * zone->managed_pages;
6202 do_div(tmp, lowmem_pages);
6203 if (is_highmem(zone)) {
6205 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6206 * need highmem pages, so cap pages_min to a small
6209 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6210 * deltas control asynch page reclaim, and so should
6211 * not be capped for highmem.
6213 unsigned long min_pages;
6215 min_pages = zone->managed_pages / 1024;
6216 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6217 zone->watermark[WMARK_MIN] = min_pages;
6220 * If it's a lowmem zone, reserve a number of pages
6221 * proportionate to the zone's size.
6223 zone->watermark[WMARK_MIN] = tmp;
6226 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6227 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6229 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6230 high_wmark_pages(zone) - low_wmark_pages(zone) -
6231 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6233 setup_zone_migrate_reserve(zone);
6234 spin_unlock_irqrestore(&zone->lock, flags);
6237 /* update totalreserve_pages */
6238 calculate_totalreserve_pages();
6242 * setup_per_zone_wmarks - called when min_free_kbytes changes
6243 * or when memory is hot-{added|removed}
6245 * Ensures that the watermark[min,low,high] values for each zone are set
6246 * correctly with respect to min_free_kbytes.
6248 void setup_per_zone_wmarks(void)
6250 mutex_lock(&zonelists_mutex);
6251 __setup_per_zone_wmarks();
6252 mutex_unlock(&zonelists_mutex);
6256 * The inactive anon list should be small enough that the VM never has to
6257 * do too much work, but large enough that each inactive page has a chance
6258 * to be referenced again before it is swapped out.
6260 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6261 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6262 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6263 * the anonymous pages are kept on the inactive list.
6266 * memory ratio inactive anon
6267 * -------------------------------------
6276 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6278 unsigned int gb, ratio;
6280 /* Zone size in gigabytes */
6281 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6283 ratio = int_sqrt(10 * gb);
6287 zone->inactive_ratio = ratio;
6290 static void __meminit setup_per_zone_inactive_ratio(void)
6295 calculate_zone_inactive_ratio(zone);
6299 * Initialise min_free_kbytes.
6301 * For small machines we want it small (128k min). For large machines
6302 * we want it large (64MB max). But it is not linear, because network
6303 * bandwidth does not increase linearly with machine size. We use
6305 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6306 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6322 int __meminit init_per_zone_wmark_min(void)
6324 unsigned long lowmem_kbytes;
6325 int new_min_free_kbytes;
6327 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6328 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6330 if (new_min_free_kbytes > user_min_free_kbytes) {
6331 min_free_kbytes = new_min_free_kbytes;
6332 if (min_free_kbytes < 128)
6333 min_free_kbytes = 128;
6334 if (min_free_kbytes > 65536)
6335 min_free_kbytes = 65536;
6337 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6338 new_min_free_kbytes, user_min_free_kbytes);
6340 setup_per_zone_wmarks();
6341 refresh_zone_stat_thresholds();
6342 setup_per_zone_lowmem_reserve();
6343 setup_per_zone_inactive_ratio();
6346 module_init(init_per_zone_wmark_min)
6349 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6350 * that we can call two helper functions whenever min_free_kbytes
6353 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6354 void __user *buffer, size_t *length, loff_t *ppos)
6358 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6363 user_min_free_kbytes = min_free_kbytes;
6364 setup_per_zone_wmarks();
6370 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6371 void __user *buffer, size_t *length, loff_t *ppos)
6376 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6381 zone->min_unmapped_pages = (zone->managed_pages *
6382 sysctl_min_unmapped_ratio) / 100;
6386 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6387 void __user *buffer, size_t *length, loff_t *ppos)
6392 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6397 zone->min_slab_pages = (zone->managed_pages *
6398 sysctl_min_slab_ratio) / 100;
6404 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6405 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6406 * whenever sysctl_lowmem_reserve_ratio changes.
6408 * The reserve ratio obviously has absolutely no relation with the
6409 * minimum watermarks. The lowmem reserve ratio can only make sense
6410 * if in function of the boot time zone sizes.
6412 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6413 void __user *buffer, size_t *length, loff_t *ppos)
6415 proc_dointvec_minmax(table, write, buffer, length, ppos);
6416 setup_per_zone_lowmem_reserve();
6421 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6422 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6423 * pagelist can have before it gets flushed back to buddy allocator.
6425 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6426 void __user *buffer, size_t *length, loff_t *ppos)
6429 int old_percpu_pagelist_fraction;
6432 mutex_lock(&pcp_batch_high_lock);
6433 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6435 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6436 if (!write || ret < 0)
6439 /* Sanity checking to avoid pcp imbalance */
6440 if (percpu_pagelist_fraction &&
6441 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6442 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6448 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6451 for_each_populated_zone(zone) {
6454 for_each_possible_cpu(cpu)
6455 pageset_set_high_and_batch(zone,
6456 per_cpu_ptr(zone->pageset, cpu));
6459 mutex_unlock(&pcp_batch_high_lock);
6464 int hashdist = HASHDIST_DEFAULT;
6466 static int __init set_hashdist(char *str)
6470 hashdist = simple_strtoul(str, &str, 0);
6473 __setup("hashdist=", set_hashdist);
6477 * allocate a large system hash table from bootmem
6478 * - it is assumed that the hash table must contain an exact power-of-2
6479 * quantity of entries
6480 * - limit is the number of hash buckets, not the total allocation size
6482 void *__init alloc_large_system_hash(const char *tablename,
6483 unsigned long bucketsize,
6484 unsigned long numentries,
6487 unsigned int *_hash_shift,
6488 unsigned int *_hash_mask,
6489 unsigned long low_limit,
6490 unsigned long high_limit)
6492 unsigned long long max = high_limit;
6493 unsigned long log2qty, size;
6496 /* allow the kernel cmdline to have a say */
6498 /* round applicable memory size up to nearest megabyte */
6499 numentries = nr_kernel_pages;
6501 /* It isn't necessary when PAGE_SIZE >= 1MB */
6502 if (PAGE_SHIFT < 20)
6503 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6505 /* limit to 1 bucket per 2^scale bytes of low memory */
6506 if (scale > PAGE_SHIFT)
6507 numentries >>= (scale - PAGE_SHIFT);
6509 numentries <<= (PAGE_SHIFT - scale);
6511 /* Make sure we've got at least a 0-order allocation.. */
6512 if (unlikely(flags & HASH_SMALL)) {
6513 /* Makes no sense without HASH_EARLY */
6514 WARN_ON(!(flags & HASH_EARLY));
6515 if (!(numentries >> *_hash_shift)) {
6516 numentries = 1UL << *_hash_shift;
6517 BUG_ON(!numentries);
6519 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6520 numentries = PAGE_SIZE / bucketsize;
6522 numentries = roundup_pow_of_two(numentries);
6524 /* limit allocation size to 1/16 total memory by default */
6526 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6527 do_div(max, bucketsize);
6529 max = min(max, 0x80000000ULL);
6531 if (numentries < low_limit)
6532 numentries = low_limit;
6533 if (numentries > max)
6536 log2qty = ilog2(numentries);
6539 size = bucketsize << log2qty;
6540 if (flags & HASH_EARLY)
6541 table = memblock_virt_alloc_nopanic(size, 0);
6543 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6546 * If bucketsize is not a power-of-two, we may free
6547 * some pages at the end of hash table which
6548 * alloc_pages_exact() automatically does
6550 if (get_order(size) < MAX_ORDER) {
6551 table = alloc_pages_exact(size, GFP_ATOMIC);
6552 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6555 } while (!table && size > PAGE_SIZE && --log2qty);
6558 panic("Failed to allocate %s hash table\n", tablename);
6560 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6563 ilog2(size) - PAGE_SHIFT,
6567 *_hash_shift = log2qty;
6569 *_hash_mask = (1 << log2qty) - 1;
6574 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6575 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6578 #ifdef CONFIG_SPARSEMEM
6579 return __pfn_to_section(pfn)->pageblock_flags;
6581 return zone->pageblock_flags;
6582 #endif /* CONFIG_SPARSEMEM */
6585 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6587 #ifdef CONFIG_SPARSEMEM
6588 pfn &= (PAGES_PER_SECTION-1);
6589 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6591 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6592 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6593 #endif /* CONFIG_SPARSEMEM */
6597 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6598 * @page: The page within the block of interest
6599 * @pfn: The target page frame number
6600 * @end_bitidx: The last bit of interest to retrieve
6601 * @mask: mask of bits that the caller is interested in
6603 * Return: pageblock_bits flags
6605 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6606 unsigned long end_bitidx,
6610 unsigned long *bitmap;
6611 unsigned long bitidx, word_bitidx;
6614 zone = page_zone(page);
6615 bitmap = get_pageblock_bitmap(zone, pfn);
6616 bitidx = pfn_to_bitidx(zone, pfn);
6617 word_bitidx = bitidx / BITS_PER_LONG;
6618 bitidx &= (BITS_PER_LONG-1);
6620 word = bitmap[word_bitidx];
6621 bitidx += end_bitidx;
6622 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6626 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6627 * @page: The page within the block of interest
6628 * @flags: The flags to set
6629 * @pfn: The target page frame number
6630 * @end_bitidx: The last bit of interest
6631 * @mask: mask of bits that the caller is interested in
6633 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6635 unsigned long end_bitidx,
6639 unsigned long *bitmap;
6640 unsigned long bitidx, word_bitidx;
6641 unsigned long old_word, word;
6643 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6645 zone = page_zone(page);
6646 bitmap = get_pageblock_bitmap(zone, pfn);
6647 bitidx = pfn_to_bitidx(zone, pfn);
6648 word_bitidx = bitidx / BITS_PER_LONG;
6649 bitidx &= (BITS_PER_LONG-1);
6651 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6653 bitidx += end_bitidx;
6654 mask <<= (BITS_PER_LONG - bitidx - 1);
6655 flags <<= (BITS_PER_LONG - bitidx - 1);
6657 word = READ_ONCE(bitmap[word_bitidx]);
6659 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6660 if (word == old_word)
6667 * This function checks whether pageblock includes unmovable pages or not.
6668 * If @count is not zero, it is okay to include less @count unmovable pages
6670 * PageLRU check without isolation or lru_lock could race so that
6671 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6672 * expect this function should be exact.
6674 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6675 bool skip_hwpoisoned_pages)
6677 unsigned long pfn, iter, found;
6681 * For avoiding noise data, lru_add_drain_all() should be called
6682 * If ZONE_MOVABLE, the zone never contains unmovable pages
6684 if (zone_idx(zone) == ZONE_MOVABLE)
6686 mt = get_pageblock_migratetype(page);
6687 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6690 pfn = page_to_pfn(page);
6691 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6692 unsigned long check = pfn + iter;
6694 if (!pfn_valid_within(check))
6697 page = pfn_to_page(check);
6700 * Hugepages are not in LRU lists, but they're movable.
6701 * We need not scan over tail pages bacause we don't
6702 * handle each tail page individually in migration.
6704 if (PageHuge(page)) {
6705 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6710 * We can't use page_count without pin a page
6711 * because another CPU can free compound page.
6712 * This check already skips compound tails of THP
6713 * because their page->_count is zero at all time.
6715 if (!atomic_read(&page->_count)) {
6716 if (PageBuddy(page))
6717 iter += (1 << page_order(page)) - 1;
6722 * The HWPoisoned page may be not in buddy system, and
6723 * page_count() is not 0.
6725 if (skip_hwpoisoned_pages && PageHWPoison(page))
6731 * If there are RECLAIMABLE pages, we need to check
6732 * it. But now, memory offline itself doesn't call
6733 * shrink_node_slabs() and it still to be fixed.
6736 * If the page is not RAM, page_count()should be 0.
6737 * we don't need more check. This is an _used_ not-movable page.
6739 * The problematic thing here is PG_reserved pages. PG_reserved
6740 * is set to both of a memory hole page and a _used_ kernel
6749 bool is_pageblock_removable_nolock(struct page *page)
6755 * We have to be careful here because we are iterating over memory
6756 * sections which are not zone aware so we might end up outside of
6757 * the zone but still within the section.
6758 * We have to take care about the node as well. If the node is offline
6759 * its NODE_DATA will be NULL - see page_zone.
6761 if (!node_online(page_to_nid(page)))
6764 zone = page_zone(page);
6765 pfn = page_to_pfn(page);
6766 if (!zone_spans_pfn(zone, pfn))
6769 return !has_unmovable_pages(zone, page, 0, true);
6774 static unsigned long pfn_max_align_down(unsigned long pfn)
6776 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6777 pageblock_nr_pages) - 1);
6780 static unsigned long pfn_max_align_up(unsigned long pfn)
6782 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6783 pageblock_nr_pages));
6786 /* [start, end) must belong to a single zone. */
6787 static int __alloc_contig_migrate_range(struct compact_control *cc,
6788 unsigned long start, unsigned long end)
6790 /* This function is based on compact_zone() from compaction.c. */
6791 unsigned long nr_reclaimed;
6792 unsigned long pfn = start;
6793 unsigned int tries = 0;
6798 while (pfn < end || !list_empty(&cc->migratepages)) {
6799 if (fatal_signal_pending(current)) {
6804 if (list_empty(&cc->migratepages)) {
6805 cc->nr_migratepages = 0;
6806 pfn = isolate_migratepages_range(cc, pfn, end);
6812 } else if (++tries == 5) {
6813 ret = ret < 0 ? ret : -EBUSY;
6817 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6819 cc->nr_migratepages -= nr_reclaimed;
6821 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6822 NULL, 0, cc->mode, MR_CMA);
6825 putback_movable_pages(&cc->migratepages);
6832 * alloc_contig_range() -- tries to allocate given range of pages
6833 * @start: start PFN to allocate
6834 * @end: one-past-the-last PFN to allocate
6835 * @migratetype: migratetype of the underlaying pageblocks (either
6836 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6837 * in range must have the same migratetype and it must
6838 * be either of the two.
6840 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6841 * aligned, however it's the caller's responsibility to guarantee that
6842 * we are the only thread that changes migrate type of pageblocks the
6845 * The PFN range must belong to a single zone.
6847 * Returns zero on success or negative error code. On success all
6848 * pages which PFN is in [start, end) are allocated for the caller and
6849 * need to be freed with free_contig_range().
6851 int alloc_contig_range(unsigned long start, unsigned long end,
6852 unsigned migratetype)
6854 unsigned long outer_start, outer_end;
6857 struct compact_control cc = {
6858 .nr_migratepages = 0,
6860 .zone = page_zone(pfn_to_page(start)),
6861 .mode = MIGRATE_SYNC,
6862 .ignore_skip_hint = true,
6864 INIT_LIST_HEAD(&cc.migratepages);
6867 * What we do here is we mark all pageblocks in range as
6868 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6869 * have different sizes, and due to the way page allocator
6870 * work, we align the range to biggest of the two pages so
6871 * that page allocator won't try to merge buddies from
6872 * different pageblocks and change MIGRATE_ISOLATE to some
6873 * other migration type.
6875 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6876 * migrate the pages from an unaligned range (ie. pages that
6877 * we are interested in). This will put all the pages in
6878 * range back to page allocator as MIGRATE_ISOLATE.
6880 * When this is done, we take the pages in range from page
6881 * allocator removing them from the buddy system. This way
6882 * page allocator will never consider using them.
6884 * This lets us mark the pageblocks back as
6885 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6886 * aligned range but not in the unaligned, original range are
6887 * put back to page allocator so that buddy can use them.
6890 ret = start_isolate_page_range(pfn_max_align_down(start),
6891 pfn_max_align_up(end), migratetype,
6896 ret = __alloc_contig_migrate_range(&cc, start, end);
6901 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6902 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6903 * more, all pages in [start, end) are free in page allocator.
6904 * What we are going to do is to allocate all pages from
6905 * [start, end) (that is remove them from page allocator).
6907 * The only problem is that pages at the beginning and at the
6908 * end of interesting range may be not aligned with pages that
6909 * page allocator holds, ie. they can be part of higher order
6910 * pages. Because of this, we reserve the bigger range and
6911 * once this is done free the pages we are not interested in.
6913 * We don't have to hold zone->lock here because the pages are
6914 * isolated thus they won't get removed from buddy.
6917 lru_add_drain_all();
6918 drain_all_pages(cc.zone);
6921 outer_start = start;
6922 while (!PageBuddy(pfn_to_page(outer_start))) {
6923 if (++order >= MAX_ORDER) {
6927 outer_start &= ~0UL << order;
6930 /* Make sure the range is really isolated. */
6931 if (test_pages_isolated(outer_start, end, false)) {
6932 pr_info("%s: [%lx, %lx) PFNs busy\n",
6933 __func__, outer_start, end);
6938 /* Grab isolated pages from freelists. */
6939 outer_end = isolate_freepages_range(&cc, outer_start, end);
6945 /* Free head and tail (if any) */
6946 if (start != outer_start)
6947 free_contig_range(outer_start, start - outer_start);
6948 if (end != outer_end)
6949 free_contig_range(end, outer_end - end);
6952 undo_isolate_page_range(pfn_max_align_down(start),
6953 pfn_max_align_up(end), migratetype);
6957 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6959 unsigned int count = 0;
6961 for (; nr_pages--; pfn++) {
6962 struct page *page = pfn_to_page(pfn);
6964 count += page_count(page) != 1;
6967 WARN(count != 0, "%d pages are still in use!\n", count);
6971 #ifdef CONFIG_MEMORY_HOTPLUG
6973 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6974 * page high values need to be recalulated.
6976 void __meminit zone_pcp_update(struct zone *zone)
6979 mutex_lock(&pcp_batch_high_lock);
6980 for_each_possible_cpu(cpu)
6981 pageset_set_high_and_batch(zone,
6982 per_cpu_ptr(zone->pageset, cpu));
6983 mutex_unlock(&pcp_batch_high_lock);
6987 void zone_pcp_reset(struct zone *zone)
6989 unsigned long flags;
6991 struct per_cpu_pageset *pset;
6993 /* avoid races with drain_pages() */
6994 local_irq_save(flags);
6995 if (zone->pageset != &boot_pageset) {
6996 for_each_online_cpu(cpu) {
6997 pset = per_cpu_ptr(zone->pageset, cpu);
6998 drain_zonestat(zone, pset);
7000 free_percpu(zone->pageset);
7001 zone->pageset = &boot_pageset;
7003 local_irq_restore(flags);
7006 #ifdef CONFIG_MEMORY_HOTREMOVE
7008 * All pages in the range must be isolated before calling this.
7011 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7015 unsigned int order, i;
7017 unsigned long flags;
7018 /* find the first valid pfn */
7019 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7024 zone = page_zone(pfn_to_page(pfn));
7025 spin_lock_irqsave(&zone->lock, flags);
7027 while (pfn < end_pfn) {
7028 if (!pfn_valid(pfn)) {
7032 page = pfn_to_page(pfn);
7034 * The HWPoisoned page may be not in buddy system, and
7035 * page_count() is not 0.
7037 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7039 SetPageReserved(page);
7043 BUG_ON(page_count(page));
7044 BUG_ON(!PageBuddy(page));
7045 order = page_order(page);
7046 #ifdef CONFIG_DEBUG_VM
7047 printk(KERN_INFO "remove from free list %lx %d %lx\n",
7048 pfn, 1 << order, end_pfn);
7050 list_del(&page->lru);
7051 rmv_page_order(page);
7052 zone->free_area[order].nr_free--;
7053 for (i = 0; i < (1 << order); i++)
7054 SetPageReserved((page+i));
7055 pfn += (1 << order);
7057 spin_unlock_irqrestore(&zone->lock, flags);
7061 #ifdef CONFIG_MEMORY_FAILURE
7062 bool is_free_buddy_page(struct page *page)
7064 struct zone *zone = page_zone(page);
7065 unsigned long pfn = page_to_pfn(page);
7066 unsigned long flags;
7069 spin_lock_irqsave(&zone->lock, flags);
7070 for (order = 0; order < MAX_ORDER; order++) {
7071 struct page *page_head = page - (pfn & ((1 << order) - 1));
7073 if (PageBuddy(page_head) && page_order(page_head) >= order)
7076 spin_unlock_irqrestore(&zone->lock, flags);
7078 return order < MAX_ORDER;