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 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
821 trace_mm_page_pcpu_drain(page, 0, mt);
822 } while (--to_free && --batch_free && !list_empty(list));
824 spin_unlock(&zone->lock);
827 static void free_one_page(struct zone *zone,
828 struct page *page, unsigned long pfn,
832 unsigned long nr_scanned;
833 spin_lock(&zone->lock);
834 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
836 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
838 if (unlikely(has_isolate_pageblock(zone) ||
839 is_migrate_isolate(migratetype))) {
840 migratetype = get_pfnblock_migratetype(page, pfn);
842 __free_one_page(page, pfn, zone, order, migratetype);
843 spin_unlock(&zone->lock);
846 static int free_tail_pages_check(struct page *head_page, struct page *page)
848 if (!IS_ENABLED(CONFIG_DEBUG_VM))
850 if (unlikely(!PageTail(page))) {
851 bad_page(page, "PageTail not set", 0);
854 if (unlikely(page->first_page != head_page)) {
855 bad_page(page, "first_page not consistent", 0);
861 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
862 unsigned long zone, int nid)
864 set_page_links(page, zone, nid, pfn);
865 init_page_count(page);
866 page_mapcount_reset(page);
867 page_cpupid_reset_last(page);
869 INIT_LIST_HEAD(&page->lru);
870 #ifdef WANT_PAGE_VIRTUAL
871 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
872 if (!is_highmem_idx(zone))
873 set_page_address(page, __va(pfn << PAGE_SHIFT));
877 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
880 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
883 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
884 static void init_reserved_page(unsigned long pfn)
889 if (!early_page_uninitialised(pfn))
892 nid = early_pfn_to_nid(pfn);
893 pgdat = NODE_DATA(nid);
895 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
896 struct zone *zone = &pgdat->node_zones[zid];
898 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
901 __init_single_pfn(pfn, zid, nid);
904 static inline void init_reserved_page(unsigned long pfn)
907 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
910 * Initialised pages do not have PageReserved set. This function is
911 * called for each range allocated by the bootmem allocator and
912 * marks the pages PageReserved. The remaining valid pages are later
913 * sent to the buddy page allocator.
915 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
917 unsigned long start_pfn = PFN_DOWN(start);
918 unsigned long end_pfn = PFN_UP(end);
920 for (; start_pfn < end_pfn; start_pfn++) {
921 if (pfn_valid(start_pfn)) {
922 struct page *page = pfn_to_page(start_pfn);
924 init_reserved_page(start_pfn);
925 SetPageReserved(page);
930 static bool free_pages_prepare(struct page *page, unsigned int order)
932 bool compound = PageCompound(page);
935 VM_BUG_ON_PAGE(PageTail(page), page);
936 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
938 trace_mm_page_free(page, order);
939 kmemcheck_free_shadow(page, order);
940 kasan_free_pages(page, order);
943 page->mapping = NULL;
944 bad += free_pages_check(page);
945 for (i = 1; i < (1 << order); i++) {
947 bad += free_tail_pages_check(page, page + i);
948 bad += free_pages_check(page + i);
953 reset_page_owner(page, order);
955 if (!PageHighMem(page)) {
956 debug_check_no_locks_freed(page_address(page),
958 debug_check_no_obj_freed(page_address(page),
961 arch_free_page(page, order);
962 kernel_map_pages(page, 1 << order, 0);
967 static void __free_pages_ok(struct page *page, unsigned int order)
971 unsigned long pfn = page_to_pfn(page);
973 if (!free_pages_prepare(page, order))
976 migratetype = get_pfnblock_migratetype(page, pfn);
977 local_irq_save(flags);
978 __count_vm_events(PGFREE, 1 << order);
979 free_one_page(page_zone(page), page, pfn, order, migratetype);
980 local_irq_restore(flags);
983 static void __init __free_pages_boot_core(struct page *page,
984 unsigned long pfn, unsigned int order)
986 unsigned int nr_pages = 1 << order;
987 struct page *p = page;
991 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
993 __ClearPageReserved(p);
994 set_page_count(p, 0);
996 __ClearPageReserved(p);
997 set_page_count(p, 0);
999 page_zone(page)->managed_pages += nr_pages;
1000 set_page_refcounted(page);
1001 __free_pages(page, order);
1004 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1005 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1007 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1009 int __meminit early_pfn_to_nid(unsigned long pfn)
1011 static DEFINE_SPINLOCK(early_pfn_lock);
1014 spin_lock(&early_pfn_lock);
1015 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1018 spin_unlock(&early_pfn_lock);
1024 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1025 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1026 struct mminit_pfnnid_cache *state)
1030 nid = __early_pfn_to_nid(pfn, state);
1031 if (nid >= 0 && nid != node)
1036 /* Only safe to use early in boot when initialisation is single-threaded */
1037 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1039 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1044 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1048 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1049 struct mminit_pfnnid_cache *state)
1056 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1059 if (early_page_uninitialised(pfn))
1061 return __free_pages_boot_core(page, pfn, order);
1064 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1065 static void __init deferred_free_range(struct page *page,
1066 unsigned long pfn, int nr_pages)
1073 /* Free a large naturally-aligned chunk if possible */
1074 if (nr_pages == MAX_ORDER_NR_PAGES &&
1075 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1076 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1077 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1081 for (i = 0; i < nr_pages; i++, page++, pfn++)
1082 __free_pages_boot_core(page, pfn, 0);
1085 /* Completion tracking for deferred_init_memmap() threads */
1086 static atomic_t pgdat_init_n_undone __initdata;
1087 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1089 static inline void __init pgdat_init_report_one_done(void)
1091 if (atomic_dec_and_test(&pgdat_init_n_undone))
1092 complete(&pgdat_init_all_done_comp);
1095 /* Initialise remaining memory on a node */
1096 static int __init deferred_init_memmap(void *data)
1098 pg_data_t *pgdat = data;
1099 int nid = pgdat->node_id;
1100 struct mminit_pfnnid_cache nid_init_state = { };
1101 unsigned long start = jiffies;
1102 unsigned long nr_pages = 0;
1103 unsigned long walk_start, walk_end;
1106 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1107 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1109 if (first_init_pfn == ULONG_MAX) {
1110 pgdat_init_report_one_done();
1114 /* Bind memory initialisation thread to a local node if possible */
1115 if (!cpumask_empty(cpumask))
1116 set_cpus_allowed_ptr(current, cpumask);
1118 /* Sanity check boundaries */
1119 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1120 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1121 pgdat->first_deferred_pfn = ULONG_MAX;
1123 /* Only the highest zone is deferred so find it */
1124 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1125 zone = pgdat->node_zones + zid;
1126 if (first_init_pfn < zone_end_pfn(zone))
1130 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1131 unsigned long pfn, end_pfn;
1132 struct page *page = NULL;
1133 struct page *free_base_page = NULL;
1134 unsigned long free_base_pfn = 0;
1137 end_pfn = min(walk_end, zone_end_pfn(zone));
1138 pfn = first_init_pfn;
1139 if (pfn < walk_start)
1141 if (pfn < zone->zone_start_pfn)
1142 pfn = zone->zone_start_pfn;
1144 for (; pfn < end_pfn; pfn++) {
1145 if (!pfn_valid_within(pfn))
1149 * Ensure pfn_valid is checked every
1150 * MAX_ORDER_NR_PAGES for memory holes
1152 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1153 if (!pfn_valid(pfn)) {
1159 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1164 /* Minimise pfn page lookups and scheduler checks */
1165 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1168 nr_pages += nr_to_free;
1169 deferred_free_range(free_base_page,
1170 free_base_pfn, nr_to_free);
1171 free_base_page = NULL;
1172 free_base_pfn = nr_to_free = 0;
1174 page = pfn_to_page(pfn);
1179 VM_BUG_ON(page_zone(page) != zone);
1183 __init_single_page(page, pfn, zid, nid);
1184 if (!free_base_page) {
1185 free_base_page = page;
1186 free_base_pfn = pfn;
1191 /* Where possible, batch up pages for a single free */
1194 /* Free the current block of pages to allocator */
1195 nr_pages += nr_to_free;
1196 deferred_free_range(free_base_page, free_base_pfn,
1198 free_base_page = NULL;
1199 free_base_pfn = nr_to_free = 0;
1202 first_init_pfn = max(end_pfn, first_init_pfn);
1205 /* Sanity check that the next zone really is unpopulated */
1206 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1208 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1209 jiffies_to_msecs(jiffies - start));
1211 pgdat_init_report_one_done();
1215 void __init page_alloc_init_late(void)
1219 /* There will be num_node_state(N_MEMORY) threads */
1220 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1221 for_each_node_state(nid, N_MEMORY) {
1222 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1225 /* Block until all are initialised */
1226 wait_for_completion(&pgdat_init_all_done_comp);
1228 /* Reinit limits that are based on free pages after the kernel is up */
1229 files_maxfiles_init();
1231 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1234 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1235 void __init init_cma_reserved_pageblock(struct page *page)
1237 unsigned i = pageblock_nr_pages;
1238 struct page *p = page;
1241 __ClearPageReserved(p);
1242 set_page_count(p, 0);
1245 set_pageblock_migratetype(page, MIGRATE_CMA);
1247 if (pageblock_order >= MAX_ORDER) {
1248 i = pageblock_nr_pages;
1251 set_page_refcounted(p);
1252 __free_pages(p, MAX_ORDER - 1);
1253 p += MAX_ORDER_NR_PAGES;
1254 } while (i -= MAX_ORDER_NR_PAGES);
1256 set_page_refcounted(page);
1257 __free_pages(page, pageblock_order);
1260 adjust_managed_page_count(page, pageblock_nr_pages);
1265 * The order of subdivision here is critical for the IO subsystem.
1266 * Please do not alter this order without good reasons and regression
1267 * testing. Specifically, as large blocks of memory are subdivided,
1268 * the order in which smaller blocks are delivered depends on the order
1269 * they're subdivided in this function. This is the primary factor
1270 * influencing the order in which pages are delivered to the IO
1271 * subsystem according to empirical testing, and this is also justified
1272 * by considering the behavior of a buddy system containing a single
1273 * large block of memory acted on by a series of small allocations.
1274 * This behavior is a critical factor in sglist merging's success.
1278 static inline void expand(struct zone *zone, struct page *page,
1279 int low, int high, struct free_area *area,
1282 unsigned long size = 1 << high;
1284 while (high > low) {
1288 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1290 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1291 debug_guardpage_enabled() &&
1292 high < debug_guardpage_minorder()) {
1294 * Mark as guard pages (or page), that will allow to
1295 * merge back to allocator when buddy will be freed.
1296 * Corresponding page table entries will not be touched,
1297 * pages will stay not present in virtual address space
1299 set_page_guard(zone, &page[size], high, migratetype);
1302 list_add(&page[size].lru, &area->free_list[migratetype]);
1304 set_page_order(&page[size], high);
1309 * This page is about to be returned from the page allocator
1311 static inline int check_new_page(struct page *page)
1313 const char *bad_reason = NULL;
1314 unsigned long bad_flags = 0;
1316 if (unlikely(page_mapcount(page)))
1317 bad_reason = "nonzero mapcount";
1318 if (unlikely(page->mapping != NULL))
1319 bad_reason = "non-NULL mapping";
1320 if (unlikely(atomic_read(&page->_count) != 0))
1321 bad_reason = "nonzero _count";
1322 if (unlikely(page->flags & __PG_HWPOISON)) {
1323 bad_reason = "HWPoisoned (hardware-corrupted)";
1324 bad_flags = __PG_HWPOISON;
1326 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1327 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1328 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1331 if (unlikely(page->mem_cgroup))
1332 bad_reason = "page still charged to cgroup";
1334 if (unlikely(bad_reason)) {
1335 bad_page(page, bad_reason, bad_flags);
1341 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1346 for (i = 0; i < (1 << order); i++) {
1347 struct page *p = page + i;
1348 if (unlikely(check_new_page(p)))
1352 set_page_private(page, 0);
1353 set_page_refcounted(page);
1355 arch_alloc_page(page, order);
1356 kernel_map_pages(page, 1 << order, 1);
1357 kasan_alloc_pages(page, order);
1359 if (gfp_flags & __GFP_ZERO)
1360 for (i = 0; i < (1 << order); i++)
1361 clear_highpage(page + i);
1363 if (order && (gfp_flags & __GFP_COMP))
1364 prep_compound_page(page, order);
1366 set_page_owner(page, order, gfp_flags);
1369 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1370 * allocate the page. The expectation is that the caller is taking
1371 * steps that will free more memory. The caller should avoid the page
1372 * being used for !PFMEMALLOC purposes.
1374 if (alloc_flags & ALLOC_NO_WATERMARKS)
1375 set_page_pfmemalloc(page);
1377 clear_page_pfmemalloc(page);
1383 * Go through the free lists for the given migratetype and remove
1384 * the smallest available page from the freelists
1387 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1390 unsigned int current_order;
1391 struct free_area *area;
1394 /* Find a page of the appropriate size in the preferred list */
1395 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1396 area = &(zone->free_area[current_order]);
1397 if (list_empty(&area->free_list[migratetype]))
1400 page = list_entry(area->free_list[migratetype].next,
1402 list_del(&page->lru);
1403 rmv_page_order(page);
1405 expand(zone, page, order, current_order, area, migratetype);
1406 set_pcppage_migratetype(page, migratetype);
1415 * This array describes the order lists are fallen back to when
1416 * the free lists for the desirable migrate type are depleted
1418 static int fallbacks[MIGRATE_TYPES][4] = {
1419 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1420 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1421 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1423 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1425 #ifdef CONFIG_MEMORY_ISOLATION
1426 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1431 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1434 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1437 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1438 unsigned int order) { return NULL; }
1442 * Move the free pages in a range to the free lists of the requested type.
1443 * Note that start_page and end_pages are not aligned on a pageblock
1444 * boundary. If alignment is required, use move_freepages_block()
1446 int move_freepages(struct zone *zone,
1447 struct page *start_page, struct page *end_page,
1451 unsigned long order;
1452 int pages_moved = 0;
1454 #ifndef CONFIG_HOLES_IN_ZONE
1456 * page_zone is not safe to call in this context when
1457 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1458 * anyway as we check zone boundaries in move_freepages_block().
1459 * Remove at a later date when no bug reports exist related to
1460 * grouping pages by mobility
1462 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1465 for (page = start_page; page <= end_page;) {
1466 /* Make sure we are not inadvertently changing nodes */
1467 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1469 if (!pfn_valid_within(page_to_pfn(page))) {
1474 if (!PageBuddy(page)) {
1479 order = page_order(page);
1480 list_move(&page->lru,
1481 &zone->free_area[order].free_list[migratetype]);
1483 pages_moved += 1 << order;
1489 int move_freepages_block(struct zone *zone, struct page *page,
1492 unsigned long start_pfn, end_pfn;
1493 struct page *start_page, *end_page;
1495 start_pfn = page_to_pfn(page);
1496 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1497 start_page = pfn_to_page(start_pfn);
1498 end_page = start_page + pageblock_nr_pages - 1;
1499 end_pfn = start_pfn + pageblock_nr_pages - 1;
1501 /* Do not cross zone boundaries */
1502 if (!zone_spans_pfn(zone, start_pfn))
1504 if (!zone_spans_pfn(zone, end_pfn))
1507 return move_freepages(zone, start_page, end_page, migratetype);
1510 static void change_pageblock_range(struct page *pageblock_page,
1511 int start_order, int migratetype)
1513 int nr_pageblocks = 1 << (start_order - pageblock_order);
1515 while (nr_pageblocks--) {
1516 set_pageblock_migratetype(pageblock_page, migratetype);
1517 pageblock_page += pageblock_nr_pages;
1522 * When we are falling back to another migratetype during allocation, try to
1523 * steal extra free pages from the same pageblocks to satisfy further
1524 * allocations, instead of polluting multiple pageblocks.
1526 * If we are stealing a relatively large buddy page, it is likely there will
1527 * be more free pages in the pageblock, so try to steal them all. For
1528 * reclaimable and unmovable allocations, we steal regardless of page size,
1529 * as fragmentation caused by those allocations polluting movable pageblocks
1530 * is worse than movable allocations stealing from unmovable and reclaimable
1533 static bool can_steal_fallback(unsigned int order, int start_mt)
1536 * Leaving this order check is intended, although there is
1537 * relaxed order check in next check. The reason is that
1538 * we can actually steal whole pageblock if this condition met,
1539 * but, below check doesn't guarantee it and that is just heuristic
1540 * so could be changed anytime.
1542 if (order >= pageblock_order)
1545 if (order >= pageblock_order / 2 ||
1546 start_mt == MIGRATE_RECLAIMABLE ||
1547 start_mt == MIGRATE_UNMOVABLE ||
1548 page_group_by_mobility_disabled)
1555 * This function implements actual steal behaviour. If order is large enough,
1556 * we can steal whole pageblock. If not, we first move freepages in this
1557 * pageblock and check whether half of pages are moved or not. If half of
1558 * pages are moved, we can change migratetype of pageblock and permanently
1559 * use it's pages as requested migratetype in the future.
1561 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1564 int current_order = page_order(page);
1567 /* Take ownership for orders >= pageblock_order */
1568 if (current_order >= pageblock_order) {
1569 change_pageblock_range(page, current_order, start_type);
1573 pages = move_freepages_block(zone, page, start_type);
1575 /* Claim the whole block if over half of it is free */
1576 if (pages >= (1 << (pageblock_order-1)) ||
1577 page_group_by_mobility_disabled)
1578 set_pageblock_migratetype(page, start_type);
1582 * Check whether there is a suitable fallback freepage with requested order.
1583 * If only_stealable is true, this function returns fallback_mt only if
1584 * we can steal other freepages all together. This would help to reduce
1585 * fragmentation due to mixed migratetype pages in one pageblock.
1587 int find_suitable_fallback(struct free_area *area, unsigned int order,
1588 int migratetype, bool only_stealable, bool *can_steal)
1593 if (area->nr_free == 0)
1598 fallback_mt = fallbacks[migratetype][i];
1599 if (fallback_mt == MIGRATE_TYPES)
1602 if (list_empty(&area->free_list[fallback_mt]))
1605 if (can_steal_fallback(order, migratetype))
1608 if (!only_stealable)
1619 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1620 * there are no empty page blocks that contain a page with a suitable order
1622 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1623 unsigned int alloc_order)
1626 unsigned long max_managed, flags;
1629 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1630 * Check is race-prone but harmless.
1632 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1633 if (zone->nr_reserved_highatomic >= max_managed)
1636 spin_lock_irqsave(&zone->lock, flags);
1638 /* Recheck the nr_reserved_highatomic limit under the lock */
1639 if (zone->nr_reserved_highatomic >= max_managed)
1643 mt = get_pageblock_migratetype(page);
1644 if (mt != MIGRATE_HIGHATOMIC &&
1645 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1646 zone->nr_reserved_highatomic += pageblock_nr_pages;
1647 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1648 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1652 spin_unlock_irqrestore(&zone->lock, flags);
1656 * Used when an allocation is about to fail under memory pressure. This
1657 * potentially hurts the reliability of high-order allocations when under
1658 * intense memory pressure but failed atomic allocations should be easier
1659 * to recover from than an OOM.
1661 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1663 struct zonelist *zonelist = ac->zonelist;
1664 unsigned long flags;
1670 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1672 /* Preserve at least one pageblock */
1673 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1676 spin_lock_irqsave(&zone->lock, flags);
1677 for (order = 0; order < MAX_ORDER; order++) {
1678 struct free_area *area = &(zone->free_area[order]);
1680 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1683 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1687 * It should never happen but changes to locking could
1688 * inadvertently allow a per-cpu drain to add pages
1689 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1690 * and watch for underflows.
1692 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1693 zone->nr_reserved_highatomic);
1696 * Convert to ac->migratetype and avoid the normal
1697 * pageblock stealing heuristics. Minimally, the caller
1698 * is doing the work and needs the pages. More
1699 * importantly, if the block was always converted to
1700 * MIGRATE_UNMOVABLE or another type then the number
1701 * of pageblocks that cannot be completely freed
1704 set_pageblock_migratetype(page, ac->migratetype);
1705 move_freepages_block(zone, page, ac->migratetype);
1706 spin_unlock_irqrestore(&zone->lock, flags);
1709 spin_unlock_irqrestore(&zone->lock, flags);
1713 /* Remove an element from the buddy allocator from the fallback list */
1714 static inline struct page *
1715 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1717 struct free_area *area;
1718 unsigned int current_order;
1723 /* Find the largest possible block of pages in the other list */
1724 for (current_order = MAX_ORDER-1;
1725 current_order >= order && current_order <= MAX_ORDER-1;
1727 area = &(zone->free_area[current_order]);
1728 fallback_mt = find_suitable_fallback(area, current_order,
1729 start_migratetype, false, &can_steal);
1730 if (fallback_mt == -1)
1733 page = list_entry(area->free_list[fallback_mt].next,
1736 steal_suitable_fallback(zone, page, start_migratetype);
1738 /* Remove the page from the freelists */
1740 list_del(&page->lru);
1741 rmv_page_order(page);
1743 expand(zone, page, order, current_order, area,
1746 * The pcppage_migratetype may differ from pageblock's
1747 * migratetype depending on the decisions in
1748 * find_suitable_fallback(). This is OK as long as it does not
1749 * differ for MIGRATE_CMA pageblocks. Those can be used as
1750 * fallback only via special __rmqueue_cma_fallback() function
1752 set_pcppage_migratetype(page, start_migratetype);
1754 trace_mm_page_alloc_extfrag(page, order, current_order,
1755 start_migratetype, fallback_mt);
1764 * Do the hard work of removing an element from the buddy allocator.
1765 * Call me with the zone->lock already held.
1767 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1768 int migratetype, gfp_t gfp_flags)
1772 page = __rmqueue_smallest(zone, order, migratetype);
1773 if (unlikely(!page)) {
1774 if (migratetype == MIGRATE_MOVABLE)
1775 page = __rmqueue_cma_fallback(zone, order);
1778 page = __rmqueue_fallback(zone, order, migratetype);
1781 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1786 * Obtain a specified number of elements from the buddy allocator, all under
1787 * a single hold of the lock, for efficiency. Add them to the supplied list.
1788 * Returns the number of new pages which were placed at *list.
1790 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1791 unsigned long count, struct list_head *list,
1792 int migratetype, bool cold)
1796 spin_lock(&zone->lock);
1797 for (i = 0; i < count; ++i) {
1798 struct page *page = __rmqueue(zone, order, migratetype, 0);
1799 if (unlikely(page == NULL))
1803 * Split buddy pages returned by expand() are received here
1804 * in physical page order. The page is added to the callers and
1805 * list and the list head then moves forward. From the callers
1806 * perspective, the linked list is ordered by page number in
1807 * some conditions. This is useful for IO devices that can
1808 * merge IO requests if the physical pages are ordered
1812 list_add(&page->lru, list);
1814 list_add_tail(&page->lru, list);
1816 if (is_migrate_cma(get_pcppage_migratetype(page)))
1817 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1820 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1821 spin_unlock(&zone->lock);
1827 * Called from the vmstat counter updater to drain pagesets of this
1828 * currently executing processor on remote nodes after they have
1831 * Note that this function must be called with the thread pinned to
1832 * a single processor.
1834 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1836 unsigned long flags;
1837 int to_drain, batch;
1839 local_irq_save(flags);
1840 batch = READ_ONCE(pcp->batch);
1841 to_drain = min(pcp->count, batch);
1843 free_pcppages_bulk(zone, to_drain, pcp);
1844 pcp->count -= to_drain;
1846 local_irq_restore(flags);
1851 * Drain pcplists of the indicated processor and zone.
1853 * The processor must either be the current processor and the
1854 * thread pinned to the current processor or a processor that
1857 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1859 unsigned long flags;
1860 struct per_cpu_pageset *pset;
1861 struct per_cpu_pages *pcp;
1863 local_irq_save(flags);
1864 pset = per_cpu_ptr(zone->pageset, cpu);
1868 free_pcppages_bulk(zone, pcp->count, pcp);
1871 local_irq_restore(flags);
1875 * Drain pcplists of all zones on the indicated processor.
1877 * The processor must either be the current processor and the
1878 * thread pinned to the current processor or a processor that
1881 static void drain_pages(unsigned int cpu)
1885 for_each_populated_zone(zone) {
1886 drain_pages_zone(cpu, zone);
1891 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1893 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1894 * the single zone's pages.
1896 void drain_local_pages(struct zone *zone)
1898 int cpu = smp_processor_id();
1901 drain_pages_zone(cpu, zone);
1907 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1909 * When zone parameter is non-NULL, spill just the single zone's pages.
1911 * Note that this code is protected against sending an IPI to an offline
1912 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1913 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1914 * nothing keeps CPUs from showing up after we populated the cpumask and
1915 * before the call to on_each_cpu_mask().
1917 void drain_all_pages(struct zone *zone)
1922 * Allocate in the BSS so we wont require allocation in
1923 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1925 static cpumask_t cpus_with_pcps;
1928 * We don't care about racing with CPU hotplug event
1929 * as offline notification will cause the notified
1930 * cpu to drain that CPU pcps and on_each_cpu_mask
1931 * disables preemption as part of its processing
1933 for_each_online_cpu(cpu) {
1934 struct per_cpu_pageset *pcp;
1936 bool has_pcps = false;
1939 pcp = per_cpu_ptr(zone->pageset, cpu);
1943 for_each_populated_zone(z) {
1944 pcp = per_cpu_ptr(z->pageset, cpu);
1945 if (pcp->pcp.count) {
1953 cpumask_set_cpu(cpu, &cpus_with_pcps);
1955 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1957 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1961 #ifdef CONFIG_HIBERNATION
1963 void mark_free_pages(struct zone *zone)
1965 unsigned long pfn, max_zone_pfn;
1966 unsigned long flags;
1967 unsigned int order, t;
1968 struct list_head *curr;
1970 if (zone_is_empty(zone))
1973 spin_lock_irqsave(&zone->lock, flags);
1975 max_zone_pfn = zone_end_pfn(zone);
1976 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1977 if (pfn_valid(pfn)) {
1978 struct page *page = pfn_to_page(pfn);
1980 if (!swsusp_page_is_forbidden(page))
1981 swsusp_unset_page_free(page);
1984 for_each_migratetype_order(order, t) {
1985 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1988 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1989 for (i = 0; i < (1UL << order); i++)
1990 swsusp_set_page_free(pfn_to_page(pfn + i));
1993 spin_unlock_irqrestore(&zone->lock, flags);
1995 #endif /* CONFIG_PM */
1998 * Free a 0-order page
1999 * cold == true ? free a cold page : free a hot page
2001 void free_hot_cold_page(struct page *page, bool cold)
2003 struct zone *zone = page_zone(page);
2004 struct per_cpu_pages *pcp;
2005 unsigned long flags;
2006 unsigned long pfn = page_to_pfn(page);
2009 if (!free_pages_prepare(page, 0))
2012 migratetype = get_pfnblock_migratetype(page, pfn);
2013 set_pcppage_migratetype(page, migratetype);
2014 local_irq_save(flags);
2015 __count_vm_event(PGFREE);
2018 * We only track unmovable, reclaimable and movable on pcp lists.
2019 * Free ISOLATE pages back to the allocator because they are being
2020 * offlined but treat RESERVE as movable pages so we can get those
2021 * areas back if necessary. Otherwise, we may have to free
2022 * excessively into the page allocator
2024 if (migratetype >= MIGRATE_PCPTYPES) {
2025 if (unlikely(is_migrate_isolate(migratetype))) {
2026 free_one_page(zone, page, pfn, 0, migratetype);
2029 migratetype = MIGRATE_MOVABLE;
2032 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2034 list_add(&page->lru, &pcp->lists[migratetype]);
2036 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2038 if (pcp->count >= pcp->high) {
2039 unsigned long batch = READ_ONCE(pcp->batch);
2040 free_pcppages_bulk(zone, batch, pcp);
2041 pcp->count -= batch;
2045 local_irq_restore(flags);
2049 * Free a list of 0-order pages
2051 void free_hot_cold_page_list(struct list_head *list, bool cold)
2053 struct page *page, *next;
2055 list_for_each_entry_safe(page, next, list, lru) {
2056 trace_mm_page_free_batched(page, cold);
2057 free_hot_cold_page(page, cold);
2062 * split_page takes a non-compound higher-order page, and splits it into
2063 * n (1<<order) sub-pages: page[0..n]
2064 * Each sub-page must be freed individually.
2066 * Note: this is probably too low level an operation for use in drivers.
2067 * Please consult with lkml before using this in your driver.
2069 void split_page(struct page *page, unsigned int order)
2074 VM_BUG_ON_PAGE(PageCompound(page), page);
2075 VM_BUG_ON_PAGE(!page_count(page), page);
2077 #ifdef CONFIG_KMEMCHECK
2079 * Split shadow pages too, because free(page[0]) would
2080 * otherwise free the whole shadow.
2082 if (kmemcheck_page_is_tracked(page))
2083 split_page(virt_to_page(page[0].shadow), order);
2086 gfp_mask = get_page_owner_gfp(page);
2087 set_page_owner(page, 0, gfp_mask);
2088 for (i = 1; i < (1 << order); i++) {
2089 set_page_refcounted(page + i);
2090 set_page_owner(page + i, 0, gfp_mask);
2093 EXPORT_SYMBOL_GPL(split_page);
2095 int __isolate_free_page(struct page *page, unsigned int order)
2097 unsigned long watermark;
2101 BUG_ON(!PageBuddy(page));
2103 zone = page_zone(page);
2104 mt = get_pageblock_migratetype(page);
2106 if (!is_migrate_isolate(mt)) {
2107 /* Obey watermarks as if the page was being allocated */
2108 watermark = low_wmark_pages(zone) + (1 << order);
2109 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2112 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2115 /* Remove page from free list */
2116 list_del(&page->lru);
2117 zone->free_area[order].nr_free--;
2118 rmv_page_order(page);
2120 set_page_owner(page, order, __GFP_MOVABLE);
2122 /* Set the pageblock if the isolated page is at least a pageblock */
2123 if (order >= pageblock_order - 1) {
2124 struct page *endpage = page + (1 << order) - 1;
2125 for (; page < endpage; page += pageblock_nr_pages) {
2126 int mt = get_pageblock_migratetype(page);
2127 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2128 set_pageblock_migratetype(page,
2134 return 1UL << order;
2138 * Similar to split_page except the page is already free. As this is only
2139 * being used for migration, the migratetype of the block also changes.
2140 * As this is called with interrupts disabled, the caller is responsible
2141 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2144 * Note: this is probably too low level an operation for use in drivers.
2145 * Please consult with lkml before using this in your driver.
2147 int split_free_page(struct page *page)
2152 order = page_order(page);
2154 nr_pages = __isolate_free_page(page, order);
2158 /* Split into individual pages */
2159 set_page_refcounted(page);
2160 split_page(page, order);
2165 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2168 struct page *buffered_rmqueue(struct zone *preferred_zone,
2169 struct zone *zone, unsigned int order,
2170 gfp_t gfp_flags, int alloc_flags, int migratetype)
2172 unsigned long flags;
2174 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2176 if (likely(order == 0)) {
2177 struct per_cpu_pages *pcp;
2178 struct list_head *list;
2180 local_irq_save(flags);
2181 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2182 list = &pcp->lists[migratetype];
2183 if (list_empty(list)) {
2184 pcp->count += rmqueue_bulk(zone, 0,
2187 if (unlikely(list_empty(list)))
2192 page = list_entry(list->prev, struct page, lru);
2194 page = list_entry(list->next, struct page, lru);
2196 list_del(&page->lru);
2199 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2201 * __GFP_NOFAIL is not to be used in new code.
2203 * All __GFP_NOFAIL callers should be fixed so that they
2204 * properly detect and handle allocation failures.
2206 * We most definitely don't want callers attempting to
2207 * allocate greater than order-1 page units with
2210 WARN_ON_ONCE(order > 1);
2212 spin_lock_irqsave(&zone->lock, flags);
2215 if (unlikely(order) && (alloc_flags & ALLOC_HARDER)) {
2216 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2218 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2221 page = __rmqueue(zone, order, migratetype, gfp_flags);
2222 spin_unlock(&zone->lock);
2225 __mod_zone_freepage_state(zone, -(1 << order),
2226 get_pcppage_migratetype(page));
2229 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2230 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2231 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2232 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2234 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2235 zone_statistics(preferred_zone, zone, gfp_flags);
2236 local_irq_restore(flags);
2238 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2242 local_irq_restore(flags);
2246 #ifdef CONFIG_FAIL_PAGE_ALLOC
2249 struct fault_attr attr;
2251 u32 ignore_gfp_highmem;
2252 u32 ignore_gfp_reclaim;
2254 } fail_page_alloc = {
2255 .attr = FAULT_ATTR_INITIALIZER,
2256 .ignore_gfp_reclaim = 1,
2257 .ignore_gfp_highmem = 1,
2261 static int __init setup_fail_page_alloc(char *str)
2263 return setup_fault_attr(&fail_page_alloc.attr, str);
2265 __setup("fail_page_alloc=", setup_fail_page_alloc);
2267 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2269 if (order < fail_page_alloc.min_order)
2271 if (gfp_mask & __GFP_NOFAIL)
2273 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2275 if (fail_page_alloc.ignore_gfp_reclaim &&
2276 (gfp_mask & __GFP_DIRECT_RECLAIM))
2279 return should_fail(&fail_page_alloc.attr, 1 << order);
2282 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2284 static int __init fail_page_alloc_debugfs(void)
2286 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2289 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2290 &fail_page_alloc.attr);
2292 return PTR_ERR(dir);
2294 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2295 &fail_page_alloc.ignore_gfp_reclaim))
2297 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2298 &fail_page_alloc.ignore_gfp_highmem))
2300 if (!debugfs_create_u32("min-order", mode, dir,
2301 &fail_page_alloc.min_order))
2306 debugfs_remove_recursive(dir);
2311 late_initcall(fail_page_alloc_debugfs);
2313 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2315 #else /* CONFIG_FAIL_PAGE_ALLOC */
2317 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2322 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2325 * Return true if free pages are above 'mark'. This takes into account the order
2326 * of the allocation.
2328 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2329 unsigned long mark, int classzone_idx, int alloc_flags,
2336 /* free_pages may go negative - that's OK */
2337 free_pages -= (1 << order) - 1;
2339 if (alloc_flags & ALLOC_HIGH)
2343 * If the caller does not have rights to ALLOC_HARDER then subtract
2344 * the high-atomic reserves. This will over-estimate the size of the
2345 * atomic reserve but it avoids a search.
2347 if (likely(!(alloc_flags & ALLOC_HARDER)))
2348 free_pages -= z->nr_reserved_highatomic;
2353 /* If allocation can't use CMA areas don't use free CMA pages */
2354 if (!(alloc_flags & ALLOC_CMA))
2355 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
2358 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
2360 for (o = 0; o < order; o++) {
2361 /* At the next order, this order's pages become unavailable */
2362 free_pages -= z->free_area[o].nr_free << o;
2364 /* Require fewer higher order pages to be free */
2367 if (free_pages <= min)
2373 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2374 int classzone_idx, int alloc_flags)
2376 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2377 zone_page_state(z, NR_FREE_PAGES));
2380 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2381 unsigned long mark, int classzone_idx)
2383 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2385 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2386 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2388 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2393 static bool zone_local(struct zone *local_zone, struct zone *zone)
2395 return local_zone->node == zone->node;
2398 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2400 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2403 #else /* CONFIG_NUMA */
2404 static bool zone_local(struct zone *local_zone, struct zone *zone)
2409 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2413 #endif /* CONFIG_NUMA */
2415 static void reset_alloc_batches(struct zone *preferred_zone)
2417 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2420 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2421 high_wmark_pages(zone) - low_wmark_pages(zone) -
2422 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2423 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2424 } while (zone++ != preferred_zone);
2428 * get_page_from_freelist goes through the zonelist trying to allocate
2431 static struct page *
2432 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2433 const struct alloc_context *ac)
2435 struct zonelist *zonelist = ac->zonelist;
2437 struct page *page = NULL;
2439 int nr_fair_skipped = 0;
2440 bool zonelist_rescan;
2443 zonelist_rescan = false;
2446 * Scan zonelist, looking for a zone with enough free.
2447 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2449 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2453 if (cpusets_enabled() &&
2454 (alloc_flags & ALLOC_CPUSET) &&
2455 !cpuset_zone_allowed(zone, gfp_mask))
2458 * Distribute pages in proportion to the individual
2459 * zone size to ensure fair page aging. The zone a
2460 * page was allocated in should have no effect on the
2461 * time the page has in memory before being reclaimed.
2463 if (alloc_flags & ALLOC_FAIR) {
2464 if (!zone_local(ac->preferred_zone, zone))
2466 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2472 * When allocating a page cache page for writing, we
2473 * want to get it from a zone that is within its dirty
2474 * limit, such that no single zone holds more than its
2475 * proportional share of globally allowed dirty pages.
2476 * The dirty limits take into account the zone's
2477 * lowmem reserves and high watermark so that kswapd
2478 * should be able to balance it without having to
2479 * write pages from its LRU list.
2481 * This may look like it could increase pressure on
2482 * lower zones by failing allocations in higher zones
2483 * before they are full. But the pages that do spill
2484 * over are limited as the lower zones are protected
2485 * by this very same mechanism. It should not become
2486 * a practical burden to them.
2488 * XXX: For now, allow allocations to potentially
2489 * exceed the per-zone dirty limit in the slowpath
2490 * (spread_dirty_pages unset) before going into reclaim,
2491 * which is important when on a NUMA setup the allowed
2492 * zones are together not big enough to reach the
2493 * global limit. The proper fix for these situations
2494 * will require awareness of zones in the
2495 * dirty-throttling and the flusher threads.
2497 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2500 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2501 if (!zone_watermark_ok(zone, order, mark,
2502 ac->classzone_idx, alloc_flags)) {
2505 /* Checked here to keep the fast path fast */
2506 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2507 if (alloc_flags & ALLOC_NO_WATERMARKS)
2510 if (zone_reclaim_mode == 0 ||
2511 !zone_allows_reclaim(ac->preferred_zone, zone))
2514 ret = zone_reclaim(zone, gfp_mask, order);
2516 case ZONE_RECLAIM_NOSCAN:
2519 case ZONE_RECLAIM_FULL:
2520 /* scanned but unreclaimable */
2523 /* did we reclaim enough */
2524 if (zone_watermark_ok(zone, order, mark,
2525 ac->classzone_idx, alloc_flags))
2533 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2534 gfp_mask, alloc_flags, ac->migratetype);
2536 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2540 * If this is a high-order atomic allocation then check
2541 * if the pageblock should be reserved for the future
2543 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2544 reserve_highatomic_pageblock(page, zone, order);
2551 * The first pass makes sure allocations are spread fairly within the
2552 * local node. However, the local node might have free pages left
2553 * after the fairness batches are exhausted, and remote zones haven't
2554 * even been considered yet. Try once more without fairness, and
2555 * include remote zones now, before entering the slowpath and waking
2556 * kswapd: prefer spilling to a remote zone over swapping locally.
2558 if (alloc_flags & ALLOC_FAIR) {
2559 alloc_flags &= ~ALLOC_FAIR;
2560 if (nr_fair_skipped) {
2561 zonelist_rescan = true;
2562 reset_alloc_batches(ac->preferred_zone);
2564 if (nr_online_nodes > 1)
2565 zonelist_rescan = true;
2568 if (zonelist_rescan)
2575 * Large machines with many possible nodes should not always dump per-node
2576 * meminfo in irq context.
2578 static inline bool should_suppress_show_mem(void)
2583 ret = in_interrupt();
2588 static DEFINE_RATELIMIT_STATE(nopage_rs,
2589 DEFAULT_RATELIMIT_INTERVAL,
2590 DEFAULT_RATELIMIT_BURST);
2592 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2594 unsigned int filter = SHOW_MEM_FILTER_NODES;
2596 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2597 debug_guardpage_minorder() > 0)
2601 * This documents exceptions given to allocations in certain
2602 * contexts that are allowed to allocate outside current's set
2605 if (!(gfp_mask & __GFP_NOMEMALLOC))
2606 if (test_thread_flag(TIF_MEMDIE) ||
2607 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2608 filter &= ~SHOW_MEM_FILTER_NODES;
2609 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2610 filter &= ~SHOW_MEM_FILTER_NODES;
2613 struct va_format vaf;
2616 va_start(args, fmt);
2621 pr_warn("%pV", &vaf);
2626 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2627 current->comm, order, gfp_mask);
2630 if (!should_suppress_show_mem())
2634 static inline struct page *
2635 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2636 const struct alloc_context *ac, unsigned long *did_some_progress)
2638 struct oom_control oc = {
2639 .zonelist = ac->zonelist,
2640 .nodemask = ac->nodemask,
2641 .gfp_mask = gfp_mask,
2646 *did_some_progress = 0;
2649 * Acquire the oom lock. If that fails, somebody else is
2650 * making progress for us.
2652 if (!mutex_trylock(&oom_lock)) {
2653 *did_some_progress = 1;
2654 schedule_timeout_uninterruptible(1);
2659 * Go through the zonelist yet one more time, keep very high watermark
2660 * here, this is only to catch a parallel oom killing, we must fail if
2661 * we're still under heavy pressure.
2663 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2664 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2668 if (!(gfp_mask & __GFP_NOFAIL)) {
2669 /* Coredumps can quickly deplete all memory reserves */
2670 if (current->flags & PF_DUMPCORE)
2672 /* The OOM killer will not help higher order allocs */
2673 if (order > PAGE_ALLOC_COSTLY_ORDER)
2675 /* The OOM killer does not needlessly kill tasks for lowmem */
2676 if (ac->high_zoneidx < ZONE_NORMAL)
2678 /* The OOM killer does not compensate for IO-less reclaim */
2679 if (!(gfp_mask & __GFP_FS)) {
2681 * XXX: Page reclaim didn't yield anything,
2682 * and the OOM killer can't be invoked, but
2683 * keep looping as per tradition.
2685 *did_some_progress = 1;
2688 if (pm_suspended_storage())
2690 /* The OOM killer may not free memory on a specific node */
2691 if (gfp_mask & __GFP_THISNODE)
2694 /* Exhausted what can be done so it's blamo time */
2695 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2696 *did_some_progress = 1;
2698 mutex_unlock(&oom_lock);
2702 #ifdef CONFIG_COMPACTION
2703 /* Try memory compaction for high-order allocations before reclaim */
2704 static struct page *
2705 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2706 int alloc_flags, const struct alloc_context *ac,
2707 enum migrate_mode mode, int *contended_compaction,
2708 bool *deferred_compaction)
2710 unsigned long compact_result;
2716 current->flags |= PF_MEMALLOC;
2717 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2718 mode, contended_compaction);
2719 current->flags &= ~PF_MEMALLOC;
2721 switch (compact_result) {
2722 case COMPACT_DEFERRED:
2723 *deferred_compaction = true;
2725 case COMPACT_SKIPPED:
2732 * At least in one zone compaction wasn't deferred or skipped, so let's
2733 * count a compaction stall
2735 count_vm_event(COMPACTSTALL);
2737 page = get_page_from_freelist(gfp_mask, order,
2738 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2741 struct zone *zone = page_zone(page);
2743 zone->compact_blockskip_flush = false;
2744 compaction_defer_reset(zone, order, true);
2745 count_vm_event(COMPACTSUCCESS);
2750 * It's bad if compaction run occurs and fails. The most likely reason
2751 * is that pages exist, but not enough to satisfy watermarks.
2753 count_vm_event(COMPACTFAIL);
2760 static inline struct page *
2761 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2762 int alloc_flags, const struct alloc_context *ac,
2763 enum migrate_mode mode, int *contended_compaction,
2764 bool *deferred_compaction)
2768 #endif /* CONFIG_COMPACTION */
2770 /* Perform direct synchronous page reclaim */
2772 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2773 const struct alloc_context *ac)
2775 struct reclaim_state reclaim_state;
2780 /* We now go into synchronous reclaim */
2781 cpuset_memory_pressure_bump();
2782 current->flags |= PF_MEMALLOC;
2783 lockdep_set_current_reclaim_state(gfp_mask);
2784 reclaim_state.reclaimed_slab = 0;
2785 current->reclaim_state = &reclaim_state;
2787 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2790 current->reclaim_state = NULL;
2791 lockdep_clear_current_reclaim_state();
2792 current->flags &= ~PF_MEMALLOC;
2799 /* The really slow allocator path where we enter direct reclaim */
2800 static inline struct page *
2801 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2802 int alloc_flags, const struct alloc_context *ac,
2803 unsigned long *did_some_progress)
2805 struct page *page = NULL;
2806 bool drained = false;
2808 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2809 if (unlikely(!(*did_some_progress)))
2813 page = get_page_from_freelist(gfp_mask, order,
2814 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2817 * If an allocation failed after direct reclaim, it could be because
2818 * pages are pinned on the per-cpu lists or in high alloc reserves.
2819 * Shrink them them and try again
2821 if (!page && !drained) {
2822 unreserve_highatomic_pageblock(ac);
2823 drain_all_pages(NULL);
2832 * This is called in the allocator slow-path if the allocation request is of
2833 * sufficient urgency to ignore watermarks and take other desperate measures
2835 static inline struct page *
2836 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2837 const struct alloc_context *ac)
2842 page = get_page_from_freelist(gfp_mask, order,
2843 ALLOC_NO_WATERMARKS, ac);
2845 if (!page && gfp_mask & __GFP_NOFAIL)
2846 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2848 } while (!page && (gfp_mask & __GFP_NOFAIL));
2853 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2858 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2859 ac->high_zoneidx, ac->nodemask)
2860 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2864 gfp_to_alloc_flags(gfp_t gfp_mask)
2866 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2868 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2869 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2872 * The caller may dip into page reserves a bit more if the caller
2873 * cannot run direct reclaim, or if the caller has realtime scheduling
2874 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2875 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2877 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2879 if (gfp_mask & __GFP_ATOMIC) {
2881 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2882 * if it can't schedule.
2884 if (!(gfp_mask & __GFP_NOMEMALLOC))
2885 alloc_flags |= ALLOC_HARDER;
2887 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2888 * comment for __cpuset_node_allowed().
2890 alloc_flags &= ~ALLOC_CPUSET;
2891 } else if (unlikely(rt_task(current)) && !in_interrupt())
2892 alloc_flags |= ALLOC_HARDER;
2894 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2895 if (gfp_mask & __GFP_MEMALLOC)
2896 alloc_flags |= ALLOC_NO_WATERMARKS;
2897 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2898 alloc_flags |= ALLOC_NO_WATERMARKS;
2899 else if (!in_interrupt() &&
2900 ((current->flags & PF_MEMALLOC) ||
2901 unlikely(test_thread_flag(TIF_MEMDIE))))
2902 alloc_flags |= ALLOC_NO_WATERMARKS;
2905 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2906 alloc_flags |= ALLOC_CMA;
2911 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2913 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2916 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2918 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2921 static inline struct page *
2922 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2923 struct alloc_context *ac)
2925 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2926 struct page *page = NULL;
2928 unsigned long pages_reclaimed = 0;
2929 unsigned long did_some_progress;
2930 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2931 bool deferred_compaction = false;
2932 int contended_compaction = COMPACT_CONTENDED_NONE;
2935 * In the slowpath, we sanity check order to avoid ever trying to
2936 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2937 * be using allocators in order of preference for an area that is
2940 if (order >= MAX_ORDER) {
2941 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2946 * We also sanity check to catch abuse of atomic reserves being used by
2947 * callers that are not in atomic context.
2949 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
2950 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
2951 gfp_mask &= ~__GFP_ATOMIC;
2954 * If this allocation cannot block and it is for a specific node, then
2955 * fail early. There's no need to wakeup kswapd or retry for a
2956 * speculative node-specific allocation.
2958 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
2962 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
2963 wake_all_kswapds(order, ac);
2966 * OK, we're below the kswapd watermark and have kicked background
2967 * reclaim. Now things get more complex, so set up alloc_flags according
2968 * to how we want to proceed.
2970 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2973 * Find the true preferred zone if the allocation is unconstrained by
2976 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2977 struct zoneref *preferred_zoneref;
2978 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2979 ac->high_zoneidx, NULL, &ac->preferred_zone);
2980 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2983 /* This is the last chance, in general, before the goto nopage. */
2984 page = get_page_from_freelist(gfp_mask, order,
2985 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2989 /* Allocate without watermarks if the context allows */
2990 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2992 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2993 * the allocation is high priority and these type of
2994 * allocations are system rather than user orientated
2996 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2998 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3005 /* Caller is not willing to reclaim, we can't balance anything */
3006 if (!can_direct_reclaim) {
3008 * All existing users of the deprecated __GFP_NOFAIL are
3009 * blockable, so warn of any new users that actually allow this
3010 * type of allocation to fail.
3012 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3016 /* Avoid recursion of direct reclaim */
3017 if (current->flags & PF_MEMALLOC)
3020 /* Avoid allocations with no watermarks from looping endlessly */
3021 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3025 * Try direct compaction. The first pass is asynchronous. Subsequent
3026 * attempts after direct reclaim are synchronous
3028 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3030 &contended_compaction,
3031 &deferred_compaction);
3035 /* Checks for THP-specific high-order allocations */
3036 if (is_thp_gfp_mask(gfp_mask)) {
3038 * If compaction is deferred for high-order allocations, it is
3039 * because sync compaction recently failed. If this is the case
3040 * and the caller requested a THP allocation, we do not want
3041 * to heavily disrupt the system, so we fail the allocation
3042 * instead of entering direct reclaim.
3044 if (deferred_compaction)
3048 * In all zones where compaction was attempted (and not
3049 * deferred or skipped), lock contention has been detected.
3050 * For THP allocation we do not want to disrupt the others
3051 * so we fallback to base pages instead.
3053 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3057 * If compaction was aborted due to need_resched(), we do not
3058 * want to further increase allocation latency, unless it is
3059 * khugepaged trying to collapse.
3061 if (contended_compaction == COMPACT_CONTENDED_SCHED
3062 && !(current->flags & PF_KTHREAD))
3067 * It can become very expensive to allocate transparent hugepages at
3068 * fault, so use asynchronous memory compaction for THP unless it is
3069 * khugepaged trying to collapse.
3071 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3072 migration_mode = MIGRATE_SYNC_LIGHT;
3074 /* Try direct reclaim and then allocating */
3075 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3076 &did_some_progress);
3080 /* Do not loop if specifically requested */
3081 if (gfp_mask & __GFP_NORETRY)
3084 /* Keep reclaiming pages as long as there is reasonable progress */
3085 pages_reclaimed += did_some_progress;
3086 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3087 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3088 /* Wait for some write requests to complete then retry */
3089 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3093 /* Reclaim has failed us, start killing things */
3094 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3098 /* Retry as long as the OOM killer is making progress */
3099 if (did_some_progress)
3104 * High-order allocations do not necessarily loop after
3105 * direct reclaim and reclaim/compaction depends on compaction
3106 * being called after reclaim so call directly if necessary
3108 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3110 &contended_compaction,
3111 &deferred_compaction);
3115 warn_alloc_failed(gfp_mask, order, NULL);
3121 * This is the 'heart' of the zoned buddy allocator.
3124 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3125 struct zonelist *zonelist, nodemask_t *nodemask)
3127 struct zoneref *preferred_zoneref;
3128 struct page *page = NULL;
3129 unsigned int cpuset_mems_cookie;
3130 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3131 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3132 struct alloc_context ac = {
3133 .high_zoneidx = gfp_zone(gfp_mask),
3134 .nodemask = nodemask,
3135 .migratetype = gfpflags_to_migratetype(gfp_mask),
3138 gfp_mask &= gfp_allowed_mask;
3140 lockdep_trace_alloc(gfp_mask);
3142 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3144 if (should_fail_alloc_page(gfp_mask, order))
3148 * Check the zones suitable for the gfp_mask contain at least one
3149 * valid zone. It's possible to have an empty zonelist as a result
3150 * of __GFP_THISNODE and a memoryless node
3152 if (unlikely(!zonelist->_zonerefs->zone))
3155 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3156 alloc_flags |= ALLOC_CMA;
3159 cpuset_mems_cookie = read_mems_allowed_begin();
3161 /* We set it here, as __alloc_pages_slowpath might have changed it */
3162 ac.zonelist = zonelist;
3164 /* Dirty zone balancing only done in the fast path */
3165 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3167 /* The preferred zone is used for statistics later */
3168 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3169 ac.nodemask ? : &cpuset_current_mems_allowed,
3170 &ac.preferred_zone);
3171 if (!ac.preferred_zone)
3173 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3175 /* First allocation attempt */
3176 alloc_mask = gfp_mask|__GFP_HARDWALL;
3177 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3178 if (unlikely(!page)) {
3180 * Runtime PM, block IO and its error handling path
3181 * can deadlock because I/O on the device might not
3184 alloc_mask = memalloc_noio_flags(gfp_mask);
3185 ac.spread_dirty_pages = false;
3187 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3190 if (kmemcheck_enabled && page)
3191 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3193 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3197 * When updating a task's mems_allowed, it is possible to race with
3198 * parallel threads in such a way that an allocation can fail while
3199 * the mask is being updated. If a page allocation is about to fail,
3200 * check if the cpuset changed during allocation and if so, retry.
3202 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3207 EXPORT_SYMBOL(__alloc_pages_nodemask);
3210 * Common helper functions.
3212 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3217 * __get_free_pages() returns a 32-bit address, which cannot represent
3220 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3222 page = alloc_pages(gfp_mask, order);
3225 return (unsigned long) page_address(page);
3227 EXPORT_SYMBOL(__get_free_pages);
3229 unsigned long get_zeroed_page(gfp_t gfp_mask)
3231 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3233 EXPORT_SYMBOL(get_zeroed_page);
3235 void __free_pages(struct page *page, unsigned int order)
3237 if (put_page_testzero(page)) {
3239 free_hot_cold_page(page, false);
3241 __free_pages_ok(page, order);
3245 EXPORT_SYMBOL(__free_pages);
3247 void free_pages(unsigned long addr, unsigned int order)
3250 VM_BUG_ON(!virt_addr_valid((void *)addr));
3251 __free_pages(virt_to_page((void *)addr), order);
3255 EXPORT_SYMBOL(free_pages);
3259 * An arbitrary-length arbitrary-offset area of memory which resides
3260 * within a 0 or higher order page. Multiple fragments within that page
3261 * are individually refcounted, in the page's reference counter.
3263 * The page_frag functions below provide a simple allocation framework for
3264 * page fragments. This is used by the network stack and network device
3265 * drivers to provide a backing region of memory for use as either an
3266 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3268 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3271 struct page *page = NULL;
3272 gfp_t gfp = gfp_mask;
3274 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3275 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3277 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3278 PAGE_FRAG_CACHE_MAX_ORDER);
3279 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3281 if (unlikely(!page))
3282 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3284 nc->va = page ? page_address(page) : NULL;
3289 void *__alloc_page_frag(struct page_frag_cache *nc,
3290 unsigned int fragsz, gfp_t gfp_mask)
3292 unsigned int size = PAGE_SIZE;
3296 if (unlikely(!nc->va)) {
3298 page = __page_frag_refill(nc, gfp_mask);
3302 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3303 /* if size can vary use size else just use PAGE_SIZE */
3306 /* Even if we own the page, we do not use atomic_set().
3307 * This would break get_page_unless_zero() users.
3309 atomic_add(size - 1, &page->_count);
3311 /* reset page count bias and offset to start of new frag */
3312 nc->pfmemalloc = page_is_pfmemalloc(page);
3313 nc->pagecnt_bias = size;
3317 offset = nc->offset - fragsz;
3318 if (unlikely(offset < 0)) {
3319 page = virt_to_page(nc->va);
3321 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3324 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3325 /* if size can vary use size else just use PAGE_SIZE */
3328 /* OK, page count is 0, we can safely set it */
3329 atomic_set(&page->_count, size);
3331 /* reset page count bias and offset to start of new frag */
3332 nc->pagecnt_bias = size;
3333 offset = size - fragsz;
3337 nc->offset = offset;
3339 return nc->va + offset;
3341 EXPORT_SYMBOL(__alloc_page_frag);
3344 * Frees a page fragment allocated out of either a compound or order 0 page.
3346 void __free_page_frag(void *addr)
3348 struct page *page = virt_to_head_page(addr);
3350 if (unlikely(put_page_testzero(page)))
3351 __free_pages_ok(page, compound_order(page));
3353 EXPORT_SYMBOL(__free_page_frag);
3356 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3357 * of the current memory cgroup.
3359 * It should be used when the caller would like to use kmalloc, but since the
3360 * allocation is large, it has to fall back to the page allocator.
3362 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3365 struct mem_cgroup *memcg = NULL;
3367 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3369 page = alloc_pages(gfp_mask, order);
3370 memcg_kmem_commit_charge(page, memcg, order);
3374 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3377 struct mem_cgroup *memcg = NULL;
3379 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3381 page = alloc_pages_node(nid, gfp_mask, order);
3382 memcg_kmem_commit_charge(page, memcg, order);
3387 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3390 void __free_kmem_pages(struct page *page, unsigned int order)
3392 memcg_kmem_uncharge_pages(page, order);
3393 __free_pages(page, order);
3396 void free_kmem_pages(unsigned long addr, unsigned int order)
3399 VM_BUG_ON(!virt_addr_valid((void *)addr));
3400 __free_kmem_pages(virt_to_page((void *)addr), order);
3404 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3407 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3408 unsigned long used = addr + PAGE_ALIGN(size);
3410 split_page(virt_to_page((void *)addr), order);
3411 while (used < alloc_end) {
3416 return (void *)addr;
3420 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3421 * @size: the number of bytes to allocate
3422 * @gfp_mask: GFP flags for the allocation
3424 * This function is similar to alloc_pages(), except that it allocates the
3425 * minimum number of pages to satisfy the request. alloc_pages() can only
3426 * allocate memory in power-of-two pages.
3428 * This function is also limited by MAX_ORDER.
3430 * Memory allocated by this function must be released by free_pages_exact().
3432 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3434 unsigned int order = get_order(size);
3437 addr = __get_free_pages(gfp_mask, order);
3438 return make_alloc_exact(addr, order, size);
3440 EXPORT_SYMBOL(alloc_pages_exact);
3443 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3445 * @nid: the preferred node ID where memory should be allocated
3446 * @size: the number of bytes to allocate
3447 * @gfp_mask: GFP flags for the allocation
3449 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3452 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3454 unsigned order = get_order(size);
3455 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3458 return make_alloc_exact((unsigned long)page_address(p), order, size);
3462 * free_pages_exact - release memory allocated via alloc_pages_exact()
3463 * @virt: the value returned by alloc_pages_exact.
3464 * @size: size of allocation, same value as passed to alloc_pages_exact().
3466 * Release the memory allocated by a previous call to alloc_pages_exact.
3468 void free_pages_exact(void *virt, size_t size)
3470 unsigned long addr = (unsigned long)virt;
3471 unsigned long end = addr + PAGE_ALIGN(size);
3473 while (addr < end) {
3478 EXPORT_SYMBOL(free_pages_exact);
3481 * nr_free_zone_pages - count number of pages beyond high watermark
3482 * @offset: The zone index of the highest zone
3484 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3485 * high watermark within all zones at or below a given zone index. For each
3486 * zone, the number of pages is calculated as:
3487 * managed_pages - high_pages
3489 static unsigned long nr_free_zone_pages(int offset)
3494 /* Just pick one node, since fallback list is circular */
3495 unsigned long sum = 0;
3497 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3499 for_each_zone_zonelist(zone, z, zonelist, offset) {
3500 unsigned long size = zone->managed_pages;
3501 unsigned long high = high_wmark_pages(zone);
3510 * nr_free_buffer_pages - count number of pages beyond high watermark
3512 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3513 * watermark within ZONE_DMA and ZONE_NORMAL.
3515 unsigned long nr_free_buffer_pages(void)
3517 return nr_free_zone_pages(gfp_zone(GFP_USER));
3519 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3522 * nr_free_pagecache_pages - count number of pages beyond high watermark
3524 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3525 * high watermark within all zones.
3527 unsigned long nr_free_pagecache_pages(void)
3529 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3532 static inline void show_node(struct zone *zone)
3534 if (IS_ENABLED(CONFIG_NUMA))
3535 printk("Node %d ", zone_to_nid(zone));
3538 void si_meminfo(struct sysinfo *val)
3540 val->totalram = totalram_pages;
3541 val->sharedram = global_page_state(NR_SHMEM);
3542 val->freeram = global_page_state(NR_FREE_PAGES);
3543 val->bufferram = nr_blockdev_pages();
3544 val->totalhigh = totalhigh_pages;
3545 val->freehigh = nr_free_highpages();
3546 val->mem_unit = PAGE_SIZE;
3549 EXPORT_SYMBOL(si_meminfo);
3552 void si_meminfo_node(struct sysinfo *val, int nid)
3554 int zone_type; /* needs to be signed */
3555 unsigned long managed_pages = 0;
3556 pg_data_t *pgdat = NODE_DATA(nid);
3558 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3559 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3560 val->totalram = managed_pages;
3561 val->sharedram = node_page_state(nid, NR_SHMEM);
3562 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3563 #ifdef CONFIG_HIGHMEM
3564 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3565 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3571 val->mem_unit = PAGE_SIZE;
3576 * Determine whether the node should be displayed or not, depending on whether
3577 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3579 bool skip_free_areas_node(unsigned int flags, int nid)
3582 unsigned int cpuset_mems_cookie;
3584 if (!(flags & SHOW_MEM_FILTER_NODES))
3588 cpuset_mems_cookie = read_mems_allowed_begin();
3589 ret = !node_isset(nid, cpuset_current_mems_allowed);
3590 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3595 #define K(x) ((x) << (PAGE_SHIFT-10))
3597 static void show_migration_types(unsigned char type)
3599 static const char types[MIGRATE_TYPES] = {
3600 [MIGRATE_UNMOVABLE] = 'U',
3601 [MIGRATE_RECLAIMABLE] = 'E',
3602 [MIGRATE_MOVABLE] = 'M',
3604 [MIGRATE_CMA] = 'C',
3606 #ifdef CONFIG_MEMORY_ISOLATION
3607 [MIGRATE_ISOLATE] = 'I',
3610 char tmp[MIGRATE_TYPES + 1];
3614 for (i = 0; i < MIGRATE_TYPES; i++) {
3615 if (type & (1 << i))
3620 printk("(%s) ", tmp);
3624 * Show free area list (used inside shift_scroll-lock stuff)
3625 * We also calculate the percentage fragmentation. We do this by counting the
3626 * memory on each free list with the exception of the first item on the list.
3629 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3632 void show_free_areas(unsigned int filter)
3634 unsigned long free_pcp = 0;
3638 for_each_populated_zone(zone) {
3639 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3642 for_each_online_cpu(cpu)
3643 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3646 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3647 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3648 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3649 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3650 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3651 " free:%lu free_pcp:%lu free_cma:%lu\n",
3652 global_page_state(NR_ACTIVE_ANON),
3653 global_page_state(NR_INACTIVE_ANON),
3654 global_page_state(NR_ISOLATED_ANON),
3655 global_page_state(NR_ACTIVE_FILE),
3656 global_page_state(NR_INACTIVE_FILE),
3657 global_page_state(NR_ISOLATED_FILE),
3658 global_page_state(NR_UNEVICTABLE),
3659 global_page_state(NR_FILE_DIRTY),
3660 global_page_state(NR_WRITEBACK),
3661 global_page_state(NR_UNSTABLE_NFS),
3662 global_page_state(NR_SLAB_RECLAIMABLE),
3663 global_page_state(NR_SLAB_UNRECLAIMABLE),
3664 global_page_state(NR_FILE_MAPPED),
3665 global_page_state(NR_SHMEM),
3666 global_page_state(NR_PAGETABLE),
3667 global_page_state(NR_BOUNCE),
3668 global_page_state(NR_FREE_PAGES),
3670 global_page_state(NR_FREE_CMA_PAGES));
3672 for_each_populated_zone(zone) {
3675 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3679 for_each_online_cpu(cpu)
3680 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3688 " active_anon:%lukB"
3689 " inactive_anon:%lukB"
3690 " active_file:%lukB"
3691 " inactive_file:%lukB"
3692 " unevictable:%lukB"
3693 " isolated(anon):%lukB"
3694 " isolated(file):%lukB"
3702 " slab_reclaimable:%lukB"
3703 " slab_unreclaimable:%lukB"
3704 " kernel_stack:%lukB"
3711 " writeback_tmp:%lukB"
3712 " pages_scanned:%lu"
3713 " all_unreclaimable? %s"
3716 K(zone_page_state(zone, NR_FREE_PAGES)),
3717 K(min_wmark_pages(zone)),
3718 K(low_wmark_pages(zone)),
3719 K(high_wmark_pages(zone)),
3720 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3721 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3722 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3723 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3724 K(zone_page_state(zone, NR_UNEVICTABLE)),
3725 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3726 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3727 K(zone->present_pages),
3728 K(zone->managed_pages),
3729 K(zone_page_state(zone, NR_MLOCK)),
3730 K(zone_page_state(zone, NR_FILE_DIRTY)),
3731 K(zone_page_state(zone, NR_WRITEBACK)),
3732 K(zone_page_state(zone, NR_FILE_MAPPED)),
3733 K(zone_page_state(zone, NR_SHMEM)),
3734 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3735 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3736 zone_page_state(zone, NR_KERNEL_STACK) *
3738 K(zone_page_state(zone, NR_PAGETABLE)),
3739 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3740 K(zone_page_state(zone, NR_BOUNCE)),
3742 K(this_cpu_read(zone->pageset->pcp.count)),
3743 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3744 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3745 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3746 (!zone_reclaimable(zone) ? "yes" : "no")
3748 printk("lowmem_reserve[]:");
3749 for (i = 0; i < MAX_NR_ZONES; i++)
3750 printk(" %ld", zone->lowmem_reserve[i]);
3754 for_each_populated_zone(zone) {
3755 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3756 unsigned char types[MAX_ORDER];
3758 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3761 printk("%s: ", zone->name);
3763 spin_lock_irqsave(&zone->lock, flags);
3764 for (order = 0; order < MAX_ORDER; order++) {
3765 struct free_area *area = &zone->free_area[order];
3768 nr[order] = area->nr_free;
3769 total += nr[order] << order;
3772 for (type = 0; type < MIGRATE_TYPES; type++) {
3773 if (!list_empty(&area->free_list[type]))
3774 types[order] |= 1 << type;
3777 spin_unlock_irqrestore(&zone->lock, flags);
3778 for (order = 0; order < MAX_ORDER; order++) {
3779 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3781 show_migration_types(types[order]);
3783 printk("= %lukB\n", K(total));
3786 hugetlb_show_meminfo();
3788 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3790 show_swap_cache_info();
3793 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3795 zoneref->zone = zone;
3796 zoneref->zone_idx = zone_idx(zone);
3800 * Builds allocation fallback zone lists.
3802 * Add all populated zones of a node to the zonelist.
3804 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3808 enum zone_type zone_type = MAX_NR_ZONES;
3812 zone = pgdat->node_zones + zone_type;
3813 if (populated_zone(zone)) {
3814 zoneref_set_zone(zone,
3815 &zonelist->_zonerefs[nr_zones++]);
3816 check_highest_zone(zone_type);
3818 } while (zone_type);
3826 * 0 = automatic detection of better ordering.
3827 * 1 = order by ([node] distance, -zonetype)
3828 * 2 = order by (-zonetype, [node] distance)
3830 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3831 * the same zonelist. So only NUMA can configure this param.
3833 #define ZONELIST_ORDER_DEFAULT 0
3834 #define ZONELIST_ORDER_NODE 1
3835 #define ZONELIST_ORDER_ZONE 2
3837 /* zonelist order in the kernel.
3838 * set_zonelist_order() will set this to NODE or ZONE.
3840 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3841 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3845 /* The value user specified ....changed by config */
3846 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3847 /* string for sysctl */
3848 #define NUMA_ZONELIST_ORDER_LEN 16
3849 char numa_zonelist_order[16] = "default";
3852 * interface for configure zonelist ordering.
3853 * command line option "numa_zonelist_order"
3854 * = "[dD]efault - default, automatic configuration.
3855 * = "[nN]ode - order by node locality, then by zone within node
3856 * = "[zZ]one - order by zone, then by locality within zone
3859 static int __parse_numa_zonelist_order(char *s)
3861 if (*s == 'd' || *s == 'D') {
3862 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3863 } else if (*s == 'n' || *s == 'N') {
3864 user_zonelist_order = ZONELIST_ORDER_NODE;
3865 } else if (*s == 'z' || *s == 'Z') {
3866 user_zonelist_order = ZONELIST_ORDER_ZONE;
3869 "Ignoring invalid numa_zonelist_order value: "
3876 static __init int setup_numa_zonelist_order(char *s)
3883 ret = __parse_numa_zonelist_order(s);
3885 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3889 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3892 * sysctl handler for numa_zonelist_order
3894 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3895 void __user *buffer, size_t *length,
3898 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3900 static DEFINE_MUTEX(zl_order_mutex);
3902 mutex_lock(&zl_order_mutex);
3904 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3908 strcpy(saved_string, (char *)table->data);
3910 ret = proc_dostring(table, write, buffer, length, ppos);
3914 int oldval = user_zonelist_order;
3916 ret = __parse_numa_zonelist_order((char *)table->data);
3919 * bogus value. restore saved string
3921 strncpy((char *)table->data, saved_string,
3922 NUMA_ZONELIST_ORDER_LEN);
3923 user_zonelist_order = oldval;
3924 } else if (oldval != user_zonelist_order) {
3925 mutex_lock(&zonelists_mutex);
3926 build_all_zonelists(NULL, NULL);
3927 mutex_unlock(&zonelists_mutex);
3931 mutex_unlock(&zl_order_mutex);
3936 #define MAX_NODE_LOAD (nr_online_nodes)
3937 static int node_load[MAX_NUMNODES];
3940 * find_next_best_node - find the next node that should appear in a given node's fallback list
3941 * @node: node whose fallback list we're appending
3942 * @used_node_mask: nodemask_t of already used nodes
3944 * We use a number of factors to determine which is the next node that should
3945 * appear on a given node's fallback list. The node should not have appeared
3946 * already in @node's fallback list, and it should be the next closest node
3947 * according to the distance array (which contains arbitrary distance values
3948 * from each node to each node in the system), and should also prefer nodes
3949 * with no CPUs, since presumably they'll have very little allocation pressure
3950 * on them otherwise.
3951 * It returns -1 if no node is found.
3953 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3956 int min_val = INT_MAX;
3957 int best_node = NUMA_NO_NODE;
3958 const struct cpumask *tmp = cpumask_of_node(0);
3960 /* Use the local node if we haven't already */
3961 if (!node_isset(node, *used_node_mask)) {
3962 node_set(node, *used_node_mask);
3966 for_each_node_state(n, N_MEMORY) {
3968 /* Don't want a node to appear more than once */
3969 if (node_isset(n, *used_node_mask))
3972 /* Use the distance array to find the distance */
3973 val = node_distance(node, n);
3975 /* Penalize nodes under us ("prefer the next node") */
3978 /* Give preference to headless and unused nodes */
3979 tmp = cpumask_of_node(n);
3980 if (!cpumask_empty(tmp))
3981 val += PENALTY_FOR_NODE_WITH_CPUS;
3983 /* Slight preference for less loaded node */
3984 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3985 val += node_load[n];
3987 if (val < min_val) {
3994 node_set(best_node, *used_node_mask);
4001 * Build zonelists ordered by node and zones within node.
4002 * This results in maximum locality--normal zone overflows into local
4003 * DMA zone, if any--but risks exhausting DMA zone.
4005 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4008 struct zonelist *zonelist;
4010 zonelist = &pgdat->node_zonelists[0];
4011 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4013 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4014 zonelist->_zonerefs[j].zone = NULL;
4015 zonelist->_zonerefs[j].zone_idx = 0;
4019 * Build gfp_thisnode zonelists
4021 static void build_thisnode_zonelists(pg_data_t *pgdat)
4024 struct zonelist *zonelist;
4026 zonelist = &pgdat->node_zonelists[1];
4027 j = build_zonelists_node(pgdat, zonelist, 0);
4028 zonelist->_zonerefs[j].zone = NULL;
4029 zonelist->_zonerefs[j].zone_idx = 0;
4033 * Build zonelists ordered by zone and nodes within zones.
4034 * This results in conserving DMA zone[s] until all Normal memory is
4035 * exhausted, but results in overflowing to remote node while memory
4036 * may still exist in local DMA zone.
4038 static int node_order[MAX_NUMNODES];
4040 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4043 int zone_type; /* needs to be signed */
4045 struct zonelist *zonelist;
4047 zonelist = &pgdat->node_zonelists[0];
4049 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4050 for (j = 0; j < nr_nodes; j++) {
4051 node = node_order[j];
4052 z = &NODE_DATA(node)->node_zones[zone_type];
4053 if (populated_zone(z)) {
4055 &zonelist->_zonerefs[pos++]);
4056 check_highest_zone(zone_type);
4060 zonelist->_zonerefs[pos].zone = NULL;
4061 zonelist->_zonerefs[pos].zone_idx = 0;
4064 #if defined(CONFIG_64BIT)
4066 * Devices that require DMA32/DMA are relatively rare and do not justify a
4067 * penalty to every machine in case the specialised case applies. Default
4068 * to Node-ordering on 64-bit NUMA machines
4070 static int default_zonelist_order(void)
4072 return ZONELIST_ORDER_NODE;
4076 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4077 * by the kernel. If processes running on node 0 deplete the low memory zone
4078 * then reclaim will occur more frequency increasing stalls and potentially
4079 * be easier to OOM if a large percentage of the zone is under writeback or
4080 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4081 * Hence, default to zone ordering on 32-bit.
4083 static int default_zonelist_order(void)
4085 return ZONELIST_ORDER_ZONE;
4087 #endif /* CONFIG_64BIT */
4089 static void set_zonelist_order(void)
4091 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4092 current_zonelist_order = default_zonelist_order();
4094 current_zonelist_order = user_zonelist_order;
4097 static void build_zonelists(pg_data_t *pgdat)
4101 nodemask_t used_mask;
4102 int local_node, prev_node;
4103 struct zonelist *zonelist;
4104 int order = current_zonelist_order;
4106 /* initialize zonelists */
4107 for (i = 0; i < MAX_ZONELISTS; i++) {
4108 zonelist = pgdat->node_zonelists + i;
4109 zonelist->_zonerefs[0].zone = NULL;
4110 zonelist->_zonerefs[0].zone_idx = 0;
4113 /* NUMA-aware ordering of nodes */
4114 local_node = pgdat->node_id;
4115 load = nr_online_nodes;
4116 prev_node = local_node;
4117 nodes_clear(used_mask);
4119 memset(node_order, 0, sizeof(node_order));
4122 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4124 * We don't want to pressure a particular node.
4125 * So adding penalty to the first node in same
4126 * distance group to make it round-robin.
4128 if (node_distance(local_node, node) !=
4129 node_distance(local_node, prev_node))
4130 node_load[node] = load;
4134 if (order == ZONELIST_ORDER_NODE)
4135 build_zonelists_in_node_order(pgdat, node);
4137 node_order[j++] = node; /* remember order */
4140 if (order == ZONELIST_ORDER_ZONE) {
4141 /* calculate node order -- i.e., DMA last! */
4142 build_zonelists_in_zone_order(pgdat, j);
4145 build_thisnode_zonelists(pgdat);
4148 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4150 * Return node id of node used for "local" allocations.
4151 * I.e., first node id of first zone in arg node's generic zonelist.
4152 * Used for initializing percpu 'numa_mem', which is used primarily
4153 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4155 int local_memory_node(int node)
4159 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4160 gfp_zone(GFP_KERNEL),
4167 #else /* CONFIG_NUMA */
4169 static void set_zonelist_order(void)
4171 current_zonelist_order = ZONELIST_ORDER_ZONE;
4174 static void build_zonelists(pg_data_t *pgdat)
4176 int node, local_node;
4178 struct zonelist *zonelist;
4180 local_node = pgdat->node_id;
4182 zonelist = &pgdat->node_zonelists[0];
4183 j = build_zonelists_node(pgdat, zonelist, 0);
4186 * Now we build the zonelist so that it contains the zones
4187 * of all the other nodes.
4188 * We don't want to pressure a particular node, so when
4189 * building the zones for node N, we make sure that the
4190 * zones coming right after the local ones are those from
4191 * node N+1 (modulo N)
4193 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4194 if (!node_online(node))
4196 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4198 for (node = 0; node < local_node; node++) {
4199 if (!node_online(node))
4201 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4204 zonelist->_zonerefs[j].zone = NULL;
4205 zonelist->_zonerefs[j].zone_idx = 0;
4208 #endif /* CONFIG_NUMA */
4211 * Boot pageset table. One per cpu which is going to be used for all
4212 * zones and all nodes. The parameters will be set in such a way
4213 * that an item put on a list will immediately be handed over to
4214 * the buddy list. This is safe since pageset manipulation is done
4215 * with interrupts disabled.
4217 * The boot_pagesets must be kept even after bootup is complete for
4218 * unused processors and/or zones. They do play a role for bootstrapping
4219 * hotplugged processors.
4221 * zoneinfo_show() and maybe other functions do
4222 * not check if the processor is online before following the pageset pointer.
4223 * Other parts of the kernel may not check if the zone is available.
4225 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4226 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4227 static void setup_zone_pageset(struct zone *zone);
4230 * Global mutex to protect against size modification of zonelists
4231 * as well as to serialize pageset setup for the new populated zone.
4233 DEFINE_MUTEX(zonelists_mutex);
4235 /* return values int ....just for stop_machine() */
4236 static int __build_all_zonelists(void *data)
4240 pg_data_t *self = data;
4243 memset(node_load, 0, sizeof(node_load));
4246 if (self && !node_online(self->node_id)) {
4247 build_zonelists(self);
4250 for_each_online_node(nid) {
4251 pg_data_t *pgdat = NODE_DATA(nid);
4253 build_zonelists(pgdat);
4257 * Initialize the boot_pagesets that are going to be used
4258 * for bootstrapping processors. The real pagesets for
4259 * each zone will be allocated later when the per cpu
4260 * allocator is available.
4262 * boot_pagesets are used also for bootstrapping offline
4263 * cpus if the system is already booted because the pagesets
4264 * are needed to initialize allocators on a specific cpu too.
4265 * F.e. the percpu allocator needs the page allocator which
4266 * needs the percpu allocator in order to allocate its pagesets
4267 * (a chicken-egg dilemma).
4269 for_each_possible_cpu(cpu) {
4270 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4272 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4274 * We now know the "local memory node" for each node--
4275 * i.e., the node of the first zone in the generic zonelist.
4276 * Set up numa_mem percpu variable for all possible cpus
4277 * if associated node has been onlined.
4279 if (node_online(cpu_to_node(cpu)))
4280 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4282 set_cpu_numa_mem(cpu, NUMA_NO_NODE);
4289 static noinline void __init
4290 build_all_zonelists_init(void)
4292 __build_all_zonelists(NULL);
4293 mminit_verify_zonelist();
4294 cpuset_init_current_mems_allowed();
4298 * Called with zonelists_mutex held always
4299 * unless system_state == SYSTEM_BOOTING.
4301 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4302 * [we're only called with non-NULL zone through __meminit paths] and
4303 * (2) call of __init annotated helper build_all_zonelists_init
4304 * [protected by SYSTEM_BOOTING].
4306 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4308 set_zonelist_order();
4310 if (system_state == SYSTEM_BOOTING) {
4311 build_all_zonelists_init();
4313 #ifdef CONFIG_MEMORY_HOTPLUG
4315 setup_zone_pageset(zone);
4317 /* we have to stop all cpus to guarantee there is no user
4319 stop_machine(__build_all_zonelists, pgdat, NULL);
4320 /* cpuset refresh routine should be here */
4322 vm_total_pages = nr_free_pagecache_pages();
4324 * Disable grouping by mobility if the number of pages in the
4325 * system is too low to allow the mechanism to work. It would be
4326 * more accurate, but expensive to check per-zone. This check is
4327 * made on memory-hotadd so a system can start with mobility
4328 * disabled and enable it later
4330 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4331 page_group_by_mobility_disabled = 1;
4333 page_group_by_mobility_disabled = 0;
4335 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4336 "Total pages: %ld\n",
4338 zonelist_order_name[current_zonelist_order],
4339 page_group_by_mobility_disabled ? "off" : "on",
4342 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4347 * Helper functions to size the waitqueue hash table.
4348 * Essentially these want to choose hash table sizes sufficiently
4349 * large so that collisions trying to wait on pages are rare.
4350 * But in fact, the number of active page waitqueues on typical
4351 * systems is ridiculously low, less than 200. So this is even
4352 * conservative, even though it seems large.
4354 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4355 * waitqueues, i.e. the size of the waitq table given the number of pages.
4357 #define PAGES_PER_WAITQUEUE 256
4359 #ifndef CONFIG_MEMORY_HOTPLUG
4360 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4362 unsigned long size = 1;
4364 pages /= PAGES_PER_WAITQUEUE;
4366 while (size < pages)
4370 * Once we have dozens or even hundreds of threads sleeping
4371 * on IO we've got bigger problems than wait queue collision.
4372 * Limit the size of the wait table to a reasonable size.
4374 size = min(size, 4096UL);
4376 return max(size, 4UL);
4380 * A zone's size might be changed by hot-add, so it is not possible to determine
4381 * a suitable size for its wait_table. So we use the maximum size now.
4383 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4385 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4386 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4387 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4389 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4390 * or more by the traditional way. (See above). It equals:
4392 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4393 * ia64(16K page size) : = ( 8G + 4M)byte.
4394 * powerpc (64K page size) : = (32G +16M)byte.
4396 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4403 * This is an integer logarithm so that shifts can be used later
4404 * to extract the more random high bits from the multiplicative
4405 * hash function before the remainder is taken.
4407 static inline unsigned long wait_table_bits(unsigned long size)
4413 * Initially all pages are reserved - free ones are freed
4414 * up by free_all_bootmem() once the early boot process is
4415 * done. Non-atomic initialization, single-pass.
4417 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4418 unsigned long start_pfn, enum memmap_context context)
4420 pg_data_t *pgdat = NODE_DATA(nid);
4421 unsigned long end_pfn = start_pfn + size;
4424 unsigned long nr_initialised = 0;
4426 if (highest_memmap_pfn < end_pfn - 1)
4427 highest_memmap_pfn = end_pfn - 1;
4429 z = &pgdat->node_zones[zone];
4430 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4432 * There can be holes in boot-time mem_map[]s
4433 * handed to this function. They do not
4434 * exist on hotplugged memory.
4436 if (context == MEMMAP_EARLY) {
4437 if (!early_pfn_valid(pfn))
4439 if (!early_pfn_in_nid(pfn, nid))
4441 if (!update_defer_init(pgdat, pfn, end_pfn,
4447 * Mark the block movable so that blocks are reserved for
4448 * movable at startup. This will force kernel allocations
4449 * to reserve their blocks rather than leaking throughout
4450 * the address space during boot when many long-lived
4451 * kernel allocations are made.
4453 * bitmap is created for zone's valid pfn range. but memmap
4454 * can be created for invalid pages (for alignment)
4455 * check here not to call set_pageblock_migratetype() against
4458 if (!(pfn & (pageblock_nr_pages - 1))) {
4459 struct page *page = pfn_to_page(pfn);
4461 __init_single_page(page, pfn, zone, nid);
4462 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4464 __init_single_pfn(pfn, zone, nid);
4469 static void __meminit zone_init_free_lists(struct zone *zone)
4471 unsigned int order, t;
4472 for_each_migratetype_order(order, t) {
4473 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4474 zone->free_area[order].nr_free = 0;
4478 #ifndef __HAVE_ARCH_MEMMAP_INIT
4479 #define memmap_init(size, nid, zone, start_pfn) \
4480 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4483 static int zone_batchsize(struct zone *zone)
4489 * The per-cpu-pages pools are set to around 1000th of the
4490 * size of the zone. But no more than 1/2 of a meg.
4492 * OK, so we don't know how big the cache is. So guess.
4494 batch = zone->managed_pages / 1024;
4495 if (batch * PAGE_SIZE > 512 * 1024)
4496 batch = (512 * 1024) / PAGE_SIZE;
4497 batch /= 4; /* We effectively *= 4 below */
4502 * Clamp the batch to a 2^n - 1 value. Having a power
4503 * of 2 value was found to be more likely to have
4504 * suboptimal cache aliasing properties in some cases.
4506 * For example if 2 tasks are alternately allocating
4507 * batches of pages, one task can end up with a lot
4508 * of pages of one half of the possible page colors
4509 * and the other with pages of the other colors.
4511 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4516 /* The deferral and batching of frees should be suppressed under NOMMU
4519 * The problem is that NOMMU needs to be able to allocate large chunks
4520 * of contiguous memory as there's no hardware page translation to
4521 * assemble apparent contiguous memory from discontiguous pages.
4523 * Queueing large contiguous runs of pages for batching, however,
4524 * causes the pages to actually be freed in smaller chunks. As there
4525 * can be a significant delay between the individual batches being
4526 * recycled, this leads to the once large chunks of space being
4527 * fragmented and becoming unavailable for high-order allocations.
4534 * pcp->high and pcp->batch values are related and dependent on one another:
4535 * ->batch must never be higher then ->high.
4536 * The following function updates them in a safe manner without read side
4539 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4540 * those fields changing asynchronously (acording the the above rule).
4542 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4543 * outside of boot time (or some other assurance that no concurrent updaters
4546 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4547 unsigned long batch)
4549 /* start with a fail safe value for batch */
4553 /* Update high, then batch, in order */
4560 /* a companion to pageset_set_high() */
4561 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4563 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4566 static void pageset_init(struct per_cpu_pageset *p)
4568 struct per_cpu_pages *pcp;
4571 memset(p, 0, sizeof(*p));
4575 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4576 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4579 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4582 pageset_set_batch(p, batch);
4586 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4587 * to the value high for the pageset p.
4589 static void pageset_set_high(struct per_cpu_pageset *p,
4592 unsigned long batch = max(1UL, high / 4);
4593 if ((high / 4) > (PAGE_SHIFT * 8))
4594 batch = PAGE_SHIFT * 8;
4596 pageset_update(&p->pcp, high, batch);
4599 static void pageset_set_high_and_batch(struct zone *zone,
4600 struct per_cpu_pageset *pcp)
4602 if (percpu_pagelist_fraction)
4603 pageset_set_high(pcp,
4604 (zone->managed_pages /
4605 percpu_pagelist_fraction));
4607 pageset_set_batch(pcp, zone_batchsize(zone));
4610 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4612 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4615 pageset_set_high_and_batch(zone, pcp);
4618 static void __meminit setup_zone_pageset(struct zone *zone)
4621 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4622 for_each_possible_cpu(cpu)
4623 zone_pageset_init(zone, cpu);
4627 * Allocate per cpu pagesets and initialize them.
4628 * Before this call only boot pagesets were available.
4630 void __init setup_per_cpu_pageset(void)
4634 for_each_populated_zone(zone)
4635 setup_zone_pageset(zone);
4638 static noinline __init_refok
4639 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4645 * The per-page waitqueue mechanism uses hashed waitqueues
4648 zone->wait_table_hash_nr_entries =
4649 wait_table_hash_nr_entries(zone_size_pages);
4650 zone->wait_table_bits =
4651 wait_table_bits(zone->wait_table_hash_nr_entries);
4652 alloc_size = zone->wait_table_hash_nr_entries
4653 * sizeof(wait_queue_head_t);
4655 if (!slab_is_available()) {
4656 zone->wait_table = (wait_queue_head_t *)
4657 memblock_virt_alloc_node_nopanic(
4658 alloc_size, zone->zone_pgdat->node_id);
4661 * This case means that a zone whose size was 0 gets new memory
4662 * via memory hot-add.
4663 * But it may be the case that a new node was hot-added. In
4664 * this case vmalloc() will not be able to use this new node's
4665 * memory - this wait_table must be initialized to use this new
4666 * node itself as well.
4667 * To use this new node's memory, further consideration will be
4670 zone->wait_table = vmalloc(alloc_size);
4672 if (!zone->wait_table)
4675 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4676 init_waitqueue_head(zone->wait_table + i);
4681 static __meminit void zone_pcp_init(struct zone *zone)
4684 * per cpu subsystem is not up at this point. The following code
4685 * relies on the ability of the linker to provide the
4686 * offset of a (static) per cpu variable into the per cpu area.
4688 zone->pageset = &boot_pageset;
4690 if (populated_zone(zone))
4691 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4692 zone->name, zone->present_pages,
4693 zone_batchsize(zone));
4696 int __meminit init_currently_empty_zone(struct zone *zone,
4697 unsigned long zone_start_pfn,
4700 struct pglist_data *pgdat = zone->zone_pgdat;
4702 ret = zone_wait_table_init(zone, size);
4705 pgdat->nr_zones = zone_idx(zone) + 1;
4707 zone->zone_start_pfn = zone_start_pfn;
4709 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4710 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4712 (unsigned long)zone_idx(zone),
4713 zone_start_pfn, (zone_start_pfn + size));
4715 zone_init_free_lists(zone);
4720 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4721 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4724 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4726 int __meminit __early_pfn_to_nid(unsigned long pfn,
4727 struct mminit_pfnnid_cache *state)
4729 unsigned long start_pfn, end_pfn;
4732 if (state->last_start <= pfn && pfn < state->last_end)
4733 return state->last_nid;
4735 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4737 state->last_start = start_pfn;
4738 state->last_end = end_pfn;
4739 state->last_nid = nid;
4744 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4747 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4748 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4749 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4751 * If an architecture guarantees that all ranges registered contain no holes
4752 * and may be freed, this this function may be used instead of calling
4753 * memblock_free_early_nid() manually.
4755 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4757 unsigned long start_pfn, end_pfn;
4760 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4761 start_pfn = min(start_pfn, max_low_pfn);
4762 end_pfn = min(end_pfn, max_low_pfn);
4764 if (start_pfn < end_pfn)
4765 memblock_free_early_nid(PFN_PHYS(start_pfn),
4766 (end_pfn - start_pfn) << PAGE_SHIFT,
4772 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4773 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4775 * If an architecture guarantees that all ranges registered contain no holes and may
4776 * be freed, this function may be used instead of calling memory_present() manually.
4778 void __init sparse_memory_present_with_active_regions(int nid)
4780 unsigned long start_pfn, end_pfn;
4783 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4784 memory_present(this_nid, start_pfn, end_pfn);
4788 * get_pfn_range_for_nid - Return the start and end page frames for a node
4789 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4790 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4791 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4793 * It returns the start and end page frame of a node based on information
4794 * provided by memblock_set_node(). If called for a node
4795 * with no available memory, a warning is printed and the start and end
4798 void __meminit get_pfn_range_for_nid(unsigned int nid,
4799 unsigned long *start_pfn, unsigned long *end_pfn)
4801 unsigned long this_start_pfn, this_end_pfn;
4807 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4808 *start_pfn = min(*start_pfn, this_start_pfn);
4809 *end_pfn = max(*end_pfn, this_end_pfn);
4812 if (*start_pfn == -1UL)
4817 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4818 * assumption is made that zones within a node are ordered in monotonic
4819 * increasing memory addresses so that the "highest" populated zone is used
4821 static void __init find_usable_zone_for_movable(void)
4824 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4825 if (zone_index == ZONE_MOVABLE)
4828 if (arch_zone_highest_possible_pfn[zone_index] >
4829 arch_zone_lowest_possible_pfn[zone_index])
4833 VM_BUG_ON(zone_index == -1);
4834 movable_zone = zone_index;
4838 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4839 * because it is sized independent of architecture. Unlike the other zones,
4840 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4841 * in each node depending on the size of each node and how evenly kernelcore
4842 * is distributed. This helper function adjusts the zone ranges
4843 * provided by the architecture for a given node by using the end of the
4844 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4845 * zones within a node are in order of monotonic increases memory addresses
4847 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4848 unsigned long zone_type,
4849 unsigned long node_start_pfn,
4850 unsigned long node_end_pfn,
4851 unsigned long *zone_start_pfn,
4852 unsigned long *zone_end_pfn)
4854 /* Only adjust if ZONE_MOVABLE is on this node */
4855 if (zone_movable_pfn[nid]) {
4856 /* Size ZONE_MOVABLE */
4857 if (zone_type == ZONE_MOVABLE) {
4858 *zone_start_pfn = zone_movable_pfn[nid];
4859 *zone_end_pfn = min(node_end_pfn,
4860 arch_zone_highest_possible_pfn[movable_zone]);
4862 /* Adjust for ZONE_MOVABLE starting within this range */
4863 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4864 *zone_end_pfn > zone_movable_pfn[nid]) {
4865 *zone_end_pfn = zone_movable_pfn[nid];
4867 /* Check if this whole range is within ZONE_MOVABLE */
4868 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4869 *zone_start_pfn = *zone_end_pfn;
4874 * Return the number of pages a zone spans in a node, including holes
4875 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4877 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4878 unsigned long zone_type,
4879 unsigned long node_start_pfn,
4880 unsigned long node_end_pfn,
4881 unsigned long *ignored)
4883 unsigned long zone_start_pfn, zone_end_pfn;
4885 /* When hotadd a new node from cpu_up(), the node should be empty */
4886 if (!node_start_pfn && !node_end_pfn)
4889 /* Get the start and end of the zone */
4890 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4891 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4892 adjust_zone_range_for_zone_movable(nid, zone_type,
4893 node_start_pfn, node_end_pfn,
4894 &zone_start_pfn, &zone_end_pfn);
4896 /* Check that this node has pages within the zone's required range */
4897 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4900 /* Move the zone boundaries inside the node if necessary */
4901 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4902 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4904 /* Return the spanned pages */
4905 return zone_end_pfn - zone_start_pfn;
4909 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4910 * then all holes in the requested range will be accounted for.
4912 unsigned long __meminit __absent_pages_in_range(int nid,
4913 unsigned long range_start_pfn,
4914 unsigned long range_end_pfn)
4916 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4917 unsigned long start_pfn, end_pfn;
4920 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4921 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4922 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4923 nr_absent -= end_pfn - start_pfn;
4929 * absent_pages_in_range - Return number of page frames in holes within a range
4930 * @start_pfn: The start PFN to start searching for holes
4931 * @end_pfn: The end PFN to stop searching for holes
4933 * It returns the number of pages frames in memory holes within a range.
4935 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4936 unsigned long end_pfn)
4938 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4941 /* Return the number of page frames in holes in a zone on a node */
4942 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4943 unsigned long zone_type,
4944 unsigned long node_start_pfn,
4945 unsigned long node_end_pfn,
4946 unsigned long *ignored)
4948 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4949 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4950 unsigned long zone_start_pfn, zone_end_pfn;
4952 /* When hotadd a new node from cpu_up(), the node should be empty */
4953 if (!node_start_pfn && !node_end_pfn)
4956 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4957 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4959 adjust_zone_range_for_zone_movable(nid, zone_type,
4960 node_start_pfn, node_end_pfn,
4961 &zone_start_pfn, &zone_end_pfn);
4962 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4965 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4966 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4967 unsigned long zone_type,
4968 unsigned long node_start_pfn,
4969 unsigned long node_end_pfn,
4970 unsigned long *zones_size)
4972 return zones_size[zone_type];
4975 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4976 unsigned long zone_type,
4977 unsigned long node_start_pfn,
4978 unsigned long node_end_pfn,
4979 unsigned long *zholes_size)
4984 return zholes_size[zone_type];
4987 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4989 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4990 unsigned long node_start_pfn,
4991 unsigned long node_end_pfn,
4992 unsigned long *zones_size,
4993 unsigned long *zholes_size)
4995 unsigned long realtotalpages = 0, totalpages = 0;
4998 for (i = 0; i < MAX_NR_ZONES; i++) {
4999 struct zone *zone = pgdat->node_zones + i;
5000 unsigned long size, real_size;
5002 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5006 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5007 node_start_pfn, node_end_pfn,
5009 zone->spanned_pages = size;
5010 zone->present_pages = real_size;
5013 realtotalpages += real_size;
5016 pgdat->node_spanned_pages = totalpages;
5017 pgdat->node_present_pages = realtotalpages;
5018 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5022 #ifndef CONFIG_SPARSEMEM
5024 * Calculate the size of the zone->blockflags rounded to an unsigned long
5025 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5026 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5027 * round what is now in bits to nearest long in bits, then return it in
5030 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5032 unsigned long usemapsize;
5034 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5035 usemapsize = roundup(zonesize, pageblock_nr_pages);
5036 usemapsize = usemapsize >> pageblock_order;
5037 usemapsize *= NR_PAGEBLOCK_BITS;
5038 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5040 return usemapsize / 8;
5043 static void __init setup_usemap(struct pglist_data *pgdat,
5045 unsigned long zone_start_pfn,
5046 unsigned long zonesize)
5048 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5049 zone->pageblock_flags = NULL;
5051 zone->pageblock_flags =
5052 memblock_virt_alloc_node_nopanic(usemapsize,
5056 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5057 unsigned long zone_start_pfn, unsigned long zonesize) {}
5058 #endif /* CONFIG_SPARSEMEM */
5060 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5062 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5063 void __paginginit set_pageblock_order(void)
5067 /* Check that pageblock_nr_pages has not already been setup */
5068 if (pageblock_order)
5071 if (HPAGE_SHIFT > PAGE_SHIFT)
5072 order = HUGETLB_PAGE_ORDER;
5074 order = MAX_ORDER - 1;
5077 * Assume the largest contiguous order of interest is a huge page.
5078 * This value may be variable depending on boot parameters on IA64 and
5081 pageblock_order = order;
5083 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5086 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5087 * is unused as pageblock_order is set at compile-time. See
5088 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5091 void __paginginit set_pageblock_order(void)
5095 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5097 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5098 unsigned long present_pages)
5100 unsigned long pages = spanned_pages;
5103 * Provide a more accurate estimation if there are holes within
5104 * the zone and SPARSEMEM is in use. If there are holes within the
5105 * zone, each populated memory region may cost us one or two extra
5106 * memmap pages due to alignment because memmap pages for each
5107 * populated regions may not naturally algined on page boundary.
5108 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5110 if (spanned_pages > present_pages + (present_pages >> 4) &&
5111 IS_ENABLED(CONFIG_SPARSEMEM))
5112 pages = present_pages;
5114 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5118 * Set up the zone data structures:
5119 * - mark all pages reserved
5120 * - mark all memory queues empty
5121 * - clear the memory bitmaps
5123 * NOTE: pgdat should get zeroed by caller.
5125 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5128 int nid = pgdat->node_id;
5129 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5132 pgdat_resize_init(pgdat);
5133 #ifdef CONFIG_NUMA_BALANCING
5134 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5135 pgdat->numabalancing_migrate_nr_pages = 0;
5136 pgdat->numabalancing_migrate_next_window = jiffies;
5138 init_waitqueue_head(&pgdat->kswapd_wait);
5139 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5140 pgdat_page_ext_init(pgdat);
5142 for (j = 0; j < MAX_NR_ZONES; j++) {
5143 struct zone *zone = pgdat->node_zones + j;
5144 unsigned long size, realsize, freesize, memmap_pages;
5146 size = zone->spanned_pages;
5147 realsize = freesize = zone->present_pages;
5150 * Adjust freesize so that it accounts for how much memory
5151 * is used by this zone for memmap. This affects the watermark
5152 * and per-cpu initialisations
5154 memmap_pages = calc_memmap_size(size, realsize);
5155 if (!is_highmem_idx(j)) {
5156 if (freesize >= memmap_pages) {
5157 freesize -= memmap_pages;
5160 " %s zone: %lu pages used for memmap\n",
5161 zone_names[j], memmap_pages);
5164 " %s zone: %lu pages exceeds freesize %lu\n",
5165 zone_names[j], memmap_pages, freesize);
5168 /* Account for reserved pages */
5169 if (j == 0 && freesize > dma_reserve) {
5170 freesize -= dma_reserve;
5171 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5172 zone_names[0], dma_reserve);
5175 if (!is_highmem_idx(j))
5176 nr_kernel_pages += freesize;
5177 /* Charge for highmem memmap if there are enough kernel pages */
5178 else if (nr_kernel_pages > memmap_pages * 2)
5179 nr_kernel_pages -= memmap_pages;
5180 nr_all_pages += freesize;
5183 * Set an approximate value for lowmem here, it will be adjusted
5184 * when the bootmem allocator frees pages into the buddy system.
5185 * And all highmem pages will be managed by the buddy system.
5187 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5190 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5192 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5194 zone->name = zone_names[j];
5195 spin_lock_init(&zone->lock);
5196 spin_lock_init(&zone->lru_lock);
5197 zone_seqlock_init(zone);
5198 zone->zone_pgdat = pgdat;
5199 zone_pcp_init(zone);
5201 /* For bootup, initialized properly in watermark setup */
5202 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5204 lruvec_init(&zone->lruvec);
5208 set_pageblock_order();
5209 setup_usemap(pgdat, zone, zone_start_pfn, size);
5210 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5212 memmap_init(size, nid, j, zone_start_pfn);
5213 zone_start_pfn += size;
5217 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5219 /* Skip empty nodes */
5220 if (!pgdat->node_spanned_pages)
5223 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5224 /* ia64 gets its own node_mem_map, before this, without bootmem */
5225 if (!pgdat->node_mem_map) {
5226 unsigned long size, start, end;
5230 * The zone's endpoints aren't required to be MAX_ORDER
5231 * aligned but the node_mem_map endpoints must be in order
5232 * for the buddy allocator to function correctly.
5234 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5235 end = pgdat_end_pfn(pgdat);
5236 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5237 size = (end - start) * sizeof(struct page);
5238 map = alloc_remap(pgdat->node_id, size);
5240 map = memblock_virt_alloc_node_nopanic(size,
5242 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5244 #ifndef CONFIG_NEED_MULTIPLE_NODES
5246 * With no DISCONTIG, the global mem_map is just set as node 0's
5248 if (pgdat == NODE_DATA(0)) {
5249 mem_map = NODE_DATA(0)->node_mem_map;
5250 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5251 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5252 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5253 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5256 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5259 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5260 unsigned long node_start_pfn, unsigned long *zholes_size)
5262 pg_data_t *pgdat = NODE_DATA(nid);
5263 unsigned long start_pfn = 0;
5264 unsigned long end_pfn = 0;
5266 /* pg_data_t should be reset to zero when it's allocated */
5267 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5269 reset_deferred_meminit(pgdat);
5270 pgdat->node_id = nid;
5271 pgdat->node_start_pfn = node_start_pfn;
5272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5273 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5274 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5275 (u64)start_pfn << PAGE_SHIFT,
5276 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5278 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5279 zones_size, zholes_size);
5281 alloc_node_mem_map(pgdat);
5282 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5283 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5284 nid, (unsigned long)pgdat,
5285 (unsigned long)pgdat->node_mem_map);
5288 free_area_init_core(pgdat);
5291 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5293 #if MAX_NUMNODES > 1
5295 * Figure out the number of possible node ids.
5297 void __init setup_nr_node_ids(void)
5299 unsigned int highest;
5301 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5302 nr_node_ids = highest + 1;
5307 * node_map_pfn_alignment - determine the maximum internode alignment
5309 * This function should be called after node map is populated and sorted.
5310 * It calculates the maximum power of two alignment which can distinguish
5313 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5314 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5315 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5316 * shifted, 1GiB is enough and this function will indicate so.
5318 * This is used to test whether pfn -> nid mapping of the chosen memory
5319 * model has fine enough granularity to avoid incorrect mapping for the
5320 * populated node map.
5322 * Returns the determined alignment in pfn's. 0 if there is no alignment
5323 * requirement (single node).
5325 unsigned long __init node_map_pfn_alignment(void)
5327 unsigned long accl_mask = 0, last_end = 0;
5328 unsigned long start, end, mask;
5332 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5333 if (!start || last_nid < 0 || last_nid == nid) {
5340 * Start with a mask granular enough to pin-point to the
5341 * start pfn and tick off bits one-by-one until it becomes
5342 * too coarse to separate the current node from the last.
5344 mask = ~((1 << __ffs(start)) - 1);
5345 while (mask && last_end <= (start & (mask << 1)))
5348 /* accumulate all internode masks */
5352 /* convert mask to number of pages */
5353 return ~accl_mask + 1;
5356 /* Find the lowest pfn for a node */
5357 static unsigned long __init find_min_pfn_for_node(int nid)
5359 unsigned long min_pfn = ULONG_MAX;
5360 unsigned long start_pfn;
5363 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5364 min_pfn = min(min_pfn, start_pfn);
5366 if (min_pfn == ULONG_MAX) {
5368 "Could not find start_pfn for node %d\n", nid);
5376 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5378 * It returns the minimum PFN based on information provided via
5379 * memblock_set_node().
5381 unsigned long __init find_min_pfn_with_active_regions(void)
5383 return find_min_pfn_for_node(MAX_NUMNODES);
5387 * early_calculate_totalpages()
5388 * Sum pages in active regions for movable zone.
5389 * Populate N_MEMORY for calculating usable_nodes.
5391 static unsigned long __init early_calculate_totalpages(void)
5393 unsigned long totalpages = 0;
5394 unsigned long start_pfn, end_pfn;
5397 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5398 unsigned long pages = end_pfn - start_pfn;
5400 totalpages += pages;
5402 node_set_state(nid, N_MEMORY);
5408 * Find the PFN the Movable zone begins in each node. Kernel memory
5409 * is spread evenly between nodes as long as the nodes have enough
5410 * memory. When they don't, some nodes will have more kernelcore than
5413 static void __init find_zone_movable_pfns_for_nodes(void)
5416 unsigned long usable_startpfn;
5417 unsigned long kernelcore_node, kernelcore_remaining;
5418 /* save the state before borrow the nodemask */
5419 nodemask_t saved_node_state = node_states[N_MEMORY];
5420 unsigned long totalpages = early_calculate_totalpages();
5421 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5422 struct memblock_region *r;
5424 /* Need to find movable_zone earlier when movable_node is specified. */
5425 find_usable_zone_for_movable();
5428 * If movable_node is specified, ignore kernelcore and movablecore
5431 if (movable_node_is_enabled()) {
5432 for_each_memblock(memory, r) {
5433 if (!memblock_is_hotpluggable(r))
5438 usable_startpfn = PFN_DOWN(r->base);
5439 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5440 min(usable_startpfn, zone_movable_pfn[nid]) :
5448 * If movablecore=nn[KMG] was specified, calculate what size of
5449 * kernelcore that corresponds so that memory usable for
5450 * any allocation type is evenly spread. If both kernelcore
5451 * and movablecore are specified, then the value of kernelcore
5452 * will be used for required_kernelcore if it's greater than
5453 * what movablecore would have allowed.
5455 if (required_movablecore) {
5456 unsigned long corepages;
5459 * Round-up so that ZONE_MOVABLE is at least as large as what
5460 * was requested by the user
5462 required_movablecore =
5463 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5464 corepages = totalpages - required_movablecore;
5466 required_kernelcore = max(required_kernelcore, corepages);
5469 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5470 if (!required_kernelcore)
5473 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5474 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5477 /* Spread kernelcore memory as evenly as possible throughout nodes */
5478 kernelcore_node = required_kernelcore / usable_nodes;
5479 for_each_node_state(nid, N_MEMORY) {
5480 unsigned long start_pfn, end_pfn;
5483 * Recalculate kernelcore_node if the division per node
5484 * now exceeds what is necessary to satisfy the requested
5485 * amount of memory for the kernel
5487 if (required_kernelcore < kernelcore_node)
5488 kernelcore_node = required_kernelcore / usable_nodes;
5491 * As the map is walked, we track how much memory is usable
5492 * by the kernel using kernelcore_remaining. When it is
5493 * 0, the rest of the node is usable by ZONE_MOVABLE
5495 kernelcore_remaining = kernelcore_node;
5497 /* Go through each range of PFNs within this node */
5498 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5499 unsigned long size_pages;
5501 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5502 if (start_pfn >= end_pfn)
5505 /* Account for what is only usable for kernelcore */
5506 if (start_pfn < usable_startpfn) {
5507 unsigned long kernel_pages;
5508 kernel_pages = min(end_pfn, usable_startpfn)
5511 kernelcore_remaining -= min(kernel_pages,
5512 kernelcore_remaining);
5513 required_kernelcore -= min(kernel_pages,
5514 required_kernelcore);
5516 /* Continue if range is now fully accounted */
5517 if (end_pfn <= usable_startpfn) {
5520 * Push zone_movable_pfn to the end so
5521 * that if we have to rebalance
5522 * kernelcore across nodes, we will
5523 * not double account here
5525 zone_movable_pfn[nid] = end_pfn;
5528 start_pfn = usable_startpfn;
5532 * The usable PFN range for ZONE_MOVABLE is from
5533 * start_pfn->end_pfn. Calculate size_pages as the
5534 * number of pages used as kernelcore
5536 size_pages = end_pfn - start_pfn;
5537 if (size_pages > kernelcore_remaining)
5538 size_pages = kernelcore_remaining;
5539 zone_movable_pfn[nid] = start_pfn + size_pages;
5542 * Some kernelcore has been met, update counts and
5543 * break if the kernelcore for this node has been
5546 required_kernelcore -= min(required_kernelcore,
5548 kernelcore_remaining -= size_pages;
5549 if (!kernelcore_remaining)
5555 * If there is still required_kernelcore, we do another pass with one
5556 * less node in the count. This will push zone_movable_pfn[nid] further
5557 * along on the nodes that still have memory until kernelcore is
5561 if (usable_nodes && required_kernelcore > usable_nodes)
5565 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5566 for (nid = 0; nid < MAX_NUMNODES; nid++)
5567 zone_movable_pfn[nid] =
5568 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5571 /* restore the node_state */
5572 node_states[N_MEMORY] = saved_node_state;
5575 /* Any regular or high memory on that node ? */
5576 static void check_for_memory(pg_data_t *pgdat, int nid)
5578 enum zone_type zone_type;
5580 if (N_MEMORY == N_NORMAL_MEMORY)
5583 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5584 struct zone *zone = &pgdat->node_zones[zone_type];
5585 if (populated_zone(zone)) {
5586 node_set_state(nid, N_HIGH_MEMORY);
5587 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5588 zone_type <= ZONE_NORMAL)
5589 node_set_state(nid, N_NORMAL_MEMORY);
5596 * free_area_init_nodes - Initialise all pg_data_t and zone data
5597 * @max_zone_pfn: an array of max PFNs for each zone
5599 * This will call free_area_init_node() for each active node in the system.
5600 * Using the page ranges provided by memblock_set_node(), the size of each
5601 * zone in each node and their holes is calculated. If the maximum PFN
5602 * between two adjacent zones match, it is assumed that the zone is empty.
5603 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5604 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5605 * starts where the previous one ended. For example, ZONE_DMA32 starts
5606 * at arch_max_dma_pfn.
5608 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5610 unsigned long start_pfn, end_pfn;
5613 /* Record where the zone boundaries are */
5614 memset(arch_zone_lowest_possible_pfn, 0,
5615 sizeof(arch_zone_lowest_possible_pfn));
5616 memset(arch_zone_highest_possible_pfn, 0,
5617 sizeof(arch_zone_highest_possible_pfn));
5618 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5619 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5620 for (i = 1; i < MAX_NR_ZONES; i++) {
5621 if (i == ZONE_MOVABLE)
5623 arch_zone_lowest_possible_pfn[i] =
5624 arch_zone_highest_possible_pfn[i-1];
5625 arch_zone_highest_possible_pfn[i] =
5626 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5628 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5629 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5631 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5632 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5633 find_zone_movable_pfns_for_nodes();
5635 /* Print out the zone ranges */
5636 pr_info("Zone ranges:\n");
5637 for (i = 0; i < MAX_NR_ZONES; i++) {
5638 if (i == ZONE_MOVABLE)
5640 pr_info(" %-8s ", zone_names[i]);
5641 if (arch_zone_lowest_possible_pfn[i] ==
5642 arch_zone_highest_possible_pfn[i])
5645 pr_cont("[mem %#018Lx-%#018Lx]\n",
5646 (u64)arch_zone_lowest_possible_pfn[i]
5648 ((u64)arch_zone_highest_possible_pfn[i]
5649 << PAGE_SHIFT) - 1);
5652 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5653 pr_info("Movable zone start for each node\n");
5654 for (i = 0; i < MAX_NUMNODES; i++) {
5655 if (zone_movable_pfn[i])
5656 pr_info(" Node %d: %#018Lx\n", i,
5657 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5660 /* Print out the early node map */
5661 pr_info("Early memory node ranges\n");
5662 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5663 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5664 (u64)start_pfn << PAGE_SHIFT,
5665 ((u64)end_pfn << PAGE_SHIFT) - 1);
5667 /* Initialise every node */
5668 mminit_verify_pageflags_layout();
5669 setup_nr_node_ids();
5670 for_each_online_node(nid) {
5671 pg_data_t *pgdat = NODE_DATA(nid);
5672 free_area_init_node(nid, NULL,
5673 find_min_pfn_for_node(nid), NULL);
5675 /* Any memory on that node */
5676 if (pgdat->node_present_pages)
5677 node_set_state(nid, N_MEMORY);
5678 check_for_memory(pgdat, nid);
5682 static int __init cmdline_parse_core(char *p, unsigned long *core)
5684 unsigned long long coremem;
5688 coremem = memparse(p, &p);
5689 *core = coremem >> PAGE_SHIFT;
5691 /* Paranoid check that UL is enough for the coremem value */
5692 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5698 * kernelcore=size sets the amount of memory for use for allocations that
5699 * cannot be reclaimed or migrated.
5701 static int __init cmdline_parse_kernelcore(char *p)
5703 return cmdline_parse_core(p, &required_kernelcore);
5707 * movablecore=size sets the amount of memory for use for allocations that
5708 * can be reclaimed or migrated.
5710 static int __init cmdline_parse_movablecore(char *p)
5712 return cmdline_parse_core(p, &required_movablecore);
5715 early_param("kernelcore", cmdline_parse_kernelcore);
5716 early_param("movablecore", cmdline_parse_movablecore);
5718 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5720 void adjust_managed_page_count(struct page *page, long count)
5722 spin_lock(&managed_page_count_lock);
5723 page_zone(page)->managed_pages += count;
5724 totalram_pages += count;
5725 #ifdef CONFIG_HIGHMEM
5726 if (PageHighMem(page))
5727 totalhigh_pages += count;
5729 spin_unlock(&managed_page_count_lock);
5731 EXPORT_SYMBOL(adjust_managed_page_count);
5733 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5736 unsigned long pages = 0;
5738 start = (void *)PAGE_ALIGN((unsigned long)start);
5739 end = (void *)((unsigned long)end & PAGE_MASK);
5740 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5741 if ((unsigned int)poison <= 0xFF)
5742 memset(pos, poison, PAGE_SIZE);
5743 free_reserved_page(virt_to_page(pos));
5747 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5748 s, pages << (PAGE_SHIFT - 10), start, end);
5752 EXPORT_SYMBOL(free_reserved_area);
5754 #ifdef CONFIG_HIGHMEM
5755 void free_highmem_page(struct page *page)
5757 __free_reserved_page(page);
5759 page_zone(page)->managed_pages++;
5765 void __init mem_init_print_info(const char *str)
5767 unsigned long physpages, codesize, datasize, rosize, bss_size;
5768 unsigned long init_code_size, init_data_size;
5770 physpages = get_num_physpages();
5771 codesize = _etext - _stext;
5772 datasize = _edata - _sdata;
5773 rosize = __end_rodata - __start_rodata;
5774 bss_size = __bss_stop - __bss_start;
5775 init_data_size = __init_end - __init_begin;
5776 init_code_size = _einittext - _sinittext;
5779 * Detect special cases and adjust section sizes accordingly:
5780 * 1) .init.* may be embedded into .data sections
5781 * 2) .init.text.* may be out of [__init_begin, __init_end],
5782 * please refer to arch/tile/kernel/vmlinux.lds.S.
5783 * 3) .rodata.* may be embedded into .text or .data sections.
5785 #define adj_init_size(start, end, size, pos, adj) \
5787 if (start <= pos && pos < end && size > adj) \
5791 adj_init_size(__init_begin, __init_end, init_data_size,
5792 _sinittext, init_code_size);
5793 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5794 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5795 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5796 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5798 #undef adj_init_size
5800 pr_info("Memory: %luK/%luK available "
5801 "(%luK kernel code, %luK rwdata, %luK rodata, "
5802 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5803 #ifdef CONFIG_HIGHMEM
5807 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5808 codesize >> 10, datasize >> 10, rosize >> 10,
5809 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5810 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5811 totalcma_pages << (PAGE_SHIFT-10),
5812 #ifdef CONFIG_HIGHMEM
5813 totalhigh_pages << (PAGE_SHIFT-10),
5815 str ? ", " : "", str ? str : "");
5819 * set_dma_reserve - set the specified number of pages reserved in the first zone
5820 * @new_dma_reserve: The number of pages to mark reserved
5822 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5823 * In the DMA zone, a significant percentage may be consumed by kernel image
5824 * and other unfreeable allocations which can skew the watermarks badly. This
5825 * function may optionally be used to account for unfreeable pages in the
5826 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5827 * smaller per-cpu batchsize.
5829 void __init set_dma_reserve(unsigned long new_dma_reserve)
5831 dma_reserve = new_dma_reserve;
5834 void __init free_area_init(unsigned long *zones_size)
5836 free_area_init_node(0, zones_size,
5837 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5840 static int page_alloc_cpu_notify(struct notifier_block *self,
5841 unsigned long action, void *hcpu)
5843 int cpu = (unsigned long)hcpu;
5845 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5846 lru_add_drain_cpu(cpu);
5850 * Spill the event counters of the dead processor
5851 * into the current processors event counters.
5852 * This artificially elevates the count of the current
5855 vm_events_fold_cpu(cpu);
5858 * Zero the differential counters of the dead processor
5859 * so that the vm statistics are consistent.
5861 * This is only okay since the processor is dead and cannot
5862 * race with what we are doing.
5864 cpu_vm_stats_fold(cpu);
5869 void __init page_alloc_init(void)
5871 hotcpu_notifier(page_alloc_cpu_notify, 0);
5875 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5876 * or min_free_kbytes changes.
5878 static void calculate_totalreserve_pages(void)
5880 struct pglist_data *pgdat;
5881 unsigned long reserve_pages = 0;
5882 enum zone_type i, j;
5884 for_each_online_pgdat(pgdat) {
5885 for (i = 0; i < MAX_NR_ZONES; i++) {
5886 struct zone *zone = pgdat->node_zones + i;
5889 /* Find valid and maximum lowmem_reserve in the zone */
5890 for (j = i; j < MAX_NR_ZONES; j++) {
5891 if (zone->lowmem_reserve[j] > max)
5892 max = zone->lowmem_reserve[j];
5895 /* we treat the high watermark as reserved pages. */
5896 max += high_wmark_pages(zone);
5898 if (max > zone->managed_pages)
5899 max = zone->managed_pages;
5900 reserve_pages += max;
5902 * Lowmem reserves are not available to
5903 * GFP_HIGHUSER page cache allocations and
5904 * kswapd tries to balance zones to their high
5905 * watermark. As a result, neither should be
5906 * regarded as dirtyable memory, to prevent a
5907 * situation where reclaim has to clean pages
5908 * in order to balance the zones.
5910 zone->dirty_balance_reserve = max;
5913 dirty_balance_reserve = reserve_pages;
5914 totalreserve_pages = reserve_pages;
5918 * setup_per_zone_lowmem_reserve - called whenever
5919 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5920 * has a correct pages reserved value, so an adequate number of
5921 * pages are left in the zone after a successful __alloc_pages().
5923 static void setup_per_zone_lowmem_reserve(void)
5925 struct pglist_data *pgdat;
5926 enum zone_type j, idx;
5928 for_each_online_pgdat(pgdat) {
5929 for (j = 0; j < MAX_NR_ZONES; j++) {
5930 struct zone *zone = pgdat->node_zones + j;
5931 unsigned long managed_pages = zone->managed_pages;
5933 zone->lowmem_reserve[j] = 0;
5937 struct zone *lower_zone;
5941 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5942 sysctl_lowmem_reserve_ratio[idx] = 1;
5944 lower_zone = pgdat->node_zones + idx;
5945 lower_zone->lowmem_reserve[j] = managed_pages /
5946 sysctl_lowmem_reserve_ratio[idx];
5947 managed_pages += lower_zone->managed_pages;
5952 /* update totalreserve_pages */
5953 calculate_totalreserve_pages();
5956 static void __setup_per_zone_wmarks(void)
5958 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5959 unsigned long lowmem_pages = 0;
5961 unsigned long flags;
5963 /* Calculate total number of !ZONE_HIGHMEM pages */
5964 for_each_zone(zone) {
5965 if (!is_highmem(zone))
5966 lowmem_pages += zone->managed_pages;
5969 for_each_zone(zone) {
5972 spin_lock_irqsave(&zone->lock, flags);
5973 tmp = (u64)pages_min * zone->managed_pages;
5974 do_div(tmp, lowmem_pages);
5975 if (is_highmem(zone)) {
5977 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5978 * need highmem pages, so cap pages_min to a small
5981 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5982 * deltas control asynch page reclaim, and so should
5983 * not be capped for highmem.
5985 unsigned long min_pages;
5987 min_pages = zone->managed_pages / 1024;
5988 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5989 zone->watermark[WMARK_MIN] = min_pages;
5992 * If it's a lowmem zone, reserve a number of pages
5993 * proportionate to the zone's size.
5995 zone->watermark[WMARK_MIN] = tmp;
5998 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5999 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6001 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6002 high_wmark_pages(zone) - low_wmark_pages(zone) -
6003 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6005 spin_unlock_irqrestore(&zone->lock, flags);
6008 /* update totalreserve_pages */
6009 calculate_totalreserve_pages();
6013 * setup_per_zone_wmarks - called when min_free_kbytes changes
6014 * or when memory is hot-{added|removed}
6016 * Ensures that the watermark[min,low,high] values for each zone are set
6017 * correctly with respect to min_free_kbytes.
6019 void setup_per_zone_wmarks(void)
6021 mutex_lock(&zonelists_mutex);
6022 __setup_per_zone_wmarks();
6023 mutex_unlock(&zonelists_mutex);
6027 * The inactive anon list should be small enough that the VM never has to
6028 * do too much work, but large enough that each inactive page has a chance
6029 * to be referenced again before it is swapped out.
6031 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6032 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6033 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6034 * the anonymous pages are kept on the inactive list.
6037 * memory ratio inactive anon
6038 * -------------------------------------
6047 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6049 unsigned int gb, ratio;
6051 /* Zone size in gigabytes */
6052 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6054 ratio = int_sqrt(10 * gb);
6058 zone->inactive_ratio = ratio;
6061 static void __meminit setup_per_zone_inactive_ratio(void)
6066 calculate_zone_inactive_ratio(zone);
6070 * Initialise min_free_kbytes.
6072 * For small machines we want it small (128k min). For large machines
6073 * we want it large (64MB max). But it is not linear, because network
6074 * bandwidth does not increase linearly with machine size. We use
6076 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6077 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6093 int __meminit init_per_zone_wmark_min(void)
6095 unsigned long lowmem_kbytes;
6096 int new_min_free_kbytes;
6098 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6099 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6101 if (new_min_free_kbytes > user_min_free_kbytes) {
6102 min_free_kbytes = new_min_free_kbytes;
6103 if (min_free_kbytes < 128)
6104 min_free_kbytes = 128;
6105 if (min_free_kbytes > 65536)
6106 min_free_kbytes = 65536;
6108 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6109 new_min_free_kbytes, user_min_free_kbytes);
6111 setup_per_zone_wmarks();
6112 refresh_zone_stat_thresholds();
6113 setup_per_zone_lowmem_reserve();
6114 setup_per_zone_inactive_ratio();
6117 module_init(init_per_zone_wmark_min)
6120 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6121 * that we can call two helper functions whenever min_free_kbytes
6124 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6125 void __user *buffer, size_t *length, loff_t *ppos)
6129 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6134 user_min_free_kbytes = min_free_kbytes;
6135 setup_per_zone_wmarks();
6141 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6142 void __user *buffer, size_t *length, loff_t *ppos)
6147 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6152 zone->min_unmapped_pages = (zone->managed_pages *
6153 sysctl_min_unmapped_ratio) / 100;
6157 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6158 void __user *buffer, size_t *length, loff_t *ppos)
6163 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6168 zone->min_slab_pages = (zone->managed_pages *
6169 sysctl_min_slab_ratio) / 100;
6175 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6176 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6177 * whenever sysctl_lowmem_reserve_ratio changes.
6179 * The reserve ratio obviously has absolutely no relation with the
6180 * minimum watermarks. The lowmem reserve ratio can only make sense
6181 * if in function of the boot time zone sizes.
6183 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6184 void __user *buffer, size_t *length, loff_t *ppos)
6186 proc_dointvec_minmax(table, write, buffer, length, ppos);
6187 setup_per_zone_lowmem_reserve();
6192 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6193 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6194 * pagelist can have before it gets flushed back to buddy allocator.
6196 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6197 void __user *buffer, size_t *length, loff_t *ppos)
6200 int old_percpu_pagelist_fraction;
6203 mutex_lock(&pcp_batch_high_lock);
6204 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6206 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6207 if (!write || ret < 0)
6210 /* Sanity checking to avoid pcp imbalance */
6211 if (percpu_pagelist_fraction &&
6212 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6213 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6219 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6222 for_each_populated_zone(zone) {
6225 for_each_possible_cpu(cpu)
6226 pageset_set_high_and_batch(zone,
6227 per_cpu_ptr(zone->pageset, cpu));
6230 mutex_unlock(&pcp_batch_high_lock);
6235 int hashdist = HASHDIST_DEFAULT;
6237 static int __init set_hashdist(char *str)
6241 hashdist = simple_strtoul(str, &str, 0);
6244 __setup("hashdist=", set_hashdist);
6248 * allocate a large system hash table from bootmem
6249 * - it is assumed that the hash table must contain an exact power-of-2
6250 * quantity of entries
6251 * - limit is the number of hash buckets, not the total allocation size
6253 void *__init alloc_large_system_hash(const char *tablename,
6254 unsigned long bucketsize,
6255 unsigned long numentries,
6258 unsigned int *_hash_shift,
6259 unsigned int *_hash_mask,
6260 unsigned long low_limit,
6261 unsigned long high_limit)
6263 unsigned long long max = high_limit;
6264 unsigned long log2qty, size;
6267 /* allow the kernel cmdline to have a say */
6269 /* round applicable memory size up to nearest megabyte */
6270 numentries = nr_kernel_pages;
6272 /* It isn't necessary when PAGE_SIZE >= 1MB */
6273 if (PAGE_SHIFT < 20)
6274 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6276 /* limit to 1 bucket per 2^scale bytes of low memory */
6277 if (scale > PAGE_SHIFT)
6278 numentries >>= (scale - PAGE_SHIFT);
6280 numentries <<= (PAGE_SHIFT - scale);
6282 /* Make sure we've got at least a 0-order allocation.. */
6283 if (unlikely(flags & HASH_SMALL)) {
6284 /* Makes no sense without HASH_EARLY */
6285 WARN_ON(!(flags & HASH_EARLY));
6286 if (!(numentries >> *_hash_shift)) {
6287 numentries = 1UL << *_hash_shift;
6288 BUG_ON(!numentries);
6290 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6291 numentries = PAGE_SIZE / bucketsize;
6293 numentries = roundup_pow_of_two(numentries);
6295 /* limit allocation size to 1/16 total memory by default */
6297 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6298 do_div(max, bucketsize);
6300 max = min(max, 0x80000000ULL);
6302 if (numentries < low_limit)
6303 numentries = low_limit;
6304 if (numentries > max)
6307 log2qty = ilog2(numentries);
6310 size = bucketsize << log2qty;
6311 if (flags & HASH_EARLY)
6312 table = memblock_virt_alloc_nopanic(size, 0);
6314 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6317 * If bucketsize is not a power-of-two, we may free
6318 * some pages at the end of hash table which
6319 * alloc_pages_exact() automatically does
6321 if (get_order(size) < MAX_ORDER) {
6322 table = alloc_pages_exact(size, GFP_ATOMIC);
6323 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6326 } while (!table && size > PAGE_SIZE && --log2qty);
6329 panic("Failed to allocate %s hash table\n", tablename);
6331 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6334 ilog2(size) - PAGE_SHIFT,
6338 *_hash_shift = log2qty;
6340 *_hash_mask = (1 << log2qty) - 1;
6345 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6346 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6349 #ifdef CONFIG_SPARSEMEM
6350 return __pfn_to_section(pfn)->pageblock_flags;
6352 return zone->pageblock_flags;
6353 #endif /* CONFIG_SPARSEMEM */
6356 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6358 #ifdef CONFIG_SPARSEMEM
6359 pfn &= (PAGES_PER_SECTION-1);
6360 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6362 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6363 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6364 #endif /* CONFIG_SPARSEMEM */
6368 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6369 * @page: The page within the block of interest
6370 * @pfn: The target page frame number
6371 * @end_bitidx: The last bit of interest to retrieve
6372 * @mask: mask of bits that the caller is interested in
6374 * Return: pageblock_bits flags
6376 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6377 unsigned long end_bitidx,
6381 unsigned long *bitmap;
6382 unsigned long bitidx, word_bitidx;
6385 zone = page_zone(page);
6386 bitmap = get_pageblock_bitmap(zone, pfn);
6387 bitidx = pfn_to_bitidx(zone, pfn);
6388 word_bitidx = bitidx / BITS_PER_LONG;
6389 bitidx &= (BITS_PER_LONG-1);
6391 word = bitmap[word_bitidx];
6392 bitidx += end_bitidx;
6393 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6397 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6398 * @page: The page within the block of interest
6399 * @flags: The flags to set
6400 * @pfn: The target page frame number
6401 * @end_bitidx: The last bit of interest
6402 * @mask: mask of bits that the caller is interested in
6404 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6406 unsigned long end_bitidx,
6410 unsigned long *bitmap;
6411 unsigned long bitidx, word_bitidx;
6412 unsigned long old_word, word;
6414 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6416 zone = page_zone(page);
6417 bitmap = get_pageblock_bitmap(zone, pfn);
6418 bitidx = pfn_to_bitidx(zone, pfn);
6419 word_bitidx = bitidx / BITS_PER_LONG;
6420 bitidx &= (BITS_PER_LONG-1);
6422 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6424 bitidx += end_bitidx;
6425 mask <<= (BITS_PER_LONG - bitidx - 1);
6426 flags <<= (BITS_PER_LONG - bitidx - 1);
6428 word = READ_ONCE(bitmap[word_bitidx]);
6430 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6431 if (word == old_word)
6438 * This function checks whether pageblock includes unmovable pages or not.
6439 * If @count is not zero, it is okay to include less @count unmovable pages
6441 * PageLRU check without isolation or lru_lock could race so that
6442 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6443 * expect this function should be exact.
6445 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6446 bool skip_hwpoisoned_pages)
6448 unsigned long pfn, iter, found;
6452 * For avoiding noise data, lru_add_drain_all() should be called
6453 * If ZONE_MOVABLE, the zone never contains unmovable pages
6455 if (zone_idx(zone) == ZONE_MOVABLE)
6457 mt = get_pageblock_migratetype(page);
6458 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6461 pfn = page_to_pfn(page);
6462 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6463 unsigned long check = pfn + iter;
6465 if (!pfn_valid_within(check))
6468 page = pfn_to_page(check);
6471 * Hugepages are not in LRU lists, but they're movable.
6472 * We need not scan over tail pages bacause we don't
6473 * handle each tail page individually in migration.
6475 if (PageHuge(page)) {
6476 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6481 * We can't use page_count without pin a page
6482 * because another CPU can free compound page.
6483 * This check already skips compound tails of THP
6484 * because their page->_count is zero at all time.
6486 if (!atomic_read(&page->_count)) {
6487 if (PageBuddy(page))
6488 iter += (1 << page_order(page)) - 1;
6493 * The HWPoisoned page may be not in buddy system, and
6494 * page_count() is not 0.
6496 if (skip_hwpoisoned_pages && PageHWPoison(page))
6502 * If there are RECLAIMABLE pages, we need to check
6503 * it. But now, memory offline itself doesn't call
6504 * shrink_node_slabs() and it still to be fixed.
6507 * If the page is not RAM, page_count()should be 0.
6508 * we don't need more check. This is an _used_ not-movable page.
6510 * The problematic thing here is PG_reserved pages. PG_reserved
6511 * is set to both of a memory hole page and a _used_ kernel
6520 bool is_pageblock_removable_nolock(struct page *page)
6526 * We have to be careful here because we are iterating over memory
6527 * sections which are not zone aware so we might end up outside of
6528 * the zone but still within the section.
6529 * We have to take care about the node as well. If the node is offline
6530 * its NODE_DATA will be NULL - see page_zone.
6532 if (!node_online(page_to_nid(page)))
6535 zone = page_zone(page);
6536 pfn = page_to_pfn(page);
6537 if (!zone_spans_pfn(zone, pfn))
6540 return !has_unmovable_pages(zone, page, 0, true);
6545 static unsigned long pfn_max_align_down(unsigned long pfn)
6547 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6548 pageblock_nr_pages) - 1);
6551 static unsigned long pfn_max_align_up(unsigned long pfn)
6553 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6554 pageblock_nr_pages));
6557 /* [start, end) must belong to a single zone. */
6558 static int __alloc_contig_migrate_range(struct compact_control *cc,
6559 unsigned long start, unsigned long end)
6561 /* This function is based on compact_zone() from compaction.c. */
6562 unsigned long nr_reclaimed;
6563 unsigned long pfn = start;
6564 unsigned int tries = 0;
6569 while (pfn < end || !list_empty(&cc->migratepages)) {
6570 if (fatal_signal_pending(current)) {
6575 if (list_empty(&cc->migratepages)) {
6576 cc->nr_migratepages = 0;
6577 pfn = isolate_migratepages_range(cc, pfn, end);
6583 } else if (++tries == 5) {
6584 ret = ret < 0 ? ret : -EBUSY;
6588 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6590 cc->nr_migratepages -= nr_reclaimed;
6592 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6593 NULL, 0, cc->mode, MR_CMA);
6596 putback_movable_pages(&cc->migratepages);
6603 * alloc_contig_range() -- tries to allocate given range of pages
6604 * @start: start PFN to allocate
6605 * @end: one-past-the-last PFN to allocate
6606 * @migratetype: migratetype of the underlaying pageblocks (either
6607 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6608 * in range must have the same migratetype and it must
6609 * be either of the two.
6611 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6612 * aligned, however it's the caller's responsibility to guarantee that
6613 * we are the only thread that changes migrate type of pageblocks the
6616 * The PFN range must belong to a single zone.
6618 * Returns zero on success or negative error code. On success all
6619 * pages which PFN is in [start, end) are allocated for the caller and
6620 * need to be freed with free_contig_range().
6622 int alloc_contig_range(unsigned long start, unsigned long end,
6623 unsigned migratetype)
6625 unsigned long outer_start, outer_end;
6628 struct compact_control cc = {
6629 .nr_migratepages = 0,
6631 .zone = page_zone(pfn_to_page(start)),
6632 .mode = MIGRATE_SYNC,
6633 .ignore_skip_hint = true,
6635 INIT_LIST_HEAD(&cc.migratepages);
6638 * What we do here is we mark all pageblocks in range as
6639 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6640 * have different sizes, and due to the way page allocator
6641 * work, we align the range to biggest of the two pages so
6642 * that page allocator won't try to merge buddies from
6643 * different pageblocks and change MIGRATE_ISOLATE to some
6644 * other migration type.
6646 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6647 * migrate the pages from an unaligned range (ie. pages that
6648 * we are interested in). This will put all the pages in
6649 * range back to page allocator as MIGRATE_ISOLATE.
6651 * When this is done, we take the pages in range from page
6652 * allocator removing them from the buddy system. This way
6653 * page allocator will never consider using them.
6655 * This lets us mark the pageblocks back as
6656 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6657 * aligned range but not in the unaligned, original range are
6658 * put back to page allocator so that buddy can use them.
6661 ret = start_isolate_page_range(pfn_max_align_down(start),
6662 pfn_max_align_up(end), migratetype,
6667 ret = __alloc_contig_migrate_range(&cc, start, end);
6672 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6673 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6674 * more, all pages in [start, end) are free in page allocator.
6675 * What we are going to do is to allocate all pages from
6676 * [start, end) (that is remove them from page allocator).
6678 * The only problem is that pages at the beginning and at the
6679 * end of interesting range may be not aligned with pages that
6680 * page allocator holds, ie. they can be part of higher order
6681 * pages. Because of this, we reserve the bigger range and
6682 * once this is done free the pages we are not interested in.
6684 * We don't have to hold zone->lock here because the pages are
6685 * isolated thus they won't get removed from buddy.
6688 lru_add_drain_all();
6689 drain_all_pages(cc.zone);
6692 outer_start = start;
6693 while (!PageBuddy(pfn_to_page(outer_start))) {
6694 if (++order >= MAX_ORDER) {
6698 outer_start &= ~0UL << order;
6701 /* Make sure the range is really isolated. */
6702 if (test_pages_isolated(outer_start, end, false)) {
6703 pr_info("%s: [%lx, %lx) PFNs busy\n",
6704 __func__, outer_start, end);
6709 /* Grab isolated pages from freelists. */
6710 outer_end = isolate_freepages_range(&cc, outer_start, end);
6716 /* Free head and tail (if any) */
6717 if (start != outer_start)
6718 free_contig_range(outer_start, start - outer_start);
6719 if (end != outer_end)
6720 free_contig_range(end, outer_end - end);
6723 undo_isolate_page_range(pfn_max_align_down(start),
6724 pfn_max_align_up(end), migratetype);
6728 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6730 unsigned int count = 0;
6732 for (; nr_pages--; pfn++) {
6733 struct page *page = pfn_to_page(pfn);
6735 count += page_count(page) != 1;
6738 WARN(count != 0, "%d pages are still in use!\n", count);
6742 #ifdef CONFIG_MEMORY_HOTPLUG
6744 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6745 * page high values need to be recalulated.
6747 void __meminit zone_pcp_update(struct zone *zone)
6750 mutex_lock(&pcp_batch_high_lock);
6751 for_each_possible_cpu(cpu)
6752 pageset_set_high_and_batch(zone,
6753 per_cpu_ptr(zone->pageset, cpu));
6754 mutex_unlock(&pcp_batch_high_lock);
6758 void zone_pcp_reset(struct zone *zone)
6760 unsigned long flags;
6762 struct per_cpu_pageset *pset;
6764 /* avoid races with drain_pages() */
6765 local_irq_save(flags);
6766 if (zone->pageset != &boot_pageset) {
6767 for_each_online_cpu(cpu) {
6768 pset = per_cpu_ptr(zone->pageset, cpu);
6769 drain_zonestat(zone, pset);
6771 free_percpu(zone->pageset);
6772 zone->pageset = &boot_pageset;
6774 local_irq_restore(flags);
6777 #ifdef CONFIG_MEMORY_HOTREMOVE
6779 * All pages in the range must be isolated before calling this.
6782 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6786 unsigned int order, i;
6788 unsigned long flags;
6789 /* find the first valid pfn */
6790 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6795 zone = page_zone(pfn_to_page(pfn));
6796 spin_lock_irqsave(&zone->lock, flags);
6798 while (pfn < end_pfn) {
6799 if (!pfn_valid(pfn)) {
6803 page = pfn_to_page(pfn);
6805 * The HWPoisoned page may be not in buddy system, and
6806 * page_count() is not 0.
6808 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6810 SetPageReserved(page);
6814 BUG_ON(page_count(page));
6815 BUG_ON(!PageBuddy(page));
6816 order = page_order(page);
6817 #ifdef CONFIG_DEBUG_VM
6818 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6819 pfn, 1 << order, end_pfn);
6821 list_del(&page->lru);
6822 rmv_page_order(page);
6823 zone->free_area[order].nr_free--;
6824 for (i = 0; i < (1 << order); i++)
6825 SetPageReserved((page+i));
6826 pfn += (1 << order);
6828 spin_unlock_irqrestore(&zone->lock, flags);
6832 #ifdef CONFIG_MEMORY_FAILURE
6833 bool is_free_buddy_page(struct page *page)
6835 struct zone *zone = page_zone(page);
6836 unsigned long pfn = page_to_pfn(page);
6837 unsigned long flags;
6840 spin_lock_irqsave(&zone->lock, flags);
6841 for (order = 0; order < MAX_ORDER; order++) {
6842 struct page *page_head = page - (pfn & ((1 << order) - 1));
6844 if (PageBuddy(page_head) && page_order(page_head) >= order)
6847 spin_unlock_irqrestore(&zone->lock, flags);
6849 return order < MAX_ORDER;