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/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
69 #include <asm/sections.h>
70 #include <asm/tlbflush.h>
71 #include <asm/div64.h>
74 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
75 static DEFINE_MUTEX(pcp_batch_high_lock
);
76 #define MIN_PERCPU_PAGELIST_FRACTION (8)
78 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
79 DEFINE_PER_CPU(int, numa_node
);
80 EXPORT_PER_CPU_SYMBOL(numa_node
);
83 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
85 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
86 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
87 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
88 * defined in <linux/topology.h>.
90 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
91 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
92 int _node_numa_mem_
[MAX_NUMNODES
];
95 /* work_structs for global per-cpu drains */
96 DEFINE_MUTEX(pcpu_drain_mutex
);
97 DEFINE_PER_CPU(struct work_struct
, pcpu_drain
);
99 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
100 volatile unsigned long latent_entropy __latent_entropy
;
101 EXPORT_SYMBOL(latent_entropy
);
105 * Array of node states.
107 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
108 [N_POSSIBLE
] = NODE_MASK_ALL
,
109 [N_ONLINE
] = { { [0] = 1UL } },
111 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
112 #ifdef CONFIG_HIGHMEM
113 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
115 #ifdef CONFIG_MOVABLE_NODE
116 [N_MEMORY
] = { { [0] = 1UL } },
118 [N_CPU
] = { { [0] = 1UL } },
121 EXPORT_SYMBOL(node_states
);
123 /* Protect totalram_pages and zone->managed_pages */
124 static DEFINE_SPINLOCK(managed_page_count_lock
);
126 unsigned long totalram_pages __read_mostly
;
127 unsigned long totalreserve_pages __read_mostly
;
128 unsigned long totalcma_pages __read_mostly
;
130 int percpu_pagelist_fraction
;
131 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
134 * A cached value of the page's pageblock's migratetype, used when the page is
135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
136 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
137 * Also the migratetype set in the page does not necessarily match the pcplist
138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
139 * other index - this ensures that it will be put on the correct CMA freelist.
141 static inline int get_pcppage_migratetype(struct page
*page
)
146 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
148 page
->index
= migratetype
;
151 #ifdef CONFIG_PM_SLEEP
153 * The following functions are used by the suspend/hibernate code to temporarily
154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
155 * while devices are suspended. To avoid races with the suspend/hibernate code,
156 * they should always be called with pm_mutex held (gfp_allowed_mask also should
157 * only be modified with pm_mutex held, unless the suspend/hibernate code is
158 * guaranteed not to run in parallel with that modification).
161 static gfp_t saved_gfp_mask
;
163 void pm_restore_gfp_mask(void)
165 WARN_ON(!mutex_is_locked(&pm_mutex
));
166 if (saved_gfp_mask
) {
167 gfp_allowed_mask
= saved_gfp_mask
;
172 void pm_restrict_gfp_mask(void)
174 WARN_ON(!mutex_is_locked(&pm_mutex
));
175 WARN_ON(saved_gfp_mask
);
176 saved_gfp_mask
= gfp_allowed_mask
;
177 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
180 bool pm_suspended_storage(void)
182 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
186 #endif /* CONFIG_PM_SLEEP */
188 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
189 unsigned int pageblock_order __read_mostly
;
192 static void __free_pages_ok(struct page
*page
, unsigned int order
);
195 * results with 256, 32 in the lowmem_reserve sysctl:
196 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
197 * 1G machine -> (16M dma, 784M normal, 224M high)
198 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
199 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
200 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
202 * TBD: should special case ZONE_DMA32 machines here - in those we normally
203 * don't need any ZONE_NORMAL reservation
205 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
206 #ifdef CONFIG_ZONE_DMA
209 #ifdef CONFIG_ZONE_DMA32
212 #ifdef CONFIG_HIGHMEM
218 EXPORT_SYMBOL(totalram_pages
);
220 static char * const zone_names
[MAX_NR_ZONES
] = {
221 #ifdef CONFIG_ZONE_DMA
224 #ifdef CONFIG_ZONE_DMA32
228 #ifdef CONFIG_HIGHMEM
232 #ifdef CONFIG_ZONE_DEVICE
237 char * const migratetype_names
[MIGRATE_TYPES
] = {
245 #ifdef CONFIG_MEMORY_ISOLATION
250 compound_page_dtor
* const compound_page_dtors
[] = {
253 #ifdef CONFIG_HUGETLB_PAGE
256 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
261 int min_free_kbytes
= 1024;
262 int user_min_free_kbytes
= -1;
263 int watermark_scale_factor
= 10;
265 static unsigned long __meminitdata nr_kernel_pages
;
266 static unsigned long __meminitdata nr_all_pages
;
267 static unsigned long __meminitdata dma_reserve
;
269 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
270 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
271 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
272 static unsigned long __initdata required_kernelcore
;
273 static unsigned long __initdata required_movablecore
;
274 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
275 static bool mirrored_kernelcore
;
277 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
279 EXPORT_SYMBOL(movable_zone
);
280 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
283 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
284 int nr_online_nodes __read_mostly
= 1;
285 EXPORT_SYMBOL(nr_node_ids
);
286 EXPORT_SYMBOL(nr_online_nodes
);
289 int page_group_by_mobility_disabled __read_mostly
;
291 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
292 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
294 pgdat
->first_deferred_pfn
= ULONG_MAX
;
297 /* Returns true if the struct page for the pfn is uninitialised */
298 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
300 int nid
= early_pfn_to_nid(pfn
);
302 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
309 * Returns false when the remaining initialisation should be deferred until
310 * later in the boot cycle when it can be parallelised.
312 static inline bool update_defer_init(pg_data_t
*pgdat
,
313 unsigned long pfn
, unsigned long zone_end
,
314 unsigned long *nr_initialised
)
316 unsigned long max_initialise
;
318 /* Always populate low zones for address-contrained allocations */
319 if (zone_end
< pgdat_end_pfn(pgdat
))
322 * Initialise at least 2G of a node but also take into account that
323 * two large system hashes that can take up 1GB for 0.25TB/node.
325 max_initialise
= max(2UL << (30 - PAGE_SHIFT
),
326 (pgdat
->node_spanned_pages
>> 8));
329 if ((*nr_initialised
> max_initialise
) &&
330 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
331 pgdat
->first_deferred_pfn
= pfn
;
338 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
342 static inline bool early_page_uninitialised(unsigned long pfn
)
347 static inline bool update_defer_init(pg_data_t
*pgdat
,
348 unsigned long pfn
, unsigned long zone_end
,
349 unsigned long *nr_initialised
)
355 /* Return a pointer to the bitmap storing bits affecting a block of pages */
356 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
359 #ifdef CONFIG_SPARSEMEM
360 return __pfn_to_section(pfn
)->pageblock_flags
;
362 return page_zone(page
)->pageblock_flags
;
363 #endif /* CONFIG_SPARSEMEM */
366 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
368 #ifdef CONFIG_SPARSEMEM
369 pfn
&= (PAGES_PER_SECTION
-1);
370 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
372 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
373 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
374 #endif /* CONFIG_SPARSEMEM */
378 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
379 * @page: The page within the block of interest
380 * @pfn: The target page frame number
381 * @end_bitidx: The last bit of interest to retrieve
382 * @mask: mask of bits that the caller is interested in
384 * Return: pageblock_bits flags
386 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
388 unsigned long end_bitidx
,
391 unsigned long *bitmap
;
392 unsigned long bitidx
, word_bitidx
;
395 bitmap
= get_pageblock_bitmap(page
, pfn
);
396 bitidx
= pfn_to_bitidx(page
, pfn
);
397 word_bitidx
= bitidx
/ BITS_PER_LONG
;
398 bitidx
&= (BITS_PER_LONG
-1);
400 word
= bitmap
[word_bitidx
];
401 bitidx
+= end_bitidx
;
402 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
405 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
406 unsigned long end_bitidx
,
409 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
412 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
414 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
418 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
419 * @page: The page within the block of interest
420 * @flags: The flags to set
421 * @pfn: The target page frame number
422 * @end_bitidx: The last bit of interest
423 * @mask: mask of bits that the caller is interested in
425 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
427 unsigned long end_bitidx
,
430 unsigned long *bitmap
;
431 unsigned long bitidx
, word_bitidx
;
432 unsigned long old_word
, word
;
434 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
436 bitmap
= get_pageblock_bitmap(page
, pfn
);
437 bitidx
= pfn_to_bitidx(page
, pfn
);
438 word_bitidx
= bitidx
/ BITS_PER_LONG
;
439 bitidx
&= (BITS_PER_LONG
-1);
441 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
443 bitidx
+= end_bitidx
;
444 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
445 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
447 word
= READ_ONCE(bitmap
[word_bitidx
]);
449 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
450 if (word
== old_word
)
456 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
458 if (unlikely(page_group_by_mobility_disabled
&&
459 migratetype
< MIGRATE_PCPTYPES
))
460 migratetype
= MIGRATE_UNMOVABLE
;
462 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
463 PB_migrate
, PB_migrate_end
);
466 #ifdef CONFIG_DEBUG_VM
467 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
471 unsigned long pfn
= page_to_pfn(page
);
472 unsigned long sp
, start_pfn
;
475 seq
= zone_span_seqbegin(zone
);
476 start_pfn
= zone
->zone_start_pfn
;
477 sp
= zone
->spanned_pages
;
478 if (!zone_spans_pfn(zone
, pfn
))
480 } while (zone_span_seqretry(zone
, seq
));
483 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
484 pfn
, zone_to_nid(zone
), zone
->name
,
485 start_pfn
, start_pfn
+ sp
);
490 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
492 if (!pfn_valid_within(page_to_pfn(page
)))
494 if (zone
!= page_zone(page
))
500 * Temporary debugging check for pages not lying within a given zone.
502 static int bad_range(struct zone
*zone
, struct page
*page
)
504 if (page_outside_zone_boundaries(zone
, page
))
506 if (!page_is_consistent(zone
, page
))
512 static inline int bad_range(struct zone
*zone
, struct page
*page
)
518 static void bad_page(struct page
*page
, const char *reason
,
519 unsigned long bad_flags
)
521 static unsigned long resume
;
522 static unsigned long nr_shown
;
523 static unsigned long nr_unshown
;
526 * Allow a burst of 60 reports, then keep quiet for that minute;
527 * or allow a steady drip of one report per second.
529 if (nr_shown
== 60) {
530 if (time_before(jiffies
, resume
)) {
536 "BUG: Bad page state: %lu messages suppressed\n",
543 resume
= jiffies
+ 60 * HZ
;
545 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
546 current
->comm
, page_to_pfn(page
));
547 __dump_page(page
, reason
);
548 bad_flags
&= page
->flags
;
550 pr_alert("bad because of flags: %#lx(%pGp)\n",
551 bad_flags
, &bad_flags
);
552 dump_page_owner(page
);
557 /* Leave bad fields for debug, except PageBuddy could make trouble */
558 page_mapcount_reset(page
); /* remove PageBuddy */
559 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
563 * Higher-order pages are called "compound pages". They are structured thusly:
565 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
567 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
568 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
570 * The first tail page's ->compound_dtor holds the offset in array of compound
571 * page destructors. See compound_page_dtors.
573 * The first tail page's ->compound_order holds the order of allocation.
574 * This usage means that zero-order pages may not be compound.
577 void free_compound_page(struct page
*page
)
579 __free_pages_ok(page
, compound_order(page
));
582 void prep_compound_page(struct page
*page
, unsigned int order
)
585 int nr_pages
= 1 << order
;
587 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
588 set_compound_order(page
, order
);
590 for (i
= 1; i
< nr_pages
; i
++) {
591 struct page
*p
= page
+ i
;
592 set_page_count(p
, 0);
593 p
->mapping
= TAIL_MAPPING
;
594 set_compound_head(p
, page
);
596 atomic_set(compound_mapcount_ptr(page
), -1);
599 #ifdef CONFIG_DEBUG_PAGEALLOC
600 unsigned int _debug_guardpage_minorder
;
601 bool _debug_pagealloc_enabled __read_mostly
602 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
603 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
604 bool _debug_guardpage_enabled __read_mostly
;
606 static int __init
early_debug_pagealloc(char *buf
)
610 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
612 early_param("debug_pagealloc", early_debug_pagealloc
);
614 static bool need_debug_guardpage(void)
616 /* If we don't use debug_pagealloc, we don't need guard page */
617 if (!debug_pagealloc_enabled())
620 if (!debug_guardpage_minorder())
626 static void init_debug_guardpage(void)
628 if (!debug_pagealloc_enabled())
631 if (!debug_guardpage_minorder())
634 _debug_guardpage_enabled
= true;
637 struct page_ext_operations debug_guardpage_ops
= {
638 .need
= need_debug_guardpage
,
639 .init
= init_debug_guardpage
,
642 static int __init
debug_guardpage_minorder_setup(char *buf
)
646 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
647 pr_err("Bad debug_guardpage_minorder value\n");
650 _debug_guardpage_minorder
= res
;
651 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
654 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
656 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
657 unsigned int order
, int migratetype
)
659 struct page_ext
*page_ext
;
661 if (!debug_guardpage_enabled())
664 if (order
>= debug_guardpage_minorder())
667 page_ext
= lookup_page_ext(page
);
668 if (unlikely(!page_ext
))
671 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
673 INIT_LIST_HEAD(&page
->lru
);
674 set_page_private(page
, order
);
675 /* Guard pages are not available for any usage */
676 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
681 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
682 unsigned int order
, int migratetype
)
684 struct page_ext
*page_ext
;
686 if (!debug_guardpage_enabled())
689 page_ext
= lookup_page_ext(page
);
690 if (unlikely(!page_ext
))
693 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
695 set_page_private(page
, 0);
696 if (!is_migrate_isolate(migratetype
))
697 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
700 struct page_ext_operations debug_guardpage_ops
;
701 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
702 unsigned int order
, int migratetype
) { return false; }
703 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
704 unsigned int order
, int migratetype
) {}
707 static inline void set_page_order(struct page
*page
, unsigned int order
)
709 set_page_private(page
, order
);
710 __SetPageBuddy(page
);
713 static inline void rmv_page_order(struct page
*page
)
715 __ClearPageBuddy(page
);
716 set_page_private(page
, 0);
720 * This function checks whether a page is free && is the buddy
721 * we can do coalesce a page and its buddy if
722 * (a) the buddy is not in a hole (check before calling!) &&
723 * (b) the buddy is in the buddy system &&
724 * (c) a page and its buddy have the same order &&
725 * (d) a page and its buddy are in the same zone.
727 * For recording whether a page is in the buddy system, we set ->_mapcount
728 * PAGE_BUDDY_MAPCOUNT_VALUE.
729 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
730 * serialized by zone->lock.
732 * For recording page's order, we use page_private(page).
734 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
737 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
738 if (page_zone_id(page
) != page_zone_id(buddy
))
741 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
746 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
748 * zone check is done late to avoid uselessly
749 * calculating zone/node ids for pages that could
752 if (page_zone_id(page
) != page_zone_id(buddy
))
755 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
763 * Freeing function for a buddy system allocator.
765 * The concept of a buddy system is to maintain direct-mapped table
766 * (containing bit values) for memory blocks of various "orders".
767 * The bottom level table contains the map for the smallest allocatable
768 * units of memory (here, pages), and each level above it describes
769 * pairs of units from the levels below, hence, "buddies".
770 * At a high level, all that happens here is marking the table entry
771 * at the bottom level available, and propagating the changes upward
772 * as necessary, plus some accounting needed to play nicely with other
773 * parts of the VM system.
774 * At each level, we keep a list of pages, which are heads of continuous
775 * free pages of length of (1 << order) and marked with _mapcount
776 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
778 * So when we are allocating or freeing one, we can derive the state of the
779 * other. That is, if we allocate a small block, and both were
780 * free, the remainder of the region must be split into blocks.
781 * If a block is freed, and its buddy is also free, then this
782 * triggers coalescing into a block of larger size.
787 static inline void __free_one_page(struct page
*page
,
789 struct zone
*zone
, unsigned int order
,
792 unsigned long combined_pfn
;
793 unsigned long uninitialized_var(buddy_pfn
);
795 unsigned int max_order
;
797 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
799 VM_BUG_ON(!zone_is_initialized(zone
));
800 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
802 VM_BUG_ON(migratetype
== -1);
803 if (likely(!is_migrate_isolate(migratetype
)))
804 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
806 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
807 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
810 while (order
< max_order
- 1) {
811 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
812 buddy
= page
+ (buddy_pfn
- pfn
);
814 if (!pfn_valid_within(buddy_pfn
))
816 if (!page_is_buddy(page
, buddy
, order
))
819 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
820 * merge with it and move up one order.
822 if (page_is_guard(buddy
)) {
823 clear_page_guard(zone
, buddy
, order
, migratetype
);
825 list_del(&buddy
->lru
);
826 zone
->free_area
[order
].nr_free
--;
827 rmv_page_order(buddy
);
829 combined_pfn
= buddy_pfn
& pfn
;
830 page
= page
+ (combined_pfn
- pfn
);
834 if (max_order
< MAX_ORDER
) {
835 /* If we are here, it means order is >= pageblock_order.
836 * We want to prevent merge between freepages on isolate
837 * pageblock and normal pageblock. Without this, pageblock
838 * isolation could cause incorrect freepage or CMA accounting.
840 * We don't want to hit this code for the more frequent
843 if (unlikely(has_isolate_pageblock(zone
))) {
846 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
847 buddy
= page
+ (buddy_pfn
- pfn
);
848 buddy_mt
= get_pageblock_migratetype(buddy
);
850 if (migratetype
!= buddy_mt
851 && (is_migrate_isolate(migratetype
) ||
852 is_migrate_isolate(buddy_mt
)))
856 goto continue_merging
;
860 set_page_order(page
, order
);
863 * If this is not the largest possible page, check if the buddy
864 * of the next-highest order is free. If it is, it's possible
865 * that pages are being freed that will coalesce soon. In case,
866 * that is happening, add the free page to the tail of the list
867 * so it's less likely to be used soon and more likely to be merged
868 * as a higher order page
870 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
871 struct page
*higher_page
, *higher_buddy
;
872 combined_pfn
= buddy_pfn
& pfn
;
873 higher_page
= page
+ (combined_pfn
- pfn
);
874 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
875 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
876 if (pfn_valid_within(buddy_pfn
) &&
877 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
878 list_add_tail(&page
->lru
,
879 &zone
->free_area
[order
].free_list
[migratetype
]);
884 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
886 zone
->free_area
[order
].nr_free
++;
890 * A bad page could be due to a number of fields. Instead of multiple branches,
891 * try and check multiple fields with one check. The caller must do a detailed
892 * check if necessary.
894 static inline bool page_expected_state(struct page
*page
,
895 unsigned long check_flags
)
897 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
900 if (unlikely((unsigned long)page
->mapping
|
901 page_ref_count(page
) |
903 (unsigned long)page
->mem_cgroup
|
905 (page
->flags
& check_flags
)))
911 static void free_pages_check_bad(struct page
*page
)
913 const char *bad_reason
;
914 unsigned long bad_flags
;
919 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
920 bad_reason
= "nonzero mapcount";
921 if (unlikely(page
->mapping
!= NULL
))
922 bad_reason
= "non-NULL mapping";
923 if (unlikely(page_ref_count(page
) != 0))
924 bad_reason
= "nonzero _refcount";
925 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
926 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
927 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
930 if (unlikely(page
->mem_cgroup
))
931 bad_reason
= "page still charged to cgroup";
933 bad_page(page
, bad_reason
, bad_flags
);
936 static inline int free_pages_check(struct page
*page
)
938 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
941 /* Something has gone sideways, find it */
942 free_pages_check_bad(page
);
946 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
951 * We rely page->lru.next never has bit 0 set, unless the page
952 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
954 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
956 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
960 switch (page
- head_page
) {
962 /* the first tail page: ->mapping is compound_mapcount() */
963 if (unlikely(compound_mapcount(page
))) {
964 bad_page(page
, "nonzero compound_mapcount", 0);
970 * the second tail page: ->mapping is
971 * page_deferred_list().next -- ignore value.
975 if (page
->mapping
!= TAIL_MAPPING
) {
976 bad_page(page
, "corrupted mapping in tail page", 0);
981 if (unlikely(!PageTail(page
))) {
982 bad_page(page
, "PageTail not set", 0);
985 if (unlikely(compound_head(page
) != head_page
)) {
986 bad_page(page
, "compound_head not consistent", 0);
991 page
->mapping
= NULL
;
992 clear_compound_head(page
);
996 static __always_inline
bool free_pages_prepare(struct page
*page
,
997 unsigned int order
, bool check_free
)
1001 VM_BUG_ON_PAGE(PageTail(page
), page
);
1003 trace_mm_page_free(page
, order
);
1004 kmemcheck_free_shadow(page
, order
);
1007 * Check tail pages before head page information is cleared to
1008 * avoid checking PageCompound for order-0 pages.
1010 if (unlikely(order
)) {
1011 bool compound
= PageCompound(page
);
1014 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1017 ClearPageDoubleMap(page
);
1018 for (i
= 1; i
< (1 << order
); i
++) {
1020 bad
+= free_tail_pages_check(page
, page
+ i
);
1021 if (unlikely(free_pages_check(page
+ i
))) {
1025 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1028 if (PageMappingFlags(page
))
1029 page
->mapping
= NULL
;
1030 if (memcg_kmem_enabled() && PageKmemcg(page
))
1031 memcg_kmem_uncharge(page
, order
);
1033 bad
+= free_pages_check(page
);
1037 page_cpupid_reset_last(page
);
1038 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1039 reset_page_owner(page
, order
);
1041 if (!PageHighMem(page
)) {
1042 debug_check_no_locks_freed(page_address(page
),
1043 PAGE_SIZE
<< order
);
1044 debug_check_no_obj_freed(page_address(page
),
1045 PAGE_SIZE
<< order
);
1047 arch_free_page(page
, order
);
1048 kernel_poison_pages(page
, 1 << order
, 0);
1049 kernel_map_pages(page
, 1 << order
, 0);
1050 kasan_free_pages(page
, order
);
1055 #ifdef CONFIG_DEBUG_VM
1056 static inline bool free_pcp_prepare(struct page
*page
)
1058 return free_pages_prepare(page
, 0, true);
1061 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1066 static bool free_pcp_prepare(struct page
*page
)
1068 return free_pages_prepare(page
, 0, false);
1071 static bool bulkfree_pcp_prepare(struct page
*page
)
1073 return free_pages_check(page
);
1075 #endif /* CONFIG_DEBUG_VM */
1078 * Frees a number of pages from the PCP lists
1079 * Assumes all pages on list are in same zone, and of same order.
1080 * count is the number of pages to free.
1082 * If the zone was previously in an "all pages pinned" state then look to
1083 * see if this freeing clears that state.
1085 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1086 * pinned" detection logic.
1088 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1089 struct per_cpu_pages
*pcp
)
1091 int migratetype
= 0;
1093 unsigned long nr_scanned
, flags
;
1094 bool isolated_pageblocks
;
1096 spin_lock_irqsave(&zone
->lock
, flags
);
1097 isolated_pageblocks
= has_isolate_pageblock(zone
);
1098 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1100 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1104 struct list_head
*list
;
1107 * Remove pages from lists in a round-robin fashion. A
1108 * batch_free count is maintained that is incremented when an
1109 * empty list is encountered. This is so more pages are freed
1110 * off fuller lists instead of spinning excessively around empty
1115 if (++migratetype
== MIGRATE_PCPTYPES
)
1117 list
= &pcp
->lists
[migratetype
];
1118 } while (list_empty(list
));
1120 /* This is the only non-empty list. Free them all. */
1121 if (batch_free
== MIGRATE_PCPTYPES
)
1125 int mt
; /* migratetype of the to-be-freed page */
1127 page
= list_last_entry(list
, struct page
, lru
);
1128 /* must delete as __free_one_page list manipulates */
1129 list_del(&page
->lru
);
1131 mt
= get_pcppage_migratetype(page
);
1132 /* MIGRATE_ISOLATE page should not go to pcplists */
1133 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1134 /* Pageblock could have been isolated meanwhile */
1135 if (unlikely(isolated_pageblocks
))
1136 mt
= get_pageblock_migratetype(page
);
1138 if (bulkfree_pcp_prepare(page
))
1141 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1142 trace_mm_page_pcpu_drain(page
, 0, mt
);
1143 } while (--count
&& --batch_free
&& !list_empty(list
));
1145 spin_unlock_irqrestore(&zone
->lock
, flags
);
1148 static void free_one_page(struct zone
*zone
,
1149 struct page
*page
, unsigned long pfn
,
1153 unsigned long nr_scanned
, flags
;
1154 spin_lock_irqsave(&zone
->lock
, flags
);
1155 __count_vm_events(PGFREE
, 1 << order
);
1156 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1158 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1160 if (unlikely(has_isolate_pageblock(zone
) ||
1161 is_migrate_isolate(migratetype
))) {
1162 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1164 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1165 spin_unlock_irqrestore(&zone
->lock
, flags
);
1168 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1169 unsigned long zone
, int nid
)
1171 set_page_links(page
, zone
, nid
, pfn
);
1172 init_page_count(page
);
1173 page_mapcount_reset(page
);
1174 page_cpupid_reset_last(page
);
1176 INIT_LIST_HEAD(&page
->lru
);
1177 #ifdef WANT_PAGE_VIRTUAL
1178 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1179 if (!is_highmem_idx(zone
))
1180 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1184 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1187 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1190 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1191 static void init_reserved_page(unsigned long pfn
)
1196 if (!early_page_uninitialised(pfn
))
1199 nid
= early_pfn_to_nid(pfn
);
1200 pgdat
= NODE_DATA(nid
);
1202 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1203 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1205 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1208 __init_single_pfn(pfn
, zid
, nid
);
1211 static inline void init_reserved_page(unsigned long pfn
)
1214 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1217 * Initialised pages do not have PageReserved set. This function is
1218 * called for each range allocated by the bootmem allocator and
1219 * marks the pages PageReserved. The remaining valid pages are later
1220 * sent to the buddy page allocator.
1222 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1224 unsigned long start_pfn
= PFN_DOWN(start
);
1225 unsigned long end_pfn
= PFN_UP(end
);
1227 for (; start_pfn
< end_pfn
; start_pfn
++) {
1228 if (pfn_valid(start_pfn
)) {
1229 struct page
*page
= pfn_to_page(start_pfn
);
1231 init_reserved_page(start_pfn
);
1233 /* Avoid false-positive PageTail() */
1234 INIT_LIST_HEAD(&page
->lru
);
1236 SetPageReserved(page
);
1241 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1244 unsigned long pfn
= page_to_pfn(page
);
1246 if (!free_pages_prepare(page
, order
, true))
1249 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1250 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1253 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1255 unsigned int nr_pages
= 1 << order
;
1256 struct page
*p
= page
;
1260 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1262 __ClearPageReserved(p
);
1263 set_page_count(p
, 0);
1265 __ClearPageReserved(p
);
1266 set_page_count(p
, 0);
1268 page_zone(page
)->managed_pages
+= nr_pages
;
1269 set_page_refcounted(page
);
1270 __free_pages(page
, order
);
1273 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1274 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1276 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1278 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1280 static DEFINE_SPINLOCK(early_pfn_lock
);
1283 spin_lock(&early_pfn_lock
);
1284 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1286 nid
= first_online_node
;
1287 spin_unlock(&early_pfn_lock
);
1293 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1294 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1295 struct mminit_pfnnid_cache
*state
)
1299 nid
= __early_pfn_to_nid(pfn
, state
);
1300 if (nid
>= 0 && nid
!= node
)
1305 /* Only safe to use early in boot when initialisation is single-threaded */
1306 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1308 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1313 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1317 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1318 struct mminit_pfnnid_cache
*state
)
1325 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1328 if (early_page_uninitialised(pfn
))
1330 return __free_pages_boot_core(page
, order
);
1334 * Check that the whole (or subset of) a pageblock given by the interval of
1335 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1336 * with the migration of free compaction scanner. The scanners then need to
1337 * use only pfn_valid_within() check for arches that allow holes within
1340 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1342 * It's possible on some configurations to have a setup like node0 node1 node0
1343 * i.e. it's possible that all pages within a zones range of pages do not
1344 * belong to a single zone. We assume that a border between node0 and node1
1345 * can occur within a single pageblock, but not a node0 node1 node0
1346 * interleaving within a single pageblock. It is therefore sufficient to check
1347 * the first and last page of a pageblock and avoid checking each individual
1348 * page in a pageblock.
1350 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1351 unsigned long end_pfn
, struct zone
*zone
)
1353 struct page
*start_page
;
1354 struct page
*end_page
;
1356 /* end_pfn is one past the range we are checking */
1359 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1362 start_page
= pfn_to_page(start_pfn
);
1364 if (page_zone(start_page
) != zone
)
1367 end_page
= pfn_to_page(end_pfn
);
1369 /* This gives a shorter code than deriving page_zone(end_page) */
1370 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1376 void set_zone_contiguous(struct zone
*zone
)
1378 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1379 unsigned long block_end_pfn
;
1381 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1382 for (; block_start_pfn
< zone_end_pfn(zone
);
1383 block_start_pfn
= block_end_pfn
,
1384 block_end_pfn
+= pageblock_nr_pages
) {
1386 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1388 if (!__pageblock_pfn_to_page(block_start_pfn
,
1389 block_end_pfn
, zone
))
1393 /* We confirm that there is no hole */
1394 zone
->contiguous
= true;
1397 void clear_zone_contiguous(struct zone
*zone
)
1399 zone
->contiguous
= false;
1402 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1403 static void __init
deferred_free_range(struct page
*page
,
1404 unsigned long pfn
, int nr_pages
)
1411 /* Free a large naturally-aligned chunk if possible */
1412 if (nr_pages
== pageblock_nr_pages
&&
1413 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1414 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1415 __free_pages_boot_core(page
, pageblock_order
);
1419 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1420 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1421 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1422 __free_pages_boot_core(page
, 0);
1426 /* Completion tracking for deferred_init_memmap() threads */
1427 static atomic_t pgdat_init_n_undone __initdata
;
1428 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1430 static inline void __init
pgdat_init_report_one_done(void)
1432 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1433 complete(&pgdat_init_all_done_comp
);
1436 /* Initialise remaining memory on a node */
1437 static int __init
deferred_init_memmap(void *data
)
1439 pg_data_t
*pgdat
= data
;
1440 int nid
= pgdat
->node_id
;
1441 struct mminit_pfnnid_cache nid_init_state
= { };
1442 unsigned long start
= jiffies
;
1443 unsigned long nr_pages
= 0;
1444 unsigned long walk_start
, walk_end
;
1447 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1448 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1450 if (first_init_pfn
== ULONG_MAX
) {
1451 pgdat_init_report_one_done();
1455 /* Bind memory initialisation thread to a local node if possible */
1456 if (!cpumask_empty(cpumask
))
1457 set_cpus_allowed_ptr(current
, cpumask
);
1459 /* Sanity check boundaries */
1460 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1461 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1462 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1464 /* Only the highest zone is deferred so find it */
1465 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1466 zone
= pgdat
->node_zones
+ zid
;
1467 if (first_init_pfn
< zone_end_pfn(zone
))
1471 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1472 unsigned long pfn
, end_pfn
;
1473 struct page
*page
= NULL
;
1474 struct page
*free_base_page
= NULL
;
1475 unsigned long free_base_pfn
= 0;
1478 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1479 pfn
= first_init_pfn
;
1480 if (pfn
< walk_start
)
1482 if (pfn
< zone
->zone_start_pfn
)
1483 pfn
= zone
->zone_start_pfn
;
1485 for (; pfn
< end_pfn
; pfn
++) {
1486 if (!pfn_valid_within(pfn
))
1490 * Ensure pfn_valid is checked every
1491 * pageblock_nr_pages for memory holes
1493 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1494 if (!pfn_valid(pfn
)) {
1500 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1505 /* Minimise pfn page lookups and scheduler checks */
1506 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1509 nr_pages
+= nr_to_free
;
1510 deferred_free_range(free_base_page
,
1511 free_base_pfn
, nr_to_free
);
1512 free_base_page
= NULL
;
1513 free_base_pfn
= nr_to_free
= 0;
1515 page
= pfn_to_page(pfn
);
1520 VM_BUG_ON(page_zone(page
) != zone
);
1524 __init_single_page(page
, pfn
, zid
, nid
);
1525 if (!free_base_page
) {
1526 free_base_page
= page
;
1527 free_base_pfn
= pfn
;
1532 /* Where possible, batch up pages for a single free */
1535 /* Free the current block of pages to allocator */
1536 nr_pages
+= nr_to_free
;
1537 deferred_free_range(free_base_page
, free_base_pfn
,
1539 free_base_page
= NULL
;
1540 free_base_pfn
= nr_to_free
= 0;
1542 /* Free the last block of pages to allocator */
1543 nr_pages
+= nr_to_free
;
1544 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1546 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1549 /* Sanity check that the next zone really is unpopulated */
1550 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1552 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1553 jiffies_to_msecs(jiffies
- start
));
1555 pgdat_init_report_one_done();
1558 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1560 void __init
page_alloc_init_late(void)
1564 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1567 /* There will be num_node_state(N_MEMORY) threads */
1568 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1569 for_each_node_state(nid
, N_MEMORY
) {
1570 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1573 /* Block until all are initialised */
1574 wait_for_completion(&pgdat_init_all_done_comp
);
1576 /* Reinit limits that are based on free pages after the kernel is up */
1577 files_maxfiles_init();
1580 for_each_populated_zone(zone
)
1581 set_zone_contiguous(zone
);
1585 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1586 void __init
init_cma_reserved_pageblock(struct page
*page
)
1588 unsigned i
= pageblock_nr_pages
;
1589 struct page
*p
= page
;
1592 __ClearPageReserved(p
);
1593 set_page_count(p
, 0);
1596 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1598 if (pageblock_order
>= MAX_ORDER
) {
1599 i
= pageblock_nr_pages
;
1602 set_page_refcounted(p
);
1603 __free_pages(p
, MAX_ORDER
- 1);
1604 p
+= MAX_ORDER_NR_PAGES
;
1605 } while (i
-= MAX_ORDER_NR_PAGES
);
1607 set_page_refcounted(page
);
1608 __free_pages(page
, pageblock_order
);
1611 adjust_managed_page_count(page
, pageblock_nr_pages
);
1616 * The order of subdivision here is critical for the IO subsystem.
1617 * Please do not alter this order without good reasons and regression
1618 * testing. Specifically, as large blocks of memory are subdivided,
1619 * the order in which smaller blocks are delivered depends on the order
1620 * they're subdivided in this function. This is the primary factor
1621 * influencing the order in which pages are delivered to the IO
1622 * subsystem according to empirical testing, and this is also justified
1623 * by considering the behavior of a buddy system containing a single
1624 * large block of memory acted on by a series of small allocations.
1625 * This behavior is a critical factor in sglist merging's success.
1629 static inline void expand(struct zone
*zone
, struct page
*page
,
1630 int low
, int high
, struct free_area
*area
,
1633 unsigned long size
= 1 << high
;
1635 while (high
> low
) {
1639 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1642 * Mark as guard pages (or page), that will allow to
1643 * merge back to allocator when buddy will be freed.
1644 * Corresponding page table entries will not be touched,
1645 * pages will stay not present in virtual address space
1647 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1650 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1652 set_page_order(&page
[size
], high
);
1656 static void check_new_page_bad(struct page
*page
)
1658 const char *bad_reason
= NULL
;
1659 unsigned long bad_flags
= 0;
1661 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1662 bad_reason
= "nonzero mapcount";
1663 if (unlikely(page
->mapping
!= NULL
))
1664 bad_reason
= "non-NULL mapping";
1665 if (unlikely(page_ref_count(page
) != 0))
1666 bad_reason
= "nonzero _count";
1667 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1668 bad_reason
= "HWPoisoned (hardware-corrupted)";
1669 bad_flags
= __PG_HWPOISON
;
1670 /* Don't complain about hwpoisoned pages */
1671 page_mapcount_reset(page
); /* remove PageBuddy */
1674 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1675 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1676 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1679 if (unlikely(page
->mem_cgroup
))
1680 bad_reason
= "page still charged to cgroup";
1682 bad_page(page
, bad_reason
, bad_flags
);
1686 * This page is about to be returned from the page allocator
1688 static inline int check_new_page(struct page
*page
)
1690 if (likely(page_expected_state(page
,
1691 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1694 check_new_page_bad(page
);
1698 static inline bool free_pages_prezeroed(bool poisoned
)
1700 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1701 page_poisoning_enabled() && poisoned
;
1704 #ifdef CONFIG_DEBUG_VM
1705 static bool check_pcp_refill(struct page
*page
)
1710 static bool check_new_pcp(struct page
*page
)
1712 return check_new_page(page
);
1715 static bool check_pcp_refill(struct page
*page
)
1717 return check_new_page(page
);
1719 static bool check_new_pcp(struct page
*page
)
1723 #endif /* CONFIG_DEBUG_VM */
1725 static bool check_new_pages(struct page
*page
, unsigned int order
)
1728 for (i
= 0; i
< (1 << order
); i
++) {
1729 struct page
*p
= page
+ i
;
1731 if (unlikely(check_new_page(p
)))
1738 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1741 set_page_private(page
, 0);
1742 set_page_refcounted(page
);
1744 arch_alloc_page(page
, order
);
1745 kernel_map_pages(page
, 1 << order
, 1);
1746 kernel_poison_pages(page
, 1 << order
, 1);
1747 kasan_alloc_pages(page
, order
);
1748 set_page_owner(page
, order
, gfp_flags
);
1751 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1752 unsigned int alloc_flags
)
1755 bool poisoned
= true;
1757 for (i
= 0; i
< (1 << order
); i
++) {
1758 struct page
*p
= page
+ i
;
1760 poisoned
&= page_is_poisoned(p
);
1763 post_alloc_hook(page
, order
, gfp_flags
);
1765 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1766 for (i
= 0; i
< (1 << order
); i
++)
1767 clear_highpage(page
+ i
);
1769 if (order
&& (gfp_flags
& __GFP_COMP
))
1770 prep_compound_page(page
, order
);
1773 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1774 * allocate the page. The expectation is that the caller is taking
1775 * steps that will free more memory. The caller should avoid the page
1776 * being used for !PFMEMALLOC purposes.
1778 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1779 set_page_pfmemalloc(page
);
1781 clear_page_pfmemalloc(page
);
1785 * Go through the free lists for the given migratetype and remove
1786 * the smallest available page from the freelists
1789 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1792 unsigned int current_order
;
1793 struct free_area
*area
;
1796 /* Find a page of the appropriate size in the preferred list */
1797 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1798 area
= &(zone
->free_area
[current_order
]);
1799 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1803 list_del(&page
->lru
);
1804 rmv_page_order(page
);
1806 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1807 set_pcppage_migratetype(page
, migratetype
);
1816 * This array describes the order lists are fallen back to when
1817 * the free lists for the desirable migrate type are depleted
1819 static int fallbacks
[MIGRATE_TYPES
][4] = {
1820 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1821 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1822 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1824 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1826 #ifdef CONFIG_MEMORY_ISOLATION
1827 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1832 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1835 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1838 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1839 unsigned int order
) { return NULL
; }
1843 * Move the free pages in a range to the free lists of the requested type.
1844 * Note that start_page and end_pages are not aligned on a pageblock
1845 * boundary. If alignment is required, use move_freepages_block()
1847 int move_freepages(struct zone
*zone
,
1848 struct page
*start_page
, struct page
*end_page
,
1853 int pages_moved
= 0;
1855 #ifndef CONFIG_HOLES_IN_ZONE
1857 * page_zone is not safe to call in this context when
1858 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1859 * anyway as we check zone boundaries in move_freepages_block().
1860 * Remove at a later date when no bug reports exist related to
1861 * grouping pages by mobility
1863 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1866 for (page
= start_page
; page
<= end_page
;) {
1867 if (!pfn_valid_within(page_to_pfn(page
))) {
1872 /* Make sure we are not inadvertently changing nodes */
1873 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1875 if (!PageBuddy(page
)) {
1880 order
= page_order(page
);
1881 list_move(&page
->lru
,
1882 &zone
->free_area
[order
].free_list
[migratetype
]);
1884 pages_moved
+= 1 << order
;
1890 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1893 unsigned long start_pfn
, end_pfn
;
1894 struct page
*start_page
, *end_page
;
1896 start_pfn
= page_to_pfn(page
);
1897 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1898 start_page
= pfn_to_page(start_pfn
);
1899 end_page
= start_page
+ pageblock_nr_pages
- 1;
1900 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1902 /* Do not cross zone boundaries */
1903 if (!zone_spans_pfn(zone
, start_pfn
))
1905 if (!zone_spans_pfn(zone
, end_pfn
))
1908 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1911 static void change_pageblock_range(struct page
*pageblock_page
,
1912 int start_order
, int migratetype
)
1914 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1916 while (nr_pageblocks
--) {
1917 set_pageblock_migratetype(pageblock_page
, migratetype
);
1918 pageblock_page
+= pageblock_nr_pages
;
1923 * When we are falling back to another migratetype during allocation, try to
1924 * steal extra free pages from the same pageblocks to satisfy further
1925 * allocations, instead of polluting multiple pageblocks.
1927 * If we are stealing a relatively large buddy page, it is likely there will
1928 * be more free pages in the pageblock, so try to steal them all. For
1929 * reclaimable and unmovable allocations, we steal regardless of page size,
1930 * as fragmentation caused by those allocations polluting movable pageblocks
1931 * is worse than movable allocations stealing from unmovable and reclaimable
1934 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1937 * Leaving this order check is intended, although there is
1938 * relaxed order check in next check. The reason is that
1939 * we can actually steal whole pageblock if this condition met,
1940 * but, below check doesn't guarantee it and that is just heuristic
1941 * so could be changed anytime.
1943 if (order
>= pageblock_order
)
1946 if (order
>= pageblock_order
/ 2 ||
1947 start_mt
== MIGRATE_RECLAIMABLE
||
1948 start_mt
== MIGRATE_UNMOVABLE
||
1949 page_group_by_mobility_disabled
)
1956 * This function implements actual steal behaviour. If order is large enough,
1957 * we can steal whole pageblock. If not, we first move freepages in this
1958 * pageblock and check whether half of pages are moved or not. If half of
1959 * pages are moved, we can change migratetype of pageblock and permanently
1960 * use it's pages as requested migratetype in the future.
1962 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1965 unsigned int current_order
= page_order(page
);
1968 /* Take ownership for orders >= pageblock_order */
1969 if (current_order
>= pageblock_order
) {
1970 change_pageblock_range(page
, current_order
, start_type
);
1974 pages
= move_freepages_block(zone
, page
, start_type
);
1976 /* Claim the whole block if over half of it is free */
1977 if (pages
>= (1 << (pageblock_order
-1)) ||
1978 page_group_by_mobility_disabled
)
1979 set_pageblock_migratetype(page
, start_type
);
1983 * Check whether there is a suitable fallback freepage with requested order.
1984 * If only_stealable is true, this function returns fallback_mt only if
1985 * we can steal other freepages all together. This would help to reduce
1986 * fragmentation due to mixed migratetype pages in one pageblock.
1988 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1989 int migratetype
, bool only_stealable
, bool *can_steal
)
1994 if (area
->nr_free
== 0)
1999 fallback_mt
= fallbacks
[migratetype
][i
];
2000 if (fallback_mt
== MIGRATE_TYPES
)
2003 if (list_empty(&area
->free_list
[fallback_mt
]))
2006 if (can_steal_fallback(order
, migratetype
))
2009 if (!only_stealable
)
2020 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2021 * there are no empty page blocks that contain a page with a suitable order
2023 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2024 unsigned int alloc_order
)
2027 unsigned long max_managed
, flags
;
2030 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2031 * Check is race-prone but harmless.
2033 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2034 if (zone
->nr_reserved_highatomic
>= max_managed
)
2037 spin_lock_irqsave(&zone
->lock
, flags
);
2039 /* Recheck the nr_reserved_highatomic limit under the lock */
2040 if (zone
->nr_reserved_highatomic
>= max_managed
)
2044 mt
= get_pageblock_migratetype(page
);
2045 if (mt
!= MIGRATE_HIGHATOMIC
&&
2046 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
2047 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2048 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2049 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
2053 spin_unlock_irqrestore(&zone
->lock
, flags
);
2057 * Used when an allocation is about to fail under memory pressure. This
2058 * potentially hurts the reliability of high-order allocations when under
2059 * intense memory pressure but failed atomic allocations should be easier
2060 * to recover from than an OOM.
2062 * If @force is true, try to unreserve a pageblock even though highatomic
2063 * pageblock is exhausted.
2065 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2068 struct zonelist
*zonelist
= ac
->zonelist
;
2069 unsigned long flags
;
2076 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2079 * Preserve at least one pageblock unless memory pressure
2082 if (!force
&& zone
->nr_reserved_highatomic
<=
2086 spin_lock_irqsave(&zone
->lock
, flags
);
2087 for (order
= 0; order
< MAX_ORDER
; order
++) {
2088 struct free_area
*area
= &(zone
->free_area
[order
]);
2090 page
= list_first_entry_or_null(
2091 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2097 * In page freeing path, migratetype change is racy so
2098 * we can counter several free pages in a pageblock
2099 * in this loop althoug we changed the pageblock type
2100 * from highatomic to ac->migratetype. So we should
2101 * adjust the count once.
2103 if (get_pageblock_migratetype(page
) ==
2104 MIGRATE_HIGHATOMIC
) {
2106 * It should never happen but changes to
2107 * locking could inadvertently allow a per-cpu
2108 * drain to add pages to MIGRATE_HIGHATOMIC
2109 * while unreserving so be safe and watch for
2112 zone
->nr_reserved_highatomic
-= min(
2114 zone
->nr_reserved_highatomic
);
2118 * Convert to ac->migratetype and avoid the normal
2119 * pageblock stealing heuristics. Minimally, the caller
2120 * is doing the work and needs the pages. More
2121 * importantly, if the block was always converted to
2122 * MIGRATE_UNMOVABLE or another type then the number
2123 * of pageblocks that cannot be completely freed
2126 set_pageblock_migratetype(page
, ac
->migratetype
);
2127 ret
= move_freepages_block(zone
, page
, ac
->migratetype
);
2129 spin_unlock_irqrestore(&zone
->lock
, flags
);
2133 spin_unlock_irqrestore(&zone
->lock
, flags
);
2139 /* Remove an element from the buddy allocator from the fallback list */
2140 static inline struct page
*
2141 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2143 struct free_area
*area
;
2144 unsigned int current_order
;
2149 /* Find the largest possible block of pages in the other list */
2150 for (current_order
= MAX_ORDER
-1;
2151 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2153 area
= &(zone
->free_area
[current_order
]);
2154 fallback_mt
= find_suitable_fallback(area
, current_order
,
2155 start_migratetype
, false, &can_steal
);
2156 if (fallback_mt
== -1)
2159 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2162 get_pageblock_migratetype(page
) != MIGRATE_HIGHATOMIC
)
2163 steal_suitable_fallback(zone
, page
, start_migratetype
);
2165 /* Remove the page from the freelists */
2167 list_del(&page
->lru
);
2168 rmv_page_order(page
);
2170 expand(zone
, page
, order
, current_order
, area
,
2173 * The pcppage_migratetype may differ from pageblock's
2174 * migratetype depending on the decisions in
2175 * find_suitable_fallback(). This is OK as long as it does not
2176 * differ for MIGRATE_CMA pageblocks. Those can be used as
2177 * fallback only via special __rmqueue_cma_fallback() function
2179 set_pcppage_migratetype(page
, start_migratetype
);
2181 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2182 start_migratetype
, fallback_mt
);
2191 * Do the hard work of removing an element from the buddy allocator.
2192 * Call me with the zone->lock already held.
2194 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2199 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2200 if (unlikely(!page
)) {
2201 if (migratetype
== MIGRATE_MOVABLE
)
2202 page
= __rmqueue_cma_fallback(zone
, order
);
2205 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2208 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2213 * Obtain a specified number of elements from the buddy allocator, all under
2214 * a single hold of the lock, for efficiency. Add them to the supplied list.
2215 * Returns the number of new pages which were placed at *list.
2217 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2218 unsigned long count
, struct list_head
*list
,
2219 int migratetype
, bool cold
)
2222 unsigned long flags
;
2224 spin_lock_irqsave(&zone
->lock
, flags
);
2225 for (i
= 0; i
< count
; ++i
) {
2226 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2227 if (unlikely(page
== NULL
))
2230 if (unlikely(check_pcp_refill(page
)))
2234 * Split buddy pages returned by expand() are received here
2235 * in physical page order. The page is added to the callers and
2236 * list and the list head then moves forward. From the callers
2237 * perspective, the linked list is ordered by page number in
2238 * some conditions. This is useful for IO devices that can
2239 * merge IO requests if the physical pages are ordered
2243 list_add(&page
->lru
, list
);
2245 list_add_tail(&page
->lru
, list
);
2248 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2249 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2254 * i pages were removed from the buddy list even if some leak due
2255 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2256 * on i. Do not confuse with 'alloced' which is the number of
2257 * pages added to the pcp list.
2259 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2260 spin_unlock_irqrestore(&zone
->lock
, flags
);
2266 * Called from the vmstat counter updater to drain pagesets of this
2267 * currently executing processor on remote nodes after they have
2270 * Note that this function must be called with the thread pinned to
2271 * a single processor.
2273 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2275 unsigned long flags
;
2276 int to_drain
, batch
;
2278 local_irq_save(flags
);
2279 batch
= READ_ONCE(pcp
->batch
);
2280 to_drain
= min(pcp
->count
, batch
);
2282 free_pcppages_bulk(zone
, to_drain
, pcp
);
2283 pcp
->count
-= to_drain
;
2285 local_irq_restore(flags
);
2290 * Drain pcplists of the indicated processor and zone.
2292 * The processor must either be the current processor and the
2293 * thread pinned to the current processor or a processor that
2296 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2298 unsigned long flags
;
2299 struct per_cpu_pageset
*pset
;
2300 struct per_cpu_pages
*pcp
;
2302 local_irq_save(flags
);
2303 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2307 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2310 local_irq_restore(flags
);
2314 * Drain pcplists of all zones on the indicated processor.
2316 * The processor must either be the current processor and the
2317 * thread pinned to the current processor or a processor that
2320 static void drain_pages(unsigned int cpu
)
2324 for_each_populated_zone(zone
) {
2325 drain_pages_zone(cpu
, zone
);
2330 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2332 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2333 * the single zone's pages.
2335 void drain_local_pages(struct zone
*zone
)
2337 int cpu
= smp_processor_id();
2340 drain_pages_zone(cpu
, zone
);
2345 static void drain_local_pages_wq(struct work_struct
*work
)
2348 * drain_all_pages doesn't use proper cpu hotplug protection so
2349 * we can race with cpu offline when the WQ can move this from
2350 * a cpu pinned worker to an unbound one. We can operate on a different
2351 * cpu which is allright but we also have to make sure to not move to
2355 drain_local_pages(NULL
);
2360 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2362 * When zone parameter is non-NULL, spill just the single zone's pages.
2364 * Note that this can be extremely slow as the draining happens in a workqueue.
2366 void drain_all_pages(struct zone
*zone
)
2371 * Allocate in the BSS so we wont require allocation in
2372 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2374 static cpumask_t cpus_with_pcps
;
2377 * Make sure nobody triggers this path before mm_percpu_wq is fully
2380 if (WARN_ON_ONCE(!mm_percpu_wq
))
2383 /* Workqueues cannot recurse */
2384 if (current
->flags
& PF_WQ_WORKER
)
2388 * Do not drain if one is already in progress unless it's specific to
2389 * a zone. Such callers are primarily CMA and memory hotplug and need
2390 * the drain to be complete when the call returns.
2392 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2395 mutex_lock(&pcpu_drain_mutex
);
2399 * We don't care about racing with CPU hotplug event
2400 * as offline notification will cause the notified
2401 * cpu to drain that CPU pcps and on_each_cpu_mask
2402 * disables preemption as part of its processing
2404 for_each_online_cpu(cpu
) {
2405 struct per_cpu_pageset
*pcp
;
2407 bool has_pcps
= false;
2410 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2414 for_each_populated_zone(z
) {
2415 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2416 if (pcp
->pcp
.count
) {
2424 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2426 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2429 for_each_cpu(cpu
, &cpus_with_pcps
) {
2430 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2431 INIT_WORK(work
, drain_local_pages_wq
);
2432 queue_work_on(cpu
, mm_percpu_wq
, work
);
2434 for_each_cpu(cpu
, &cpus_with_pcps
)
2435 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2437 mutex_unlock(&pcpu_drain_mutex
);
2440 #ifdef CONFIG_HIBERNATION
2442 void mark_free_pages(struct zone
*zone
)
2444 unsigned long pfn
, max_zone_pfn
;
2445 unsigned long flags
;
2446 unsigned int order
, t
;
2449 if (zone_is_empty(zone
))
2452 spin_lock_irqsave(&zone
->lock
, flags
);
2454 max_zone_pfn
= zone_end_pfn(zone
);
2455 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2456 if (pfn_valid(pfn
)) {
2457 page
= pfn_to_page(pfn
);
2459 if (page_zone(page
) != zone
)
2462 if (!swsusp_page_is_forbidden(page
))
2463 swsusp_unset_page_free(page
);
2466 for_each_migratetype_order(order
, t
) {
2467 list_for_each_entry(page
,
2468 &zone
->free_area
[order
].free_list
[t
], lru
) {
2471 pfn
= page_to_pfn(page
);
2472 for (i
= 0; i
< (1UL << order
); i
++)
2473 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2476 spin_unlock_irqrestore(&zone
->lock
, flags
);
2478 #endif /* CONFIG_PM */
2481 * Free a 0-order page
2482 * cold == true ? free a cold page : free a hot page
2484 void free_hot_cold_page(struct page
*page
, bool cold
)
2486 struct zone
*zone
= page_zone(page
);
2487 struct per_cpu_pages
*pcp
;
2488 unsigned long pfn
= page_to_pfn(page
);
2491 if (in_interrupt()) {
2492 __free_pages_ok(page
, 0);
2496 if (!free_pcp_prepare(page
))
2499 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2500 set_pcppage_migratetype(page
, migratetype
);
2504 * We only track unmovable, reclaimable and movable on pcp lists.
2505 * Free ISOLATE pages back to the allocator because they are being
2506 * offlined but treat RESERVE as movable pages so we can get those
2507 * areas back if necessary. Otherwise, we may have to free
2508 * excessively into the page allocator
2510 if (migratetype
>= MIGRATE_PCPTYPES
) {
2511 if (unlikely(is_migrate_isolate(migratetype
))) {
2512 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2515 migratetype
= MIGRATE_MOVABLE
;
2518 __count_vm_event(PGFREE
);
2519 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2521 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2523 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2525 if (pcp
->count
>= pcp
->high
) {
2526 unsigned long batch
= READ_ONCE(pcp
->batch
);
2527 free_pcppages_bulk(zone
, batch
, pcp
);
2528 pcp
->count
-= batch
;
2536 * Free a list of 0-order pages
2538 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2540 struct page
*page
, *next
;
2542 list_for_each_entry_safe(page
, next
, list
, lru
) {
2543 trace_mm_page_free_batched(page
, cold
);
2544 free_hot_cold_page(page
, cold
);
2549 * split_page takes a non-compound higher-order page, and splits it into
2550 * n (1<<order) sub-pages: page[0..n]
2551 * Each sub-page must be freed individually.
2553 * Note: this is probably too low level an operation for use in drivers.
2554 * Please consult with lkml before using this in your driver.
2556 void split_page(struct page
*page
, unsigned int order
)
2560 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2561 VM_BUG_ON_PAGE(!page_count(page
), page
);
2563 #ifdef CONFIG_KMEMCHECK
2565 * Split shadow pages too, because free(page[0]) would
2566 * otherwise free the whole shadow.
2568 if (kmemcheck_page_is_tracked(page
))
2569 split_page(virt_to_page(page
[0].shadow
), order
);
2572 for (i
= 1; i
< (1 << order
); i
++)
2573 set_page_refcounted(page
+ i
);
2574 split_page_owner(page
, order
);
2576 EXPORT_SYMBOL_GPL(split_page
);
2578 int __isolate_free_page(struct page
*page
, unsigned int order
)
2580 unsigned long watermark
;
2584 BUG_ON(!PageBuddy(page
));
2586 zone
= page_zone(page
);
2587 mt
= get_pageblock_migratetype(page
);
2589 if (!is_migrate_isolate(mt
)) {
2591 * Obey watermarks as if the page was being allocated. We can
2592 * emulate a high-order watermark check with a raised order-0
2593 * watermark, because we already know our high-order page
2596 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2597 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2600 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2603 /* Remove page from free list */
2604 list_del(&page
->lru
);
2605 zone
->free_area
[order
].nr_free
--;
2606 rmv_page_order(page
);
2609 * Set the pageblock if the isolated page is at least half of a
2612 if (order
>= pageblock_order
- 1) {
2613 struct page
*endpage
= page
+ (1 << order
) - 1;
2614 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2615 int mt
= get_pageblock_migratetype(page
);
2616 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2617 && mt
!= MIGRATE_HIGHATOMIC
)
2618 set_pageblock_migratetype(page
,
2624 return 1UL << order
;
2628 * Update NUMA hit/miss statistics
2630 * Must be called with interrupts disabled.
2632 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2635 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2637 if (z
->node
!= numa_node_id())
2638 local_stat
= NUMA_OTHER
;
2640 if (z
->node
== preferred_zone
->node
)
2641 __inc_zone_state(z
, NUMA_HIT
);
2643 __inc_zone_state(z
, NUMA_MISS
);
2644 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2646 __inc_zone_state(z
, local_stat
);
2650 /* Remove page from the per-cpu list, caller must protect the list */
2651 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2652 bool cold
, struct per_cpu_pages
*pcp
,
2653 struct list_head
*list
)
2657 VM_BUG_ON(in_interrupt());
2660 if (list_empty(list
)) {
2661 pcp
->count
+= rmqueue_bulk(zone
, 0,
2664 if (unlikely(list_empty(list
)))
2669 page
= list_last_entry(list
, struct page
, lru
);
2671 page
= list_first_entry(list
, struct page
, lru
);
2673 list_del(&page
->lru
);
2675 } while (check_new_pcp(page
));
2680 /* Lock and remove page from the per-cpu list */
2681 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2682 struct zone
*zone
, unsigned int order
,
2683 gfp_t gfp_flags
, int migratetype
)
2685 struct per_cpu_pages
*pcp
;
2686 struct list_head
*list
;
2687 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2691 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2692 list
= &pcp
->lists
[migratetype
];
2693 page
= __rmqueue_pcplist(zone
, migratetype
, cold
, pcp
, list
);
2695 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2696 zone_statistics(preferred_zone
, zone
);
2703 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2706 struct page
*rmqueue(struct zone
*preferred_zone
,
2707 struct zone
*zone
, unsigned int order
,
2708 gfp_t gfp_flags
, unsigned int alloc_flags
,
2711 unsigned long flags
;
2714 if (likely(order
== 0) && !in_interrupt()) {
2715 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
2716 gfp_flags
, migratetype
);
2721 * We most definitely don't want callers attempting to
2722 * allocate greater than order-1 page units with __GFP_NOFAIL.
2724 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2725 spin_lock_irqsave(&zone
->lock
, flags
);
2729 if (alloc_flags
& ALLOC_HARDER
) {
2730 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2732 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2735 page
= __rmqueue(zone
, order
, migratetype
);
2736 } while (page
&& check_new_pages(page
, order
));
2737 spin_unlock(&zone
->lock
);
2740 __mod_zone_freepage_state(zone
, -(1 << order
),
2741 get_pcppage_migratetype(page
));
2743 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2744 zone_statistics(preferred_zone
, zone
);
2745 local_irq_restore(flags
);
2748 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
2752 local_irq_restore(flags
);
2756 #ifdef CONFIG_FAIL_PAGE_ALLOC
2759 struct fault_attr attr
;
2761 bool ignore_gfp_highmem
;
2762 bool ignore_gfp_reclaim
;
2764 } fail_page_alloc
= {
2765 .attr
= FAULT_ATTR_INITIALIZER
,
2766 .ignore_gfp_reclaim
= true,
2767 .ignore_gfp_highmem
= true,
2771 static int __init
setup_fail_page_alloc(char *str
)
2773 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2775 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2777 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2779 if (order
< fail_page_alloc
.min_order
)
2781 if (gfp_mask
& __GFP_NOFAIL
)
2783 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2785 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2786 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2789 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2792 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2794 static int __init
fail_page_alloc_debugfs(void)
2796 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2799 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2800 &fail_page_alloc
.attr
);
2802 return PTR_ERR(dir
);
2804 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2805 &fail_page_alloc
.ignore_gfp_reclaim
))
2807 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2808 &fail_page_alloc
.ignore_gfp_highmem
))
2810 if (!debugfs_create_u32("min-order", mode
, dir
,
2811 &fail_page_alloc
.min_order
))
2816 debugfs_remove_recursive(dir
);
2821 late_initcall(fail_page_alloc_debugfs
);
2823 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2825 #else /* CONFIG_FAIL_PAGE_ALLOC */
2827 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2832 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2835 * Return true if free base pages are above 'mark'. For high-order checks it
2836 * will return true of the order-0 watermark is reached and there is at least
2837 * one free page of a suitable size. Checking now avoids taking the zone lock
2838 * to check in the allocation paths if no pages are free.
2840 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2841 int classzone_idx
, unsigned int alloc_flags
,
2846 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2848 /* free_pages may go negative - that's OK */
2849 free_pages
-= (1 << order
) - 1;
2851 if (alloc_flags
& ALLOC_HIGH
)
2855 * If the caller does not have rights to ALLOC_HARDER then subtract
2856 * the high-atomic reserves. This will over-estimate the size of the
2857 * atomic reserve but it avoids a search.
2859 if (likely(!alloc_harder
))
2860 free_pages
-= z
->nr_reserved_highatomic
;
2865 /* If allocation can't use CMA areas don't use free CMA pages */
2866 if (!(alloc_flags
& ALLOC_CMA
))
2867 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2871 * Check watermarks for an order-0 allocation request. If these
2872 * are not met, then a high-order request also cannot go ahead
2873 * even if a suitable page happened to be free.
2875 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2878 /* If this is an order-0 request then the watermark is fine */
2882 /* For a high-order request, check at least one suitable page is free */
2883 for (o
= order
; o
< MAX_ORDER
; o
++) {
2884 struct free_area
*area
= &z
->free_area
[o
];
2893 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2894 if (!list_empty(&area
->free_list
[mt
]))
2899 if ((alloc_flags
& ALLOC_CMA
) &&
2900 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2908 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2909 int classzone_idx
, unsigned int alloc_flags
)
2911 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2912 zone_page_state(z
, NR_FREE_PAGES
));
2915 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2916 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2918 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2922 /* If allocation can't use CMA areas don't use free CMA pages */
2923 if (!(alloc_flags
& ALLOC_CMA
))
2924 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2928 * Fast check for order-0 only. If this fails then the reserves
2929 * need to be calculated. There is a corner case where the check
2930 * passes but only the high-order atomic reserve are free. If
2931 * the caller is !atomic then it'll uselessly search the free
2932 * list. That corner case is then slower but it is harmless.
2934 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2937 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2941 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2942 unsigned long mark
, int classzone_idx
)
2944 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2946 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2947 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2949 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2954 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2956 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
2959 #else /* CONFIG_NUMA */
2960 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2964 #endif /* CONFIG_NUMA */
2967 * get_page_from_freelist goes through the zonelist trying to allocate
2970 static struct page
*
2971 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2972 const struct alloc_context
*ac
)
2974 struct zoneref
*z
= ac
->preferred_zoneref
;
2976 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
2979 * Scan zonelist, looking for a zone with enough free.
2980 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2982 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2987 if (cpusets_enabled() &&
2988 (alloc_flags
& ALLOC_CPUSET
) &&
2989 !__cpuset_zone_allowed(zone
, gfp_mask
))
2992 * When allocating a page cache page for writing, we
2993 * want to get it from a node that is within its dirty
2994 * limit, such that no single node holds more than its
2995 * proportional share of globally allowed dirty pages.
2996 * The dirty limits take into account the node's
2997 * lowmem reserves and high watermark so that kswapd
2998 * should be able to balance it without having to
2999 * write pages from its LRU list.
3001 * XXX: For now, allow allocations to potentially
3002 * exceed the per-node dirty limit in the slowpath
3003 * (spread_dirty_pages unset) before going into reclaim,
3004 * which is important when on a NUMA setup the allowed
3005 * nodes are together not big enough to reach the
3006 * global limit. The proper fix for these situations
3007 * will require awareness of nodes in the
3008 * dirty-throttling and the flusher threads.
3010 if (ac
->spread_dirty_pages
) {
3011 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3014 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3015 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3020 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
3021 if (!zone_watermark_fast(zone
, order
, mark
,
3022 ac_classzone_idx(ac
), alloc_flags
)) {
3025 /* Checked here to keep the fast path fast */
3026 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3027 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3030 if (node_reclaim_mode
== 0 ||
3031 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3034 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3036 case NODE_RECLAIM_NOSCAN
:
3039 case NODE_RECLAIM_FULL
:
3040 /* scanned but unreclaimable */
3043 /* did we reclaim enough */
3044 if (zone_watermark_ok(zone
, order
, mark
,
3045 ac_classzone_idx(ac
), alloc_flags
))
3053 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3054 gfp_mask
, alloc_flags
, ac
->migratetype
);
3056 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3059 * If this is a high-order atomic allocation then check
3060 * if the pageblock should be reserved for the future
3062 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3063 reserve_highatomic_pageblock(page
, zone
, order
);
3073 * Large machines with many possible nodes should not always dump per-node
3074 * meminfo in irq context.
3076 static inline bool should_suppress_show_mem(void)
3081 ret
= in_interrupt();
3086 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3088 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3089 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3091 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs
))
3095 * This documents exceptions given to allocations in certain
3096 * contexts that are allowed to allocate outside current's set
3099 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3100 if (test_thread_flag(TIF_MEMDIE
) ||
3101 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3102 filter
&= ~SHOW_MEM_FILTER_NODES
;
3103 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3104 filter
&= ~SHOW_MEM_FILTER_NODES
;
3106 show_mem(filter
, nodemask
);
3109 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3111 struct va_format vaf
;
3113 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3114 DEFAULT_RATELIMIT_BURST
);
3116 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
3117 debug_guardpage_minorder() > 0)
3120 pr_warn("%s: ", current
->comm
);
3122 va_start(args
, fmt
);
3125 pr_cont("%pV", &vaf
);
3128 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask
, &gfp_mask
);
3130 pr_cont("%*pbl\n", nodemask_pr_args(nodemask
));
3132 pr_cont("(null)\n");
3134 cpuset_print_current_mems_allowed();
3137 warn_alloc_show_mem(gfp_mask
, nodemask
);
3140 static inline struct page
*
3141 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3142 unsigned int alloc_flags
,
3143 const struct alloc_context
*ac
)
3147 page
= get_page_from_freelist(gfp_mask
, order
,
3148 alloc_flags
|ALLOC_CPUSET
, ac
);
3150 * fallback to ignore cpuset restriction if our nodes
3154 page
= get_page_from_freelist(gfp_mask
, order
,
3160 static inline struct page
*
3161 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3162 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3164 struct oom_control oc
= {
3165 .zonelist
= ac
->zonelist
,
3166 .nodemask
= ac
->nodemask
,
3168 .gfp_mask
= gfp_mask
,
3173 *did_some_progress
= 0;
3176 * Acquire the oom lock. If that fails, somebody else is
3177 * making progress for us.
3179 if (!mutex_trylock(&oom_lock
)) {
3180 *did_some_progress
= 1;
3181 schedule_timeout_uninterruptible(1);
3186 * Go through the zonelist yet one more time, keep very high watermark
3187 * here, this is only to catch a parallel oom killing, we must fail if
3188 * we're still under heavy pressure.
3190 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3191 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3195 /* Coredumps can quickly deplete all memory reserves */
3196 if (current
->flags
& PF_DUMPCORE
)
3198 /* The OOM killer will not help higher order allocs */
3199 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3201 /* The OOM killer does not needlessly kill tasks for lowmem */
3202 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3204 if (pm_suspended_storage())
3207 * XXX: GFP_NOFS allocations should rather fail than rely on
3208 * other request to make a forward progress.
3209 * We are in an unfortunate situation where out_of_memory cannot
3210 * do much for this context but let's try it to at least get
3211 * access to memory reserved if the current task is killed (see
3212 * out_of_memory). Once filesystems are ready to handle allocation
3213 * failures more gracefully we should just bail out here.
3216 /* The OOM killer may not free memory on a specific node */
3217 if (gfp_mask
& __GFP_THISNODE
)
3220 /* Exhausted what can be done so it's blamo time */
3221 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3222 *did_some_progress
= 1;
3225 * Help non-failing allocations by giving them access to memory
3228 if (gfp_mask
& __GFP_NOFAIL
)
3229 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3230 ALLOC_NO_WATERMARKS
, ac
);
3233 mutex_unlock(&oom_lock
);
3238 * Maximum number of compaction retries wit a progress before OOM
3239 * killer is consider as the only way to move forward.
3241 #define MAX_COMPACT_RETRIES 16
3243 #ifdef CONFIG_COMPACTION
3244 /* Try memory compaction for high-order allocations before reclaim */
3245 static struct page
*
3246 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3247 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3248 enum compact_priority prio
, enum compact_result
*compact_result
)
3255 current
->flags
|= PF_MEMALLOC
;
3256 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3258 current
->flags
&= ~PF_MEMALLOC
;
3260 if (*compact_result
<= COMPACT_INACTIVE
)
3264 * At least in one zone compaction wasn't deferred or skipped, so let's
3265 * count a compaction stall
3267 count_vm_event(COMPACTSTALL
);
3269 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3272 struct zone
*zone
= page_zone(page
);
3274 zone
->compact_blockskip_flush
= false;
3275 compaction_defer_reset(zone
, order
, true);
3276 count_vm_event(COMPACTSUCCESS
);
3281 * It's bad if compaction run occurs and fails. The most likely reason
3282 * is that pages exist, but not enough to satisfy watermarks.
3284 count_vm_event(COMPACTFAIL
);
3292 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3293 enum compact_result compact_result
,
3294 enum compact_priority
*compact_priority
,
3295 int *compaction_retries
)
3297 int max_retries
= MAX_COMPACT_RETRIES
;
3300 int retries
= *compaction_retries
;
3301 enum compact_priority priority
= *compact_priority
;
3306 if (compaction_made_progress(compact_result
))
3307 (*compaction_retries
)++;
3310 * compaction considers all the zone as desperately out of memory
3311 * so it doesn't really make much sense to retry except when the
3312 * failure could be caused by insufficient priority
3314 if (compaction_failed(compact_result
))
3315 goto check_priority
;
3318 * make sure the compaction wasn't deferred or didn't bail out early
3319 * due to locks contention before we declare that we should give up.
3320 * But do not retry if the given zonelist is not suitable for
3323 if (compaction_withdrawn(compact_result
)) {
3324 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3329 * !costly requests are much more important than __GFP_REPEAT
3330 * costly ones because they are de facto nofail and invoke OOM
3331 * killer to move on while costly can fail and users are ready
3332 * to cope with that. 1/4 retries is rather arbitrary but we
3333 * would need much more detailed feedback from compaction to
3334 * make a better decision.
3336 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3338 if (*compaction_retries
<= max_retries
) {
3344 * Make sure there are attempts at the highest priority if we exhausted
3345 * all retries or failed at the lower priorities.
3348 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3349 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3351 if (*compact_priority
> min_priority
) {
3352 (*compact_priority
)--;
3353 *compaction_retries
= 0;
3357 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3361 static inline struct page
*
3362 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3363 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3364 enum compact_priority prio
, enum compact_result
*compact_result
)
3366 *compact_result
= COMPACT_SKIPPED
;
3371 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3372 enum compact_result compact_result
,
3373 enum compact_priority
*compact_priority
,
3374 int *compaction_retries
)
3379 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3383 * There are setups with compaction disabled which would prefer to loop
3384 * inside the allocator rather than hit the oom killer prematurely.
3385 * Let's give them a good hope and keep retrying while the order-0
3386 * watermarks are OK.
3388 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3390 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3391 ac_classzone_idx(ac
), alloc_flags
))
3396 #endif /* CONFIG_COMPACTION */
3398 /* Perform direct synchronous page reclaim */
3400 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3401 const struct alloc_context
*ac
)
3403 struct reclaim_state reclaim_state
;
3408 /* We now go into synchronous reclaim */
3409 cpuset_memory_pressure_bump();
3410 current
->flags
|= PF_MEMALLOC
;
3411 lockdep_set_current_reclaim_state(gfp_mask
);
3412 reclaim_state
.reclaimed_slab
= 0;
3413 current
->reclaim_state
= &reclaim_state
;
3415 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3418 current
->reclaim_state
= NULL
;
3419 lockdep_clear_current_reclaim_state();
3420 current
->flags
&= ~PF_MEMALLOC
;
3427 /* The really slow allocator path where we enter direct reclaim */
3428 static inline struct page
*
3429 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3430 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3431 unsigned long *did_some_progress
)
3433 struct page
*page
= NULL
;
3434 bool drained
= false;
3436 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3437 if (unlikely(!(*did_some_progress
)))
3441 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3444 * If an allocation failed after direct reclaim, it could be because
3445 * pages are pinned on the per-cpu lists or in high alloc reserves.
3446 * Shrink them them and try again
3448 if (!page
&& !drained
) {
3449 unreserve_highatomic_pageblock(ac
, false);
3450 drain_all_pages(NULL
);
3458 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3462 pg_data_t
*last_pgdat
= NULL
;
3464 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3465 ac
->high_zoneidx
, ac
->nodemask
) {
3466 if (last_pgdat
!= zone
->zone_pgdat
)
3467 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3468 last_pgdat
= zone
->zone_pgdat
;
3472 static inline unsigned int
3473 gfp_to_alloc_flags(gfp_t gfp_mask
)
3475 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3477 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3478 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3481 * The caller may dip into page reserves a bit more if the caller
3482 * cannot run direct reclaim, or if the caller has realtime scheduling
3483 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3484 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3486 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3488 if (gfp_mask
& __GFP_ATOMIC
) {
3490 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3491 * if it can't schedule.
3493 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3494 alloc_flags
|= ALLOC_HARDER
;
3496 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3497 * comment for __cpuset_node_allowed().
3499 alloc_flags
&= ~ALLOC_CPUSET
;
3500 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3501 alloc_flags
|= ALLOC_HARDER
;
3504 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3505 alloc_flags
|= ALLOC_CMA
;
3510 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3512 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3515 if (gfp_mask
& __GFP_MEMALLOC
)
3517 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3519 if (!in_interrupt() &&
3520 ((current
->flags
& PF_MEMALLOC
) ||
3521 unlikely(test_thread_flag(TIF_MEMDIE
))))
3528 * Maximum number of reclaim retries without any progress before OOM killer
3529 * is consider as the only way to move forward.
3531 #define MAX_RECLAIM_RETRIES 16
3534 * Checks whether it makes sense to retry the reclaim to make a forward progress
3535 * for the given allocation request.
3536 * The reclaim feedback represented by did_some_progress (any progress during
3537 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3538 * any progress in a row) is considered as well as the reclaimable pages on the
3539 * applicable zone list (with a backoff mechanism which is a function of
3540 * no_progress_loops).
3542 * Returns true if a retry is viable or false to enter the oom path.
3545 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3546 struct alloc_context
*ac
, int alloc_flags
,
3547 bool did_some_progress
, int *no_progress_loops
)
3553 * Costly allocations might have made a progress but this doesn't mean
3554 * their order will become available due to high fragmentation so
3555 * always increment the no progress counter for them
3557 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3558 *no_progress_loops
= 0;
3560 (*no_progress_loops
)++;
3563 * Make sure we converge to OOM if we cannot make any progress
3564 * several times in the row.
3566 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3567 /* Before OOM, exhaust highatomic_reserve */
3568 return unreserve_highatomic_pageblock(ac
, true);
3572 * Keep reclaiming pages while there is a chance this will lead
3573 * somewhere. If none of the target zones can satisfy our allocation
3574 * request even if all reclaimable pages are considered then we are
3575 * screwed and have to go OOM.
3577 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3579 unsigned long available
;
3580 unsigned long reclaimable
;
3581 unsigned long min_wmark
= min_wmark_pages(zone
);
3584 available
= reclaimable
= zone_reclaimable_pages(zone
);
3585 available
-= DIV_ROUND_UP((*no_progress_loops
) * available
,
3586 MAX_RECLAIM_RETRIES
);
3587 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3590 * Would the allocation succeed if we reclaimed the whole
3593 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3594 ac_classzone_idx(ac
), alloc_flags
, available
);
3595 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3596 available
, min_wmark
, *no_progress_loops
, wmark
);
3599 * If we didn't make any progress and have a lot of
3600 * dirty + writeback pages then we should wait for
3601 * an IO to complete to slow down the reclaim and
3602 * prevent from pre mature OOM
3604 if (!did_some_progress
) {
3605 unsigned long write_pending
;
3607 write_pending
= zone_page_state_snapshot(zone
,
3608 NR_ZONE_WRITE_PENDING
);
3610 if (2 * write_pending
> reclaimable
) {
3611 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3617 * Memory allocation/reclaim might be called from a WQ
3618 * context and the current implementation of the WQ
3619 * concurrency control doesn't recognize that
3620 * a particular WQ is congested if the worker thread is
3621 * looping without ever sleeping. Therefore we have to
3622 * do a short sleep here rather than calling
3625 if (current
->flags
& PF_WQ_WORKER
)
3626 schedule_timeout_uninterruptible(1);
3637 static inline struct page
*
3638 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3639 struct alloc_context
*ac
)
3641 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3642 struct page
*page
= NULL
;
3643 unsigned int alloc_flags
;
3644 unsigned long did_some_progress
;
3645 enum compact_priority compact_priority
;
3646 enum compact_result compact_result
;
3647 int compaction_retries
;
3648 int no_progress_loops
;
3649 unsigned long alloc_start
= jiffies
;
3650 unsigned int stall_timeout
= 10 * HZ
;
3651 unsigned int cpuset_mems_cookie
;
3654 * In the slowpath, we sanity check order to avoid ever trying to
3655 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3656 * be using allocators in order of preference for an area that is
3659 if (order
>= MAX_ORDER
) {
3660 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3665 * We also sanity check to catch abuse of atomic reserves being used by
3666 * callers that are not in atomic context.
3668 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3669 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3670 gfp_mask
&= ~__GFP_ATOMIC
;
3673 compaction_retries
= 0;
3674 no_progress_loops
= 0;
3675 compact_priority
= DEF_COMPACT_PRIORITY
;
3676 cpuset_mems_cookie
= read_mems_allowed_begin();
3679 * The fast path uses conservative alloc_flags to succeed only until
3680 * kswapd needs to be woken up, and to avoid the cost of setting up
3681 * alloc_flags precisely. So we do that now.
3683 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3686 * We need to recalculate the starting point for the zonelist iterator
3687 * because we might have used different nodemask in the fast path, or
3688 * there was a cpuset modification and we are retrying - otherwise we
3689 * could end up iterating over non-eligible zones endlessly.
3691 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3692 ac
->high_zoneidx
, ac
->nodemask
);
3693 if (!ac
->preferred_zoneref
->zone
)
3696 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3697 wake_all_kswapds(order
, ac
);
3700 * The adjusted alloc_flags might result in immediate success, so try
3703 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3708 * For costly allocations, try direct compaction first, as it's likely
3709 * that we have enough base pages and don't need to reclaim. Don't try
3710 * that for allocations that are allowed to ignore watermarks, as the
3711 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3713 if (can_direct_reclaim
&& order
> PAGE_ALLOC_COSTLY_ORDER
&&
3714 !gfp_pfmemalloc_allowed(gfp_mask
)) {
3715 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3717 INIT_COMPACT_PRIORITY
,
3723 * Checks for costly allocations with __GFP_NORETRY, which
3724 * includes THP page fault allocations
3726 if (gfp_mask
& __GFP_NORETRY
) {
3728 * If compaction is deferred for high-order allocations,
3729 * it is because sync compaction recently failed. If
3730 * this is the case and the caller requested a THP
3731 * allocation, we do not want to heavily disrupt the
3732 * system, so we fail the allocation instead of entering
3735 if (compact_result
== COMPACT_DEFERRED
)
3739 * Looks like reclaim/compaction is worth trying, but
3740 * sync compaction could be very expensive, so keep
3741 * using async compaction.
3743 compact_priority
= INIT_COMPACT_PRIORITY
;
3748 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3749 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3750 wake_all_kswapds(order
, ac
);
3752 if (gfp_pfmemalloc_allowed(gfp_mask
))
3753 alloc_flags
= ALLOC_NO_WATERMARKS
;
3756 * Reset the zonelist iterators if memory policies can be ignored.
3757 * These allocations are high priority and system rather than user
3760 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3761 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3762 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3763 ac
->high_zoneidx
, ac
->nodemask
);
3766 /* Attempt with potentially adjusted zonelist and alloc_flags */
3767 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3771 /* Caller is not willing to reclaim, we can't balance anything */
3772 if (!can_direct_reclaim
)
3775 /* Make sure we know about allocations which stall for too long */
3776 if (time_after(jiffies
, alloc_start
+ stall_timeout
)) {
3777 warn_alloc(gfp_mask
, ac
->nodemask
,
3778 "page allocation stalls for %ums, order:%u",
3779 jiffies_to_msecs(jiffies
-alloc_start
), order
);
3780 stall_timeout
+= 10 * HZ
;
3783 /* Avoid recursion of direct reclaim */
3784 if (current
->flags
& PF_MEMALLOC
)
3787 /* Try direct reclaim and then allocating */
3788 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3789 &did_some_progress
);
3793 /* Try direct compaction and then allocating */
3794 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3795 compact_priority
, &compact_result
);
3799 /* Do not loop if specifically requested */
3800 if (gfp_mask
& __GFP_NORETRY
)
3804 * Do not retry costly high order allocations unless they are
3807 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_REPEAT
))
3810 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3811 did_some_progress
> 0, &no_progress_loops
))
3815 * It doesn't make any sense to retry for the compaction if the order-0
3816 * reclaim is not able to make any progress because the current
3817 * implementation of the compaction depends on the sufficient amount
3818 * of free memory (see __compaction_suitable)
3820 if (did_some_progress
> 0 &&
3821 should_compact_retry(ac
, order
, alloc_flags
,
3822 compact_result
, &compact_priority
,
3823 &compaction_retries
))
3827 * It's possible we raced with cpuset update so the OOM would be
3828 * premature (see below the nopage: label for full explanation).
3830 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3833 /* Reclaim has failed us, start killing things */
3834 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3838 /* Avoid allocations with no watermarks from looping endlessly */
3839 if (test_thread_flag(TIF_MEMDIE
))
3842 /* Retry as long as the OOM killer is making progress */
3843 if (did_some_progress
) {
3844 no_progress_loops
= 0;
3850 * When updating a task's mems_allowed or mempolicy nodemask, it is
3851 * possible to race with parallel threads in such a way that our
3852 * allocation can fail while the mask is being updated. If we are about
3853 * to fail, check if the cpuset changed during allocation and if so,
3856 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3860 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3863 if (gfp_mask
& __GFP_NOFAIL
) {
3865 * All existing users of the __GFP_NOFAIL are blockable, so warn
3866 * of any new users that actually require GFP_NOWAIT
3868 if (WARN_ON_ONCE(!can_direct_reclaim
))
3872 * PF_MEMALLOC request from this context is rather bizarre
3873 * because we cannot reclaim anything and only can loop waiting
3874 * for somebody to do a work for us
3876 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
3879 * non failing costly orders are a hard requirement which we
3880 * are not prepared for much so let's warn about these users
3881 * so that we can identify them and convert them to something
3884 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
3887 * Help non-failing allocations by giving them access to memory
3888 * reserves but do not use ALLOC_NO_WATERMARKS because this
3889 * could deplete whole memory reserves which would just make
3890 * the situation worse
3892 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
3900 warn_alloc(gfp_mask
, ac
->nodemask
,
3901 "page allocation failure: order:%u", order
);
3906 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
3907 struct zonelist
*zonelist
, nodemask_t
*nodemask
,
3908 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
3909 unsigned int *alloc_flags
)
3911 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
3912 ac
->zonelist
= zonelist
;
3913 ac
->nodemask
= nodemask
;
3914 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
3916 if (cpusets_enabled()) {
3917 *alloc_mask
|= __GFP_HARDWALL
;
3919 ac
->nodemask
= &cpuset_current_mems_allowed
;
3921 *alloc_flags
|= ALLOC_CPUSET
;
3924 lockdep_trace_alloc(gfp_mask
);
3926 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3928 if (should_fail_alloc_page(gfp_mask
, order
))
3931 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
3932 *alloc_flags
|= ALLOC_CMA
;
3937 /* Determine whether to spread dirty pages and what the first usable zone */
3938 static inline void finalise_ac(gfp_t gfp_mask
,
3939 unsigned int order
, struct alloc_context
*ac
)
3941 /* Dirty zone balancing only done in the fast path */
3942 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3945 * The preferred zone is used for statistics but crucially it is
3946 * also used as the starting point for the zonelist iterator. It
3947 * may get reset for allocations that ignore memory policies.
3949 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3950 ac
->high_zoneidx
, ac
->nodemask
);
3954 * This is the 'heart' of the zoned buddy allocator.
3957 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3958 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3961 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
3962 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3963 struct alloc_context ac
= { };
3965 gfp_mask
&= gfp_allowed_mask
;
3966 if (!prepare_alloc_pages(gfp_mask
, order
, zonelist
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
3969 finalise_ac(gfp_mask
, order
, &ac
);
3971 /* First allocation attempt */
3972 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3977 * Runtime PM, block IO and its error handling path can deadlock
3978 * because I/O on the device might not complete.
3980 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3981 ac
.spread_dirty_pages
= false;
3984 * Restore the original nodemask if it was potentially replaced with
3985 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3987 if (unlikely(ac
.nodemask
!= nodemask
))
3988 ac
.nodemask
= nodemask
;
3990 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3993 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
3994 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
3995 __free_pages(page
, order
);
3999 if (kmemcheck_enabled
&& page
)
4000 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
4002 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4006 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4009 * Common helper functions.
4011 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4016 * __get_free_pages() returns a 32-bit address, which cannot represent
4019 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
4021 page
= alloc_pages(gfp_mask
, order
);
4024 return (unsigned long) page_address(page
);
4026 EXPORT_SYMBOL(__get_free_pages
);
4028 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4030 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4032 EXPORT_SYMBOL(get_zeroed_page
);
4034 void __free_pages(struct page
*page
, unsigned int order
)
4036 if (put_page_testzero(page
)) {
4038 free_hot_cold_page(page
, false);
4040 __free_pages_ok(page
, order
);
4044 EXPORT_SYMBOL(__free_pages
);
4046 void free_pages(unsigned long addr
, unsigned int order
)
4049 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4050 __free_pages(virt_to_page((void *)addr
), order
);
4054 EXPORT_SYMBOL(free_pages
);
4058 * An arbitrary-length arbitrary-offset area of memory which resides
4059 * within a 0 or higher order page. Multiple fragments within that page
4060 * are individually refcounted, in the page's reference counter.
4062 * The page_frag functions below provide a simple allocation framework for
4063 * page fragments. This is used by the network stack and network device
4064 * drivers to provide a backing region of memory for use as either an
4065 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4067 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4070 struct page
*page
= NULL
;
4071 gfp_t gfp
= gfp_mask
;
4073 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4074 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4076 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4077 PAGE_FRAG_CACHE_MAX_ORDER
);
4078 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4080 if (unlikely(!page
))
4081 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4083 nc
->va
= page
? page_address(page
) : NULL
;
4088 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4090 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4092 if (page_ref_sub_and_test(page
, count
)) {
4093 unsigned int order
= compound_order(page
);
4096 free_hot_cold_page(page
, false);
4098 __free_pages_ok(page
, order
);
4101 EXPORT_SYMBOL(__page_frag_cache_drain
);
4103 void *page_frag_alloc(struct page_frag_cache
*nc
,
4104 unsigned int fragsz
, gfp_t gfp_mask
)
4106 unsigned int size
= PAGE_SIZE
;
4110 if (unlikely(!nc
->va
)) {
4112 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4116 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4117 /* if size can vary use size else just use PAGE_SIZE */
4120 /* Even if we own the page, we do not use atomic_set().
4121 * This would break get_page_unless_zero() users.
4123 page_ref_add(page
, size
- 1);
4125 /* reset page count bias and offset to start of new frag */
4126 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4127 nc
->pagecnt_bias
= size
;
4131 offset
= nc
->offset
- fragsz
;
4132 if (unlikely(offset
< 0)) {
4133 page
= virt_to_page(nc
->va
);
4135 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4138 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4139 /* if size can vary use size else just use PAGE_SIZE */
4142 /* OK, page count is 0, we can safely set it */
4143 set_page_count(page
, size
);
4145 /* reset page count bias and offset to start of new frag */
4146 nc
->pagecnt_bias
= size
;
4147 offset
= size
- fragsz
;
4151 nc
->offset
= offset
;
4153 return nc
->va
+ offset
;
4155 EXPORT_SYMBOL(page_frag_alloc
);
4158 * Frees a page fragment allocated out of either a compound or order 0 page.
4160 void page_frag_free(void *addr
)
4162 struct page
*page
= virt_to_head_page(addr
);
4164 if (unlikely(put_page_testzero(page
)))
4165 __free_pages_ok(page
, compound_order(page
));
4167 EXPORT_SYMBOL(page_frag_free
);
4169 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4173 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4174 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4176 split_page(virt_to_page((void *)addr
), order
);
4177 while (used
< alloc_end
) {
4182 return (void *)addr
;
4186 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4187 * @size: the number of bytes to allocate
4188 * @gfp_mask: GFP flags for the allocation
4190 * This function is similar to alloc_pages(), except that it allocates the
4191 * minimum number of pages to satisfy the request. alloc_pages() can only
4192 * allocate memory in power-of-two pages.
4194 * This function is also limited by MAX_ORDER.
4196 * Memory allocated by this function must be released by free_pages_exact().
4198 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4200 unsigned int order
= get_order(size
);
4203 addr
= __get_free_pages(gfp_mask
, order
);
4204 return make_alloc_exact(addr
, order
, size
);
4206 EXPORT_SYMBOL(alloc_pages_exact
);
4209 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4211 * @nid: the preferred node ID where memory should be allocated
4212 * @size: the number of bytes to allocate
4213 * @gfp_mask: GFP flags for the allocation
4215 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4218 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4220 unsigned int order
= get_order(size
);
4221 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4224 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4228 * free_pages_exact - release memory allocated via alloc_pages_exact()
4229 * @virt: the value returned by alloc_pages_exact.
4230 * @size: size of allocation, same value as passed to alloc_pages_exact().
4232 * Release the memory allocated by a previous call to alloc_pages_exact.
4234 void free_pages_exact(void *virt
, size_t size
)
4236 unsigned long addr
= (unsigned long)virt
;
4237 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4239 while (addr
< end
) {
4244 EXPORT_SYMBOL(free_pages_exact
);
4247 * nr_free_zone_pages - count number of pages beyond high watermark
4248 * @offset: The zone index of the highest zone
4250 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4251 * high watermark within all zones at or below a given zone index. For each
4252 * zone, the number of pages is calculated as:
4253 * managed_pages - high_pages
4255 static unsigned long nr_free_zone_pages(int offset
)
4260 /* Just pick one node, since fallback list is circular */
4261 unsigned long sum
= 0;
4263 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4265 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4266 unsigned long size
= zone
->managed_pages
;
4267 unsigned long high
= high_wmark_pages(zone
);
4276 * nr_free_buffer_pages - count number of pages beyond high watermark
4278 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4279 * watermark within ZONE_DMA and ZONE_NORMAL.
4281 unsigned long nr_free_buffer_pages(void)
4283 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4285 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4288 * nr_free_pagecache_pages - count number of pages beyond high watermark
4290 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4291 * high watermark within all zones.
4293 unsigned long nr_free_pagecache_pages(void)
4295 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4298 static inline void show_node(struct zone
*zone
)
4300 if (IS_ENABLED(CONFIG_NUMA
))
4301 printk("Node %d ", zone_to_nid(zone
));
4304 long si_mem_available(void)
4307 unsigned long pagecache
;
4308 unsigned long wmark_low
= 0;
4309 unsigned long pages
[NR_LRU_LISTS
];
4313 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4314 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4317 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4320 * Estimate the amount of memory available for userspace allocations,
4321 * without causing swapping.
4323 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4326 * Not all the page cache can be freed, otherwise the system will
4327 * start swapping. Assume at least half of the page cache, or the
4328 * low watermark worth of cache, needs to stay.
4330 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4331 pagecache
-= min(pagecache
/ 2, wmark_low
);
4332 available
+= pagecache
;
4335 * Part of the reclaimable slab consists of items that are in use,
4336 * and cannot be freed. Cap this estimate at the low watermark.
4338 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4339 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4345 EXPORT_SYMBOL_GPL(si_mem_available
);
4347 void si_meminfo(struct sysinfo
*val
)
4349 val
->totalram
= totalram_pages
;
4350 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4351 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4352 val
->bufferram
= nr_blockdev_pages();
4353 val
->totalhigh
= totalhigh_pages
;
4354 val
->freehigh
= nr_free_highpages();
4355 val
->mem_unit
= PAGE_SIZE
;
4358 EXPORT_SYMBOL(si_meminfo
);
4361 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4363 int zone_type
; /* needs to be signed */
4364 unsigned long managed_pages
= 0;
4365 unsigned long managed_highpages
= 0;
4366 unsigned long free_highpages
= 0;
4367 pg_data_t
*pgdat
= NODE_DATA(nid
);
4369 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4370 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4371 val
->totalram
= managed_pages
;
4372 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4373 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4374 #ifdef CONFIG_HIGHMEM
4375 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4376 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4378 if (is_highmem(zone
)) {
4379 managed_highpages
+= zone
->managed_pages
;
4380 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4383 val
->totalhigh
= managed_highpages
;
4384 val
->freehigh
= free_highpages
;
4386 val
->totalhigh
= managed_highpages
;
4387 val
->freehigh
= free_highpages
;
4389 val
->mem_unit
= PAGE_SIZE
;
4394 * Determine whether the node should be displayed or not, depending on whether
4395 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4397 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4399 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4403 * no node mask - aka implicit memory numa policy. Do not bother with
4404 * the synchronization - read_mems_allowed_begin - because we do not
4405 * have to be precise here.
4408 nodemask
= &cpuset_current_mems_allowed
;
4410 return !node_isset(nid
, *nodemask
);
4413 #define K(x) ((x) << (PAGE_SHIFT-10))
4415 static void show_migration_types(unsigned char type
)
4417 static const char types
[MIGRATE_TYPES
] = {
4418 [MIGRATE_UNMOVABLE
] = 'U',
4419 [MIGRATE_MOVABLE
] = 'M',
4420 [MIGRATE_RECLAIMABLE
] = 'E',
4421 [MIGRATE_HIGHATOMIC
] = 'H',
4423 [MIGRATE_CMA
] = 'C',
4425 #ifdef CONFIG_MEMORY_ISOLATION
4426 [MIGRATE_ISOLATE
] = 'I',
4429 char tmp
[MIGRATE_TYPES
+ 1];
4433 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4434 if (type
& (1 << i
))
4439 printk(KERN_CONT
"(%s) ", tmp
);
4443 * Show free area list (used inside shift_scroll-lock stuff)
4444 * We also calculate the percentage fragmentation. We do this by counting the
4445 * memory on each free list with the exception of the first item on the list.
4448 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4451 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4453 unsigned long free_pcp
= 0;
4458 for_each_populated_zone(zone
) {
4459 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4462 for_each_online_cpu(cpu
)
4463 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4466 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4467 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4468 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4469 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4470 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4471 " free:%lu free_pcp:%lu free_cma:%lu\n",
4472 global_node_page_state(NR_ACTIVE_ANON
),
4473 global_node_page_state(NR_INACTIVE_ANON
),
4474 global_node_page_state(NR_ISOLATED_ANON
),
4475 global_node_page_state(NR_ACTIVE_FILE
),
4476 global_node_page_state(NR_INACTIVE_FILE
),
4477 global_node_page_state(NR_ISOLATED_FILE
),
4478 global_node_page_state(NR_UNEVICTABLE
),
4479 global_node_page_state(NR_FILE_DIRTY
),
4480 global_node_page_state(NR_WRITEBACK
),
4481 global_node_page_state(NR_UNSTABLE_NFS
),
4482 global_page_state(NR_SLAB_RECLAIMABLE
),
4483 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4484 global_node_page_state(NR_FILE_MAPPED
),
4485 global_node_page_state(NR_SHMEM
),
4486 global_page_state(NR_PAGETABLE
),
4487 global_page_state(NR_BOUNCE
),
4488 global_page_state(NR_FREE_PAGES
),
4490 global_page_state(NR_FREE_CMA_PAGES
));
4492 for_each_online_pgdat(pgdat
) {
4493 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
4497 " active_anon:%lukB"
4498 " inactive_anon:%lukB"
4499 " active_file:%lukB"
4500 " inactive_file:%lukB"
4501 " unevictable:%lukB"
4502 " isolated(anon):%lukB"
4503 " isolated(file):%lukB"
4508 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4510 " shmem_pmdmapped: %lukB"
4513 " writeback_tmp:%lukB"
4515 " pages_scanned:%lu"
4516 " all_unreclaimable? %s"
4519 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4520 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4521 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4522 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4523 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4524 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4525 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4526 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4527 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4528 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4529 K(node_page_state(pgdat
, NR_SHMEM
)),
4530 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4531 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4532 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4534 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4536 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4537 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4538 node_page_state(pgdat
, NR_PAGES_SCANNED
),
4539 !pgdat_reclaimable(pgdat
) ? "yes" : "no");
4542 for_each_populated_zone(zone
) {
4545 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4549 for_each_online_cpu(cpu
)
4550 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4559 " active_anon:%lukB"
4560 " inactive_anon:%lukB"
4561 " active_file:%lukB"
4562 " inactive_file:%lukB"
4563 " unevictable:%lukB"
4564 " writepending:%lukB"
4568 " slab_reclaimable:%lukB"
4569 " slab_unreclaimable:%lukB"
4570 " kernel_stack:%lukB"
4578 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4579 K(min_wmark_pages(zone
)),
4580 K(low_wmark_pages(zone
)),
4581 K(high_wmark_pages(zone
)),
4582 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4583 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4584 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4585 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4586 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4587 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4588 K(zone
->present_pages
),
4589 K(zone
->managed_pages
),
4590 K(zone_page_state(zone
, NR_MLOCK
)),
4591 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4592 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4593 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4594 K(zone_page_state(zone
, NR_PAGETABLE
)),
4595 K(zone_page_state(zone
, NR_BOUNCE
)),
4597 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4598 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4599 printk("lowmem_reserve[]:");
4600 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4601 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
4602 printk(KERN_CONT
"\n");
4605 for_each_populated_zone(zone
) {
4607 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4608 unsigned char types
[MAX_ORDER
];
4610 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4613 printk(KERN_CONT
"%s: ", zone
->name
);
4615 spin_lock_irqsave(&zone
->lock
, flags
);
4616 for (order
= 0; order
< MAX_ORDER
; order
++) {
4617 struct free_area
*area
= &zone
->free_area
[order
];
4620 nr
[order
] = area
->nr_free
;
4621 total
+= nr
[order
] << order
;
4624 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4625 if (!list_empty(&area
->free_list
[type
]))
4626 types
[order
] |= 1 << type
;
4629 spin_unlock_irqrestore(&zone
->lock
, flags
);
4630 for (order
= 0; order
< MAX_ORDER
; order
++) {
4631 printk(KERN_CONT
"%lu*%lukB ",
4632 nr
[order
], K(1UL) << order
);
4634 show_migration_types(types
[order
]);
4636 printk(KERN_CONT
"= %lukB\n", K(total
));
4639 hugetlb_show_meminfo();
4641 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4643 show_swap_cache_info();
4646 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4648 zoneref
->zone
= zone
;
4649 zoneref
->zone_idx
= zone_idx(zone
);
4653 * Builds allocation fallback zone lists.
4655 * Add all populated zones of a node to the zonelist.
4657 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4661 enum zone_type zone_type
= MAX_NR_ZONES
;
4665 zone
= pgdat
->node_zones
+ zone_type
;
4666 if (managed_zone(zone
)) {
4667 zoneref_set_zone(zone
,
4668 &zonelist
->_zonerefs
[nr_zones
++]);
4669 check_highest_zone(zone_type
);
4671 } while (zone_type
);
4679 * 0 = automatic detection of better ordering.
4680 * 1 = order by ([node] distance, -zonetype)
4681 * 2 = order by (-zonetype, [node] distance)
4683 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4684 * the same zonelist. So only NUMA can configure this param.
4686 #define ZONELIST_ORDER_DEFAULT 0
4687 #define ZONELIST_ORDER_NODE 1
4688 #define ZONELIST_ORDER_ZONE 2
4690 /* zonelist order in the kernel.
4691 * set_zonelist_order() will set this to NODE or ZONE.
4693 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4694 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4698 /* The value user specified ....changed by config */
4699 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4700 /* string for sysctl */
4701 #define NUMA_ZONELIST_ORDER_LEN 16
4702 char numa_zonelist_order
[16] = "default";
4705 * interface for configure zonelist ordering.
4706 * command line option "numa_zonelist_order"
4707 * = "[dD]efault - default, automatic configuration.
4708 * = "[nN]ode - order by node locality, then by zone within node
4709 * = "[zZ]one - order by zone, then by locality within zone
4712 static int __parse_numa_zonelist_order(char *s
)
4714 if (*s
== 'd' || *s
== 'D') {
4715 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4716 } else if (*s
== 'n' || *s
== 'N') {
4717 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4718 } else if (*s
== 'z' || *s
== 'Z') {
4719 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4721 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4727 static __init
int setup_numa_zonelist_order(char *s
)
4734 ret
= __parse_numa_zonelist_order(s
);
4736 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4740 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4743 * sysctl handler for numa_zonelist_order
4745 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4746 void __user
*buffer
, size_t *length
,
4749 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4751 static DEFINE_MUTEX(zl_order_mutex
);
4753 mutex_lock(&zl_order_mutex
);
4755 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4759 strcpy(saved_string
, (char *)table
->data
);
4761 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4765 int oldval
= user_zonelist_order
;
4767 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4770 * bogus value. restore saved string
4772 strncpy((char *)table
->data
, saved_string
,
4773 NUMA_ZONELIST_ORDER_LEN
);
4774 user_zonelist_order
= oldval
;
4775 } else if (oldval
!= user_zonelist_order
) {
4776 mutex_lock(&zonelists_mutex
);
4777 build_all_zonelists(NULL
, NULL
);
4778 mutex_unlock(&zonelists_mutex
);
4782 mutex_unlock(&zl_order_mutex
);
4787 #define MAX_NODE_LOAD (nr_online_nodes)
4788 static int node_load
[MAX_NUMNODES
];
4791 * find_next_best_node - find the next node that should appear in a given node's fallback list
4792 * @node: node whose fallback list we're appending
4793 * @used_node_mask: nodemask_t of already used nodes
4795 * We use a number of factors to determine which is the next node that should
4796 * appear on a given node's fallback list. The node should not have appeared
4797 * already in @node's fallback list, and it should be the next closest node
4798 * according to the distance array (which contains arbitrary distance values
4799 * from each node to each node in the system), and should also prefer nodes
4800 * with no CPUs, since presumably they'll have very little allocation pressure
4801 * on them otherwise.
4802 * It returns -1 if no node is found.
4804 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4807 int min_val
= INT_MAX
;
4808 int best_node
= NUMA_NO_NODE
;
4809 const struct cpumask
*tmp
= cpumask_of_node(0);
4811 /* Use the local node if we haven't already */
4812 if (!node_isset(node
, *used_node_mask
)) {
4813 node_set(node
, *used_node_mask
);
4817 for_each_node_state(n
, N_MEMORY
) {
4819 /* Don't want a node to appear more than once */
4820 if (node_isset(n
, *used_node_mask
))
4823 /* Use the distance array to find the distance */
4824 val
= node_distance(node
, n
);
4826 /* Penalize nodes under us ("prefer the next node") */
4829 /* Give preference to headless and unused nodes */
4830 tmp
= cpumask_of_node(n
);
4831 if (!cpumask_empty(tmp
))
4832 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4834 /* Slight preference for less loaded node */
4835 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4836 val
+= node_load
[n
];
4838 if (val
< min_val
) {
4845 node_set(best_node
, *used_node_mask
);
4852 * Build zonelists ordered by node and zones within node.
4853 * This results in maximum locality--normal zone overflows into local
4854 * DMA zone, if any--but risks exhausting DMA zone.
4856 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4859 struct zonelist
*zonelist
;
4861 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4862 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4864 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4865 zonelist
->_zonerefs
[j
].zone
= NULL
;
4866 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4870 * Build gfp_thisnode zonelists
4872 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4875 struct zonelist
*zonelist
;
4877 zonelist
= &pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
];
4878 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4879 zonelist
->_zonerefs
[j
].zone
= NULL
;
4880 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4884 * Build zonelists ordered by zone and nodes within zones.
4885 * This results in conserving DMA zone[s] until all Normal memory is
4886 * exhausted, but results in overflowing to remote node while memory
4887 * may still exist in local DMA zone.
4889 static int node_order
[MAX_NUMNODES
];
4891 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4894 int zone_type
; /* needs to be signed */
4896 struct zonelist
*zonelist
;
4898 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4900 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4901 for (j
= 0; j
< nr_nodes
; j
++) {
4902 node
= node_order
[j
];
4903 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4904 if (managed_zone(z
)) {
4906 &zonelist
->_zonerefs
[pos
++]);
4907 check_highest_zone(zone_type
);
4911 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4912 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4915 #if defined(CONFIG_64BIT)
4917 * Devices that require DMA32/DMA are relatively rare and do not justify a
4918 * penalty to every machine in case the specialised case applies. Default
4919 * to Node-ordering on 64-bit NUMA machines
4921 static int default_zonelist_order(void)
4923 return ZONELIST_ORDER_NODE
;
4927 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4928 * by the kernel. If processes running on node 0 deplete the low memory zone
4929 * then reclaim will occur more frequency increasing stalls and potentially
4930 * be easier to OOM if a large percentage of the zone is under writeback or
4931 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4932 * Hence, default to zone ordering on 32-bit.
4934 static int default_zonelist_order(void)
4936 return ZONELIST_ORDER_ZONE
;
4938 #endif /* CONFIG_64BIT */
4940 static void set_zonelist_order(void)
4942 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4943 current_zonelist_order
= default_zonelist_order();
4945 current_zonelist_order
= user_zonelist_order
;
4948 static void build_zonelists(pg_data_t
*pgdat
)
4951 nodemask_t used_mask
;
4952 int local_node
, prev_node
;
4953 struct zonelist
*zonelist
;
4954 unsigned int order
= current_zonelist_order
;
4956 /* initialize zonelists */
4957 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4958 zonelist
= pgdat
->node_zonelists
+ i
;
4959 zonelist
->_zonerefs
[0].zone
= NULL
;
4960 zonelist
->_zonerefs
[0].zone_idx
= 0;
4963 /* NUMA-aware ordering of nodes */
4964 local_node
= pgdat
->node_id
;
4965 load
= nr_online_nodes
;
4966 prev_node
= local_node
;
4967 nodes_clear(used_mask
);
4969 memset(node_order
, 0, sizeof(node_order
));
4972 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4974 * We don't want to pressure a particular node.
4975 * So adding penalty to the first node in same
4976 * distance group to make it round-robin.
4978 if (node_distance(local_node
, node
) !=
4979 node_distance(local_node
, prev_node
))
4980 node_load
[node
] = load
;
4984 if (order
== ZONELIST_ORDER_NODE
)
4985 build_zonelists_in_node_order(pgdat
, node
);
4987 node_order
[i
++] = node
; /* remember order */
4990 if (order
== ZONELIST_ORDER_ZONE
) {
4991 /* calculate node order -- i.e., DMA last! */
4992 build_zonelists_in_zone_order(pgdat
, i
);
4995 build_thisnode_zonelists(pgdat
);
4998 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5000 * Return node id of node used for "local" allocations.
5001 * I.e., first node id of first zone in arg node's generic zonelist.
5002 * Used for initializing percpu 'numa_mem', which is used primarily
5003 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5005 int local_memory_node(int node
)
5009 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5010 gfp_zone(GFP_KERNEL
),
5012 return z
->zone
->node
;
5016 static void setup_min_unmapped_ratio(void);
5017 static void setup_min_slab_ratio(void);
5018 #else /* CONFIG_NUMA */
5020 static void set_zonelist_order(void)
5022 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
5025 static void build_zonelists(pg_data_t
*pgdat
)
5027 int node
, local_node
;
5029 struct zonelist
*zonelist
;
5031 local_node
= pgdat
->node_id
;
5033 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
5034 j
= build_zonelists_node(pgdat
, zonelist
, 0);
5037 * Now we build the zonelist so that it contains the zones
5038 * of all the other nodes.
5039 * We don't want to pressure a particular node, so when
5040 * building the zones for node N, we make sure that the
5041 * zones coming right after the local ones are those from
5042 * node N+1 (modulo N)
5044 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5045 if (!node_online(node
))
5047 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
5049 for (node
= 0; node
< local_node
; node
++) {
5050 if (!node_online(node
))
5052 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
5055 zonelist
->_zonerefs
[j
].zone
= NULL
;
5056 zonelist
->_zonerefs
[j
].zone_idx
= 0;
5059 #endif /* CONFIG_NUMA */
5062 * Boot pageset table. One per cpu which is going to be used for all
5063 * zones and all nodes. The parameters will be set in such a way
5064 * that an item put on a list will immediately be handed over to
5065 * the buddy list. This is safe since pageset manipulation is done
5066 * with interrupts disabled.
5068 * The boot_pagesets must be kept even after bootup is complete for
5069 * unused processors and/or zones. They do play a role for bootstrapping
5070 * hotplugged processors.
5072 * zoneinfo_show() and maybe other functions do
5073 * not check if the processor is online before following the pageset pointer.
5074 * Other parts of the kernel may not check if the zone is available.
5076 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5077 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5078 static void setup_zone_pageset(struct zone
*zone
);
5081 * Global mutex to protect against size modification of zonelists
5082 * as well as to serialize pageset setup for the new populated zone.
5084 DEFINE_MUTEX(zonelists_mutex
);
5086 /* return values int ....just for stop_machine() */
5087 static int __build_all_zonelists(void *data
)
5091 pg_data_t
*self
= data
;
5094 memset(node_load
, 0, sizeof(node_load
));
5097 if (self
&& !node_online(self
->node_id
)) {
5098 build_zonelists(self
);
5101 for_each_online_node(nid
) {
5102 pg_data_t
*pgdat
= NODE_DATA(nid
);
5104 build_zonelists(pgdat
);
5108 * Initialize the boot_pagesets that are going to be used
5109 * for bootstrapping processors. The real pagesets for
5110 * each zone will be allocated later when the per cpu
5111 * allocator is available.
5113 * boot_pagesets are used also for bootstrapping offline
5114 * cpus if the system is already booted because the pagesets
5115 * are needed to initialize allocators on a specific cpu too.
5116 * F.e. the percpu allocator needs the page allocator which
5117 * needs the percpu allocator in order to allocate its pagesets
5118 * (a chicken-egg dilemma).
5120 for_each_possible_cpu(cpu
) {
5121 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5123 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5125 * We now know the "local memory node" for each node--
5126 * i.e., the node of the first zone in the generic zonelist.
5127 * Set up numa_mem percpu variable for on-line cpus. During
5128 * boot, only the boot cpu should be on-line; we'll init the
5129 * secondary cpus' numa_mem as they come on-line. During
5130 * node/memory hotplug, we'll fixup all on-line cpus.
5132 if (cpu_online(cpu
))
5133 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5140 static noinline
void __init
5141 build_all_zonelists_init(void)
5143 __build_all_zonelists(NULL
);
5144 mminit_verify_zonelist();
5145 cpuset_init_current_mems_allowed();
5149 * Called with zonelists_mutex held always
5150 * unless system_state == SYSTEM_BOOTING.
5152 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5153 * [we're only called with non-NULL zone through __meminit paths] and
5154 * (2) call of __init annotated helper build_all_zonelists_init
5155 * [protected by SYSTEM_BOOTING].
5157 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
5159 set_zonelist_order();
5161 if (system_state
== SYSTEM_BOOTING
) {
5162 build_all_zonelists_init();
5164 #ifdef CONFIG_MEMORY_HOTPLUG
5166 setup_zone_pageset(zone
);
5168 /* we have to stop all cpus to guarantee there is no user
5170 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
5171 /* cpuset refresh routine should be here */
5173 vm_total_pages
= nr_free_pagecache_pages();
5175 * Disable grouping by mobility if the number of pages in the
5176 * system is too low to allow the mechanism to work. It would be
5177 * more accurate, but expensive to check per-zone. This check is
5178 * made on memory-hotadd so a system can start with mobility
5179 * disabled and enable it later
5181 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5182 page_group_by_mobility_disabled
= 1;
5184 page_group_by_mobility_disabled
= 0;
5186 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5188 zonelist_order_name
[current_zonelist_order
],
5189 page_group_by_mobility_disabled
? "off" : "on",
5192 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5197 * Initially all pages are reserved - free ones are freed
5198 * up by free_all_bootmem() once the early boot process is
5199 * done. Non-atomic initialization, single-pass.
5201 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5202 unsigned long start_pfn
, enum memmap_context context
)
5204 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5205 unsigned long end_pfn
= start_pfn
+ size
;
5206 pg_data_t
*pgdat
= NODE_DATA(nid
);
5208 unsigned long nr_initialised
= 0;
5209 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5210 struct memblock_region
*r
= NULL
, *tmp
;
5213 if (highest_memmap_pfn
< end_pfn
- 1)
5214 highest_memmap_pfn
= end_pfn
- 1;
5217 * Honor reservation requested by the driver for this ZONE_DEVICE
5220 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5221 start_pfn
+= altmap
->reserve
;
5223 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5225 * There can be holes in boot-time mem_map[]s handed to this
5226 * function. They do not exist on hotplugged memory.
5228 if (context
!= MEMMAP_EARLY
)
5231 if (!early_pfn_valid(pfn
)) {
5232 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5234 * Skip to the pfn preceding the next valid one (or
5235 * end_pfn), such that we hit a valid pfn (or end_pfn)
5236 * on our next iteration of the loop.
5238 pfn
= memblock_next_valid_pfn(pfn
, end_pfn
) - 1;
5242 if (!early_pfn_in_nid(pfn
, nid
))
5244 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5247 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5249 * Check given memblock attribute by firmware which can affect
5250 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5251 * mirrored, it's an overlapped memmap init. skip it.
5253 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5254 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5255 for_each_memblock(memory
, tmp
)
5256 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5260 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5261 memblock_is_mirror(r
)) {
5262 /* already initialized as NORMAL */
5263 pfn
= memblock_region_memory_end_pfn(r
);
5271 * Mark the block movable so that blocks are reserved for
5272 * movable at startup. This will force kernel allocations
5273 * to reserve their blocks rather than leaking throughout
5274 * the address space during boot when many long-lived
5275 * kernel allocations are made.
5277 * bitmap is created for zone's valid pfn range. but memmap
5278 * can be created for invalid pages (for alignment)
5279 * check here not to call set_pageblock_migratetype() against
5282 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5283 struct page
*page
= pfn_to_page(pfn
);
5285 __init_single_page(page
, pfn
, zone
, nid
);
5286 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5288 __init_single_pfn(pfn
, zone
, nid
);
5293 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5295 unsigned int order
, t
;
5296 for_each_migratetype_order(order
, t
) {
5297 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5298 zone
->free_area
[order
].nr_free
= 0;
5302 #ifndef __HAVE_ARCH_MEMMAP_INIT
5303 #define memmap_init(size, nid, zone, start_pfn) \
5304 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5307 static int zone_batchsize(struct zone
*zone
)
5313 * The per-cpu-pages pools are set to around 1000th of the
5314 * size of the zone. But no more than 1/2 of a meg.
5316 * OK, so we don't know how big the cache is. So guess.
5318 batch
= zone
->managed_pages
/ 1024;
5319 if (batch
* PAGE_SIZE
> 512 * 1024)
5320 batch
= (512 * 1024) / PAGE_SIZE
;
5321 batch
/= 4; /* We effectively *= 4 below */
5326 * Clamp the batch to a 2^n - 1 value. Having a power
5327 * of 2 value was found to be more likely to have
5328 * suboptimal cache aliasing properties in some cases.
5330 * For example if 2 tasks are alternately allocating
5331 * batches of pages, one task can end up with a lot
5332 * of pages of one half of the possible page colors
5333 * and the other with pages of the other colors.
5335 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5340 /* The deferral and batching of frees should be suppressed under NOMMU
5343 * The problem is that NOMMU needs to be able to allocate large chunks
5344 * of contiguous memory as there's no hardware page translation to
5345 * assemble apparent contiguous memory from discontiguous pages.
5347 * Queueing large contiguous runs of pages for batching, however,
5348 * causes the pages to actually be freed in smaller chunks. As there
5349 * can be a significant delay between the individual batches being
5350 * recycled, this leads to the once large chunks of space being
5351 * fragmented and becoming unavailable for high-order allocations.
5358 * pcp->high and pcp->batch values are related and dependent on one another:
5359 * ->batch must never be higher then ->high.
5360 * The following function updates them in a safe manner without read side
5363 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5364 * those fields changing asynchronously (acording the the above rule).
5366 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5367 * outside of boot time (or some other assurance that no concurrent updaters
5370 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5371 unsigned long batch
)
5373 /* start with a fail safe value for batch */
5377 /* Update high, then batch, in order */
5384 /* a companion to pageset_set_high() */
5385 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5387 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5390 static void pageset_init(struct per_cpu_pageset
*p
)
5392 struct per_cpu_pages
*pcp
;
5395 memset(p
, 0, sizeof(*p
));
5399 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5400 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5403 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5406 pageset_set_batch(p
, batch
);
5410 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5411 * to the value high for the pageset p.
5413 static void pageset_set_high(struct per_cpu_pageset
*p
,
5416 unsigned long batch
= max(1UL, high
/ 4);
5417 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5418 batch
= PAGE_SHIFT
* 8;
5420 pageset_update(&p
->pcp
, high
, batch
);
5423 static void pageset_set_high_and_batch(struct zone
*zone
,
5424 struct per_cpu_pageset
*pcp
)
5426 if (percpu_pagelist_fraction
)
5427 pageset_set_high(pcp
,
5428 (zone
->managed_pages
/
5429 percpu_pagelist_fraction
));
5431 pageset_set_batch(pcp
, zone_batchsize(zone
));
5434 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5436 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5439 pageset_set_high_and_batch(zone
, pcp
);
5442 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5445 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5446 for_each_possible_cpu(cpu
)
5447 zone_pageset_init(zone
, cpu
);
5451 * Allocate per cpu pagesets and initialize them.
5452 * Before this call only boot pagesets were available.
5454 void __init
setup_per_cpu_pageset(void)
5456 struct pglist_data
*pgdat
;
5459 for_each_populated_zone(zone
)
5460 setup_zone_pageset(zone
);
5462 for_each_online_pgdat(pgdat
)
5463 pgdat
->per_cpu_nodestats
=
5464 alloc_percpu(struct per_cpu_nodestat
);
5467 static __meminit
void zone_pcp_init(struct zone
*zone
)
5470 * per cpu subsystem is not up at this point. The following code
5471 * relies on the ability of the linker to provide the
5472 * offset of a (static) per cpu variable into the per cpu area.
5474 zone
->pageset
= &boot_pageset
;
5476 if (populated_zone(zone
))
5477 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5478 zone
->name
, zone
->present_pages
,
5479 zone_batchsize(zone
));
5482 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5483 unsigned long zone_start_pfn
,
5486 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5488 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5490 zone
->zone_start_pfn
= zone_start_pfn
;
5492 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5493 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5495 (unsigned long)zone_idx(zone
),
5496 zone_start_pfn
, (zone_start_pfn
+ size
));
5498 zone_init_free_lists(zone
);
5499 zone
->initialized
= 1;
5504 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5505 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5508 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5510 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5511 struct mminit_pfnnid_cache
*state
)
5513 unsigned long start_pfn
, end_pfn
;
5516 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5517 return state
->last_nid
;
5519 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5521 state
->last_start
= start_pfn
;
5522 state
->last_end
= end_pfn
;
5523 state
->last_nid
= nid
;
5528 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5531 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5532 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5533 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5535 * If an architecture guarantees that all ranges registered contain no holes
5536 * and may be freed, this this function may be used instead of calling
5537 * memblock_free_early_nid() manually.
5539 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5541 unsigned long start_pfn
, end_pfn
;
5544 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5545 start_pfn
= min(start_pfn
, max_low_pfn
);
5546 end_pfn
= min(end_pfn
, max_low_pfn
);
5548 if (start_pfn
< end_pfn
)
5549 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5550 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5556 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5557 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5559 * If an architecture guarantees that all ranges registered contain no holes and may
5560 * be freed, this function may be used instead of calling memory_present() manually.
5562 void __init
sparse_memory_present_with_active_regions(int nid
)
5564 unsigned long start_pfn
, end_pfn
;
5567 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5568 memory_present(this_nid
, start_pfn
, end_pfn
);
5572 * get_pfn_range_for_nid - Return the start and end page frames for a node
5573 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5574 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5575 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5577 * It returns the start and end page frame of a node based on information
5578 * provided by memblock_set_node(). If called for a node
5579 * with no available memory, a warning is printed and the start and end
5582 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5583 unsigned long *start_pfn
, unsigned long *end_pfn
)
5585 unsigned long this_start_pfn
, this_end_pfn
;
5591 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5592 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5593 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5596 if (*start_pfn
== -1UL)
5601 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5602 * assumption is made that zones within a node are ordered in monotonic
5603 * increasing memory addresses so that the "highest" populated zone is used
5605 static void __init
find_usable_zone_for_movable(void)
5608 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5609 if (zone_index
== ZONE_MOVABLE
)
5612 if (arch_zone_highest_possible_pfn
[zone_index
] >
5613 arch_zone_lowest_possible_pfn
[zone_index
])
5617 VM_BUG_ON(zone_index
== -1);
5618 movable_zone
= zone_index
;
5622 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5623 * because it is sized independent of architecture. Unlike the other zones,
5624 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5625 * in each node depending on the size of each node and how evenly kernelcore
5626 * is distributed. This helper function adjusts the zone ranges
5627 * provided by the architecture for a given node by using the end of the
5628 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5629 * zones within a node are in order of monotonic increases memory addresses
5631 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5632 unsigned long zone_type
,
5633 unsigned long node_start_pfn
,
5634 unsigned long node_end_pfn
,
5635 unsigned long *zone_start_pfn
,
5636 unsigned long *zone_end_pfn
)
5638 /* Only adjust if ZONE_MOVABLE is on this node */
5639 if (zone_movable_pfn
[nid
]) {
5640 /* Size ZONE_MOVABLE */
5641 if (zone_type
== ZONE_MOVABLE
) {
5642 *zone_start_pfn
= zone_movable_pfn
[nid
];
5643 *zone_end_pfn
= min(node_end_pfn
,
5644 arch_zone_highest_possible_pfn
[movable_zone
]);
5646 /* Adjust for ZONE_MOVABLE starting within this range */
5647 } else if (!mirrored_kernelcore
&&
5648 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5649 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5650 *zone_end_pfn
= zone_movable_pfn
[nid
];
5652 /* Check if this whole range is within ZONE_MOVABLE */
5653 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5654 *zone_start_pfn
= *zone_end_pfn
;
5659 * Return the number of pages a zone spans in a node, including holes
5660 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5662 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5663 unsigned long zone_type
,
5664 unsigned long node_start_pfn
,
5665 unsigned long node_end_pfn
,
5666 unsigned long *zone_start_pfn
,
5667 unsigned long *zone_end_pfn
,
5668 unsigned long *ignored
)
5670 /* When hotadd a new node from cpu_up(), the node should be empty */
5671 if (!node_start_pfn
&& !node_end_pfn
)
5674 /* Get the start and end of the zone */
5675 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5676 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5677 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5678 node_start_pfn
, node_end_pfn
,
5679 zone_start_pfn
, zone_end_pfn
);
5681 /* Check that this node has pages within the zone's required range */
5682 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5685 /* Move the zone boundaries inside the node if necessary */
5686 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5687 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5689 /* Return the spanned pages */
5690 return *zone_end_pfn
- *zone_start_pfn
;
5694 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5695 * then all holes in the requested range will be accounted for.
5697 unsigned long __meminit
__absent_pages_in_range(int nid
,
5698 unsigned long range_start_pfn
,
5699 unsigned long range_end_pfn
)
5701 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5702 unsigned long start_pfn
, end_pfn
;
5705 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5706 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5707 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5708 nr_absent
-= end_pfn
- start_pfn
;
5714 * absent_pages_in_range - Return number of page frames in holes within a range
5715 * @start_pfn: The start PFN to start searching for holes
5716 * @end_pfn: The end PFN to stop searching for holes
5718 * It returns the number of pages frames in memory holes within a range.
5720 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5721 unsigned long end_pfn
)
5723 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5726 /* Return the number of page frames in holes in a zone on a node */
5727 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5728 unsigned long zone_type
,
5729 unsigned long node_start_pfn
,
5730 unsigned long node_end_pfn
,
5731 unsigned long *ignored
)
5733 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5734 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5735 unsigned long zone_start_pfn
, zone_end_pfn
;
5736 unsigned long nr_absent
;
5738 /* When hotadd a new node from cpu_up(), the node should be empty */
5739 if (!node_start_pfn
&& !node_end_pfn
)
5742 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5743 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5745 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5746 node_start_pfn
, node_end_pfn
,
5747 &zone_start_pfn
, &zone_end_pfn
);
5748 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5751 * ZONE_MOVABLE handling.
5752 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5755 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5756 unsigned long start_pfn
, end_pfn
;
5757 struct memblock_region
*r
;
5759 for_each_memblock(memory
, r
) {
5760 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5761 zone_start_pfn
, zone_end_pfn
);
5762 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5763 zone_start_pfn
, zone_end_pfn
);
5765 if (zone_type
== ZONE_MOVABLE
&&
5766 memblock_is_mirror(r
))
5767 nr_absent
+= end_pfn
- start_pfn
;
5769 if (zone_type
== ZONE_NORMAL
&&
5770 !memblock_is_mirror(r
))
5771 nr_absent
+= end_pfn
- start_pfn
;
5778 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5779 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5780 unsigned long zone_type
,
5781 unsigned long node_start_pfn
,
5782 unsigned long node_end_pfn
,
5783 unsigned long *zone_start_pfn
,
5784 unsigned long *zone_end_pfn
,
5785 unsigned long *zones_size
)
5789 *zone_start_pfn
= node_start_pfn
;
5790 for (zone
= 0; zone
< zone_type
; zone
++)
5791 *zone_start_pfn
+= zones_size
[zone
];
5793 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5795 return zones_size
[zone_type
];
5798 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5799 unsigned long zone_type
,
5800 unsigned long node_start_pfn
,
5801 unsigned long node_end_pfn
,
5802 unsigned long *zholes_size
)
5807 return zholes_size
[zone_type
];
5810 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5812 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5813 unsigned long node_start_pfn
,
5814 unsigned long node_end_pfn
,
5815 unsigned long *zones_size
,
5816 unsigned long *zholes_size
)
5818 unsigned long realtotalpages
= 0, totalpages
= 0;
5821 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5822 struct zone
*zone
= pgdat
->node_zones
+ i
;
5823 unsigned long zone_start_pfn
, zone_end_pfn
;
5824 unsigned long size
, real_size
;
5826 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5832 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5833 node_start_pfn
, node_end_pfn
,
5836 zone
->zone_start_pfn
= zone_start_pfn
;
5838 zone
->zone_start_pfn
= 0;
5839 zone
->spanned_pages
= size
;
5840 zone
->present_pages
= real_size
;
5843 realtotalpages
+= real_size
;
5846 pgdat
->node_spanned_pages
= totalpages
;
5847 pgdat
->node_present_pages
= realtotalpages
;
5848 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5852 #ifndef CONFIG_SPARSEMEM
5854 * Calculate the size of the zone->blockflags rounded to an unsigned long
5855 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5856 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5857 * round what is now in bits to nearest long in bits, then return it in
5860 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5862 unsigned long usemapsize
;
5864 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5865 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5866 usemapsize
= usemapsize
>> pageblock_order
;
5867 usemapsize
*= NR_PAGEBLOCK_BITS
;
5868 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5870 return usemapsize
/ 8;
5873 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5875 unsigned long zone_start_pfn
,
5876 unsigned long zonesize
)
5878 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5879 zone
->pageblock_flags
= NULL
;
5881 zone
->pageblock_flags
=
5882 memblock_virt_alloc_node_nopanic(usemapsize
,
5886 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5887 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5888 #endif /* CONFIG_SPARSEMEM */
5890 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5892 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5893 void __paginginit
set_pageblock_order(void)
5897 /* Check that pageblock_nr_pages has not already been setup */
5898 if (pageblock_order
)
5901 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5902 order
= HUGETLB_PAGE_ORDER
;
5904 order
= MAX_ORDER
- 1;
5907 * Assume the largest contiguous order of interest is a huge page.
5908 * This value may be variable depending on boot parameters on IA64 and
5911 pageblock_order
= order
;
5913 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5916 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5917 * is unused as pageblock_order is set at compile-time. See
5918 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5921 void __paginginit
set_pageblock_order(void)
5925 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5927 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5928 unsigned long present_pages
)
5930 unsigned long pages
= spanned_pages
;
5933 * Provide a more accurate estimation if there are holes within
5934 * the zone and SPARSEMEM is in use. If there are holes within the
5935 * zone, each populated memory region may cost us one or two extra
5936 * memmap pages due to alignment because memmap pages for each
5937 * populated regions may not be naturally aligned on page boundary.
5938 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5940 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5941 IS_ENABLED(CONFIG_SPARSEMEM
))
5942 pages
= present_pages
;
5944 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5948 * Set up the zone data structures:
5949 * - mark all pages reserved
5950 * - mark all memory queues empty
5951 * - clear the memory bitmaps
5953 * NOTE: pgdat should get zeroed by caller.
5955 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5958 int nid
= pgdat
->node_id
;
5961 pgdat_resize_init(pgdat
);
5962 #ifdef CONFIG_NUMA_BALANCING
5963 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5964 pgdat
->numabalancing_migrate_nr_pages
= 0;
5965 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5967 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5968 spin_lock_init(&pgdat
->split_queue_lock
);
5969 INIT_LIST_HEAD(&pgdat
->split_queue
);
5970 pgdat
->split_queue_len
= 0;
5972 init_waitqueue_head(&pgdat
->kswapd_wait
);
5973 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5974 #ifdef CONFIG_COMPACTION
5975 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5977 pgdat_page_ext_init(pgdat
);
5978 spin_lock_init(&pgdat
->lru_lock
);
5979 lruvec_init(node_lruvec(pgdat
));
5981 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5982 struct zone
*zone
= pgdat
->node_zones
+ j
;
5983 unsigned long size
, realsize
, freesize
, memmap_pages
;
5984 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5986 size
= zone
->spanned_pages
;
5987 realsize
= freesize
= zone
->present_pages
;
5990 * Adjust freesize so that it accounts for how much memory
5991 * is used by this zone for memmap. This affects the watermark
5992 * and per-cpu initialisations
5994 memmap_pages
= calc_memmap_size(size
, realsize
);
5995 if (!is_highmem_idx(j
)) {
5996 if (freesize
>= memmap_pages
) {
5997 freesize
-= memmap_pages
;
6000 " %s zone: %lu pages used for memmap\n",
6001 zone_names
[j
], memmap_pages
);
6003 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6004 zone_names
[j
], memmap_pages
, freesize
);
6007 /* Account for reserved pages */
6008 if (j
== 0 && freesize
> dma_reserve
) {
6009 freesize
-= dma_reserve
;
6010 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6011 zone_names
[0], dma_reserve
);
6014 if (!is_highmem_idx(j
))
6015 nr_kernel_pages
+= freesize
;
6016 /* Charge for highmem memmap if there are enough kernel pages */
6017 else if (nr_kernel_pages
> memmap_pages
* 2)
6018 nr_kernel_pages
-= memmap_pages
;
6019 nr_all_pages
+= freesize
;
6022 * Set an approximate value for lowmem here, it will be adjusted
6023 * when the bootmem allocator frees pages into the buddy system.
6024 * And all highmem pages will be managed by the buddy system.
6026 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
6030 zone
->name
= zone_names
[j
];
6031 zone
->zone_pgdat
= pgdat
;
6032 spin_lock_init(&zone
->lock
);
6033 zone_seqlock_init(zone
);
6034 zone_pcp_init(zone
);
6039 set_pageblock_order();
6040 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6041 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6043 memmap_init(size
, nid
, j
, zone_start_pfn
);
6047 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6049 unsigned long __maybe_unused start
= 0;
6050 unsigned long __maybe_unused offset
= 0;
6052 /* Skip empty nodes */
6053 if (!pgdat
->node_spanned_pages
)
6056 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6057 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6058 offset
= pgdat
->node_start_pfn
- start
;
6059 /* ia64 gets its own node_mem_map, before this, without bootmem */
6060 if (!pgdat
->node_mem_map
) {
6061 unsigned long size
, end
;
6065 * The zone's endpoints aren't required to be MAX_ORDER
6066 * aligned but the node_mem_map endpoints must be in order
6067 * for the buddy allocator to function correctly.
6069 end
= pgdat_end_pfn(pgdat
);
6070 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6071 size
= (end
- start
) * sizeof(struct page
);
6072 map
= alloc_remap(pgdat
->node_id
, size
);
6074 map
= memblock_virt_alloc_node_nopanic(size
,
6076 pgdat
->node_mem_map
= map
+ offset
;
6078 #ifndef CONFIG_NEED_MULTIPLE_NODES
6080 * With no DISCONTIG, the global mem_map is just set as node 0's
6082 if (pgdat
== NODE_DATA(0)) {
6083 mem_map
= NODE_DATA(0)->node_mem_map
;
6084 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6085 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6087 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6090 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6093 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
6094 unsigned long node_start_pfn
, unsigned long *zholes_size
)
6096 pg_data_t
*pgdat
= NODE_DATA(nid
);
6097 unsigned long start_pfn
= 0;
6098 unsigned long end_pfn
= 0;
6100 /* pg_data_t should be reset to zero when it's allocated */
6101 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6103 reset_deferred_meminit(pgdat
);
6104 pgdat
->node_id
= nid
;
6105 pgdat
->node_start_pfn
= node_start_pfn
;
6106 pgdat
->per_cpu_nodestats
= NULL
;
6107 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6108 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6109 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6110 (u64
)start_pfn
<< PAGE_SHIFT
,
6111 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6113 start_pfn
= node_start_pfn
;
6115 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6116 zones_size
, zholes_size
);
6118 alloc_node_mem_map(pgdat
);
6119 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6120 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6121 nid
, (unsigned long)pgdat
,
6122 (unsigned long)pgdat
->node_mem_map
);
6125 free_area_init_core(pgdat
);
6128 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6130 #if MAX_NUMNODES > 1
6132 * Figure out the number of possible node ids.
6134 void __init
setup_nr_node_ids(void)
6136 unsigned int highest
;
6138 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6139 nr_node_ids
= highest
+ 1;
6144 * node_map_pfn_alignment - determine the maximum internode alignment
6146 * This function should be called after node map is populated and sorted.
6147 * It calculates the maximum power of two alignment which can distinguish
6150 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6151 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6152 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6153 * shifted, 1GiB is enough and this function will indicate so.
6155 * This is used to test whether pfn -> nid mapping of the chosen memory
6156 * model has fine enough granularity to avoid incorrect mapping for the
6157 * populated node map.
6159 * Returns the determined alignment in pfn's. 0 if there is no alignment
6160 * requirement (single node).
6162 unsigned long __init
node_map_pfn_alignment(void)
6164 unsigned long accl_mask
= 0, last_end
= 0;
6165 unsigned long start
, end
, mask
;
6169 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6170 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6177 * Start with a mask granular enough to pin-point to the
6178 * start pfn and tick off bits one-by-one until it becomes
6179 * too coarse to separate the current node from the last.
6181 mask
= ~((1 << __ffs(start
)) - 1);
6182 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6185 /* accumulate all internode masks */
6189 /* convert mask to number of pages */
6190 return ~accl_mask
+ 1;
6193 /* Find the lowest pfn for a node */
6194 static unsigned long __init
find_min_pfn_for_node(int nid
)
6196 unsigned long min_pfn
= ULONG_MAX
;
6197 unsigned long start_pfn
;
6200 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6201 min_pfn
= min(min_pfn
, start_pfn
);
6203 if (min_pfn
== ULONG_MAX
) {
6204 pr_warn("Could not find start_pfn for node %d\n", nid
);
6212 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6214 * It returns the minimum PFN based on information provided via
6215 * memblock_set_node().
6217 unsigned long __init
find_min_pfn_with_active_regions(void)
6219 return find_min_pfn_for_node(MAX_NUMNODES
);
6223 * early_calculate_totalpages()
6224 * Sum pages in active regions for movable zone.
6225 * Populate N_MEMORY for calculating usable_nodes.
6227 static unsigned long __init
early_calculate_totalpages(void)
6229 unsigned long totalpages
= 0;
6230 unsigned long start_pfn
, end_pfn
;
6233 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6234 unsigned long pages
= end_pfn
- start_pfn
;
6236 totalpages
+= pages
;
6238 node_set_state(nid
, N_MEMORY
);
6244 * Find the PFN the Movable zone begins in each node. Kernel memory
6245 * is spread evenly between nodes as long as the nodes have enough
6246 * memory. When they don't, some nodes will have more kernelcore than
6249 static void __init
find_zone_movable_pfns_for_nodes(void)
6252 unsigned long usable_startpfn
;
6253 unsigned long kernelcore_node
, kernelcore_remaining
;
6254 /* save the state before borrow the nodemask */
6255 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6256 unsigned long totalpages
= early_calculate_totalpages();
6257 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6258 struct memblock_region
*r
;
6260 /* Need to find movable_zone earlier when movable_node is specified. */
6261 find_usable_zone_for_movable();
6264 * If movable_node is specified, ignore kernelcore and movablecore
6267 if (movable_node_is_enabled()) {
6268 for_each_memblock(memory
, r
) {
6269 if (!memblock_is_hotpluggable(r
))
6274 usable_startpfn
= PFN_DOWN(r
->base
);
6275 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6276 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6284 * If kernelcore=mirror is specified, ignore movablecore option
6286 if (mirrored_kernelcore
) {
6287 bool mem_below_4gb_not_mirrored
= false;
6289 for_each_memblock(memory
, r
) {
6290 if (memblock_is_mirror(r
))
6295 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6297 if (usable_startpfn
< 0x100000) {
6298 mem_below_4gb_not_mirrored
= true;
6302 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6303 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6307 if (mem_below_4gb_not_mirrored
)
6308 pr_warn("This configuration results in unmirrored kernel memory.");
6314 * If movablecore=nn[KMG] was specified, calculate what size of
6315 * kernelcore that corresponds so that memory usable for
6316 * any allocation type is evenly spread. If both kernelcore
6317 * and movablecore are specified, then the value of kernelcore
6318 * will be used for required_kernelcore if it's greater than
6319 * what movablecore would have allowed.
6321 if (required_movablecore
) {
6322 unsigned long corepages
;
6325 * Round-up so that ZONE_MOVABLE is at least as large as what
6326 * was requested by the user
6328 required_movablecore
=
6329 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6330 required_movablecore
= min(totalpages
, required_movablecore
);
6331 corepages
= totalpages
- required_movablecore
;
6333 required_kernelcore
= max(required_kernelcore
, corepages
);
6337 * If kernelcore was not specified or kernelcore size is larger
6338 * than totalpages, there is no ZONE_MOVABLE.
6340 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6343 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6344 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6347 /* Spread kernelcore memory as evenly as possible throughout nodes */
6348 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6349 for_each_node_state(nid
, N_MEMORY
) {
6350 unsigned long start_pfn
, end_pfn
;
6353 * Recalculate kernelcore_node if the division per node
6354 * now exceeds what is necessary to satisfy the requested
6355 * amount of memory for the kernel
6357 if (required_kernelcore
< kernelcore_node
)
6358 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6361 * As the map is walked, we track how much memory is usable
6362 * by the kernel using kernelcore_remaining. When it is
6363 * 0, the rest of the node is usable by ZONE_MOVABLE
6365 kernelcore_remaining
= kernelcore_node
;
6367 /* Go through each range of PFNs within this node */
6368 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6369 unsigned long size_pages
;
6371 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6372 if (start_pfn
>= end_pfn
)
6375 /* Account for what is only usable for kernelcore */
6376 if (start_pfn
< usable_startpfn
) {
6377 unsigned long kernel_pages
;
6378 kernel_pages
= min(end_pfn
, usable_startpfn
)
6381 kernelcore_remaining
-= min(kernel_pages
,
6382 kernelcore_remaining
);
6383 required_kernelcore
-= min(kernel_pages
,
6384 required_kernelcore
);
6386 /* Continue if range is now fully accounted */
6387 if (end_pfn
<= usable_startpfn
) {
6390 * Push zone_movable_pfn to the end so
6391 * that if we have to rebalance
6392 * kernelcore across nodes, we will
6393 * not double account here
6395 zone_movable_pfn
[nid
] = end_pfn
;
6398 start_pfn
= usable_startpfn
;
6402 * The usable PFN range for ZONE_MOVABLE is from
6403 * start_pfn->end_pfn. Calculate size_pages as the
6404 * number of pages used as kernelcore
6406 size_pages
= end_pfn
- start_pfn
;
6407 if (size_pages
> kernelcore_remaining
)
6408 size_pages
= kernelcore_remaining
;
6409 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6412 * Some kernelcore has been met, update counts and
6413 * break if the kernelcore for this node has been
6416 required_kernelcore
-= min(required_kernelcore
,
6418 kernelcore_remaining
-= size_pages
;
6419 if (!kernelcore_remaining
)
6425 * If there is still required_kernelcore, we do another pass with one
6426 * less node in the count. This will push zone_movable_pfn[nid] further
6427 * along on the nodes that still have memory until kernelcore is
6431 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6435 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6436 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6437 zone_movable_pfn
[nid
] =
6438 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6441 /* restore the node_state */
6442 node_states
[N_MEMORY
] = saved_node_state
;
6445 /* Any regular or high memory on that node ? */
6446 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6448 enum zone_type zone_type
;
6450 if (N_MEMORY
== N_NORMAL_MEMORY
)
6453 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6454 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6455 if (populated_zone(zone
)) {
6456 node_set_state(nid
, N_HIGH_MEMORY
);
6457 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6458 zone_type
<= ZONE_NORMAL
)
6459 node_set_state(nid
, N_NORMAL_MEMORY
);
6466 * free_area_init_nodes - Initialise all pg_data_t and zone data
6467 * @max_zone_pfn: an array of max PFNs for each zone
6469 * This will call free_area_init_node() for each active node in the system.
6470 * Using the page ranges provided by memblock_set_node(), the size of each
6471 * zone in each node and their holes is calculated. If the maximum PFN
6472 * between two adjacent zones match, it is assumed that the zone is empty.
6473 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6474 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6475 * starts where the previous one ended. For example, ZONE_DMA32 starts
6476 * at arch_max_dma_pfn.
6478 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6480 unsigned long start_pfn
, end_pfn
;
6483 /* Record where the zone boundaries are */
6484 memset(arch_zone_lowest_possible_pfn
, 0,
6485 sizeof(arch_zone_lowest_possible_pfn
));
6486 memset(arch_zone_highest_possible_pfn
, 0,
6487 sizeof(arch_zone_highest_possible_pfn
));
6489 start_pfn
= find_min_pfn_with_active_regions();
6491 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6492 if (i
== ZONE_MOVABLE
)
6495 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6496 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6497 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6499 start_pfn
= end_pfn
;
6502 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6503 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6504 find_zone_movable_pfns_for_nodes();
6506 /* Print out the zone ranges */
6507 pr_info("Zone ranges:\n");
6508 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6509 if (i
== ZONE_MOVABLE
)
6511 pr_info(" %-8s ", zone_names
[i
]);
6512 if (arch_zone_lowest_possible_pfn
[i
] ==
6513 arch_zone_highest_possible_pfn
[i
])
6516 pr_cont("[mem %#018Lx-%#018Lx]\n",
6517 (u64
)arch_zone_lowest_possible_pfn
[i
]
6519 ((u64
)arch_zone_highest_possible_pfn
[i
]
6520 << PAGE_SHIFT
) - 1);
6523 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6524 pr_info("Movable zone start for each node\n");
6525 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6526 if (zone_movable_pfn
[i
])
6527 pr_info(" Node %d: %#018Lx\n", i
,
6528 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6531 /* Print out the early node map */
6532 pr_info("Early memory node ranges\n");
6533 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6534 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6535 (u64
)start_pfn
<< PAGE_SHIFT
,
6536 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6538 /* Initialise every node */
6539 mminit_verify_pageflags_layout();
6540 setup_nr_node_ids();
6541 for_each_online_node(nid
) {
6542 pg_data_t
*pgdat
= NODE_DATA(nid
);
6543 free_area_init_node(nid
, NULL
,
6544 find_min_pfn_for_node(nid
), NULL
);
6546 /* Any memory on that node */
6547 if (pgdat
->node_present_pages
)
6548 node_set_state(nid
, N_MEMORY
);
6549 check_for_memory(pgdat
, nid
);
6553 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6555 unsigned long long coremem
;
6559 coremem
= memparse(p
, &p
);
6560 *core
= coremem
>> PAGE_SHIFT
;
6562 /* Paranoid check that UL is enough for the coremem value */
6563 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6569 * kernelcore=size sets the amount of memory for use for allocations that
6570 * cannot be reclaimed or migrated.
6572 static int __init
cmdline_parse_kernelcore(char *p
)
6574 /* parse kernelcore=mirror */
6575 if (parse_option_str(p
, "mirror")) {
6576 mirrored_kernelcore
= true;
6580 return cmdline_parse_core(p
, &required_kernelcore
);
6584 * movablecore=size sets the amount of memory for use for allocations that
6585 * can be reclaimed or migrated.
6587 static int __init
cmdline_parse_movablecore(char *p
)
6589 return cmdline_parse_core(p
, &required_movablecore
);
6592 early_param("kernelcore", cmdline_parse_kernelcore
);
6593 early_param("movablecore", cmdline_parse_movablecore
);
6595 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6597 void adjust_managed_page_count(struct page
*page
, long count
)
6599 spin_lock(&managed_page_count_lock
);
6600 page_zone(page
)->managed_pages
+= count
;
6601 totalram_pages
+= count
;
6602 #ifdef CONFIG_HIGHMEM
6603 if (PageHighMem(page
))
6604 totalhigh_pages
+= count
;
6606 spin_unlock(&managed_page_count_lock
);
6608 EXPORT_SYMBOL(adjust_managed_page_count
);
6610 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6613 unsigned long pages
= 0;
6615 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6616 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6617 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6618 if ((unsigned int)poison
<= 0xFF)
6619 memset(pos
, poison
, PAGE_SIZE
);
6620 free_reserved_page(virt_to_page(pos
));
6624 pr_info("Freeing %s memory: %ldK\n",
6625 s
, pages
<< (PAGE_SHIFT
- 10));
6629 EXPORT_SYMBOL(free_reserved_area
);
6631 #ifdef CONFIG_HIGHMEM
6632 void free_highmem_page(struct page
*page
)
6634 __free_reserved_page(page
);
6636 page_zone(page
)->managed_pages
++;
6642 void __init
mem_init_print_info(const char *str
)
6644 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6645 unsigned long init_code_size
, init_data_size
;
6647 physpages
= get_num_physpages();
6648 codesize
= _etext
- _stext
;
6649 datasize
= _edata
- _sdata
;
6650 rosize
= __end_rodata
- __start_rodata
;
6651 bss_size
= __bss_stop
- __bss_start
;
6652 init_data_size
= __init_end
- __init_begin
;
6653 init_code_size
= _einittext
- _sinittext
;
6656 * Detect special cases and adjust section sizes accordingly:
6657 * 1) .init.* may be embedded into .data sections
6658 * 2) .init.text.* may be out of [__init_begin, __init_end],
6659 * please refer to arch/tile/kernel/vmlinux.lds.S.
6660 * 3) .rodata.* may be embedded into .text or .data sections.
6662 #define adj_init_size(start, end, size, pos, adj) \
6664 if (start <= pos && pos < end && size > adj) \
6668 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6669 _sinittext
, init_code_size
);
6670 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6671 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6672 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6673 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6675 #undef adj_init_size
6677 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6678 #ifdef CONFIG_HIGHMEM
6682 nr_free_pages() << (PAGE_SHIFT
- 10),
6683 physpages
<< (PAGE_SHIFT
- 10),
6684 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6685 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6686 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6687 totalcma_pages
<< (PAGE_SHIFT
- 10),
6688 #ifdef CONFIG_HIGHMEM
6689 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6691 str
? ", " : "", str
? str
: "");
6695 * set_dma_reserve - set the specified number of pages reserved in the first zone
6696 * @new_dma_reserve: The number of pages to mark reserved
6698 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6699 * In the DMA zone, a significant percentage may be consumed by kernel image
6700 * and other unfreeable allocations which can skew the watermarks badly. This
6701 * function may optionally be used to account for unfreeable pages in the
6702 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6703 * smaller per-cpu batchsize.
6705 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6707 dma_reserve
= new_dma_reserve
;
6710 void __init
free_area_init(unsigned long *zones_size
)
6712 free_area_init_node(0, zones_size
,
6713 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6716 static int page_alloc_cpu_dead(unsigned int cpu
)
6719 lru_add_drain_cpu(cpu
);
6723 * Spill the event counters of the dead processor
6724 * into the current processors event counters.
6725 * This artificially elevates the count of the current
6728 vm_events_fold_cpu(cpu
);
6731 * Zero the differential counters of the dead processor
6732 * so that the vm statistics are consistent.
6734 * This is only okay since the processor is dead and cannot
6735 * race with what we are doing.
6737 cpu_vm_stats_fold(cpu
);
6741 void __init
page_alloc_init(void)
6745 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
6746 "mm/page_alloc:dead", NULL
,
6747 page_alloc_cpu_dead
);
6752 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6753 * or min_free_kbytes changes.
6755 static void calculate_totalreserve_pages(void)
6757 struct pglist_data
*pgdat
;
6758 unsigned long reserve_pages
= 0;
6759 enum zone_type i
, j
;
6761 for_each_online_pgdat(pgdat
) {
6763 pgdat
->totalreserve_pages
= 0;
6765 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6766 struct zone
*zone
= pgdat
->node_zones
+ i
;
6769 /* Find valid and maximum lowmem_reserve in the zone */
6770 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6771 if (zone
->lowmem_reserve
[j
] > max
)
6772 max
= zone
->lowmem_reserve
[j
];
6775 /* we treat the high watermark as reserved pages. */
6776 max
+= high_wmark_pages(zone
);
6778 if (max
> zone
->managed_pages
)
6779 max
= zone
->managed_pages
;
6781 pgdat
->totalreserve_pages
+= max
;
6783 reserve_pages
+= max
;
6786 totalreserve_pages
= reserve_pages
;
6790 * setup_per_zone_lowmem_reserve - called whenever
6791 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6792 * has a correct pages reserved value, so an adequate number of
6793 * pages are left in the zone after a successful __alloc_pages().
6795 static void setup_per_zone_lowmem_reserve(void)
6797 struct pglist_data
*pgdat
;
6798 enum zone_type j
, idx
;
6800 for_each_online_pgdat(pgdat
) {
6801 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6802 struct zone
*zone
= pgdat
->node_zones
+ j
;
6803 unsigned long managed_pages
= zone
->managed_pages
;
6805 zone
->lowmem_reserve
[j
] = 0;
6809 struct zone
*lower_zone
;
6813 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6814 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6816 lower_zone
= pgdat
->node_zones
+ idx
;
6817 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6818 sysctl_lowmem_reserve_ratio
[idx
];
6819 managed_pages
+= lower_zone
->managed_pages
;
6824 /* update totalreserve_pages */
6825 calculate_totalreserve_pages();
6828 static void __setup_per_zone_wmarks(void)
6830 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6831 unsigned long lowmem_pages
= 0;
6833 unsigned long flags
;
6835 /* Calculate total number of !ZONE_HIGHMEM pages */
6836 for_each_zone(zone
) {
6837 if (!is_highmem(zone
))
6838 lowmem_pages
+= zone
->managed_pages
;
6841 for_each_zone(zone
) {
6844 spin_lock_irqsave(&zone
->lock
, flags
);
6845 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6846 do_div(tmp
, lowmem_pages
);
6847 if (is_highmem(zone
)) {
6849 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6850 * need highmem pages, so cap pages_min to a small
6853 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6854 * deltas control asynch page reclaim, and so should
6855 * not be capped for highmem.
6857 unsigned long min_pages
;
6859 min_pages
= zone
->managed_pages
/ 1024;
6860 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6861 zone
->watermark
[WMARK_MIN
] = min_pages
;
6864 * If it's a lowmem zone, reserve a number of pages
6865 * proportionate to the zone's size.
6867 zone
->watermark
[WMARK_MIN
] = tmp
;
6871 * Set the kswapd watermarks distance according to the
6872 * scale factor in proportion to available memory, but
6873 * ensure a minimum size on small systems.
6875 tmp
= max_t(u64
, tmp
>> 2,
6876 mult_frac(zone
->managed_pages
,
6877 watermark_scale_factor
, 10000));
6879 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6880 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6882 spin_unlock_irqrestore(&zone
->lock
, flags
);
6885 /* update totalreserve_pages */
6886 calculate_totalreserve_pages();
6890 * setup_per_zone_wmarks - called when min_free_kbytes changes
6891 * or when memory is hot-{added|removed}
6893 * Ensures that the watermark[min,low,high] values for each zone are set
6894 * correctly with respect to min_free_kbytes.
6896 void setup_per_zone_wmarks(void)
6898 mutex_lock(&zonelists_mutex
);
6899 __setup_per_zone_wmarks();
6900 mutex_unlock(&zonelists_mutex
);
6904 * Initialise min_free_kbytes.
6906 * For small machines we want it small (128k min). For large machines
6907 * we want it large (64MB max). But it is not linear, because network
6908 * bandwidth does not increase linearly with machine size. We use
6910 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6911 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6927 int __meminit
init_per_zone_wmark_min(void)
6929 unsigned long lowmem_kbytes
;
6930 int new_min_free_kbytes
;
6932 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6933 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6935 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6936 min_free_kbytes
= new_min_free_kbytes
;
6937 if (min_free_kbytes
< 128)
6938 min_free_kbytes
= 128;
6939 if (min_free_kbytes
> 65536)
6940 min_free_kbytes
= 65536;
6942 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6943 new_min_free_kbytes
, user_min_free_kbytes
);
6945 setup_per_zone_wmarks();
6946 refresh_zone_stat_thresholds();
6947 setup_per_zone_lowmem_reserve();
6950 setup_min_unmapped_ratio();
6951 setup_min_slab_ratio();
6956 core_initcall(init_per_zone_wmark_min
)
6959 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6960 * that we can call two helper functions whenever min_free_kbytes
6963 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6964 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6968 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6973 user_min_free_kbytes
= min_free_kbytes
;
6974 setup_per_zone_wmarks();
6979 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6980 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6984 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6989 setup_per_zone_wmarks();
6995 static void setup_min_unmapped_ratio(void)
7000 for_each_online_pgdat(pgdat
)
7001 pgdat
->min_unmapped_pages
= 0;
7004 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
7005 sysctl_min_unmapped_ratio
) / 100;
7009 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7010 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7014 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7018 setup_min_unmapped_ratio();
7023 static void setup_min_slab_ratio(void)
7028 for_each_online_pgdat(pgdat
)
7029 pgdat
->min_slab_pages
= 0;
7032 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
7033 sysctl_min_slab_ratio
) / 100;
7036 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7037 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7041 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7045 setup_min_slab_ratio();
7052 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7053 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7054 * whenever sysctl_lowmem_reserve_ratio changes.
7056 * The reserve ratio obviously has absolutely no relation with the
7057 * minimum watermarks. The lowmem reserve ratio can only make sense
7058 * if in function of the boot time zone sizes.
7060 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7061 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7063 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7064 setup_per_zone_lowmem_reserve();
7069 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7070 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7071 * pagelist can have before it gets flushed back to buddy allocator.
7073 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7074 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7077 int old_percpu_pagelist_fraction
;
7080 mutex_lock(&pcp_batch_high_lock
);
7081 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7083 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7084 if (!write
|| ret
< 0)
7087 /* Sanity checking to avoid pcp imbalance */
7088 if (percpu_pagelist_fraction
&&
7089 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7090 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7096 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7099 for_each_populated_zone(zone
) {
7102 for_each_possible_cpu(cpu
)
7103 pageset_set_high_and_batch(zone
,
7104 per_cpu_ptr(zone
->pageset
, cpu
));
7107 mutex_unlock(&pcp_batch_high_lock
);
7112 int hashdist
= HASHDIST_DEFAULT
;
7114 static int __init
set_hashdist(char *str
)
7118 hashdist
= simple_strtoul(str
, &str
, 0);
7121 __setup("hashdist=", set_hashdist
);
7124 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7126 * Returns the number of pages that arch has reserved but
7127 * is not known to alloc_large_system_hash().
7129 static unsigned long __init
arch_reserved_kernel_pages(void)
7136 * allocate a large system hash table from bootmem
7137 * - it is assumed that the hash table must contain an exact power-of-2
7138 * quantity of entries
7139 * - limit is the number of hash buckets, not the total allocation size
7141 void *__init
alloc_large_system_hash(const char *tablename
,
7142 unsigned long bucketsize
,
7143 unsigned long numentries
,
7146 unsigned int *_hash_shift
,
7147 unsigned int *_hash_mask
,
7148 unsigned long low_limit
,
7149 unsigned long high_limit
)
7151 unsigned long long max
= high_limit
;
7152 unsigned long log2qty
, size
;
7155 /* allow the kernel cmdline to have a say */
7157 /* round applicable memory size up to nearest megabyte */
7158 numentries
= nr_kernel_pages
;
7159 numentries
-= arch_reserved_kernel_pages();
7161 /* It isn't necessary when PAGE_SIZE >= 1MB */
7162 if (PAGE_SHIFT
< 20)
7163 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7165 /* limit to 1 bucket per 2^scale bytes of low memory */
7166 if (scale
> PAGE_SHIFT
)
7167 numentries
>>= (scale
- PAGE_SHIFT
);
7169 numentries
<<= (PAGE_SHIFT
- scale
);
7171 /* Make sure we've got at least a 0-order allocation.. */
7172 if (unlikely(flags
& HASH_SMALL
)) {
7173 /* Makes no sense without HASH_EARLY */
7174 WARN_ON(!(flags
& HASH_EARLY
));
7175 if (!(numentries
>> *_hash_shift
)) {
7176 numentries
= 1UL << *_hash_shift
;
7177 BUG_ON(!numentries
);
7179 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7180 numentries
= PAGE_SIZE
/ bucketsize
;
7182 numentries
= roundup_pow_of_two(numentries
);
7184 /* limit allocation size to 1/16 total memory by default */
7186 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7187 do_div(max
, bucketsize
);
7189 max
= min(max
, 0x80000000ULL
);
7191 if (numentries
< low_limit
)
7192 numentries
= low_limit
;
7193 if (numentries
> max
)
7196 log2qty
= ilog2(numentries
);
7199 size
= bucketsize
<< log2qty
;
7200 if (flags
& HASH_EARLY
)
7201 table
= memblock_virt_alloc_nopanic(size
, 0);
7203 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
7206 * If bucketsize is not a power-of-two, we may free
7207 * some pages at the end of hash table which
7208 * alloc_pages_exact() automatically does
7210 if (get_order(size
) < MAX_ORDER
) {
7211 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
7212 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
7215 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7218 panic("Failed to allocate %s hash table\n", tablename
);
7220 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7221 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7224 *_hash_shift
= log2qty
;
7226 *_hash_mask
= (1 << log2qty
) - 1;
7232 * This function checks whether pageblock includes unmovable pages or not.
7233 * If @count is not zero, it is okay to include less @count unmovable pages
7235 * PageLRU check without isolation or lru_lock could race so that
7236 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7237 * check without lock_page also may miss some movable non-lru pages at
7238 * race condition. So you can't expect this function should be exact.
7240 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7241 bool skip_hwpoisoned_pages
)
7243 unsigned long pfn
, iter
, found
;
7247 * For avoiding noise data, lru_add_drain_all() should be called
7248 * If ZONE_MOVABLE, the zone never contains unmovable pages
7250 if (zone_idx(zone
) == ZONE_MOVABLE
)
7252 mt
= get_pageblock_migratetype(page
);
7253 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7256 pfn
= page_to_pfn(page
);
7257 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7258 unsigned long check
= pfn
+ iter
;
7260 if (!pfn_valid_within(check
))
7263 page
= pfn_to_page(check
);
7266 * Hugepages are not in LRU lists, but they're movable.
7267 * We need not scan over tail pages bacause we don't
7268 * handle each tail page individually in migration.
7270 if (PageHuge(page
)) {
7271 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7276 * We can't use page_count without pin a page
7277 * because another CPU can free compound page.
7278 * This check already skips compound tails of THP
7279 * because their page->_refcount is zero at all time.
7281 if (!page_ref_count(page
)) {
7282 if (PageBuddy(page
))
7283 iter
+= (1 << page_order(page
)) - 1;
7288 * The HWPoisoned page may be not in buddy system, and
7289 * page_count() is not 0.
7291 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7294 if (__PageMovable(page
))
7300 * If there are RECLAIMABLE pages, we need to check
7301 * it. But now, memory offline itself doesn't call
7302 * shrink_node_slabs() and it still to be fixed.
7305 * If the page is not RAM, page_count()should be 0.
7306 * we don't need more check. This is an _used_ not-movable page.
7308 * The problematic thing here is PG_reserved pages. PG_reserved
7309 * is set to both of a memory hole page and a _used_ kernel
7318 bool is_pageblock_removable_nolock(struct page
*page
)
7324 * We have to be careful here because we are iterating over memory
7325 * sections which are not zone aware so we might end up outside of
7326 * the zone but still within the section.
7327 * We have to take care about the node as well. If the node is offline
7328 * its NODE_DATA will be NULL - see page_zone.
7330 if (!node_online(page_to_nid(page
)))
7333 zone
= page_zone(page
);
7334 pfn
= page_to_pfn(page
);
7335 if (!zone_spans_pfn(zone
, pfn
))
7338 return !has_unmovable_pages(zone
, page
, 0, true);
7341 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7343 static unsigned long pfn_max_align_down(unsigned long pfn
)
7345 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7346 pageblock_nr_pages
) - 1);
7349 static unsigned long pfn_max_align_up(unsigned long pfn
)
7351 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7352 pageblock_nr_pages
));
7355 /* [start, end) must belong to a single zone. */
7356 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7357 unsigned long start
, unsigned long end
)
7359 /* This function is based on compact_zone() from compaction.c. */
7360 unsigned long nr_reclaimed
;
7361 unsigned long pfn
= start
;
7362 unsigned int tries
= 0;
7367 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7368 if (fatal_signal_pending(current
)) {
7373 if (list_empty(&cc
->migratepages
)) {
7374 cc
->nr_migratepages
= 0;
7375 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7381 } else if (++tries
== 5) {
7382 ret
= ret
< 0 ? ret
: -EBUSY
;
7386 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7388 cc
->nr_migratepages
-= nr_reclaimed
;
7390 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7391 NULL
, 0, cc
->mode
, MR_CMA
);
7394 putback_movable_pages(&cc
->migratepages
);
7401 * alloc_contig_range() -- tries to allocate given range of pages
7402 * @start: start PFN to allocate
7403 * @end: one-past-the-last PFN to allocate
7404 * @migratetype: migratetype of the underlaying pageblocks (either
7405 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7406 * in range must have the same migratetype and it must
7407 * be either of the two.
7408 * @gfp_mask: GFP mask to use during compaction
7410 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7411 * aligned, however it's the caller's responsibility to guarantee that
7412 * we are the only thread that changes migrate type of pageblocks the
7415 * The PFN range must belong to a single zone.
7417 * Returns zero on success or negative error code. On success all
7418 * pages which PFN is in [start, end) are allocated for the caller and
7419 * need to be freed with free_contig_range().
7421 int alloc_contig_range(unsigned long start
, unsigned long end
,
7422 unsigned migratetype
, gfp_t gfp_mask
)
7424 unsigned long outer_start
, outer_end
;
7428 struct compact_control cc
= {
7429 .nr_migratepages
= 0,
7431 .zone
= page_zone(pfn_to_page(start
)),
7432 .mode
= MIGRATE_SYNC
,
7433 .ignore_skip_hint
= true,
7434 .gfp_mask
= memalloc_noio_flags(gfp_mask
),
7436 INIT_LIST_HEAD(&cc
.migratepages
);
7439 * What we do here is we mark all pageblocks in range as
7440 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7441 * have different sizes, and due to the way page allocator
7442 * work, we align the range to biggest of the two pages so
7443 * that page allocator won't try to merge buddies from
7444 * different pageblocks and change MIGRATE_ISOLATE to some
7445 * other migration type.
7447 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7448 * migrate the pages from an unaligned range (ie. pages that
7449 * we are interested in). This will put all the pages in
7450 * range back to page allocator as MIGRATE_ISOLATE.
7452 * When this is done, we take the pages in range from page
7453 * allocator removing them from the buddy system. This way
7454 * page allocator will never consider using them.
7456 * This lets us mark the pageblocks back as
7457 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7458 * aligned range but not in the unaligned, original range are
7459 * put back to page allocator so that buddy can use them.
7462 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7463 pfn_max_align_up(end
), migratetype
,
7469 * In case of -EBUSY, we'd like to know which page causes problem.
7470 * So, just fall through. We will check it in test_pages_isolated().
7472 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7473 if (ret
&& ret
!= -EBUSY
)
7477 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7478 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7479 * more, all pages in [start, end) are free in page allocator.
7480 * What we are going to do is to allocate all pages from
7481 * [start, end) (that is remove them from page allocator).
7483 * The only problem is that pages at the beginning and at the
7484 * end of interesting range may be not aligned with pages that
7485 * page allocator holds, ie. they can be part of higher order
7486 * pages. Because of this, we reserve the bigger range and
7487 * once this is done free the pages we are not interested in.
7489 * We don't have to hold zone->lock here because the pages are
7490 * isolated thus they won't get removed from buddy.
7493 lru_add_drain_all();
7494 drain_all_pages(cc
.zone
);
7497 outer_start
= start
;
7498 while (!PageBuddy(pfn_to_page(outer_start
))) {
7499 if (++order
>= MAX_ORDER
) {
7500 outer_start
= start
;
7503 outer_start
&= ~0UL << order
;
7506 if (outer_start
!= start
) {
7507 order
= page_order(pfn_to_page(outer_start
));
7510 * outer_start page could be small order buddy page and
7511 * it doesn't include start page. Adjust outer_start
7512 * in this case to report failed page properly
7513 * on tracepoint in test_pages_isolated()
7515 if (outer_start
+ (1UL << order
) <= start
)
7516 outer_start
= start
;
7519 /* Make sure the range is really isolated. */
7520 if (test_pages_isolated(outer_start
, end
, false)) {
7521 pr_info("%s: [%lx, %lx) PFNs busy\n",
7522 __func__
, outer_start
, end
);
7527 /* Grab isolated pages from freelists. */
7528 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7534 /* Free head and tail (if any) */
7535 if (start
!= outer_start
)
7536 free_contig_range(outer_start
, start
- outer_start
);
7537 if (end
!= outer_end
)
7538 free_contig_range(end
, outer_end
- end
);
7541 undo_isolate_page_range(pfn_max_align_down(start
),
7542 pfn_max_align_up(end
), migratetype
);
7546 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7548 unsigned int count
= 0;
7550 for (; nr_pages
--; pfn
++) {
7551 struct page
*page
= pfn_to_page(pfn
);
7553 count
+= page_count(page
) != 1;
7556 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7560 #ifdef CONFIG_MEMORY_HOTPLUG
7562 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7563 * page high values need to be recalulated.
7565 void __meminit
zone_pcp_update(struct zone
*zone
)
7568 mutex_lock(&pcp_batch_high_lock
);
7569 for_each_possible_cpu(cpu
)
7570 pageset_set_high_and_batch(zone
,
7571 per_cpu_ptr(zone
->pageset
, cpu
));
7572 mutex_unlock(&pcp_batch_high_lock
);
7576 void zone_pcp_reset(struct zone
*zone
)
7578 unsigned long flags
;
7580 struct per_cpu_pageset
*pset
;
7582 /* avoid races with drain_pages() */
7583 local_irq_save(flags
);
7584 if (zone
->pageset
!= &boot_pageset
) {
7585 for_each_online_cpu(cpu
) {
7586 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7587 drain_zonestat(zone
, pset
);
7589 free_percpu(zone
->pageset
);
7590 zone
->pageset
= &boot_pageset
;
7592 local_irq_restore(flags
);
7595 #ifdef CONFIG_MEMORY_HOTREMOVE
7597 * All pages in the range must be in a single zone and isolated
7598 * before calling this.
7601 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7605 unsigned int order
, i
;
7607 unsigned long flags
;
7608 /* find the first valid pfn */
7609 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7614 zone
= page_zone(pfn_to_page(pfn
));
7615 spin_lock_irqsave(&zone
->lock
, flags
);
7617 while (pfn
< end_pfn
) {
7618 if (!pfn_valid(pfn
)) {
7622 page
= pfn_to_page(pfn
);
7624 * The HWPoisoned page may be not in buddy system, and
7625 * page_count() is not 0.
7627 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7629 SetPageReserved(page
);
7633 BUG_ON(page_count(page
));
7634 BUG_ON(!PageBuddy(page
));
7635 order
= page_order(page
);
7636 #ifdef CONFIG_DEBUG_VM
7637 pr_info("remove from free list %lx %d %lx\n",
7638 pfn
, 1 << order
, end_pfn
);
7640 list_del(&page
->lru
);
7641 rmv_page_order(page
);
7642 zone
->free_area
[order
].nr_free
--;
7643 for (i
= 0; i
< (1 << order
); i
++)
7644 SetPageReserved((page
+i
));
7645 pfn
+= (1 << order
);
7647 spin_unlock_irqrestore(&zone
->lock
, flags
);
7651 bool is_free_buddy_page(struct page
*page
)
7653 struct zone
*zone
= page_zone(page
);
7654 unsigned long pfn
= page_to_pfn(page
);
7655 unsigned long flags
;
7658 spin_lock_irqsave(&zone
->lock
, flags
);
7659 for (order
= 0; order
< MAX_ORDER
; order
++) {
7660 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7662 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7665 spin_unlock_irqrestore(&zone
->lock
, flags
);
7667 return order
< MAX_ORDER
;