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
;
1094 bool isolated_pageblocks
;
1096 spin_lock(&zone
->lock
);
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(&zone
->lock
);
1148 static void free_one_page(struct zone
*zone
,
1149 struct page
*page
, unsigned long pfn
,
1153 unsigned long nr_scanned
;
1154 spin_lock(&zone
->lock
);
1155 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1157 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1159 if (unlikely(has_isolate_pageblock(zone
) ||
1160 is_migrate_isolate(migratetype
))) {
1161 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1163 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1164 spin_unlock(&zone
->lock
);
1167 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1168 unsigned long zone
, int nid
)
1170 set_page_links(page
, zone
, nid
, pfn
);
1171 init_page_count(page
);
1172 page_mapcount_reset(page
);
1173 page_cpupid_reset_last(page
);
1175 INIT_LIST_HEAD(&page
->lru
);
1176 #ifdef WANT_PAGE_VIRTUAL
1177 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1178 if (!is_highmem_idx(zone
))
1179 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1183 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1186 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1189 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1190 static void init_reserved_page(unsigned long pfn
)
1195 if (!early_page_uninitialised(pfn
))
1198 nid
= early_pfn_to_nid(pfn
);
1199 pgdat
= NODE_DATA(nid
);
1201 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1202 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1204 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1207 __init_single_pfn(pfn
, zid
, nid
);
1210 static inline void init_reserved_page(unsigned long pfn
)
1213 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1216 * Initialised pages do not have PageReserved set. This function is
1217 * called for each range allocated by the bootmem allocator and
1218 * marks the pages PageReserved. The remaining valid pages are later
1219 * sent to the buddy page allocator.
1221 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1223 unsigned long start_pfn
= PFN_DOWN(start
);
1224 unsigned long end_pfn
= PFN_UP(end
);
1226 for (; start_pfn
< end_pfn
; start_pfn
++) {
1227 if (pfn_valid(start_pfn
)) {
1228 struct page
*page
= pfn_to_page(start_pfn
);
1230 init_reserved_page(start_pfn
);
1232 /* Avoid false-positive PageTail() */
1233 INIT_LIST_HEAD(&page
->lru
);
1235 SetPageReserved(page
);
1240 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1242 unsigned long flags
;
1244 unsigned long pfn
= page_to_pfn(page
);
1246 if (!free_pages_prepare(page
, order
, true))
1249 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1250 local_irq_save(flags
);
1251 __count_vm_events(PGFREE
, 1 << order
);
1252 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1253 local_irq_restore(flags
);
1256 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1258 unsigned int nr_pages
= 1 << order
;
1259 struct page
*p
= page
;
1263 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1265 __ClearPageReserved(p
);
1266 set_page_count(p
, 0);
1268 __ClearPageReserved(p
);
1269 set_page_count(p
, 0);
1271 page_zone(page
)->managed_pages
+= nr_pages
;
1272 set_page_refcounted(page
);
1273 __free_pages(page
, order
);
1276 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1277 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1279 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1281 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1283 static DEFINE_SPINLOCK(early_pfn_lock
);
1286 spin_lock(&early_pfn_lock
);
1287 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1289 nid
= first_online_node
;
1290 spin_unlock(&early_pfn_lock
);
1296 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1297 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1298 struct mminit_pfnnid_cache
*state
)
1302 nid
= __early_pfn_to_nid(pfn
, state
);
1303 if (nid
>= 0 && nid
!= node
)
1308 /* Only safe to use early in boot when initialisation is single-threaded */
1309 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1311 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1316 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1320 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1321 struct mminit_pfnnid_cache
*state
)
1328 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1331 if (early_page_uninitialised(pfn
))
1333 return __free_pages_boot_core(page
, order
);
1337 * Check that the whole (or subset of) a pageblock given by the interval of
1338 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1339 * with the migration of free compaction scanner. The scanners then need to
1340 * use only pfn_valid_within() check for arches that allow holes within
1343 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1345 * It's possible on some configurations to have a setup like node0 node1 node0
1346 * i.e. it's possible that all pages within a zones range of pages do not
1347 * belong to a single zone. We assume that a border between node0 and node1
1348 * can occur within a single pageblock, but not a node0 node1 node0
1349 * interleaving within a single pageblock. It is therefore sufficient to check
1350 * the first and last page of a pageblock and avoid checking each individual
1351 * page in a pageblock.
1353 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1354 unsigned long end_pfn
, struct zone
*zone
)
1356 struct page
*start_page
;
1357 struct page
*end_page
;
1359 /* end_pfn is one past the range we are checking */
1362 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1365 start_page
= pfn_to_page(start_pfn
);
1367 if (page_zone(start_page
) != zone
)
1370 end_page
= pfn_to_page(end_pfn
);
1372 /* This gives a shorter code than deriving page_zone(end_page) */
1373 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1379 void set_zone_contiguous(struct zone
*zone
)
1381 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1382 unsigned long block_end_pfn
;
1384 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1385 for (; block_start_pfn
< zone_end_pfn(zone
);
1386 block_start_pfn
= block_end_pfn
,
1387 block_end_pfn
+= pageblock_nr_pages
) {
1389 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1391 if (!__pageblock_pfn_to_page(block_start_pfn
,
1392 block_end_pfn
, zone
))
1396 /* We confirm that there is no hole */
1397 zone
->contiguous
= true;
1400 void clear_zone_contiguous(struct zone
*zone
)
1402 zone
->contiguous
= false;
1405 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1406 static void __init
deferred_free_range(struct page
*page
,
1407 unsigned long pfn
, int nr_pages
)
1414 /* Free a large naturally-aligned chunk if possible */
1415 if (nr_pages
== pageblock_nr_pages
&&
1416 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1417 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1418 __free_pages_boot_core(page
, pageblock_order
);
1422 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1423 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1424 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1425 __free_pages_boot_core(page
, 0);
1429 /* Completion tracking for deferred_init_memmap() threads */
1430 static atomic_t pgdat_init_n_undone __initdata
;
1431 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1433 static inline void __init
pgdat_init_report_one_done(void)
1435 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1436 complete(&pgdat_init_all_done_comp
);
1439 /* Initialise remaining memory on a node */
1440 static int __init
deferred_init_memmap(void *data
)
1442 pg_data_t
*pgdat
= data
;
1443 int nid
= pgdat
->node_id
;
1444 struct mminit_pfnnid_cache nid_init_state
= { };
1445 unsigned long start
= jiffies
;
1446 unsigned long nr_pages
= 0;
1447 unsigned long walk_start
, walk_end
;
1450 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1451 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1453 if (first_init_pfn
== ULONG_MAX
) {
1454 pgdat_init_report_one_done();
1458 /* Bind memory initialisation thread to a local node if possible */
1459 if (!cpumask_empty(cpumask
))
1460 set_cpus_allowed_ptr(current
, cpumask
);
1462 /* Sanity check boundaries */
1463 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1464 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1465 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1467 /* Only the highest zone is deferred so find it */
1468 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1469 zone
= pgdat
->node_zones
+ zid
;
1470 if (first_init_pfn
< zone_end_pfn(zone
))
1474 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1475 unsigned long pfn
, end_pfn
;
1476 struct page
*page
= NULL
;
1477 struct page
*free_base_page
= NULL
;
1478 unsigned long free_base_pfn
= 0;
1481 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1482 pfn
= first_init_pfn
;
1483 if (pfn
< walk_start
)
1485 if (pfn
< zone
->zone_start_pfn
)
1486 pfn
= zone
->zone_start_pfn
;
1488 for (; pfn
< end_pfn
; pfn
++) {
1489 if (!pfn_valid_within(pfn
))
1493 * Ensure pfn_valid is checked every
1494 * pageblock_nr_pages for memory holes
1496 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1497 if (!pfn_valid(pfn
)) {
1503 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1508 /* Minimise pfn page lookups and scheduler checks */
1509 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1512 nr_pages
+= nr_to_free
;
1513 deferred_free_range(free_base_page
,
1514 free_base_pfn
, nr_to_free
);
1515 free_base_page
= NULL
;
1516 free_base_pfn
= nr_to_free
= 0;
1518 page
= pfn_to_page(pfn
);
1523 VM_BUG_ON(page_zone(page
) != zone
);
1527 __init_single_page(page
, pfn
, zid
, nid
);
1528 if (!free_base_page
) {
1529 free_base_page
= page
;
1530 free_base_pfn
= pfn
;
1535 /* Where possible, batch up pages for a single free */
1538 /* Free the current block of pages to allocator */
1539 nr_pages
+= nr_to_free
;
1540 deferred_free_range(free_base_page
, free_base_pfn
,
1542 free_base_page
= NULL
;
1543 free_base_pfn
= nr_to_free
= 0;
1545 /* Free the last block of pages to allocator */
1546 nr_pages
+= nr_to_free
;
1547 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1549 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1552 /* Sanity check that the next zone really is unpopulated */
1553 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1555 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1556 jiffies_to_msecs(jiffies
- start
));
1558 pgdat_init_report_one_done();
1561 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1563 void __init
page_alloc_init_late(void)
1567 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1570 /* There will be num_node_state(N_MEMORY) threads */
1571 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1572 for_each_node_state(nid
, N_MEMORY
) {
1573 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1576 /* Block until all are initialised */
1577 wait_for_completion(&pgdat_init_all_done_comp
);
1579 /* Reinit limits that are based on free pages after the kernel is up */
1580 files_maxfiles_init();
1583 for_each_populated_zone(zone
)
1584 set_zone_contiguous(zone
);
1588 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1589 void __init
init_cma_reserved_pageblock(struct page
*page
)
1591 unsigned i
= pageblock_nr_pages
;
1592 struct page
*p
= page
;
1595 __ClearPageReserved(p
);
1596 set_page_count(p
, 0);
1599 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1601 if (pageblock_order
>= MAX_ORDER
) {
1602 i
= pageblock_nr_pages
;
1605 set_page_refcounted(p
);
1606 __free_pages(p
, MAX_ORDER
- 1);
1607 p
+= MAX_ORDER_NR_PAGES
;
1608 } while (i
-= MAX_ORDER_NR_PAGES
);
1610 set_page_refcounted(page
);
1611 __free_pages(page
, pageblock_order
);
1614 adjust_managed_page_count(page
, pageblock_nr_pages
);
1619 * The order of subdivision here is critical for the IO subsystem.
1620 * Please do not alter this order without good reasons and regression
1621 * testing. Specifically, as large blocks of memory are subdivided,
1622 * the order in which smaller blocks are delivered depends on the order
1623 * they're subdivided in this function. This is the primary factor
1624 * influencing the order in which pages are delivered to the IO
1625 * subsystem according to empirical testing, and this is also justified
1626 * by considering the behavior of a buddy system containing a single
1627 * large block of memory acted on by a series of small allocations.
1628 * This behavior is a critical factor in sglist merging's success.
1632 static inline void expand(struct zone
*zone
, struct page
*page
,
1633 int low
, int high
, struct free_area
*area
,
1636 unsigned long size
= 1 << high
;
1638 while (high
> low
) {
1642 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1645 * Mark as guard pages (or page), that will allow to
1646 * merge back to allocator when buddy will be freed.
1647 * Corresponding page table entries will not be touched,
1648 * pages will stay not present in virtual address space
1650 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1653 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1655 set_page_order(&page
[size
], high
);
1659 static void check_new_page_bad(struct page
*page
)
1661 const char *bad_reason
= NULL
;
1662 unsigned long bad_flags
= 0;
1664 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1665 bad_reason
= "nonzero mapcount";
1666 if (unlikely(page
->mapping
!= NULL
))
1667 bad_reason
= "non-NULL mapping";
1668 if (unlikely(page_ref_count(page
) != 0))
1669 bad_reason
= "nonzero _count";
1670 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1671 bad_reason
= "HWPoisoned (hardware-corrupted)";
1672 bad_flags
= __PG_HWPOISON
;
1673 /* Don't complain about hwpoisoned pages */
1674 page_mapcount_reset(page
); /* remove PageBuddy */
1677 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1678 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1679 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1682 if (unlikely(page
->mem_cgroup
))
1683 bad_reason
= "page still charged to cgroup";
1685 bad_page(page
, bad_reason
, bad_flags
);
1689 * This page is about to be returned from the page allocator
1691 static inline int check_new_page(struct page
*page
)
1693 if (likely(page_expected_state(page
,
1694 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1697 check_new_page_bad(page
);
1701 static inline bool free_pages_prezeroed(bool poisoned
)
1703 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1704 page_poisoning_enabled() && poisoned
;
1707 #ifdef CONFIG_DEBUG_VM
1708 static bool check_pcp_refill(struct page
*page
)
1713 static bool check_new_pcp(struct page
*page
)
1715 return check_new_page(page
);
1718 static bool check_pcp_refill(struct page
*page
)
1720 return check_new_page(page
);
1722 static bool check_new_pcp(struct page
*page
)
1726 #endif /* CONFIG_DEBUG_VM */
1728 static bool check_new_pages(struct page
*page
, unsigned int order
)
1731 for (i
= 0; i
< (1 << order
); i
++) {
1732 struct page
*p
= page
+ i
;
1734 if (unlikely(check_new_page(p
)))
1741 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1744 set_page_private(page
, 0);
1745 set_page_refcounted(page
);
1747 arch_alloc_page(page
, order
);
1748 kernel_map_pages(page
, 1 << order
, 1);
1749 kernel_poison_pages(page
, 1 << order
, 1);
1750 kasan_alloc_pages(page
, order
);
1751 set_page_owner(page
, order
, gfp_flags
);
1754 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1755 unsigned int alloc_flags
)
1758 bool poisoned
= true;
1760 for (i
= 0; i
< (1 << order
); i
++) {
1761 struct page
*p
= page
+ i
;
1763 poisoned
&= page_is_poisoned(p
);
1766 post_alloc_hook(page
, order
, gfp_flags
);
1768 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1769 for (i
= 0; i
< (1 << order
); i
++)
1770 clear_highpage(page
+ i
);
1772 if (order
&& (gfp_flags
& __GFP_COMP
))
1773 prep_compound_page(page
, order
);
1776 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1777 * allocate the page. The expectation is that the caller is taking
1778 * steps that will free more memory. The caller should avoid the page
1779 * being used for !PFMEMALLOC purposes.
1781 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1782 set_page_pfmemalloc(page
);
1784 clear_page_pfmemalloc(page
);
1788 * Go through the free lists for the given migratetype and remove
1789 * the smallest available page from the freelists
1792 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1795 unsigned int current_order
;
1796 struct free_area
*area
;
1799 /* Find a page of the appropriate size in the preferred list */
1800 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1801 area
= &(zone
->free_area
[current_order
]);
1802 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1806 list_del(&page
->lru
);
1807 rmv_page_order(page
);
1809 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1810 set_pcppage_migratetype(page
, migratetype
);
1819 * This array describes the order lists are fallen back to when
1820 * the free lists for the desirable migrate type are depleted
1822 static int fallbacks
[MIGRATE_TYPES
][4] = {
1823 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1824 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1825 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1827 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1829 #ifdef CONFIG_MEMORY_ISOLATION
1830 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1835 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1838 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1841 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1842 unsigned int order
) { return NULL
; }
1846 * Move the free pages in a range to the free lists of the requested type.
1847 * Note that start_page and end_pages are not aligned on a pageblock
1848 * boundary. If alignment is required, use move_freepages_block()
1850 int move_freepages(struct zone
*zone
,
1851 struct page
*start_page
, struct page
*end_page
,
1856 int pages_moved
= 0;
1858 #ifndef CONFIG_HOLES_IN_ZONE
1860 * page_zone is not safe to call in this context when
1861 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1862 * anyway as we check zone boundaries in move_freepages_block().
1863 * Remove at a later date when no bug reports exist related to
1864 * grouping pages by mobility
1866 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1869 for (page
= start_page
; page
<= end_page
;) {
1870 if (!pfn_valid_within(page_to_pfn(page
))) {
1875 /* Make sure we are not inadvertently changing nodes */
1876 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1878 if (!PageBuddy(page
)) {
1883 order
= page_order(page
);
1884 list_move(&page
->lru
,
1885 &zone
->free_area
[order
].free_list
[migratetype
]);
1887 pages_moved
+= 1 << order
;
1893 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1896 unsigned long start_pfn
, end_pfn
;
1897 struct page
*start_page
, *end_page
;
1899 start_pfn
= page_to_pfn(page
);
1900 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1901 start_page
= pfn_to_page(start_pfn
);
1902 end_page
= start_page
+ pageblock_nr_pages
- 1;
1903 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1905 /* Do not cross zone boundaries */
1906 if (!zone_spans_pfn(zone
, start_pfn
))
1908 if (!zone_spans_pfn(zone
, end_pfn
))
1911 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1914 static void change_pageblock_range(struct page
*pageblock_page
,
1915 int start_order
, int migratetype
)
1917 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1919 while (nr_pageblocks
--) {
1920 set_pageblock_migratetype(pageblock_page
, migratetype
);
1921 pageblock_page
+= pageblock_nr_pages
;
1926 * When we are falling back to another migratetype during allocation, try to
1927 * steal extra free pages from the same pageblocks to satisfy further
1928 * allocations, instead of polluting multiple pageblocks.
1930 * If we are stealing a relatively large buddy page, it is likely there will
1931 * be more free pages in the pageblock, so try to steal them all. For
1932 * reclaimable and unmovable allocations, we steal regardless of page size,
1933 * as fragmentation caused by those allocations polluting movable pageblocks
1934 * is worse than movable allocations stealing from unmovable and reclaimable
1937 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1940 * Leaving this order check is intended, although there is
1941 * relaxed order check in next check. The reason is that
1942 * we can actually steal whole pageblock if this condition met,
1943 * but, below check doesn't guarantee it and that is just heuristic
1944 * so could be changed anytime.
1946 if (order
>= pageblock_order
)
1949 if (order
>= pageblock_order
/ 2 ||
1950 start_mt
== MIGRATE_RECLAIMABLE
||
1951 start_mt
== MIGRATE_UNMOVABLE
||
1952 page_group_by_mobility_disabled
)
1959 * This function implements actual steal behaviour. If order is large enough,
1960 * we can steal whole pageblock. If not, we first move freepages in this
1961 * pageblock and check whether half of pages are moved or not. If half of
1962 * pages are moved, we can change migratetype of pageblock and permanently
1963 * use it's pages as requested migratetype in the future.
1965 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1968 unsigned int current_order
= page_order(page
);
1971 /* Take ownership for orders >= pageblock_order */
1972 if (current_order
>= pageblock_order
) {
1973 change_pageblock_range(page
, current_order
, start_type
);
1977 pages
= move_freepages_block(zone
, page
, start_type
);
1979 /* Claim the whole block if over half of it is free */
1980 if (pages
>= (1 << (pageblock_order
-1)) ||
1981 page_group_by_mobility_disabled
)
1982 set_pageblock_migratetype(page
, start_type
);
1986 * Check whether there is a suitable fallback freepage with requested order.
1987 * If only_stealable is true, this function returns fallback_mt only if
1988 * we can steal other freepages all together. This would help to reduce
1989 * fragmentation due to mixed migratetype pages in one pageblock.
1991 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1992 int migratetype
, bool only_stealable
, bool *can_steal
)
1997 if (area
->nr_free
== 0)
2002 fallback_mt
= fallbacks
[migratetype
][i
];
2003 if (fallback_mt
== MIGRATE_TYPES
)
2006 if (list_empty(&area
->free_list
[fallback_mt
]))
2009 if (can_steal_fallback(order
, migratetype
))
2012 if (!only_stealable
)
2023 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2024 * there are no empty page blocks that contain a page with a suitable order
2026 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2027 unsigned int alloc_order
)
2030 unsigned long max_managed
, flags
;
2033 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2034 * Check is race-prone but harmless.
2036 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2037 if (zone
->nr_reserved_highatomic
>= max_managed
)
2040 spin_lock_irqsave(&zone
->lock
, flags
);
2042 /* Recheck the nr_reserved_highatomic limit under the lock */
2043 if (zone
->nr_reserved_highatomic
>= max_managed
)
2047 mt
= get_pageblock_migratetype(page
);
2048 if (mt
!= MIGRATE_HIGHATOMIC
&&
2049 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
2050 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2051 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2052 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
2056 spin_unlock_irqrestore(&zone
->lock
, flags
);
2060 * Used when an allocation is about to fail under memory pressure. This
2061 * potentially hurts the reliability of high-order allocations when under
2062 * intense memory pressure but failed atomic allocations should be easier
2063 * to recover from than an OOM.
2065 * If @force is true, try to unreserve a pageblock even though highatomic
2066 * pageblock is exhausted.
2068 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2071 struct zonelist
*zonelist
= ac
->zonelist
;
2072 unsigned long flags
;
2079 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2082 * Preserve at least one pageblock unless memory pressure
2085 if (!force
&& zone
->nr_reserved_highatomic
<=
2089 spin_lock_irqsave(&zone
->lock
, flags
);
2090 for (order
= 0; order
< MAX_ORDER
; order
++) {
2091 struct free_area
*area
= &(zone
->free_area
[order
]);
2093 page
= list_first_entry_or_null(
2094 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2100 * In page freeing path, migratetype change is racy so
2101 * we can counter several free pages in a pageblock
2102 * in this loop althoug we changed the pageblock type
2103 * from highatomic to ac->migratetype. So we should
2104 * adjust the count once.
2106 if (get_pageblock_migratetype(page
) ==
2107 MIGRATE_HIGHATOMIC
) {
2109 * It should never happen but changes to
2110 * locking could inadvertently allow a per-cpu
2111 * drain to add pages to MIGRATE_HIGHATOMIC
2112 * while unreserving so be safe and watch for
2115 zone
->nr_reserved_highatomic
-= min(
2117 zone
->nr_reserved_highatomic
);
2121 * Convert to ac->migratetype and avoid the normal
2122 * pageblock stealing heuristics. Minimally, the caller
2123 * is doing the work and needs the pages. More
2124 * importantly, if the block was always converted to
2125 * MIGRATE_UNMOVABLE or another type then the number
2126 * of pageblocks that cannot be completely freed
2129 set_pageblock_migratetype(page
, ac
->migratetype
);
2130 ret
= move_freepages_block(zone
, page
, ac
->migratetype
);
2132 spin_unlock_irqrestore(&zone
->lock
, flags
);
2136 spin_unlock_irqrestore(&zone
->lock
, flags
);
2142 /* Remove an element from the buddy allocator from the fallback list */
2143 static inline struct page
*
2144 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2146 struct free_area
*area
;
2147 unsigned int current_order
;
2152 /* Find the largest possible block of pages in the other list */
2153 for (current_order
= MAX_ORDER
-1;
2154 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2156 area
= &(zone
->free_area
[current_order
]);
2157 fallback_mt
= find_suitable_fallback(area
, current_order
,
2158 start_migratetype
, false, &can_steal
);
2159 if (fallback_mt
== -1)
2162 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2165 get_pageblock_migratetype(page
) != MIGRATE_HIGHATOMIC
)
2166 steal_suitable_fallback(zone
, page
, start_migratetype
);
2168 /* Remove the page from the freelists */
2170 list_del(&page
->lru
);
2171 rmv_page_order(page
);
2173 expand(zone
, page
, order
, current_order
, area
,
2176 * The pcppage_migratetype may differ from pageblock's
2177 * migratetype depending on the decisions in
2178 * find_suitable_fallback(). This is OK as long as it does not
2179 * differ for MIGRATE_CMA pageblocks. Those can be used as
2180 * fallback only via special __rmqueue_cma_fallback() function
2182 set_pcppage_migratetype(page
, start_migratetype
);
2184 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2185 start_migratetype
, fallback_mt
);
2194 * Do the hard work of removing an element from the buddy allocator.
2195 * Call me with the zone->lock already held.
2197 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2202 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2203 if (unlikely(!page
)) {
2204 if (migratetype
== MIGRATE_MOVABLE
)
2205 page
= __rmqueue_cma_fallback(zone
, order
);
2208 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2211 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2216 * Obtain a specified number of elements from the buddy allocator, all under
2217 * a single hold of the lock, for efficiency. Add them to the supplied list.
2218 * Returns the number of new pages which were placed at *list.
2220 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2221 unsigned long count
, struct list_head
*list
,
2222 int migratetype
, bool cold
)
2226 spin_lock(&zone
->lock
);
2227 for (i
= 0; i
< count
; ++i
) {
2228 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2229 if (unlikely(page
== NULL
))
2232 if (unlikely(check_pcp_refill(page
)))
2236 * Split buddy pages returned by expand() are received here
2237 * in physical page order. The page is added to the callers and
2238 * list and the list head then moves forward. From the callers
2239 * perspective, the linked list is ordered by page number in
2240 * some conditions. This is useful for IO devices that can
2241 * merge IO requests if the physical pages are ordered
2245 list_add(&page
->lru
, list
);
2247 list_add_tail(&page
->lru
, list
);
2250 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2251 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2256 * i pages were removed from the buddy list even if some leak due
2257 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2258 * on i. Do not confuse with 'alloced' which is the number of
2259 * pages added to the pcp list.
2261 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2262 spin_unlock(&zone
->lock
);
2268 * Called from the vmstat counter updater to drain pagesets of this
2269 * currently executing processor on remote nodes after they have
2272 * Note that this function must be called with the thread pinned to
2273 * a single processor.
2275 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2277 unsigned long flags
;
2278 int to_drain
, batch
;
2280 local_irq_save(flags
);
2281 batch
= READ_ONCE(pcp
->batch
);
2282 to_drain
= min(pcp
->count
, batch
);
2284 free_pcppages_bulk(zone
, to_drain
, pcp
);
2285 pcp
->count
-= to_drain
;
2287 local_irq_restore(flags
);
2292 * Drain pcplists of the indicated processor and zone.
2294 * The processor must either be the current processor and the
2295 * thread pinned to the current processor or a processor that
2298 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2300 unsigned long flags
;
2301 struct per_cpu_pageset
*pset
;
2302 struct per_cpu_pages
*pcp
;
2304 local_irq_save(flags
);
2305 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2309 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2312 local_irq_restore(flags
);
2316 * Drain pcplists of all zones on the indicated processor.
2318 * The processor must either be the current processor and the
2319 * thread pinned to the current processor or a processor that
2322 static void drain_pages(unsigned int cpu
)
2326 for_each_populated_zone(zone
) {
2327 drain_pages_zone(cpu
, zone
);
2332 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2334 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2335 * the single zone's pages.
2337 void drain_local_pages(struct zone
*zone
)
2339 int cpu
= smp_processor_id();
2342 drain_pages_zone(cpu
, zone
);
2347 static void drain_local_pages_wq(struct work_struct
*work
)
2350 * drain_all_pages doesn't use proper cpu hotplug protection so
2351 * we can race with cpu offline when the WQ can move this from
2352 * a cpu pinned worker to an unbound one. We can operate on a different
2353 * cpu which is allright but we also have to make sure to not move to
2357 drain_local_pages(NULL
);
2362 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2364 * When zone parameter is non-NULL, spill just the single zone's pages.
2366 * Note that this can be extremely slow as the draining happens in a workqueue.
2368 void drain_all_pages(struct zone
*zone
)
2373 * Allocate in the BSS so we wont require allocation in
2374 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2376 static cpumask_t cpus_with_pcps
;
2379 * Make sure nobody triggers this path before mm_percpu_wq is fully
2382 if (WARN_ON_ONCE(!mm_percpu_wq
))
2385 /* Workqueues cannot recurse */
2386 if (current
->flags
& PF_WQ_WORKER
)
2390 * Do not drain if one is already in progress unless it's specific to
2391 * a zone. Such callers are primarily CMA and memory hotplug and need
2392 * the drain to be complete when the call returns.
2394 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2397 mutex_lock(&pcpu_drain_mutex
);
2401 * We don't care about racing with CPU hotplug event
2402 * as offline notification will cause the notified
2403 * cpu to drain that CPU pcps and on_each_cpu_mask
2404 * disables preemption as part of its processing
2406 for_each_online_cpu(cpu
) {
2407 struct per_cpu_pageset
*pcp
;
2409 bool has_pcps
= false;
2412 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2416 for_each_populated_zone(z
) {
2417 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2418 if (pcp
->pcp
.count
) {
2426 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2428 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2431 for_each_cpu(cpu
, &cpus_with_pcps
) {
2432 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2433 INIT_WORK(work
, drain_local_pages_wq
);
2434 queue_work_on(cpu
, mm_percpu_wq
, work
);
2436 for_each_cpu(cpu
, &cpus_with_pcps
)
2437 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2439 mutex_unlock(&pcpu_drain_mutex
);
2442 #ifdef CONFIG_HIBERNATION
2444 void mark_free_pages(struct zone
*zone
)
2446 unsigned long pfn
, max_zone_pfn
;
2447 unsigned long flags
;
2448 unsigned int order
, t
;
2451 if (zone_is_empty(zone
))
2454 spin_lock_irqsave(&zone
->lock
, flags
);
2456 max_zone_pfn
= zone_end_pfn(zone
);
2457 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2458 if (pfn_valid(pfn
)) {
2459 page
= pfn_to_page(pfn
);
2461 if (page_zone(page
) != zone
)
2464 if (!swsusp_page_is_forbidden(page
))
2465 swsusp_unset_page_free(page
);
2468 for_each_migratetype_order(order
, t
) {
2469 list_for_each_entry(page
,
2470 &zone
->free_area
[order
].free_list
[t
], lru
) {
2473 pfn
= page_to_pfn(page
);
2474 for (i
= 0; i
< (1UL << order
); i
++)
2475 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2478 spin_unlock_irqrestore(&zone
->lock
, flags
);
2480 #endif /* CONFIG_PM */
2483 * Free a 0-order page
2484 * cold == true ? free a cold page : free a hot page
2486 void free_hot_cold_page(struct page
*page
, bool cold
)
2488 struct zone
*zone
= page_zone(page
);
2489 struct per_cpu_pages
*pcp
;
2490 unsigned long flags
;
2491 unsigned long pfn
= page_to_pfn(page
);
2494 if (!free_pcp_prepare(page
))
2497 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2498 set_pcppage_migratetype(page
, migratetype
);
2499 local_irq_save(flags
);
2500 __count_vm_event(PGFREE
);
2503 * We only track unmovable, reclaimable and movable on pcp lists.
2504 * Free ISOLATE pages back to the allocator because they are being
2505 * offlined but treat RESERVE as movable pages so we can get those
2506 * areas back if necessary. Otherwise, we may have to free
2507 * excessively into the page allocator
2509 if (migratetype
>= MIGRATE_PCPTYPES
) {
2510 if (unlikely(is_migrate_isolate(migratetype
))) {
2511 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2514 migratetype
= MIGRATE_MOVABLE
;
2517 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2519 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2521 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2523 if (pcp
->count
>= pcp
->high
) {
2524 unsigned long batch
= READ_ONCE(pcp
->batch
);
2525 free_pcppages_bulk(zone
, batch
, pcp
);
2526 pcp
->count
-= batch
;
2530 local_irq_restore(flags
);
2534 * Free a list of 0-order pages
2536 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2538 struct page
*page
, *next
;
2540 list_for_each_entry_safe(page
, next
, list
, lru
) {
2541 trace_mm_page_free_batched(page
, cold
);
2542 free_hot_cold_page(page
, cold
);
2547 * split_page takes a non-compound higher-order page, and splits it into
2548 * n (1<<order) sub-pages: page[0..n]
2549 * Each sub-page must be freed individually.
2551 * Note: this is probably too low level an operation for use in drivers.
2552 * Please consult with lkml before using this in your driver.
2554 void split_page(struct page
*page
, unsigned int order
)
2558 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2559 VM_BUG_ON_PAGE(!page_count(page
), page
);
2561 #ifdef CONFIG_KMEMCHECK
2563 * Split shadow pages too, because free(page[0]) would
2564 * otherwise free the whole shadow.
2566 if (kmemcheck_page_is_tracked(page
))
2567 split_page(virt_to_page(page
[0].shadow
), order
);
2570 for (i
= 1; i
< (1 << order
); i
++)
2571 set_page_refcounted(page
+ i
);
2572 split_page_owner(page
, order
);
2574 EXPORT_SYMBOL_GPL(split_page
);
2576 int __isolate_free_page(struct page
*page
, unsigned int order
)
2578 unsigned long watermark
;
2582 BUG_ON(!PageBuddy(page
));
2584 zone
= page_zone(page
);
2585 mt
= get_pageblock_migratetype(page
);
2587 if (!is_migrate_isolate(mt
)) {
2589 * Obey watermarks as if the page was being allocated. We can
2590 * emulate a high-order watermark check with a raised order-0
2591 * watermark, because we already know our high-order page
2594 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2595 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2598 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2601 /* Remove page from free list */
2602 list_del(&page
->lru
);
2603 zone
->free_area
[order
].nr_free
--;
2604 rmv_page_order(page
);
2607 * Set the pageblock if the isolated page is at least half of a
2610 if (order
>= pageblock_order
- 1) {
2611 struct page
*endpage
= page
+ (1 << order
) - 1;
2612 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2613 int mt
= get_pageblock_migratetype(page
);
2614 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2615 && mt
!= MIGRATE_HIGHATOMIC
)
2616 set_pageblock_migratetype(page
,
2622 return 1UL << order
;
2626 * Update NUMA hit/miss statistics
2628 * Must be called with interrupts disabled.
2630 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2633 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2635 if (z
->node
!= numa_node_id())
2636 local_stat
= NUMA_OTHER
;
2638 if (z
->node
== preferred_zone
->node
)
2639 __inc_zone_state(z
, NUMA_HIT
);
2641 __inc_zone_state(z
, NUMA_MISS
);
2642 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2644 __inc_zone_state(z
, local_stat
);
2648 /* Remove page from the per-cpu list, caller must protect the list */
2649 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2650 bool cold
, struct per_cpu_pages
*pcp
,
2651 struct list_head
*list
)
2656 if (list_empty(list
)) {
2657 pcp
->count
+= rmqueue_bulk(zone
, 0,
2660 if (unlikely(list_empty(list
)))
2665 page
= list_last_entry(list
, struct page
, lru
);
2667 page
= list_first_entry(list
, struct page
, lru
);
2669 list_del(&page
->lru
);
2671 } while (check_new_pcp(page
));
2676 /* Lock and remove page from the per-cpu list */
2677 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2678 struct zone
*zone
, unsigned int order
,
2679 gfp_t gfp_flags
, int migratetype
)
2681 struct per_cpu_pages
*pcp
;
2682 struct list_head
*list
;
2683 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2685 unsigned long flags
;
2687 local_irq_save(flags
);
2688 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2689 list
= &pcp
->lists
[migratetype
];
2690 page
= __rmqueue_pcplist(zone
, migratetype
, cold
, pcp
, list
);
2692 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2693 zone_statistics(preferred_zone
, zone
);
2695 local_irq_restore(flags
);
2700 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2703 struct page
*rmqueue(struct zone
*preferred_zone
,
2704 struct zone
*zone
, unsigned int order
,
2705 gfp_t gfp_flags
, unsigned int alloc_flags
,
2708 unsigned long flags
;
2711 if (likely(order
== 0)) {
2712 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
2713 gfp_flags
, migratetype
);
2718 * We most definitely don't want callers attempting to
2719 * allocate greater than order-1 page units with __GFP_NOFAIL.
2721 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2722 spin_lock_irqsave(&zone
->lock
, flags
);
2726 if (alloc_flags
& ALLOC_HARDER
) {
2727 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2729 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2732 page
= __rmqueue(zone
, order
, migratetype
);
2733 } while (page
&& check_new_pages(page
, order
));
2734 spin_unlock(&zone
->lock
);
2737 __mod_zone_freepage_state(zone
, -(1 << order
),
2738 get_pcppage_migratetype(page
));
2740 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2741 zone_statistics(preferred_zone
, zone
);
2742 local_irq_restore(flags
);
2745 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
2749 local_irq_restore(flags
);
2753 #ifdef CONFIG_FAIL_PAGE_ALLOC
2756 struct fault_attr attr
;
2758 bool ignore_gfp_highmem
;
2759 bool ignore_gfp_reclaim
;
2761 } fail_page_alloc
= {
2762 .attr
= FAULT_ATTR_INITIALIZER
,
2763 .ignore_gfp_reclaim
= true,
2764 .ignore_gfp_highmem
= true,
2768 static int __init
setup_fail_page_alloc(char *str
)
2770 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2772 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2774 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2776 if (order
< fail_page_alloc
.min_order
)
2778 if (gfp_mask
& __GFP_NOFAIL
)
2780 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2782 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2783 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2786 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2789 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2791 static int __init
fail_page_alloc_debugfs(void)
2793 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2796 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2797 &fail_page_alloc
.attr
);
2799 return PTR_ERR(dir
);
2801 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2802 &fail_page_alloc
.ignore_gfp_reclaim
))
2804 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2805 &fail_page_alloc
.ignore_gfp_highmem
))
2807 if (!debugfs_create_u32("min-order", mode
, dir
,
2808 &fail_page_alloc
.min_order
))
2813 debugfs_remove_recursive(dir
);
2818 late_initcall(fail_page_alloc_debugfs
);
2820 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2822 #else /* CONFIG_FAIL_PAGE_ALLOC */
2824 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2829 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2832 * Return true if free base pages are above 'mark'. For high-order checks it
2833 * will return true of the order-0 watermark is reached and there is at least
2834 * one free page of a suitable size. Checking now avoids taking the zone lock
2835 * to check in the allocation paths if no pages are free.
2837 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2838 int classzone_idx
, unsigned int alloc_flags
,
2843 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2845 /* free_pages may go negative - that's OK */
2846 free_pages
-= (1 << order
) - 1;
2848 if (alloc_flags
& ALLOC_HIGH
)
2852 * If the caller does not have rights to ALLOC_HARDER then subtract
2853 * the high-atomic reserves. This will over-estimate the size of the
2854 * atomic reserve but it avoids a search.
2856 if (likely(!alloc_harder
))
2857 free_pages
-= z
->nr_reserved_highatomic
;
2862 /* If allocation can't use CMA areas don't use free CMA pages */
2863 if (!(alloc_flags
& ALLOC_CMA
))
2864 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2868 * Check watermarks for an order-0 allocation request. If these
2869 * are not met, then a high-order request also cannot go ahead
2870 * even if a suitable page happened to be free.
2872 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2875 /* If this is an order-0 request then the watermark is fine */
2879 /* For a high-order request, check at least one suitable page is free */
2880 for (o
= order
; o
< MAX_ORDER
; o
++) {
2881 struct free_area
*area
= &z
->free_area
[o
];
2890 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2891 if (!list_empty(&area
->free_list
[mt
]))
2896 if ((alloc_flags
& ALLOC_CMA
) &&
2897 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2905 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2906 int classzone_idx
, unsigned int alloc_flags
)
2908 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2909 zone_page_state(z
, NR_FREE_PAGES
));
2912 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2913 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2915 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2919 /* If allocation can't use CMA areas don't use free CMA pages */
2920 if (!(alloc_flags
& ALLOC_CMA
))
2921 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2925 * Fast check for order-0 only. If this fails then the reserves
2926 * need to be calculated. There is a corner case where the check
2927 * passes but only the high-order atomic reserve are free. If
2928 * the caller is !atomic then it'll uselessly search the free
2929 * list. That corner case is then slower but it is harmless.
2931 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2934 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2938 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2939 unsigned long mark
, int classzone_idx
)
2941 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2943 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2944 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2946 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2951 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2953 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
2956 #else /* CONFIG_NUMA */
2957 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2961 #endif /* CONFIG_NUMA */
2964 * get_page_from_freelist goes through the zonelist trying to allocate
2967 static struct page
*
2968 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2969 const struct alloc_context
*ac
)
2971 struct zoneref
*z
= ac
->preferred_zoneref
;
2973 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
2976 * Scan zonelist, looking for a zone with enough free.
2977 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2979 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2984 if (cpusets_enabled() &&
2985 (alloc_flags
& ALLOC_CPUSET
) &&
2986 !__cpuset_zone_allowed(zone
, gfp_mask
))
2989 * When allocating a page cache page for writing, we
2990 * want to get it from a node that is within its dirty
2991 * limit, such that no single node holds more than its
2992 * proportional share of globally allowed dirty pages.
2993 * The dirty limits take into account the node's
2994 * lowmem reserves and high watermark so that kswapd
2995 * should be able to balance it without having to
2996 * write pages from its LRU list.
2998 * XXX: For now, allow allocations to potentially
2999 * exceed the per-node dirty limit in the slowpath
3000 * (spread_dirty_pages unset) before going into reclaim,
3001 * which is important when on a NUMA setup the allowed
3002 * nodes are together not big enough to reach the
3003 * global limit. The proper fix for these situations
3004 * will require awareness of nodes in the
3005 * dirty-throttling and the flusher threads.
3007 if (ac
->spread_dirty_pages
) {
3008 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3011 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3012 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3017 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
3018 if (!zone_watermark_fast(zone
, order
, mark
,
3019 ac_classzone_idx(ac
), alloc_flags
)) {
3022 /* Checked here to keep the fast path fast */
3023 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3024 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3027 if (node_reclaim_mode
== 0 ||
3028 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3031 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3033 case NODE_RECLAIM_NOSCAN
:
3036 case NODE_RECLAIM_FULL
:
3037 /* scanned but unreclaimable */
3040 /* did we reclaim enough */
3041 if (zone_watermark_ok(zone
, order
, mark
,
3042 ac_classzone_idx(ac
), alloc_flags
))
3050 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3051 gfp_mask
, alloc_flags
, ac
->migratetype
);
3053 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3056 * If this is a high-order atomic allocation then check
3057 * if the pageblock should be reserved for the future
3059 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3060 reserve_highatomic_pageblock(page
, zone
, order
);
3070 * Large machines with many possible nodes should not always dump per-node
3071 * meminfo in irq context.
3073 static inline bool should_suppress_show_mem(void)
3078 ret
= in_interrupt();
3083 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3085 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3086 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3088 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs
))
3092 * This documents exceptions given to allocations in certain
3093 * contexts that are allowed to allocate outside current's set
3096 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3097 if (test_thread_flag(TIF_MEMDIE
) ||
3098 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3099 filter
&= ~SHOW_MEM_FILTER_NODES
;
3100 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3101 filter
&= ~SHOW_MEM_FILTER_NODES
;
3103 show_mem(filter
, nodemask
);
3106 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3108 struct va_format vaf
;
3110 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3111 DEFAULT_RATELIMIT_BURST
);
3113 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
3114 debug_guardpage_minorder() > 0)
3117 pr_warn("%s: ", current
->comm
);
3119 va_start(args
, fmt
);
3122 pr_cont("%pV", &vaf
);
3125 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask
, &gfp_mask
);
3127 pr_cont("%*pbl\n", nodemask_pr_args(nodemask
));
3129 pr_cont("(null)\n");
3131 cpuset_print_current_mems_allowed();
3134 warn_alloc_show_mem(gfp_mask
, nodemask
);
3137 static inline struct page
*
3138 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3139 unsigned int alloc_flags
,
3140 const struct alloc_context
*ac
)
3144 page
= get_page_from_freelist(gfp_mask
, order
,
3145 alloc_flags
|ALLOC_CPUSET
, ac
);
3147 * fallback to ignore cpuset restriction if our nodes
3151 page
= get_page_from_freelist(gfp_mask
, order
,
3157 static inline struct page
*
3158 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3159 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3161 struct oom_control oc
= {
3162 .zonelist
= ac
->zonelist
,
3163 .nodemask
= ac
->nodemask
,
3165 .gfp_mask
= gfp_mask
,
3170 *did_some_progress
= 0;
3173 * Acquire the oom lock. If that fails, somebody else is
3174 * making progress for us.
3176 if (!mutex_trylock(&oom_lock
)) {
3177 *did_some_progress
= 1;
3178 schedule_timeout_uninterruptible(1);
3183 * Go through the zonelist yet one more time, keep very high watermark
3184 * here, this is only to catch a parallel oom killing, we must fail if
3185 * we're still under heavy pressure.
3187 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3188 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3192 /* Coredumps can quickly deplete all memory reserves */
3193 if (current
->flags
& PF_DUMPCORE
)
3195 /* The OOM killer will not help higher order allocs */
3196 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3198 /* The OOM killer does not needlessly kill tasks for lowmem */
3199 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3201 if (pm_suspended_storage())
3204 * XXX: GFP_NOFS allocations should rather fail than rely on
3205 * other request to make a forward progress.
3206 * We are in an unfortunate situation where out_of_memory cannot
3207 * do much for this context but let's try it to at least get
3208 * access to memory reserved if the current task is killed (see
3209 * out_of_memory). Once filesystems are ready to handle allocation
3210 * failures more gracefully we should just bail out here.
3213 /* The OOM killer may not free memory on a specific node */
3214 if (gfp_mask
& __GFP_THISNODE
)
3217 /* Exhausted what can be done so it's blamo time */
3218 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3219 *did_some_progress
= 1;
3222 * Help non-failing allocations by giving them access to memory
3225 if (gfp_mask
& __GFP_NOFAIL
)
3226 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3227 ALLOC_NO_WATERMARKS
, ac
);
3230 mutex_unlock(&oom_lock
);
3235 * Maximum number of compaction retries wit a progress before OOM
3236 * killer is consider as the only way to move forward.
3238 #define MAX_COMPACT_RETRIES 16
3240 #ifdef CONFIG_COMPACTION
3241 /* Try memory compaction for high-order allocations before reclaim */
3242 static struct page
*
3243 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3244 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3245 enum compact_priority prio
, enum compact_result
*compact_result
)
3252 current
->flags
|= PF_MEMALLOC
;
3253 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3255 current
->flags
&= ~PF_MEMALLOC
;
3257 if (*compact_result
<= COMPACT_INACTIVE
)
3261 * At least in one zone compaction wasn't deferred or skipped, so let's
3262 * count a compaction stall
3264 count_vm_event(COMPACTSTALL
);
3266 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3269 struct zone
*zone
= page_zone(page
);
3271 zone
->compact_blockskip_flush
= false;
3272 compaction_defer_reset(zone
, order
, true);
3273 count_vm_event(COMPACTSUCCESS
);
3278 * It's bad if compaction run occurs and fails. The most likely reason
3279 * is that pages exist, but not enough to satisfy watermarks.
3281 count_vm_event(COMPACTFAIL
);
3289 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3290 enum compact_result compact_result
,
3291 enum compact_priority
*compact_priority
,
3292 int *compaction_retries
)
3294 int max_retries
= MAX_COMPACT_RETRIES
;
3297 int retries
= *compaction_retries
;
3298 enum compact_priority priority
= *compact_priority
;
3303 if (compaction_made_progress(compact_result
))
3304 (*compaction_retries
)++;
3307 * compaction considers all the zone as desperately out of memory
3308 * so it doesn't really make much sense to retry except when the
3309 * failure could be caused by insufficient priority
3311 if (compaction_failed(compact_result
))
3312 goto check_priority
;
3315 * make sure the compaction wasn't deferred or didn't bail out early
3316 * due to locks contention before we declare that we should give up.
3317 * But do not retry if the given zonelist is not suitable for
3320 if (compaction_withdrawn(compact_result
)) {
3321 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3326 * !costly requests are much more important than __GFP_REPEAT
3327 * costly ones because they are de facto nofail and invoke OOM
3328 * killer to move on while costly can fail and users are ready
3329 * to cope with that. 1/4 retries is rather arbitrary but we
3330 * would need much more detailed feedback from compaction to
3331 * make a better decision.
3333 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3335 if (*compaction_retries
<= max_retries
) {
3341 * Make sure there are attempts at the highest priority if we exhausted
3342 * all retries or failed at the lower priorities.
3345 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3346 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3348 if (*compact_priority
> min_priority
) {
3349 (*compact_priority
)--;
3350 *compaction_retries
= 0;
3354 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3358 static inline struct page
*
3359 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3360 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3361 enum compact_priority prio
, enum compact_result
*compact_result
)
3363 *compact_result
= COMPACT_SKIPPED
;
3368 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3369 enum compact_result compact_result
,
3370 enum compact_priority
*compact_priority
,
3371 int *compaction_retries
)
3376 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3380 * There are setups with compaction disabled which would prefer to loop
3381 * inside the allocator rather than hit the oom killer prematurely.
3382 * Let's give them a good hope and keep retrying while the order-0
3383 * watermarks are OK.
3385 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3387 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3388 ac_classzone_idx(ac
), alloc_flags
))
3393 #endif /* CONFIG_COMPACTION */
3395 /* Perform direct synchronous page reclaim */
3397 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3398 const struct alloc_context
*ac
)
3400 struct reclaim_state reclaim_state
;
3405 /* We now go into synchronous reclaim */
3406 cpuset_memory_pressure_bump();
3407 current
->flags
|= PF_MEMALLOC
;
3408 lockdep_set_current_reclaim_state(gfp_mask
);
3409 reclaim_state
.reclaimed_slab
= 0;
3410 current
->reclaim_state
= &reclaim_state
;
3412 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3415 current
->reclaim_state
= NULL
;
3416 lockdep_clear_current_reclaim_state();
3417 current
->flags
&= ~PF_MEMALLOC
;
3424 /* The really slow allocator path where we enter direct reclaim */
3425 static inline struct page
*
3426 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3427 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3428 unsigned long *did_some_progress
)
3430 struct page
*page
= NULL
;
3431 bool drained
= false;
3433 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3434 if (unlikely(!(*did_some_progress
)))
3438 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3441 * If an allocation failed after direct reclaim, it could be because
3442 * pages are pinned on the per-cpu lists or in high alloc reserves.
3443 * Shrink them them and try again
3445 if (!page
&& !drained
) {
3446 unreserve_highatomic_pageblock(ac
, false);
3447 drain_all_pages(NULL
);
3455 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3459 pg_data_t
*last_pgdat
= NULL
;
3461 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3462 ac
->high_zoneidx
, ac
->nodemask
) {
3463 if (last_pgdat
!= zone
->zone_pgdat
)
3464 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3465 last_pgdat
= zone
->zone_pgdat
;
3469 static inline unsigned int
3470 gfp_to_alloc_flags(gfp_t gfp_mask
)
3472 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3474 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3475 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3478 * The caller may dip into page reserves a bit more if the caller
3479 * cannot run direct reclaim, or if the caller has realtime scheduling
3480 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3481 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3483 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3485 if (gfp_mask
& __GFP_ATOMIC
) {
3487 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3488 * if it can't schedule.
3490 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3491 alloc_flags
|= ALLOC_HARDER
;
3493 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3494 * comment for __cpuset_node_allowed().
3496 alloc_flags
&= ~ALLOC_CPUSET
;
3497 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3498 alloc_flags
|= ALLOC_HARDER
;
3501 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3502 alloc_flags
|= ALLOC_CMA
;
3507 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3509 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3512 if (gfp_mask
& __GFP_MEMALLOC
)
3514 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3516 if (!in_interrupt() &&
3517 ((current
->flags
& PF_MEMALLOC
) ||
3518 unlikely(test_thread_flag(TIF_MEMDIE
))))
3525 * Maximum number of reclaim retries without any progress before OOM killer
3526 * is consider as the only way to move forward.
3528 #define MAX_RECLAIM_RETRIES 16
3531 * Checks whether it makes sense to retry the reclaim to make a forward progress
3532 * for the given allocation request.
3533 * The reclaim feedback represented by did_some_progress (any progress during
3534 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3535 * any progress in a row) is considered as well as the reclaimable pages on the
3536 * applicable zone list (with a backoff mechanism which is a function of
3537 * no_progress_loops).
3539 * Returns true if a retry is viable or false to enter the oom path.
3542 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3543 struct alloc_context
*ac
, int alloc_flags
,
3544 bool did_some_progress
, int *no_progress_loops
)
3550 * Costly allocations might have made a progress but this doesn't mean
3551 * their order will become available due to high fragmentation so
3552 * always increment the no progress counter for them
3554 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3555 *no_progress_loops
= 0;
3557 (*no_progress_loops
)++;
3560 * Make sure we converge to OOM if we cannot make any progress
3561 * several times in the row.
3563 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3564 /* Before OOM, exhaust highatomic_reserve */
3565 return unreserve_highatomic_pageblock(ac
, true);
3569 * Keep reclaiming pages while there is a chance this will lead
3570 * somewhere. If none of the target zones can satisfy our allocation
3571 * request even if all reclaimable pages are considered then we are
3572 * screwed and have to go OOM.
3574 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3576 unsigned long available
;
3577 unsigned long reclaimable
;
3578 unsigned long min_wmark
= min_wmark_pages(zone
);
3581 available
= reclaimable
= zone_reclaimable_pages(zone
);
3582 available
-= DIV_ROUND_UP((*no_progress_loops
) * available
,
3583 MAX_RECLAIM_RETRIES
);
3584 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3587 * Would the allocation succeed if we reclaimed the whole
3590 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3591 ac_classzone_idx(ac
), alloc_flags
, available
);
3592 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3593 available
, min_wmark
, *no_progress_loops
, wmark
);
3596 * If we didn't make any progress and have a lot of
3597 * dirty + writeback pages then we should wait for
3598 * an IO to complete to slow down the reclaim and
3599 * prevent from pre mature OOM
3601 if (!did_some_progress
) {
3602 unsigned long write_pending
;
3604 write_pending
= zone_page_state_snapshot(zone
,
3605 NR_ZONE_WRITE_PENDING
);
3607 if (2 * write_pending
> reclaimable
) {
3608 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3614 * Memory allocation/reclaim might be called from a WQ
3615 * context and the current implementation of the WQ
3616 * concurrency control doesn't recognize that
3617 * a particular WQ is congested if the worker thread is
3618 * looping without ever sleeping. Therefore we have to
3619 * do a short sleep here rather than calling
3622 if (current
->flags
& PF_WQ_WORKER
)
3623 schedule_timeout_uninterruptible(1);
3634 static inline struct page
*
3635 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3636 struct alloc_context
*ac
)
3638 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3639 struct page
*page
= NULL
;
3640 unsigned int alloc_flags
;
3641 unsigned long did_some_progress
;
3642 enum compact_priority compact_priority
;
3643 enum compact_result compact_result
;
3644 int compaction_retries
;
3645 int no_progress_loops
;
3646 unsigned long alloc_start
= jiffies
;
3647 unsigned int stall_timeout
= 10 * HZ
;
3648 unsigned int cpuset_mems_cookie
;
3651 * In the slowpath, we sanity check order to avoid ever trying to
3652 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3653 * be using allocators in order of preference for an area that is
3656 if (order
>= MAX_ORDER
) {
3657 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3662 * We also sanity check to catch abuse of atomic reserves being used by
3663 * callers that are not in atomic context.
3665 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3666 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3667 gfp_mask
&= ~__GFP_ATOMIC
;
3670 compaction_retries
= 0;
3671 no_progress_loops
= 0;
3672 compact_priority
= DEF_COMPACT_PRIORITY
;
3673 cpuset_mems_cookie
= read_mems_allowed_begin();
3676 * The fast path uses conservative alloc_flags to succeed only until
3677 * kswapd needs to be woken up, and to avoid the cost of setting up
3678 * alloc_flags precisely. So we do that now.
3680 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3683 * We need to recalculate the starting point for the zonelist iterator
3684 * because we might have used different nodemask in the fast path, or
3685 * there was a cpuset modification and we are retrying - otherwise we
3686 * could end up iterating over non-eligible zones endlessly.
3688 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3689 ac
->high_zoneidx
, ac
->nodemask
);
3690 if (!ac
->preferred_zoneref
->zone
)
3693 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3694 wake_all_kswapds(order
, ac
);
3697 * The adjusted alloc_flags might result in immediate success, so try
3700 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3705 * For costly allocations, try direct compaction first, as it's likely
3706 * that we have enough base pages and don't need to reclaim. Don't try
3707 * that for allocations that are allowed to ignore watermarks, as the
3708 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3710 if (can_direct_reclaim
&& order
> PAGE_ALLOC_COSTLY_ORDER
&&
3711 !gfp_pfmemalloc_allowed(gfp_mask
)) {
3712 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3714 INIT_COMPACT_PRIORITY
,
3720 * Checks for costly allocations with __GFP_NORETRY, which
3721 * includes THP page fault allocations
3723 if (gfp_mask
& __GFP_NORETRY
) {
3725 * If compaction is deferred for high-order allocations,
3726 * it is because sync compaction recently failed. If
3727 * this is the case and the caller requested a THP
3728 * allocation, we do not want to heavily disrupt the
3729 * system, so we fail the allocation instead of entering
3732 if (compact_result
== COMPACT_DEFERRED
)
3736 * Looks like reclaim/compaction is worth trying, but
3737 * sync compaction could be very expensive, so keep
3738 * using async compaction.
3740 compact_priority
= INIT_COMPACT_PRIORITY
;
3745 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3746 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3747 wake_all_kswapds(order
, ac
);
3749 if (gfp_pfmemalloc_allowed(gfp_mask
))
3750 alloc_flags
= ALLOC_NO_WATERMARKS
;
3753 * Reset the zonelist iterators if memory policies can be ignored.
3754 * These allocations are high priority and system rather than user
3757 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3758 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3759 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3760 ac
->high_zoneidx
, ac
->nodemask
);
3763 /* Attempt with potentially adjusted zonelist and alloc_flags */
3764 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3768 /* Caller is not willing to reclaim, we can't balance anything */
3769 if (!can_direct_reclaim
)
3772 /* Make sure we know about allocations which stall for too long */
3773 if (time_after(jiffies
, alloc_start
+ stall_timeout
)) {
3774 warn_alloc(gfp_mask
, ac
->nodemask
,
3775 "page allocation stalls for %ums, order:%u",
3776 jiffies_to_msecs(jiffies
-alloc_start
), order
);
3777 stall_timeout
+= 10 * HZ
;
3780 /* Avoid recursion of direct reclaim */
3781 if (current
->flags
& PF_MEMALLOC
)
3784 /* Try direct reclaim and then allocating */
3785 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3786 &did_some_progress
);
3790 /* Try direct compaction and then allocating */
3791 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3792 compact_priority
, &compact_result
);
3796 /* Do not loop if specifically requested */
3797 if (gfp_mask
& __GFP_NORETRY
)
3801 * Do not retry costly high order allocations unless they are
3804 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_REPEAT
))
3807 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3808 did_some_progress
> 0, &no_progress_loops
))
3812 * It doesn't make any sense to retry for the compaction if the order-0
3813 * reclaim is not able to make any progress because the current
3814 * implementation of the compaction depends on the sufficient amount
3815 * of free memory (see __compaction_suitable)
3817 if (did_some_progress
> 0 &&
3818 should_compact_retry(ac
, order
, alloc_flags
,
3819 compact_result
, &compact_priority
,
3820 &compaction_retries
))
3824 * It's possible we raced with cpuset update so the OOM would be
3825 * premature (see below the nopage: label for full explanation).
3827 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3830 /* Reclaim has failed us, start killing things */
3831 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3835 /* Avoid allocations with no watermarks from looping endlessly */
3836 if (test_thread_flag(TIF_MEMDIE
))
3839 /* Retry as long as the OOM killer is making progress */
3840 if (did_some_progress
) {
3841 no_progress_loops
= 0;
3847 * When updating a task's mems_allowed or mempolicy nodemask, it is
3848 * possible to race with parallel threads in such a way that our
3849 * allocation can fail while the mask is being updated. If we are about
3850 * to fail, check if the cpuset changed during allocation and if so,
3853 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3857 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3860 if (gfp_mask
& __GFP_NOFAIL
) {
3862 * All existing users of the __GFP_NOFAIL are blockable, so warn
3863 * of any new users that actually require GFP_NOWAIT
3865 if (WARN_ON_ONCE(!can_direct_reclaim
))
3869 * PF_MEMALLOC request from this context is rather bizarre
3870 * because we cannot reclaim anything and only can loop waiting
3871 * for somebody to do a work for us
3873 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
3876 * non failing costly orders are a hard requirement which we
3877 * are not prepared for much so let's warn about these users
3878 * so that we can identify them and convert them to something
3881 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
3884 * Help non-failing allocations by giving them access to memory
3885 * reserves but do not use ALLOC_NO_WATERMARKS because this
3886 * could deplete whole memory reserves which would just make
3887 * the situation worse
3889 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
3897 warn_alloc(gfp_mask
, ac
->nodemask
,
3898 "page allocation failure: order:%u", order
);
3903 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
3904 struct zonelist
*zonelist
, nodemask_t
*nodemask
,
3905 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
3906 unsigned int *alloc_flags
)
3908 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
3909 ac
->zonelist
= zonelist
;
3910 ac
->nodemask
= nodemask
;
3911 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
3913 if (cpusets_enabled()) {
3914 *alloc_mask
|= __GFP_HARDWALL
;
3916 ac
->nodemask
= &cpuset_current_mems_allowed
;
3918 *alloc_flags
|= ALLOC_CPUSET
;
3921 lockdep_trace_alloc(gfp_mask
);
3923 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3925 if (should_fail_alloc_page(gfp_mask
, order
))
3928 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
3929 *alloc_flags
|= ALLOC_CMA
;
3934 /* Determine whether to spread dirty pages and what the first usable zone */
3935 static inline void finalise_ac(gfp_t gfp_mask
,
3936 unsigned int order
, struct alloc_context
*ac
)
3938 /* Dirty zone balancing only done in the fast path */
3939 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3942 * The preferred zone is used for statistics but crucially it is
3943 * also used as the starting point for the zonelist iterator. It
3944 * may get reset for allocations that ignore memory policies.
3946 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3947 ac
->high_zoneidx
, ac
->nodemask
);
3951 * This is the 'heart' of the zoned buddy allocator.
3954 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3955 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3958 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
3959 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3960 struct alloc_context ac
= { };
3962 gfp_mask
&= gfp_allowed_mask
;
3963 if (!prepare_alloc_pages(gfp_mask
, order
, zonelist
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
3966 finalise_ac(gfp_mask
, order
, &ac
);
3968 /* First allocation attempt */
3969 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3974 * Runtime PM, block IO and its error handling path can deadlock
3975 * because I/O on the device might not complete.
3977 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3978 ac
.spread_dirty_pages
= false;
3981 * Restore the original nodemask if it was potentially replaced with
3982 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3984 if (unlikely(ac
.nodemask
!= nodemask
))
3985 ac
.nodemask
= nodemask
;
3987 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3990 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
3991 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
3992 __free_pages(page
, order
);
3996 if (kmemcheck_enabled
&& page
)
3997 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3999 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4003 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4006 * Common helper functions.
4008 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4013 * __get_free_pages() returns a 32-bit address, which cannot represent
4016 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
4018 page
= alloc_pages(gfp_mask
, order
);
4021 return (unsigned long) page_address(page
);
4023 EXPORT_SYMBOL(__get_free_pages
);
4025 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4027 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4029 EXPORT_SYMBOL(get_zeroed_page
);
4031 void __free_pages(struct page
*page
, unsigned int order
)
4033 if (put_page_testzero(page
)) {
4035 free_hot_cold_page(page
, false);
4037 __free_pages_ok(page
, order
);
4041 EXPORT_SYMBOL(__free_pages
);
4043 void free_pages(unsigned long addr
, unsigned int order
)
4046 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4047 __free_pages(virt_to_page((void *)addr
), order
);
4051 EXPORT_SYMBOL(free_pages
);
4055 * An arbitrary-length arbitrary-offset area of memory which resides
4056 * within a 0 or higher order page. Multiple fragments within that page
4057 * are individually refcounted, in the page's reference counter.
4059 * The page_frag functions below provide a simple allocation framework for
4060 * page fragments. This is used by the network stack and network device
4061 * drivers to provide a backing region of memory for use as either an
4062 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4064 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4067 struct page
*page
= NULL
;
4068 gfp_t gfp
= gfp_mask
;
4070 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4071 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4073 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4074 PAGE_FRAG_CACHE_MAX_ORDER
);
4075 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4077 if (unlikely(!page
))
4078 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4080 nc
->va
= page
? page_address(page
) : NULL
;
4085 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4087 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4089 if (page_ref_sub_and_test(page
, count
)) {
4090 unsigned int order
= compound_order(page
);
4093 free_hot_cold_page(page
, false);
4095 __free_pages_ok(page
, order
);
4098 EXPORT_SYMBOL(__page_frag_cache_drain
);
4100 void *page_frag_alloc(struct page_frag_cache
*nc
,
4101 unsigned int fragsz
, gfp_t gfp_mask
)
4103 unsigned int size
= PAGE_SIZE
;
4107 if (unlikely(!nc
->va
)) {
4109 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4113 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4114 /* if size can vary use size else just use PAGE_SIZE */
4117 /* Even if we own the page, we do not use atomic_set().
4118 * This would break get_page_unless_zero() users.
4120 page_ref_add(page
, size
- 1);
4122 /* reset page count bias and offset to start of new frag */
4123 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4124 nc
->pagecnt_bias
= size
;
4128 offset
= nc
->offset
- fragsz
;
4129 if (unlikely(offset
< 0)) {
4130 page
= virt_to_page(nc
->va
);
4132 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4135 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4136 /* if size can vary use size else just use PAGE_SIZE */
4139 /* OK, page count is 0, we can safely set it */
4140 set_page_count(page
, size
);
4142 /* reset page count bias and offset to start of new frag */
4143 nc
->pagecnt_bias
= size
;
4144 offset
= size
- fragsz
;
4148 nc
->offset
= offset
;
4150 return nc
->va
+ offset
;
4152 EXPORT_SYMBOL(page_frag_alloc
);
4155 * Frees a page fragment allocated out of either a compound or order 0 page.
4157 void page_frag_free(void *addr
)
4159 struct page
*page
= virt_to_head_page(addr
);
4161 if (unlikely(put_page_testzero(page
)))
4162 __free_pages_ok(page
, compound_order(page
));
4164 EXPORT_SYMBOL(page_frag_free
);
4166 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4170 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4171 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4173 split_page(virt_to_page((void *)addr
), order
);
4174 while (used
< alloc_end
) {
4179 return (void *)addr
;
4183 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4184 * @size: the number of bytes to allocate
4185 * @gfp_mask: GFP flags for the allocation
4187 * This function is similar to alloc_pages(), except that it allocates the
4188 * minimum number of pages to satisfy the request. alloc_pages() can only
4189 * allocate memory in power-of-two pages.
4191 * This function is also limited by MAX_ORDER.
4193 * Memory allocated by this function must be released by free_pages_exact().
4195 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4197 unsigned int order
= get_order(size
);
4200 addr
= __get_free_pages(gfp_mask
, order
);
4201 return make_alloc_exact(addr
, order
, size
);
4203 EXPORT_SYMBOL(alloc_pages_exact
);
4206 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4208 * @nid: the preferred node ID where memory should be allocated
4209 * @size: the number of bytes to allocate
4210 * @gfp_mask: GFP flags for the allocation
4212 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4215 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4217 unsigned int order
= get_order(size
);
4218 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4221 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4225 * free_pages_exact - release memory allocated via alloc_pages_exact()
4226 * @virt: the value returned by alloc_pages_exact.
4227 * @size: size of allocation, same value as passed to alloc_pages_exact().
4229 * Release the memory allocated by a previous call to alloc_pages_exact.
4231 void free_pages_exact(void *virt
, size_t size
)
4233 unsigned long addr
= (unsigned long)virt
;
4234 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4236 while (addr
< end
) {
4241 EXPORT_SYMBOL(free_pages_exact
);
4244 * nr_free_zone_pages - count number of pages beyond high watermark
4245 * @offset: The zone index of the highest zone
4247 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4248 * high watermark within all zones at or below a given zone index. For each
4249 * zone, the number of pages is calculated as:
4251 * nr_free_zone_pages = managed_pages - high_pages
4253 static unsigned long nr_free_zone_pages(int offset
)
4258 /* Just pick one node, since fallback list is circular */
4259 unsigned long sum
= 0;
4261 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4263 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4264 unsigned long size
= zone
->managed_pages
;
4265 unsigned long high
= high_wmark_pages(zone
);
4274 * nr_free_buffer_pages - count number of pages beyond high watermark
4276 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4277 * watermark within ZONE_DMA and ZONE_NORMAL.
4279 unsigned long nr_free_buffer_pages(void)
4281 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4283 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4286 * nr_free_pagecache_pages - count number of pages beyond high watermark
4288 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4289 * high watermark within all zones.
4291 unsigned long nr_free_pagecache_pages(void)
4293 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4296 static inline void show_node(struct zone
*zone
)
4298 if (IS_ENABLED(CONFIG_NUMA
))
4299 printk("Node %d ", zone_to_nid(zone
));
4302 long si_mem_available(void)
4305 unsigned long pagecache
;
4306 unsigned long wmark_low
= 0;
4307 unsigned long pages
[NR_LRU_LISTS
];
4311 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4312 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4315 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4318 * Estimate the amount of memory available for userspace allocations,
4319 * without causing swapping.
4321 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4324 * Not all the page cache can be freed, otherwise the system will
4325 * start swapping. Assume at least half of the page cache, or the
4326 * low watermark worth of cache, needs to stay.
4328 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4329 pagecache
-= min(pagecache
/ 2, wmark_low
);
4330 available
+= pagecache
;
4333 * Part of the reclaimable slab consists of items that are in use,
4334 * and cannot be freed. Cap this estimate at the low watermark.
4336 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4337 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4343 EXPORT_SYMBOL_GPL(si_mem_available
);
4345 void si_meminfo(struct sysinfo
*val
)
4347 val
->totalram
= totalram_pages
;
4348 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4349 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4350 val
->bufferram
= nr_blockdev_pages();
4351 val
->totalhigh
= totalhigh_pages
;
4352 val
->freehigh
= nr_free_highpages();
4353 val
->mem_unit
= PAGE_SIZE
;
4356 EXPORT_SYMBOL(si_meminfo
);
4359 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4361 int zone_type
; /* needs to be signed */
4362 unsigned long managed_pages
= 0;
4363 unsigned long managed_highpages
= 0;
4364 unsigned long free_highpages
= 0;
4365 pg_data_t
*pgdat
= NODE_DATA(nid
);
4367 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4368 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4369 val
->totalram
= managed_pages
;
4370 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4371 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4372 #ifdef CONFIG_HIGHMEM
4373 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4374 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4376 if (is_highmem(zone
)) {
4377 managed_highpages
+= zone
->managed_pages
;
4378 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4381 val
->totalhigh
= managed_highpages
;
4382 val
->freehigh
= free_highpages
;
4384 val
->totalhigh
= managed_highpages
;
4385 val
->freehigh
= free_highpages
;
4387 val
->mem_unit
= PAGE_SIZE
;
4392 * Determine whether the node should be displayed or not, depending on whether
4393 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4395 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4397 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4401 * no node mask - aka implicit memory numa policy. Do not bother with
4402 * the synchronization - read_mems_allowed_begin - because we do not
4403 * have to be precise here.
4406 nodemask
= &cpuset_current_mems_allowed
;
4408 return !node_isset(nid
, *nodemask
);
4411 #define K(x) ((x) << (PAGE_SHIFT-10))
4413 static void show_migration_types(unsigned char type
)
4415 static const char types
[MIGRATE_TYPES
] = {
4416 [MIGRATE_UNMOVABLE
] = 'U',
4417 [MIGRATE_MOVABLE
] = 'M',
4418 [MIGRATE_RECLAIMABLE
] = 'E',
4419 [MIGRATE_HIGHATOMIC
] = 'H',
4421 [MIGRATE_CMA
] = 'C',
4423 #ifdef CONFIG_MEMORY_ISOLATION
4424 [MIGRATE_ISOLATE
] = 'I',
4427 char tmp
[MIGRATE_TYPES
+ 1];
4431 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4432 if (type
& (1 << i
))
4437 printk(KERN_CONT
"(%s) ", tmp
);
4441 * Show free area list (used inside shift_scroll-lock stuff)
4442 * We also calculate the percentage fragmentation. We do this by counting the
4443 * memory on each free list with the exception of the first item on the list.
4446 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4449 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4451 unsigned long free_pcp
= 0;
4456 for_each_populated_zone(zone
) {
4457 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4460 for_each_online_cpu(cpu
)
4461 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4464 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4465 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4466 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4467 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4468 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4469 " free:%lu free_pcp:%lu free_cma:%lu\n",
4470 global_node_page_state(NR_ACTIVE_ANON
),
4471 global_node_page_state(NR_INACTIVE_ANON
),
4472 global_node_page_state(NR_ISOLATED_ANON
),
4473 global_node_page_state(NR_ACTIVE_FILE
),
4474 global_node_page_state(NR_INACTIVE_FILE
),
4475 global_node_page_state(NR_ISOLATED_FILE
),
4476 global_node_page_state(NR_UNEVICTABLE
),
4477 global_node_page_state(NR_FILE_DIRTY
),
4478 global_node_page_state(NR_WRITEBACK
),
4479 global_node_page_state(NR_UNSTABLE_NFS
),
4480 global_page_state(NR_SLAB_RECLAIMABLE
),
4481 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4482 global_node_page_state(NR_FILE_MAPPED
),
4483 global_node_page_state(NR_SHMEM
),
4484 global_page_state(NR_PAGETABLE
),
4485 global_page_state(NR_BOUNCE
),
4486 global_page_state(NR_FREE_PAGES
),
4488 global_page_state(NR_FREE_CMA_PAGES
));
4490 for_each_online_pgdat(pgdat
) {
4491 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
4495 " active_anon:%lukB"
4496 " inactive_anon:%lukB"
4497 " active_file:%lukB"
4498 " inactive_file:%lukB"
4499 " unevictable:%lukB"
4500 " isolated(anon):%lukB"
4501 " isolated(file):%lukB"
4506 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4508 " shmem_pmdmapped: %lukB"
4511 " writeback_tmp:%lukB"
4513 " pages_scanned:%lu"
4514 " all_unreclaimable? %s"
4517 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4518 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4519 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4520 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4521 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4522 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4523 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4524 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4525 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4526 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4527 K(node_page_state(pgdat
, NR_SHMEM
)),
4528 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4529 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4530 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4532 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4534 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4535 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4536 node_page_state(pgdat
, NR_PAGES_SCANNED
),
4537 !pgdat_reclaimable(pgdat
) ? "yes" : "no");
4540 for_each_populated_zone(zone
) {
4543 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4547 for_each_online_cpu(cpu
)
4548 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4557 " active_anon:%lukB"
4558 " inactive_anon:%lukB"
4559 " active_file:%lukB"
4560 " inactive_file:%lukB"
4561 " unevictable:%lukB"
4562 " writepending:%lukB"
4566 " slab_reclaimable:%lukB"
4567 " slab_unreclaimable:%lukB"
4568 " kernel_stack:%lukB"
4576 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4577 K(min_wmark_pages(zone
)),
4578 K(low_wmark_pages(zone
)),
4579 K(high_wmark_pages(zone
)),
4580 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4581 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4582 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4583 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4584 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4585 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4586 K(zone
->present_pages
),
4587 K(zone
->managed_pages
),
4588 K(zone_page_state(zone
, NR_MLOCK
)),
4589 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4590 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4591 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4592 K(zone_page_state(zone
, NR_PAGETABLE
)),
4593 K(zone_page_state(zone
, NR_BOUNCE
)),
4595 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4596 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4597 printk("lowmem_reserve[]:");
4598 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4599 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
4600 printk(KERN_CONT
"\n");
4603 for_each_populated_zone(zone
) {
4605 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4606 unsigned char types
[MAX_ORDER
];
4608 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4611 printk(KERN_CONT
"%s: ", zone
->name
);
4613 spin_lock_irqsave(&zone
->lock
, flags
);
4614 for (order
= 0; order
< MAX_ORDER
; order
++) {
4615 struct free_area
*area
= &zone
->free_area
[order
];
4618 nr
[order
] = area
->nr_free
;
4619 total
+= nr
[order
] << order
;
4622 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4623 if (!list_empty(&area
->free_list
[type
]))
4624 types
[order
] |= 1 << type
;
4627 spin_unlock_irqrestore(&zone
->lock
, flags
);
4628 for (order
= 0; order
< MAX_ORDER
; order
++) {
4629 printk(KERN_CONT
"%lu*%lukB ",
4630 nr
[order
], K(1UL) << order
);
4632 show_migration_types(types
[order
]);
4634 printk(KERN_CONT
"= %lukB\n", K(total
));
4637 hugetlb_show_meminfo();
4639 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4641 show_swap_cache_info();
4644 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4646 zoneref
->zone
= zone
;
4647 zoneref
->zone_idx
= zone_idx(zone
);
4651 * Builds allocation fallback zone lists.
4653 * Add all populated zones of a node to the zonelist.
4655 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4659 enum zone_type zone_type
= MAX_NR_ZONES
;
4663 zone
= pgdat
->node_zones
+ zone_type
;
4664 if (managed_zone(zone
)) {
4665 zoneref_set_zone(zone
,
4666 &zonelist
->_zonerefs
[nr_zones
++]);
4667 check_highest_zone(zone_type
);
4669 } while (zone_type
);
4677 * 0 = automatic detection of better ordering.
4678 * 1 = order by ([node] distance, -zonetype)
4679 * 2 = order by (-zonetype, [node] distance)
4681 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4682 * the same zonelist. So only NUMA can configure this param.
4684 #define ZONELIST_ORDER_DEFAULT 0
4685 #define ZONELIST_ORDER_NODE 1
4686 #define ZONELIST_ORDER_ZONE 2
4688 /* zonelist order in the kernel.
4689 * set_zonelist_order() will set this to NODE or ZONE.
4691 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4692 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4696 /* The value user specified ....changed by config */
4697 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4698 /* string for sysctl */
4699 #define NUMA_ZONELIST_ORDER_LEN 16
4700 char numa_zonelist_order
[16] = "default";
4703 * interface for configure zonelist ordering.
4704 * command line option "numa_zonelist_order"
4705 * = "[dD]efault - default, automatic configuration.
4706 * = "[nN]ode - order by node locality, then by zone within node
4707 * = "[zZ]one - order by zone, then by locality within zone
4710 static int __parse_numa_zonelist_order(char *s
)
4712 if (*s
== 'd' || *s
== 'D') {
4713 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4714 } else if (*s
== 'n' || *s
== 'N') {
4715 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4716 } else if (*s
== 'z' || *s
== 'Z') {
4717 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4719 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4725 static __init
int setup_numa_zonelist_order(char *s
)
4732 ret
= __parse_numa_zonelist_order(s
);
4734 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4738 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4741 * sysctl handler for numa_zonelist_order
4743 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4744 void __user
*buffer
, size_t *length
,
4747 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4749 static DEFINE_MUTEX(zl_order_mutex
);
4751 mutex_lock(&zl_order_mutex
);
4753 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4757 strcpy(saved_string
, (char *)table
->data
);
4759 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4763 int oldval
= user_zonelist_order
;
4765 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4768 * bogus value. restore saved string
4770 strncpy((char *)table
->data
, saved_string
,
4771 NUMA_ZONELIST_ORDER_LEN
);
4772 user_zonelist_order
= oldval
;
4773 } else if (oldval
!= user_zonelist_order
) {
4774 mutex_lock(&zonelists_mutex
);
4775 build_all_zonelists(NULL
, NULL
);
4776 mutex_unlock(&zonelists_mutex
);
4780 mutex_unlock(&zl_order_mutex
);
4785 #define MAX_NODE_LOAD (nr_online_nodes)
4786 static int node_load
[MAX_NUMNODES
];
4789 * find_next_best_node - find the next node that should appear in a given node's fallback list
4790 * @node: node whose fallback list we're appending
4791 * @used_node_mask: nodemask_t of already used nodes
4793 * We use a number of factors to determine which is the next node that should
4794 * appear on a given node's fallback list. The node should not have appeared
4795 * already in @node's fallback list, and it should be the next closest node
4796 * according to the distance array (which contains arbitrary distance values
4797 * from each node to each node in the system), and should also prefer nodes
4798 * with no CPUs, since presumably they'll have very little allocation pressure
4799 * on them otherwise.
4800 * It returns -1 if no node is found.
4802 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4805 int min_val
= INT_MAX
;
4806 int best_node
= NUMA_NO_NODE
;
4807 const struct cpumask
*tmp
= cpumask_of_node(0);
4809 /* Use the local node if we haven't already */
4810 if (!node_isset(node
, *used_node_mask
)) {
4811 node_set(node
, *used_node_mask
);
4815 for_each_node_state(n
, N_MEMORY
) {
4817 /* Don't want a node to appear more than once */
4818 if (node_isset(n
, *used_node_mask
))
4821 /* Use the distance array to find the distance */
4822 val
= node_distance(node
, n
);
4824 /* Penalize nodes under us ("prefer the next node") */
4827 /* Give preference to headless and unused nodes */
4828 tmp
= cpumask_of_node(n
);
4829 if (!cpumask_empty(tmp
))
4830 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4832 /* Slight preference for less loaded node */
4833 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4834 val
+= node_load
[n
];
4836 if (val
< min_val
) {
4843 node_set(best_node
, *used_node_mask
);
4850 * Build zonelists ordered by node and zones within node.
4851 * This results in maximum locality--normal zone overflows into local
4852 * DMA zone, if any--but risks exhausting DMA zone.
4854 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4857 struct zonelist
*zonelist
;
4859 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4860 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4862 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4863 zonelist
->_zonerefs
[j
].zone
= NULL
;
4864 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4868 * Build gfp_thisnode zonelists
4870 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4873 struct zonelist
*zonelist
;
4875 zonelist
= &pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
];
4876 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4877 zonelist
->_zonerefs
[j
].zone
= NULL
;
4878 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4882 * Build zonelists ordered by zone and nodes within zones.
4883 * This results in conserving DMA zone[s] until all Normal memory is
4884 * exhausted, but results in overflowing to remote node while memory
4885 * may still exist in local DMA zone.
4887 static int node_order
[MAX_NUMNODES
];
4889 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4892 int zone_type
; /* needs to be signed */
4894 struct zonelist
*zonelist
;
4896 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4898 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4899 for (j
= 0; j
< nr_nodes
; j
++) {
4900 node
= node_order
[j
];
4901 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4902 if (managed_zone(z
)) {
4904 &zonelist
->_zonerefs
[pos
++]);
4905 check_highest_zone(zone_type
);
4909 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4910 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4913 #if defined(CONFIG_64BIT)
4915 * Devices that require DMA32/DMA are relatively rare and do not justify a
4916 * penalty to every machine in case the specialised case applies. Default
4917 * to Node-ordering on 64-bit NUMA machines
4919 static int default_zonelist_order(void)
4921 return ZONELIST_ORDER_NODE
;
4925 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4926 * by the kernel. If processes running on node 0 deplete the low memory zone
4927 * then reclaim will occur more frequency increasing stalls and potentially
4928 * be easier to OOM if a large percentage of the zone is under writeback or
4929 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4930 * Hence, default to zone ordering on 32-bit.
4932 static int default_zonelist_order(void)
4934 return ZONELIST_ORDER_ZONE
;
4936 #endif /* CONFIG_64BIT */
4938 static void set_zonelist_order(void)
4940 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4941 current_zonelist_order
= default_zonelist_order();
4943 current_zonelist_order
= user_zonelist_order
;
4946 static void build_zonelists(pg_data_t
*pgdat
)
4949 nodemask_t used_mask
;
4950 int local_node
, prev_node
;
4951 struct zonelist
*zonelist
;
4952 unsigned int order
= current_zonelist_order
;
4954 /* initialize zonelists */
4955 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4956 zonelist
= pgdat
->node_zonelists
+ i
;
4957 zonelist
->_zonerefs
[0].zone
= NULL
;
4958 zonelist
->_zonerefs
[0].zone_idx
= 0;
4961 /* NUMA-aware ordering of nodes */
4962 local_node
= pgdat
->node_id
;
4963 load
= nr_online_nodes
;
4964 prev_node
= local_node
;
4965 nodes_clear(used_mask
);
4967 memset(node_order
, 0, sizeof(node_order
));
4970 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4972 * We don't want to pressure a particular node.
4973 * So adding penalty to the first node in same
4974 * distance group to make it round-robin.
4976 if (node_distance(local_node
, node
) !=
4977 node_distance(local_node
, prev_node
))
4978 node_load
[node
] = load
;
4982 if (order
== ZONELIST_ORDER_NODE
)
4983 build_zonelists_in_node_order(pgdat
, node
);
4985 node_order
[i
++] = node
; /* remember order */
4988 if (order
== ZONELIST_ORDER_ZONE
) {
4989 /* calculate node order -- i.e., DMA last! */
4990 build_zonelists_in_zone_order(pgdat
, i
);
4993 build_thisnode_zonelists(pgdat
);
4996 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4998 * Return node id of node used for "local" allocations.
4999 * I.e., first node id of first zone in arg node's generic zonelist.
5000 * Used for initializing percpu 'numa_mem', which is used primarily
5001 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5003 int local_memory_node(int node
)
5007 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5008 gfp_zone(GFP_KERNEL
),
5010 return z
->zone
->node
;
5014 static void setup_min_unmapped_ratio(void);
5015 static void setup_min_slab_ratio(void);
5016 #else /* CONFIG_NUMA */
5018 static void set_zonelist_order(void)
5020 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
5023 static void build_zonelists(pg_data_t
*pgdat
)
5025 int node
, local_node
;
5027 struct zonelist
*zonelist
;
5029 local_node
= pgdat
->node_id
;
5031 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
5032 j
= build_zonelists_node(pgdat
, zonelist
, 0);
5035 * Now we build the zonelist so that it contains the zones
5036 * of all the other nodes.
5037 * We don't want to pressure a particular node, so when
5038 * building the zones for node N, we make sure that the
5039 * zones coming right after the local ones are those from
5040 * node N+1 (modulo N)
5042 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5043 if (!node_online(node
))
5045 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
5047 for (node
= 0; node
< local_node
; node
++) {
5048 if (!node_online(node
))
5050 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
5053 zonelist
->_zonerefs
[j
].zone
= NULL
;
5054 zonelist
->_zonerefs
[j
].zone_idx
= 0;
5057 #endif /* CONFIG_NUMA */
5060 * Boot pageset table. One per cpu which is going to be used for all
5061 * zones and all nodes. The parameters will be set in such a way
5062 * that an item put on a list will immediately be handed over to
5063 * the buddy list. This is safe since pageset manipulation is done
5064 * with interrupts disabled.
5066 * The boot_pagesets must be kept even after bootup is complete for
5067 * unused processors and/or zones. They do play a role for bootstrapping
5068 * hotplugged processors.
5070 * zoneinfo_show() and maybe other functions do
5071 * not check if the processor is online before following the pageset pointer.
5072 * Other parts of the kernel may not check if the zone is available.
5074 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5075 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5076 static void setup_zone_pageset(struct zone
*zone
);
5079 * Global mutex to protect against size modification of zonelists
5080 * as well as to serialize pageset setup for the new populated zone.
5082 DEFINE_MUTEX(zonelists_mutex
);
5084 /* return values int ....just for stop_machine() */
5085 static int __build_all_zonelists(void *data
)
5089 pg_data_t
*self
= data
;
5092 memset(node_load
, 0, sizeof(node_load
));
5095 if (self
&& !node_online(self
->node_id
)) {
5096 build_zonelists(self
);
5099 for_each_online_node(nid
) {
5100 pg_data_t
*pgdat
= NODE_DATA(nid
);
5102 build_zonelists(pgdat
);
5106 * Initialize the boot_pagesets that are going to be used
5107 * for bootstrapping processors. The real pagesets for
5108 * each zone will be allocated later when the per cpu
5109 * allocator is available.
5111 * boot_pagesets are used also for bootstrapping offline
5112 * cpus if the system is already booted because the pagesets
5113 * are needed to initialize allocators on a specific cpu too.
5114 * F.e. the percpu allocator needs the page allocator which
5115 * needs the percpu allocator in order to allocate its pagesets
5116 * (a chicken-egg dilemma).
5118 for_each_possible_cpu(cpu
) {
5119 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5121 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5123 * We now know the "local memory node" for each node--
5124 * i.e., the node of the first zone in the generic zonelist.
5125 * Set up numa_mem percpu variable for on-line cpus. During
5126 * boot, only the boot cpu should be on-line; we'll init the
5127 * secondary cpus' numa_mem as they come on-line. During
5128 * node/memory hotplug, we'll fixup all on-line cpus.
5130 if (cpu_online(cpu
))
5131 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5138 static noinline
void __init
5139 build_all_zonelists_init(void)
5141 __build_all_zonelists(NULL
);
5142 mminit_verify_zonelist();
5143 cpuset_init_current_mems_allowed();
5147 * Called with zonelists_mutex held always
5148 * unless system_state == SYSTEM_BOOTING.
5150 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5151 * [we're only called with non-NULL zone through __meminit paths] and
5152 * (2) call of __init annotated helper build_all_zonelists_init
5153 * [protected by SYSTEM_BOOTING].
5155 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
5157 set_zonelist_order();
5159 if (system_state
== SYSTEM_BOOTING
) {
5160 build_all_zonelists_init();
5162 #ifdef CONFIG_MEMORY_HOTPLUG
5164 setup_zone_pageset(zone
);
5166 /* we have to stop all cpus to guarantee there is no user
5168 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
5169 /* cpuset refresh routine should be here */
5171 vm_total_pages
= nr_free_pagecache_pages();
5173 * Disable grouping by mobility if the number of pages in the
5174 * system is too low to allow the mechanism to work. It would be
5175 * more accurate, but expensive to check per-zone. This check is
5176 * made on memory-hotadd so a system can start with mobility
5177 * disabled and enable it later
5179 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5180 page_group_by_mobility_disabled
= 1;
5182 page_group_by_mobility_disabled
= 0;
5184 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5186 zonelist_order_name
[current_zonelist_order
],
5187 page_group_by_mobility_disabled
? "off" : "on",
5190 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5195 * Initially all pages are reserved - free ones are freed
5196 * up by free_all_bootmem() once the early boot process is
5197 * done. Non-atomic initialization, single-pass.
5199 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5200 unsigned long start_pfn
, enum memmap_context context
)
5202 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5203 unsigned long end_pfn
= start_pfn
+ size
;
5204 pg_data_t
*pgdat
= NODE_DATA(nid
);
5206 unsigned long nr_initialised
= 0;
5207 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5208 struct memblock_region
*r
= NULL
, *tmp
;
5211 if (highest_memmap_pfn
< end_pfn
- 1)
5212 highest_memmap_pfn
= end_pfn
- 1;
5215 * Honor reservation requested by the driver for this ZONE_DEVICE
5218 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5219 start_pfn
+= altmap
->reserve
;
5221 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5223 * There can be holes in boot-time mem_map[]s handed to this
5224 * function. They do not exist on hotplugged memory.
5226 if (context
!= MEMMAP_EARLY
)
5229 if (!early_pfn_valid(pfn
)) {
5230 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5232 * Skip to the pfn preceding the next valid one (or
5233 * end_pfn), such that we hit a valid pfn (or end_pfn)
5234 * on our next iteration of the loop.
5236 pfn
= memblock_next_valid_pfn(pfn
, end_pfn
) - 1;
5240 if (!early_pfn_in_nid(pfn
, nid
))
5242 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5245 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5247 * Check given memblock attribute by firmware which can affect
5248 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5249 * mirrored, it's an overlapped memmap init. skip it.
5251 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5252 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5253 for_each_memblock(memory
, tmp
)
5254 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5258 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5259 memblock_is_mirror(r
)) {
5260 /* already initialized as NORMAL */
5261 pfn
= memblock_region_memory_end_pfn(r
);
5269 * Mark the block movable so that blocks are reserved for
5270 * movable at startup. This will force kernel allocations
5271 * to reserve their blocks rather than leaking throughout
5272 * the address space during boot when many long-lived
5273 * kernel allocations are made.
5275 * bitmap is created for zone's valid pfn range. but memmap
5276 * can be created for invalid pages (for alignment)
5277 * check here not to call set_pageblock_migratetype() against
5280 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5281 struct page
*page
= pfn_to_page(pfn
);
5283 __init_single_page(page
, pfn
, zone
, nid
);
5284 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5286 __init_single_pfn(pfn
, zone
, nid
);
5291 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5293 unsigned int order
, t
;
5294 for_each_migratetype_order(order
, t
) {
5295 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5296 zone
->free_area
[order
].nr_free
= 0;
5300 #ifndef __HAVE_ARCH_MEMMAP_INIT
5301 #define memmap_init(size, nid, zone, start_pfn) \
5302 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5305 static int zone_batchsize(struct zone
*zone
)
5311 * The per-cpu-pages pools are set to around 1000th of the
5312 * size of the zone. But no more than 1/2 of a meg.
5314 * OK, so we don't know how big the cache is. So guess.
5316 batch
= zone
->managed_pages
/ 1024;
5317 if (batch
* PAGE_SIZE
> 512 * 1024)
5318 batch
= (512 * 1024) / PAGE_SIZE
;
5319 batch
/= 4; /* We effectively *= 4 below */
5324 * Clamp the batch to a 2^n - 1 value. Having a power
5325 * of 2 value was found to be more likely to have
5326 * suboptimal cache aliasing properties in some cases.
5328 * For example if 2 tasks are alternately allocating
5329 * batches of pages, one task can end up with a lot
5330 * of pages of one half of the possible page colors
5331 * and the other with pages of the other colors.
5333 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5338 /* The deferral and batching of frees should be suppressed under NOMMU
5341 * The problem is that NOMMU needs to be able to allocate large chunks
5342 * of contiguous memory as there's no hardware page translation to
5343 * assemble apparent contiguous memory from discontiguous pages.
5345 * Queueing large contiguous runs of pages for batching, however,
5346 * causes the pages to actually be freed in smaller chunks. As there
5347 * can be a significant delay between the individual batches being
5348 * recycled, this leads to the once large chunks of space being
5349 * fragmented and becoming unavailable for high-order allocations.
5356 * pcp->high and pcp->batch values are related and dependent on one another:
5357 * ->batch must never be higher then ->high.
5358 * The following function updates them in a safe manner without read side
5361 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5362 * those fields changing asynchronously (acording the the above rule).
5364 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5365 * outside of boot time (or some other assurance that no concurrent updaters
5368 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5369 unsigned long batch
)
5371 /* start with a fail safe value for batch */
5375 /* Update high, then batch, in order */
5382 /* a companion to pageset_set_high() */
5383 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5385 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5388 static void pageset_init(struct per_cpu_pageset
*p
)
5390 struct per_cpu_pages
*pcp
;
5393 memset(p
, 0, sizeof(*p
));
5397 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5398 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5401 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5404 pageset_set_batch(p
, batch
);
5408 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5409 * to the value high for the pageset p.
5411 static void pageset_set_high(struct per_cpu_pageset
*p
,
5414 unsigned long batch
= max(1UL, high
/ 4);
5415 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5416 batch
= PAGE_SHIFT
* 8;
5418 pageset_update(&p
->pcp
, high
, batch
);
5421 static void pageset_set_high_and_batch(struct zone
*zone
,
5422 struct per_cpu_pageset
*pcp
)
5424 if (percpu_pagelist_fraction
)
5425 pageset_set_high(pcp
,
5426 (zone
->managed_pages
/
5427 percpu_pagelist_fraction
));
5429 pageset_set_batch(pcp
, zone_batchsize(zone
));
5432 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5434 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5437 pageset_set_high_and_batch(zone
, pcp
);
5440 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5443 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5444 for_each_possible_cpu(cpu
)
5445 zone_pageset_init(zone
, cpu
);
5449 * Allocate per cpu pagesets and initialize them.
5450 * Before this call only boot pagesets were available.
5452 void __init
setup_per_cpu_pageset(void)
5454 struct pglist_data
*pgdat
;
5457 for_each_populated_zone(zone
)
5458 setup_zone_pageset(zone
);
5460 for_each_online_pgdat(pgdat
)
5461 pgdat
->per_cpu_nodestats
=
5462 alloc_percpu(struct per_cpu_nodestat
);
5465 static __meminit
void zone_pcp_init(struct zone
*zone
)
5468 * per cpu subsystem is not up at this point. The following code
5469 * relies on the ability of the linker to provide the
5470 * offset of a (static) per cpu variable into the per cpu area.
5472 zone
->pageset
= &boot_pageset
;
5474 if (populated_zone(zone
))
5475 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5476 zone
->name
, zone
->present_pages
,
5477 zone_batchsize(zone
));
5480 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5481 unsigned long zone_start_pfn
,
5484 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5486 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5488 zone
->zone_start_pfn
= zone_start_pfn
;
5490 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5491 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5493 (unsigned long)zone_idx(zone
),
5494 zone_start_pfn
, (zone_start_pfn
+ size
));
5496 zone_init_free_lists(zone
);
5497 zone
->initialized
= 1;
5502 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5503 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5506 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5508 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5509 struct mminit_pfnnid_cache
*state
)
5511 unsigned long start_pfn
, end_pfn
;
5514 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5515 return state
->last_nid
;
5517 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5519 state
->last_start
= start_pfn
;
5520 state
->last_end
= end_pfn
;
5521 state
->last_nid
= nid
;
5526 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5529 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5530 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5531 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5533 * If an architecture guarantees that all ranges registered contain no holes
5534 * and may be freed, this this function may be used instead of calling
5535 * memblock_free_early_nid() manually.
5537 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5539 unsigned long start_pfn
, end_pfn
;
5542 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5543 start_pfn
= min(start_pfn
, max_low_pfn
);
5544 end_pfn
= min(end_pfn
, max_low_pfn
);
5546 if (start_pfn
< end_pfn
)
5547 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5548 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5554 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5555 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5557 * If an architecture guarantees that all ranges registered contain no holes and may
5558 * be freed, this function may be used instead of calling memory_present() manually.
5560 void __init
sparse_memory_present_with_active_regions(int nid
)
5562 unsigned long start_pfn
, end_pfn
;
5565 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5566 memory_present(this_nid
, start_pfn
, end_pfn
);
5570 * get_pfn_range_for_nid - Return the start and end page frames for a node
5571 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5572 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5573 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5575 * It returns the start and end page frame of a node based on information
5576 * provided by memblock_set_node(). If called for a node
5577 * with no available memory, a warning is printed and the start and end
5580 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5581 unsigned long *start_pfn
, unsigned long *end_pfn
)
5583 unsigned long this_start_pfn
, this_end_pfn
;
5589 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5590 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5591 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5594 if (*start_pfn
== -1UL)
5599 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5600 * assumption is made that zones within a node are ordered in monotonic
5601 * increasing memory addresses so that the "highest" populated zone is used
5603 static void __init
find_usable_zone_for_movable(void)
5606 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5607 if (zone_index
== ZONE_MOVABLE
)
5610 if (arch_zone_highest_possible_pfn
[zone_index
] >
5611 arch_zone_lowest_possible_pfn
[zone_index
])
5615 VM_BUG_ON(zone_index
== -1);
5616 movable_zone
= zone_index
;
5620 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5621 * because it is sized independent of architecture. Unlike the other zones,
5622 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5623 * in each node depending on the size of each node and how evenly kernelcore
5624 * is distributed. This helper function adjusts the zone ranges
5625 * provided by the architecture for a given node by using the end of the
5626 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5627 * zones within a node are in order of monotonic increases memory addresses
5629 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5630 unsigned long zone_type
,
5631 unsigned long node_start_pfn
,
5632 unsigned long node_end_pfn
,
5633 unsigned long *zone_start_pfn
,
5634 unsigned long *zone_end_pfn
)
5636 /* Only adjust if ZONE_MOVABLE is on this node */
5637 if (zone_movable_pfn
[nid
]) {
5638 /* Size ZONE_MOVABLE */
5639 if (zone_type
== ZONE_MOVABLE
) {
5640 *zone_start_pfn
= zone_movable_pfn
[nid
];
5641 *zone_end_pfn
= min(node_end_pfn
,
5642 arch_zone_highest_possible_pfn
[movable_zone
]);
5644 /* Adjust for ZONE_MOVABLE starting within this range */
5645 } else if (!mirrored_kernelcore
&&
5646 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5647 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5648 *zone_end_pfn
= zone_movable_pfn
[nid
];
5650 /* Check if this whole range is within ZONE_MOVABLE */
5651 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5652 *zone_start_pfn
= *zone_end_pfn
;
5657 * Return the number of pages a zone spans in a node, including holes
5658 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5660 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5661 unsigned long zone_type
,
5662 unsigned long node_start_pfn
,
5663 unsigned long node_end_pfn
,
5664 unsigned long *zone_start_pfn
,
5665 unsigned long *zone_end_pfn
,
5666 unsigned long *ignored
)
5668 /* When hotadd a new node from cpu_up(), the node should be empty */
5669 if (!node_start_pfn
&& !node_end_pfn
)
5672 /* Get the start and end of the zone */
5673 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5674 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5675 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5676 node_start_pfn
, node_end_pfn
,
5677 zone_start_pfn
, zone_end_pfn
);
5679 /* Check that this node has pages within the zone's required range */
5680 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5683 /* Move the zone boundaries inside the node if necessary */
5684 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5685 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5687 /* Return the spanned pages */
5688 return *zone_end_pfn
- *zone_start_pfn
;
5692 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5693 * then all holes in the requested range will be accounted for.
5695 unsigned long __meminit
__absent_pages_in_range(int nid
,
5696 unsigned long range_start_pfn
,
5697 unsigned long range_end_pfn
)
5699 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5700 unsigned long start_pfn
, end_pfn
;
5703 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5704 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5705 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5706 nr_absent
-= end_pfn
- start_pfn
;
5712 * absent_pages_in_range - Return number of page frames in holes within a range
5713 * @start_pfn: The start PFN to start searching for holes
5714 * @end_pfn: The end PFN to stop searching for holes
5716 * It returns the number of pages frames in memory holes within a range.
5718 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5719 unsigned long end_pfn
)
5721 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5724 /* Return the number of page frames in holes in a zone on a node */
5725 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5726 unsigned long zone_type
,
5727 unsigned long node_start_pfn
,
5728 unsigned long node_end_pfn
,
5729 unsigned long *ignored
)
5731 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5732 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5733 unsigned long zone_start_pfn
, zone_end_pfn
;
5734 unsigned long nr_absent
;
5736 /* When hotadd a new node from cpu_up(), the node should be empty */
5737 if (!node_start_pfn
&& !node_end_pfn
)
5740 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5741 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5743 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5744 node_start_pfn
, node_end_pfn
,
5745 &zone_start_pfn
, &zone_end_pfn
);
5746 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5749 * ZONE_MOVABLE handling.
5750 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5753 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5754 unsigned long start_pfn
, end_pfn
;
5755 struct memblock_region
*r
;
5757 for_each_memblock(memory
, r
) {
5758 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5759 zone_start_pfn
, zone_end_pfn
);
5760 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5761 zone_start_pfn
, zone_end_pfn
);
5763 if (zone_type
== ZONE_MOVABLE
&&
5764 memblock_is_mirror(r
))
5765 nr_absent
+= end_pfn
- start_pfn
;
5767 if (zone_type
== ZONE_NORMAL
&&
5768 !memblock_is_mirror(r
))
5769 nr_absent
+= end_pfn
- start_pfn
;
5776 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5777 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5778 unsigned long zone_type
,
5779 unsigned long node_start_pfn
,
5780 unsigned long node_end_pfn
,
5781 unsigned long *zone_start_pfn
,
5782 unsigned long *zone_end_pfn
,
5783 unsigned long *zones_size
)
5787 *zone_start_pfn
= node_start_pfn
;
5788 for (zone
= 0; zone
< zone_type
; zone
++)
5789 *zone_start_pfn
+= zones_size
[zone
];
5791 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5793 return zones_size
[zone_type
];
5796 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5797 unsigned long zone_type
,
5798 unsigned long node_start_pfn
,
5799 unsigned long node_end_pfn
,
5800 unsigned long *zholes_size
)
5805 return zholes_size
[zone_type
];
5808 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5810 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5811 unsigned long node_start_pfn
,
5812 unsigned long node_end_pfn
,
5813 unsigned long *zones_size
,
5814 unsigned long *zholes_size
)
5816 unsigned long realtotalpages
= 0, totalpages
= 0;
5819 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5820 struct zone
*zone
= pgdat
->node_zones
+ i
;
5821 unsigned long zone_start_pfn
, zone_end_pfn
;
5822 unsigned long size
, real_size
;
5824 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5830 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5831 node_start_pfn
, node_end_pfn
,
5834 zone
->zone_start_pfn
= zone_start_pfn
;
5836 zone
->zone_start_pfn
= 0;
5837 zone
->spanned_pages
= size
;
5838 zone
->present_pages
= real_size
;
5841 realtotalpages
+= real_size
;
5844 pgdat
->node_spanned_pages
= totalpages
;
5845 pgdat
->node_present_pages
= realtotalpages
;
5846 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5850 #ifndef CONFIG_SPARSEMEM
5852 * Calculate the size of the zone->blockflags rounded to an unsigned long
5853 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5854 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5855 * round what is now in bits to nearest long in bits, then return it in
5858 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5860 unsigned long usemapsize
;
5862 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5863 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5864 usemapsize
= usemapsize
>> pageblock_order
;
5865 usemapsize
*= NR_PAGEBLOCK_BITS
;
5866 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5868 return usemapsize
/ 8;
5871 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5873 unsigned long zone_start_pfn
,
5874 unsigned long zonesize
)
5876 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5877 zone
->pageblock_flags
= NULL
;
5879 zone
->pageblock_flags
=
5880 memblock_virt_alloc_node_nopanic(usemapsize
,
5884 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5885 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5886 #endif /* CONFIG_SPARSEMEM */
5888 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5890 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5891 void __paginginit
set_pageblock_order(void)
5895 /* Check that pageblock_nr_pages has not already been setup */
5896 if (pageblock_order
)
5899 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5900 order
= HUGETLB_PAGE_ORDER
;
5902 order
= MAX_ORDER
- 1;
5905 * Assume the largest contiguous order of interest is a huge page.
5906 * This value may be variable depending on boot parameters on IA64 and
5909 pageblock_order
= order
;
5911 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5914 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5915 * is unused as pageblock_order is set at compile-time. See
5916 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5919 void __paginginit
set_pageblock_order(void)
5923 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5925 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5926 unsigned long present_pages
)
5928 unsigned long pages
= spanned_pages
;
5931 * Provide a more accurate estimation if there are holes within
5932 * the zone and SPARSEMEM is in use. If there are holes within the
5933 * zone, each populated memory region may cost us one or two extra
5934 * memmap pages due to alignment because memmap pages for each
5935 * populated regions may not be naturally aligned on page boundary.
5936 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5938 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5939 IS_ENABLED(CONFIG_SPARSEMEM
))
5940 pages
= present_pages
;
5942 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5946 * Set up the zone data structures:
5947 * - mark all pages reserved
5948 * - mark all memory queues empty
5949 * - clear the memory bitmaps
5951 * NOTE: pgdat should get zeroed by caller.
5953 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5956 int nid
= pgdat
->node_id
;
5959 pgdat_resize_init(pgdat
);
5960 #ifdef CONFIG_NUMA_BALANCING
5961 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5962 pgdat
->numabalancing_migrate_nr_pages
= 0;
5963 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5965 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5966 spin_lock_init(&pgdat
->split_queue_lock
);
5967 INIT_LIST_HEAD(&pgdat
->split_queue
);
5968 pgdat
->split_queue_len
= 0;
5970 init_waitqueue_head(&pgdat
->kswapd_wait
);
5971 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5972 #ifdef CONFIG_COMPACTION
5973 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5975 pgdat_page_ext_init(pgdat
);
5976 spin_lock_init(&pgdat
->lru_lock
);
5977 lruvec_init(node_lruvec(pgdat
));
5979 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5980 struct zone
*zone
= pgdat
->node_zones
+ j
;
5981 unsigned long size
, realsize
, freesize
, memmap_pages
;
5982 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5984 size
= zone
->spanned_pages
;
5985 realsize
= freesize
= zone
->present_pages
;
5988 * Adjust freesize so that it accounts for how much memory
5989 * is used by this zone for memmap. This affects the watermark
5990 * and per-cpu initialisations
5992 memmap_pages
= calc_memmap_size(size
, realsize
);
5993 if (!is_highmem_idx(j
)) {
5994 if (freesize
>= memmap_pages
) {
5995 freesize
-= memmap_pages
;
5998 " %s zone: %lu pages used for memmap\n",
5999 zone_names
[j
], memmap_pages
);
6001 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6002 zone_names
[j
], memmap_pages
, freesize
);
6005 /* Account for reserved pages */
6006 if (j
== 0 && freesize
> dma_reserve
) {
6007 freesize
-= dma_reserve
;
6008 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6009 zone_names
[0], dma_reserve
);
6012 if (!is_highmem_idx(j
))
6013 nr_kernel_pages
+= freesize
;
6014 /* Charge for highmem memmap if there are enough kernel pages */
6015 else if (nr_kernel_pages
> memmap_pages
* 2)
6016 nr_kernel_pages
-= memmap_pages
;
6017 nr_all_pages
+= freesize
;
6020 * Set an approximate value for lowmem here, it will be adjusted
6021 * when the bootmem allocator frees pages into the buddy system.
6022 * And all highmem pages will be managed by the buddy system.
6024 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
6028 zone
->name
= zone_names
[j
];
6029 zone
->zone_pgdat
= pgdat
;
6030 spin_lock_init(&zone
->lock
);
6031 zone_seqlock_init(zone
);
6032 zone_pcp_init(zone
);
6037 set_pageblock_order();
6038 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6039 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6041 memmap_init(size
, nid
, j
, zone_start_pfn
);
6045 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6047 unsigned long __maybe_unused start
= 0;
6048 unsigned long __maybe_unused offset
= 0;
6050 /* Skip empty nodes */
6051 if (!pgdat
->node_spanned_pages
)
6054 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6055 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6056 offset
= pgdat
->node_start_pfn
- start
;
6057 /* ia64 gets its own node_mem_map, before this, without bootmem */
6058 if (!pgdat
->node_mem_map
) {
6059 unsigned long size
, end
;
6063 * The zone's endpoints aren't required to be MAX_ORDER
6064 * aligned but the node_mem_map endpoints must be in order
6065 * for the buddy allocator to function correctly.
6067 end
= pgdat_end_pfn(pgdat
);
6068 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6069 size
= (end
- start
) * sizeof(struct page
);
6070 map
= alloc_remap(pgdat
->node_id
, size
);
6072 map
= memblock_virt_alloc_node_nopanic(size
,
6074 pgdat
->node_mem_map
= map
+ offset
;
6076 #ifndef CONFIG_NEED_MULTIPLE_NODES
6078 * With no DISCONTIG, the global mem_map is just set as node 0's
6080 if (pgdat
== NODE_DATA(0)) {
6081 mem_map
= NODE_DATA(0)->node_mem_map
;
6082 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6083 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6085 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6088 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6091 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
6092 unsigned long node_start_pfn
, unsigned long *zholes_size
)
6094 pg_data_t
*pgdat
= NODE_DATA(nid
);
6095 unsigned long start_pfn
= 0;
6096 unsigned long end_pfn
= 0;
6098 /* pg_data_t should be reset to zero when it's allocated */
6099 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6101 reset_deferred_meminit(pgdat
);
6102 pgdat
->node_id
= nid
;
6103 pgdat
->node_start_pfn
= node_start_pfn
;
6104 pgdat
->per_cpu_nodestats
= NULL
;
6105 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6106 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6107 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6108 (u64
)start_pfn
<< PAGE_SHIFT
,
6109 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6111 start_pfn
= node_start_pfn
;
6113 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6114 zones_size
, zholes_size
);
6116 alloc_node_mem_map(pgdat
);
6117 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6118 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6119 nid
, (unsigned long)pgdat
,
6120 (unsigned long)pgdat
->node_mem_map
);
6123 free_area_init_core(pgdat
);
6126 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6128 #if MAX_NUMNODES > 1
6130 * Figure out the number of possible node ids.
6132 void __init
setup_nr_node_ids(void)
6134 unsigned int highest
;
6136 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6137 nr_node_ids
= highest
+ 1;
6142 * node_map_pfn_alignment - determine the maximum internode alignment
6144 * This function should be called after node map is populated and sorted.
6145 * It calculates the maximum power of two alignment which can distinguish
6148 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6149 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6150 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6151 * shifted, 1GiB is enough and this function will indicate so.
6153 * This is used to test whether pfn -> nid mapping of the chosen memory
6154 * model has fine enough granularity to avoid incorrect mapping for the
6155 * populated node map.
6157 * Returns the determined alignment in pfn's. 0 if there is no alignment
6158 * requirement (single node).
6160 unsigned long __init
node_map_pfn_alignment(void)
6162 unsigned long accl_mask
= 0, last_end
= 0;
6163 unsigned long start
, end
, mask
;
6167 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6168 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6175 * Start with a mask granular enough to pin-point to the
6176 * start pfn and tick off bits one-by-one until it becomes
6177 * too coarse to separate the current node from the last.
6179 mask
= ~((1 << __ffs(start
)) - 1);
6180 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6183 /* accumulate all internode masks */
6187 /* convert mask to number of pages */
6188 return ~accl_mask
+ 1;
6191 /* Find the lowest pfn for a node */
6192 static unsigned long __init
find_min_pfn_for_node(int nid
)
6194 unsigned long min_pfn
= ULONG_MAX
;
6195 unsigned long start_pfn
;
6198 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6199 min_pfn
= min(min_pfn
, start_pfn
);
6201 if (min_pfn
== ULONG_MAX
) {
6202 pr_warn("Could not find start_pfn for node %d\n", nid
);
6210 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6212 * It returns the minimum PFN based on information provided via
6213 * memblock_set_node().
6215 unsigned long __init
find_min_pfn_with_active_regions(void)
6217 return find_min_pfn_for_node(MAX_NUMNODES
);
6221 * early_calculate_totalpages()
6222 * Sum pages in active regions for movable zone.
6223 * Populate N_MEMORY for calculating usable_nodes.
6225 static unsigned long __init
early_calculate_totalpages(void)
6227 unsigned long totalpages
= 0;
6228 unsigned long start_pfn
, end_pfn
;
6231 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6232 unsigned long pages
= end_pfn
- start_pfn
;
6234 totalpages
+= pages
;
6236 node_set_state(nid
, N_MEMORY
);
6242 * Find the PFN the Movable zone begins in each node. Kernel memory
6243 * is spread evenly between nodes as long as the nodes have enough
6244 * memory. When they don't, some nodes will have more kernelcore than
6247 static void __init
find_zone_movable_pfns_for_nodes(void)
6250 unsigned long usable_startpfn
;
6251 unsigned long kernelcore_node
, kernelcore_remaining
;
6252 /* save the state before borrow the nodemask */
6253 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6254 unsigned long totalpages
= early_calculate_totalpages();
6255 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6256 struct memblock_region
*r
;
6258 /* Need to find movable_zone earlier when movable_node is specified. */
6259 find_usable_zone_for_movable();
6262 * If movable_node is specified, ignore kernelcore and movablecore
6265 if (movable_node_is_enabled()) {
6266 for_each_memblock(memory
, r
) {
6267 if (!memblock_is_hotpluggable(r
))
6272 usable_startpfn
= PFN_DOWN(r
->base
);
6273 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6274 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6282 * If kernelcore=mirror is specified, ignore movablecore option
6284 if (mirrored_kernelcore
) {
6285 bool mem_below_4gb_not_mirrored
= false;
6287 for_each_memblock(memory
, r
) {
6288 if (memblock_is_mirror(r
))
6293 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6295 if (usable_startpfn
< 0x100000) {
6296 mem_below_4gb_not_mirrored
= true;
6300 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6301 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6305 if (mem_below_4gb_not_mirrored
)
6306 pr_warn("This configuration results in unmirrored kernel memory.");
6312 * If movablecore=nn[KMG] was specified, calculate what size of
6313 * kernelcore that corresponds so that memory usable for
6314 * any allocation type is evenly spread. If both kernelcore
6315 * and movablecore are specified, then the value of kernelcore
6316 * will be used for required_kernelcore if it's greater than
6317 * what movablecore would have allowed.
6319 if (required_movablecore
) {
6320 unsigned long corepages
;
6323 * Round-up so that ZONE_MOVABLE is at least as large as what
6324 * was requested by the user
6326 required_movablecore
=
6327 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6328 required_movablecore
= min(totalpages
, required_movablecore
);
6329 corepages
= totalpages
- required_movablecore
;
6331 required_kernelcore
= max(required_kernelcore
, corepages
);
6335 * If kernelcore was not specified or kernelcore size is larger
6336 * than totalpages, there is no ZONE_MOVABLE.
6338 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6341 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6342 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6345 /* Spread kernelcore memory as evenly as possible throughout nodes */
6346 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6347 for_each_node_state(nid
, N_MEMORY
) {
6348 unsigned long start_pfn
, end_pfn
;
6351 * Recalculate kernelcore_node if the division per node
6352 * now exceeds what is necessary to satisfy the requested
6353 * amount of memory for the kernel
6355 if (required_kernelcore
< kernelcore_node
)
6356 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6359 * As the map is walked, we track how much memory is usable
6360 * by the kernel using kernelcore_remaining. When it is
6361 * 0, the rest of the node is usable by ZONE_MOVABLE
6363 kernelcore_remaining
= kernelcore_node
;
6365 /* Go through each range of PFNs within this node */
6366 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6367 unsigned long size_pages
;
6369 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6370 if (start_pfn
>= end_pfn
)
6373 /* Account for what is only usable for kernelcore */
6374 if (start_pfn
< usable_startpfn
) {
6375 unsigned long kernel_pages
;
6376 kernel_pages
= min(end_pfn
, usable_startpfn
)
6379 kernelcore_remaining
-= min(kernel_pages
,
6380 kernelcore_remaining
);
6381 required_kernelcore
-= min(kernel_pages
,
6382 required_kernelcore
);
6384 /* Continue if range is now fully accounted */
6385 if (end_pfn
<= usable_startpfn
) {
6388 * Push zone_movable_pfn to the end so
6389 * that if we have to rebalance
6390 * kernelcore across nodes, we will
6391 * not double account here
6393 zone_movable_pfn
[nid
] = end_pfn
;
6396 start_pfn
= usable_startpfn
;
6400 * The usable PFN range for ZONE_MOVABLE is from
6401 * start_pfn->end_pfn. Calculate size_pages as the
6402 * number of pages used as kernelcore
6404 size_pages
= end_pfn
- start_pfn
;
6405 if (size_pages
> kernelcore_remaining
)
6406 size_pages
= kernelcore_remaining
;
6407 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6410 * Some kernelcore has been met, update counts and
6411 * break if the kernelcore for this node has been
6414 required_kernelcore
-= min(required_kernelcore
,
6416 kernelcore_remaining
-= size_pages
;
6417 if (!kernelcore_remaining
)
6423 * If there is still required_kernelcore, we do another pass with one
6424 * less node in the count. This will push zone_movable_pfn[nid] further
6425 * along on the nodes that still have memory until kernelcore is
6429 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6433 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6434 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6435 zone_movable_pfn
[nid
] =
6436 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6439 /* restore the node_state */
6440 node_states
[N_MEMORY
] = saved_node_state
;
6443 /* Any regular or high memory on that node ? */
6444 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6446 enum zone_type zone_type
;
6448 if (N_MEMORY
== N_NORMAL_MEMORY
)
6451 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6452 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6453 if (populated_zone(zone
)) {
6454 node_set_state(nid
, N_HIGH_MEMORY
);
6455 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6456 zone_type
<= ZONE_NORMAL
)
6457 node_set_state(nid
, N_NORMAL_MEMORY
);
6464 * free_area_init_nodes - Initialise all pg_data_t and zone data
6465 * @max_zone_pfn: an array of max PFNs for each zone
6467 * This will call free_area_init_node() for each active node in the system.
6468 * Using the page ranges provided by memblock_set_node(), the size of each
6469 * zone in each node and their holes is calculated. If the maximum PFN
6470 * between two adjacent zones match, it is assumed that the zone is empty.
6471 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6472 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6473 * starts where the previous one ended. For example, ZONE_DMA32 starts
6474 * at arch_max_dma_pfn.
6476 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6478 unsigned long start_pfn
, end_pfn
;
6481 /* Record where the zone boundaries are */
6482 memset(arch_zone_lowest_possible_pfn
, 0,
6483 sizeof(arch_zone_lowest_possible_pfn
));
6484 memset(arch_zone_highest_possible_pfn
, 0,
6485 sizeof(arch_zone_highest_possible_pfn
));
6487 start_pfn
= find_min_pfn_with_active_regions();
6489 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6490 if (i
== ZONE_MOVABLE
)
6493 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6494 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6495 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6497 start_pfn
= end_pfn
;
6500 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6501 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6502 find_zone_movable_pfns_for_nodes();
6504 /* Print out the zone ranges */
6505 pr_info("Zone ranges:\n");
6506 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6507 if (i
== ZONE_MOVABLE
)
6509 pr_info(" %-8s ", zone_names
[i
]);
6510 if (arch_zone_lowest_possible_pfn
[i
] ==
6511 arch_zone_highest_possible_pfn
[i
])
6514 pr_cont("[mem %#018Lx-%#018Lx]\n",
6515 (u64
)arch_zone_lowest_possible_pfn
[i
]
6517 ((u64
)arch_zone_highest_possible_pfn
[i
]
6518 << PAGE_SHIFT
) - 1);
6521 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6522 pr_info("Movable zone start for each node\n");
6523 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6524 if (zone_movable_pfn
[i
])
6525 pr_info(" Node %d: %#018Lx\n", i
,
6526 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6529 /* Print out the early node map */
6530 pr_info("Early memory node ranges\n");
6531 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6532 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6533 (u64
)start_pfn
<< PAGE_SHIFT
,
6534 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6536 /* Initialise every node */
6537 mminit_verify_pageflags_layout();
6538 setup_nr_node_ids();
6539 for_each_online_node(nid
) {
6540 pg_data_t
*pgdat
= NODE_DATA(nid
);
6541 free_area_init_node(nid
, NULL
,
6542 find_min_pfn_for_node(nid
), NULL
);
6544 /* Any memory on that node */
6545 if (pgdat
->node_present_pages
)
6546 node_set_state(nid
, N_MEMORY
);
6547 check_for_memory(pgdat
, nid
);
6551 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6553 unsigned long long coremem
;
6557 coremem
= memparse(p
, &p
);
6558 *core
= coremem
>> PAGE_SHIFT
;
6560 /* Paranoid check that UL is enough for the coremem value */
6561 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6567 * kernelcore=size sets the amount of memory for use for allocations that
6568 * cannot be reclaimed or migrated.
6570 static int __init
cmdline_parse_kernelcore(char *p
)
6572 /* parse kernelcore=mirror */
6573 if (parse_option_str(p
, "mirror")) {
6574 mirrored_kernelcore
= true;
6578 return cmdline_parse_core(p
, &required_kernelcore
);
6582 * movablecore=size sets the amount of memory for use for allocations that
6583 * can be reclaimed or migrated.
6585 static int __init
cmdline_parse_movablecore(char *p
)
6587 return cmdline_parse_core(p
, &required_movablecore
);
6590 early_param("kernelcore", cmdline_parse_kernelcore
);
6591 early_param("movablecore", cmdline_parse_movablecore
);
6593 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6595 void adjust_managed_page_count(struct page
*page
, long count
)
6597 spin_lock(&managed_page_count_lock
);
6598 page_zone(page
)->managed_pages
+= count
;
6599 totalram_pages
+= count
;
6600 #ifdef CONFIG_HIGHMEM
6601 if (PageHighMem(page
))
6602 totalhigh_pages
+= count
;
6604 spin_unlock(&managed_page_count_lock
);
6606 EXPORT_SYMBOL(adjust_managed_page_count
);
6608 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6611 unsigned long pages
= 0;
6613 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6614 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6615 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6616 if ((unsigned int)poison
<= 0xFF)
6617 memset(pos
, poison
, PAGE_SIZE
);
6618 free_reserved_page(virt_to_page(pos
));
6622 pr_info("Freeing %s memory: %ldK\n",
6623 s
, pages
<< (PAGE_SHIFT
- 10));
6627 EXPORT_SYMBOL(free_reserved_area
);
6629 #ifdef CONFIG_HIGHMEM
6630 void free_highmem_page(struct page
*page
)
6632 __free_reserved_page(page
);
6634 page_zone(page
)->managed_pages
++;
6640 void __init
mem_init_print_info(const char *str
)
6642 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6643 unsigned long init_code_size
, init_data_size
;
6645 physpages
= get_num_physpages();
6646 codesize
= _etext
- _stext
;
6647 datasize
= _edata
- _sdata
;
6648 rosize
= __end_rodata
- __start_rodata
;
6649 bss_size
= __bss_stop
- __bss_start
;
6650 init_data_size
= __init_end
- __init_begin
;
6651 init_code_size
= _einittext
- _sinittext
;
6654 * Detect special cases and adjust section sizes accordingly:
6655 * 1) .init.* may be embedded into .data sections
6656 * 2) .init.text.* may be out of [__init_begin, __init_end],
6657 * please refer to arch/tile/kernel/vmlinux.lds.S.
6658 * 3) .rodata.* may be embedded into .text or .data sections.
6660 #define adj_init_size(start, end, size, pos, adj) \
6662 if (start <= pos && pos < end && size > adj) \
6666 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6667 _sinittext
, init_code_size
);
6668 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6669 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6670 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6671 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6673 #undef adj_init_size
6675 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6676 #ifdef CONFIG_HIGHMEM
6680 nr_free_pages() << (PAGE_SHIFT
- 10),
6681 physpages
<< (PAGE_SHIFT
- 10),
6682 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6683 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6684 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6685 totalcma_pages
<< (PAGE_SHIFT
- 10),
6686 #ifdef CONFIG_HIGHMEM
6687 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6689 str
? ", " : "", str
? str
: "");
6693 * set_dma_reserve - set the specified number of pages reserved in the first zone
6694 * @new_dma_reserve: The number of pages to mark reserved
6696 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6697 * In the DMA zone, a significant percentage may be consumed by kernel image
6698 * and other unfreeable allocations which can skew the watermarks badly. This
6699 * function may optionally be used to account for unfreeable pages in the
6700 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6701 * smaller per-cpu batchsize.
6703 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6705 dma_reserve
= new_dma_reserve
;
6708 void __init
free_area_init(unsigned long *zones_size
)
6710 free_area_init_node(0, zones_size
,
6711 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6714 static int page_alloc_cpu_dead(unsigned int cpu
)
6717 lru_add_drain_cpu(cpu
);
6721 * Spill the event counters of the dead processor
6722 * into the current processors event counters.
6723 * This artificially elevates the count of the current
6726 vm_events_fold_cpu(cpu
);
6729 * Zero the differential counters of the dead processor
6730 * so that the vm statistics are consistent.
6732 * This is only okay since the processor is dead and cannot
6733 * race with what we are doing.
6735 cpu_vm_stats_fold(cpu
);
6739 void __init
page_alloc_init(void)
6743 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
6744 "mm/page_alloc:dead", NULL
,
6745 page_alloc_cpu_dead
);
6750 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6751 * or min_free_kbytes changes.
6753 static void calculate_totalreserve_pages(void)
6755 struct pglist_data
*pgdat
;
6756 unsigned long reserve_pages
= 0;
6757 enum zone_type i
, j
;
6759 for_each_online_pgdat(pgdat
) {
6761 pgdat
->totalreserve_pages
= 0;
6763 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6764 struct zone
*zone
= pgdat
->node_zones
+ i
;
6767 /* Find valid and maximum lowmem_reserve in the zone */
6768 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6769 if (zone
->lowmem_reserve
[j
] > max
)
6770 max
= zone
->lowmem_reserve
[j
];
6773 /* we treat the high watermark as reserved pages. */
6774 max
+= high_wmark_pages(zone
);
6776 if (max
> zone
->managed_pages
)
6777 max
= zone
->managed_pages
;
6779 pgdat
->totalreserve_pages
+= max
;
6781 reserve_pages
+= max
;
6784 totalreserve_pages
= reserve_pages
;
6788 * setup_per_zone_lowmem_reserve - called whenever
6789 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6790 * has a correct pages reserved value, so an adequate number of
6791 * pages are left in the zone after a successful __alloc_pages().
6793 static void setup_per_zone_lowmem_reserve(void)
6795 struct pglist_data
*pgdat
;
6796 enum zone_type j
, idx
;
6798 for_each_online_pgdat(pgdat
) {
6799 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6800 struct zone
*zone
= pgdat
->node_zones
+ j
;
6801 unsigned long managed_pages
= zone
->managed_pages
;
6803 zone
->lowmem_reserve
[j
] = 0;
6807 struct zone
*lower_zone
;
6811 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6812 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6814 lower_zone
= pgdat
->node_zones
+ idx
;
6815 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6816 sysctl_lowmem_reserve_ratio
[idx
];
6817 managed_pages
+= lower_zone
->managed_pages
;
6822 /* update totalreserve_pages */
6823 calculate_totalreserve_pages();
6826 static void __setup_per_zone_wmarks(void)
6828 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6829 unsigned long lowmem_pages
= 0;
6831 unsigned long flags
;
6833 /* Calculate total number of !ZONE_HIGHMEM pages */
6834 for_each_zone(zone
) {
6835 if (!is_highmem(zone
))
6836 lowmem_pages
+= zone
->managed_pages
;
6839 for_each_zone(zone
) {
6842 spin_lock_irqsave(&zone
->lock
, flags
);
6843 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6844 do_div(tmp
, lowmem_pages
);
6845 if (is_highmem(zone
)) {
6847 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6848 * need highmem pages, so cap pages_min to a small
6851 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6852 * deltas control asynch page reclaim, and so should
6853 * not be capped for highmem.
6855 unsigned long min_pages
;
6857 min_pages
= zone
->managed_pages
/ 1024;
6858 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6859 zone
->watermark
[WMARK_MIN
] = min_pages
;
6862 * If it's a lowmem zone, reserve a number of pages
6863 * proportionate to the zone's size.
6865 zone
->watermark
[WMARK_MIN
] = tmp
;
6869 * Set the kswapd watermarks distance according to the
6870 * scale factor in proportion to available memory, but
6871 * ensure a minimum size on small systems.
6873 tmp
= max_t(u64
, tmp
>> 2,
6874 mult_frac(zone
->managed_pages
,
6875 watermark_scale_factor
, 10000));
6877 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6878 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6880 spin_unlock_irqrestore(&zone
->lock
, flags
);
6883 /* update totalreserve_pages */
6884 calculate_totalreserve_pages();
6888 * setup_per_zone_wmarks - called when min_free_kbytes changes
6889 * or when memory is hot-{added|removed}
6891 * Ensures that the watermark[min,low,high] values for each zone are set
6892 * correctly with respect to min_free_kbytes.
6894 void setup_per_zone_wmarks(void)
6896 mutex_lock(&zonelists_mutex
);
6897 __setup_per_zone_wmarks();
6898 mutex_unlock(&zonelists_mutex
);
6902 * Initialise min_free_kbytes.
6904 * For small machines we want it small (128k min). For large machines
6905 * we want it large (64MB max). But it is not linear, because network
6906 * bandwidth does not increase linearly with machine size. We use
6908 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6909 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6925 int __meminit
init_per_zone_wmark_min(void)
6927 unsigned long lowmem_kbytes
;
6928 int new_min_free_kbytes
;
6930 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6931 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6933 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6934 min_free_kbytes
= new_min_free_kbytes
;
6935 if (min_free_kbytes
< 128)
6936 min_free_kbytes
= 128;
6937 if (min_free_kbytes
> 65536)
6938 min_free_kbytes
= 65536;
6940 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6941 new_min_free_kbytes
, user_min_free_kbytes
);
6943 setup_per_zone_wmarks();
6944 refresh_zone_stat_thresholds();
6945 setup_per_zone_lowmem_reserve();
6948 setup_min_unmapped_ratio();
6949 setup_min_slab_ratio();
6954 core_initcall(init_per_zone_wmark_min
)
6957 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6958 * that we can call two helper functions whenever min_free_kbytes
6961 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6962 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6966 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6971 user_min_free_kbytes
= min_free_kbytes
;
6972 setup_per_zone_wmarks();
6977 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6978 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6982 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6987 setup_per_zone_wmarks();
6993 static void setup_min_unmapped_ratio(void)
6998 for_each_online_pgdat(pgdat
)
6999 pgdat
->min_unmapped_pages
= 0;
7002 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
7003 sysctl_min_unmapped_ratio
) / 100;
7007 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7008 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7012 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7016 setup_min_unmapped_ratio();
7021 static void setup_min_slab_ratio(void)
7026 for_each_online_pgdat(pgdat
)
7027 pgdat
->min_slab_pages
= 0;
7030 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
7031 sysctl_min_slab_ratio
) / 100;
7034 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7035 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7039 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7043 setup_min_slab_ratio();
7050 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7051 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7052 * whenever sysctl_lowmem_reserve_ratio changes.
7054 * The reserve ratio obviously has absolutely no relation with the
7055 * minimum watermarks. The lowmem reserve ratio can only make sense
7056 * if in function of the boot time zone sizes.
7058 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7059 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7061 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7062 setup_per_zone_lowmem_reserve();
7067 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7068 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7069 * pagelist can have before it gets flushed back to buddy allocator.
7071 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7072 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7075 int old_percpu_pagelist_fraction
;
7078 mutex_lock(&pcp_batch_high_lock
);
7079 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7081 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7082 if (!write
|| ret
< 0)
7085 /* Sanity checking to avoid pcp imbalance */
7086 if (percpu_pagelist_fraction
&&
7087 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7088 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7094 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7097 for_each_populated_zone(zone
) {
7100 for_each_possible_cpu(cpu
)
7101 pageset_set_high_and_batch(zone
,
7102 per_cpu_ptr(zone
->pageset
, cpu
));
7105 mutex_unlock(&pcp_batch_high_lock
);
7110 int hashdist
= HASHDIST_DEFAULT
;
7112 static int __init
set_hashdist(char *str
)
7116 hashdist
= simple_strtoul(str
, &str
, 0);
7119 __setup("hashdist=", set_hashdist
);
7122 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7124 * Returns the number of pages that arch has reserved but
7125 * is not known to alloc_large_system_hash().
7127 static unsigned long __init
arch_reserved_kernel_pages(void)
7134 * allocate a large system hash table from bootmem
7135 * - it is assumed that the hash table must contain an exact power-of-2
7136 * quantity of entries
7137 * - limit is the number of hash buckets, not the total allocation size
7139 void *__init
alloc_large_system_hash(const char *tablename
,
7140 unsigned long bucketsize
,
7141 unsigned long numentries
,
7144 unsigned int *_hash_shift
,
7145 unsigned int *_hash_mask
,
7146 unsigned long low_limit
,
7147 unsigned long high_limit
)
7149 unsigned long long max
= high_limit
;
7150 unsigned long log2qty
, size
;
7153 /* allow the kernel cmdline to have a say */
7155 /* round applicable memory size up to nearest megabyte */
7156 numentries
= nr_kernel_pages
;
7157 numentries
-= arch_reserved_kernel_pages();
7159 /* It isn't necessary when PAGE_SIZE >= 1MB */
7160 if (PAGE_SHIFT
< 20)
7161 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7163 /* limit to 1 bucket per 2^scale bytes of low memory */
7164 if (scale
> PAGE_SHIFT
)
7165 numentries
>>= (scale
- PAGE_SHIFT
);
7167 numentries
<<= (PAGE_SHIFT
- scale
);
7169 /* Make sure we've got at least a 0-order allocation.. */
7170 if (unlikely(flags
& HASH_SMALL
)) {
7171 /* Makes no sense without HASH_EARLY */
7172 WARN_ON(!(flags
& HASH_EARLY
));
7173 if (!(numentries
>> *_hash_shift
)) {
7174 numentries
= 1UL << *_hash_shift
;
7175 BUG_ON(!numentries
);
7177 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7178 numentries
= PAGE_SIZE
/ bucketsize
;
7180 numentries
= roundup_pow_of_two(numentries
);
7182 /* limit allocation size to 1/16 total memory by default */
7184 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7185 do_div(max
, bucketsize
);
7187 max
= min(max
, 0x80000000ULL
);
7189 if (numentries
< low_limit
)
7190 numentries
= low_limit
;
7191 if (numentries
> max
)
7194 log2qty
= ilog2(numentries
);
7197 size
= bucketsize
<< log2qty
;
7198 if (flags
& HASH_EARLY
)
7199 table
= memblock_virt_alloc_nopanic(size
, 0);
7201 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
7204 * If bucketsize is not a power-of-two, we may free
7205 * some pages at the end of hash table which
7206 * alloc_pages_exact() automatically does
7208 if (get_order(size
) < MAX_ORDER
) {
7209 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
7210 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
7213 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7216 panic("Failed to allocate %s hash table\n", tablename
);
7218 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7219 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7222 *_hash_shift
= log2qty
;
7224 *_hash_mask
= (1 << log2qty
) - 1;
7230 * This function checks whether pageblock includes unmovable pages or not.
7231 * If @count is not zero, it is okay to include less @count unmovable pages
7233 * PageLRU check without isolation or lru_lock could race so that
7234 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7235 * check without lock_page also may miss some movable non-lru pages at
7236 * race condition. So you can't expect this function should be exact.
7238 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7239 bool skip_hwpoisoned_pages
)
7241 unsigned long pfn
, iter
, found
;
7245 * For avoiding noise data, lru_add_drain_all() should be called
7246 * If ZONE_MOVABLE, the zone never contains unmovable pages
7248 if (zone_idx(zone
) == ZONE_MOVABLE
)
7250 mt
= get_pageblock_migratetype(page
);
7251 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7254 pfn
= page_to_pfn(page
);
7255 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7256 unsigned long check
= pfn
+ iter
;
7258 if (!pfn_valid_within(check
))
7261 page
= pfn_to_page(check
);
7264 * Hugepages are not in LRU lists, but they're movable.
7265 * We need not scan over tail pages bacause we don't
7266 * handle each tail page individually in migration.
7268 if (PageHuge(page
)) {
7269 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7274 * We can't use page_count without pin a page
7275 * because another CPU can free compound page.
7276 * This check already skips compound tails of THP
7277 * because their page->_refcount is zero at all time.
7279 if (!page_ref_count(page
)) {
7280 if (PageBuddy(page
))
7281 iter
+= (1 << page_order(page
)) - 1;
7286 * The HWPoisoned page may be not in buddy system, and
7287 * page_count() is not 0.
7289 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7292 if (__PageMovable(page
))
7298 * If there are RECLAIMABLE pages, we need to check
7299 * it. But now, memory offline itself doesn't call
7300 * shrink_node_slabs() and it still to be fixed.
7303 * If the page is not RAM, page_count()should be 0.
7304 * we don't need more check. This is an _used_ not-movable page.
7306 * The problematic thing here is PG_reserved pages. PG_reserved
7307 * is set to both of a memory hole page and a _used_ kernel
7316 bool is_pageblock_removable_nolock(struct page
*page
)
7322 * We have to be careful here because we are iterating over memory
7323 * sections which are not zone aware so we might end up outside of
7324 * the zone but still within the section.
7325 * We have to take care about the node as well. If the node is offline
7326 * its NODE_DATA will be NULL - see page_zone.
7328 if (!node_online(page_to_nid(page
)))
7331 zone
= page_zone(page
);
7332 pfn
= page_to_pfn(page
);
7333 if (!zone_spans_pfn(zone
, pfn
))
7336 return !has_unmovable_pages(zone
, page
, 0, true);
7339 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7341 static unsigned long pfn_max_align_down(unsigned long pfn
)
7343 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7344 pageblock_nr_pages
) - 1);
7347 static unsigned long pfn_max_align_up(unsigned long pfn
)
7349 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7350 pageblock_nr_pages
));
7353 /* [start, end) must belong to a single zone. */
7354 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7355 unsigned long start
, unsigned long end
)
7357 /* This function is based on compact_zone() from compaction.c. */
7358 unsigned long nr_reclaimed
;
7359 unsigned long pfn
= start
;
7360 unsigned int tries
= 0;
7365 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7366 if (fatal_signal_pending(current
)) {
7371 if (list_empty(&cc
->migratepages
)) {
7372 cc
->nr_migratepages
= 0;
7373 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7379 } else if (++tries
== 5) {
7380 ret
= ret
< 0 ? ret
: -EBUSY
;
7384 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7386 cc
->nr_migratepages
-= nr_reclaimed
;
7388 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7389 NULL
, 0, cc
->mode
, MR_CMA
);
7392 putback_movable_pages(&cc
->migratepages
);
7399 * alloc_contig_range() -- tries to allocate given range of pages
7400 * @start: start PFN to allocate
7401 * @end: one-past-the-last PFN to allocate
7402 * @migratetype: migratetype of the underlaying pageblocks (either
7403 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7404 * in range must have the same migratetype and it must
7405 * be either of the two.
7406 * @gfp_mask: GFP mask to use during compaction
7408 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7409 * aligned, however it's the caller's responsibility to guarantee that
7410 * we are the only thread that changes migrate type of pageblocks the
7413 * The PFN range must belong to a single zone.
7415 * Returns zero on success or negative error code. On success all
7416 * pages which PFN is in [start, end) are allocated for the caller and
7417 * need to be freed with free_contig_range().
7419 int alloc_contig_range(unsigned long start
, unsigned long end
,
7420 unsigned migratetype
, gfp_t gfp_mask
)
7422 unsigned long outer_start
, outer_end
;
7426 struct compact_control cc
= {
7427 .nr_migratepages
= 0,
7429 .zone
= page_zone(pfn_to_page(start
)),
7430 .mode
= MIGRATE_SYNC
,
7431 .ignore_skip_hint
= true,
7432 .gfp_mask
= memalloc_noio_flags(gfp_mask
),
7434 INIT_LIST_HEAD(&cc
.migratepages
);
7437 * What we do here is we mark all pageblocks in range as
7438 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7439 * have different sizes, and due to the way page allocator
7440 * work, we align the range to biggest of the two pages so
7441 * that page allocator won't try to merge buddies from
7442 * different pageblocks and change MIGRATE_ISOLATE to some
7443 * other migration type.
7445 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7446 * migrate the pages from an unaligned range (ie. pages that
7447 * we are interested in). This will put all the pages in
7448 * range back to page allocator as MIGRATE_ISOLATE.
7450 * When this is done, we take the pages in range from page
7451 * allocator removing them from the buddy system. This way
7452 * page allocator will never consider using them.
7454 * This lets us mark the pageblocks back as
7455 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7456 * aligned range but not in the unaligned, original range are
7457 * put back to page allocator so that buddy can use them.
7460 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7461 pfn_max_align_up(end
), migratetype
,
7467 * In case of -EBUSY, we'd like to know which page causes problem.
7468 * So, just fall through. We will check it in test_pages_isolated().
7470 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7471 if (ret
&& ret
!= -EBUSY
)
7475 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7476 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7477 * more, all pages in [start, end) are free in page allocator.
7478 * What we are going to do is to allocate all pages from
7479 * [start, end) (that is remove them from page allocator).
7481 * The only problem is that pages at the beginning and at the
7482 * end of interesting range may be not aligned with pages that
7483 * page allocator holds, ie. they can be part of higher order
7484 * pages. Because of this, we reserve the bigger range and
7485 * once this is done free the pages we are not interested in.
7487 * We don't have to hold zone->lock here because the pages are
7488 * isolated thus they won't get removed from buddy.
7491 lru_add_drain_all();
7492 drain_all_pages(cc
.zone
);
7495 outer_start
= start
;
7496 while (!PageBuddy(pfn_to_page(outer_start
))) {
7497 if (++order
>= MAX_ORDER
) {
7498 outer_start
= start
;
7501 outer_start
&= ~0UL << order
;
7504 if (outer_start
!= start
) {
7505 order
= page_order(pfn_to_page(outer_start
));
7508 * outer_start page could be small order buddy page and
7509 * it doesn't include start page. Adjust outer_start
7510 * in this case to report failed page properly
7511 * on tracepoint in test_pages_isolated()
7513 if (outer_start
+ (1UL << order
) <= start
)
7514 outer_start
= start
;
7517 /* Make sure the range is really isolated. */
7518 if (test_pages_isolated(outer_start
, end
, false)) {
7519 pr_info("%s: [%lx, %lx) PFNs busy\n",
7520 __func__
, outer_start
, end
);
7525 /* Grab isolated pages from freelists. */
7526 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7532 /* Free head and tail (if any) */
7533 if (start
!= outer_start
)
7534 free_contig_range(outer_start
, start
- outer_start
);
7535 if (end
!= outer_end
)
7536 free_contig_range(end
, outer_end
- end
);
7539 undo_isolate_page_range(pfn_max_align_down(start
),
7540 pfn_max_align_up(end
), migratetype
);
7544 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7546 unsigned int count
= 0;
7548 for (; nr_pages
--; pfn
++) {
7549 struct page
*page
= pfn_to_page(pfn
);
7551 count
+= page_count(page
) != 1;
7554 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7558 #ifdef CONFIG_MEMORY_HOTPLUG
7560 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7561 * page high values need to be recalulated.
7563 void __meminit
zone_pcp_update(struct zone
*zone
)
7566 mutex_lock(&pcp_batch_high_lock
);
7567 for_each_possible_cpu(cpu
)
7568 pageset_set_high_and_batch(zone
,
7569 per_cpu_ptr(zone
->pageset
, cpu
));
7570 mutex_unlock(&pcp_batch_high_lock
);
7574 void zone_pcp_reset(struct zone
*zone
)
7576 unsigned long flags
;
7578 struct per_cpu_pageset
*pset
;
7580 /* avoid races with drain_pages() */
7581 local_irq_save(flags
);
7582 if (zone
->pageset
!= &boot_pageset
) {
7583 for_each_online_cpu(cpu
) {
7584 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7585 drain_zonestat(zone
, pset
);
7587 free_percpu(zone
->pageset
);
7588 zone
->pageset
= &boot_pageset
;
7590 local_irq_restore(flags
);
7593 #ifdef CONFIG_MEMORY_HOTREMOVE
7595 * All pages in the range must be in a single zone and isolated
7596 * before calling this.
7599 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7603 unsigned int order
, i
;
7605 unsigned long flags
;
7606 /* find the first valid pfn */
7607 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7612 zone
= page_zone(pfn_to_page(pfn
));
7613 spin_lock_irqsave(&zone
->lock
, flags
);
7615 while (pfn
< end_pfn
) {
7616 if (!pfn_valid(pfn
)) {
7620 page
= pfn_to_page(pfn
);
7622 * The HWPoisoned page may be not in buddy system, and
7623 * page_count() is not 0.
7625 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7627 SetPageReserved(page
);
7631 BUG_ON(page_count(page
));
7632 BUG_ON(!PageBuddy(page
));
7633 order
= page_order(page
);
7634 #ifdef CONFIG_DEBUG_VM
7635 pr_info("remove from free list %lx %d %lx\n",
7636 pfn
, 1 << order
, end_pfn
);
7638 list_del(&page
->lru
);
7639 rmv_page_order(page
);
7640 zone
->free_area
[order
].nr_free
--;
7641 for (i
= 0; i
< (1 << order
); i
++)
7642 SetPageReserved((page
+i
));
7643 pfn
+= (1 << order
);
7645 spin_unlock_irqrestore(&zone
->lock
, flags
);
7649 bool is_free_buddy_page(struct page
*page
)
7651 struct zone
*zone
= page_zone(page
);
7652 unsigned long pfn
= page_to_pfn(page
);
7653 unsigned long flags
;
7656 spin_lock_irqsave(&zone
->lock
, flags
);
7657 for (order
= 0; order
< MAX_ORDER
; order
++) {
7658 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7660 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7663 spin_unlock_irqrestore(&zone
->lock
, flags
);
7665 return order
< MAX_ORDER
;