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>
68 #include <linux/ftrace.h>
70 #include <asm/sections.h>
71 #include <asm/tlbflush.h>
72 #include <asm/div64.h>
75 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
76 static DEFINE_MUTEX(pcp_batch_high_lock
);
77 #define MIN_PERCPU_PAGELIST_FRACTION (8)
79 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
80 DEFINE_PER_CPU(int, numa_node
);
81 EXPORT_PER_CPU_SYMBOL(numa_node
);
84 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
86 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
87 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
88 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
89 * defined in <linux/topology.h>.
91 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
92 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
93 int _node_numa_mem_
[MAX_NUMNODES
];
96 /* work_structs for global per-cpu drains */
97 DEFINE_MUTEX(pcpu_drain_mutex
);
98 DEFINE_PER_CPU(struct work_struct
, pcpu_drain
);
100 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
101 volatile unsigned long latent_entropy __latent_entropy
;
102 EXPORT_SYMBOL(latent_entropy
);
106 * Array of node states.
108 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
109 [N_POSSIBLE
] = NODE_MASK_ALL
,
110 [N_ONLINE
] = { { [0] = 1UL } },
112 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
113 #ifdef CONFIG_HIGHMEM
114 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
116 #ifdef CONFIG_MOVABLE_NODE
117 [N_MEMORY
] = { { [0] = 1UL } },
119 [N_CPU
] = { { [0] = 1UL } },
122 EXPORT_SYMBOL(node_states
);
124 /* Protect totalram_pages and zone->managed_pages */
125 static DEFINE_SPINLOCK(managed_page_count_lock
);
127 unsigned long totalram_pages __read_mostly
;
128 unsigned long totalreserve_pages __read_mostly
;
129 unsigned long totalcma_pages __read_mostly
;
131 int percpu_pagelist_fraction
;
132 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
135 * A cached value of the page's pageblock's migratetype, used when the page is
136 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
137 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
138 * Also the migratetype set in the page does not necessarily match the pcplist
139 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
140 * other index - this ensures that it will be put on the correct CMA freelist.
142 static inline int get_pcppage_migratetype(struct page
*page
)
147 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
149 page
->index
= migratetype
;
152 #ifdef CONFIG_PM_SLEEP
154 * The following functions are used by the suspend/hibernate code to temporarily
155 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
156 * while devices are suspended. To avoid races with the suspend/hibernate code,
157 * they should always be called with pm_mutex held (gfp_allowed_mask also should
158 * only be modified with pm_mutex held, unless the suspend/hibernate code is
159 * guaranteed not to run in parallel with that modification).
162 static gfp_t saved_gfp_mask
;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&pm_mutex
));
167 if (saved_gfp_mask
) {
168 gfp_allowed_mask
= saved_gfp_mask
;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&pm_mutex
));
176 WARN_ON(saved_gfp_mask
);
177 saved_gfp_mask
= gfp_allowed_mask
;
178 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly
;
193 static void __free_pages_ok(struct page
*page
, unsigned int order
);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
213 #ifdef CONFIG_HIGHMEM
219 EXPORT_SYMBOL(totalram_pages
);
221 static char * const zone_names
[MAX_NR_ZONES
] = {
222 #ifdef CONFIG_ZONE_DMA
225 #ifdef CONFIG_ZONE_DMA32
229 #ifdef CONFIG_HIGHMEM
233 #ifdef CONFIG_ZONE_DEVICE
238 char * const migratetype_names
[MIGRATE_TYPES
] = {
246 #ifdef CONFIG_MEMORY_ISOLATION
251 compound_page_dtor
* const compound_page_dtors
[] = {
254 #ifdef CONFIG_HUGETLB_PAGE
257 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
262 int min_free_kbytes
= 1024;
263 int user_min_free_kbytes
= -1;
264 int watermark_scale_factor
= 10;
266 static unsigned long __meminitdata nr_kernel_pages
;
267 static unsigned long __meminitdata nr_all_pages
;
268 static unsigned long __meminitdata dma_reserve
;
270 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
271 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
272 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
273 static unsigned long __initdata required_kernelcore
;
274 static unsigned long __initdata required_movablecore
;
275 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
276 static bool mirrored_kernelcore
;
278 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
280 EXPORT_SYMBOL(movable_zone
);
281 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
284 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
285 int nr_online_nodes __read_mostly
= 1;
286 EXPORT_SYMBOL(nr_node_ids
);
287 EXPORT_SYMBOL(nr_online_nodes
);
290 int page_group_by_mobility_disabled __read_mostly
;
292 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
293 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
295 unsigned long max_initialise
;
296 unsigned long reserved_lowmem
;
299 * Initialise at least 2G of a node but also take into account that
300 * two large system hashes that can take up 1GB for 0.25TB/node.
302 max_initialise
= max(2UL << (30 - PAGE_SHIFT
),
303 (pgdat
->node_spanned_pages
>> 8));
306 * Compensate the all the memblock reservations (e.g. crash kernel)
307 * from the initial estimation to make sure we will initialize enough
310 reserved_lowmem
= memblock_reserved_memory_within(pgdat
->node_start_pfn
,
311 pgdat
->node_start_pfn
+ max_initialise
);
312 max_initialise
+= reserved_lowmem
;
314 pgdat
->static_init_size
= min(max_initialise
, pgdat
->node_spanned_pages
);
315 pgdat
->first_deferred_pfn
= ULONG_MAX
;
318 /* Returns true if the struct page for the pfn is uninitialised */
319 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
321 int nid
= early_pfn_to_nid(pfn
);
323 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
330 * Returns false when the remaining initialisation should be deferred until
331 * later in the boot cycle when it can be parallelised.
333 static inline bool update_defer_init(pg_data_t
*pgdat
,
334 unsigned long pfn
, unsigned long zone_end
,
335 unsigned long *nr_initialised
)
337 /* Always populate low zones for address-contrained allocations */
338 if (zone_end
< pgdat_end_pfn(pgdat
))
341 if ((*nr_initialised
> pgdat
->static_init_size
) &&
342 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
343 pgdat
->first_deferred_pfn
= pfn
;
350 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
354 static inline bool early_page_uninitialised(unsigned long pfn
)
359 static inline bool update_defer_init(pg_data_t
*pgdat
,
360 unsigned long pfn
, unsigned long zone_end
,
361 unsigned long *nr_initialised
)
367 /* Return a pointer to the bitmap storing bits affecting a block of pages */
368 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
371 #ifdef CONFIG_SPARSEMEM
372 return __pfn_to_section(pfn
)->pageblock_flags
;
374 return page_zone(page
)->pageblock_flags
;
375 #endif /* CONFIG_SPARSEMEM */
378 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
380 #ifdef CONFIG_SPARSEMEM
381 pfn
&= (PAGES_PER_SECTION
-1);
382 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
384 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
385 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
386 #endif /* CONFIG_SPARSEMEM */
390 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
391 * @page: The page within the block of interest
392 * @pfn: The target page frame number
393 * @end_bitidx: The last bit of interest to retrieve
394 * @mask: mask of bits that the caller is interested in
396 * Return: pageblock_bits flags
398 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
400 unsigned long end_bitidx
,
403 unsigned long *bitmap
;
404 unsigned long bitidx
, word_bitidx
;
407 bitmap
= get_pageblock_bitmap(page
, pfn
);
408 bitidx
= pfn_to_bitidx(page
, pfn
);
409 word_bitidx
= bitidx
/ BITS_PER_LONG
;
410 bitidx
&= (BITS_PER_LONG
-1);
412 word
= bitmap
[word_bitidx
];
413 bitidx
+= end_bitidx
;
414 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
417 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
418 unsigned long end_bitidx
,
421 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
424 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
426 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
430 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
431 * @page: The page within the block of interest
432 * @flags: The flags to set
433 * @pfn: The target page frame number
434 * @end_bitidx: The last bit of interest
435 * @mask: mask of bits that the caller is interested in
437 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
439 unsigned long end_bitidx
,
442 unsigned long *bitmap
;
443 unsigned long bitidx
, word_bitidx
;
444 unsigned long old_word
, word
;
446 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
448 bitmap
= get_pageblock_bitmap(page
, pfn
);
449 bitidx
= pfn_to_bitidx(page
, pfn
);
450 word_bitidx
= bitidx
/ BITS_PER_LONG
;
451 bitidx
&= (BITS_PER_LONG
-1);
453 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
455 bitidx
+= end_bitidx
;
456 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
457 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
459 word
= READ_ONCE(bitmap
[word_bitidx
]);
461 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
462 if (word
== old_word
)
468 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
470 if (unlikely(page_group_by_mobility_disabled
&&
471 migratetype
< MIGRATE_PCPTYPES
))
472 migratetype
= MIGRATE_UNMOVABLE
;
474 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
475 PB_migrate
, PB_migrate_end
);
478 #ifdef CONFIG_DEBUG_VM
479 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
483 unsigned long pfn
= page_to_pfn(page
);
484 unsigned long sp
, start_pfn
;
487 seq
= zone_span_seqbegin(zone
);
488 start_pfn
= zone
->zone_start_pfn
;
489 sp
= zone
->spanned_pages
;
490 if (!zone_spans_pfn(zone
, pfn
))
492 } while (zone_span_seqretry(zone
, seq
));
495 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
496 pfn
, zone_to_nid(zone
), zone
->name
,
497 start_pfn
, start_pfn
+ sp
);
502 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
504 if (!pfn_valid_within(page_to_pfn(page
)))
506 if (zone
!= page_zone(page
))
512 * Temporary debugging check for pages not lying within a given zone.
514 static int bad_range(struct zone
*zone
, struct page
*page
)
516 if (page_outside_zone_boundaries(zone
, page
))
518 if (!page_is_consistent(zone
, page
))
524 static inline int bad_range(struct zone
*zone
, struct page
*page
)
530 static void bad_page(struct page
*page
, const char *reason
,
531 unsigned long bad_flags
)
533 static unsigned long resume
;
534 static unsigned long nr_shown
;
535 static unsigned long nr_unshown
;
538 * Allow a burst of 60 reports, then keep quiet for that minute;
539 * or allow a steady drip of one report per second.
541 if (nr_shown
== 60) {
542 if (time_before(jiffies
, resume
)) {
548 "BUG: Bad page state: %lu messages suppressed\n",
555 resume
= jiffies
+ 60 * HZ
;
557 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
558 current
->comm
, page_to_pfn(page
));
559 __dump_page(page
, reason
);
560 bad_flags
&= page
->flags
;
562 pr_alert("bad because of flags: %#lx(%pGp)\n",
563 bad_flags
, &bad_flags
);
564 dump_page_owner(page
);
569 /* Leave bad fields for debug, except PageBuddy could make trouble */
570 page_mapcount_reset(page
); /* remove PageBuddy */
571 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
575 * Higher-order pages are called "compound pages". They are structured thusly:
577 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
579 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
580 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
582 * The first tail page's ->compound_dtor holds the offset in array of compound
583 * page destructors. See compound_page_dtors.
585 * The first tail page's ->compound_order holds the order of allocation.
586 * This usage means that zero-order pages may not be compound.
589 void free_compound_page(struct page
*page
)
591 __free_pages_ok(page
, compound_order(page
));
594 void prep_compound_page(struct page
*page
, unsigned int order
)
597 int nr_pages
= 1 << order
;
599 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
600 set_compound_order(page
, order
);
602 for (i
= 1; i
< nr_pages
; i
++) {
603 struct page
*p
= page
+ i
;
604 set_page_count(p
, 0);
605 p
->mapping
= TAIL_MAPPING
;
606 set_compound_head(p
, page
);
608 atomic_set(compound_mapcount_ptr(page
), -1);
611 #ifdef CONFIG_DEBUG_PAGEALLOC
612 unsigned int _debug_guardpage_minorder
;
613 bool _debug_pagealloc_enabled __read_mostly
614 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
615 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
616 bool _debug_guardpage_enabled __read_mostly
;
618 static int __init
early_debug_pagealloc(char *buf
)
622 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
624 early_param("debug_pagealloc", early_debug_pagealloc
);
626 static bool need_debug_guardpage(void)
628 /* If we don't use debug_pagealloc, we don't need guard page */
629 if (!debug_pagealloc_enabled())
632 if (!debug_guardpage_minorder())
638 static void init_debug_guardpage(void)
640 if (!debug_pagealloc_enabled())
643 if (!debug_guardpage_minorder())
646 _debug_guardpage_enabled
= true;
649 struct page_ext_operations debug_guardpage_ops
= {
650 .need
= need_debug_guardpage
,
651 .init
= init_debug_guardpage
,
654 static int __init
debug_guardpage_minorder_setup(char *buf
)
658 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
659 pr_err("Bad debug_guardpage_minorder value\n");
662 _debug_guardpage_minorder
= res
;
663 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
666 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
668 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
669 unsigned int order
, int migratetype
)
671 struct page_ext
*page_ext
;
673 if (!debug_guardpage_enabled())
676 if (order
>= debug_guardpage_minorder())
679 page_ext
= lookup_page_ext(page
);
680 if (unlikely(!page_ext
))
683 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
685 INIT_LIST_HEAD(&page
->lru
);
686 set_page_private(page
, order
);
687 /* Guard pages are not available for any usage */
688 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
693 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
694 unsigned int order
, int migratetype
)
696 struct page_ext
*page_ext
;
698 if (!debug_guardpage_enabled())
701 page_ext
= lookup_page_ext(page
);
702 if (unlikely(!page_ext
))
705 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
707 set_page_private(page
, 0);
708 if (!is_migrate_isolate(migratetype
))
709 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
712 struct page_ext_operations debug_guardpage_ops
;
713 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
714 unsigned int order
, int migratetype
) { return false; }
715 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
716 unsigned int order
, int migratetype
) {}
719 static inline void set_page_order(struct page
*page
, unsigned int order
)
721 set_page_private(page
, order
);
722 __SetPageBuddy(page
);
725 static inline void rmv_page_order(struct page
*page
)
727 __ClearPageBuddy(page
);
728 set_page_private(page
, 0);
732 * This function checks whether a page is free && is the buddy
733 * we can do coalesce a page and its buddy if
734 * (a) the buddy is not in a hole (check before calling!) &&
735 * (b) the buddy is in the buddy system &&
736 * (c) a page and its buddy have the same order &&
737 * (d) a page and its buddy are in the same zone.
739 * For recording whether a page is in the buddy system, we set ->_mapcount
740 * PAGE_BUDDY_MAPCOUNT_VALUE.
741 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
742 * serialized by zone->lock.
744 * For recording page's order, we use page_private(page).
746 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
749 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
750 if (page_zone_id(page
) != page_zone_id(buddy
))
753 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
758 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
760 * zone check is done late to avoid uselessly
761 * calculating zone/node ids for pages that could
764 if (page_zone_id(page
) != page_zone_id(buddy
))
767 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
775 * Freeing function for a buddy system allocator.
777 * The concept of a buddy system is to maintain direct-mapped table
778 * (containing bit values) for memory blocks of various "orders".
779 * The bottom level table contains the map for the smallest allocatable
780 * units of memory (here, pages), and each level above it describes
781 * pairs of units from the levels below, hence, "buddies".
782 * At a high level, all that happens here is marking the table entry
783 * at the bottom level available, and propagating the changes upward
784 * as necessary, plus some accounting needed to play nicely with other
785 * parts of the VM system.
786 * At each level, we keep a list of pages, which are heads of continuous
787 * free pages of length of (1 << order) and marked with _mapcount
788 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
790 * So when we are allocating or freeing one, we can derive the state of the
791 * other. That is, if we allocate a small block, and both were
792 * free, the remainder of the region must be split into blocks.
793 * If a block is freed, and its buddy is also free, then this
794 * triggers coalescing into a block of larger size.
799 static inline void __free_one_page(struct page
*page
,
801 struct zone
*zone
, unsigned int order
,
804 unsigned long combined_pfn
;
805 unsigned long uninitialized_var(buddy_pfn
);
807 unsigned int max_order
;
809 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
811 VM_BUG_ON(!zone_is_initialized(zone
));
812 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
814 VM_BUG_ON(migratetype
== -1);
815 if (likely(!is_migrate_isolate(migratetype
)))
816 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
818 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
819 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
822 while (order
< max_order
- 1) {
823 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
824 buddy
= page
+ (buddy_pfn
- pfn
);
826 if (!pfn_valid_within(buddy_pfn
))
828 if (!page_is_buddy(page
, buddy
, order
))
831 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
832 * merge with it and move up one order.
834 if (page_is_guard(buddy
)) {
835 clear_page_guard(zone
, buddy
, order
, migratetype
);
837 list_del(&buddy
->lru
);
838 zone
->free_area
[order
].nr_free
--;
839 rmv_page_order(buddy
);
841 combined_pfn
= buddy_pfn
& pfn
;
842 page
= page
+ (combined_pfn
- pfn
);
846 if (max_order
< MAX_ORDER
) {
847 /* If we are here, it means order is >= pageblock_order.
848 * We want to prevent merge between freepages on isolate
849 * pageblock and normal pageblock. Without this, pageblock
850 * isolation could cause incorrect freepage or CMA accounting.
852 * We don't want to hit this code for the more frequent
855 if (unlikely(has_isolate_pageblock(zone
))) {
858 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
859 buddy
= page
+ (buddy_pfn
- pfn
);
860 buddy_mt
= get_pageblock_migratetype(buddy
);
862 if (migratetype
!= buddy_mt
863 && (is_migrate_isolate(migratetype
) ||
864 is_migrate_isolate(buddy_mt
)))
868 goto continue_merging
;
872 set_page_order(page
, order
);
875 * If this is not the largest possible page, check if the buddy
876 * of the next-highest order is free. If it is, it's possible
877 * that pages are being freed that will coalesce soon. In case,
878 * that is happening, add the free page to the tail of the list
879 * so it's less likely to be used soon and more likely to be merged
880 * as a higher order page
882 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
883 struct page
*higher_page
, *higher_buddy
;
884 combined_pfn
= buddy_pfn
& pfn
;
885 higher_page
= page
+ (combined_pfn
- pfn
);
886 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
887 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
888 if (pfn_valid_within(buddy_pfn
) &&
889 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
890 list_add_tail(&page
->lru
,
891 &zone
->free_area
[order
].free_list
[migratetype
]);
896 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
898 zone
->free_area
[order
].nr_free
++;
902 * A bad page could be due to a number of fields. Instead of multiple branches,
903 * try and check multiple fields with one check. The caller must do a detailed
904 * check if necessary.
906 static inline bool page_expected_state(struct page
*page
,
907 unsigned long check_flags
)
909 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
912 if (unlikely((unsigned long)page
->mapping
|
913 page_ref_count(page
) |
915 (unsigned long)page
->mem_cgroup
|
917 (page
->flags
& check_flags
)))
923 static void free_pages_check_bad(struct page
*page
)
925 const char *bad_reason
;
926 unsigned long bad_flags
;
931 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
932 bad_reason
= "nonzero mapcount";
933 if (unlikely(page
->mapping
!= NULL
))
934 bad_reason
= "non-NULL mapping";
935 if (unlikely(page_ref_count(page
) != 0))
936 bad_reason
= "nonzero _refcount";
937 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
938 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
939 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
942 if (unlikely(page
->mem_cgroup
))
943 bad_reason
= "page still charged to cgroup";
945 bad_page(page
, bad_reason
, bad_flags
);
948 static inline int free_pages_check(struct page
*page
)
950 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
953 /* Something has gone sideways, find it */
954 free_pages_check_bad(page
);
958 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
963 * We rely page->lru.next never has bit 0 set, unless the page
964 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
966 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
968 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
972 switch (page
- head_page
) {
974 /* the first tail page: ->mapping is compound_mapcount() */
975 if (unlikely(compound_mapcount(page
))) {
976 bad_page(page
, "nonzero compound_mapcount", 0);
982 * the second tail page: ->mapping is
983 * page_deferred_list().next -- ignore value.
987 if (page
->mapping
!= TAIL_MAPPING
) {
988 bad_page(page
, "corrupted mapping in tail page", 0);
993 if (unlikely(!PageTail(page
))) {
994 bad_page(page
, "PageTail not set", 0);
997 if (unlikely(compound_head(page
) != head_page
)) {
998 bad_page(page
, "compound_head not consistent", 0);
1003 page
->mapping
= NULL
;
1004 clear_compound_head(page
);
1008 static __always_inline
bool free_pages_prepare(struct page
*page
,
1009 unsigned int order
, bool check_free
)
1013 VM_BUG_ON_PAGE(PageTail(page
), page
);
1015 trace_mm_page_free(page
, order
);
1016 kmemcheck_free_shadow(page
, order
);
1019 * Check tail pages before head page information is cleared to
1020 * avoid checking PageCompound for order-0 pages.
1022 if (unlikely(order
)) {
1023 bool compound
= PageCompound(page
);
1026 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1029 ClearPageDoubleMap(page
);
1030 for (i
= 1; i
< (1 << order
); i
++) {
1032 bad
+= free_tail_pages_check(page
, page
+ i
);
1033 if (unlikely(free_pages_check(page
+ i
))) {
1037 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1040 if (PageMappingFlags(page
))
1041 page
->mapping
= NULL
;
1042 if (memcg_kmem_enabled() && PageKmemcg(page
))
1043 memcg_kmem_uncharge(page
, order
);
1045 bad
+= free_pages_check(page
);
1049 page_cpupid_reset_last(page
);
1050 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1051 reset_page_owner(page
, order
);
1053 if (!PageHighMem(page
)) {
1054 debug_check_no_locks_freed(page_address(page
),
1055 PAGE_SIZE
<< order
);
1056 debug_check_no_obj_freed(page_address(page
),
1057 PAGE_SIZE
<< order
);
1059 arch_free_page(page
, order
);
1060 kernel_poison_pages(page
, 1 << order
, 0);
1061 kernel_map_pages(page
, 1 << order
, 0);
1062 kasan_free_pages(page
, order
);
1067 #ifdef CONFIG_DEBUG_VM
1068 static inline bool free_pcp_prepare(struct page
*page
)
1070 return free_pages_prepare(page
, 0, true);
1073 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1078 static bool free_pcp_prepare(struct page
*page
)
1080 return free_pages_prepare(page
, 0, false);
1083 static bool bulkfree_pcp_prepare(struct page
*page
)
1085 return free_pages_check(page
);
1087 #endif /* CONFIG_DEBUG_VM */
1090 * Frees a number of pages from the PCP lists
1091 * Assumes all pages on list are in same zone, and of same order.
1092 * count is the number of pages to free.
1094 * If the zone was previously in an "all pages pinned" state then look to
1095 * see if this freeing clears that state.
1097 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1098 * pinned" detection logic.
1100 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1101 struct per_cpu_pages
*pcp
)
1103 int migratetype
= 0;
1105 bool isolated_pageblocks
;
1107 spin_lock(&zone
->lock
);
1108 isolated_pageblocks
= has_isolate_pageblock(zone
);
1112 struct list_head
*list
;
1115 * Remove pages from lists in a round-robin fashion. A
1116 * batch_free count is maintained that is incremented when an
1117 * empty list is encountered. This is so more pages are freed
1118 * off fuller lists instead of spinning excessively around empty
1123 if (++migratetype
== MIGRATE_PCPTYPES
)
1125 list
= &pcp
->lists
[migratetype
];
1126 } while (list_empty(list
));
1128 /* This is the only non-empty list. Free them all. */
1129 if (batch_free
== MIGRATE_PCPTYPES
)
1133 int mt
; /* migratetype of the to-be-freed page */
1135 page
= list_last_entry(list
, struct page
, lru
);
1136 /* must delete as __free_one_page list manipulates */
1137 list_del(&page
->lru
);
1139 mt
= get_pcppage_migratetype(page
);
1140 /* MIGRATE_ISOLATE page should not go to pcplists */
1141 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1142 /* Pageblock could have been isolated meanwhile */
1143 if (unlikely(isolated_pageblocks
))
1144 mt
= get_pageblock_migratetype(page
);
1146 if (bulkfree_pcp_prepare(page
))
1149 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1150 trace_mm_page_pcpu_drain(page
, 0, mt
);
1151 } while (--count
&& --batch_free
&& !list_empty(list
));
1153 spin_unlock(&zone
->lock
);
1156 static void free_one_page(struct zone
*zone
,
1157 struct page
*page
, unsigned long pfn
,
1161 spin_lock(&zone
->lock
);
1162 if (unlikely(has_isolate_pageblock(zone
) ||
1163 is_migrate_isolate(migratetype
))) {
1164 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1166 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1167 spin_unlock(&zone
->lock
);
1170 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1171 unsigned long zone
, int nid
)
1173 set_page_links(page
, zone
, nid
, pfn
);
1174 init_page_count(page
);
1175 page_mapcount_reset(page
);
1176 page_cpupid_reset_last(page
);
1178 INIT_LIST_HEAD(&page
->lru
);
1179 #ifdef WANT_PAGE_VIRTUAL
1180 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1181 if (!is_highmem_idx(zone
))
1182 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1186 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1189 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1192 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1193 static void init_reserved_page(unsigned long pfn
)
1198 if (!early_page_uninitialised(pfn
))
1201 nid
= early_pfn_to_nid(pfn
);
1202 pgdat
= NODE_DATA(nid
);
1204 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1205 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1207 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1210 __init_single_pfn(pfn
, zid
, nid
);
1213 static inline void init_reserved_page(unsigned long pfn
)
1216 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1219 * Initialised pages do not have PageReserved set. This function is
1220 * called for each range allocated by the bootmem allocator and
1221 * marks the pages PageReserved. The remaining valid pages are later
1222 * sent to the buddy page allocator.
1224 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1226 unsigned long start_pfn
= PFN_DOWN(start
);
1227 unsigned long end_pfn
= PFN_UP(end
);
1229 for (; start_pfn
< end_pfn
; start_pfn
++) {
1230 if (pfn_valid(start_pfn
)) {
1231 struct page
*page
= pfn_to_page(start_pfn
);
1233 init_reserved_page(start_pfn
);
1235 /* Avoid false-positive PageTail() */
1236 INIT_LIST_HEAD(&page
->lru
);
1238 SetPageReserved(page
);
1243 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1245 unsigned long flags
;
1247 unsigned long pfn
= page_to_pfn(page
);
1249 if (!free_pages_prepare(page
, order
, true))
1252 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1253 local_irq_save(flags
);
1254 __count_vm_events(PGFREE
, 1 << order
);
1255 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1256 local_irq_restore(flags
);
1259 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1261 unsigned int nr_pages
= 1 << order
;
1262 struct page
*p
= page
;
1266 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1268 __ClearPageReserved(p
);
1269 set_page_count(p
, 0);
1271 __ClearPageReserved(p
);
1272 set_page_count(p
, 0);
1274 page_zone(page
)->managed_pages
+= nr_pages
;
1275 set_page_refcounted(page
);
1276 __free_pages(page
, order
);
1279 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1280 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1282 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1284 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1286 static DEFINE_SPINLOCK(early_pfn_lock
);
1289 spin_lock(&early_pfn_lock
);
1290 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1292 nid
= first_online_node
;
1293 spin_unlock(&early_pfn_lock
);
1299 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1300 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1301 struct mminit_pfnnid_cache
*state
)
1305 nid
= __early_pfn_to_nid(pfn
, state
);
1306 if (nid
>= 0 && nid
!= node
)
1311 /* Only safe to use early in boot when initialisation is single-threaded */
1312 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1314 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1319 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1323 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1324 struct mminit_pfnnid_cache
*state
)
1331 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1334 if (early_page_uninitialised(pfn
))
1336 return __free_pages_boot_core(page
, order
);
1340 * Check that the whole (or subset of) a pageblock given by the interval of
1341 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1342 * with the migration of free compaction scanner. The scanners then need to
1343 * use only pfn_valid_within() check for arches that allow holes within
1346 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1348 * It's possible on some configurations to have a setup like node0 node1 node0
1349 * i.e. it's possible that all pages within a zones range of pages do not
1350 * belong to a single zone. We assume that a border between node0 and node1
1351 * can occur within a single pageblock, but not a node0 node1 node0
1352 * interleaving within a single pageblock. It is therefore sufficient to check
1353 * the first and last page of a pageblock and avoid checking each individual
1354 * page in a pageblock.
1356 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1357 unsigned long end_pfn
, struct zone
*zone
)
1359 struct page
*start_page
;
1360 struct page
*end_page
;
1362 /* end_pfn is one past the range we are checking */
1365 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1368 start_page
= pfn_to_page(start_pfn
);
1370 if (page_zone(start_page
) != zone
)
1373 end_page
= pfn_to_page(end_pfn
);
1375 /* This gives a shorter code than deriving page_zone(end_page) */
1376 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1382 void set_zone_contiguous(struct zone
*zone
)
1384 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1385 unsigned long block_end_pfn
;
1387 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1388 for (; block_start_pfn
< zone_end_pfn(zone
);
1389 block_start_pfn
= block_end_pfn
,
1390 block_end_pfn
+= pageblock_nr_pages
) {
1392 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1394 if (!__pageblock_pfn_to_page(block_start_pfn
,
1395 block_end_pfn
, zone
))
1399 /* We confirm that there is no hole */
1400 zone
->contiguous
= true;
1403 void clear_zone_contiguous(struct zone
*zone
)
1405 zone
->contiguous
= false;
1408 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1409 static void __init
deferred_free_range(struct page
*page
,
1410 unsigned long pfn
, int nr_pages
)
1417 /* Free a large naturally-aligned chunk if possible */
1418 if (nr_pages
== pageblock_nr_pages
&&
1419 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1420 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1421 __free_pages_boot_core(page
, pageblock_order
);
1425 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1426 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1427 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1428 __free_pages_boot_core(page
, 0);
1432 /* Completion tracking for deferred_init_memmap() threads */
1433 static atomic_t pgdat_init_n_undone __initdata
;
1434 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1436 static inline void __init
pgdat_init_report_one_done(void)
1438 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1439 complete(&pgdat_init_all_done_comp
);
1442 /* Initialise remaining memory on a node */
1443 static int __init
deferred_init_memmap(void *data
)
1445 pg_data_t
*pgdat
= data
;
1446 int nid
= pgdat
->node_id
;
1447 struct mminit_pfnnid_cache nid_init_state
= { };
1448 unsigned long start
= jiffies
;
1449 unsigned long nr_pages
= 0;
1450 unsigned long walk_start
, walk_end
;
1453 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1454 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1456 if (first_init_pfn
== ULONG_MAX
) {
1457 pgdat_init_report_one_done();
1461 /* Bind memory initialisation thread to a local node if possible */
1462 if (!cpumask_empty(cpumask
))
1463 set_cpus_allowed_ptr(current
, cpumask
);
1465 /* Sanity check boundaries */
1466 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1467 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1468 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1470 /* Only the highest zone is deferred so find it */
1471 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1472 zone
= pgdat
->node_zones
+ zid
;
1473 if (first_init_pfn
< zone_end_pfn(zone
))
1477 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1478 unsigned long pfn
, end_pfn
;
1479 struct page
*page
= NULL
;
1480 struct page
*free_base_page
= NULL
;
1481 unsigned long free_base_pfn
= 0;
1484 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1485 pfn
= first_init_pfn
;
1486 if (pfn
< walk_start
)
1488 if (pfn
< zone
->zone_start_pfn
)
1489 pfn
= zone
->zone_start_pfn
;
1491 for (; pfn
< end_pfn
; pfn
++) {
1492 if (!pfn_valid_within(pfn
))
1496 * Ensure pfn_valid is checked every
1497 * pageblock_nr_pages for memory holes
1499 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1500 if (!pfn_valid(pfn
)) {
1506 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1511 /* Minimise pfn page lookups and scheduler checks */
1512 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1515 nr_pages
+= nr_to_free
;
1516 deferred_free_range(free_base_page
,
1517 free_base_pfn
, nr_to_free
);
1518 free_base_page
= NULL
;
1519 free_base_pfn
= nr_to_free
= 0;
1521 page
= pfn_to_page(pfn
);
1526 VM_BUG_ON(page_zone(page
) != zone
);
1530 __init_single_page(page
, pfn
, zid
, nid
);
1531 if (!free_base_page
) {
1532 free_base_page
= page
;
1533 free_base_pfn
= pfn
;
1538 /* Where possible, batch up pages for a single free */
1541 /* Free the current block of pages to allocator */
1542 nr_pages
+= nr_to_free
;
1543 deferred_free_range(free_base_page
, free_base_pfn
,
1545 free_base_page
= NULL
;
1546 free_base_pfn
= nr_to_free
= 0;
1548 /* Free the last block of pages to allocator */
1549 nr_pages
+= nr_to_free
;
1550 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1552 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1555 /* Sanity check that the next zone really is unpopulated */
1556 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1558 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1559 jiffies_to_msecs(jiffies
- start
));
1561 pgdat_init_report_one_done();
1564 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1566 void __init
page_alloc_init_late(void)
1570 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1573 /* There will be num_node_state(N_MEMORY) threads */
1574 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1575 for_each_node_state(nid
, N_MEMORY
) {
1576 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1579 /* Block until all are initialised */
1580 wait_for_completion(&pgdat_init_all_done_comp
);
1582 /* Reinit limits that are based on free pages after the kernel is up */
1583 files_maxfiles_init();
1586 for_each_populated_zone(zone
)
1587 set_zone_contiguous(zone
);
1591 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1592 void __init
init_cma_reserved_pageblock(struct page
*page
)
1594 unsigned i
= pageblock_nr_pages
;
1595 struct page
*p
= page
;
1598 __ClearPageReserved(p
);
1599 set_page_count(p
, 0);
1602 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1604 if (pageblock_order
>= MAX_ORDER
) {
1605 i
= pageblock_nr_pages
;
1608 set_page_refcounted(p
);
1609 __free_pages(p
, MAX_ORDER
- 1);
1610 p
+= MAX_ORDER_NR_PAGES
;
1611 } while (i
-= MAX_ORDER_NR_PAGES
);
1613 set_page_refcounted(page
);
1614 __free_pages(page
, pageblock_order
);
1617 adjust_managed_page_count(page
, pageblock_nr_pages
);
1622 * The order of subdivision here is critical for the IO subsystem.
1623 * Please do not alter this order without good reasons and regression
1624 * testing. Specifically, as large blocks of memory are subdivided,
1625 * the order in which smaller blocks are delivered depends on the order
1626 * they're subdivided in this function. This is the primary factor
1627 * influencing the order in which pages are delivered to the IO
1628 * subsystem according to empirical testing, and this is also justified
1629 * by considering the behavior of a buddy system containing a single
1630 * large block of memory acted on by a series of small allocations.
1631 * This behavior is a critical factor in sglist merging's success.
1635 static inline void expand(struct zone
*zone
, struct page
*page
,
1636 int low
, int high
, struct free_area
*area
,
1639 unsigned long size
= 1 << high
;
1641 while (high
> low
) {
1645 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1648 * Mark as guard pages (or page), that will allow to
1649 * merge back to allocator when buddy will be freed.
1650 * Corresponding page table entries will not be touched,
1651 * pages will stay not present in virtual address space
1653 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1656 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1658 set_page_order(&page
[size
], high
);
1662 static void check_new_page_bad(struct page
*page
)
1664 const char *bad_reason
= NULL
;
1665 unsigned long bad_flags
= 0;
1667 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1668 bad_reason
= "nonzero mapcount";
1669 if (unlikely(page
->mapping
!= NULL
))
1670 bad_reason
= "non-NULL mapping";
1671 if (unlikely(page_ref_count(page
) != 0))
1672 bad_reason
= "nonzero _count";
1673 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1674 bad_reason
= "HWPoisoned (hardware-corrupted)";
1675 bad_flags
= __PG_HWPOISON
;
1676 /* Don't complain about hwpoisoned pages */
1677 page_mapcount_reset(page
); /* remove PageBuddy */
1680 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1681 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1682 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1685 if (unlikely(page
->mem_cgroup
))
1686 bad_reason
= "page still charged to cgroup";
1688 bad_page(page
, bad_reason
, bad_flags
);
1692 * This page is about to be returned from the page allocator
1694 static inline int check_new_page(struct page
*page
)
1696 if (likely(page_expected_state(page
,
1697 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1700 check_new_page_bad(page
);
1704 static inline bool free_pages_prezeroed(void)
1706 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1707 page_poisoning_enabled();
1710 #ifdef CONFIG_DEBUG_VM
1711 static bool check_pcp_refill(struct page
*page
)
1716 static bool check_new_pcp(struct page
*page
)
1718 return check_new_page(page
);
1721 static bool check_pcp_refill(struct page
*page
)
1723 return check_new_page(page
);
1725 static bool check_new_pcp(struct page
*page
)
1729 #endif /* CONFIG_DEBUG_VM */
1731 static bool check_new_pages(struct page
*page
, unsigned int order
)
1734 for (i
= 0; i
< (1 << order
); i
++) {
1735 struct page
*p
= page
+ i
;
1737 if (unlikely(check_new_page(p
)))
1744 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1747 set_page_private(page
, 0);
1748 set_page_refcounted(page
);
1750 arch_alloc_page(page
, order
);
1751 kernel_map_pages(page
, 1 << order
, 1);
1752 kernel_poison_pages(page
, 1 << order
, 1);
1753 kasan_alloc_pages(page
, order
);
1754 set_page_owner(page
, order
, gfp_flags
);
1757 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1758 unsigned int alloc_flags
)
1762 post_alloc_hook(page
, order
, gfp_flags
);
1764 if (!free_pages_prezeroed() && (gfp_flags
& __GFP_ZERO
))
1765 for (i
= 0; i
< (1 << order
); i
++)
1766 clear_highpage(page
+ i
);
1768 if (order
&& (gfp_flags
& __GFP_COMP
))
1769 prep_compound_page(page
, order
);
1772 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1773 * allocate the page. The expectation is that the caller is taking
1774 * steps that will free more memory. The caller should avoid the page
1775 * being used for !PFMEMALLOC purposes.
1777 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1778 set_page_pfmemalloc(page
);
1780 clear_page_pfmemalloc(page
);
1784 * Go through the free lists for the given migratetype and remove
1785 * the smallest available page from the freelists
1788 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1791 unsigned int current_order
;
1792 struct free_area
*area
;
1795 /* Find a page of the appropriate size in the preferred list */
1796 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1797 area
= &(zone
->free_area
[current_order
]);
1798 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1802 list_del(&page
->lru
);
1803 rmv_page_order(page
);
1805 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1806 set_pcppage_migratetype(page
, migratetype
);
1815 * This array describes the order lists are fallen back to when
1816 * the free lists for the desirable migrate type are depleted
1818 static int fallbacks
[MIGRATE_TYPES
][4] = {
1819 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1820 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1821 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1823 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1825 #ifdef CONFIG_MEMORY_ISOLATION
1826 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1831 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1834 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1837 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1838 unsigned int order
) { return NULL
; }
1842 * Move the free pages in a range to the free lists of the requested type.
1843 * Note that start_page and end_pages are not aligned on a pageblock
1844 * boundary. If alignment is required, use move_freepages_block()
1846 static int move_freepages(struct zone
*zone
,
1847 struct page
*start_page
, struct page
*end_page
,
1848 int migratetype
, int *num_movable
)
1852 int pages_moved
= 0;
1854 #ifndef CONFIG_HOLES_IN_ZONE
1856 * page_zone is not safe to call in this context when
1857 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1858 * anyway as we check zone boundaries in move_freepages_block().
1859 * Remove at a later date when no bug reports exist related to
1860 * grouping pages by mobility
1862 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1868 for (page
= start_page
; page
<= end_page
;) {
1869 if (!pfn_valid_within(page_to_pfn(page
))) {
1874 /* Make sure we are not inadvertently changing nodes */
1875 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1877 if (!PageBuddy(page
)) {
1879 * We assume that pages that could be isolated for
1880 * migration are movable. But we don't actually try
1881 * isolating, as that would be expensive.
1884 (PageLRU(page
) || __PageMovable(page
)))
1891 order
= page_order(page
);
1892 list_move(&page
->lru
,
1893 &zone
->free_area
[order
].free_list
[migratetype
]);
1895 pages_moved
+= 1 << order
;
1901 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1902 int migratetype
, int *num_movable
)
1904 unsigned long start_pfn
, end_pfn
;
1905 struct page
*start_page
, *end_page
;
1907 start_pfn
= page_to_pfn(page
);
1908 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1909 start_page
= pfn_to_page(start_pfn
);
1910 end_page
= start_page
+ pageblock_nr_pages
- 1;
1911 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1913 /* Do not cross zone boundaries */
1914 if (!zone_spans_pfn(zone
, start_pfn
))
1916 if (!zone_spans_pfn(zone
, end_pfn
))
1919 return move_freepages(zone
, start_page
, end_page
, migratetype
,
1923 static void change_pageblock_range(struct page
*pageblock_page
,
1924 int start_order
, int migratetype
)
1926 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1928 while (nr_pageblocks
--) {
1929 set_pageblock_migratetype(pageblock_page
, migratetype
);
1930 pageblock_page
+= pageblock_nr_pages
;
1935 * When we are falling back to another migratetype during allocation, try to
1936 * steal extra free pages from the same pageblocks to satisfy further
1937 * allocations, instead of polluting multiple pageblocks.
1939 * If we are stealing a relatively large buddy page, it is likely there will
1940 * be more free pages in the pageblock, so try to steal them all. For
1941 * reclaimable and unmovable allocations, we steal regardless of page size,
1942 * as fragmentation caused by those allocations polluting movable pageblocks
1943 * is worse than movable allocations stealing from unmovable and reclaimable
1946 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1949 * Leaving this order check is intended, although there is
1950 * relaxed order check in next check. The reason is that
1951 * we can actually steal whole pageblock if this condition met,
1952 * but, below check doesn't guarantee it and that is just heuristic
1953 * so could be changed anytime.
1955 if (order
>= pageblock_order
)
1958 if (order
>= pageblock_order
/ 2 ||
1959 start_mt
== MIGRATE_RECLAIMABLE
||
1960 start_mt
== MIGRATE_UNMOVABLE
||
1961 page_group_by_mobility_disabled
)
1968 * This function implements actual steal behaviour. If order is large enough,
1969 * we can steal whole pageblock. If not, we first move freepages in this
1970 * pageblock to our migratetype and determine how many already-allocated pages
1971 * are there in the pageblock with a compatible migratetype. If at least half
1972 * of pages are free or compatible, we can change migratetype of the pageblock
1973 * itself, so pages freed in the future will be put on the correct free list.
1975 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1976 int start_type
, bool whole_block
)
1978 unsigned int current_order
= page_order(page
);
1979 struct free_area
*area
;
1980 int free_pages
, movable_pages
, alike_pages
;
1983 old_block_type
= get_pageblock_migratetype(page
);
1986 * This can happen due to races and we want to prevent broken
1987 * highatomic accounting.
1989 if (is_migrate_highatomic(old_block_type
))
1992 /* Take ownership for orders >= pageblock_order */
1993 if (current_order
>= pageblock_order
) {
1994 change_pageblock_range(page
, current_order
, start_type
);
1998 /* We are not allowed to try stealing from the whole block */
2002 free_pages
= move_freepages_block(zone
, page
, start_type
,
2005 * Determine how many pages are compatible with our allocation.
2006 * For movable allocation, it's the number of movable pages which
2007 * we just obtained. For other types it's a bit more tricky.
2009 if (start_type
== MIGRATE_MOVABLE
) {
2010 alike_pages
= movable_pages
;
2013 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2014 * to MOVABLE pageblock, consider all non-movable pages as
2015 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2016 * vice versa, be conservative since we can't distinguish the
2017 * exact migratetype of non-movable pages.
2019 if (old_block_type
== MIGRATE_MOVABLE
)
2020 alike_pages
= pageblock_nr_pages
2021 - (free_pages
+ movable_pages
);
2026 /* moving whole block can fail due to zone boundary conditions */
2031 * If a sufficient number of pages in the block are either free or of
2032 * comparable migratability as our allocation, claim the whole block.
2034 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2035 page_group_by_mobility_disabled
)
2036 set_pageblock_migratetype(page
, start_type
);
2041 area
= &zone
->free_area
[current_order
];
2042 list_move(&page
->lru
, &area
->free_list
[start_type
]);
2046 * Check whether there is a suitable fallback freepage with requested order.
2047 * If only_stealable is true, this function returns fallback_mt only if
2048 * we can steal other freepages all together. This would help to reduce
2049 * fragmentation due to mixed migratetype pages in one pageblock.
2051 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2052 int migratetype
, bool only_stealable
, bool *can_steal
)
2057 if (area
->nr_free
== 0)
2062 fallback_mt
= fallbacks
[migratetype
][i
];
2063 if (fallback_mt
== MIGRATE_TYPES
)
2066 if (list_empty(&area
->free_list
[fallback_mt
]))
2069 if (can_steal_fallback(order
, migratetype
))
2072 if (!only_stealable
)
2083 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2084 * there are no empty page blocks that contain a page with a suitable order
2086 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2087 unsigned int alloc_order
)
2090 unsigned long max_managed
, flags
;
2093 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2094 * Check is race-prone but harmless.
2096 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2097 if (zone
->nr_reserved_highatomic
>= max_managed
)
2100 spin_lock_irqsave(&zone
->lock
, flags
);
2102 /* Recheck the nr_reserved_highatomic limit under the lock */
2103 if (zone
->nr_reserved_highatomic
>= max_managed
)
2107 mt
= get_pageblock_migratetype(page
);
2108 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2109 && !is_migrate_cma(mt
)) {
2110 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2111 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2112 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2116 spin_unlock_irqrestore(&zone
->lock
, flags
);
2120 * Used when an allocation is about to fail under memory pressure. This
2121 * potentially hurts the reliability of high-order allocations when under
2122 * intense memory pressure but failed atomic allocations should be easier
2123 * to recover from than an OOM.
2125 * If @force is true, try to unreserve a pageblock even though highatomic
2126 * pageblock is exhausted.
2128 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2131 struct zonelist
*zonelist
= ac
->zonelist
;
2132 unsigned long flags
;
2139 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2142 * Preserve at least one pageblock unless memory pressure
2145 if (!force
&& zone
->nr_reserved_highatomic
<=
2149 spin_lock_irqsave(&zone
->lock
, flags
);
2150 for (order
= 0; order
< MAX_ORDER
; order
++) {
2151 struct free_area
*area
= &(zone
->free_area
[order
]);
2153 page
= list_first_entry_or_null(
2154 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2160 * In page freeing path, migratetype change is racy so
2161 * we can counter several free pages in a pageblock
2162 * in this loop althoug we changed the pageblock type
2163 * from highatomic to ac->migratetype. So we should
2164 * adjust the count once.
2166 if (is_migrate_highatomic_page(page
)) {
2168 * It should never happen but changes to
2169 * locking could inadvertently allow a per-cpu
2170 * drain to add pages to MIGRATE_HIGHATOMIC
2171 * while unreserving so be safe and watch for
2174 zone
->nr_reserved_highatomic
-= min(
2176 zone
->nr_reserved_highatomic
);
2180 * Convert to ac->migratetype and avoid the normal
2181 * pageblock stealing heuristics. Minimally, the caller
2182 * is doing the work and needs the pages. More
2183 * importantly, if the block was always converted to
2184 * MIGRATE_UNMOVABLE or another type then the number
2185 * of pageblocks that cannot be completely freed
2188 set_pageblock_migratetype(page
, ac
->migratetype
);
2189 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2192 spin_unlock_irqrestore(&zone
->lock
, flags
);
2196 spin_unlock_irqrestore(&zone
->lock
, flags
);
2203 * Try finding a free buddy page on the fallback list and put it on the free
2204 * list of requested migratetype, possibly along with other pages from the same
2205 * block, depending on fragmentation avoidance heuristics. Returns true if
2206 * fallback was found so that __rmqueue_smallest() can grab it.
2209 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2211 struct free_area
*area
;
2212 unsigned int current_order
;
2217 /* Find the largest possible block of pages in the other list */
2218 for (current_order
= MAX_ORDER
-1;
2219 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2221 area
= &(zone
->free_area
[current_order
]);
2222 fallback_mt
= find_suitable_fallback(area
, current_order
,
2223 start_migratetype
, false, &can_steal
);
2224 if (fallback_mt
== -1)
2227 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2230 steal_suitable_fallback(zone
, page
, start_migratetype
,
2233 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2234 start_migratetype
, fallback_mt
);
2243 * Do the hard work of removing an element from the buddy allocator.
2244 * Call me with the zone->lock already held.
2246 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2252 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2253 if (unlikely(!page
)) {
2254 if (migratetype
== MIGRATE_MOVABLE
)
2255 page
= __rmqueue_cma_fallback(zone
, order
);
2257 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
))
2261 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2266 * Obtain a specified number of elements from the buddy allocator, all under
2267 * a single hold of the lock, for efficiency. Add them to the supplied list.
2268 * Returns the number of new pages which were placed at *list.
2270 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2271 unsigned long count
, struct list_head
*list
,
2272 int migratetype
, bool cold
)
2276 spin_lock(&zone
->lock
);
2277 for (i
= 0; i
< count
; ++i
) {
2278 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2279 if (unlikely(page
== NULL
))
2282 if (unlikely(check_pcp_refill(page
)))
2286 * Split buddy pages returned by expand() are received here
2287 * in physical page order. The page is added to the callers and
2288 * list and the list head then moves forward. From the callers
2289 * perspective, the linked list is ordered by page number in
2290 * some conditions. This is useful for IO devices that can
2291 * merge IO requests if the physical pages are ordered
2295 list_add(&page
->lru
, list
);
2297 list_add_tail(&page
->lru
, list
);
2300 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2301 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2306 * i pages were removed from the buddy list even if some leak due
2307 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2308 * on i. Do not confuse with 'alloced' which is the number of
2309 * pages added to the pcp list.
2311 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2312 spin_unlock(&zone
->lock
);
2318 * Called from the vmstat counter updater to drain pagesets of this
2319 * currently executing processor on remote nodes after they have
2322 * Note that this function must be called with the thread pinned to
2323 * a single processor.
2325 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2327 unsigned long flags
;
2328 int to_drain
, batch
;
2330 local_irq_save(flags
);
2331 batch
= READ_ONCE(pcp
->batch
);
2332 to_drain
= min(pcp
->count
, batch
);
2334 free_pcppages_bulk(zone
, to_drain
, pcp
);
2335 pcp
->count
-= to_drain
;
2337 local_irq_restore(flags
);
2342 * Drain pcplists of the indicated processor and zone.
2344 * The processor must either be the current processor and the
2345 * thread pinned to the current processor or a processor that
2348 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2350 unsigned long flags
;
2351 struct per_cpu_pageset
*pset
;
2352 struct per_cpu_pages
*pcp
;
2354 local_irq_save(flags
);
2355 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2359 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2362 local_irq_restore(flags
);
2366 * Drain pcplists of all zones on the indicated processor.
2368 * The processor must either be the current processor and the
2369 * thread pinned to the current processor or a processor that
2372 static void drain_pages(unsigned int cpu
)
2376 for_each_populated_zone(zone
) {
2377 drain_pages_zone(cpu
, zone
);
2382 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2384 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2385 * the single zone's pages.
2387 void drain_local_pages(struct zone
*zone
)
2389 int cpu
= smp_processor_id();
2392 drain_pages_zone(cpu
, zone
);
2397 static void drain_local_pages_wq(struct work_struct
*work
)
2400 * drain_all_pages doesn't use proper cpu hotplug protection so
2401 * we can race with cpu offline when the WQ can move this from
2402 * a cpu pinned worker to an unbound one. We can operate on a different
2403 * cpu which is allright but we also have to make sure to not move to
2407 drain_local_pages(NULL
);
2412 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2414 * When zone parameter is non-NULL, spill just the single zone's pages.
2416 * Note that this can be extremely slow as the draining happens in a workqueue.
2418 void drain_all_pages(struct zone
*zone
)
2423 * Allocate in the BSS so we wont require allocation in
2424 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2426 static cpumask_t cpus_with_pcps
;
2429 * Make sure nobody triggers this path before mm_percpu_wq is fully
2432 if (WARN_ON_ONCE(!mm_percpu_wq
))
2435 /* Workqueues cannot recurse */
2436 if (current
->flags
& PF_WQ_WORKER
)
2440 * Do not drain if one is already in progress unless it's specific to
2441 * a zone. Such callers are primarily CMA and memory hotplug and need
2442 * the drain to be complete when the call returns.
2444 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2447 mutex_lock(&pcpu_drain_mutex
);
2451 * We don't care about racing with CPU hotplug event
2452 * as offline notification will cause the notified
2453 * cpu to drain that CPU pcps and on_each_cpu_mask
2454 * disables preemption as part of its processing
2456 for_each_online_cpu(cpu
) {
2457 struct per_cpu_pageset
*pcp
;
2459 bool has_pcps
= false;
2462 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2466 for_each_populated_zone(z
) {
2467 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2468 if (pcp
->pcp
.count
) {
2476 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2478 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2481 for_each_cpu(cpu
, &cpus_with_pcps
) {
2482 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2483 INIT_WORK(work
, drain_local_pages_wq
);
2484 queue_work_on(cpu
, mm_percpu_wq
, work
);
2486 for_each_cpu(cpu
, &cpus_with_pcps
)
2487 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2489 mutex_unlock(&pcpu_drain_mutex
);
2492 #ifdef CONFIG_HIBERNATION
2494 void mark_free_pages(struct zone
*zone
)
2496 unsigned long pfn
, max_zone_pfn
;
2497 unsigned long flags
;
2498 unsigned int order
, t
;
2501 if (zone_is_empty(zone
))
2504 spin_lock_irqsave(&zone
->lock
, flags
);
2506 max_zone_pfn
= zone_end_pfn(zone
);
2507 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2508 if (pfn_valid(pfn
)) {
2509 page
= pfn_to_page(pfn
);
2511 if (page_zone(page
) != zone
)
2514 if (!swsusp_page_is_forbidden(page
))
2515 swsusp_unset_page_free(page
);
2518 for_each_migratetype_order(order
, t
) {
2519 list_for_each_entry(page
,
2520 &zone
->free_area
[order
].free_list
[t
], lru
) {
2523 pfn
= page_to_pfn(page
);
2524 for (i
= 0; i
< (1UL << order
); i
++)
2525 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2528 spin_unlock_irqrestore(&zone
->lock
, flags
);
2530 #endif /* CONFIG_PM */
2533 * Free a 0-order page
2534 * cold == true ? free a cold page : free a hot page
2536 void free_hot_cold_page(struct page
*page
, bool cold
)
2538 struct zone
*zone
= page_zone(page
);
2539 struct per_cpu_pages
*pcp
;
2540 unsigned long flags
;
2541 unsigned long pfn
= page_to_pfn(page
);
2544 if (!free_pcp_prepare(page
))
2547 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2548 set_pcppage_migratetype(page
, migratetype
);
2549 local_irq_save(flags
);
2550 __count_vm_event(PGFREE
);
2553 * We only track unmovable, reclaimable and movable on pcp lists.
2554 * Free ISOLATE pages back to the allocator because they are being
2555 * offlined but treat HIGHATOMIC as movable pages so we can get those
2556 * areas back if necessary. Otherwise, we may have to free
2557 * excessively into the page allocator
2559 if (migratetype
>= MIGRATE_PCPTYPES
) {
2560 if (unlikely(is_migrate_isolate(migratetype
))) {
2561 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2564 migratetype
= MIGRATE_MOVABLE
;
2567 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2569 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2571 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2573 if (pcp
->count
>= pcp
->high
) {
2574 unsigned long batch
= READ_ONCE(pcp
->batch
);
2575 free_pcppages_bulk(zone
, batch
, pcp
);
2576 pcp
->count
-= batch
;
2580 local_irq_restore(flags
);
2584 * Free a list of 0-order pages
2586 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2588 struct page
*page
, *next
;
2590 list_for_each_entry_safe(page
, next
, list
, lru
) {
2591 trace_mm_page_free_batched(page
, cold
);
2592 free_hot_cold_page(page
, cold
);
2597 * split_page takes a non-compound higher-order page, and splits it into
2598 * n (1<<order) sub-pages: page[0..n]
2599 * Each sub-page must be freed individually.
2601 * Note: this is probably too low level an operation for use in drivers.
2602 * Please consult with lkml before using this in your driver.
2604 void split_page(struct page
*page
, unsigned int order
)
2608 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2609 VM_BUG_ON_PAGE(!page_count(page
), page
);
2611 #ifdef CONFIG_KMEMCHECK
2613 * Split shadow pages too, because free(page[0]) would
2614 * otherwise free the whole shadow.
2616 if (kmemcheck_page_is_tracked(page
))
2617 split_page(virt_to_page(page
[0].shadow
), order
);
2620 for (i
= 1; i
< (1 << order
); i
++)
2621 set_page_refcounted(page
+ i
);
2622 split_page_owner(page
, order
);
2624 EXPORT_SYMBOL_GPL(split_page
);
2626 int __isolate_free_page(struct page
*page
, unsigned int order
)
2628 unsigned long watermark
;
2632 BUG_ON(!PageBuddy(page
));
2634 zone
= page_zone(page
);
2635 mt
= get_pageblock_migratetype(page
);
2637 if (!is_migrate_isolate(mt
)) {
2639 * Obey watermarks as if the page was being allocated. We can
2640 * emulate a high-order watermark check with a raised order-0
2641 * watermark, because we already know our high-order page
2644 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2645 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2648 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2651 /* Remove page from free list */
2652 list_del(&page
->lru
);
2653 zone
->free_area
[order
].nr_free
--;
2654 rmv_page_order(page
);
2657 * Set the pageblock if the isolated page is at least half of a
2660 if (order
>= pageblock_order
- 1) {
2661 struct page
*endpage
= page
+ (1 << order
) - 1;
2662 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2663 int mt
= get_pageblock_migratetype(page
);
2664 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2665 && !is_migrate_highatomic(mt
))
2666 set_pageblock_migratetype(page
,
2672 return 1UL << order
;
2676 * Update NUMA hit/miss statistics
2678 * Must be called with interrupts disabled.
2680 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2683 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2685 if (z
->node
!= numa_node_id())
2686 local_stat
= NUMA_OTHER
;
2688 if (z
->node
== preferred_zone
->node
)
2689 __inc_zone_state(z
, NUMA_HIT
);
2691 __inc_zone_state(z
, NUMA_MISS
);
2692 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2694 __inc_zone_state(z
, local_stat
);
2698 /* Remove page from the per-cpu list, caller must protect the list */
2699 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2700 bool cold
, struct per_cpu_pages
*pcp
,
2701 struct list_head
*list
)
2706 if (list_empty(list
)) {
2707 pcp
->count
+= rmqueue_bulk(zone
, 0,
2710 if (unlikely(list_empty(list
)))
2715 page
= list_last_entry(list
, struct page
, lru
);
2717 page
= list_first_entry(list
, struct page
, lru
);
2719 list_del(&page
->lru
);
2721 } while (check_new_pcp(page
));
2726 /* Lock and remove page from the per-cpu list */
2727 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2728 struct zone
*zone
, unsigned int order
,
2729 gfp_t gfp_flags
, int migratetype
)
2731 struct per_cpu_pages
*pcp
;
2732 struct list_head
*list
;
2733 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2735 unsigned long flags
;
2737 local_irq_save(flags
);
2738 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2739 list
= &pcp
->lists
[migratetype
];
2740 page
= __rmqueue_pcplist(zone
, migratetype
, cold
, pcp
, list
);
2742 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2743 zone_statistics(preferred_zone
, zone
);
2745 local_irq_restore(flags
);
2750 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2753 struct page
*rmqueue(struct zone
*preferred_zone
,
2754 struct zone
*zone
, unsigned int order
,
2755 gfp_t gfp_flags
, unsigned int alloc_flags
,
2758 unsigned long flags
;
2761 if (likely(order
== 0)) {
2762 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
2763 gfp_flags
, migratetype
);
2768 * We most definitely don't want callers attempting to
2769 * allocate greater than order-1 page units with __GFP_NOFAIL.
2771 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2772 spin_lock_irqsave(&zone
->lock
, flags
);
2776 if (alloc_flags
& ALLOC_HARDER
) {
2777 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2779 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2782 page
= __rmqueue(zone
, order
, migratetype
);
2783 } while (page
&& check_new_pages(page
, order
));
2784 spin_unlock(&zone
->lock
);
2787 __mod_zone_freepage_state(zone
, -(1 << order
),
2788 get_pcppage_migratetype(page
));
2790 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2791 zone_statistics(preferred_zone
, zone
);
2792 local_irq_restore(flags
);
2795 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
2799 local_irq_restore(flags
);
2803 #ifdef CONFIG_FAIL_PAGE_ALLOC
2806 struct fault_attr attr
;
2808 bool ignore_gfp_highmem
;
2809 bool ignore_gfp_reclaim
;
2811 } fail_page_alloc
= {
2812 .attr
= FAULT_ATTR_INITIALIZER
,
2813 .ignore_gfp_reclaim
= true,
2814 .ignore_gfp_highmem
= true,
2818 static int __init
setup_fail_page_alloc(char *str
)
2820 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2822 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2824 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2826 if (order
< fail_page_alloc
.min_order
)
2828 if (gfp_mask
& __GFP_NOFAIL
)
2830 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2832 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2833 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2836 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2839 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2841 static int __init
fail_page_alloc_debugfs(void)
2843 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2846 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2847 &fail_page_alloc
.attr
);
2849 return PTR_ERR(dir
);
2851 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2852 &fail_page_alloc
.ignore_gfp_reclaim
))
2854 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2855 &fail_page_alloc
.ignore_gfp_highmem
))
2857 if (!debugfs_create_u32("min-order", mode
, dir
,
2858 &fail_page_alloc
.min_order
))
2863 debugfs_remove_recursive(dir
);
2868 late_initcall(fail_page_alloc_debugfs
);
2870 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2872 #else /* CONFIG_FAIL_PAGE_ALLOC */
2874 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2879 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2882 * Return true if free base pages are above 'mark'. For high-order checks it
2883 * will return true of the order-0 watermark is reached and there is at least
2884 * one free page of a suitable size. Checking now avoids taking the zone lock
2885 * to check in the allocation paths if no pages are free.
2887 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2888 int classzone_idx
, unsigned int alloc_flags
,
2893 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2895 /* free_pages may go negative - that's OK */
2896 free_pages
-= (1 << order
) - 1;
2898 if (alloc_flags
& ALLOC_HIGH
)
2902 * If the caller does not have rights to ALLOC_HARDER then subtract
2903 * the high-atomic reserves. This will over-estimate the size of the
2904 * atomic reserve but it avoids a search.
2906 if (likely(!alloc_harder
))
2907 free_pages
-= z
->nr_reserved_highatomic
;
2912 /* If allocation can't use CMA areas don't use free CMA pages */
2913 if (!(alloc_flags
& ALLOC_CMA
))
2914 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2918 * Check watermarks for an order-0 allocation request. If these
2919 * are not met, then a high-order request also cannot go ahead
2920 * even if a suitable page happened to be free.
2922 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2925 /* If this is an order-0 request then the watermark is fine */
2929 /* For a high-order request, check at least one suitable page is free */
2930 for (o
= order
; o
< MAX_ORDER
; o
++) {
2931 struct free_area
*area
= &z
->free_area
[o
];
2940 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2941 if (!list_empty(&area
->free_list
[mt
]))
2946 if ((alloc_flags
& ALLOC_CMA
) &&
2947 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2955 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2956 int classzone_idx
, unsigned int alloc_flags
)
2958 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2959 zone_page_state(z
, NR_FREE_PAGES
));
2962 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2963 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2965 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2969 /* If allocation can't use CMA areas don't use free CMA pages */
2970 if (!(alloc_flags
& ALLOC_CMA
))
2971 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2975 * Fast check for order-0 only. If this fails then the reserves
2976 * need to be calculated. There is a corner case where the check
2977 * passes but only the high-order atomic reserve are free. If
2978 * the caller is !atomic then it'll uselessly search the free
2979 * list. That corner case is then slower but it is harmless.
2981 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2984 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2988 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2989 unsigned long mark
, int classzone_idx
)
2991 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2993 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2994 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2996 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3001 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3003 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3006 #else /* CONFIG_NUMA */
3007 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3011 #endif /* CONFIG_NUMA */
3014 * get_page_from_freelist goes through the zonelist trying to allocate
3017 static struct page
*
3018 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3019 const struct alloc_context
*ac
)
3021 struct zoneref
*z
= ac
->preferred_zoneref
;
3023 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3026 * Scan zonelist, looking for a zone with enough free.
3027 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3029 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3034 if (cpusets_enabled() &&
3035 (alloc_flags
& ALLOC_CPUSET
) &&
3036 !__cpuset_zone_allowed(zone
, gfp_mask
))
3039 * When allocating a page cache page for writing, we
3040 * want to get it from a node that is within its dirty
3041 * limit, such that no single node holds more than its
3042 * proportional share of globally allowed dirty pages.
3043 * The dirty limits take into account the node's
3044 * lowmem reserves and high watermark so that kswapd
3045 * should be able to balance it without having to
3046 * write pages from its LRU list.
3048 * XXX: For now, allow allocations to potentially
3049 * exceed the per-node dirty limit in the slowpath
3050 * (spread_dirty_pages unset) before going into reclaim,
3051 * which is important when on a NUMA setup the allowed
3052 * nodes are together not big enough to reach the
3053 * global limit. The proper fix for these situations
3054 * will require awareness of nodes in the
3055 * dirty-throttling and the flusher threads.
3057 if (ac
->spread_dirty_pages
) {
3058 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3061 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3062 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3067 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
3068 if (!zone_watermark_fast(zone
, order
, mark
,
3069 ac_classzone_idx(ac
), alloc_flags
)) {
3072 /* Checked here to keep the fast path fast */
3073 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3074 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3077 if (node_reclaim_mode
== 0 ||
3078 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3081 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3083 case NODE_RECLAIM_NOSCAN
:
3086 case NODE_RECLAIM_FULL
:
3087 /* scanned but unreclaimable */
3090 /* did we reclaim enough */
3091 if (zone_watermark_ok(zone
, order
, mark
,
3092 ac_classzone_idx(ac
), alloc_flags
))
3100 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3101 gfp_mask
, alloc_flags
, ac
->migratetype
);
3103 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3106 * If this is a high-order atomic allocation then check
3107 * if the pageblock should be reserved for the future
3109 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3110 reserve_highatomic_pageblock(page
, zone
, order
);
3120 * Large machines with many possible nodes should not always dump per-node
3121 * meminfo in irq context.
3123 static inline bool should_suppress_show_mem(void)
3128 ret
= in_interrupt();
3133 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3135 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3136 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3138 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs
))
3142 * This documents exceptions given to allocations in certain
3143 * contexts that are allowed to allocate outside current's set
3146 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3147 if (test_thread_flag(TIF_MEMDIE
) ||
3148 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3149 filter
&= ~SHOW_MEM_FILTER_NODES
;
3150 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3151 filter
&= ~SHOW_MEM_FILTER_NODES
;
3153 show_mem(filter
, nodemask
);
3156 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3158 struct va_format vaf
;
3160 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3161 DEFAULT_RATELIMIT_BURST
);
3163 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3166 pr_warn("%s: ", current
->comm
);
3168 va_start(args
, fmt
);
3171 pr_cont("%pV", &vaf
);
3174 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask
, &gfp_mask
);
3176 pr_cont("%*pbl\n", nodemask_pr_args(nodemask
));
3178 pr_cont("(null)\n");
3180 cpuset_print_current_mems_allowed();
3183 warn_alloc_show_mem(gfp_mask
, nodemask
);
3186 static inline struct page
*
3187 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3188 unsigned int alloc_flags
,
3189 const struct alloc_context
*ac
)
3193 page
= get_page_from_freelist(gfp_mask
, order
,
3194 alloc_flags
|ALLOC_CPUSET
, ac
);
3196 * fallback to ignore cpuset restriction if our nodes
3200 page
= get_page_from_freelist(gfp_mask
, order
,
3206 static inline struct page
*
3207 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3208 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3210 struct oom_control oc
= {
3211 .zonelist
= ac
->zonelist
,
3212 .nodemask
= ac
->nodemask
,
3214 .gfp_mask
= gfp_mask
,
3219 *did_some_progress
= 0;
3222 * Acquire the oom lock. If that fails, somebody else is
3223 * making progress for us.
3225 if (!mutex_trylock(&oom_lock
)) {
3226 *did_some_progress
= 1;
3227 schedule_timeout_uninterruptible(1);
3232 * Go through the zonelist yet one more time, keep very high watermark
3233 * here, this is only to catch a parallel oom killing, we must fail if
3234 * we're still under heavy pressure.
3236 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3237 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3241 /* Coredumps can quickly deplete all memory reserves */
3242 if (current
->flags
& PF_DUMPCORE
)
3244 /* The OOM killer will not help higher order allocs */
3245 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3247 /* The OOM killer does not needlessly kill tasks for lowmem */
3248 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3250 if (pm_suspended_storage())
3253 * XXX: GFP_NOFS allocations should rather fail than rely on
3254 * other request to make a forward progress.
3255 * We are in an unfortunate situation where out_of_memory cannot
3256 * do much for this context but let's try it to at least get
3257 * access to memory reserved if the current task is killed (see
3258 * out_of_memory). Once filesystems are ready to handle allocation
3259 * failures more gracefully we should just bail out here.
3262 /* The OOM killer may not free memory on a specific node */
3263 if (gfp_mask
& __GFP_THISNODE
)
3266 /* Exhausted what can be done so it's blamo time */
3267 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3268 *did_some_progress
= 1;
3271 * Help non-failing allocations by giving them access to memory
3274 if (gfp_mask
& __GFP_NOFAIL
)
3275 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3276 ALLOC_NO_WATERMARKS
, ac
);
3279 mutex_unlock(&oom_lock
);
3284 * Maximum number of compaction retries wit a progress before OOM
3285 * killer is consider as the only way to move forward.
3287 #define MAX_COMPACT_RETRIES 16
3289 #ifdef CONFIG_COMPACTION
3290 /* Try memory compaction for high-order allocations before reclaim */
3291 static struct page
*
3292 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3293 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3294 enum compact_priority prio
, enum compact_result
*compact_result
)
3297 unsigned int noreclaim_flag
;
3302 noreclaim_flag
= memalloc_noreclaim_save();
3303 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3305 memalloc_noreclaim_restore(noreclaim_flag
);
3307 if (*compact_result
<= COMPACT_INACTIVE
)
3311 * At least in one zone compaction wasn't deferred or skipped, so let's
3312 * count a compaction stall
3314 count_vm_event(COMPACTSTALL
);
3316 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3319 struct zone
*zone
= page_zone(page
);
3321 zone
->compact_blockskip_flush
= false;
3322 compaction_defer_reset(zone
, order
, true);
3323 count_vm_event(COMPACTSUCCESS
);
3328 * It's bad if compaction run occurs and fails. The most likely reason
3329 * is that pages exist, but not enough to satisfy watermarks.
3331 count_vm_event(COMPACTFAIL
);
3339 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3340 enum compact_result compact_result
,
3341 enum compact_priority
*compact_priority
,
3342 int *compaction_retries
)
3344 int max_retries
= MAX_COMPACT_RETRIES
;
3347 int retries
= *compaction_retries
;
3348 enum compact_priority priority
= *compact_priority
;
3353 if (compaction_made_progress(compact_result
))
3354 (*compaction_retries
)++;
3357 * compaction considers all the zone as desperately out of memory
3358 * so it doesn't really make much sense to retry except when the
3359 * failure could be caused by insufficient priority
3361 if (compaction_failed(compact_result
))
3362 goto check_priority
;
3365 * make sure the compaction wasn't deferred or didn't bail out early
3366 * due to locks contention before we declare that we should give up.
3367 * But do not retry if the given zonelist is not suitable for
3370 if (compaction_withdrawn(compact_result
)) {
3371 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3376 * !costly requests are much more important than __GFP_REPEAT
3377 * costly ones because they are de facto nofail and invoke OOM
3378 * killer to move on while costly can fail and users are ready
3379 * to cope with that. 1/4 retries is rather arbitrary but we
3380 * would need much more detailed feedback from compaction to
3381 * make a better decision.
3383 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3385 if (*compaction_retries
<= max_retries
) {
3391 * Make sure there are attempts at the highest priority if we exhausted
3392 * all retries or failed at the lower priorities.
3395 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3396 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3398 if (*compact_priority
> min_priority
) {
3399 (*compact_priority
)--;
3400 *compaction_retries
= 0;
3404 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3408 static inline struct page
*
3409 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3410 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3411 enum compact_priority prio
, enum compact_result
*compact_result
)
3413 *compact_result
= COMPACT_SKIPPED
;
3418 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3419 enum compact_result compact_result
,
3420 enum compact_priority
*compact_priority
,
3421 int *compaction_retries
)
3426 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3430 * There are setups with compaction disabled which would prefer to loop
3431 * inside the allocator rather than hit the oom killer prematurely.
3432 * Let's give them a good hope and keep retrying while the order-0
3433 * watermarks are OK.
3435 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3437 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3438 ac_classzone_idx(ac
), alloc_flags
))
3443 #endif /* CONFIG_COMPACTION */
3445 /* Perform direct synchronous page reclaim */
3447 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3448 const struct alloc_context
*ac
)
3450 struct reclaim_state reclaim_state
;
3452 unsigned int noreclaim_flag
;
3456 /* We now go into synchronous reclaim */
3457 cpuset_memory_pressure_bump();
3458 noreclaim_flag
= memalloc_noreclaim_save();
3459 lockdep_set_current_reclaim_state(gfp_mask
);
3460 reclaim_state
.reclaimed_slab
= 0;
3461 current
->reclaim_state
= &reclaim_state
;
3463 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3466 current
->reclaim_state
= NULL
;
3467 lockdep_clear_current_reclaim_state();
3468 memalloc_noreclaim_restore(noreclaim_flag
);
3475 /* The really slow allocator path where we enter direct reclaim */
3476 static inline struct page
*
3477 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3478 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3479 unsigned long *did_some_progress
)
3481 struct page
*page
= NULL
;
3482 bool drained
= false;
3484 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3485 if (unlikely(!(*did_some_progress
)))
3489 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3492 * If an allocation failed after direct reclaim, it could be because
3493 * pages are pinned on the per-cpu lists or in high alloc reserves.
3494 * Shrink them them and try again
3496 if (!page
&& !drained
) {
3497 unreserve_highatomic_pageblock(ac
, false);
3498 drain_all_pages(NULL
);
3506 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3510 pg_data_t
*last_pgdat
= NULL
;
3512 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3513 ac
->high_zoneidx
, ac
->nodemask
) {
3514 if (last_pgdat
!= zone
->zone_pgdat
)
3515 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3516 last_pgdat
= zone
->zone_pgdat
;
3520 static inline unsigned int
3521 gfp_to_alloc_flags(gfp_t gfp_mask
)
3523 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3525 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3526 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3529 * The caller may dip into page reserves a bit more if the caller
3530 * cannot run direct reclaim, or if the caller has realtime scheduling
3531 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3532 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3534 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3536 if (gfp_mask
& __GFP_ATOMIC
) {
3538 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3539 * if it can't schedule.
3541 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3542 alloc_flags
|= ALLOC_HARDER
;
3544 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3545 * comment for __cpuset_node_allowed().
3547 alloc_flags
&= ~ALLOC_CPUSET
;
3548 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3549 alloc_flags
|= ALLOC_HARDER
;
3552 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3553 alloc_flags
|= ALLOC_CMA
;
3558 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3560 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3563 if (gfp_mask
& __GFP_MEMALLOC
)
3565 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3567 if (!in_interrupt() &&
3568 ((current
->flags
& PF_MEMALLOC
) ||
3569 unlikely(test_thread_flag(TIF_MEMDIE
))))
3576 * Checks whether it makes sense to retry the reclaim to make a forward progress
3577 * for the given allocation request.
3579 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3580 * without success, or when we couldn't even meet the watermark if we
3581 * reclaimed all remaining pages on the LRU lists.
3583 * Returns true if a retry is viable or false to enter the oom path.
3586 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3587 struct alloc_context
*ac
, int alloc_flags
,
3588 bool did_some_progress
, int *no_progress_loops
)
3594 * Costly allocations might have made a progress but this doesn't mean
3595 * their order will become available due to high fragmentation so
3596 * always increment the no progress counter for them
3598 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3599 *no_progress_loops
= 0;
3601 (*no_progress_loops
)++;
3604 * Make sure we converge to OOM if we cannot make any progress
3605 * several times in the row.
3607 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3608 /* Before OOM, exhaust highatomic_reserve */
3609 return unreserve_highatomic_pageblock(ac
, true);
3613 * Keep reclaiming pages while there is a chance this will lead
3614 * somewhere. If none of the target zones can satisfy our allocation
3615 * request even if all reclaimable pages are considered then we are
3616 * screwed and have to go OOM.
3618 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3620 unsigned long available
;
3621 unsigned long reclaimable
;
3622 unsigned long min_wmark
= min_wmark_pages(zone
);
3625 available
= reclaimable
= zone_reclaimable_pages(zone
);
3626 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3629 * Would the allocation succeed if we reclaimed all
3630 * reclaimable pages?
3632 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3633 ac_classzone_idx(ac
), alloc_flags
, available
);
3634 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3635 available
, min_wmark
, *no_progress_loops
, wmark
);
3638 * If we didn't make any progress and have a lot of
3639 * dirty + writeback pages then we should wait for
3640 * an IO to complete to slow down the reclaim and
3641 * prevent from pre mature OOM
3643 if (!did_some_progress
) {
3644 unsigned long write_pending
;
3646 write_pending
= zone_page_state_snapshot(zone
,
3647 NR_ZONE_WRITE_PENDING
);
3649 if (2 * write_pending
> reclaimable
) {
3650 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3656 * Memory allocation/reclaim might be called from a WQ
3657 * context and the current implementation of the WQ
3658 * concurrency control doesn't recognize that
3659 * a particular WQ is congested if the worker thread is
3660 * looping without ever sleeping. Therefore we have to
3661 * do a short sleep here rather than calling
3664 if (current
->flags
& PF_WQ_WORKER
)
3665 schedule_timeout_uninterruptible(1);
3676 static inline struct page
*
3677 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3678 struct alloc_context
*ac
)
3680 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3681 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
3682 struct page
*page
= NULL
;
3683 unsigned int alloc_flags
;
3684 unsigned long did_some_progress
;
3685 enum compact_priority compact_priority
;
3686 enum compact_result compact_result
;
3687 int compaction_retries
;
3688 int no_progress_loops
;
3689 unsigned long alloc_start
= jiffies
;
3690 unsigned int stall_timeout
= 10 * HZ
;
3691 unsigned int cpuset_mems_cookie
;
3694 * In the slowpath, we sanity check order to avoid ever trying to
3695 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3696 * be using allocators in order of preference for an area that is
3699 if (order
>= MAX_ORDER
) {
3700 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3705 * We also sanity check to catch abuse of atomic reserves being used by
3706 * callers that are not in atomic context.
3708 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3709 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3710 gfp_mask
&= ~__GFP_ATOMIC
;
3713 compaction_retries
= 0;
3714 no_progress_loops
= 0;
3715 compact_priority
= DEF_COMPACT_PRIORITY
;
3716 cpuset_mems_cookie
= read_mems_allowed_begin();
3719 * The fast path uses conservative alloc_flags to succeed only until
3720 * kswapd needs to be woken up, and to avoid the cost of setting up
3721 * alloc_flags precisely. So we do that now.
3723 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3726 * We need to recalculate the starting point for the zonelist iterator
3727 * because we might have used different nodemask in the fast path, or
3728 * there was a cpuset modification and we are retrying - otherwise we
3729 * could end up iterating over non-eligible zones endlessly.
3731 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3732 ac
->high_zoneidx
, ac
->nodemask
);
3733 if (!ac
->preferred_zoneref
->zone
)
3736 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3737 wake_all_kswapds(order
, ac
);
3740 * The adjusted alloc_flags might result in immediate success, so try
3743 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3748 * For costly allocations, try direct compaction first, as it's likely
3749 * that we have enough base pages and don't need to reclaim. For non-
3750 * movable high-order allocations, do that as well, as compaction will
3751 * try prevent permanent fragmentation by migrating from blocks of the
3753 * Don't try this for allocations that are allowed to ignore
3754 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3756 if (can_direct_reclaim
&&
3758 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
3759 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
3760 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3762 INIT_COMPACT_PRIORITY
,
3768 * Checks for costly allocations with __GFP_NORETRY, which
3769 * includes THP page fault allocations
3771 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
3773 * If compaction is deferred for high-order allocations,
3774 * it is because sync compaction recently failed. If
3775 * this is the case and the caller requested a THP
3776 * allocation, we do not want to heavily disrupt the
3777 * system, so we fail the allocation instead of entering
3780 if (compact_result
== COMPACT_DEFERRED
)
3784 * Looks like reclaim/compaction is worth trying, but
3785 * sync compaction could be very expensive, so keep
3786 * using async compaction.
3788 compact_priority
= INIT_COMPACT_PRIORITY
;
3793 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3794 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3795 wake_all_kswapds(order
, ac
);
3797 if (gfp_pfmemalloc_allowed(gfp_mask
))
3798 alloc_flags
= ALLOC_NO_WATERMARKS
;
3801 * Reset the zonelist iterators if memory policies can be ignored.
3802 * These allocations are high priority and system rather than user
3805 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3806 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3807 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3808 ac
->high_zoneidx
, ac
->nodemask
);
3811 /* Attempt with potentially adjusted zonelist and alloc_flags */
3812 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3816 /* Caller is not willing to reclaim, we can't balance anything */
3817 if (!can_direct_reclaim
)
3820 /* Make sure we know about allocations which stall for too long */
3821 if (time_after(jiffies
, alloc_start
+ stall_timeout
)) {
3822 warn_alloc(gfp_mask
& ~__GFP_NOWARN
, ac
->nodemask
,
3823 "page allocation stalls for %ums, order:%u",
3824 jiffies_to_msecs(jiffies
-alloc_start
), order
);
3825 stall_timeout
+= 10 * HZ
;
3828 /* Avoid recursion of direct reclaim */
3829 if (current
->flags
& PF_MEMALLOC
)
3832 /* Try direct reclaim and then allocating */
3833 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3834 &did_some_progress
);
3838 /* Try direct compaction and then allocating */
3839 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3840 compact_priority
, &compact_result
);
3844 /* Do not loop if specifically requested */
3845 if (gfp_mask
& __GFP_NORETRY
)
3849 * Do not retry costly high order allocations unless they are
3852 if (costly_order
&& !(gfp_mask
& __GFP_REPEAT
))
3855 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3856 did_some_progress
> 0, &no_progress_loops
))
3860 * It doesn't make any sense to retry for the compaction if the order-0
3861 * reclaim is not able to make any progress because the current
3862 * implementation of the compaction depends on the sufficient amount
3863 * of free memory (see __compaction_suitable)
3865 if (did_some_progress
> 0 &&
3866 should_compact_retry(ac
, order
, alloc_flags
,
3867 compact_result
, &compact_priority
,
3868 &compaction_retries
))
3872 * It's possible we raced with cpuset update so the OOM would be
3873 * premature (see below the nopage: label for full explanation).
3875 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3878 /* Reclaim has failed us, start killing things */
3879 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3883 /* Avoid allocations with no watermarks from looping endlessly */
3884 if (test_thread_flag(TIF_MEMDIE
) &&
3885 (alloc_flags
== ALLOC_NO_WATERMARKS
||
3886 (gfp_mask
& __GFP_NOMEMALLOC
)))
3889 /* Retry as long as the OOM killer is making progress */
3890 if (did_some_progress
) {
3891 no_progress_loops
= 0;
3897 * When updating a task's mems_allowed or mempolicy nodemask, it is
3898 * possible to race with parallel threads in such a way that our
3899 * allocation can fail while the mask is being updated. If we are about
3900 * to fail, check if the cpuset changed during allocation and if so,
3903 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3907 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3910 if (gfp_mask
& __GFP_NOFAIL
) {
3912 * All existing users of the __GFP_NOFAIL are blockable, so warn
3913 * of any new users that actually require GFP_NOWAIT
3915 if (WARN_ON_ONCE(!can_direct_reclaim
))
3919 * PF_MEMALLOC request from this context is rather bizarre
3920 * because we cannot reclaim anything and only can loop waiting
3921 * for somebody to do a work for us
3923 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
3926 * non failing costly orders are a hard requirement which we
3927 * are not prepared for much so let's warn about these users
3928 * so that we can identify them and convert them to something
3931 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
3934 * Help non-failing allocations by giving them access to memory
3935 * reserves but do not use ALLOC_NO_WATERMARKS because this
3936 * could deplete whole memory reserves which would just make
3937 * the situation worse
3939 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
3947 warn_alloc(gfp_mask
, ac
->nodemask
,
3948 "page allocation failure: order:%u", order
);
3953 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
3954 struct zonelist
*zonelist
, nodemask_t
*nodemask
,
3955 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
3956 unsigned int *alloc_flags
)
3958 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
3959 ac
->zonelist
= zonelist
;
3960 ac
->nodemask
= nodemask
;
3961 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
3963 if (cpusets_enabled()) {
3964 *alloc_mask
|= __GFP_HARDWALL
;
3966 ac
->nodemask
= &cpuset_current_mems_allowed
;
3968 *alloc_flags
|= ALLOC_CPUSET
;
3971 lockdep_trace_alloc(gfp_mask
);
3973 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3975 if (should_fail_alloc_page(gfp_mask
, order
))
3978 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
3979 *alloc_flags
|= ALLOC_CMA
;
3984 /* Determine whether to spread dirty pages and what the first usable zone */
3985 static inline void finalise_ac(gfp_t gfp_mask
,
3986 unsigned int order
, struct alloc_context
*ac
)
3988 /* Dirty zone balancing only done in the fast path */
3989 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3992 * The preferred zone is used for statistics but crucially it is
3993 * also used as the starting point for the zonelist iterator. It
3994 * may get reset for allocations that ignore memory policies.
3996 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3997 ac
->high_zoneidx
, ac
->nodemask
);
4001 * This is the 'heart' of the zoned buddy allocator.
4004 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
4005 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
4008 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4009 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
4010 struct alloc_context ac
= { };
4012 gfp_mask
&= gfp_allowed_mask
;
4013 if (!prepare_alloc_pages(gfp_mask
, order
, zonelist
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4016 finalise_ac(gfp_mask
, order
, &ac
);
4018 /* First allocation attempt */
4019 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4024 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4025 * resp. GFP_NOIO which has to be inherited for all allocation requests
4026 * from a particular context which has been marked by
4027 * memalloc_no{fs,io}_{save,restore}.
4029 alloc_mask
= current_gfp_context(gfp_mask
);
4030 ac
.spread_dirty_pages
= false;
4033 * Restore the original nodemask if it was potentially replaced with
4034 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4036 if (unlikely(ac
.nodemask
!= nodemask
))
4037 ac
.nodemask
= nodemask
;
4039 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4042 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4043 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4044 __free_pages(page
, order
);
4048 if (kmemcheck_enabled
&& page
)
4049 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
4051 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4055 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4058 * Common helper functions.
4060 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4065 * __get_free_pages() returns a 32-bit address, which cannot represent
4068 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
4070 page
= alloc_pages(gfp_mask
, order
);
4073 return (unsigned long) page_address(page
);
4075 EXPORT_SYMBOL(__get_free_pages
);
4077 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4079 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4081 EXPORT_SYMBOL(get_zeroed_page
);
4083 void __free_pages(struct page
*page
, unsigned int order
)
4085 if (put_page_testzero(page
)) {
4087 free_hot_cold_page(page
, false);
4089 __free_pages_ok(page
, order
);
4093 EXPORT_SYMBOL(__free_pages
);
4095 void free_pages(unsigned long addr
, unsigned int order
)
4098 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4099 __free_pages(virt_to_page((void *)addr
), order
);
4103 EXPORT_SYMBOL(free_pages
);
4107 * An arbitrary-length arbitrary-offset area of memory which resides
4108 * within a 0 or higher order page. Multiple fragments within that page
4109 * are individually refcounted, in the page's reference counter.
4111 * The page_frag functions below provide a simple allocation framework for
4112 * page fragments. This is used by the network stack and network device
4113 * drivers to provide a backing region of memory for use as either an
4114 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4116 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4119 struct page
*page
= NULL
;
4120 gfp_t gfp
= gfp_mask
;
4122 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4123 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4125 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4126 PAGE_FRAG_CACHE_MAX_ORDER
);
4127 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4129 if (unlikely(!page
))
4130 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4132 nc
->va
= page
? page_address(page
) : NULL
;
4137 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4139 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4141 if (page_ref_sub_and_test(page
, count
)) {
4142 unsigned int order
= compound_order(page
);
4145 free_hot_cold_page(page
, false);
4147 __free_pages_ok(page
, order
);
4150 EXPORT_SYMBOL(__page_frag_cache_drain
);
4152 void *page_frag_alloc(struct page_frag_cache
*nc
,
4153 unsigned int fragsz
, gfp_t gfp_mask
)
4155 unsigned int size
= PAGE_SIZE
;
4159 if (unlikely(!nc
->va
)) {
4161 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4165 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4166 /* if size can vary use size else just use PAGE_SIZE */
4169 /* Even if we own the page, we do not use atomic_set().
4170 * This would break get_page_unless_zero() users.
4172 page_ref_add(page
, size
- 1);
4174 /* reset page count bias and offset to start of new frag */
4175 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4176 nc
->pagecnt_bias
= size
;
4180 offset
= nc
->offset
- fragsz
;
4181 if (unlikely(offset
< 0)) {
4182 page
= virt_to_page(nc
->va
);
4184 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4187 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4188 /* if size can vary use size else just use PAGE_SIZE */
4191 /* OK, page count is 0, we can safely set it */
4192 set_page_count(page
, size
);
4194 /* reset page count bias and offset to start of new frag */
4195 nc
->pagecnt_bias
= size
;
4196 offset
= size
- fragsz
;
4200 nc
->offset
= offset
;
4202 return nc
->va
+ offset
;
4204 EXPORT_SYMBOL(page_frag_alloc
);
4207 * Frees a page fragment allocated out of either a compound or order 0 page.
4209 void page_frag_free(void *addr
)
4211 struct page
*page
= virt_to_head_page(addr
);
4213 if (unlikely(put_page_testzero(page
)))
4214 __free_pages_ok(page
, compound_order(page
));
4216 EXPORT_SYMBOL(page_frag_free
);
4218 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4222 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4223 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4225 split_page(virt_to_page((void *)addr
), order
);
4226 while (used
< alloc_end
) {
4231 return (void *)addr
;
4235 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4236 * @size: the number of bytes to allocate
4237 * @gfp_mask: GFP flags for the allocation
4239 * This function is similar to alloc_pages(), except that it allocates the
4240 * minimum number of pages to satisfy the request. alloc_pages() can only
4241 * allocate memory in power-of-two pages.
4243 * This function is also limited by MAX_ORDER.
4245 * Memory allocated by this function must be released by free_pages_exact().
4247 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4249 unsigned int order
= get_order(size
);
4252 addr
= __get_free_pages(gfp_mask
, order
);
4253 return make_alloc_exact(addr
, order
, size
);
4255 EXPORT_SYMBOL(alloc_pages_exact
);
4258 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4260 * @nid: the preferred node ID where memory should be allocated
4261 * @size: the number of bytes to allocate
4262 * @gfp_mask: GFP flags for the allocation
4264 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4267 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4269 unsigned int order
= get_order(size
);
4270 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4273 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4277 * free_pages_exact - release memory allocated via alloc_pages_exact()
4278 * @virt: the value returned by alloc_pages_exact.
4279 * @size: size of allocation, same value as passed to alloc_pages_exact().
4281 * Release the memory allocated by a previous call to alloc_pages_exact.
4283 void free_pages_exact(void *virt
, size_t size
)
4285 unsigned long addr
= (unsigned long)virt
;
4286 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4288 while (addr
< end
) {
4293 EXPORT_SYMBOL(free_pages_exact
);
4296 * nr_free_zone_pages - count number of pages beyond high watermark
4297 * @offset: The zone index of the highest zone
4299 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4300 * high watermark within all zones at or below a given zone index. For each
4301 * zone, the number of pages is calculated as:
4303 * nr_free_zone_pages = managed_pages - high_pages
4305 static unsigned long nr_free_zone_pages(int offset
)
4310 /* Just pick one node, since fallback list is circular */
4311 unsigned long sum
= 0;
4313 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4315 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4316 unsigned long size
= zone
->managed_pages
;
4317 unsigned long high
= high_wmark_pages(zone
);
4326 * nr_free_buffer_pages - count number of pages beyond high watermark
4328 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4329 * watermark within ZONE_DMA and ZONE_NORMAL.
4331 unsigned long nr_free_buffer_pages(void)
4333 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4335 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4338 * nr_free_pagecache_pages - count number of pages beyond high watermark
4340 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4341 * high watermark within all zones.
4343 unsigned long nr_free_pagecache_pages(void)
4345 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4348 static inline void show_node(struct zone
*zone
)
4350 if (IS_ENABLED(CONFIG_NUMA
))
4351 printk("Node %d ", zone_to_nid(zone
));
4354 long si_mem_available(void)
4357 unsigned long pagecache
;
4358 unsigned long wmark_low
= 0;
4359 unsigned long pages
[NR_LRU_LISTS
];
4363 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4364 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4367 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4370 * Estimate the amount of memory available for userspace allocations,
4371 * without causing swapping.
4373 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4376 * Not all the page cache can be freed, otherwise the system will
4377 * start swapping. Assume at least half of the page cache, or the
4378 * low watermark worth of cache, needs to stay.
4380 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4381 pagecache
-= min(pagecache
/ 2, wmark_low
);
4382 available
+= pagecache
;
4385 * Part of the reclaimable slab consists of items that are in use,
4386 * and cannot be freed. Cap this estimate at the low watermark.
4388 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4389 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4395 EXPORT_SYMBOL_GPL(si_mem_available
);
4397 void si_meminfo(struct sysinfo
*val
)
4399 val
->totalram
= totalram_pages
;
4400 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4401 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4402 val
->bufferram
= nr_blockdev_pages();
4403 val
->totalhigh
= totalhigh_pages
;
4404 val
->freehigh
= nr_free_highpages();
4405 val
->mem_unit
= PAGE_SIZE
;
4408 EXPORT_SYMBOL(si_meminfo
);
4411 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4413 int zone_type
; /* needs to be signed */
4414 unsigned long managed_pages
= 0;
4415 unsigned long managed_highpages
= 0;
4416 unsigned long free_highpages
= 0;
4417 pg_data_t
*pgdat
= NODE_DATA(nid
);
4419 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4420 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4421 val
->totalram
= managed_pages
;
4422 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4423 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4424 #ifdef CONFIG_HIGHMEM
4425 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4426 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4428 if (is_highmem(zone
)) {
4429 managed_highpages
+= zone
->managed_pages
;
4430 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4433 val
->totalhigh
= managed_highpages
;
4434 val
->freehigh
= free_highpages
;
4436 val
->totalhigh
= managed_highpages
;
4437 val
->freehigh
= free_highpages
;
4439 val
->mem_unit
= PAGE_SIZE
;
4444 * Determine whether the node should be displayed or not, depending on whether
4445 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4447 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4449 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4453 * no node mask - aka implicit memory numa policy. Do not bother with
4454 * the synchronization - read_mems_allowed_begin - because we do not
4455 * have to be precise here.
4458 nodemask
= &cpuset_current_mems_allowed
;
4460 return !node_isset(nid
, *nodemask
);
4463 #define K(x) ((x) << (PAGE_SHIFT-10))
4465 static void show_migration_types(unsigned char type
)
4467 static const char types
[MIGRATE_TYPES
] = {
4468 [MIGRATE_UNMOVABLE
] = 'U',
4469 [MIGRATE_MOVABLE
] = 'M',
4470 [MIGRATE_RECLAIMABLE
] = 'E',
4471 [MIGRATE_HIGHATOMIC
] = 'H',
4473 [MIGRATE_CMA
] = 'C',
4475 #ifdef CONFIG_MEMORY_ISOLATION
4476 [MIGRATE_ISOLATE
] = 'I',
4479 char tmp
[MIGRATE_TYPES
+ 1];
4483 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4484 if (type
& (1 << i
))
4489 printk(KERN_CONT
"(%s) ", tmp
);
4493 * Show free area list (used inside shift_scroll-lock stuff)
4494 * We also calculate the percentage fragmentation. We do this by counting the
4495 * memory on each free list with the exception of the first item on the list.
4498 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4501 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4503 unsigned long free_pcp
= 0;
4508 for_each_populated_zone(zone
) {
4509 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4512 for_each_online_cpu(cpu
)
4513 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4516 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4517 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4518 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4519 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4520 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4521 " free:%lu free_pcp:%lu free_cma:%lu\n",
4522 global_node_page_state(NR_ACTIVE_ANON
),
4523 global_node_page_state(NR_INACTIVE_ANON
),
4524 global_node_page_state(NR_ISOLATED_ANON
),
4525 global_node_page_state(NR_ACTIVE_FILE
),
4526 global_node_page_state(NR_INACTIVE_FILE
),
4527 global_node_page_state(NR_ISOLATED_FILE
),
4528 global_node_page_state(NR_UNEVICTABLE
),
4529 global_node_page_state(NR_FILE_DIRTY
),
4530 global_node_page_state(NR_WRITEBACK
),
4531 global_node_page_state(NR_UNSTABLE_NFS
),
4532 global_page_state(NR_SLAB_RECLAIMABLE
),
4533 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4534 global_node_page_state(NR_FILE_MAPPED
),
4535 global_node_page_state(NR_SHMEM
),
4536 global_page_state(NR_PAGETABLE
),
4537 global_page_state(NR_BOUNCE
),
4538 global_page_state(NR_FREE_PAGES
),
4540 global_page_state(NR_FREE_CMA_PAGES
));
4542 for_each_online_pgdat(pgdat
) {
4543 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
4547 " active_anon:%lukB"
4548 " inactive_anon:%lukB"
4549 " active_file:%lukB"
4550 " inactive_file:%lukB"
4551 " unevictable:%lukB"
4552 " isolated(anon):%lukB"
4553 " isolated(file):%lukB"
4558 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4560 " shmem_pmdmapped: %lukB"
4563 " writeback_tmp:%lukB"
4565 " all_unreclaimable? %s"
4568 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4569 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4570 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4571 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4572 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4573 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4574 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4575 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4576 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4577 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4578 K(node_page_state(pgdat
, NR_SHMEM
)),
4579 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4580 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4581 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4583 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4585 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4586 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4587 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
4591 for_each_populated_zone(zone
) {
4594 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4598 for_each_online_cpu(cpu
)
4599 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4608 " active_anon:%lukB"
4609 " inactive_anon:%lukB"
4610 " active_file:%lukB"
4611 " inactive_file:%lukB"
4612 " unevictable:%lukB"
4613 " writepending:%lukB"
4617 " slab_reclaimable:%lukB"
4618 " slab_unreclaimable:%lukB"
4619 " kernel_stack:%lukB"
4627 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4628 K(min_wmark_pages(zone
)),
4629 K(low_wmark_pages(zone
)),
4630 K(high_wmark_pages(zone
)),
4631 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4632 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4633 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4634 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4635 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4636 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4637 K(zone
->present_pages
),
4638 K(zone
->managed_pages
),
4639 K(zone_page_state(zone
, NR_MLOCK
)),
4640 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4641 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4642 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4643 K(zone_page_state(zone
, NR_PAGETABLE
)),
4644 K(zone_page_state(zone
, NR_BOUNCE
)),
4646 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4647 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4648 printk("lowmem_reserve[]:");
4649 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4650 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
4651 printk(KERN_CONT
"\n");
4654 for_each_populated_zone(zone
) {
4656 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4657 unsigned char types
[MAX_ORDER
];
4659 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4662 printk(KERN_CONT
"%s: ", zone
->name
);
4664 spin_lock_irqsave(&zone
->lock
, flags
);
4665 for (order
= 0; order
< MAX_ORDER
; order
++) {
4666 struct free_area
*area
= &zone
->free_area
[order
];
4669 nr
[order
] = area
->nr_free
;
4670 total
+= nr
[order
] << order
;
4673 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4674 if (!list_empty(&area
->free_list
[type
]))
4675 types
[order
] |= 1 << type
;
4678 spin_unlock_irqrestore(&zone
->lock
, flags
);
4679 for (order
= 0; order
< MAX_ORDER
; order
++) {
4680 printk(KERN_CONT
"%lu*%lukB ",
4681 nr
[order
], K(1UL) << order
);
4683 show_migration_types(types
[order
]);
4685 printk(KERN_CONT
"= %lukB\n", K(total
));
4688 hugetlb_show_meminfo();
4690 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4692 show_swap_cache_info();
4695 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4697 zoneref
->zone
= zone
;
4698 zoneref
->zone_idx
= zone_idx(zone
);
4702 * Builds allocation fallback zone lists.
4704 * Add all populated zones of a node to the zonelist.
4706 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4710 enum zone_type zone_type
= MAX_NR_ZONES
;
4714 zone
= pgdat
->node_zones
+ zone_type
;
4715 if (managed_zone(zone
)) {
4716 zoneref_set_zone(zone
,
4717 &zonelist
->_zonerefs
[nr_zones
++]);
4718 check_highest_zone(zone_type
);
4720 } while (zone_type
);
4728 * 0 = automatic detection of better ordering.
4729 * 1 = order by ([node] distance, -zonetype)
4730 * 2 = order by (-zonetype, [node] distance)
4732 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4733 * the same zonelist. So only NUMA can configure this param.
4735 #define ZONELIST_ORDER_DEFAULT 0
4736 #define ZONELIST_ORDER_NODE 1
4737 #define ZONELIST_ORDER_ZONE 2
4739 /* zonelist order in the kernel.
4740 * set_zonelist_order() will set this to NODE or ZONE.
4742 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4743 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4747 /* The value user specified ....changed by config */
4748 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4749 /* string for sysctl */
4750 #define NUMA_ZONELIST_ORDER_LEN 16
4751 char numa_zonelist_order
[16] = "default";
4754 * interface for configure zonelist ordering.
4755 * command line option "numa_zonelist_order"
4756 * = "[dD]efault - default, automatic configuration.
4757 * = "[nN]ode - order by node locality, then by zone within node
4758 * = "[zZ]one - order by zone, then by locality within zone
4761 static int __parse_numa_zonelist_order(char *s
)
4763 if (*s
== 'd' || *s
== 'D') {
4764 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4765 } else if (*s
== 'n' || *s
== 'N') {
4766 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4767 } else if (*s
== 'z' || *s
== 'Z') {
4768 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4770 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4776 static __init
int setup_numa_zonelist_order(char *s
)
4783 ret
= __parse_numa_zonelist_order(s
);
4785 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4789 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4792 * sysctl handler for numa_zonelist_order
4794 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4795 void __user
*buffer
, size_t *length
,
4798 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4800 static DEFINE_MUTEX(zl_order_mutex
);
4802 mutex_lock(&zl_order_mutex
);
4804 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4808 strcpy(saved_string
, (char *)table
->data
);
4810 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4814 int oldval
= user_zonelist_order
;
4816 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4819 * bogus value. restore saved string
4821 strncpy((char *)table
->data
, saved_string
,
4822 NUMA_ZONELIST_ORDER_LEN
);
4823 user_zonelist_order
= oldval
;
4824 } else if (oldval
!= user_zonelist_order
) {
4825 mutex_lock(&zonelists_mutex
);
4826 build_all_zonelists(NULL
, NULL
);
4827 mutex_unlock(&zonelists_mutex
);
4831 mutex_unlock(&zl_order_mutex
);
4836 #define MAX_NODE_LOAD (nr_online_nodes)
4837 static int node_load
[MAX_NUMNODES
];
4840 * find_next_best_node - find the next node that should appear in a given node's fallback list
4841 * @node: node whose fallback list we're appending
4842 * @used_node_mask: nodemask_t of already used nodes
4844 * We use a number of factors to determine which is the next node that should
4845 * appear on a given node's fallback list. The node should not have appeared
4846 * already in @node's fallback list, and it should be the next closest node
4847 * according to the distance array (which contains arbitrary distance values
4848 * from each node to each node in the system), and should also prefer nodes
4849 * with no CPUs, since presumably they'll have very little allocation pressure
4850 * on them otherwise.
4851 * It returns -1 if no node is found.
4853 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4856 int min_val
= INT_MAX
;
4857 int best_node
= NUMA_NO_NODE
;
4858 const struct cpumask
*tmp
= cpumask_of_node(0);
4860 /* Use the local node if we haven't already */
4861 if (!node_isset(node
, *used_node_mask
)) {
4862 node_set(node
, *used_node_mask
);
4866 for_each_node_state(n
, N_MEMORY
) {
4868 /* Don't want a node to appear more than once */
4869 if (node_isset(n
, *used_node_mask
))
4872 /* Use the distance array to find the distance */
4873 val
= node_distance(node
, n
);
4875 /* Penalize nodes under us ("prefer the next node") */
4878 /* Give preference to headless and unused nodes */
4879 tmp
= cpumask_of_node(n
);
4880 if (!cpumask_empty(tmp
))
4881 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4883 /* Slight preference for less loaded node */
4884 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4885 val
+= node_load
[n
];
4887 if (val
< min_val
) {
4894 node_set(best_node
, *used_node_mask
);
4901 * Build zonelists ordered by node and zones within node.
4902 * This results in maximum locality--normal zone overflows into local
4903 * DMA zone, if any--but risks exhausting DMA zone.
4905 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4908 struct zonelist
*zonelist
;
4910 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4911 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4913 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4914 zonelist
->_zonerefs
[j
].zone
= NULL
;
4915 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4919 * Build gfp_thisnode zonelists
4921 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4924 struct zonelist
*zonelist
;
4926 zonelist
= &pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
];
4927 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4928 zonelist
->_zonerefs
[j
].zone
= NULL
;
4929 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4933 * Build zonelists ordered by zone and nodes within zones.
4934 * This results in conserving DMA zone[s] until all Normal memory is
4935 * exhausted, but results in overflowing to remote node while memory
4936 * may still exist in local DMA zone.
4938 static int node_order
[MAX_NUMNODES
];
4940 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4943 int zone_type
; /* needs to be signed */
4945 struct zonelist
*zonelist
;
4947 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4949 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4950 for (j
= 0; j
< nr_nodes
; j
++) {
4951 node
= node_order
[j
];
4952 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4953 if (managed_zone(z
)) {
4955 &zonelist
->_zonerefs
[pos
++]);
4956 check_highest_zone(zone_type
);
4960 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4961 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4964 #if defined(CONFIG_64BIT)
4966 * Devices that require DMA32/DMA are relatively rare and do not justify a
4967 * penalty to every machine in case the specialised case applies. Default
4968 * to Node-ordering on 64-bit NUMA machines
4970 static int default_zonelist_order(void)
4972 return ZONELIST_ORDER_NODE
;
4976 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4977 * by the kernel. If processes running on node 0 deplete the low memory zone
4978 * then reclaim will occur more frequency increasing stalls and potentially
4979 * be easier to OOM if a large percentage of the zone is under writeback or
4980 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4981 * Hence, default to zone ordering on 32-bit.
4983 static int default_zonelist_order(void)
4985 return ZONELIST_ORDER_ZONE
;
4987 #endif /* CONFIG_64BIT */
4989 static void set_zonelist_order(void)
4991 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4992 current_zonelist_order
= default_zonelist_order();
4994 current_zonelist_order
= user_zonelist_order
;
4997 static void build_zonelists(pg_data_t
*pgdat
)
5000 nodemask_t used_mask
;
5001 int local_node
, prev_node
;
5002 struct zonelist
*zonelist
;
5003 unsigned int order
= current_zonelist_order
;
5005 /* initialize zonelists */
5006 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
5007 zonelist
= pgdat
->node_zonelists
+ i
;
5008 zonelist
->_zonerefs
[0].zone
= NULL
;
5009 zonelist
->_zonerefs
[0].zone_idx
= 0;
5012 /* NUMA-aware ordering of nodes */
5013 local_node
= pgdat
->node_id
;
5014 load
= nr_online_nodes
;
5015 prev_node
= local_node
;
5016 nodes_clear(used_mask
);
5018 memset(node_order
, 0, sizeof(node_order
));
5021 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5023 * We don't want to pressure a particular node.
5024 * So adding penalty to the first node in same
5025 * distance group to make it round-robin.
5027 if (node_distance(local_node
, node
) !=
5028 node_distance(local_node
, prev_node
))
5029 node_load
[node
] = load
;
5033 if (order
== ZONELIST_ORDER_NODE
)
5034 build_zonelists_in_node_order(pgdat
, node
);
5036 node_order
[i
++] = node
; /* remember order */
5039 if (order
== ZONELIST_ORDER_ZONE
) {
5040 /* calculate node order -- i.e., DMA last! */
5041 build_zonelists_in_zone_order(pgdat
, i
);
5044 build_thisnode_zonelists(pgdat
);
5047 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5049 * Return node id of node used for "local" allocations.
5050 * I.e., first node id of first zone in arg node's generic zonelist.
5051 * Used for initializing percpu 'numa_mem', which is used primarily
5052 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5054 int local_memory_node(int node
)
5058 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5059 gfp_zone(GFP_KERNEL
),
5061 return z
->zone
->node
;
5065 static void setup_min_unmapped_ratio(void);
5066 static void setup_min_slab_ratio(void);
5067 #else /* CONFIG_NUMA */
5069 static void set_zonelist_order(void)
5071 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
5074 static void build_zonelists(pg_data_t
*pgdat
)
5076 int node
, local_node
;
5078 struct zonelist
*zonelist
;
5080 local_node
= pgdat
->node_id
;
5082 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
5083 j
= build_zonelists_node(pgdat
, zonelist
, 0);
5086 * Now we build the zonelist so that it contains the zones
5087 * of all the other nodes.
5088 * We don't want to pressure a particular node, so when
5089 * building the zones for node N, we make sure that the
5090 * zones coming right after the local ones are those from
5091 * node N+1 (modulo N)
5093 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5094 if (!node_online(node
))
5096 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
5098 for (node
= 0; node
< local_node
; node
++) {
5099 if (!node_online(node
))
5101 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
5104 zonelist
->_zonerefs
[j
].zone
= NULL
;
5105 zonelist
->_zonerefs
[j
].zone_idx
= 0;
5108 #endif /* CONFIG_NUMA */
5111 * Boot pageset table. One per cpu which is going to be used for all
5112 * zones and all nodes. The parameters will be set in such a way
5113 * that an item put on a list will immediately be handed over to
5114 * the buddy list. This is safe since pageset manipulation is done
5115 * with interrupts disabled.
5117 * The boot_pagesets must be kept even after bootup is complete for
5118 * unused processors and/or zones. They do play a role for bootstrapping
5119 * hotplugged processors.
5121 * zoneinfo_show() and maybe other functions do
5122 * not check if the processor is online before following the pageset pointer.
5123 * Other parts of the kernel may not check if the zone is available.
5125 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5126 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5127 static void setup_zone_pageset(struct zone
*zone
);
5130 * Global mutex to protect against size modification of zonelists
5131 * as well as to serialize pageset setup for the new populated zone.
5133 DEFINE_MUTEX(zonelists_mutex
);
5135 /* return values int ....just for stop_machine() */
5136 static int __build_all_zonelists(void *data
)
5140 pg_data_t
*self
= data
;
5143 memset(node_load
, 0, sizeof(node_load
));
5146 if (self
&& !node_online(self
->node_id
)) {
5147 build_zonelists(self
);
5150 for_each_online_node(nid
) {
5151 pg_data_t
*pgdat
= NODE_DATA(nid
);
5153 build_zonelists(pgdat
);
5157 * Initialize the boot_pagesets that are going to be used
5158 * for bootstrapping processors. The real pagesets for
5159 * each zone will be allocated later when the per cpu
5160 * allocator is available.
5162 * boot_pagesets are used also for bootstrapping offline
5163 * cpus if the system is already booted because the pagesets
5164 * are needed to initialize allocators on a specific cpu too.
5165 * F.e. the percpu allocator needs the page allocator which
5166 * needs the percpu allocator in order to allocate its pagesets
5167 * (a chicken-egg dilemma).
5169 for_each_possible_cpu(cpu
) {
5170 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5172 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5174 * We now know the "local memory node" for each node--
5175 * i.e., the node of the first zone in the generic zonelist.
5176 * Set up numa_mem percpu variable for on-line cpus. During
5177 * boot, only the boot cpu should be on-line; we'll init the
5178 * secondary cpus' numa_mem as they come on-line. During
5179 * node/memory hotplug, we'll fixup all on-line cpus.
5181 if (cpu_online(cpu
))
5182 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5189 static noinline
void __init
5190 build_all_zonelists_init(void)
5192 __build_all_zonelists(NULL
);
5193 mminit_verify_zonelist();
5194 cpuset_init_current_mems_allowed();
5198 * Called with zonelists_mutex held always
5199 * unless system_state == SYSTEM_BOOTING.
5201 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5202 * [we're only called with non-NULL zone through __meminit paths] and
5203 * (2) call of __init annotated helper build_all_zonelists_init
5204 * [protected by SYSTEM_BOOTING].
5206 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
5208 set_zonelist_order();
5210 if (system_state
== SYSTEM_BOOTING
) {
5211 build_all_zonelists_init();
5213 #ifdef CONFIG_MEMORY_HOTPLUG
5215 setup_zone_pageset(zone
);
5217 /* we have to stop all cpus to guarantee there is no user
5219 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
5220 /* cpuset refresh routine should be here */
5222 vm_total_pages
= nr_free_pagecache_pages();
5224 * Disable grouping by mobility if the number of pages in the
5225 * system is too low to allow the mechanism to work. It would be
5226 * more accurate, but expensive to check per-zone. This check is
5227 * made on memory-hotadd so a system can start with mobility
5228 * disabled and enable it later
5230 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5231 page_group_by_mobility_disabled
= 1;
5233 page_group_by_mobility_disabled
= 0;
5235 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5237 zonelist_order_name
[current_zonelist_order
],
5238 page_group_by_mobility_disabled
? "off" : "on",
5241 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5246 * Initially all pages are reserved - free ones are freed
5247 * up by free_all_bootmem() once the early boot process is
5248 * done. Non-atomic initialization, single-pass.
5250 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5251 unsigned long start_pfn
, enum memmap_context context
)
5253 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5254 unsigned long end_pfn
= start_pfn
+ size
;
5255 pg_data_t
*pgdat
= NODE_DATA(nid
);
5257 unsigned long nr_initialised
= 0;
5258 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5259 struct memblock_region
*r
= NULL
, *tmp
;
5262 if (highest_memmap_pfn
< end_pfn
- 1)
5263 highest_memmap_pfn
= end_pfn
- 1;
5266 * Honor reservation requested by the driver for this ZONE_DEVICE
5269 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5270 start_pfn
+= altmap
->reserve
;
5272 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5274 * There can be holes in boot-time mem_map[]s handed to this
5275 * function. They do not exist on hotplugged memory.
5277 if (context
!= MEMMAP_EARLY
)
5280 if (!early_pfn_valid(pfn
)) {
5281 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5283 * Skip to the pfn preceding the next valid one (or
5284 * end_pfn), such that we hit a valid pfn (or end_pfn)
5285 * on our next iteration of the loop.
5287 pfn
= memblock_next_valid_pfn(pfn
, end_pfn
) - 1;
5291 if (!early_pfn_in_nid(pfn
, nid
))
5293 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5296 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5298 * Check given memblock attribute by firmware which can affect
5299 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5300 * mirrored, it's an overlapped memmap init. skip it.
5302 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5303 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5304 for_each_memblock(memory
, tmp
)
5305 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5309 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5310 memblock_is_mirror(r
)) {
5311 /* already initialized as NORMAL */
5312 pfn
= memblock_region_memory_end_pfn(r
);
5320 * Mark the block movable so that blocks are reserved for
5321 * movable at startup. This will force kernel allocations
5322 * to reserve their blocks rather than leaking throughout
5323 * the address space during boot when many long-lived
5324 * kernel allocations are made.
5326 * bitmap is created for zone's valid pfn range. but memmap
5327 * can be created for invalid pages (for alignment)
5328 * check here not to call set_pageblock_migratetype() against
5331 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5332 struct page
*page
= pfn_to_page(pfn
);
5334 __init_single_page(page
, pfn
, zone
, nid
);
5335 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5337 __init_single_pfn(pfn
, zone
, nid
);
5342 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5344 unsigned int order
, t
;
5345 for_each_migratetype_order(order
, t
) {
5346 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5347 zone
->free_area
[order
].nr_free
= 0;
5351 #ifndef __HAVE_ARCH_MEMMAP_INIT
5352 #define memmap_init(size, nid, zone, start_pfn) \
5353 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5356 static int zone_batchsize(struct zone
*zone
)
5362 * The per-cpu-pages pools are set to around 1000th of the
5363 * size of the zone. But no more than 1/2 of a meg.
5365 * OK, so we don't know how big the cache is. So guess.
5367 batch
= zone
->managed_pages
/ 1024;
5368 if (batch
* PAGE_SIZE
> 512 * 1024)
5369 batch
= (512 * 1024) / PAGE_SIZE
;
5370 batch
/= 4; /* We effectively *= 4 below */
5375 * Clamp the batch to a 2^n - 1 value. Having a power
5376 * of 2 value was found to be more likely to have
5377 * suboptimal cache aliasing properties in some cases.
5379 * For example if 2 tasks are alternately allocating
5380 * batches of pages, one task can end up with a lot
5381 * of pages of one half of the possible page colors
5382 * and the other with pages of the other colors.
5384 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5389 /* The deferral and batching of frees should be suppressed under NOMMU
5392 * The problem is that NOMMU needs to be able to allocate large chunks
5393 * of contiguous memory as there's no hardware page translation to
5394 * assemble apparent contiguous memory from discontiguous pages.
5396 * Queueing large contiguous runs of pages for batching, however,
5397 * causes the pages to actually be freed in smaller chunks. As there
5398 * can be a significant delay between the individual batches being
5399 * recycled, this leads to the once large chunks of space being
5400 * fragmented and becoming unavailable for high-order allocations.
5407 * pcp->high and pcp->batch values are related and dependent on one another:
5408 * ->batch must never be higher then ->high.
5409 * The following function updates them in a safe manner without read side
5412 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5413 * those fields changing asynchronously (acording the the above rule).
5415 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5416 * outside of boot time (or some other assurance that no concurrent updaters
5419 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5420 unsigned long batch
)
5422 /* start with a fail safe value for batch */
5426 /* Update high, then batch, in order */
5433 /* a companion to pageset_set_high() */
5434 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5436 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5439 static void pageset_init(struct per_cpu_pageset
*p
)
5441 struct per_cpu_pages
*pcp
;
5444 memset(p
, 0, sizeof(*p
));
5448 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5449 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5452 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5455 pageset_set_batch(p
, batch
);
5459 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5460 * to the value high for the pageset p.
5462 static void pageset_set_high(struct per_cpu_pageset
*p
,
5465 unsigned long batch
= max(1UL, high
/ 4);
5466 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5467 batch
= PAGE_SHIFT
* 8;
5469 pageset_update(&p
->pcp
, high
, batch
);
5472 static void pageset_set_high_and_batch(struct zone
*zone
,
5473 struct per_cpu_pageset
*pcp
)
5475 if (percpu_pagelist_fraction
)
5476 pageset_set_high(pcp
,
5477 (zone
->managed_pages
/
5478 percpu_pagelist_fraction
));
5480 pageset_set_batch(pcp
, zone_batchsize(zone
));
5483 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5485 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5488 pageset_set_high_and_batch(zone
, pcp
);
5491 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5494 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5495 for_each_possible_cpu(cpu
)
5496 zone_pageset_init(zone
, cpu
);
5500 * Allocate per cpu pagesets and initialize them.
5501 * Before this call only boot pagesets were available.
5503 void __init
setup_per_cpu_pageset(void)
5505 struct pglist_data
*pgdat
;
5508 for_each_populated_zone(zone
)
5509 setup_zone_pageset(zone
);
5511 for_each_online_pgdat(pgdat
)
5512 pgdat
->per_cpu_nodestats
=
5513 alloc_percpu(struct per_cpu_nodestat
);
5516 static __meminit
void zone_pcp_init(struct zone
*zone
)
5519 * per cpu subsystem is not up at this point. The following code
5520 * relies on the ability of the linker to provide the
5521 * offset of a (static) per cpu variable into the per cpu area.
5523 zone
->pageset
= &boot_pageset
;
5525 if (populated_zone(zone
))
5526 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5527 zone
->name
, zone
->present_pages
,
5528 zone_batchsize(zone
));
5531 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5532 unsigned long zone_start_pfn
,
5535 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5537 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5539 zone
->zone_start_pfn
= zone_start_pfn
;
5541 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5542 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5544 (unsigned long)zone_idx(zone
),
5545 zone_start_pfn
, (zone_start_pfn
+ size
));
5547 zone_init_free_lists(zone
);
5548 zone
->initialized
= 1;
5553 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5554 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5557 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5559 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5560 struct mminit_pfnnid_cache
*state
)
5562 unsigned long start_pfn
, end_pfn
;
5565 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5566 return state
->last_nid
;
5568 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5570 state
->last_start
= start_pfn
;
5571 state
->last_end
= end_pfn
;
5572 state
->last_nid
= nid
;
5577 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5580 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5581 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5582 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5584 * If an architecture guarantees that all ranges registered contain no holes
5585 * and may be freed, this this function may be used instead of calling
5586 * memblock_free_early_nid() manually.
5588 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5590 unsigned long start_pfn
, end_pfn
;
5593 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5594 start_pfn
= min(start_pfn
, max_low_pfn
);
5595 end_pfn
= min(end_pfn
, max_low_pfn
);
5597 if (start_pfn
< end_pfn
)
5598 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5599 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5605 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5606 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5608 * If an architecture guarantees that all ranges registered contain no holes and may
5609 * be freed, this function may be used instead of calling memory_present() manually.
5611 void __init
sparse_memory_present_with_active_regions(int nid
)
5613 unsigned long start_pfn
, end_pfn
;
5616 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5617 memory_present(this_nid
, start_pfn
, end_pfn
);
5621 * get_pfn_range_for_nid - Return the start and end page frames for a node
5622 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5623 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5624 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5626 * It returns the start and end page frame of a node based on information
5627 * provided by memblock_set_node(). If called for a node
5628 * with no available memory, a warning is printed and the start and end
5631 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5632 unsigned long *start_pfn
, unsigned long *end_pfn
)
5634 unsigned long this_start_pfn
, this_end_pfn
;
5640 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5641 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5642 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5645 if (*start_pfn
== -1UL)
5650 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5651 * assumption is made that zones within a node are ordered in monotonic
5652 * increasing memory addresses so that the "highest" populated zone is used
5654 static void __init
find_usable_zone_for_movable(void)
5657 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5658 if (zone_index
== ZONE_MOVABLE
)
5661 if (arch_zone_highest_possible_pfn
[zone_index
] >
5662 arch_zone_lowest_possible_pfn
[zone_index
])
5666 VM_BUG_ON(zone_index
== -1);
5667 movable_zone
= zone_index
;
5671 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5672 * because it is sized independent of architecture. Unlike the other zones,
5673 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5674 * in each node depending on the size of each node and how evenly kernelcore
5675 * is distributed. This helper function adjusts the zone ranges
5676 * provided by the architecture for a given node by using the end of the
5677 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5678 * zones within a node are in order of monotonic increases memory addresses
5680 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5681 unsigned long zone_type
,
5682 unsigned long node_start_pfn
,
5683 unsigned long node_end_pfn
,
5684 unsigned long *zone_start_pfn
,
5685 unsigned long *zone_end_pfn
)
5687 /* Only adjust if ZONE_MOVABLE is on this node */
5688 if (zone_movable_pfn
[nid
]) {
5689 /* Size ZONE_MOVABLE */
5690 if (zone_type
== ZONE_MOVABLE
) {
5691 *zone_start_pfn
= zone_movable_pfn
[nid
];
5692 *zone_end_pfn
= min(node_end_pfn
,
5693 arch_zone_highest_possible_pfn
[movable_zone
]);
5695 /* Adjust for ZONE_MOVABLE starting within this range */
5696 } else if (!mirrored_kernelcore
&&
5697 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5698 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5699 *zone_end_pfn
= zone_movable_pfn
[nid
];
5701 /* Check if this whole range is within ZONE_MOVABLE */
5702 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5703 *zone_start_pfn
= *zone_end_pfn
;
5708 * Return the number of pages a zone spans in a node, including holes
5709 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5711 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5712 unsigned long zone_type
,
5713 unsigned long node_start_pfn
,
5714 unsigned long node_end_pfn
,
5715 unsigned long *zone_start_pfn
,
5716 unsigned long *zone_end_pfn
,
5717 unsigned long *ignored
)
5719 /* When hotadd a new node from cpu_up(), the node should be empty */
5720 if (!node_start_pfn
&& !node_end_pfn
)
5723 /* Get the start and end of the zone */
5724 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5725 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5726 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5727 node_start_pfn
, node_end_pfn
,
5728 zone_start_pfn
, zone_end_pfn
);
5730 /* Check that this node has pages within the zone's required range */
5731 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5734 /* Move the zone boundaries inside the node if necessary */
5735 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5736 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5738 /* Return the spanned pages */
5739 return *zone_end_pfn
- *zone_start_pfn
;
5743 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5744 * then all holes in the requested range will be accounted for.
5746 unsigned long __meminit
__absent_pages_in_range(int nid
,
5747 unsigned long range_start_pfn
,
5748 unsigned long range_end_pfn
)
5750 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5751 unsigned long start_pfn
, end_pfn
;
5754 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5755 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5756 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5757 nr_absent
-= end_pfn
- start_pfn
;
5763 * absent_pages_in_range - Return number of page frames in holes within a range
5764 * @start_pfn: The start PFN to start searching for holes
5765 * @end_pfn: The end PFN to stop searching for holes
5767 * It returns the number of pages frames in memory holes within a range.
5769 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5770 unsigned long end_pfn
)
5772 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5775 /* Return the number of page frames in holes in a zone on a node */
5776 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5777 unsigned long zone_type
,
5778 unsigned long node_start_pfn
,
5779 unsigned long node_end_pfn
,
5780 unsigned long *ignored
)
5782 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5783 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5784 unsigned long zone_start_pfn
, zone_end_pfn
;
5785 unsigned long nr_absent
;
5787 /* When hotadd a new node from cpu_up(), the node should be empty */
5788 if (!node_start_pfn
&& !node_end_pfn
)
5791 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5792 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5794 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5795 node_start_pfn
, node_end_pfn
,
5796 &zone_start_pfn
, &zone_end_pfn
);
5797 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5800 * ZONE_MOVABLE handling.
5801 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5804 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5805 unsigned long start_pfn
, end_pfn
;
5806 struct memblock_region
*r
;
5808 for_each_memblock(memory
, r
) {
5809 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5810 zone_start_pfn
, zone_end_pfn
);
5811 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5812 zone_start_pfn
, zone_end_pfn
);
5814 if (zone_type
== ZONE_MOVABLE
&&
5815 memblock_is_mirror(r
))
5816 nr_absent
+= end_pfn
- start_pfn
;
5818 if (zone_type
== ZONE_NORMAL
&&
5819 !memblock_is_mirror(r
))
5820 nr_absent
+= end_pfn
- start_pfn
;
5827 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5828 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5829 unsigned long zone_type
,
5830 unsigned long node_start_pfn
,
5831 unsigned long node_end_pfn
,
5832 unsigned long *zone_start_pfn
,
5833 unsigned long *zone_end_pfn
,
5834 unsigned long *zones_size
)
5838 *zone_start_pfn
= node_start_pfn
;
5839 for (zone
= 0; zone
< zone_type
; zone
++)
5840 *zone_start_pfn
+= zones_size
[zone
];
5842 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5844 return zones_size
[zone_type
];
5847 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5848 unsigned long zone_type
,
5849 unsigned long node_start_pfn
,
5850 unsigned long node_end_pfn
,
5851 unsigned long *zholes_size
)
5856 return zholes_size
[zone_type
];
5859 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5861 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5862 unsigned long node_start_pfn
,
5863 unsigned long node_end_pfn
,
5864 unsigned long *zones_size
,
5865 unsigned long *zholes_size
)
5867 unsigned long realtotalpages
= 0, totalpages
= 0;
5870 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5871 struct zone
*zone
= pgdat
->node_zones
+ i
;
5872 unsigned long zone_start_pfn
, zone_end_pfn
;
5873 unsigned long size
, real_size
;
5875 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5881 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5882 node_start_pfn
, node_end_pfn
,
5885 zone
->zone_start_pfn
= zone_start_pfn
;
5887 zone
->zone_start_pfn
= 0;
5888 zone
->spanned_pages
= size
;
5889 zone
->present_pages
= real_size
;
5892 realtotalpages
+= real_size
;
5895 pgdat
->node_spanned_pages
= totalpages
;
5896 pgdat
->node_present_pages
= realtotalpages
;
5897 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5901 #ifndef CONFIG_SPARSEMEM
5903 * Calculate the size of the zone->blockflags rounded to an unsigned long
5904 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5905 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5906 * round what is now in bits to nearest long in bits, then return it in
5909 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5911 unsigned long usemapsize
;
5913 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5914 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5915 usemapsize
= usemapsize
>> pageblock_order
;
5916 usemapsize
*= NR_PAGEBLOCK_BITS
;
5917 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5919 return usemapsize
/ 8;
5922 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5924 unsigned long zone_start_pfn
,
5925 unsigned long zonesize
)
5927 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5928 zone
->pageblock_flags
= NULL
;
5930 zone
->pageblock_flags
=
5931 memblock_virt_alloc_node_nopanic(usemapsize
,
5935 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5936 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5937 #endif /* CONFIG_SPARSEMEM */
5939 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5941 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5942 void __paginginit
set_pageblock_order(void)
5946 /* Check that pageblock_nr_pages has not already been setup */
5947 if (pageblock_order
)
5950 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5951 order
= HUGETLB_PAGE_ORDER
;
5953 order
= MAX_ORDER
- 1;
5956 * Assume the largest contiguous order of interest is a huge page.
5957 * This value may be variable depending on boot parameters on IA64 and
5960 pageblock_order
= order
;
5962 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5965 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5966 * is unused as pageblock_order is set at compile-time. See
5967 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5970 void __paginginit
set_pageblock_order(void)
5974 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5976 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5977 unsigned long present_pages
)
5979 unsigned long pages
= spanned_pages
;
5982 * Provide a more accurate estimation if there are holes within
5983 * the zone and SPARSEMEM is in use. If there are holes within the
5984 * zone, each populated memory region may cost us one or two extra
5985 * memmap pages due to alignment because memmap pages for each
5986 * populated regions may not be naturally aligned on page boundary.
5987 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5989 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5990 IS_ENABLED(CONFIG_SPARSEMEM
))
5991 pages
= present_pages
;
5993 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5997 * Set up the zone data structures:
5998 * - mark all pages reserved
5999 * - mark all memory queues empty
6000 * - clear the memory bitmaps
6002 * NOTE: pgdat should get zeroed by caller.
6004 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
6007 int nid
= pgdat
->node_id
;
6010 pgdat_resize_init(pgdat
);
6011 #ifdef CONFIG_NUMA_BALANCING
6012 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
6013 pgdat
->numabalancing_migrate_nr_pages
= 0;
6014 pgdat
->numabalancing_migrate_next_window
= jiffies
;
6016 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6017 spin_lock_init(&pgdat
->split_queue_lock
);
6018 INIT_LIST_HEAD(&pgdat
->split_queue
);
6019 pgdat
->split_queue_len
= 0;
6021 init_waitqueue_head(&pgdat
->kswapd_wait
);
6022 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6023 #ifdef CONFIG_COMPACTION
6024 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6026 pgdat_page_ext_init(pgdat
);
6027 spin_lock_init(&pgdat
->lru_lock
);
6028 lruvec_init(node_lruvec(pgdat
));
6030 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6031 struct zone
*zone
= pgdat
->node_zones
+ j
;
6032 unsigned long size
, realsize
, freesize
, memmap_pages
;
6033 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6035 size
= zone
->spanned_pages
;
6036 realsize
= freesize
= zone
->present_pages
;
6039 * Adjust freesize so that it accounts for how much memory
6040 * is used by this zone for memmap. This affects the watermark
6041 * and per-cpu initialisations
6043 memmap_pages
= calc_memmap_size(size
, realsize
);
6044 if (!is_highmem_idx(j
)) {
6045 if (freesize
>= memmap_pages
) {
6046 freesize
-= memmap_pages
;
6049 " %s zone: %lu pages used for memmap\n",
6050 zone_names
[j
], memmap_pages
);
6052 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6053 zone_names
[j
], memmap_pages
, freesize
);
6056 /* Account for reserved pages */
6057 if (j
== 0 && freesize
> dma_reserve
) {
6058 freesize
-= dma_reserve
;
6059 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6060 zone_names
[0], dma_reserve
);
6063 if (!is_highmem_idx(j
))
6064 nr_kernel_pages
+= freesize
;
6065 /* Charge for highmem memmap if there are enough kernel pages */
6066 else if (nr_kernel_pages
> memmap_pages
* 2)
6067 nr_kernel_pages
-= memmap_pages
;
6068 nr_all_pages
+= freesize
;
6071 * Set an approximate value for lowmem here, it will be adjusted
6072 * when the bootmem allocator frees pages into the buddy system.
6073 * And all highmem pages will be managed by the buddy system.
6075 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
6079 zone
->name
= zone_names
[j
];
6080 zone
->zone_pgdat
= pgdat
;
6081 spin_lock_init(&zone
->lock
);
6082 zone_seqlock_init(zone
);
6083 zone_pcp_init(zone
);
6088 set_pageblock_order();
6089 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6090 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6092 memmap_init(size
, nid
, j
, zone_start_pfn
);
6096 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6098 unsigned long __maybe_unused start
= 0;
6099 unsigned long __maybe_unused offset
= 0;
6101 /* Skip empty nodes */
6102 if (!pgdat
->node_spanned_pages
)
6105 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6106 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6107 offset
= pgdat
->node_start_pfn
- start
;
6108 /* ia64 gets its own node_mem_map, before this, without bootmem */
6109 if (!pgdat
->node_mem_map
) {
6110 unsigned long size
, end
;
6114 * The zone's endpoints aren't required to be MAX_ORDER
6115 * aligned but the node_mem_map endpoints must be in order
6116 * for the buddy allocator to function correctly.
6118 end
= pgdat_end_pfn(pgdat
);
6119 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6120 size
= (end
- start
) * sizeof(struct page
);
6121 map
= alloc_remap(pgdat
->node_id
, size
);
6123 map
= memblock_virt_alloc_node_nopanic(size
,
6125 pgdat
->node_mem_map
= map
+ offset
;
6127 #ifndef CONFIG_NEED_MULTIPLE_NODES
6129 * With no DISCONTIG, the global mem_map is just set as node 0's
6131 if (pgdat
== NODE_DATA(0)) {
6132 mem_map
= NODE_DATA(0)->node_mem_map
;
6133 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6134 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6136 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6139 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6142 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
6143 unsigned long node_start_pfn
, unsigned long *zholes_size
)
6145 pg_data_t
*pgdat
= NODE_DATA(nid
);
6146 unsigned long start_pfn
= 0;
6147 unsigned long end_pfn
= 0;
6149 /* pg_data_t should be reset to zero when it's allocated */
6150 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6152 pgdat
->node_id
= nid
;
6153 pgdat
->node_start_pfn
= node_start_pfn
;
6154 pgdat
->per_cpu_nodestats
= NULL
;
6155 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6156 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6157 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6158 (u64
)start_pfn
<< PAGE_SHIFT
,
6159 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6161 start_pfn
= node_start_pfn
;
6163 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6164 zones_size
, zholes_size
);
6166 alloc_node_mem_map(pgdat
);
6167 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6168 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6169 nid
, (unsigned long)pgdat
,
6170 (unsigned long)pgdat
->node_mem_map
);
6173 reset_deferred_meminit(pgdat
);
6174 free_area_init_core(pgdat
);
6177 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6179 #if MAX_NUMNODES > 1
6181 * Figure out the number of possible node ids.
6183 void __init
setup_nr_node_ids(void)
6185 unsigned int highest
;
6187 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6188 nr_node_ids
= highest
+ 1;
6193 * node_map_pfn_alignment - determine the maximum internode alignment
6195 * This function should be called after node map is populated and sorted.
6196 * It calculates the maximum power of two alignment which can distinguish
6199 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6200 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6201 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6202 * shifted, 1GiB is enough and this function will indicate so.
6204 * This is used to test whether pfn -> nid mapping of the chosen memory
6205 * model has fine enough granularity to avoid incorrect mapping for the
6206 * populated node map.
6208 * Returns the determined alignment in pfn's. 0 if there is no alignment
6209 * requirement (single node).
6211 unsigned long __init
node_map_pfn_alignment(void)
6213 unsigned long accl_mask
= 0, last_end
= 0;
6214 unsigned long start
, end
, mask
;
6218 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6219 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6226 * Start with a mask granular enough to pin-point to the
6227 * start pfn and tick off bits one-by-one until it becomes
6228 * too coarse to separate the current node from the last.
6230 mask
= ~((1 << __ffs(start
)) - 1);
6231 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6234 /* accumulate all internode masks */
6238 /* convert mask to number of pages */
6239 return ~accl_mask
+ 1;
6242 /* Find the lowest pfn for a node */
6243 static unsigned long __init
find_min_pfn_for_node(int nid
)
6245 unsigned long min_pfn
= ULONG_MAX
;
6246 unsigned long start_pfn
;
6249 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6250 min_pfn
= min(min_pfn
, start_pfn
);
6252 if (min_pfn
== ULONG_MAX
) {
6253 pr_warn("Could not find start_pfn for node %d\n", nid
);
6261 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6263 * It returns the minimum PFN based on information provided via
6264 * memblock_set_node().
6266 unsigned long __init
find_min_pfn_with_active_regions(void)
6268 return find_min_pfn_for_node(MAX_NUMNODES
);
6272 * early_calculate_totalpages()
6273 * Sum pages in active regions for movable zone.
6274 * Populate N_MEMORY for calculating usable_nodes.
6276 static unsigned long __init
early_calculate_totalpages(void)
6278 unsigned long totalpages
= 0;
6279 unsigned long start_pfn
, end_pfn
;
6282 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6283 unsigned long pages
= end_pfn
- start_pfn
;
6285 totalpages
+= pages
;
6287 node_set_state(nid
, N_MEMORY
);
6293 * Find the PFN the Movable zone begins in each node. Kernel memory
6294 * is spread evenly between nodes as long as the nodes have enough
6295 * memory. When they don't, some nodes will have more kernelcore than
6298 static void __init
find_zone_movable_pfns_for_nodes(void)
6301 unsigned long usable_startpfn
;
6302 unsigned long kernelcore_node
, kernelcore_remaining
;
6303 /* save the state before borrow the nodemask */
6304 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6305 unsigned long totalpages
= early_calculate_totalpages();
6306 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6307 struct memblock_region
*r
;
6309 /* Need to find movable_zone earlier when movable_node is specified. */
6310 find_usable_zone_for_movable();
6313 * If movable_node is specified, ignore kernelcore and movablecore
6316 if (movable_node_is_enabled()) {
6317 for_each_memblock(memory
, r
) {
6318 if (!memblock_is_hotpluggable(r
))
6323 usable_startpfn
= PFN_DOWN(r
->base
);
6324 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6325 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6333 * If kernelcore=mirror is specified, ignore movablecore option
6335 if (mirrored_kernelcore
) {
6336 bool mem_below_4gb_not_mirrored
= false;
6338 for_each_memblock(memory
, r
) {
6339 if (memblock_is_mirror(r
))
6344 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6346 if (usable_startpfn
< 0x100000) {
6347 mem_below_4gb_not_mirrored
= true;
6351 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6352 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6356 if (mem_below_4gb_not_mirrored
)
6357 pr_warn("This configuration results in unmirrored kernel memory.");
6363 * If movablecore=nn[KMG] was specified, calculate what size of
6364 * kernelcore that corresponds so that memory usable for
6365 * any allocation type is evenly spread. If both kernelcore
6366 * and movablecore are specified, then the value of kernelcore
6367 * will be used for required_kernelcore if it's greater than
6368 * what movablecore would have allowed.
6370 if (required_movablecore
) {
6371 unsigned long corepages
;
6374 * Round-up so that ZONE_MOVABLE is at least as large as what
6375 * was requested by the user
6377 required_movablecore
=
6378 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6379 required_movablecore
= min(totalpages
, required_movablecore
);
6380 corepages
= totalpages
- required_movablecore
;
6382 required_kernelcore
= max(required_kernelcore
, corepages
);
6386 * If kernelcore was not specified or kernelcore size is larger
6387 * than totalpages, there is no ZONE_MOVABLE.
6389 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6392 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6393 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6396 /* Spread kernelcore memory as evenly as possible throughout nodes */
6397 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6398 for_each_node_state(nid
, N_MEMORY
) {
6399 unsigned long start_pfn
, end_pfn
;
6402 * Recalculate kernelcore_node if the division per node
6403 * now exceeds what is necessary to satisfy the requested
6404 * amount of memory for the kernel
6406 if (required_kernelcore
< kernelcore_node
)
6407 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6410 * As the map is walked, we track how much memory is usable
6411 * by the kernel using kernelcore_remaining. When it is
6412 * 0, the rest of the node is usable by ZONE_MOVABLE
6414 kernelcore_remaining
= kernelcore_node
;
6416 /* Go through each range of PFNs within this node */
6417 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6418 unsigned long size_pages
;
6420 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6421 if (start_pfn
>= end_pfn
)
6424 /* Account for what is only usable for kernelcore */
6425 if (start_pfn
< usable_startpfn
) {
6426 unsigned long kernel_pages
;
6427 kernel_pages
= min(end_pfn
, usable_startpfn
)
6430 kernelcore_remaining
-= min(kernel_pages
,
6431 kernelcore_remaining
);
6432 required_kernelcore
-= min(kernel_pages
,
6433 required_kernelcore
);
6435 /* Continue if range is now fully accounted */
6436 if (end_pfn
<= usable_startpfn
) {
6439 * Push zone_movable_pfn to the end so
6440 * that if we have to rebalance
6441 * kernelcore across nodes, we will
6442 * not double account here
6444 zone_movable_pfn
[nid
] = end_pfn
;
6447 start_pfn
= usable_startpfn
;
6451 * The usable PFN range for ZONE_MOVABLE is from
6452 * start_pfn->end_pfn. Calculate size_pages as the
6453 * number of pages used as kernelcore
6455 size_pages
= end_pfn
- start_pfn
;
6456 if (size_pages
> kernelcore_remaining
)
6457 size_pages
= kernelcore_remaining
;
6458 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6461 * Some kernelcore has been met, update counts and
6462 * break if the kernelcore for this node has been
6465 required_kernelcore
-= min(required_kernelcore
,
6467 kernelcore_remaining
-= size_pages
;
6468 if (!kernelcore_remaining
)
6474 * If there is still required_kernelcore, we do another pass with one
6475 * less node in the count. This will push zone_movable_pfn[nid] further
6476 * along on the nodes that still have memory until kernelcore is
6480 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6484 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6485 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6486 zone_movable_pfn
[nid
] =
6487 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6490 /* restore the node_state */
6491 node_states
[N_MEMORY
] = saved_node_state
;
6494 /* Any regular or high memory on that node ? */
6495 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6497 enum zone_type zone_type
;
6499 if (N_MEMORY
== N_NORMAL_MEMORY
)
6502 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6503 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6504 if (populated_zone(zone
)) {
6505 node_set_state(nid
, N_HIGH_MEMORY
);
6506 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6507 zone_type
<= ZONE_NORMAL
)
6508 node_set_state(nid
, N_NORMAL_MEMORY
);
6515 * free_area_init_nodes - Initialise all pg_data_t and zone data
6516 * @max_zone_pfn: an array of max PFNs for each zone
6518 * This will call free_area_init_node() for each active node in the system.
6519 * Using the page ranges provided by memblock_set_node(), the size of each
6520 * zone in each node and their holes is calculated. If the maximum PFN
6521 * between two adjacent zones match, it is assumed that the zone is empty.
6522 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6523 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6524 * starts where the previous one ended. For example, ZONE_DMA32 starts
6525 * at arch_max_dma_pfn.
6527 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6529 unsigned long start_pfn
, end_pfn
;
6532 /* Record where the zone boundaries are */
6533 memset(arch_zone_lowest_possible_pfn
, 0,
6534 sizeof(arch_zone_lowest_possible_pfn
));
6535 memset(arch_zone_highest_possible_pfn
, 0,
6536 sizeof(arch_zone_highest_possible_pfn
));
6538 start_pfn
= find_min_pfn_with_active_regions();
6540 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6541 if (i
== ZONE_MOVABLE
)
6544 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6545 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6546 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6548 start_pfn
= end_pfn
;
6551 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6552 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6553 find_zone_movable_pfns_for_nodes();
6555 /* Print out the zone ranges */
6556 pr_info("Zone ranges:\n");
6557 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6558 if (i
== ZONE_MOVABLE
)
6560 pr_info(" %-8s ", zone_names
[i
]);
6561 if (arch_zone_lowest_possible_pfn
[i
] ==
6562 arch_zone_highest_possible_pfn
[i
])
6565 pr_cont("[mem %#018Lx-%#018Lx]\n",
6566 (u64
)arch_zone_lowest_possible_pfn
[i
]
6568 ((u64
)arch_zone_highest_possible_pfn
[i
]
6569 << PAGE_SHIFT
) - 1);
6572 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6573 pr_info("Movable zone start for each node\n");
6574 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6575 if (zone_movable_pfn
[i
])
6576 pr_info(" Node %d: %#018Lx\n", i
,
6577 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6580 /* Print out the early node map */
6581 pr_info("Early memory node ranges\n");
6582 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6583 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6584 (u64
)start_pfn
<< PAGE_SHIFT
,
6585 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6587 /* Initialise every node */
6588 mminit_verify_pageflags_layout();
6589 setup_nr_node_ids();
6590 for_each_online_node(nid
) {
6591 pg_data_t
*pgdat
= NODE_DATA(nid
);
6592 free_area_init_node(nid
, NULL
,
6593 find_min_pfn_for_node(nid
), NULL
);
6595 /* Any memory on that node */
6596 if (pgdat
->node_present_pages
)
6597 node_set_state(nid
, N_MEMORY
);
6598 check_for_memory(pgdat
, nid
);
6602 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6604 unsigned long long coremem
;
6608 coremem
= memparse(p
, &p
);
6609 *core
= coremem
>> PAGE_SHIFT
;
6611 /* Paranoid check that UL is enough for the coremem value */
6612 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6618 * kernelcore=size sets the amount of memory for use for allocations that
6619 * cannot be reclaimed or migrated.
6621 static int __init
cmdline_parse_kernelcore(char *p
)
6623 /* parse kernelcore=mirror */
6624 if (parse_option_str(p
, "mirror")) {
6625 mirrored_kernelcore
= true;
6629 return cmdline_parse_core(p
, &required_kernelcore
);
6633 * movablecore=size sets the amount of memory for use for allocations that
6634 * can be reclaimed or migrated.
6636 static int __init
cmdline_parse_movablecore(char *p
)
6638 return cmdline_parse_core(p
, &required_movablecore
);
6641 early_param("kernelcore", cmdline_parse_kernelcore
);
6642 early_param("movablecore", cmdline_parse_movablecore
);
6644 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6646 void adjust_managed_page_count(struct page
*page
, long count
)
6648 spin_lock(&managed_page_count_lock
);
6649 page_zone(page
)->managed_pages
+= count
;
6650 totalram_pages
+= count
;
6651 #ifdef CONFIG_HIGHMEM
6652 if (PageHighMem(page
))
6653 totalhigh_pages
+= count
;
6655 spin_unlock(&managed_page_count_lock
);
6657 EXPORT_SYMBOL(adjust_managed_page_count
);
6659 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6662 unsigned long pages
= 0;
6664 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6665 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6666 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6667 if ((unsigned int)poison
<= 0xFF)
6668 memset(pos
, poison
, PAGE_SIZE
);
6669 free_reserved_page(virt_to_page(pos
));
6673 pr_info("Freeing %s memory: %ldK\n",
6674 s
, pages
<< (PAGE_SHIFT
- 10));
6678 EXPORT_SYMBOL(free_reserved_area
);
6680 #ifdef CONFIG_HIGHMEM
6681 void free_highmem_page(struct page
*page
)
6683 __free_reserved_page(page
);
6685 page_zone(page
)->managed_pages
++;
6691 void __init
mem_init_print_info(const char *str
)
6693 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6694 unsigned long init_code_size
, init_data_size
;
6696 physpages
= get_num_physpages();
6697 codesize
= _etext
- _stext
;
6698 datasize
= _edata
- _sdata
;
6699 rosize
= __end_rodata
- __start_rodata
;
6700 bss_size
= __bss_stop
- __bss_start
;
6701 init_data_size
= __init_end
- __init_begin
;
6702 init_code_size
= _einittext
- _sinittext
;
6705 * Detect special cases and adjust section sizes accordingly:
6706 * 1) .init.* may be embedded into .data sections
6707 * 2) .init.text.* may be out of [__init_begin, __init_end],
6708 * please refer to arch/tile/kernel/vmlinux.lds.S.
6709 * 3) .rodata.* may be embedded into .text or .data sections.
6711 #define adj_init_size(start, end, size, pos, adj) \
6713 if (start <= pos && pos < end && size > adj) \
6717 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6718 _sinittext
, init_code_size
);
6719 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6720 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6721 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6722 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6724 #undef adj_init_size
6726 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6727 #ifdef CONFIG_HIGHMEM
6731 nr_free_pages() << (PAGE_SHIFT
- 10),
6732 physpages
<< (PAGE_SHIFT
- 10),
6733 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6734 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6735 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6736 totalcma_pages
<< (PAGE_SHIFT
- 10),
6737 #ifdef CONFIG_HIGHMEM
6738 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6740 str
? ", " : "", str
? str
: "");
6744 * set_dma_reserve - set the specified number of pages reserved in the first zone
6745 * @new_dma_reserve: The number of pages to mark reserved
6747 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6748 * In the DMA zone, a significant percentage may be consumed by kernel image
6749 * and other unfreeable allocations which can skew the watermarks badly. This
6750 * function may optionally be used to account for unfreeable pages in the
6751 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6752 * smaller per-cpu batchsize.
6754 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6756 dma_reserve
= new_dma_reserve
;
6759 void __init
free_area_init(unsigned long *zones_size
)
6761 free_area_init_node(0, zones_size
,
6762 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6765 static int page_alloc_cpu_dead(unsigned int cpu
)
6768 lru_add_drain_cpu(cpu
);
6772 * Spill the event counters of the dead processor
6773 * into the current processors event counters.
6774 * This artificially elevates the count of the current
6777 vm_events_fold_cpu(cpu
);
6780 * Zero the differential counters of the dead processor
6781 * so that the vm statistics are consistent.
6783 * This is only okay since the processor is dead and cannot
6784 * race with what we are doing.
6786 cpu_vm_stats_fold(cpu
);
6790 void __init
page_alloc_init(void)
6794 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
6795 "mm/page_alloc:dead", NULL
,
6796 page_alloc_cpu_dead
);
6801 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6802 * or min_free_kbytes changes.
6804 static void calculate_totalreserve_pages(void)
6806 struct pglist_data
*pgdat
;
6807 unsigned long reserve_pages
= 0;
6808 enum zone_type i
, j
;
6810 for_each_online_pgdat(pgdat
) {
6812 pgdat
->totalreserve_pages
= 0;
6814 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6815 struct zone
*zone
= pgdat
->node_zones
+ i
;
6818 /* Find valid and maximum lowmem_reserve in the zone */
6819 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6820 if (zone
->lowmem_reserve
[j
] > max
)
6821 max
= zone
->lowmem_reserve
[j
];
6824 /* we treat the high watermark as reserved pages. */
6825 max
+= high_wmark_pages(zone
);
6827 if (max
> zone
->managed_pages
)
6828 max
= zone
->managed_pages
;
6830 pgdat
->totalreserve_pages
+= max
;
6832 reserve_pages
+= max
;
6835 totalreserve_pages
= reserve_pages
;
6839 * setup_per_zone_lowmem_reserve - called whenever
6840 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6841 * has a correct pages reserved value, so an adequate number of
6842 * pages are left in the zone after a successful __alloc_pages().
6844 static void setup_per_zone_lowmem_reserve(void)
6846 struct pglist_data
*pgdat
;
6847 enum zone_type j
, idx
;
6849 for_each_online_pgdat(pgdat
) {
6850 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6851 struct zone
*zone
= pgdat
->node_zones
+ j
;
6852 unsigned long managed_pages
= zone
->managed_pages
;
6854 zone
->lowmem_reserve
[j
] = 0;
6858 struct zone
*lower_zone
;
6862 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6863 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6865 lower_zone
= pgdat
->node_zones
+ idx
;
6866 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6867 sysctl_lowmem_reserve_ratio
[idx
];
6868 managed_pages
+= lower_zone
->managed_pages
;
6873 /* update totalreserve_pages */
6874 calculate_totalreserve_pages();
6877 static void __setup_per_zone_wmarks(void)
6879 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6880 unsigned long lowmem_pages
= 0;
6882 unsigned long flags
;
6884 /* Calculate total number of !ZONE_HIGHMEM pages */
6885 for_each_zone(zone
) {
6886 if (!is_highmem(zone
))
6887 lowmem_pages
+= zone
->managed_pages
;
6890 for_each_zone(zone
) {
6893 spin_lock_irqsave(&zone
->lock
, flags
);
6894 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6895 do_div(tmp
, lowmem_pages
);
6896 if (is_highmem(zone
)) {
6898 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6899 * need highmem pages, so cap pages_min to a small
6902 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6903 * deltas control asynch page reclaim, and so should
6904 * not be capped for highmem.
6906 unsigned long min_pages
;
6908 min_pages
= zone
->managed_pages
/ 1024;
6909 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6910 zone
->watermark
[WMARK_MIN
] = min_pages
;
6913 * If it's a lowmem zone, reserve a number of pages
6914 * proportionate to the zone's size.
6916 zone
->watermark
[WMARK_MIN
] = tmp
;
6920 * Set the kswapd watermarks distance according to the
6921 * scale factor in proportion to available memory, but
6922 * ensure a minimum size on small systems.
6924 tmp
= max_t(u64
, tmp
>> 2,
6925 mult_frac(zone
->managed_pages
,
6926 watermark_scale_factor
, 10000));
6928 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6929 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6931 spin_unlock_irqrestore(&zone
->lock
, flags
);
6934 /* update totalreserve_pages */
6935 calculate_totalreserve_pages();
6939 * setup_per_zone_wmarks - called when min_free_kbytes changes
6940 * or when memory is hot-{added|removed}
6942 * Ensures that the watermark[min,low,high] values for each zone are set
6943 * correctly with respect to min_free_kbytes.
6945 void setup_per_zone_wmarks(void)
6947 mutex_lock(&zonelists_mutex
);
6948 __setup_per_zone_wmarks();
6949 mutex_unlock(&zonelists_mutex
);
6953 * Initialise min_free_kbytes.
6955 * For small machines we want it small (128k min). For large machines
6956 * we want it large (64MB max). But it is not linear, because network
6957 * bandwidth does not increase linearly with machine size. We use
6959 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6960 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6976 int __meminit
init_per_zone_wmark_min(void)
6978 unsigned long lowmem_kbytes
;
6979 int new_min_free_kbytes
;
6981 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6982 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6984 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6985 min_free_kbytes
= new_min_free_kbytes
;
6986 if (min_free_kbytes
< 128)
6987 min_free_kbytes
= 128;
6988 if (min_free_kbytes
> 65536)
6989 min_free_kbytes
= 65536;
6991 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6992 new_min_free_kbytes
, user_min_free_kbytes
);
6994 setup_per_zone_wmarks();
6995 refresh_zone_stat_thresholds();
6996 setup_per_zone_lowmem_reserve();
6999 setup_min_unmapped_ratio();
7000 setup_min_slab_ratio();
7005 core_initcall(init_per_zone_wmark_min
)
7008 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7009 * that we can call two helper functions whenever min_free_kbytes
7012 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7013 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7017 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7022 user_min_free_kbytes
= min_free_kbytes
;
7023 setup_per_zone_wmarks();
7028 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7029 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7033 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7038 setup_per_zone_wmarks();
7044 static void setup_min_unmapped_ratio(void)
7049 for_each_online_pgdat(pgdat
)
7050 pgdat
->min_unmapped_pages
= 0;
7053 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
7054 sysctl_min_unmapped_ratio
) / 100;
7058 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7059 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7063 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7067 setup_min_unmapped_ratio();
7072 static void setup_min_slab_ratio(void)
7077 for_each_online_pgdat(pgdat
)
7078 pgdat
->min_slab_pages
= 0;
7081 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
7082 sysctl_min_slab_ratio
) / 100;
7085 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7086 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7090 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7094 setup_min_slab_ratio();
7101 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7102 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7103 * whenever sysctl_lowmem_reserve_ratio changes.
7105 * The reserve ratio obviously has absolutely no relation with the
7106 * minimum watermarks. The lowmem reserve ratio can only make sense
7107 * if in function of the boot time zone sizes.
7109 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7110 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7112 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7113 setup_per_zone_lowmem_reserve();
7118 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7119 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7120 * pagelist can have before it gets flushed back to buddy allocator.
7122 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7123 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7126 int old_percpu_pagelist_fraction
;
7129 mutex_lock(&pcp_batch_high_lock
);
7130 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7132 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7133 if (!write
|| ret
< 0)
7136 /* Sanity checking to avoid pcp imbalance */
7137 if (percpu_pagelist_fraction
&&
7138 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7139 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7145 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7148 for_each_populated_zone(zone
) {
7151 for_each_possible_cpu(cpu
)
7152 pageset_set_high_and_batch(zone
,
7153 per_cpu_ptr(zone
->pageset
, cpu
));
7156 mutex_unlock(&pcp_batch_high_lock
);
7161 int hashdist
= HASHDIST_DEFAULT
;
7163 static int __init
set_hashdist(char *str
)
7167 hashdist
= simple_strtoul(str
, &str
, 0);
7170 __setup("hashdist=", set_hashdist
);
7173 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7175 * Returns the number of pages that arch has reserved but
7176 * is not known to alloc_large_system_hash().
7178 static unsigned long __init
arch_reserved_kernel_pages(void)
7185 * allocate a large system hash table from bootmem
7186 * - it is assumed that the hash table must contain an exact power-of-2
7187 * quantity of entries
7188 * - limit is the number of hash buckets, not the total allocation size
7190 void *__init
alloc_large_system_hash(const char *tablename
,
7191 unsigned long bucketsize
,
7192 unsigned long numentries
,
7195 unsigned int *_hash_shift
,
7196 unsigned int *_hash_mask
,
7197 unsigned long low_limit
,
7198 unsigned long high_limit
)
7200 unsigned long long max
= high_limit
;
7201 unsigned long log2qty
, size
;
7204 /* allow the kernel cmdline to have a say */
7206 /* round applicable memory size up to nearest megabyte */
7207 numentries
= nr_kernel_pages
;
7208 numentries
-= arch_reserved_kernel_pages();
7210 /* It isn't necessary when PAGE_SIZE >= 1MB */
7211 if (PAGE_SHIFT
< 20)
7212 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7214 /* limit to 1 bucket per 2^scale bytes of low memory */
7215 if (scale
> PAGE_SHIFT
)
7216 numentries
>>= (scale
- PAGE_SHIFT
);
7218 numentries
<<= (PAGE_SHIFT
- scale
);
7220 /* Make sure we've got at least a 0-order allocation.. */
7221 if (unlikely(flags
& HASH_SMALL
)) {
7222 /* Makes no sense without HASH_EARLY */
7223 WARN_ON(!(flags
& HASH_EARLY
));
7224 if (!(numentries
>> *_hash_shift
)) {
7225 numentries
= 1UL << *_hash_shift
;
7226 BUG_ON(!numentries
);
7228 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7229 numentries
= PAGE_SIZE
/ bucketsize
;
7231 numentries
= roundup_pow_of_two(numentries
);
7233 /* limit allocation size to 1/16 total memory by default */
7235 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7236 do_div(max
, bucketsize
);
7238 max
= min(max
, 0x80000000ULL
);
7240 if (numentries
< low_limit
)
7241 numentries
= low_limit
;
7242 if (numentries
> max
)
7245 log2qty
= ilog2(numentries
);
7248 size
= bucketsize
<< log2qty
;
7249 if (flags
& HASH_EARLY
)
7250 table
= memblock_virt_alloc_nopanic(size
, 0);
7252 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
7255 * If bucketsize is not a power-of-two, we may free
7256 * some pages at the end of hash table which
7257 * alloc_pages_exact() automatically does
7259 if (get_order(size
) < MAX_ORDER
) {
7260 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
7261 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
7264 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7267 panic("Failed to allocate %s hash table\n", tablename
);
7269 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7270 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7273 *_hash_shift
= log2qty
;
7275 *_hash_mask
= (1 << log2qty
) - 1;
7281 * This function checks whether pageblock includes unmovable pages or not.
7282 * If @count is not zero, it is okay to include less @count unmovable pages
7284 * PageLRU check without isolation or lru_lock could race so that
7285 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7286 * check without lock_page also may miss some movable non-lru pages at
7287 * race condition. So you can't expect this function should be exact.
7289 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7290 bool skip_hwpoisoned_pages
)
7292 unsigned long pfn
, iter
, found
;
7296 * For avoiding noise data, lru_add_drain_all() should be called
7297 * If ZONE_MOVABLE, the zone never contains unmovable pages
7299 if (zone_idx(zone
) == ZONE_MOVABLE
)
7301 mt
= get_pageblock_migratetype(page
);
7302 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7305 pfn
= page_to_pfn(page
);
7306 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7307 unsigned long check
= pfn
+ iter
;
7309 if (!pfn_valid_within(check
))
7312 page
= pfn_to_page(check
);
7315 * Hugepages are not in LRU lists, but they're movable.
7316 * We need not scan over tail pages bacause we don't
7317 * handle each tail page individually in migration.
7319 if (PageHuge(page
)) {
7320 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7325 * We can't use page_count without pin a page
7326 * because another CPU can free compound page.
7327 * This check already skips compound tails of THP
7328 * because their page->_refcount is zero at all time.
7330 if (!page_ref_count(page
)) {
7331 if (PageBuddy(page
))
7332 iter
+= (1 << page_order(page
)) - 1;
7337 * The HWPoisoned page may be not in buddy system, and
7338 * page_count() is not 0.
7340 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7343 if (__PageMovable(page
))
7349 * If there are RECLAIMABLE pages, we need to check
7350 * it. But now, memory offline itself doesn't call
7351 * shrink_node_slabs() and it still to be fixed.
7354 * If the page is not RAM, page_count()should be 0.
7355 * we don't need more check. This is an _used_ not-movable page.
7357 * The problematic thing here is PG_reserved pages. PG_reserved
7358 * is set to both of a memory hole page and a _used_ kernel
7367 bool is_pageblock_removable_nolock(struct page
*page
)
7373 * We have to be careful here because we are iterating over memory
7374 * sections which are not zone aware so we might end up outside of
7375 * the zone but still within the section.
7376 * We have to take care about the node as well. If the node is offline
7377 * its NODE_DATA will be NULL - see page_zone.
7379 if (!node_online(page_to_nid(page
)))
7382 zone
= page_zone(page
);
7383 pfn
= page_to_pfn(page
);
7384 if (!zone_spans_pfn(zone
, pfn
))
7387 return !has_unmovable_pages(zone
, page
, 0, true);
7390 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7392 static unsigned long pfn_max_align_down(unsigned long pfn
)
7394 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7395 pageblock_nr_pages
) - 1);
7398 static unsigned long pfn_max_align_up(unsigned long pfn
)
7400 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7401 pageblock_nr_pages
));
7404 /* [start, end) must belong to a single zone. */
7405 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7406 unsigned long start
, unsigned long end
)
7408 /* This function is based on compact_zone() from compaction.c. */
7409 unsigned long nr_reclaimed
;
7410 unsigned long pfn
= start
;
7411 unsigned int tries
= 0;
7416 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7417 if (fatal_signal_pending(current
)) {
7422 if (list_empty(&cc
->migratepages
)) {
7423 cc
->nr_migratepages
= 0;
7424 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7430 } else if (++tries
== 5) {
7431 ret
= ret
< 0 ? ret
: -EBUSY
;
7435 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7437 cc
->nr_migratepages
-= nr_reclaimed
;
7439 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7440 NULL
, 0, cc
->mode
, MR_CMA
);
7443 putback_movable_pages(&cc
->migratepages
);
7450 * alloc_contig_range() -- tries to allocate given range of pages
7451 * @start: start PFN to allocate
7452 * @end: one-past-the-last PFN to allocate
7453 * @migratetype: migratetype of the underlaying pageblocks (either
7454 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7455 * in range must have the same migratetype and it must
7456 * be either of the two.
7457 * @gfp_mask: GFP mask to use during compaction
7459 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7460 * aligned, however it's the caller's responsibility to guarantee that
7461 * we are the only thread that changes migrate type of pageblocks the
7464 * The PFN range must belong to a single zone.
7466 * Returns zero on success or negative error code. On success all
7467 * pages which PFN is in [start, end) are allocated for the caller and
7468 * need to be freed with free_contig_range().
7470 int alloc_contig_range(unsigned long start
, unsigned long end
,
7471 unsigned migratetype
, gfp_t gfp_mask
)
7473 unsigned long outer_start
, outer_end
;
7477 struct compact_control cc
= {
7478 .nr_migratepages
= 0,
7480 .zone
= page_zone(pfn_to_page(start
)),
7481 .mode
= MIGRATE_SYNC
,
7482 .ignore_skip_hint
= true,
7483 .gfp_mask
= current_gfp_context(gfp_mask
),
7485 INIT_LIST_HEAD(&cc
.migratepages
);
7488 * What we do here is we mark all pageblocks in range as
7489 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7490 * have different sizes, and due to the way page allocator
7491 * work, we align the range to biggest of the two pages so
7492 * that page allocator won't try to merge buddies from
7493 * different pageblocks and change MIGRATE_ISOLATE to some
7494 * other migration type.
7496 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7497 * migrate the pages from an unaligned range (ie. pages that
7498 * we are interested in). This will put all the pages in
7499 * range back to page allocator as MIGRATE_ISOLATE.
7501 * When this is done, we take the pages in range from page
7502 * allocator removing them from the buddy system. This way
7503 * page allocator will never consider using them.
7505 * This lets us mark the pageblocks back as
7506 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7507 * aligned range but not in the unaligned, original range are
7508 * put back to page allocator so that buddy can use them.
7511 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7512 pfn_max_align_up(end
), migratetype
,
7518 * In case of -EBUSY, we'd like to know which page causes problem.
7519 * So, just fall through. We will check it in test_pages_isolated().
7521 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7522 if (ret
&& ret
!= -EBUSY
)
7526 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7527 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7528 * more, all pages in [start, end) are free in page allocator.
7529 * What we are going to do is to allocate all pages from
7530 * [start, end) (that is remove them from page allocator).
7532 * The only problem is that pages at the beginning and at the
7533 * end of interesting range may be not aligned with pages that
7534 * page allocator holds, ie. they can be part of higher order
7535 * pages. Because of this, we reserve the bigger range and
7536 * once this is done free the pages we are not interested in.
7538 * We don't have to hold zone->lock here because the pages are
7539 * isolated thus they won't get removed from buddy.
7542 lru_add_drain_all();
7543 drain_all_pages(cc
.zone
);
7546 outer_start
= start
;
7547 while (!PageBuddy(pfn_to_page(outer_start
))) {
7548 if (++order
>= MAX_ORDER
) {
7549 outer_start
= start
;
7552 outer_start
&= ~0UL << order
;
7555 if (outer_start
!= start
) {
7556 order
= page_order(pfn_to_page(outer_start
));
7559 * outer_start page could be small order buddy page and
7560 * it doesn't include start page. Adjust outer_start
7561 * in this case to report failed page properly
7562 * on tracepoint in test_pages_isolated()
7564 if (outer_start
+ (1UL << order
) <= start
)
7565 outer_start
= start
;
7568 /* Make sure the range is really isolated. */
7569 if (test_pages_isolated(outer_start
, end
, false)) {
7570 pr_info("%s: [%lx, %lx) PFNs busy\n",
7571 __func__
, outer_start
, end
);
7576 /* Grab isolated pages from freelists. */
7577 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7583 /* Free head and tail (if any) */
7584 if (start
!= outer_start
)
7585 free_contig_range(outer_start
, start
- outer_start
);
7586 if (end
!= outer_end
)
7587 free_contig_range(end
, outer_end
- end
);
7590 undo_isolate_page_range(pfn_max_align_down(start
),
7591 pfn_max_align_up(end
), migratetype
);
7595 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7597 unsigned int count
= 0;
7599 for (; nr_pages
--; pfn
++) {
7600 struct page
*page
= pfn_to_page(pfn
);
7602 count
+= page_count(page
) != 1;
7605 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7609 #ifdef CONFIG_MEMORY_HOTPLUG
7611 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7612 * page high values need to be recalulated.
7614 void __meminit
zone_pcp_update(struct zone
*zone
)
7617 mutex_lock(&pcp_batch_high_lock
);
7618 for_each_possible_cpu(cpu
)
7619 pageset_set_high_and_batch(zone
,
7620 per_cpu_ptr(zone
->pageset
, cpu
));
7621 mutex_unlock(&pcp_batch_high_lock
);
7625 void zone_pcp_reset(struct zone
*zone
)
7627 unsigned long flags
;
7629 struct per_cpu_pageset
*pset
;
7631 /* avoid races with drain_pages() */
7632 local_irq_save(flags
);
7633 if (zone
->pageset
!= &boot_pageset
) {
7634 for_each_online_cpu(cpu
) {
7635 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7636 drain_zonestat(zone
, pset
);
7638 free_percpu(zone
->pageset
);
7639 zone
->pageset
= &boot_pageset
;
7641 local_irq_restore(flags
);
7644 #ifdef CONFIG_MEMORY_HOTREMOVE
7646 * All pages in the range must be in a single zone and isolated
7647 * before calling this.
7650 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7654 unsigned int order
, i
;
7656 unsigned long flags
;
7657 /* find the first valid pfn */
7658 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7663 zone
= page_zone(pfn_to_page(pfn
));
7664 spin_lock_irqsave(&zone
->lock
, flags
);
7666 while (pfn
< end_pfn
) {
7667 if (!pfn_valid(pfn
)) {
7671 page
= pfn_to_page(pfn
);
7673 * The HWPoisoned page may be not in buddy system, and
7674 * page_count() is not 0.
7676 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7678 SetPageReserved(page
);
7682 BUG_ON(page_count(page
));
7683 BUG_ON(!PageBuddy(page
));
7684 order
= page_order(page
);
7685 #ifdef CONFIG_DEBUG_VM
7686 pr_info("remove from free list %lx %d %lx\n",
7687 pfn
, 1 << order
, end_pfn
);
7689 list_del(&page
->lru
);
7690 rmv_page_order(page
);
7691 zone
->free_area
[order
].nr_free
--;
7692 for (i
= 0; i
< (1 << order
); i
++)
7693 SetPageReserved((page
+i
));
7694 pfn
+= (1 << order
);
7696 spin_unlock_irqrestore(&zone
->lock
, flags
);
7700 bool is_free_buddy_page(struct page
*page
)
7702 struct zone
*zone
= page_zone(page
);
7703 unsigned long pfn
= page_to_pfn(page
);
7704 unsigned long flags
;
7707 spin_lock_irqsave(&zone
->lock
, flags
);
7708 for (order
= 0; order
< MAX_ORDER
; order
++) {
7709 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7711 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7714 spin_unlock_irqrestore(&zone
->lock
, flags
);
7716 return order
< MAX_ORDER
;