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 <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
67 #include <asm/sections.h>
68 #include <asm/tlbflush.h>
69 #include <asm/div64.h>
72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73 static DEFINE_MUTEX(pcp_batch_high_lock
);
74 #define MIN_PERCPU_PAGELIST_FRACTION (8)
76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77 DEFINE_PER_CPU(int, numa_node
);
78 EXPORT_PER_CPU_SYMBOL(numa_node
);
81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
88 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
89 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
90 int _node_numa_mem_
[MAX_NUMNODES
];
94 * Array of node states.
96 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
97 [N_POSSIBLE
] = NODE_MASK_ALL
,
98 [N_ONLINE
] = { { [0] = 1UL } },
100 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
104 #ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY
] = { { [0] = 1UL } },
107 [N_CPU
] = { { [0] = 1UL } },
110 EXPORT_SYMBOL(node_states
);
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock
);
115 unsigned long totalram_pages __read_mostly
;
116 unsigned long totalreserve_pages __read_mostly
;
117 unsigned long totalcma_pages __read_mostly
;
119 int percpu_pagelist_fraction
;
120 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
130 static inline int get_pcppage_migratetype(struct page
*page
)
135 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
137 page
->index
= migratetype
;
140 #ifdef CONFIG_PM_SLEEP
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
150 static gfp_t saved_gfp_mask
;
152 void pm_restore_gfp_mask(void)
154 WARN_ON(!mutex_is_locked(&pm_mutex
));
155 if (saved_gfp_mask
) {
156 gfp_allowed_mask
= saved_gfp_mask
;
161 void pm_restrict_gfp_mask(void)
163 WARN_ON(!mutex_is_locked(&pm_mutex
));
164 WARN_ON(saved_gfp_mask
);
165 saved_gfp_mask
= gfp_allowed_mask
;
166 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
169 bool pm_suspended_storage(void)
171 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
175 #endif /* CONFIG_PM_SLEEP */
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly
;
181 static void __free_pages_ok(struct page
*page
, unsigned int order
);
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
194 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 EXPORT_SYMBOL(totalram_pages
);
209 static char * const zone_names
[MAX_NR_ZONES
] = {
210 #ifdef CONFIG_ZONE_DMA
213 #ifdef CONFIG_ZONE_DMA32
217 #ifdef CONFIG_HIGHMEM
221 #ifdef CONFIG_ZONE_DEVICE
226 char * const migratetype_names
[MIGRATE_TYPES
] = {
234 #ifdef CONFIG_MEMORY_ISOLATION
239 compound_page_dtor
* const compound_page_dtors
[] = {
242 #ifdef CONFIG_HUGETLB_PAGE
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
250 int min_free_kbytes
= 1024;
251 int user_min_free_kbytes
= -1;
252 int watermark_scale_factor
= 10;
254 static unsigned long __meminitdata nr_kernel_pages
;
255 static unsigned long __meminitdata nr_all_pages
;
256 static unsigned long __meminitdata dma_reserve
;
258 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
259 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
260 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
261 static unsigned long __initdata required_kernelcore
;
262 static unsigned long __initdata required_movablecore
;
263 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
264 static bool mirrored_kernelcore
;
266 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
268 EXPORT_SYMBOL(movable_zone
);
269 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
272 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
273 int nr_online_nodes __read_mostly
= 1;
274 EXPORT_SYMBOL(nr_node_ids
);
275 EXPORT_SYMBOL(nr_online_nodes
);
278 int page_group_by_mobility_disabled __read_mostly
;
280 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
281 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
283 pgdat
->first_deferred_pfn
= ULONG_MAX
;
286 /* Returns true if the struct page for the pfn is uninitialised */
287 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
289 if (pfn
>= NODE_DATA(early_pfn_to_nid(pfn
))->first_deferred_pfn
)
295 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
297 if (pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
307 static inline bool update_defer_init(pg_data_t
*pgdat
,
308 unsigned long pfn
, unsigned long zone_end
,
309 unsigned long *nr_initialised
)
311 unsigned long max_initialise
;
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end
< pgdat_end_pfn(pgdat
))
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
320 max_initialise
= max(2UL << (30 - PAGE_SHIFT
),
321 (pgdat
->node_spanned_pages
>> 8));
324 if ((*nr_initialised
> max_initialise
) &&
325 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
326 pgdat
->first_deferred_pfn
= pfn
;
333 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
337 static inline bool early_page_uninitialised(unsigned long pfn
)
342 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
347 static inline bool update_defer_init(pg_data_t
*pgdat
,
348 unsigned long pfn
, unsigned long zone_end
,
349 unsigned long *nr_initialised
)
355 /* Return a pointer to the bitmap storing bits affecting a block of pages */
356 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
359 #ifdef CONFIG_SPARSEMEM
360 return __pfn_to_section(pfn
)->pageblock_flags
;
362 return page_zone(page
)->pageblock_flags
;
363 #endif /* CONFIG_SPARSEMEM */
366 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
368 #ifdef CONFIG_SPARSEMEM
369 pfn
&= (PAGES_PER_SECTION
-1);
370 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
372 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
373 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
374 #endif /* CONFIG_SPARSEMEM */
378 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
379 * @page: The page within the block of interest
380 * @pfn: The target page frame number
381 * @end_bitidx: The last bit of interest to retrieve
382 * @mask: mask of bits that the caller is interested in
384 * Return: pageblock_bits flags
386 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
388 unsigned long end_bitidx
,
391 unsigned long *bitmap
;
392 unsigned long bitidx
, word_bitidx
;
395 bitmap
= get_pageblock_bitmap(page
, pfn
);
396 bitidx
= pfn_to_bitidx(page
, pfn
);
397 word_bitidx
= bitidx
/ BITS_PER_LONG
;
398 bitidx
&= (BITS_PER_LONG
-1);
400 word
= bitmap
[word_bitidx
];
401 bitidx
+= end_bitidx
;
402 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
405 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
406 unsigned long end_bitidx
,
409 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
412 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
414 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
418 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
419 * @page: The page within the block of interest
420 * @flags: The flags to set
421 * @pfn: The target page frame number
422 * @end_bitidx: The last bit of interest
423 * @mask: mask of bits that the caller is interested in
425 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
427 unsigned long end_bitidx
,
430 unsigned long *bitmap
;
431 unsigned long bitidx
, word_bitidx
;
432 unsigned long old_word
, word
;
434 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
436 bitmap
= get_pageblock_bitmap(page
, pfn
);
437 bitidx
= pfn_to_bitidx(page
, pfn
);
438 word_bitidx
= bitidx
/ BITS_PER_LONG
;
439 bitidx
&= (BITS_PER_LONG
-1);
441 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
443 bitidx
+= end_bitidx
;
444 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
445 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
447 word
= READ_ONCE(bitmap
[word_bitidx
]);
449 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
450 if (word
== old_word
)
456 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
458 if (unlikely(page_group_by_mobility_disabled
&&
459 migratetype
< MIGRATE_PCPTYPES
))
460 migratetype
= MIGRATE_UNMOVABLE
;
462 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
463 PB_migrate
, PB_migrate_end
);
466 #ifdef CONFIG_DEBUG_VM
467 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
471 unsigned long pfn
= page_to_pfn(page
);
472 unsigned long sp
, start_pfn
;
475 seq
= zone_span_seqbegin(zone
);
476 start_pfn
= zone
->zone_start_pfn
;
477 sp
= zone
->spanned_pages
;
478 if (!zone_spans_pfn(zone
, pfn
))
480 } while (zone_span_seqretry(zone
, seq
));
483 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
484 pfn
, zone_to_nid(zone
), zone
->name
,
485 start_pfn
, start_pfn
+ sp
);
490 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
492 if (!pfn_valid_within(page_to_pfn(page
)))
494 if (zone
!= page_zone(page
))
500 * Temporary debugging check for pages not lying within a given zone.
502 static int bad_range(struct zone
*zone
, struct page
*page
)
504 if (page_outside_zone_boundaries(zone
, page
))
506 if (!page_is_consistent(zone
, page
))
512 static inline int bad_range(struct zone
*zone
, struct page
*page
)
518 static void bad_page(struct page
*page
, const char *reason
,
519 unsigned long bad_flags
)
521 static unsigned long resume
;
522 static unsigned long nr_shown
;
523 static unsigned long nr_unshown
;
525 /* Don't complain about poisoned pages */
526 if (PageHWPoison(page
)) {
527 page_mapcount_reset(page
); /* remove PageBuddy */
532 * Allow a burst of 60 reports, then keep quiet for that minute;
533 * or allow a steady drip of one report per second.
535 if (nr_shown
== 60) {
536 if (time_before(jiffies
, resume
)) {
542 "BUG: Bad page state: %lu messages suppressed\n",
549 resume
= jiffies
+ 60 * HZ
;
551 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
552 current
->comm
, page_to_pfn(page
));
553 __dump_page(page
, reason
);
554 bad_flags
&= page
->flags
;
556 pr_alert("bad because of flags: %#lx(%pGp)\n",
557 bad_flags
, &bad_flags
);
558 dump_page_owner(page
);
563 /* Leave bad fields for debug, except PageBuddy could make trouble */
564 page_mapcount_reset(page
); /* remove PageBuddy */
565 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
569 * Higher-order pages are called "compound pages". They are structured thusly:
571 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
573 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
574 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
576 * The first tail page's ->compound_dtor holds the offset in array of compound
577 * page destructors. See compound_page_dtors.
579 * The first tail page's ->compound_order holds the order of allocation.
580 * This usage means that zero-order pages may not be compound.
583 void free_compound_page(struct page
*page
)
585 __free_pages_ok(page
, compound_order(page
));
588 void prep_compound_page(struct page
*page
, unsigned int order
)
591 int nr_pages
= 1 << order
;
593 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
594 set_compound_order(page
, order
);
596 for (i
= 1; i
< nr_pages
; i
++) {
597 struct page
*p
= page
+ i
;
598 set_page_count(p
, 0);
599 p
->mapping
= TAIL_MAPPING
;
600 set_compound_head(p
, page
);
602 atomic_set(compound_mapcount_ptr(page
), -1);
605 #ifdef CONFIG_DEBUG_PAGEALLOC
606 unsigned int _debug_guardpage_minorder
;
607 bool _debug_pagealloc_enabled __read_mostly
608 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
609 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
610 bool _debug_guardpage_enabled __read_mostly
;
612 static int __init
early_debug_pagealloc(char *buf
)
617 if (strcmp(buf
, "on") == 0)
618 _debug_pagealloc_enabled
= true;
620 if (strcmp(buf
, "off") == 0)
621 _debug_pagealloc_enabled
= false;
625 early_param("debug_pagealloc", early_debug_pagealloc
);
627 static bool need_debug_guardpage(void)
629 /* If we don't use debug_pagealloc, we don't need guard page */
630 if (!debug_pagealloc_enabled())
636 static void init_debug_guardpage(void)
638 if (!debug_pagealloc_enabled())
641 _debug_guardpage_enabled
= true;
644 struct page_ext_operations debug_guardpage_ops
= {
645 .need
= need_debug_guardpage
,
646 .init
= init_debug_guardpage
,
649 static int __init
debug_guardpage_minorder_setup(char *buf
)
653 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
654 pr_err("Bad debug_guardpage_minorder value\n");
657 _debug_guardpage_minorder
= res
;
658 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
661 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
663 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
664 unsigned int order
, int migratetype
)
666 struct page_ext
*page_ext
;
668 if (!debug_guardpage_enabled())
671 page_ext
= lookup_page_ext(page
);
672 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
674 INIT_LIST_HEAD(&page
->lru
);
675 set_page_private(page
, order
);
676 /* Guard pages are not available for any usage */
677 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
680 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
681 unsigned int order
, int migratetype
)
683 struct page_ext
*page_ext
;
685 if (!debug_guardpage_enabled())
688 page_ext
= lookup_page_ext(page
);
689 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
691 set_page_private(page
, 0);
692 if (!is_migrate_isolate(migratetype
))
693 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
696 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
697 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
698 unsigned int order
, int migratetype
) {}
699 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
700 unsigned int order
, int migratetype
) {}
703 static inline void set_page_order(struct page
*page
, unsigned int order
)
705 set_page_private(page
, order
);
706 __SetPageBuddy(page
);
709 static inline void rmv_page_order(struct page
*page
)
711 __ClearPageBuddy(page
);
712 set_page_private(page
, 0);
716 * This function checks whether a page is free && is the buddy
717 * we can do coalesce a page and its buddy if
718 * (a) the buddy is not in a hole &&
719 * (b) the buddy is in the buddy system &&
720 * (c) a page and its buddy have the same order &&
721 * (d) a page and its buddy are in the same zone.
723 * For recording whether a page is in the buddy system, we set ->_mapcount
724 * PAGE_BUDDY_MAPCOUNT_VALUE.
725 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
726 * serialized by zone->lock.
728 * For recording page's order, we use page_private(page).
730 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
733 if (!pfn_valid_within(page_to_pfn(buddy
)))
736 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
737 if (page_zone_id(page
) != page_zone_id(buddy
))
740 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
745 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
747 * zone check is done late to avoid uselessly
748 * calculating zone/node ids for pages that could
751 if (page_zone_id(page
) != page_zone_id(buddy
))
754 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
762 * Freeing function for a buddy system allocator.
764 * The concept of a buddy system is to maintain direct-mapped table
765 * (containing bit values) for memory blocks of various "orders".
766 * The bottom level table contains the map for the smallest allocatable
767 * units of memory (here, pages), and each level above it describes
768 * pairs of units from the levels below, hence, "buddies".
769 * At a high level, all that happens here is marking the table entry
770 * at the bottom level available, and propagating the changes upward
771 * as necessary, plus some accounting needed to play nicely with other
772 * parts of the VM system.
773 * At each level, we keep a list of pages, which are heads of continuous
774 * free pages of length of (1 << order) and marked with _mapcount
775 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
777 * So when we are allocating or freeing one, we can derive the state of the
778 * other. That is, if we allocate a small block, and both were
779 * free, the remainder of the region must be split into blocks.
780 * If a block is freed, and its buddy is also free, then this
781 * triggers coalescing into a block of larger size.
786 static inline void __free_one_page(struct page
*page
,
788 struct zone
*zone
, unsigned int order
,
791 unsigned long page_idx
;
792 unsigned long combined_idx
;
793 unsigned long uninitialized_var(buddy_idx
);
795 unsigned int max_order
;
797 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
799 VM_BUG_ON(!zone_is_initialized(zone
));
800 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
802 VM_BUG_ON(migratetype
== -1);
803 if (likely(!is_migrate_isolate(migratetype
)))
804 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
806 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
808 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
809 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
812 while (order
< max_order
- 1) {
813 buddy_idx
= __find_buddy_index(page_idx
, order
);
814 buddy
= page
+ (buddy_idx
- page_idx
);
815 if (!page_is_buddy(page
, buddy
, order
))
818 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
819 * merge with it and move up one order.
821 if (page_is_guard(buddy
)) {
822 clear_page_guard(zone
, buddy
, order
, migratetype
);
824 list_del(&buddy
->lru
);
825 zone
->free_area
[order
].nr_free
--;
826 rmv_page_order(buddy
);
828 combined_idx
= buddy_idx
& page_idx
;
829 page
= page
+ (combined_idx
- page_idx
);
830 page_idx
= combined_idx
;
833 if (max_order
< MAX_ORDER
) {
834 /* If we are here, it means order is >= pageblock_order.
835 * We want to prevent merge between freepages on isolate
836 * pageblock and normal pageblock. Without this, pageblock
837 * isolation could cause incorrect freepage or CMA accounting.
839 * We don't want to hit this code for the more frequent
842 if (unlikely(has_isolate_pageblock(zone
))) {
845 buddy_idx
= __find_buddy_index(page_idx
, order
);
846 buddy
= page
+ (buddy_idx
- page_idx
);
847 buddy_mt
= get_pageblock_migratetype(buddy
);
849 if (migratetype
!= buddy_mt
850 && (is_migrate_isolate(migratetype
) ||
851 is_migrate_isolate(buddy_mt
)))
855 goto continue_merging
;
859 set_page_order(page
, order
);
862 * If this is not the largest possible page, check if the buddy
863 * of the next-highest order is free. If it is, it's possible
864 * that pages are being freed that will coalesce soon. In case,
865 * that is happening, add the free page to the tail of the list
866 * so it's less likely to be used soon and more likely to be merged
867 * as a higher order page
869 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
870 struct page
*higher_page
, *higher_buddy
;
871 combined_idx
= buddy_idx
& page_idx
;
872 higher_page
= page
+ (combined_idx
- page_idx
);
873 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
874 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
875 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
876 list_add_tail(&page
->lru
,
877 &zone
->free_area
[order
].free_list
[migratetype
]);
882 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
884 zone
->free_area
[order
].nr_free
++;
888 * A bad page could be due to a number of fields. Instead of multiple branches,
889 * try and check multiple fields with one check. The caller must do a detailed
890 * check if necessary.
892 static inline bool page_expected_state(struct page
*page
,
893 unsigned long check_flags
)
895 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
898 if (unlikely((unsigned long)page
->mapping
|
899 page_ref_count(page
) |
901 (unsigned long)page
->mem_cgroup
|
903 (page
->flags
& check_flags
)))
909 static void free_pages_check_bad(struct page
*page
)
911 const char *bad_reason
;
912 unsigned long bad_flags
;
917 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
918 bad_reason
= "nonzero mapcount";
919 if (unlikely(page
->mapping
!= NULL
))
920 bad_reason
= "non-NULL mapping";
921 if (unlikely(page_ref_count(page
) != 0))
922 bad_reason
= "nonzero _refcount";
923 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
924 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
925 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
928 if (unlikely(page
->mem_cgroup
))
929 bad_reason
= "page still charged to cgroup";
931 bad_page(page
, bad_reason
, bad_flags
);
934 static inline int free_pages_check(struct page
*page
)
936 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
939 /* Something has gone sideways, find it */
940 free_pages_check_bad(page
);
945 * Frees a number of pages from the PCP lists
946 * Assumes all pages on list are in same zone, and of same order.
947 * count is the number of pages to free.
949 * If the zone was previously in an "all pages pinned" state then look to
950 * see if this freeing clears that state.
952 * And clear the zone's pages_scanned counter, to hold off the "all pages are
953 * pinned" detection logic.
955 static void free_pcppages_bulk(struct zone
*zone
, int count
,
956 struct per_cpu_pages
*pcp
)
960 unsigned long nr_scanned
;
961 bool isolated_pageblocks
;
963 spin_lock(&zone
->lock
);
964 isolated_pageblocks
= has_isolate_pageblock(zone
);
965 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
967 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
971 struct list_head
*list
;
974 * Remove pages from lists in a round-robin fashion. A
975 * batch_free count is maintained that is incremented when an
976 * empty list is encountered. This is so more pages are freed
977 * off fuller lists instead of spinning excessively around empty
982 if (++migratetype
== MIGRATE_PCPTYPES
)
984 list
= &pcp
->lists
[migratetype
];
985 } while (list_empty(list
));
987 /* This is the only non-empty list. Free them all. */
988 if (batch_free
== MIGRATE_PCPTYPES
)
992 int mt
; /* migratetype of the to-be-freed page */
994 page
= list_last_entry(list
, struct page
, lru
);
995 /* must delete as __free_one_page list manipulates */
996 list_del(&page
->lru
);
998 mt
= get_pcppage_migratetype(page
);
999 /* MIGRATE_ISOLATE page should not go to pcplists */
1000 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1001 /* Pageblock could have been isolated meanwhile */
1002 if (unlikely(isolated_pageblocks
))
1003 mt
= get_pageblock_migratetype(page
);
1005 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1006 trace_mm_page_pcpu_drain(page
, 0, mt
);
1007 } while (--count
&& --batch_free
&& !list_empty(list
));
1009 spin_unlock(&zone
->lock
);
1012 static void free_one_page(struct zone
*zone
,
1013 struct page
*page
, unsigned long pfn
,
1017 unsigned long nr_scanned
;
1018 spin_lock(&zone
->lock
);
1019 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
1021 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
1023 if (unlikely(has_isolate_pageblock(zone
) ||
1024 is_migrate_isolate(migratetype
))) {
1025 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1027 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1028 spin_unlock(&zone
->lock
);
1031 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1036 * We rely page->lru.next never has bit 0 set, unless the page
1037 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1039 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1041 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1045 switch (page
- head_page
) {
1047 /* the first tail page: ->mapping is compound_mapcount() */
1048 if (unlikely(compound_mapcount(page
))) {
1049 bad_page(page
, "nonzero compound_mapcount", 0);
1055 * the second tail page: ->mapping is
1056 * page_deferred_list().next -- ignore value.
1060 if (page
->mapping
!= TAIL_MAPPING
) {
1061 bad_page(page
, "corrupted mapping in tail page", 0);
1066 if (unlikely(!PageTail(page
))) {
1067 bad_page(page
, "PageTail not set", 0);
1070 if (unlikely(compound_head(page
) != head_page
)) {
1071 bad_page(page
, "compound_head not consistent", 0);
1076 page
->mapping
= NULL
;
1077 clear_compound_head(page
);
1081 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1082 unsigned long zone
, int nid
)
1084 set_page_links(page
, zone
, nid
, pfn
);
1085 init_page_count(page
);
1086 page_mapcount_reset(page
);
1087 page_cpupid_reset_last(page
);
1089 INIT_LIST_HEAD(&page
->lru
);
1090 #ifdef WANT_PAGE_VIRTUAL
1091 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1092 if (!is_highmem_idx(zone
))
1093 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1097 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1100 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1103 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1104 static void init_reserved_page(unsigned long pfn
)
1109 if (!early_page_uninitialised(pfn
))
1112 nid
= early_pfn_to_nid(pfn
);
1113 pgdat
= NODE_DATA(nid
);
1115 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1116 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1118 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1121 __init_single_pfn(pfn
, zid
, nid
);
1124 static inline void init_reserved_page(unsigned long pfn
)
1127 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1130 * Initialised pages do not have PageReserved set. This function is
1131 * called for each range allocated by the bootmem allocator and
1132 * marks the pages PageReserved. The remaining valid pages are later
1133 * sent to the buddy page allocator.
1135 void __meminit
reserve_bootmem_region(unsigned long start
, unsigned long end
)
1137 unsigned long start_pfn
= PFN_DOWN(start
);
1138 unsigned long end_pfn
= PFN_UP(end
);
1140 for (; start_pfn
< end_pfn
; start_pfn
++) {
1141 if (pfn_valid(start_pfn
)) {
1142 struct page
*page
= pfn_to_page(start_pfn
);
1144 init_reserved_page(start_pfn
);
1146 /* Avoid false-positive PageTail() */
1147 INIT_LIST_HEAD(&page
->lru
);
1149 SetPageReserved(page
);
1154 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
1158 VM_BUG_ON_PAGE(PageTail(page
), page
);
1160 trace_mm_page_free(page
, order
);
1161 kmemcheck_free_shadow(page
, order
);
1162 kasan_free_pages(page
, order
);
1165 * Check tail pages before head page information is cleared to
1166 * avoid checking PageCompound for order-0 pages.
1168 if (unlikely(order
)) {
1169 bool compound
= PageCompound(page
);
1172 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1174 for (i
= 1; i
< (1 << order
); i
++) {
1176 bad
+= free_tail_pages_check(page
, page
+ i
);
1177 if (unlikely(free_pages_check(page
+ i
))) {
1181 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1184 if (PageAnonHead(page
))
1185 page
->mapping
= NULL
;
1186 bad
+= free_pages_check(page
);
1190 page_cpupid_reset_last(page
);
1191 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1192 reset_page_owner(page
, order
);
1194 if (!PageHighMem(page
)) {
1195 debug_check_no_locks_freed(page_address(page
),
1196 PAGE_SIZE
<< order
);
1197 debug_check_no_obj_freed(page_address(page
),
1198 PAGE_SIZE
<< order
);
1200 arch_free_page(page
, order
);
1201 kernel_poison_pages(page
, 1 << order
, 0);
1202 kernel_map_pages(page
, 1 << order
, 0);
1207 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1209 unsigned long flags
;
1211 unsigned long pfn
= page_to_pfn(page
);
1213 if (!free_pages_prepare(page
, order
))
1216 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1217 local_irq_save(flags
);
1218 __count_vm_events(PGFREE
, 1 << order
);
1219 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1220 local_irq_restore(flags
);
1223 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1225 unsigned int nr_pages
= 1 << order
;
1226 struct page
*p
= page
;
1230 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1232 __ClearPageReserved(p
);
1233 set_page_count(p
, 0);
1235 __ClearPageReserved(p
);
1236 set_page_count(p
, 0);
1238 page_zone(page
)->managed_pages
+= nr_pages
;
1239 set_page_refcounted(page
);
1240 __free_pages(page
, order
);
1243 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1244 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1246 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1248 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1250 static DEFINE_SPINLOCK(early_pfn_lock
);
1253 spin_lock(&early_pfn_lock
);
1254 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1257 spin_unlock(&early_pfn_lock
);
1263 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1264 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1265 struct mminit_pfnnid_cache
*state
)
1269 nid
= __early_pfn_to_nid(pfn
, state
);
1270 if (nid
>= 0 && nid
!= node
)
1275 /* Only safe to use early in boot when initialisation is single-threaded */
1276 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1278 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1283 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1287 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1288 struct mminit_pfnnid_cache
*state
)
1295 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1298 if (early_page_uninitialised(pfn
))
1300 return __free_pages_boot_core(page
, order
);
1304 * Check that the whole (or subset of) a pageblock given by the interval of
1305 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1306 * with the migration of free compaction scanner. The scanners then need to
1307 * use only pfn_valid_within() check for arches that allow holes within
1310 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1312 * It's possible on some configurations to have a setup like node0 node1 node0
1313 * i.e. it's possible that all pages within a zones range of pages do not
1314 * belong to a single zone. We assume that a border between node0 and node1
1315 * can occur within a single pageblock, but not a node0 node1 node0
1316 * interleaving within a single pageblock. It is therefore sufficient to check
1317 * the first and last page of a pageblock and avoid checking each individual
1318 * page in a pageblock.
1320 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1321 unsigned long end_pfn
, struct zone
*zone
)
1323 struct page
*start_page
;
1324 struct page
*end_page
;
1326 /* end_pfn is one past the range we are checking */
1329 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1332 start_page
= pfn_to_page(start_pfn
);
1334 if (page_zone(start_page
) != zone
)
1337 end_page
= pfn_to_page(end_pfn
);
1339 /* This gives a shorter code than deriving page_zone(end_page) */
1340 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1346 void set_zone_contiguous(struct zone
*zone
)
1348 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1349 unsigned long block_end_pfn
;
1351 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1352 for (; block_start_pfn
< zone_end_pfn(zone
);
1353 block_start_pfn
= block_end_pfn
,
1354 block_end_pfn
+= pageblock_nr_pages
) {
1356 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1358 if (!__pageblock_pfn_to_page(block_start_pfn
,
1359 block_end_pfn
, zone
))
1363 /* We confirm that there is no hole */
1364 zone
->contiguous
= true;
1367 void clear_zone_contiguous(struct zone
*zone
)
1369 zone
->contiguous
= false;
1372 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1373 static void __init
deferred_free_range(struct page
*page
,
1374 unsigned long pfn
, int nr_pages
)
1381 /* Free a large naturally-aligned chunk if possible */
1382 if (nr_pages
== MAX_ORDER_NR_PAGES
&&
1383 (pfn
& (MAX_ORDER_NR_PAGES
-1)) == 0) {
1384 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1385 __free_pages_boot_core(page
, MAX_ORDER
-1);
1389 for (i
= 0; i
< nr_pages
; i
++, page
++)
1390 __free_pages_boot_core(page
, 0);
1393 /* Completion tracking for deferred_init_memmap() threads */
1394 static atomic_t pgdat_init_n_undone __initdata
;
1395 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1397 static inline void __init
pgdat_init_report_one_done(void)
1399 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1400 complete(&pgdat_init_all_done_comp
);
1403 /* Initialise remaining memory on a node */
1404 static int __init
deferred_init_memmap(void *data
)
1406 pg_data_t
*pgdat
= data
;
1407 int nid
= pgdat
->node_id
;
1408 struct mminit_pfnnid_cache nid_init_state
= { };
1409 unsigned long start
= jiffies
;
1410 unsigned long nr_pages
= 0;
1411 unsigned long walk_start
, walk_end
;
1414 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1415 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1417 if (first_init_pfn
== ULONG_MAX
) {
1418 pgdat_init_report_one_done();
1422 /* Bind memory initialisation thread to a local node if possible */
1423 if (!cpumask_empty(cpumask
))
1424 set_cpus_allowed_ptr(current
, cpumask
);
1426 /* Sanity check boundaries */
1427 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1428 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1429 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1431 /* Only the highest zone is deferred so find it */
1432 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1433 zone
= pgdat
->node_zones
+ zid
;
1434 if (first_init_pfn
< zone_end_pfn(zone
))
1438 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1439 unsigned long pfn
, end_pfn
;
1440 struct page
*page
= NULL
;
1441 struct page
*free_base_page
= NULL
;
1442 unsigned long free_base_pfn
= 0;
1445 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1446 pfn
= first_init_pfn
;
1447 if (pfn
< walk_start
)
1449 if (pfn
< zone
->zone_start_pfn
)
1450 pfn
= zone
->zone_start_pfn
;
1452 for (; pfn
< end_pfn
; pfn
++) {
1453 if (!pfn_valid_within(pfn
))
1457 * Ensure pfn_valid is checked every
1458 * MAX_ORDER_NR_PAGES for memory holes
1460 if ((pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
1461 if (!pfn_valid(pfn
)) {
1467 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1472 /* Minimise pfn page lookups and scheduler checks */
1473 if (page
&& (pfn
& (MAX_ORDER_NR_PAGES
- 1)) != 0) {
1476 nr_pages
+= nr_to_free
;
1477 deferred_free_range(free_base_page
,
1478 free_base_pfn
, nr_to_free
);
1479 free_base_page
= NULL
;
1480 free_base_pfn
= nr_to_free
= 0;
1482 page
= pfn_to_page(pfn
);
1487 VM_BUG_ON(page_zone(page
) != zone
);
1491 __init_single_page(page
, pfn
, zid
, nid
);
1492 if (!free_base_page
) {
1493 free_base_page
= page
;
1494 free_base_pfn
= pfn
;
1499 /* Where possible, batch up pages for a single free */
1502 /* Free the current block of pages to allocator */
1503 nr_pages
+= nr_to_free
;
1504 deferred_free_range(free_base_page
, free_base_pfn
,
1506 free_base_page
= NULL
;
1507 free_base_pfn
= nr_to_free
= 0;
1510 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1513 /* Sanity check that the next zone really is unpopulated */
1514 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1516 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1517 jiffies_to_msecs(jiffies
- start
));
1519 pgdat_init_report_one_done();
1522 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1524 void __init
page_alloc_init_late(void)
1528 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1531 /* There will be num_node_state(N_MEMORY) threads */
1532 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1533 for_each_node_state(nid
, N_MEMORY
) {
1534 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1537 /* Block until all are initialised */
1538 wait_for_completion(&pgdat_init_all_done_comp
);
1540 /* Reinit limits that are based on free pages after the kernel is up */
1541 files_maxfiles_init();
1544 for_each_populated_zone(zone
)
1545 set_zone_contiguous(zone
);
1549 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1550 void __init
init_cma_reserved_pageblock(struct page
*page
)
1552 unsigned i
= pageblock_nr_pages
;
1553 struct page
*p
= page
;
1556 __ClearPageReserved(p
);
1557 set_page_count(p
, 0);
1560 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1562 if (pageblock_order
>= MAX_ORDER
) {
1563 i
= pageblock_nr_pages
;
1566 set_page_refcounted(p
);
1567 __free_pages(p
, MAX_ORDER
- 1);
1568 p
+= MAX_ORDER_NR_PAGES
;
1569 } while (i
-= MAX_ORDER_NR_PAGES
);
1571 set_page_refcounted(page
);
1572 __free_pages(page
, pageblock_order
);
1575 adjust_managed_page_count(page
, pageblock_nr_pages
);
1580 * The order of subdivision here is critical for the IO subsystem.
1581 * Please do not alter this order without good reasons and regression
1582 * testing. Specifically, as large blocks of memory are subdivided,
1583 * the order in which smaller blocks are delivered depends on the order
1584 * they're subdivided in this function. This is the primary factor
1585 * influencing the order in which pages are delivered to the IO
1586 * subsystem according to empirical testing, and this is also justified
1587 * by considering the behavior of a buddy system containing a single
1588 * large block of memory acted on by a series of small allocations.
1589 * This behavior is a critical factor in sglist merging's success.
1593 static inline void expand(struct zone
*zone
, struct page
*page
,
1594 int low
, int high
, struct free_area
*area
,
1597 unsigned long size
= 1 << high
;
1599 while (high
> low
) {
1603 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1605 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
1606 debug_guardpage_enabled() &&
1607 high
< debug_guardpage_minorder()) {
1609 * Mark as guard pages (or page), that will allow to
1610 * merge back to allocator when buddy will be freed.
1611 * Corresponding page table entries will not be touched,
1612 * pages will stay not present in virtual address space
1614 set_page_guard(zone
, &page
[size
], high
, migratetype
);
1617 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1619 set_page_order(&page
[size
], high
);
1624 * This page is about to be returned from the page allocator
1626 static inline int check_new_page(struct page
*page
)
1628 const char *bad_reason
;
1629 unsigned long bad_flags
;
1631 if (page_expected_state(page
, PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
))
1636 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1637 bad_reason
= "nonzero mapcount";
1638 if (unlikely(page
->mapping
!= NULL
))
1639 bad_reason
= "non-NULL mapping";
1640 if (unlikely(page_ref_count(page
) != 0))
1641 bad_reason
= "nonzero _count";
1642 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1643 bad_reason
= "HWPoisoned (hardware-corrupted)";
1644 bad_flags
= __PG_HWPOISON
;
1646 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1647 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1648 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1651 if (unlikely(page
->mem_cgroup
))
1652 bad_reason
= "page still charged to cgroup";
1654 if (unlikely(bad_reason
)) {
1655 bad_page(page
, bad_reason
, bad_flags
);
1661 static inline bool free_pages_prezeroed(bool poisoned
)
1663 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1664 page_poisoning_enabled() && poisoned
;
1667 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1668 unsigned int alloc_flags
)
1671 bool poisoned
= true;
1673 for (i
= 0; i
< (1 << order
); i
++) {
1674 struct page
*p
= page
+ i
;
1675 if (unlikely(check_new_page(p
)))
1678 poisoned
&= page_is_poisoned(p
);
1681 set_page_private(page
, 0);
1682 set_page_refcounted(page
);
1684 arch_alloc_page(page
, order
);
1685 kernel_map_pages(page
, 1 << order
, 1);
1686 kernel_poison_pages(page
, 1 << order
, 1);
1687 kasan_alloc_pages(page
, order
);
1689 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1690 for (i
= 0; i
< (1 << order
); i
++)
1691 clear_highpage(page
+ i
);
1693 if (order
&& (gfp_flags
& __GFP_COMP
))
1694 prep_compound_page(page
, order
);
1696 set_page_owner(page
, order
, gfp_flags
);
1699 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1700 * allocate the page. The expectation is that the caller is taking
1701 * steps that will free more memory. The caller should avoid the page
1702 * being used for !PFMEMALLOC purposes.
1704 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1705 set_page_pfmemalloc(page
);
1707 clear_page_pfmemalloc(page
);
1713 * Go through the free lists for the given migratetype and remove
1714 * the smallest available page from the freelists
1717 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1720 unsigned int current_order
;
1721 struct free_area
*area
;
1724 /* Find a page of the appropriate size in the preferred list */
1725 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1726 area
= &(zone
->free_area
[current_order
]);
1727 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1731 list_del(&page
->lru
);
1732 rmv_page_order(page
);
1734 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1735 set_pcppage_migratetype(page
, migratetype
);
1744 * This array describes the order lists are fallen back to when
1745 * the free lists for the desirable migrate type are depleted
1747 static int fallbacks
[MIGRATE_TYPES
][4] = {
1748 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1749 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1750 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1752 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1754 #ifdef CONFIG_MEMORY_ISOLATION
1755 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1760 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1763 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1766 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1767 unsigned int order
) { return NULL
; }
1771 * Move the free pages in a range to the free lists of the requested type.
1772 * Note that start_page and end_pages are not aligned on a pageblock
1773 * boundary. If alignment is required, use move_freepages_block()
1775 int move_freepages(struct zone
*zone
,
1776 struct page
*start_page
, struct page
*end_page
,
1781 int pages_moved
= 0;
1783 #ifndef CONFIG_HOLES_IN_ZONE
1785 * page_zone is not safe to call in this context when
1786 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1787 * anyway as we check zone boundaries in move_freepages_block().
1788 * Remove at a later date when no bug reports exist related to
1789 * grouping pages by mobility
1791 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1794 for (page
= start_page
; page
<= end_page
;) {
1795 /* Make sure we are not inadvertently changing nodes */
1796 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1798 if (!pfn_valid_within(page_to_pfn(page
))) {
1803 if (!PageBuddy(page
)) {
1808 order
= page_order(page
);
1809 list_move(&page
->lru
,
1810 &zone
->free_area
[order
].free_list
[migratetype
]);
1812 pages_moved
+= 1 << order
;
1818 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1821 unsigned long start_pfn
, end_pfn
;
1822 struct page
*start_page
, *end_page
;
1824 start_pfn
= page_to_pfn(page
);
1825 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1826 start_page
= pfn_to_page(start_pfn
);
1827 end_page
= start_page
+ pageblock_nr_pages
- 1;
1828 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1830 /* Do not cross zone boundaries */
1831 if (!zone_spans_pfn(zone
, start_pfn
))
1833 if (!zone_spans_pfn(zone
, end_pfn
))
1836 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1839 static void change_pageblock_range(struct page
*pageblock_page
,
1840 int start_order
, int migratetype
)
1842 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1844 while (nr_pageblocks
--) {
1845 set_pageblock_migratetype(pageblock_page
, migratetype
);
1846 pageblock_page
+= pageblock_nr_pages
;
1851 * When we are falling back to another migratetype during allocation, try to
1852 * steal extra free pages from the same pageblocks to satisfy further
1853 * allocations, instead of polluting multiple pageblocks.
1855 * If we are stealing a relatively large buddy page, it is likely there will
1856 * be more free pages in the pageblock, so try to steal them all. For
1857 * reclaimable and unmovable allocations, we steal regardless of page size,
1858 * as fragmentation caused by those allocations polluting movable pageblocks
1859 * is worse than movable allocations stealing from unmovable and reclaimable
1862 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1865 * Leaving this order check is intended, although there is
1866 * relaxed order check in next check. The reason is that
1867 * we can actually steal whole pageblock if this condition met,
1868 * but, below check doesn't guarantee it and that is just heuristic
1869 * so could be changed anytime.
1871 if (order
>= pageblock_order
)
1874 if (order
>= pageblock_order
/ 2 ||
1875 start_mt
== MIGRATE_RECLAIMABLE
||
1876 start_mt
== MIGRATE_UNMOVABLE
||
1877 page_group_by_mobility_disabled
)
1884 * This function implements actual steal behaviour. If order is large enough,
1885 * we can steal whole pageblock. If not, we first move freepages in this
1886 * pageblock and check whether half of pages are moved or not. If half of
1887 * pages are moved, we can change migratetype of pageblock and permanently
1888 * use it's pages as requested migratetype in the future.
1890 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1893 unsigned int current_order
= page_order(page
);
1896 /* Take ownership for orders >= pageblock_order */
1897 if (current_order
>= pageblock_order
) {
1898 change_pageblock_range(page
, current_order
, start_type
);
1902 pages
= move_freepages_block(zone
, page
, start_type
);
1904 /* Claim the whole block if over half of it is free */
1905 if (pages
>= (1 << (pageblock_order
-1)) ||
1906 page_group_by_mobility_disabled
)
1907 set_pageblock_migratetype(page
, start_type
);
1911 * Check whether there is a suitable fallback freepage with requested order.
1912 * If only_stealable is true, this function returns fallback_mt only if
1913 * we can steal other freepages all together. This would help to reduce
1914 * fragmentation due to mixed migratetype pages in one pageblock.
1916 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1917 int migratetype
, bool only_stealable
, bool *can_steal
)
1922 if (area
->nr_free
== 0)
1927 fallback_mt
= fallbacks
[migratetype
][i
];
1928 if (fallback_mt
== MIGRATE_TYPES
)
1931 if (list_empty(&area
->free_list
[fallback_mt
]))
1934 if (can_steal_fallback(order
, migratetype
))
1937 if (!only_stealable
)
1948 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1949 * there are no empty page blocks that contain a page with a suitable order
1951 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
1952 unsigned int alloc_order
)
1955 unsigned long max_managed
, flags
;
1958 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1959 * Check is race-prone but harmless.
1961 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
1962 if (zone
->nr_reserved_highatomic
>= max_managed
)
1965 spin_lock_irqsave(&zone
->lock
, flags
);
1967 /* Recheck the nr_reserved_highatomic limit under the lock */
1968 if (zone
->nr_reserved_highatomic
>= max_managed
)
1972 mt
= get_pageblock_migratetype(page
);
1973 if (mt
!= MIGRATE_HIGHATOMIC
&&
1974 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
1975 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
1976 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
1977 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
1981 spin_unlock_irqrestore(&zone
->lock
, flags
);
1985 * Used when an allocation is about to fail under memory pressure. This
1986 * potentially hurts the reliability of high-order allocations when under
1987 * intense memory pressure but failed atomic allocations should be easier
1988 * to recover from than an OOM.
1990 static void unreserve_highatomic_pageblock(const struct alloc_context
*ac
)
1992 struct zonelist
*zonelist
= ac
->zonelist
;
1993 unsigned long flags
;
1999 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2001 /* Preserve at least one pageblock */
2002 if (zone
->nr_reserved_highatomic
<= pageblock_nr_pages
)
2005 spin_lock_irqsave(&zone
->lock
, flags
);
2006 for (order
= 0; order
< MAX_ORDER
; order
++) {
2007 struct free_area
*area
= &(zone
->free_area
[order
]);
2009 page
= list_first_entry_or_null(
2010 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2016 * It should never happen but changes to locking could
2017 * inadvertently allow a per-cpu drain to add pages
2018 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2019 * and watch for underflows.
2021 zone
->nr_reserved_highatomic
-= min(pageblock_nr_pages
,
2022 zone
->nr_reserved_highatomic
);
2025 * Convert to ac->migratetype and avoid the normal
2026 * pageblock stealing heuristics. Minimally, the caller
2027 * is doing the work and needs the pages. More
2028 * importantly, if the block was always converted to
2029 * MIGRATE_UNMOVABLE or another type then the number
2030 * of pageblocks that cannot be completely freed
2033 set_pageblock_migratetype(page
, ac
->migratetype
);
2034 move_freepages_block(zone
, page
, ac
->migratetype
);
2035 spin_unlock_irqrestore(&zone
->lock
, flags
);
2038 spin_unlock_irqrestore(&zone
->lock
, flags
);
2042 /* Remove an element from the buddy allocator from the fallback list */
2043 static inline struct page
*
2044 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2046 struct free_area
*area
;
2047 unsigned int current_order
;
2052 /* Find the largest possible block of pages in the other list */
2053 for (current_order
= MAX_ORDER
-1;
2054 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2056 area
= &(zone
->free_area
[current_order
]);
2057 fallback_mt
= find_suitable_fallback(area
, current_order
,
2058 start_migratetype
, false, &can_steal
);
2059 if (fallback_mt
== -1)
2062 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2065 steal_suitable_fallback(zone
, page
, start_migratetype
);
2067 /* Remove the page from the freelists */
2069 list_del(&page
->lru
);
2070 rmv_page_order(page
);
2072 expand(zone
, page
, order
, current_order
, area
,
2075 * The pcppage_migratetype may differ from pageblock's
2076 * migratetype depending on the decisions in
2077 * find_suitable_fallback(). This is OK as long as it does not
2078 * differ for MIGRATE_CMA pageblocks. Those can be used as
2079 * fallback only via special __rmqueue_cma_fallback() function
2081 set_pcppage_migratetype(page
, start_migratetype
);
2083 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2084 start_migratetype
, fallback_mt
);
2093 * Do the hard work of removing an element from the buddy allocator.
2094 * Call me with the zone->lock already held.
2096 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2101 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2102 if (unlikely(!page
)) {
2103 if (migratetype
== MIGRATE_MOVABLE
)
2104 page
= __rmqueue_cma_fallback(zone
, order
);
2107 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2110 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2115 * Obtain a specified number of elements from the buddy allocator, all under
2116 * a single hold of the lock, for efficiency. Add them to the supplied list.
2117 * Returns the number of new pages which were placed at *list.
2119 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2120 unsigned long count
, struct list_head
*list
,
2121 int migratetype
, bool cold
)
2125 spin_lock(&zone
->lock
);
2126 for (i
= 0; i
< count
; ++i
) {
2127 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2128 if (unlikely(page
== NULL
))
2132 * Split buddy pages returned by expand() are received here
2133 * in physical page order. The page is added to the callers and
2134 * list and the list head then moves forward. From the callers
2135 * perspective, the linked list is ordered by page number in
2136 * some conditions. This is useful for IO devices that can
2137 * merge IO requests if the physical pages are ordered
2141 list_add(&page
->lru
, list
);
2143 list_add_tail(&page
->lru
, list
);
2145 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2146 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2149 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2150 spin_unlock(&zone
->lock
);
2156 * Called from the vmstat counter updater to drain pagesets of this
2157 * currently executing processor on remote nodes after they have
2160 * Note that this function must be called with the thread pinned to
2161 * a single processor.
2163 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2165 unsigned long flags
;
2166 int to_drain
, batch
;
2168 local_irq_save(flags
);
2169 batch
= READ_ONCE(pcp
->batch
);
2170 to_drain
= min(pcp
->count
, batch
);
2172 free_pcppages_bulk(zone
, to_drain
, pcp
);
2173 pcp
->count
-= to_drain
;
2175 local_irq_restore(flags
);
2180 * Drain pcplists of the indicated processor and zone.
2182 * The processor must either be the current processor and the
2183 * thread pinned to the current processor or a processor that
2186 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2188 unsigned long flags
;
2189 struct per_cpu_pageset
*pset
;
2190 struct per_cpu_pages
*pcp
;
2192 local_irq_save(flags
);
2193 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2197 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2200 local_irq_restore(flags
);
2204 * Drain pcplists of all zones on the indicated processor.
2206 * The processor must either be the current processor and the
2207 * thread pinned to the current processor or a processor that
2210 static void drain_pages(unsigned int cpu
)
2214 for_each_populated_zone(zone
) {
2215 drain_pages_zone(cpu
, zone
);
2220 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2222 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2223 * the single zone's pages.
2225 void drain_local_pages(struct zone
*zone
)
2227 int cpu
= smp_processor_id();
2230 drain_pages_zone(cpu
, zone
);
2236 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2238 * When zone parameter is non-NULL, spill just the single zone's pages.
2240 * Note that this code is protected against sending an IPI to an offline
2241 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2242 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2243 * nothing keeps CPUs from showing up after we populated the cpumask and
2244 * before the call to on_each_cpu_mask().
2246 void drain_all_pages(struct zone
*zone
)
2251 * Allocate in the BSS so we wont require allocation in
2252 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2254 static cpumask_t cpus_with_pcps
;
2257 * We don't care about racing with CPU hotplug event
2258 * as offline notification will cause the notified
2259 * cpu to drain that CPU pcps and on_each_cpu_mask
2260 * disables preemption as part of its processing
2262 for_each_online_cpu(cpu
) {
2263 struct per_cpu_pageset
*pcp
;
2265 bool has_pcps
= false;
2268 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2272 for_each_populated_zone(z
) {
2273 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2274 if (pcp
->pcp
.count
) {
2282 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2284 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2286 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
2290 #ifdef CONFIG_HIBERNATION
2292 void mark_free_pages(struct zone
*zone
)
2294 unsigned long pfn
, max_zone_pfn
;
2295 unsigned long flags
;
2296 unsigned int order
, t
;
2299 if (zone_is_empty(zone
))
2302 spin_lock_irqsave(&zone
->lock
, flags
);
2304 max_zone_pfn
= zone_end_pfn(zone
);
2305 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2306 if (pfn_valid(pfn
)) {
2307 page
= pfn_to_page(pfn
);
2309 if (page_zone(page
) != zone
)
2312 if (!swsusp_page_is_forbidden(page
))
2313 swsusp_unset_page_free(page
);
2316 for_each_migratetype_order(order
, t
) {
2317 list_for_each_entry(page
,
2318 &zone
->free_area
[order
].free_list
[t
], lru
) {
2321 pfn
= page_to_pfn(page
);
2322 for (i
= 0; i
< (1UL << order
); i
++)
2323 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2326 spin_unlock_irqrestore(&zone
->lock
, flags
);
2328 #endif /* CONFIG_PM */
2331 * Free a 0-order page
2332 * cold == true ? free a cold page : free a hot page
2334 void free_hot_cold_page(struct page
*page
, bool cold
)
2336 struct zone
*zone
= page_zone(page
);
2337 struct per_cpu_pages
*pcp
;
2338 unsigned long flags
;
2339 unsigned long pfn
= page_to_pfn(page
);
2342 if (!free_pages_prepare(page
, 0))
2345 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2346 set_pcppage_migratetype(page
, migratetype
);
2347 local_irq_save(flags
);
2348 __count_vm_event(PGFREE
);
2351 * We only track unmovable, reclaimable and movable on pcp lists.
2352 * Free ISOLATE pages back to the allocator because they are being
2353 * offlined but treat RESERVE as movable pages so we can get those
2354 * areas back if necessary. Otherwise, we may have to free
2355 * excessively into the page allocator
2357 if (migratetype
>= MIGRATE_PCPTYPES
) {
2358 if (unlikely(is_migrate_isolate(migratetype
))) {
2359 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2362 migratetype
= MIGRATE_MOVABLE
;
2365 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2367 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2369 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2371 if (pcp
->count
>= pcp
->high
) {
2372 unsigned long batch
= READ_ONCE(pcp
->batch
);
2373 free_pcppages_bulk(zone
, batch
, pcp
);
2374 pcp
->count
-= batch
;
2378 local_irq_restore(flags
);
2382 * Free a list of 0-order pages
2384 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2386 struct page
*page
, *next
;
2388 list_for_each_entry_safe(page
, next
, list
, lru
) {
2389 trace_mm_page_free_batched(page
, cold
);
2390 free_hot_cold_page(page
, cold
);
2395 * split_page takes a non-compound higher-order page, and splits it into
2396 * n (1<<order) sub-pages: page[0..n]
2397 * Each sub-page must be freed individually.
2399 * Note: this is probably too low level an operation for use in drivers.
2400 * Please consult with lkml before using this in your driver.
2402 void split_page(struct page
*page
, unsigned int order
)
2407 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2408 VM_BUG_ON_PAGE(!page_count(page
), page
);
2410 #ifdef CONFIG_KMEMCHECK
2412 * Split shadow pages too, because free(page[0]) would
2413 * otherwise free the whole shadow.
2415 if (kmemcheck_page_is_tracked(page
))
2416 split_page(virt_to_page(page
[0].shadow
), order
);
2419 gfp_mask
= get_page_owner_gfp(page
);
2420 set_page_owner(page
, 0, gfp_mask
);
2421 for (i
= 1; i
< (1 << order
); i
++) {
2422 set_page_refcounted(page
+ i
);
2423 set_page_owner(page
+ i
, 0, gfp_mask
);
2426 EXPORT_SYMBOL_GPL(split_page
);
2428 int __isolate_free_page(struct page
*page
, unsigned int order
)
2430 unsigned long watermark
;
2434 BUG_ON(!PageBuddy(page
));
2436 zone
= page_zone(page
);
2437 mt
= get_pageblock_migratetype(page
);
2439 if (!is_migrate_isolate(mt
)) {
2440 /* Obey watermarks as if the page was being allocated */
2441 watermark
= low_wmark_pages(zone
) + (1 << order
);
2442 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
2445 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2448 /* Remove page from free list */
2449 list_del(&page
->lru
);
2450 zone
->free_area
[order
].nr_free
--;
2451 rmv_page_order(page
);
2453 set_page_owner(page
, order
, __GFP_MOVABLE
);
2455 /* Set the pageblock if the isolated page is at least a pageblock */
2456 if (order
>= pageblock_order
- 1) {
2457 struct page
*endpage
= page
+ (1 << order
) - 1;
2458 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2459 int mt
= get_pageblock_migratetype(page
);
2460 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2461 set_pageblock_migratetype(page
,
2467 return 1UL << order
;
2471 * Similar to split_page except the page is already free. As this is only
2472 * being used for migration, the migratetype of the block also changes.
2473 * As this is called with interrupts disabled, the caller is responsible
2474 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2477 * Note: this is probably too low level an operation for use in drivers.
2478 * Please consult with lkml before using this in your driver.
2480 int split_free_page(struct page
*page
)
2485 order
= page_order(page
);
2487 nr_pages
= __isolate_free_page(page
, order
);
2491 /* Split into individual pages */
2492 set_page_refcounted(page
);
2493 split_page(page
, order
);
2498 * Update NUMA hit/miss statistics
2500 * Must be called with interrupts disabled.
2502 * When __GFP_OTHER_NODE is set assume the node of the preferred
2503 * zone is the local node. This is useful for daemons who allocate
2504 * memory on behalf of other processes.
2506 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2510 int local_nid
= numa_node_id();
2511 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2513 if (unlikely(flags
& __GFP_OTHER_NODE
)) {
2514 local_stat
= NUMA_OTHER
;
2515 local_nid
= preferred_zone
->node
;
2518 if (z
->node
== local_nid
) {
2519 __inc_zone_state(z
, NUMA_HIT
);
2520 __inc_zone_state(z
, local_stat
);
2522 __inc_zone_state(z
, NUMA_MISS
);
2523 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2529 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2532 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2533 struct zone
*zone
, unsigned int order
,
2534 gfp_t gfp_flags
, unsigned int alloc_flags
,
2537 unsigned long flags
;
2539 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2541 if (likely(order
== 0)) {
2542 struct per_cpu_pages
*pcp
;
2543 struct list_head
*list
;
2545 local_irq_save(flags
);
2546 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2547 list
= &pcp
->lists
[migratetype
];
2548 if (list_empty(list
)) {
2549 pcp
->count
+= rmqueue_bulk(zone
, 0,
2552 if (unlikely(list_empty(list
)))
2557 page
= list_last_entry(list
, struct page
, lru
);
2559 page
= list_first_entry(list
, struct page
, lru
);
2561 __dec_zone_state(zone
, NR_ALLOC_BATCH
);
2562 list_del(&page
->lru
);
2566 * We most definitely don't want callers attempting to
2567 * allocate greater than order-1 page units with __GFP_NOFAIL.
2569 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2570 spin_lock_irqsave(&zone
->lock
, flags
);
2573 if (alloc_flags
& ALLOC_HARDER
) {
2574 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2576 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2579 page
= __rmqueue(zone
, order
, migratetype
);
2580 spin_unlock(&zone
->lock
);
2583 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
2584 __mod_zone_freepage_state(zone
, -(1 << order
),
2585 get_pcppage_migratetype(page
));
2588 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
2589 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
2590 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2592 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
2593 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2594 local_irq_restore(flags
);
2596 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2600 local_irq_restore(flags
);
2604 #ifdef CONFIG_FAIL_PAGE_ALLOC
2607 struct fault_attr attr
;
2609 bool ignore_gfp_highmem
;
2610 bool ignore_gfp_reclaim
;
2612 } fail_page_alloc
= {
2613 .attr
= FAULT_ATTR_INITIALIZER
,
2614 .ignore_gfp_reclaim
= true,
2615 .ignore_gfp_highmem
= true,
2619 static int __init
setup_fail_page_alloc(char *str
)
2621 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2623 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2625 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2627 if (order
< fail_page_alloc
.min_order
)
2629 if (gfp_mask
& __GFP_NOFAIL
)
2631 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2633 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2634 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2637 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2640 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2642 static int __init
fail_page_alloc_debugfs(void)
2644 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2647 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2648 &fail_page_alloc
.attr
);
2650 return PTR_ERR(dir
);
2652 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2653 &fail_page_alloc
.ignore_gfp_reclaim
))
2655 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2656 &fail_page_alloc
.ignore_gfp_highmem
))
2658 if (!debugfs_create_u32("min-order", mode
, dir
,
2659 &fail_page_alloc
.min_order
))
2664 debugfs_remove_recursive(dir
);
2669 late_initcall(fail_page_alloc_debugfs
);
2671 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2673 #else /* CONFIG_FAIL_PAGE_ALLOC */
2675 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2680 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2683 * Return true if free base pages are above 'mark'. For high-order checks it
2684 * will return true of the order-0 watermark is reached and there is at least
2685 * one free page of a suitable size. Checking now avoids taking the zone lock
2686 * to check in the allocation paths if no pages are free.
2688 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
2689 unsigned long mark
, int classzone_idx
,
2690 unsigned int alloc_flags
,
2695 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2697 /* free_pages may go negative - that's OK */
2698 free_pages
-= (1 << order
) - 1;
2700 if (alloc_flags
& ALLOC_HIGH
)
2704 * If the caller does not have rights to ALLOC_HARDER then subtract
2705 * the high-atomic reserves. This will over-estimate the size of the
2706 * atomic reserve but it avoids a search.
2708 if (likely(!alloc_harder
))
2709 free_pages
-= z
->nr_reserved_highatomic
;
2714 /* If allocation can't use CMA areas don't use free CMA pages */
2715 if (!(alloc_flags
& ALLOC_CMA
))
2716 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2720 * Check watermarks for an order-0 allocation request. If these
2721 * are not met, then a high-order request also cannot go ahead
2722 * even if a suitable page happened to be free.
2724 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2727 /* If this is an order-0 request then the watermark is fine */
2731 /* For a high-order request, check at least one suitable page is free */
2732 for (o
= order
; o
< MAX_ORDER
; o
++) {
2733 struct free_area
*area
= &z
->free_area
[o
];
2742 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2743 if (!list_empty(&area
->free_list
[mt
]))
2748 if ((alloc_flags
& ALLOC_CMA
) &&
2749 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2757 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2758 int classzone_idx
, unsigned int alloc_flags
)
2760 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2761 zone_page_state(z
, NR_FREE_PAGES
));
2764 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2765 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2767 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2771 /* If allocation can't use CMA areas don't use free CMA pages */
2772 if (!(alloc_flags
& ALLOC_CMA
))
2773 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2777 * Fast check for order-0 only. If this fails then the reserves
2778 * need to be calculated. There is a corner case where the check
2779 * passes but only the high-order atomic reserve are free. If
2780 * the caller is !atomic then it'll uselessly search the free
2781 * list. That corner case is then slower but it is harmless.
2783 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2786 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2790 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2791 unsigned long mark
, int classzone_idx
)
2793 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2795 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2796 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2798 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2803 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2805 return local_zone
->node
== zone
->node
;
2808 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2810 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2813 #else /* CONFIG_NUMA */
2814 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2819 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2823 #endif /* CONFIG_NUMA */
2825 static void reset_alloc_batches(struct zone
*preferred_zone
)
2827 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2830 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2831 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2832 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2833 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2834 } while (zone
++ != preferred_zone
);
2838 * get_page_from_freelist goes through the zonelist trying to allocate
2841 static struct page
*
2842 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2843 const struct alloc_context
*ac
)
2845 struct zoneref
*z
= ac
->preferred_zoneref
;
2847 bool fair_skipped
= false;
2848 bool apply_fair
= (alloc_flags
& ALLOC_FAIR
);
2852 * Scan zonelist, looking for a zone with enough free.
2853 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2855 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2860 if (cpusets_enabled() &&
2861 (alloc_flags
& ALLOC_CPUSET
) &&
2862 !__cpuset_zone_allowed(zone
, gfp_mask
))
2865 * Distribute pages in proportion to the individual
2866 * zone size to ensure fair page aging. The zone a
2867 * page was allocated in should have no effect on the
2868 * time the page has in memory before being reclaimed.
2871 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2872 fair_skipped
= true;
2875 if (!zone_local(ac
->preferred_zoneref
->zone
, zone
)) {
2882 * When allocating a page cache page for writing, we
2883 * want to get it from a zone that is within its dirty
2884 * limit, such that no single zone holds more than its
2885 * proportional share of globally allowed dirty pages.
2886 * The dirty limits take into account the zone's
2887 * lowmem reserves and high watermark so that kswapd
2888 * should be able to balance it without having to
2889 * write pages from its LRU list.
2891 * This may look like it could increase pressure on
2892 * lower zones by failing allocations in higher zones
2893 * before they are full. But the pages that do spill
2894 * over are limited as the lower zones are protected
2895 * by this very same mechanism. It should not become
2896 * a practical burden to them.
2898 * XXX: For now, allow allocations to potentially
2899 * exceed the per-zone dirty limit in the slowpath
2900 * (spread_dirty_pages unset) before going into reclaim,
2901 * which is important when on a NUMA setup the allowed
2902 * zones are together not big enough to reach the
2903 * global limit. The proper fix for these situations
2904 * will require awareness of zones in the
2905 * dirty-throttling and the flusher threads.
2907 if (ac
->spread_dirty_pages
&& !zone_dirty_ok(zone
))
2910 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2911 if (!zone_watermark_fast(zone
, order
, mark
,
2912 ac_classzone_idx(ac
), alloc_flags
)) {
2915 /* Checked here to keep the fast path fast */
2916 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2917 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2920 if (zone_reclaim_mode
== 0 ||
2921 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
2924 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2926 case ZONE_RECLAIM_NOSCAN
:
2929 case ZONE_RECLAIM_FULL
:
2930 /* scanned but unreclaimable */
2933 /* did we reclaim enough */
2934 if (zone_watermark_ok(zone
, order
, mark
,
2935 ac_classzone_idx(ac
), alloc_flags
))
2943 page
= buffered_rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
2944 gfp_mask
, alloc_flags
, ac
->migratetype
);
2946 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2950 * If this is a high-order atomic allocation then check
2951 * if the pageblock should be reserved for the future
2953 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
2954 reserve_highatomic_pageblock(page
, zone
, order
);
2961 * The first pass makes sure allocations are spread fairly within the
2962 * local node. However, the local node might have free pages left
2963 * after the fairness batches are exhausted, and remote zones haven't
2964 * even been considered yet. Try once more without fairness, and
2965 * include remote zones now, before entering the slowpath and waking
2966 * kswapd: prefer spilling to a remote zone over swapping locally.
2971 fair_skipped
= false;
2972 reset_alloc_batches(ac
->preferred_zoneref
->zone
);
2980 * Large machines with many possible nodes should not always dump per-node
2981 * meminfo in irq context.
2983 static inline bool should_suppress_show_mem(void)
2988 ret
= in_interrupt();
2993 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2994 DEFAULT_RATELIMIT_INTERVAL
,
2995 DEFAULT_RATELIMIT_BURST
);
2997 void warn_alloc_failed(gfp_t gfp_mask
, unsigned int order
, const char *fmt
, ...)
2999 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3001 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
3002 debug_guardpage_minorder() > 0)
3006 * This documents exceptions given to allocations in certain
3007 * contexts that are allowed to allocate outside current's set
3010 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3011 if (test_thread_flag(TIF_MEMDIE
) ||
3012 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3013 filter
&= ~SHOW_MEM_FILTER_NODES
;
3014 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3015 filter
&= ~SHOW_MEM_FILTER_NODES
;
3018 struct va_format vaf
;
3021 va_start(args
, fmt
);
3026 pr_warn("%pV", &vaf
);
3031 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3032 current
->comm
, order
, gfp_mask
, &gfp_mask
);
3034 if (!should_suppress_show_mem())
3038 static inline struct page
*
3039 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3040 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3042 struct oom_control oc
= {
3043 .zonelist
= ac
->zonelist
,
3044 .nodemask
= ac
->nodemask
,
3045 .gfp_mask
= gfp_mask
,
3050 *did_some_progress
= 0;
3053 * Acquire the oom lock. If that fails, somebody else is
3054 * making progress for us.
3056 if (!mutex_trylock(&oom_lock
)) {
3057 *did_some_progress
= 1;
3058 schedule_timeout_uninterruptible(1);
3063 * Go through the zonelist yet one more time, keep very high watermark
3064 * here, this is only to catch a parallel oom killing, we must fail if
3065 * we're still under heavy pressure.
3067 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3068 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3072 if (!(gfp_mask
& __GFP_NOFAIL
)) {
3073 /* Coredumps can quickly deplete all memory reserves */
3074 if (current
->flags
& PF_DUMPCORE
)
3076 /* The OOM killer will not help higher order allocs */
3077 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3079 /* The OOM killer does not needlessly kill tasks for lowmem */
3080 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3082 if (pm_suspended_storage())
3085 * XXX: GFP_NOFS allocations should rather fail than rely on
3086 * other request to make a forward progress.
3087 * We are in an unfortunate situation where out_of_memory cannot
3088 * do much for this context but let's try it to at least get
3089 * access to memory reserved if the current task is killed (see
3090 * out_of_memory). Once filesystems are ready to handle allocation
3091 * failures more gracefully we should just bail out here.
3094 /* The OOM killer may not free memory on a specific node */
3095 if (gfp_mask
& __GFP_THISNODE
)
3098 /* Exhausted what can be done so it's blamo time */
3099 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3100 *did_some_progress
= 1;
3102 if (gfp_mask
& __GFP_NOFAIL
) {
3103 page
= get_page_from_freelist(gfp_mask
, order
,
3104 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
3106 * fallback to ignore cpuset restriction if our nodes
3110 page
= get_page_from_freelist(gfp_mask
, order
,
3111 ALLOC_NO_WATERMARKS
, ac
);
3115 mutex_unlock(&oom_lock
);
3119 #ifdef CONFIG_COMPACTION
3120 /* Try memory compaction for high-order allocations before reclaim */
3121 static struct page
*
3122 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3123 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3124 enum migrate_mode mode
, int *contended_compaction
,
3125 bool *deferred_compaction
)
3127 unsigned long compact_result
;
3133 current
->flags
|= PF_MEMALLOC
;
3134 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3135 mode
, contended_compaction
);
3136 current
->flags
&= ~PF_MEMALLOC
;
3138 switch (compact_result
) {
3139 case COMPACT_DEFERRED
:
3140 *deferred_compaction
= true;
3142 case COMPACT_SKIPPED
:
3149 * At least in one zone compaction wasn't deferred or skipped, so let's
3150 * count a compaction stall
3152 count_vm_event(COMPACTSTALL
);
3154 page
= get_page_from_freelist(gfp_mask
, order
,
3155 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3158 struct zone
*zone
= page_zone(page
);
3160 zone
->compact_blockskip_flush
= false;
3161 compaction_defer_reset(zone
, order
, true);
3162 count_vm_event(COMPACTSUCCESS
);
3167 * It's bad if compaction run occurs and fails. The most likely reason
3168 * is that pages exist, but not enough to satisfy watermarks.
3170 count_vm_event(COMPACTFAIL
);
3177 static inline struct page
*
3178 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3179 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3180 enum migrate_mode mode
, int *contended_compaction
,
3181 bool *deferred_compaction
)
3185 #endif /* CONFIG_COMPACTION */
3187 /* Perform direct synchronous page reclaim */
3189 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3190 const struct alloc_context
*ac
)
3192 struct reclaim_state reclaim_state
;
3197 /* We now go into synchronous reclaim */
3198 cpuset_memory_pressure_bump();
3199 current
->flags
|= PF_MEMALLOC
;
3200 lockdep_set_current_reclaim_state(gfp_mask
);
3201 reclaim_state
.reclaimed_slab
= 0;
3202 current
->reclaim_state
= &reclaim_state
;
3204 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3207 current
->reclaim_state
= NULL
;
3208 lockdep_clear_current_reclaim_state();
3209 current
->flags
&= ~PF_MEMALLOC
;
3216 /* The really slow allocator path where we enter direct reclaim */
3217 static inline struct page
*
3218 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3219 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3220 unsigned long *did_some_progress
)
3222 struct page
*page
= NULL
;
3223 bool drained
= false;
3225 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3226 if (unlikely(!(*did_some_progress
)))
3230 page
= get_page_from_freelist(gfp_mask
, order
,
3231 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3234 * If an allocation failed after direct reclaim, it could be because
3235 * pages are pinned on the per-cpu lists or in high alloc reserves.
3236 * Shrink them them and try again
3238 if (!page
&& !drained
) {
3239 unreserve_highatomic_pageblock(ac
);
3240 drain_all_pages(NULL
);
3248 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3253 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3254 ac
->high_zoneidx
, ac
->nodemask
)
3255 wakeup_kswapd(zone
, order
, ac_classzone_idx(ac
));
3258 static inline unsigned int
3259 gfp_to_alloc_flags(gfp_t gfp_mask
)
3261 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3263 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3264 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3267 * The caller may dip into page reserves a bit more if the caller
3268 * cannot run direct reclaim, or if the caller has realtime scheduling
3269 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3270 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3272 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3274 if (gfp_mask
& __GFP_ATOMIC
) {
3276 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3277 * if it can't schedule.
3279 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3280 alloc_flags
|= ALLOC_HARDER
;
3282 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3283 * comment for __cpuset_node_allowed().
3285 alloc_flags
&= ~ALLOC_CPUSET
;
3286 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3287 alloc_flags
|= ALLOC_HARDER
;
3289 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
3290 if (gfp_mask
& __GFP_MEMALLOC
)
3291 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3292 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3293 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3294 else if (!in_interrupt() &&
3295 ((current
->flags
& PF_MEMALLOC
) ||
3296 unlikely(test_thread_flag(TIF_MEMDIE
))))
3297 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3300 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3301 alloc_flags
|= ALLOC_CMA
;
3306 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3308 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
3311 static inline bool is_thp_gfp_mask(gfp_t gfp_mask
)
3313 return (gfp_mask
& (GFP_TRANSHUGE
| __GFP_KSWAPD_RECLAIM
)) == GFP_TRANSHUGE
;
3316 static inline struct page
*
3317 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3318 struct alloc_context
*ac
)
3320 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3321 struct page
*page
= NULL
;
3322 unsigned int alloc_flags
;
3323 unsigned long pages_reclaimed
= 0;
3324 unsigned long did_some_progress
;
3325 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
3326 bool deferred_compaction
= false;
3327 int contended_compaction
= COMPACT_CONTENDED_NONE
;
3330 * In the slowpath, we sanity check order to avoid ever trying to
3331 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3332 * be using allocators in order of preference for an area that is
3335 if (order
>= MAX_ORDER
) {
3336 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3341 * We also sanity check to catch abuse of atomic reserves being used by
3342 * callers that are not in atomic context.
3344 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3345 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3346 gfp_mask
&= ~__GFP_ATOMIC
;
3349 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3350 wake_all_kswapds(order
, ac
);
3353 * OK, we're below the kswapd watermark and have kicked background
3354 * reclaim. Now things get more complex, so set up alloc_flags according
3355 * to how we want to proceed.
3357 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3359 /* This is the last chance, in general, before the goto nopage. */
3360 page
= get_page_from_freelist(gfp_mask
, order
,
3361 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3365 /* Allocate without watermarks if the context allows */
3366 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
3368 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3369 * the allocation is high priority and these type of
3370 * allocations are system rather than user orientated
3372 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3373 page
= get_page_from_freelist(gfp_mask
, order
,
3374 ALLOC_NO_WATERMARKS
, ac
);
3379 /* Caller is not willing to reclaim, we can't balance anything */
3380 if (!can_direct_reclaim
) {
3382 * All existing users of the __GFP_NOFAIL are blockable, so warn
3383 * of any new users that actually allow this type of allocation
3386 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3390 /* Avoid recursion of direct reclaim */
3391 if (current
->flags
& PF_MEMALLOC
) {
3393 * __GFP_NOFAIL request from this context is rather bizarre
3394 * because we cannot reclaim anything and only can loop waiting
3395 * for somebody to do a work for us.
3397 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3404 /* Avoid allocations with no watermarks from looping endlessly */
3405 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3409 * Try direct compaction. The first pass is asynchronous. Subsequent
3410 * attempts after direct reclaim are synchronous
3412 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3414 &contended_compaction
,
3415 &deferred_compaction
);
3419 /* Checks for THP-specific high-order allocations */
3420 if (is_thp_gfp_mask(gfp_mask
)) {
3422 * If compaction is deferred for high-order allocations, it is
3423 * because sync compaction recently failed. If this is the case
3424 * and the caller requested a THP allocation, we do not want
3425 * to heavily disrupt the system, so we fail the allocation
3426 * instead of entering direct reclaim.
3428 if (deferred_compaction
)
3432 * In all zones where compaction was attempted (and not
3433 * deferred or skipped), lock contention has been detected.
3434 * For THP allocation we do not want to disrupt the others
3435 * so we fallback to base pages instead.
3437 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
3441 * If compaction was aborted due to need_resched(), we do not
3442 * want to further increase allocation latency, unless it is
3443 * khugepaged trying to collapse.
3445 if (contended_compaction
== COMPACT_CONTENDED_SCHED
3446 && !(current
->flags
& PF_KTHREAD
))
3451 * It can become very expensive to allocate transparent hugepages at
3452 * fault, so use asynchronous memory compaction for THP unless it is
3453 * khugepaged trying to collapse.
3455 if (!is_thp_gfp_mask(gfp_mask
) || (current
->flags
& PF_KTHREAD
))
3456 migration_mode
= MIGRATE_SYNC_LIGHT
;
3458 /* Try direct reclaim and then allocating */
3459 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3460 &did_some_progress
);
3464 /* Do not loop if specifically requested */
3465 if (gfp_mask
& __GFP_NORETRY
)
3468 /* Keep reclaiming pages as long as there is reasonable progress */
3469 pages_reclaimed
+= did_some_progress
;
3470 if ((did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
) ||
3471 ((gfp_mask
& __GFP_REPEAT
) && pages_reclaimed
< (1 << order
))) {
3472 /* Wait for some write requests to complete then retry */
3473 wait_iff_congested(ac
->preferred_zoneref
->zone
, BLK_RW_ASYNC
, HZ
/50);
3477 /* Reclaim has failed us, start killing things */
3478 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3482 /* Retry as long as the OOM killer is making progress */
3483 if (did_some_progress
)
3488 * High-order allocations do not necessarily loop after
3489 * direct reclaim and reclaim/compaction depends on compaction
3490 * being called after reclaim so call directly if necessary
3492 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
,
3494 &contended_compaction
,
3495 &deferred_compaction
);
3499 warn_alloc_failed(gfp_mask
, order
, NULL
);
3505 * This is the 'heart' of the zoned buddy allocator.
3508 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3509 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3512 unsigned int cpuset_mems_cookie
;
3513 unsigned int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_FAIR
;
3514 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3515 struct alloc_context ac
= {
3516 .high_zoneidx
= gfp_zone(gfp_mask
),
3517 .zonelist
= zonelist
,
3518 .nodemask
= nodemask
,
3519 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3522 if (cpusets_enabled()) {
3523 alloc_mask
|= __GFP_HARDWALL
;
3524 alloc_flags
|= ALLOC_CPUSET
;
3526 ac
.nodemask
= &cpuset_current_mems_allowed
;
3529 gfp_mask
&= gfp_allowed_mask
;
3531 lockdep_trace_alloc(gfp_mask
);
3533 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3535 if (should_fail_alloc_page(gfp_mask
, order
))
3539 * Check the zones suitable for the gfp_mask contain at least one
3540 * valid zone. It's possible to have an empty zonelist as a result
3541 * of __GFP_THISNODE and a memoryless node
3543 if (unlikely(!zonelist
->_zonerefs
->zone
))
3546 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3547 alloc_flags
|= ALLOC_CMA
;
3550 cpuset_mems_cookie
= read_mems_allowed_begin();
3552 /* Dirty zone balancing only done in the fast path */
3553 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3555 /* The preferred zone is used for statistics later */
3556 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3557 ac
.high_zoneidx
, ac
.nodemask
);
3558 if (!ac
.preferred_zoneref
) {
3563 /* First allocation attempt */
3564 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3569 * Runtime PM, block IO and its error handling path can deadlock
3570 * because I/O on the device might not complete.
3572 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3573 ac
.spread_dirty_pages
= false;
3575 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3579 * When updating a task's mems_allowed, it is possible to race with
3580 * parallel threads in such a way that an allocation can fail while
3581 * the mask is being updated. If a page allocation is about to fail,
3582 * check if the cpuset changed during allocation and if so, retry.
3584 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3585 alloc_mask
= gfp_mask
;
3590 if (kmemcheck_enabled
&& page
)
3591 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3593 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3597 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3600 * Common helper functions.
3602 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3607 * __get_free_pages() returns a 32-bit address, which cannot represent
3610 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3612 page
= alloc_pages(gfp_mask
, order
);
3615 return (unsigned long) page_address(page
);
3617 EXPORT_SYMBOL(__get_free_pages
);
3619 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3621 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3623 EXPORT_SYMBOL(get_zeroed_page
);
3625 void __free_pages(struct page
*page
, unsigned int order
)
3627 if (put_page_testzero(page
)) {
3629 free_hot_cold_page(page
, false);
3631 __free_pages_ok(page
, order
);
3635 EXPORT_SYMBOL(__free_pages
);
3637 void free_pages(unsigned long addr
, unsigned int order
)
3640 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3641 __free_pages(virt_to_page((void *)addr
), order
);
3645 EXPORT_SYMBOL(free_pages
);
3649 * An arbitrary-length arbitrary-offset area of memory which resides
3650 * within a 0 or higher order page. Multiple fragments within that page
3651 * are individually refcounted, in the page's reference counter.
3653 * The page_frag functions below provide a simple allocation framework for
3654 * page fragments. This is used by the network stack and network device
3655 * drivers to provide a backing region of memory for use as either an
3656 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3658 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3661 struct page
*page
= NULL
;
3662 gfp_t gfp
= gfp_mask
;
3664 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3665 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3667 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3668 PAGE_FRAG_CACHE_MAX_ORDER
);
3669 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3671 if (unlikely(!page
))
3672 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3674 nc
->va
= page
? page_address(page
) : NULL
;
3679 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3680 unsigned int fragsz
, gfp_t gfp_mask
)
3682 unsigned int size
= PAGE_SIZE
;
3686 if (unlikely(!nc
->va
)) {
3688 page
= __page_frag_refill(nc
, gfp_mask
);
3692 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3693 /* if size can vary use size else just use PAGE_SIZE */
3696 /* Even if we own the page, we do not use atomic_set().
3697 * This would break get_page_unless_zero() users.
3699 page_ref_add(page
, size
- 1);
3701 /* reset page count bias and offset to start of new frag */
3702 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3703 nc
->pagecnt_bias
= size
;
3707 offset
= nc
->offset
- fragsz
;
3708 if (unlikely(offset
< 0)) {
3709 page
= virt_to_page(nc
->va
);
3711 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3714 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3715 /* if size can vary use size else just use PAGE_SIZE */
3718 /* OK, page count is 0, we can safely set it */
3719 set_page_count(page
, size
);
3721 /* reset page count bias and offset to start of new frag */
3722 nc
->pagecnt_bias
= size
;
3723 offset
= size
- fragsz
;
3727 nc
->offset
= offset
;
3729 return nc
->va
+ offset
;
3731 EXPORT_SYMBOL(__alloc_page_frag
);
3734 * Frees a page fragment allocated out of either a compound or order 0 page.
3736 void __free_page_frag(void *addr
)
3738 struct page
*page
= virt_to_head_page(addr
);
3740 if (unlikely(put_page_testzero(page
)))
3741 __free_pages_ok(page
, compound_order(page
));
3743 EXPORT_SYMBOL(__free_page_frag
);
3746 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3747 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3748 * equivalent to alloc_pages.
3750 * It should be used when the caller would like to use kmalloc, but since the
3751 * allocation is large, it has to fall back to the page allocator.
3753 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
3757 page
= alloc_pages(gfp_mask
, order
);
3758 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3759 __free_pages(page
, order
);
3765 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
3769 page
= alloc_pages_node(nid
, gfp_mask
, order
);
3770 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3771 __free_pages(page
, order
);
3778 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3781 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3783 memcg_kmem_uncharge(page
, order
);
3784 __free_pages(page
, order
);
3787 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3790 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3791 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3795 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3799 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3800 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3802 split_page(virt_to_page((void *)addr
), order
);
3803 while (used
< alloc_end
) {
3808 return (void *)addr
;
3812 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3813 * @size: the number of bytes to allocate
3814 * @gfp_mask: GFP flags for the allocation
3816 * This function is similar to alloc_pages(), except that it allocates the
3817 * minimum number of pages to satisfy the request. alloc_pages() can only
3818 * allocate memory in power-of-two pages.
3820 * This function is also limited by MAX_ORDER.
3822 * Memory allocated by this function must be released by free_pages_exact().
3824 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3826 unsigned int order
= get_order(size
);
3829 addr
= __get_free_pages(gfp_mask
, order
);
3830 return make_alloc_exact(addr
, order
, size
);
3832 EXPORT_SYMBOL(alloc_pages_exact
);
3835 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3837 * @nid: the preferred node ID where memory should be allocated
3838 * @size: the number of bytes to allocate
3839 * @gfp_mask: GFP flags for the allocation
3841 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3844 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3846 unsigned int order
= get_order(size
);
3847 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3850 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3854 * free_pages_exact - release memory allocated via alloc_pages_exact()
3855 * @virt: the value returned by alloc_pages_exact.
3856 * @size: size of allocation, same value as passed to alloc_pages_exact().
3858 * Release the memory allocated by a previous call to alloc_pages_exact.
3860 void free_pages_exact(void *virt
, size_t size
)
3862 unsigned long addr
= (unsigned long)virt
;
3863 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3865 while (addr
< end
) {
3870 EXPORT_SYMBOL(free_pages_exact
);
3873 * nr_free_zone_pages - count number of pages beyond high watermark
3874 * @offset: The zone index of the highest zone
3876 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3877 * high watermark within all zones at or below a given zone index. For each
3878 * zone, the number of pages is calculated as:
3879 * managed_pages - high_pages
3881 static unsigned long nr_free_zone_pages(int offset
)
3886 /* Just pick one node, since fallback list is circular */
3887 unsigned long sum
= 0;
3889 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3891 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3892 unsigned long size
= zone
->managed_pages
;
3893 unsigned long high
= high_wmark_pages(zone
);
3902 * nr_free_buffer_pages - count number of pages beyond high watermark
3904 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3905 * watermark within ZONE_DMA and ZONE_NORMAL.
3907 unsigned long nr_free_buffer_pages(void)
3909 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3911 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3914 * nr_free_pagecache_pages - count number of pages beyond high watermark
3916 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3917 * high watermark within all zones.
3919 unsigned long nr_free_pagecache_pages(void)
3921 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3924 static inline void show_node(struct zone
*zone
)
3926 if (IS_ENABLED(CONFIG_NUMA
))
3927 printk("Node %d ", zone_to_nid(zone
));
3930 long si_mem_available(void)
3933 unsigned long pagecache
;
3934 unsigned long wmark_low
= 0;
3935 unsigned long pages
[NR_LRU_LISTS
];
3939 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
3940 pages
[lru
] = global_page_state(NR_LRU_BASE
+ lru
);
3943 wmark_low
+= zone
->watermark
[WMARK_LOW
];
3946 * Estimate the amount of memory available for userspace allocations,
3947 * without causing swapping.
3949 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
3952 * Not all the page cache can be freed, otherwise the system will
3953 * start swapping. Assume at least half of the page cache, or the
3954 * low watermark worth of cache, needs to stay.
3956 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
3957 pagecache
-= min(pagecache
/ 2, wmark_low
);
3958 available
+= pagecache
;
3961 * Part of the reclaimable slab consists of items that are in use,
3962 * and cannot be freed. Cap this estimate at the low watermark.
3964 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
3965 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
3971 EXPORT_SYMBOL_GPL(si_mem_available
);
3973 void si_meminfo(struct sysinfo
*val
)
3975 val
->totalram
= totalram_pages
;
3976 val
->sharedram
= global_page_state(NR_SHMEM
);
3977 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3978 val
->bufferram
= nr_blockdev_pages();
3979 val
->totalhigh
= totalhigh_pages
;
3980 val
->freehigh
= nr_free_highpages();
3981 val
->mem_unit
= PAGE_SIZE
;
3984 EXPORT_SYMBOL(si_meminfo
);
3987 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3989 int zone_type
; /* needs to be signed */
3990 unsigned long managed_pages
= 0;
3991 unsigned long managed_highpages
= 0;
3992 unsigned long free_highpages
= 0;
3993 pg_data_t
*pgdat
= NODE_DATA(nid
);
3995 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3996 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3997 val
->totalram
= managed_pages
;
3998 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3999 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
4000 #ifdef CONFIG_HIGHMEM
4001 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4002 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4004 if (is_highmem(zone
)) {
4005 managed_highpages
+= zone
->managed_pages
;
4006 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4009 val
->totalhigh
= managed_highpages
;
4010 val
->freehigh
= free_highpages
;
4012 val
->totalhigh
= managed_highpages
;
4013 val
->freehigh
= free_highpages
;
4015 val
->mem_unit
= PAGE_SIZE
;
4020 * Determine whether the node should be displayed or not, depending on whether
4021 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4023 bool skip_free_areas_node(unsigned int flags
, int nid
)
4026 unsigned int cpuset_mems_cookie
;
4028 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4032 cpuset_mems_cookie
= read_mems_allowed_begin();
4033 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
4034 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
4039 #define K(x) ((x) << (PAGE_SHIFT-10))
4041 static void show_migration_types(unsigned char type
)
4043 static const char types
[MIGRATE_TYPES
] = {
4044 [MIGRATE_UNMOVABLE
] = 'U',
4045 [MIGRATE_MOVABLE
] = 'M',
4046 [MIGRATE_RECLAIMABLE
] = 'E',
4047 [MIGRATE_HIGHATOMIC
] = 'H',
4049 [MIGRATE_CMA
] = 'C',
4051 #ifdef CONFIG_MEMORY_ISOLATION
4052 [MIGRATE_ISOLATE
] = 'I',
4055 char tmp
[MIGRATE_TYPES
+ 1];
4059 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4060 if (type
& (1 << i
))
4065 printk("(%s) ", tmp
);
4069 * Show free area list (used inside shift_scroll-lock stuff)
4070 * We also calculate the percentage fragmentation. We do this by counting the
4071 * memory on each free list with the exception of the first item on the list.
4074 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4077 void show_free_areas(unsigned int filter
)
4079 unsigned long free_pcp
= 0;
4083 for_each_populated_zone(zone
) {
4084 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4087 for_each_online_cpu(cpu
)
4088 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4091 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4092 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4093 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4094 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4095 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4096 " free:%lu free_pcp:%lu free_cma:%lu\n",
4097 global_page_state(NR_ACTIVE_ANON
),
4098 global_page_state(NR_INACTIVE_ANON
),
4099 global_page_state(NR_ISOLATED_ANON
),
4100 global_page_state(NR_ACTIVE_FILE
),
4101 global_page_state(NR_INACTIVE_FILE
),
4102 global_page_state(NR_ISOLATED_FILE
),
4103 global_page_state(NR_UNEVICTABLE
),
4104 global_page_state(NR_FILE_DIRTY
),
4105 global_page_state(NR_WRITEBACK
),
4106 global_page_state(NR_UNSTABLE_NFS
),
4107 global_page_state(NR_SLAB_RECLAIMABLE
),
4108 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4109 global_page_state(NR_FILE_MAPPED
),
4110 global_page_state(NR_SHMEM
),
4111 global_page_state(NR_PAGETABLE
),
4112 global_page_state(NR_BOUNCE
),
4113 global_page_state(NR_FREE_PAGES
),
4115 global_page_state(NR_FREE_CMA_PAGES
));
4117 for_each_populated_zone(zone
) {
4120 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4124 for_each_online_cpu(cpu
)
4125 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4133 " active_anon:%lukB"
4134 " inactive_anon:%lukB"
4135 " active_file:%lukB"
4136 " inactive_file:%lukB"
4137 " unevictable:%lukB"
4138 " isolated(anon):%lukB"
4139 " isolated(file):%lukB"
4147 " slab_reclaimable:%lukB"
4148 " slab_unreclaimable:%lukB"
4149 " kernel_stack:%lukB"
4156 " writeback_tmp:%lukB"
4157 " pages_scanned:%lu"
4158 " all_unreclaimable? %s"
4161 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4162 K(min_wmark_pages(zone
)),
4163 K(low_wmark_pages(zone
)),
4164 K(high_wmark_pages(zone
)),
4165 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
4166 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
4167 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
4168 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
4169 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
4170 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
4171 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
4172 K(zone
->present_pages
),
4173 K(zone
->managed_pages
),
4174 K(zone_page_state(zone
, NR_MLOCK
)),
4175 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
4176 K(zone_page_state(zone
, NR_WRITEBACK
)),
4177 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
4178 K(zone_page_state(zone
, NR_SHMEM
)),
4179 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4180 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4181 zone_page_state(zone
, NR_KERNEL_STACK
) *
4183 K(zone_page_state(zone
, NR_PAGETABLE
)),
4184 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
4185 K(zone_page_state(zone
, NR_BOUNCE
)),
4187 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4188 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
4189 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
4190 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
4191 (!zone_reclaimable(zone
) ? "yes" : "no")
4193 printk("lowmem_reserve[]:");
4194 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4195 printk(" %ld", zone
->lowmem_reserve
[i
]);
4199 for_each_populated_zone(zone
) {
4201 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4202 unsigned char types
[MAX_ORDER
];
4204 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4207 printk("%s: ", zone
->name
);
4209 spin_lock_irqsave(&zone
->lock
, flags
);
4210 for (order
= 0; order
< MAX_ORDER
; order
++) {
4211 struct free_area
*area
= &zone
->free_area
[order
];
4214 nr
[order
] = area
->nr_free
;
4215 total
+= nr
[order
] << order
;
4218 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4219 if (!list_empty(&area
->free_list
[type
]))
4220 types
[order
] |= 1 << type
;
4223 spin_unlock_irqrestore(&zone
->lock
, flags
);
4224 for (order
= 0; order
< MAX_ORDER
; order
++) {
4225 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
4227 show_migration_types(types
[order
]);
4229 printk("= %lukB\n", K(total
));
4232 hugetlb_show_meminfo();
4234 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
4236 show_swap_cache_info();
4239 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4241 zoneref
->zone
= zone
;
4242 zoneref
->zone_idx
= zone_idx(zone
);
4246 * Builds allocation fallback zone lists.
4248 * Add all populated zones of a node to the zonelist.
4250 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4254 enum zone_type zone_type
= MAX_NR_ZONES
;
4258 zone
= pgdat
->node_zones
+ zone_type
;
4259 if (populated_zone(zone
)) {
4260 zoneref_set_zone(zone
,
4261 &zonelist
->_zonerefs
[nr_zones
++]);
4262 check_highest_zone(zone_type
);
4264 } while (zone_type
);
4272 * 0 = automatic detection of better ordering.
4273 * 1 = order by ([node] distance, -zonetype)
4274 * 2 = order by (-zonetype, [node] distance)
4276 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4277 * the same zonelist. So only NUMA can configure this param.
4279 #define ZONELIST_ORDER_DEFAULT 0
4280 #define ZONELIST_ORDER_NODE 1
4281 #define ZONELIST_ORDER_ZONE 2
4283 /* zonelist order in the kernel.
4284 * set_zonelist_order() will set this to NODE or ZONE.
4286 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4287 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4291 /* The value user specified ....changed by config */
4292 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4293 /* string for sysctl */
4294 #define NUMA_ZONELIST_ORDER_LEN 16
4295 char numa_zonelist_order
[16] = "default";
4298 * interface for configure zonelist ordering.
4299 * command line option "numa_zonelist_order"
4300 * = "[dD]efault - default, automatic configuration.
4301 * = "[nN]ode - order by node locality, then by zone within node
4302 * = "[zZ]one - order by zone, then by locality within zone
4305 static int __parse_numa_zonelist_order(char *s
)
4307 if (*s
== 'd' || *s
== 'D') {
4308 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4309 } else if (*s
== 'n' || *s
== 'N') {
4310 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4311 } else if (*s
== 'z' || *s
== 'Z') {
4312 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4314 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4320 static __init
int setup_numa_zonelist_order(char *s
)
4327 ret
= __parse_numa_zonelist_order(s
);
4329 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4333 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4336 * sysctl handler for numa_zonelist_order
4338 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4339 void __user
*buffer
, size_t *length
,
4342 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4344 static DEFINE_MUTEX(zl_order_mutex
);
4346 mutex_lock(&zl_order_mutex
);
4348 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4352 strcpy(saved_string
, (char *)table
->data
);
4354 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4358 int oldval
= user_zonelist_order
;
4360 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4363 * bogus value. restore saved string
4365 strncpy((char *)table
->data
, saved_string
,
4366 NUMA_ZONELIST_ORDER_LEN
);
4367 user_zonelist_order
= oldval
;
4368 } else if (oldval
!= user_zonelist_order
) {
4369 mutex_lock(&zonelists_mutex
);
4370 build_all_zonelists(NULL
, NULL
);
4371 mutex_unlock(&zonelists_mutex
);
4375 mutex_unlock(&zl_order_mutex
);
4380 #define MAX_NODE_LOAD (nr_online_nodes)
4381 static int node_load
[MAX_NUMNODES
];
4384 * find_next_best_node - find the next node that should appear in a given node's fallback list
4385 * @node: node whose fallback list we're appending
4386 * @used_node_mask: nodemask_t of already used nodes
4388 * We use a number of factors to determine which is the next node that should
4389 * appear on a given node's fallback list. The node should not have appeared
4390 * already in @node's fallback list, and it should be the next closest node
4391 * according to the distance array (which contains arbitrary distance values
4392 * from each node to each node in the system), and should also prefer nodes
4393 * with no CPUs, since presumably they'll have very little allocation pressure
4394 * on them otherwise.
4395 * It returns -1 if no node is found.
4397 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4400 int min_val
= INT_MAX
;
4401 int best_node
= NUMA_NO_NODE
;
4402 const struct cpumask
*tmp
= cpumask_of_node(0);
4404 /* Use the local node if we haven't already */
4405 if (!node_isset(node
, *used_node_mask
)) {
4406 node_set(node
, *used_node_mask
);
4410 for_each_node_state(n
, N_MEMORY
) {
4412 /* Don't want a node to appear more than once */
4413 if (node_isset(n
, *used_node_mask
))
4416 /* Use the distance array to find the distance */
4417 val
= node_distance(node
, n
);
4419 /* Penalize nodes under us ("prefer the next node") */
4422 /* Give preference to headless and unused nodes */
4423 tmp
= cpumask_of_node(n
);
4424 if (!cpumask_empty(tmp
))
4425 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4427 /* Slight preference for less loaded node */
4428 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4429 val
+= node_load
[n
];
4431 if (val
< min_val
) {
4438 node_set(best_node
, *used_node_mask
);
4445 * Build zonelists ordered by node and zones within node.
4446 * This results in maximum locality--normal zone overflows into local
4447 * DMA zone, if any--but risks exhausting DMA zone.
4449 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4452 struct zonelist
*zonelist
;
4454 zonelist
= &pgdat
->node_zonelists
[0];
4455 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4457 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4458 zonelist
->_zonerefs
[j
].zone
= NULL
;
4459 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4463 * Build gfp_thisnode zonelists
4465 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4468 struct zonelist
*zonelist
;
4470 zonelist
= &pgdat
->node_zonelists
[1];
4471 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4472 zonelist
->_zonerefs
[j
].zone
= NULL
;
4473 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4477 * Build zonelists ordered by zone and nodes within zones.
4478 * This results in conserving DMA zone[s] until all Normal memory is
4479 * exhausted, but results in overflowing to remote node while memory
4480 * may still exist in local DMA zone.
4482 static int node_order
[MAX_NUMNODES
];
4484 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4487 int zone_type
; /* needs to be signed */
4489 struct zonelist
*zonelist
;
4491 zonelist
= &pgdat
->node_zonelists
[0];
4493 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4494 for (j
= 0; j
< nr_nodes
; j
++) {
4495 node
= node_order
[j
];
4496 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4497 if (populated_zone(z
)) {
4499 &zonelist
->_zonerefs
[pos
++]);
4500 check_highest_zone(zone_type
);
4504 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4505 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4508 #if defined(CONFIG_64BIT)
4510 * Devices that require DMA32/DMA are relatively rare and do not justify a
4511 * penalty to every machine in case the specialised case applies. Default
4512 * to Node-ordering on 64-bit NUMA machines
4514 static int default_zonelist_order(void)
4516 return ZONELIST_ORDER_NODE
;
4520 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4521 * by the kernel. If processes running on node 0 deplete the low memory zone
4522 * then reclaim will occur more frequency increasing stalls and potentially
4523 * be easier to OOM if a large percentage of the zone is under writeback or
4524 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4525 * Hence, default to zone ordering on 32-bit.
4527 static int default_zonelist_order(void)
4529 return ZONELIST_ORDER_ZONE
;
4531 #endif /* CONFIG_64BIT */
4533 static void set_zonelist_order(void)
4535 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4536 current_zonelist_order
= default_zonelist_order();
4538 current_zonelist_order
= user_zonelist_order
;
4541 static void build_zonelists(pg_data_t
*pgdat
)
4544 nodemask_t used_mask
;
4545 int local_node
, prev_node
;
4546 struct zonelist
*zonelist
;
4547 unsigned int order
= current_zonelist_order
;
4549 /* initialize zonelists */
4550 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4551 zonelist
= pgdat
->node_zonelists
+ i
;
4552 zonelist
->_zonerefs
[0].zone
= NULL
;
4553 zonelist
->_zonerefs
[0].zone_idx
= 0;
4556 /* NUMA-aware ordering of nodes */
4557 local_node
= pgdat
->node_id
;
4558 load
= nr_online_nodes
;
4559 prev_node
= local_node
;
4560 nodes_clear(used_mask
);
4562 memset(node_order
, 0, sizeof(node_order
));
4565 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4567 * We don't want to pressure a particular node.
4568 * So adding penalty to the first node in same
4569 * distance group to make it round-robin.
4571 if (node_distance(local_node
, node
) !=
4572 node_distance(local_node
, prev_node
))
4573 node_load
[node
] = load
;
4577 if (order
== ZONELIST_ORDER_NODE
)
4578 build_zonelists_in_node_order(pgdat
, node
);
4580 node_order
[i
++] = node
; /* remember order */
4583 if (order
== ZONELIST_ORDER_ZONE
) {
4584 /* calculate node order -- i.e., DMA last! */
4585 build_zonelists_in_zone_order(pgdat
, i
);
4588 build_thisnode_zonelists(pgdat
);
4591 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4593 * Return node id of node used for "local" allocations.
4594 * I.e., first node id of first zone in arg node's generic zonelist.
4595 * Used for initializing percpu 'numa_mem', which is used primarily
4596 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4598 int local_memory_node(int node
)
4602 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4603 gfp_zone(GFP_KERNEL
),
4605 return z
->zone
->node
;
4609 #else /* CONFIG_NUMA */
4611 static void set_zonelist_order(void)
4613 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4616 static void build_zonelists(pg_data_t
*pgdat
)
4618 int node
, local_node
;
4620 struct zonelist
*zonelist
;
4622 local_node
= pgdat
->node_id
;
4624 zonelist
= &pgdat
->node_zonelists
[0];
4625 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4628 * Now we build the zonelist so that it contains the zones
4629 * of all the other nodes.
4630 * We don't want to pressure a particular node, so when
4631 * building the zones for node N, we make sure that the
4632 * zones coming right after the local ones are those from
4633 * node N+1 (modulo N)
4635 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4636 if (!node_online(node
))
4638 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4640 for (node
= 0; node
< local_node
; node
++) {
4641 if (!node_online(node
))
4643 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4646 zonelist
->_zonerefs
[j
].zone
= NULL
;
4647 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4650 #endif /* CONFIG_NUMA */
4653 * Boot pageset table. One per cpu which is going to be used for all
4654 * zones and all nodes. The parameters will be set in such a way
4655 * that an item put on a list will immediately be handed over to
4656 * the buddy list. This is safe since pageset manipulation is done
4657 * with interrupts disabled.
4659 * The boot_pagesets must be kept even after bootup is complete for
4660 * unused processors and/or zones. They do play a role for bootstrapping
4661 * hotplugged processors.
4663 * zoneinfo_show() and maybe other functions do
4664 * not check if the processor is online before following the pageset pointer.
4665 * Other parts of the kernel may not check if the zone is available.
4667 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4668 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4669 static void setup_zone_pageset(struct zone
*zone
);
4672 * Global mutex to protect against size modification of zonelists
4673 * as well as to serialize pageset setup for the new populated zone.
4675 DEFINE_MUTEX(zonelists_mutex
);
4677 /* return values int ....just for stop_machine() */
4678 static int __build_all_zonelists(void *data
)
4682 pg_data_t
*self
= data
;
4685 memset(node_load
, 0, sizeof(node_load
));
4688 if (self
&& !node_online(self
->node_id
)) {
4689 build_zonelists(self
);
4692 for_each_online_node(nid
) {
4693 pg_data_t
*pgdat
= NODE_DATA(nid
);
4695 build_zonelists(pgdat
);
4699 * Initialize the boot_pagesets that are going to be used
4700 * for bootstrapping processors. The real pagesets for
4701 * each zone will be allocated later when the per cpu
4702 * allocator is available.
4704 * boot_pagesets are used also for bootstrapping offline
4705 * cpus if the system is already booted because the pagesets
4706 * are needed to initialize allocators on a specific cpu too.
4707 * F.e. the percpu allocator needs the page allocator which
4708 * needs the percpu allocator in order to allocate its pagesets
4709 * (a chicken-egg dilemma).
4711 for_each_possible_cpu(cpu
) {
4712 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4714 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4716 * We now know the "local memory node" for each node--
4717 * i.e., the node of the first zone in the generic zonelist.
4718 * Set up numa_mem percpu variable for on-line cpus. During
4719 * boot, only the boot cpu should be on-line; we'll init the
4720 * secondary cpus' numa_mem as they come on-line. During
4721 * node/memory hotplug, we'll fixup all on-line cpus.
4723 if (cpu_online(cpu
))
4724 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4731 static noinline
void __init
4732 build_all_zonelists_init(void)
4734 __build_all_zonelists(NULL
);
4735 mminit_verify_zonelist();
4736 cpuset_init_current_mems_allowed();
4740 * Called with zonelists_mutex held always
4741 * unless system_state == SYSTEM_BOOTING.
4743 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4744 * [we're only called with non-NULL zone through __meminit paths] and
4745 * (2) call of __init annotated helper build_all_zonelists_init
4746 * [protected by SYSTEM_BOOTING].
4748 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4750 set_zonelist_order();
4752 if (system_state
== SYSTEM_BOOTING
) {
4753 build_all_zonelists_init();
4755 #ifdef CONFIG_MEMORY_HOTPLUG
4757 setup_zone_pageset(zone
);
4759 /* we have to stop all cpus to guarantee there is no user
4761 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4762 /* cpuset refresh routine should be here */
4764 vm_total_pages
= nr_free_pagecache_pages();
4766 * Disable grouping by mobility if the number of pages in the
4767 * system is too low to allow the mechanism to work. It would be
4768 * more accurate, but expensive to check per-zone. This check is
4769 * made on memory-hotadd so a system can start with mobility
4770 * disabled and enable it later
4772 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4773 page_group_by_mobility_disabled
= 1;
4775 page_group_by_mobility_disabled
= 0;
4777 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4779 zonelist_order_name
[current_zonelist_order
],
4780 page_group_by_mobility_disabled
? "off" : "on",
4783 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4788 * Helper functions to size the waitqueue hash table.
4789 * Essentially these want to choose hash table sizes sufficiently
4790 * large so that collisions trying to wait on pages are rare.
4791 * But in fact, the number of active page waitqueues on typical
4792 * systems is ridiculously low, less than 200. So this is even
4793 * conservative, even though it seems large.
4795 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4796 * waitqueues, i.e. the size of the waitq table given the number of pages.
4798 #define PAGES_PER_WAITQUEUE 256
4800 #ifndef CONFIG_MEMORY_HOTPLUG
4801 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4803 unsigned long size
= 1;
4805 pages
/= PAGES_PER_WAITQUEUE
;
4807 while (size
< pages
)
4811 * Once we have dozens or even hundreds of threads sleeping
4812 * on IO we've got bigger problems than wait queue collision.
4813 * Limit the size of the wait table to a reasonable size.
4815 size
= min(size
, 4096UL);
4817 return max(size
, 4UL);
4821 * A zone's size might be changed by hot-add, so it is not possible to determine
4822 * a suitable size for its wait_table. So we use the maximum size now.
4824 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4826 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4827 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4828 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4830 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4831 * or more by the traditional way. (See above). It equals:
4833 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4834 * ia64(16K page size) : = ( 8G + 4M)byte.
4835 * powerpc (64K page size) : = (32G +16M)byte.
4837 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4844 * This is an integer logarithm so that shifts can be used later
4845 * to extract the more random high bits from the multiplicative
4846 * hash function before the remainder is taken.
4848 static inline unsigned long wait_table_bits(unsigned long size
)
4854 * Initially all pages are reserved - free ones are freed
4855 * up by free_all_bootmem() once the early boot process is
4856 * done. Non-atomic initialization, single-pass.
4858 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4859 unsigned long start_pfn
, enum memmap_context context
)
4861 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
4862 unsigned long end_pfn
= start_pfn
+ size
;
4863 pg_data_t
*pgdat
= NODE_DATA(nid
);
4865 unsigned long nr_initialised
= 0;
4866 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4867 struct memblock_region
*r
= NULL
, *tmp
;
4870 if (highest_memmap_pfn
< end_pfn
- 1)
4871 highest_memmap_pfn
= end_pfn
- 1;
4874 * Honor reservation requested by the driver for this ZONE_DEVICE
4877 if (altmap
&& start_pfn
== altmap
->base_pfn
)
4878 start_pfn
+= altmap
->reserve
;
4880 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4882 * There can be holes in boot-time mem_map[]s handed to this
4883 * function. They do not exist on hotplugged memory.
4885 if (context
!= MEMMAP_EARLY
)
4888 if (!early_pfn_valid(pfn
))
4890 if (!early_pfn_in_nid(pfn
, nid
))
4892 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
4895 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4897 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4898 * from zone_movable_pfn[nid] to end of each node should be
4899 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4901 if (!mirrored_kernelcore
&& zone_movable_pfn
[nid
])
4902 if (zone
== ZONE_NORMAL
&& pfn
>= zone_movable_pfn
[nid
])
4906 * Check given memblock attribute by firmware which can affect
4907 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4908 * mirrored, it's an overlapped memmap init. skip it.
4910 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
4911 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
4912 for_each_memblock(memory
, tmp
)
4913 if (pfn
< memblock_region_memory_end_pfn(tmp
))
4917 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
4918 memblock_is_mirror(r
)) {
4919 /* already initialized as NORMAL */
4920 pfn
= memblock_region_memory_end_pfn(r
);
4928 * Mark the block movable so that blocks are reserved for
4929 * movable at startup. This will force kernel allocations
4930 * to reserve their blocks rather than leaking throughout
4931 * the address space during boot when many long-lived
4932 * kernel allocations are made.
4934 * bitmap is created for zone's valid pfn range. but memmap
4935 * can be created for invalid pages (for alignment)
4936 * check here not to call set_pageblock_migratetype() against
4939 if (!(pfn
& (pageblock_nr_pages
- 1))) {
4940 struct page
*page
= pfn_to_page(pfn
);
4942 __init_single_page(page
, pfn
, zone
, nid
);
4943 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4945 __init_single_pfn(pfn
, zone
, nid
);
4950 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4952 unsigned int order
, t
;
4953 for_each_migratetype_order(order
, t
) {
4954 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4955 zone
->free_area
[order
].nr_free
= 0;
4959 #ifndef __HAVE_ARCH_MEMMAP_INIT
4960 #define memmap_init(size, nid, zone, start_pfn) \
4961 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4964 static int zone_batchsize(struct zone
*zone
)
4970 * The per-cpu-pages pools are set to around 1000th of the
4971 * size of the zone. But no more than 1/2 of a meg.
4973 * OK, so we don't know how big the cache is. So guess.
4975 batch
= zone
->managed_pages
/ 1024;
4976 if (batch
* PAGE_SIZE
> 512 * 1024)
4977 batch
= (512 * 1024) / PAGE_SIZE
;
4978 batch
/= 4; /* We effectively *= 4 below */
4983 * Clamp the batch to a 2^n - 1 value. Having a power
4984 * of 2 value was found to be more likely to have
4985 * suboptimal cache aliasing properties in some cases.
4987 * For example if 2 tasks are alternately allocating
4988 * batches of pages, one task can end up with a lot
4989 * of pages of one half of the possible page colors
4990 * and the other with pages of the other colors.
4992 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4997 /* The deferral and batching of frees should be suppressed under NOMMU
5000 * The problem is that NOMMU needs to be able to allocate large chunks
5001 * of contiguous memory as there's no hardware page translation to
5002 * assemble apparent contiguous memory from discontiguous pages.
5004 * Queueing large contiguous runs of pages for batching, however,
5005 * causes the pages to actually be freed in smaller chunks. As there
5006 * can be a significant delay between the individual batches being
5007 * recycled, this leads to the once large chunks of space being
5008 * fragmented and becoming unavailable for high-order allocations.
5015 * pcp->high and pcp->batch values are related and dependent on one another:
5016 * ->batch must never be higher then ->high.
5017 * The following function updates them in a safe manner without read side
5020 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5021 * those fields changing asynchronously (acording the the above rule).
5023 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5024 * outside of boot time (or some other assurance that no concurrent updaters
5027 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5028 unsigned long batch
)
5030 /* start with a fail safe value for batch */
5034 /* Update high, then batch, in order */
5041 /* a companion to pageset_set_high() */
5042 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5044 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5047 static void pageset_init(struct per_cpu_pageset
*p
)
5049 struct per_cpu_pages
*pcp
;
5052 memset(p
, 0, sizeof(*p
));
5056 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5057 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5060 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5063 pageset_set_batch(p
, batch
);
5067 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5068 * to the value high for the pageset p.
5070 static void pageset_set_high(struct per_cpu_pageset
*p
,
5073 unsigned long batch
= max(1UL, high
/ 4);
5074 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5075 batch
= PAGE_SHIFT
* 8;
5077 pageset_update(&p
->pcp
, high
, batch
);
5080 static void pageset_set_high_and_batch(struct zone
*zone
,
5081 struct per_cpu_pageset
*pcp
)
5083 if (percpu_pagelist_fraction
)
5084 pageset_set_high(pcp
,
5085 (zone
->managed_pages
/
5086 percpu_pagelist_fraction
));
5088 pageset_set_batch(pcp
, zone_batchsize(zone
));
5091 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5093 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5096 pageset_set_high_and_batch(zone
, pcp
);
5099 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5102 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5103 for_each_possible_cpu(cpu
)
5104 zone_pageset_init(zone
, cpu
);
5108 * Allocate per cpu pagesets and initialize them.
5109 * Before this call only boot pagesets were available.
5111 void __init
setup_per_cpu_pageset(void)
5115 for_each_populated_zone(zone
)
5116 setup_zone_pageset(zone
);
5119 static noinline __init_refok
5120 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
5126 * The per-page waitqueue mechanism uses hashed waitqueues
5129 zone
->wait_table_hash_nr_entries
=
5130 wait_table_hash_nr_entries(zone_size_pages
);
5131 zone
->wait_table_bits
=
5132 wait_table_bits(zone
->wait_table_hash_nr_entries
);
5133 alloc_size
= zone
->wait_table_hash_nr_entries
5134 * sizeof(wait_queue_head_t
);
5136 if (!slab_is_available()) {
5137 zone
->wait_table
= (wait_queue_head_t
*)
5138 memblock_virt_alloc_node_nopanic(
5139 alloc_size
, zone
->zone_pgdat
->node_id
);
5142 * This case means that a zone whose size was 0 gets new memory
5143 * via memory hot-add.
5144 * But it may be the case that a new node was hot-added. In
5145 * this case vmalloc() will not be able to use this new node's
5146 * memory - this wait_table must be initialized to use this new
5147 * node itself as well.
5148 * To use this new node's memory, further consideration will be
5151 zone
->wait_table
= vmalloc(alloc_size
);
5153 if (!zone
->wait_table
)
5156 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
5157 init_waitqueue_head(zone
->wait_table
+ i
);
5162 static __meminit
void zone_pcp_init(struct zone
*zone
)
5165 * per cpu subsystem is not up at this point. The following code
5166 * relies on the ability of the linker to provide the
5167 * offset of a (static) per cpu variable into the per cpu area.
5169 zone
->pageset
= &boot_pageset
;
5171 if (populated_zone(zone
))
5172 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5173 zone
->name
, zone
->present_pages
,
5174 zone_batchsize(zone
));
5177 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5178 unsigned long zone_start_pfn
,
5181 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5183 ret
= zone_wait_table_init(zone
, size
);
5186 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5188 zone
->zone_start_pfn
= zone_start_pfn
;
5190 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5191 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5193 (unsigned long)zone_idx(zone
),
5194 zone_start_pfn
, (zone_start_pfn
+ size
));
5196 zone_init_free_lists(zone
);
5201 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5202 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5205 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5207 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5208 struct mminit_pfnnid_cache
*state
)
5210 unsigned long start_pfn
, end_pfn
;
5213 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5214 return state
->last_nid
;
5216 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5218 state
->last_start
= start_pfn
;
5219 state
->last_end
= end_pfn
;
5220 state
->last_nid
= nid
;
5225 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5228 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5229 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5230 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5232 * If an architecture guarantees that all ranges registered contain no holes
5233 * and may be freed, this this function may be used instead of calling
5234 * memblock_free_early_nid() manually.
5236 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5238 unsigned long start_pfn
, end_pfn
;
5241 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5242 start_pfn
= min(start_pfn
, max_low_pfn
);
5243 end_pfn
= min(end_pfn
, max_low_pfn
);
5245 if (start_pfn
< end_pfn
)
5246 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5247 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5253 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5254 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5256 * If an architecture guarantees that all ranges registered contain no holes and may
5257 * be freed, this function may be used instead of calling memory_present() manually.
5259 void __init
sparse_memory_present_with_active_regions(int nid
)
5261 unsigned long start_pfn
, end_pfn
;
5264 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5265 memory_present(this_nid
, start_pfn
, end_pfn
);
5269 * get_pfn_range_for_nid - Return the start and end page frames for a node
5270 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5271 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5272 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5274 * It returns the start and end page frame of a node based on information
5275 * provided by memblock_set_node(). If called for a node
5276 * with no available memory, a warning is printed and the start and end
5279 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5280 unsigned long *start_pfn
, unsigned long *end_pfn
)
5282 unsigned long this_start_pfn
, this_end_pfn
;
5288 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5289 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5290 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5293 if (*start_pfn
== -1UL)
5298 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5299 * assumption is made that zones within a node are ordered in monotonic
5300 * increasing memory addresses so that the "highest" populated zone is used
5302 static void __init
find_usable_zone_for_movable(void)
5305 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5306 if (zone_index
== ZONE_MOVABLE
)
5309 if (arch_zone_highest_possible_pfn
[zone_index
] >
5310 arch_zone_lowest_possible_pfn
[zone_index
])
5314 VM_BUG_ON(zone_index
== -1);
5315 movable_zone
= zone_index
;
5319 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5320 * because it is sized independent of architecture. Unlike the other zones,
5321 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5322 * in each node depending on the size of each node and how evenly kernelcore
5323 * is distributed. This helper function adjusts the zone ranges
5324 * provided by the architecture for a given node by using the end of the
5325 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5326 * zones within a node are in order of monotonic increases memory addresses
5328 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5329 unsigned long zone_type
,
5330 unsigned long node_start_pfn
,
5331 unsigned long node_end_pfn
,
5332 unsigned long *zone_start_pfn
,
5333 unsigned long *zone_end_pfn
)
5335 /* Only adjust if ZONE_MOVABLE is on this node */
5336 if (zone_movable_pfn
[nid
]) {
5337 /* Size ZONE_MOVABLE */
5338 if (zone_type
== ZONE_MOVABLE
) {
5339 *zone_start_pfn
= zone_movable_pfn
[nid
];
5340 *zone_end_pfn
= min(node_end_pfn
,
5341 arch_zone_highest_possible_pfn
[movable_zone
]);
5343 /* Check if this whole range is within ZONE_MOVABLE */
5344 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5345 *zone_start_pfn
= *zone_end_pfn
;
5350 * Return the number of pages a zone spans in a node, including holes
5351 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5353 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5354 unsigned long zone_type
,
5355 unsigned long node_start_pfn
,
5356 unsigned long node_end_pfn
,
5357 unsigned long *zone_start_pfn
,
5358 unsigned long *zone_end_pfn
,
5359 unsigned long *ignored
)
5361 /* When hotadd a new node from cpu_up(), the node should be empty */
5362 if (!node_start_pfn
&& !node_end_pfn
)
5365 /* Get the start and end of the zone */
5366 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5367 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5368 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5369 node_start_pfn
, node_end_pfn
,
5370 zone_start_pfn
, zone_end_pfn
);
5372 /* Check that this node has pages within the zone's required range */
5373 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5376 /* Move the zone boundaries inside the node if necessary */
5377 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5378 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5380 /* Return the spanned pages */
5381 return *zone_end_pfn
- *zone_start_pfn
;
5385 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5386 * then all holes in the requested range will be accounted for.
5388 unsigned long __meminit
__absent_pages_in_range(int nid
,
5389 unsigned long range_start_pfn
,
5390 unsigned long range_end_pfn
)
5392 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5393 unsigned long start_pfn
, end_pfn
;
5396 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5397 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5398 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5399 nr_absent
-= end_pfn
- start_pfn
;
5405 * absent_pages_in_range - Return number of page frames in holes within a range
5406 * @start_pfn: The start PFN to start searching for holes
5407 * @end_pfn: The end PFN to stop searching for holes
5409 * It returns the number of pages frames in memory holes within a range.
5411 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5412 unsigned long end_pfn
)
5414 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5417 /* Return the number of page frames in holes in a zone on a node */
5418 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5419 unsigned long zone_type
,
5420 unsigned long node_start_pfn
,
5421 unsigned long node_end_pfn
,
5422 unsigned long *ignored
)
5424 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5425 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5426 unsigned long zone_start_pfn
, zone_end_pfn
;
5427 unsigned long nr_absent
;
5429 /* When hotadd a new node from cpu_up(), the node should be empty */
5430 if (!node_start_pfn
&& !node_end_pfn
)
5433 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5434 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5436 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5437 node_start_pfn
, node_end_pfn
,
5438 &zone_start_pfn
, &zone_end_pfn
);
5439 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5442 * ZONE_MOVABLE handling.
5443 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5446 if (zone_movable_pfn
[nid
]) {
5447 if (mirrored_kernelcore
) {
5448 unsigned long start_pfn
, end_pfn
;
5449 struct memblock_region
*r
;
5451 for_each_memblock(memory
, r
) {
5452 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5453 zone_start_pfn
, zone_end_pfn
);
5454 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5455 zone_start_pfn
, zone_end_pfn
);
5457 if (zone_type
== ZONE_MOVABLE
&&
5458 memblock_is_mirror(r
))
5459 nr_absent
+= end_pfn
- start_pfn
;
5461 if (zone_type
== ZONE_NORMAL
&&
5462 !memblock_is_mirror(r
))
5463 nr_absent
+= end_pfn
- start_pfn
;
5466 if (zone_type
== ZONE_NORMAL
)
5467 nr_absent
+= node_end_pfn
- zone_movable_pfn
[nid
];
5474 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5475 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5476 unsigned long zone_type
,
5477 unsigned long node_start_pfn
,
5478 unsigned long node_end_pfn
,
5479 unsigned long *zone_start_pfn
,
5480 unsigned long *zone_end_pfn
,
5481 unsigned long *zones_size
)
5485 *zone_start_pfn
= node_start_pfn
;
5486 for (zone
= 0; zone
< zone_type
; zone
++)
5487 *zone_start_pfn
+= zones_size
[zone
];
5489 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5491 return zones_size
[zone_type
];
5494 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5495 unsigned long zone_type
,
5496 unsigned long node_start_pfn
,
5497 unsigned long node_end_pfn
,
5498 unsigned long *zholes_size
)
5503 return zholes_size
[zone_type
];
5506 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5508 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5509 unsigned long node_start_pfn
,
5510 unsigned long node_end_pfn
,
5511 unsigned long *zones_size
,
5512 unsigned long *zholes_size
)
5514 unsigned long realtotalpages
= 0, totalpages
= 0;
5517 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5518 struct zone
*zone
= pgdat
->node_zones
+ i
;
5519 unsigned long zone_start_pfn
, zone_end_pfn
;
5520 unsigned long size
, real_size
;
5522 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5528 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5529 node_start_pfn
, node_end_pfn
,
5532 zone
->zone_start_pfn
= zone_start_pfn
;
5534 zone
->zone_start_pfn
= 0;
5535 zone
->spanned_pages
= size
;
5536 zone
->present_pages
= real_size
;
5539 realtotalpages
+= real_size
;
5542 pgdat
->node_spanned_pages
= totalpages
;
5543 pgdat
->node_present_pages
= realtotalpages
;
5544 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5548 #ifndef CONFIG_SPARSEMEM
5550 * Calculate the size of the zone->blockflags rounded to an unsigned long
5551 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5552 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5553 * round what is now in bits to nearest long in bits, then return it in
5556 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5558 unsigned long usemapsize
;
5560 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5561 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5562 usemapsize
= usemapsize
>> pageblock_order
;
5563 usemapsize
*= NR_PAGEBLOCK_BITS
;
5564 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5566 return usemapsize
/ 8;
5569 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5571 unsigned long zone_start_pfn
,
5572 unsigned long zonesize
)
5574 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5575 zone
->pageblock_flags
= NULL
;
5577 zone
->pageblock_flags
=
5578 memblock_virt_alloc_node_nopanic(usemapsize
,
5582 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5583 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5584 #endif /* CONFIG_SPARSEMEM */
5586 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5588 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5589 void __paginginit
set_pageblock_order(void)
5593 /* Check that pageblock_nr_pages has not already been setup */
5594 if (pageblock_order
)
5597 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5598 order
= HUGETLB_PAGE_ORDER
;
5600 order
= MAX_ORDER
- 1;
5603 * Assume the largest contiguous order of interest is a huge page.
5604 * This value may be variable depending on boot parameters on IA64 and
5607 pageblock_order
= order
;
5609 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5612 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5613 * is unused as pageblock_order is set at compile-time. See
5614 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5617 void __paginginit
set_pageblock_order(void)
5621 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5623 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5624 unsigned long present_pages
)
5626 unsigned long pages
= spanned_pages
;
5629 * Provide a more accurate estimation if there are holes within
5630 * the zone and SPARSEMEM is in use. If there are holes within the
5631 * zone, each populated memory region may cost us one or two extra
5632 * memmap pages due to alignment because memmap pages for each
5633 * populated regions may not naturally algined on page boundary.
5634 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5636 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5637 IS_ENABLED(CONFIG_SPARSEMEM
))
5638 pages
= present_pages
;
5640 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5644 * Set up the zone data structures:
5645 * - mark all pages reserved
5646 * - mark all memory queues empty
5647 * - clear the memory bitmaps
5649 * NOTE: pgdat should get zeroed by caller.
5651 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5654 int nid
= pgdat
->node_id
;
5657 pgdat_resize_init(pgdat
);
5658 #ifdef CONFIG_NUMA_BALANCING
5659 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5660 pgdat
->numabalancing_migrate_nr_pages
= 0;
5661 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5663 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5664 spin_lock_init(&pgdat
->split_queue_lock
);
5665 INIT_LIST_HEAD(&pgdat
->split_queue
);
5666 pgdat
->split_queue_len
= 0;
5668 init_waitqueue_head(&pgdat
->kswapd_wait
);
5669 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5670 #ifdef CONFIG_COMPACTION
5671 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5673 pgdat_page_ext_init(pgdat
);
5675 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5676 struct zone
*zone
= pgdat
->node_zones
+ j
;
5677 unsigned long size
, realsize
, freesize
, memmap_pages
;
5678 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5680 size
= zone
->spanned_pages
;
5681 realsize
= freesize
= zone
->present_pages
;
5684 * Adjust freesize so that it accounts for how much memory
5685 * is used by this zone for memmap. This affects the watermark
5686 * and per-cpu initialisations
5688 memmap_pages
= calc_memmap_size(size
, realsize
);
5689 if (!is_highmem_idx(j
)) {
5690 if (freesize
>= memmap_pages
) {
5691 freesize
-= memmap_pages
;
5694 " %s zone: %lu pages used for memmap\n",
5695 zone_names
[j
], memmap_pages
);
5697 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5698 zone_names
[j
], memmap_pages
, freesize
);
5701 /* Account for reserved pages */
5702 if (j
== 0 && freesize
> dma_reserve
) {
5703 freesize
-= dma_reserve
;
5704 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5705 zone_names
[0], dma_reserve
);
5708 if (!is_highmem_idx(j
))
5709 nr_kernel_pages
+= freesize
;
5710 /* Charge for highmem memmap if there are enough kernel pages */
5711 else if (nr_kernel_pages
> memmap_pages
* 2)
5712 nr_kernel_pages
-= memmap_pages
;
5713 nr_all_pages
+= freesize
;
5716 * Set an approximate value for lowmem here, it will be adjusted
5717 * when the bootmem allocator frees pages into the buddy system.
5718 * And all highmem pages will be managed by the buddy system.
5720 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5723 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
5725 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
5727 zone
->name
= zone_names
[j
];
5728 spin_lock_init(&zone
->lock
);
5729 spin_lock_init(&zone
->lru_lock
);
5730 zone_seqlock_init(zone
);
5731 zone
->zone_pgdat
= pgdat
;
5732 zone_pcp_init(zone
);
5734 /* For bootup, initialized properly in watermark setup */
5735 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
5737 lruvec_init(&zone
->lruvec
);
5741 set_pageblock_order();
5742 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5743 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5745 memmap_init(size
, nid
, j
, zone_start_pfn
);
5749 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
5751 unsigned long __maybe_unused start
= 0;
5752 unsigned long __maybe_unused offset
= 0;
5754 /* Skip empty nodes */
5755 if (!pgdat
->node_spanned_pages
)
5758 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5759 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5760 offset
= pgdat
->node_start_pfn
- start
;
5761 /* ia64 gets its own node_mem_map, before this, without bootmem */
5762 if (!pgdat
->node_mem_map
) {
5763 unsigned long size
, end
;
5767 * The zone's endpoints aren't required to be MAX_ORDER
5768 * aligned but the node_mem_map endpoints must be in order
5769 * for the buddy allocator to function correctly.
5771 end
= pgdat_end_pfn(pgdat
);
5772 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5773 size
= (end
- start
) * sizeof(struct page
);
5774 map
= alloc_remap(pgdat
->node_id
, size
);
5776 map
= memblock_virt_alloc_node_nopanic(size
,
5778 pgdat
->node_mem_map
= map
+ offset
;
5780 #ifndef CONFIG_NEED_MULTIPLE_NODES
5782 * With no DISCONTIG, the global mem_map is just set as node 0's
5784 if (pgdat
== NODE_DATA(0)) {
5785 mem_map
= NODE_DATA(0)->node_mem_map
;
5786 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5787 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5789 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5792 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5795 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5796 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5798 pg_data_t
*pgdat
= NODE_DATA(nid
);
5799 unsigned long start_pfn
= 0;
5800 unsigned long end_pfn
= 0;
5802 /* pg_data_t should be reset to zero when it's allocated */
5803 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5805 reset_deferred_meminit(pgdat
);
5806 pgdat
->node_id
= nid
;
5807 pgdat
->node_start_pfn
= node_start_pfn
;
5808 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5809 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5810 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5811 (u64
)start_pfn
<< PAGE_SHIFT
,
5812 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5814 start_pfn
= node_start_pfn
;
5816 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5817 zones_size
, zholes_size
);
5819 alloc_node_mem_map(pgdat
);
5820 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5821 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5822 nid
, (unsigned long)pgdat
,
5823 (unsigned long)pgdat
->node_mem_map
);
5826 free_area_init_core(pgdat
);
5829 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5831 #if MAX_NUMNODES > 1
5833 * Figure out the number of possible node ids.
5835 void __init
setup_nr_node_ids(void)
5837 unsigned int highest
;
5839 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5840 nr_node_ids
= highest
+ 1;
5845 * node_map_pfn_alignment - determine the maximum internode alignment
5847 * This function should be called after node map is populated and sorted.
5848 * It calculates the maximum power of two alignment which can distinguish
5851 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5852 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5853 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5854 * shifted, 1GiB is enough and this function will indicate so.
5856 * This is used to test whether pfn -> nid mapping of the chosen memory
5857 * model has fine enough granularity to avoid incorrect mapping for the
5858 * populated node map.
5860 * Returns the determined alignment in pfn's. 0 if there is no alignment
5861 * requirement (single node).
5863 unsigned long __init
node_map_pfn_alignment(void)
5865 unsigned long accl_mask
= 0, last_end
= 0;
5866 unsigned long start
, end
, mask
;
5870 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5871 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5878 * Start with a mask granular enough to pin-point to the
5879 * start pfn and tick off bits one-by-one until it becomes
5880 * too coarse to separate the current node from the last.
5882 mask
= ~((1 << __ffs(start
)) - 1);
5883 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5886 /* accumulate all internode masks */
5890 /* convert mask to number of pages */
5891 return ~accl_mask
+ 1;
5894 /* Find the lowest pfn for a node */
5895 static unsigned long __init
find_min_pfn_for_node(int nid
)
5897 unsigned long min_pfn
= ULONG_MAX
;
5898 unsigned long start_pfn
;
5901 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5902 min_pfn
= min(min_pfn
, start_pfn
);
5904 if (min_pfn
== ULONG_MAX
) {
5905 pr_warn("Could not find start_pfn for node %d\n", nid
);
5913 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5915 * It returns the minimum PFN based on information provided via
5916 * memblock_set_node().
5918 unsigned long __init
find_min_pfn_with_active_regions(void)
5920 return find_min_pfn_for_node(MAX_NUMNODES
);
5924 * early_calculate_totalpages()
5925 * Sum pages in active regions for movable zone.
5926 * Populate N_MEMORY for calculating usable_nodes.
5928 static unsigned long __init
early_calculate_totalpages(void)
5930 unsigned long totalpages
= 0;
5931 unsigned long start_pfn
, end_pfn
;
5934 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5935 unsigned long pages
= end_pfn
- start_pfn
;
5937 totalpages
+= pages
;
5939 node_set_state(nid
, N_MEMORY
);
5945 * Find the PFN the Movable zone begins in each node. Kernel memory
5946 * is spread evenly between nodes as long as the nodes have enough
5947 * memory. When they don't, some nodes will have more kernelcore than
5950 static void __init
find_zone_movable_pfns_for_nodes(void)
5953 unsigned long usable_startpfn
;
5954 unsigned long kernelcore_node
, kernelcore_remaining
;
5955 /* save the state before borrow the nodemask */
5956 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5957 unsigned long totalpages
= early_calculate_totalpages();
5958 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5959 struct memblock_region
*r
;
5961 /* Need to find movable_zone earlier when movable_node is specified. */
5962 find_usable_zone_for_movable();
5965 * If movable_node is specified, ignore kernelcore and movablecore
5968 if (movable_node_is_enabled()) {
5969 for_each_memblock(memory
, r
) {
5970 if (!memblock_is_hotpluggable(r
))
5975 usable_startpfn
= PFN_DOWN(r
->base
);
5976 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5977 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5985 * If kernelcore=mirror is specified, ignore movablecore option
5987 if (mirrored_kernelcore
) {
5988 bool mem_below_4gb_not_mirrored
= false;
5990 for_each_memblock(memory
, r
) {
5991 if (memblock_is_mirror(r
))
5996 usable_startpfn
= memblock_region_memory_base_pfn(r
);
5998 if (usable_startpfn
< 0x100000) {
5999 mem_below_4gb_not_mirrored
= true;
6003 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6004 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6008 if (mem_below_4gb_not_mirrored
)
6009 pr_warn("This configuration results in unmirrored kernel memory.");
6015 * If movablecore=nn[KMG] was specified, calculate what size of
6016 * kernelcore that corresponds so that memory usable for
6017 * any allocation type is evenly spread. If both kernelcore
6018 * and movablecore are specified, then the value of kernelcore
6019 * will be used for required_kernelcore if it's greater than
6020 * what movablecore would have allowed.
6022 if (required_movablecore
) {
6023 unsigned long corepages
;
6026 * Round-up so that ZONE_MOVABLE is at least as large as what
6027 * was requested by the user
6029 required_movablecore
=
6030 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6031 required_movablecore
= min(totalpages
, required_movablecore
);
6032 corepages
= totalpages
- required_movablecore
;
6034 required_kernelcore
= max(required_kernelcore
, corepages
);
6038 * If kernelcore was not specified or kernelcore size is larger
6039 * than totalpages, there is no ZONE_MOVABLE.
6041 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6044 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6045 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6048 /* Spread kernelcore memory as evenly as possible throughout nodes */
6049 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6050 for_each_node_state(nid
, N_MEMORY
) {
6051 unsigned long start_pfn
, end_pfn
;
6054 * Recalculate kernelcore_node if the division per node
6055 * now exceeds what is necessary to satisfy the requested
6056 * amount of memory for the kernel
6058 if (required_kernelcore
< kernelcore_node
)
6059 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6062 * As the map is walked, we track how much memory is usable
6063 * by the kernel using kernelcore_remaining. When it is
6064 * 0, the rest of the node is usable by ZONE_MOVABLE
6066 kernelcore_remaining
= kernelcore_node
;
6068 /* Go through each range of PFNs within this node */
6069 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6070 unsigned long size_pages
;
6072 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6073 if (start_pfn
>= end_pfn
)
6076 /* Account for what is only usable for kernelcore */
6077 if (start_pfn
< usable_startpfn
) {
6078 unsigned long kernel_pages
;
6079 kernel_pages
= min(end_pfn
, usable_startpfn
)
6082 kernelcore_remaining
-= min(kernel_pages
,
6083 kernelcore_remaining
);
6084 required_kernelcore
-= min(kernel_pages
,
6085 required_kernelcore
);
6087 /* Continue if range is now fully accounted */
6088 if (end_pfn
<= usable_startpfn
) {
6091 * Push zone_movable_pfn to the end so
6092 * that if we have to rebalance
6093 * kernelcore across nodes, we will
6094 * not double account here
6096 zone_movable_pfn
[nid
] = end_pfn
;
6099 start_pfn
= usable_startpfn
;
6103 * The usable PFN range for ZONE_MOVABLE is from
6104 * start_pfn->end_pfn. Calculate size_pages as the
6105 * number of pages used as kernelcore
6107 size_pages
= end_pfn
- start_pfn
;
6108 if (size_pages
> kernelcore_remaining
)
6109 size_pages
= kernelcore_remaining
;
6110 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6113 * Some kernelcore has been met, update counts and
6114 * break if the kernelcore for this node has been
6117 required_kernelcore
-= min(required_kernelcore
,
6119 kernelcore_remaining
-= size_pages
;
6120 if (!kernelcore_remaining
)
6126 * If there is still required_kernelcore, we do another pass with one
6127 * less node in the count. This will push zone_movable_pfn[nid] further
6128 * along on the nodes that still have memory until kernelcore is
6132 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6136 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6137 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6138 zone_movable_pfn
[nid
] =
6139 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6142 /* restore the node_state */
6143 node_states
[N_MEMORY
] = saved_node_state
;
6146 /* Any regular or high memory on that node ? */
6147 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6149 enum zone_type zone_type
;
6151 if (N_MEMORY
== N_NORMAL_MEMORY
)
6154 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6155 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6156 if (populated_zone(zone
)) {
6157 node_set_state(nid
, N_HIGH_MEMORY
);
6158 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6159 zone_type
<= ZONE_NORMAL
)
6160 node_set_state(nid
, N_NORMAL_MEMORY
);
6167 * free_area_init_nodes - Initialise all pg_data_t and zone data
6168 * @max_zone_pfn: an array of max PFNs for each zone
6170 * This will call free_area_init_node() for each active node in the system.
6171 * Using the page ranges provided by memblock_set_node(), the size of each
6172 * zone in each node and their holes is calculated. If the maximum PFN
6173 * between two adjacent zones match, it is assumed that the zone is empty.
6174 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6175 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6176 * starts where the previous one ended. For example, ZONE_DMA32 starts
6177 * at arch_max_dma_pfn.
6179 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6181 unsigned long start_pfn
, end_pfn
;
6184 /* Record where the zone boundaries are */
6185 memset(arch_zone_lowest_possible_pfn
, 0,
6186 sizeof(arch_zone_lowest_possible_pfn
));
6187 memset(arch_zone_highest_possible_pfn
, 0,
6188 sizeof(arch_zone_highest_possible_pfn
));
6189 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
6190 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
6191 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
6192 if (i
== ZONE_MOVABLE
)
6194 arch_zone_lowest_possible_pfn
[i
] =
6195 arch_zone_highest_possible_pfn
[i
-1];
6196 arch_zone_highest_possible_pfn
[i
] =
6197 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
6199 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6200 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6202 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6203 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6204 find_zone_movable_pfns_for_nodes();
6206 /* Print out the zone ranges */
6207 pr_info("Zone ranges:\n");
6208 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6209 if (i
== ZONE_MOVABLE
)
6211 pr_info(" %-8s ", zone_names
[i
]);
6212 if (arch_zone_lowest_possible_pfn
[i
] ==
6213 arch_zone_highest_possible_pfn
[i
])
6216 pr_cont("[mem %#018Lx-%#018Lx]\n",
6217 (u64
)arch_zone_lowest_possible_pfn
[i
]
6219 ((u64
)arch_zone_highest_possible_pfn
[i
]
6220 << PAGE_SHIFT
) - 1);
6223 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6224 pr_info("Movable zone start for each node\n");
6225 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6226 if (zone_movable_pfn
[i
])
6227 pr_info(" Node %d: %#018Lx\n", i
,
6228 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6231 /* Print out the early node map */
6232 pr_info("Early memory node ranges\n");
6233 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6234 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6235 (u64
)start_pfn
<< PAGE_SHIFT
,
6236 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6238 /* Initialise every node */
6239 mminit_verify_pageflags_layout();
6240 setup_nr_node_ids();
6241 for_each_online_node(nid
) {
6242 pg_data_t
*pgdat
= NODE_DATA(nid
);
6243 free_area_init_node(nid
, NULL
,
6244 find_min_pfn_for_node(nid
), NULL
);
6246 /* Any memory on that node */
6247 if (pgdat
->node_present_pages
)
6248 node_set_state(nid
, N_MEMORY
);
6249 check_for_memory(pgdat
, nid
);
6253 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6255 unsigned long long coremem
;
6259 coremem
= memparse(p
, &p
);
6260 *core
= coremem
>> PAGE_SHIFT
;
6262 /* Paranoid check that UL is enough for the coremem value */
6263 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6269 * kernelcore=size sets the amount of memory for use for allocations that
6270 * cannot be reclaimed or migrated.
6272 static int __init
cmdline_parse_kernelcore(char *p
)
6274 /* parse kernelcore=mirror */
6275 if (parse_option_str(p
, "mirror")) {
6276 mirrored_kernelcore
= true;
6280 return cmdline_parse_core(p
, &required_kernelcore
);
6284 * movablecore=size sets the amount of memory for use for allocations that
6285 * can be reclaimed or migrated.
6287 static int __init
cmdline_parse_movablecore(char *p
)
6289 return cmdline_parse_core(p
, &required_movablecore
);
6292 early_param("kernelcore", cmdline_parse_kernelcore
);
6293 early_param("movablecore", cmdline_parse_movablecore
);
6295 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6297 void adjust_managed_page_count(struct page
*page
, long count
)
6299 spin_lock(&managed_page_count_lock
);
6300 page_zone(page
)->managed_pages
+= count
;
6301 totalram_pages
+= count
;
6302 #ifdef CONFIG_HIGHMEM
6303 if (PageHighMem(page
))
6304 totalhigh_pages
+= count
;
6306 spin_unlock(&managed_page_count_lock
);
6308 EXPORT_SYMBOL(adjust_managed_page_count
);
6310 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6313 unsigned long pages
= 0;
6315 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6316 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6317 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6318 if ((unsigned int)poison
<= 0xFF)
6319 memset(pos
, poison
, PAGE_SIZE
);
6320 free_reserved_page(virt_to_page(pos
));
6324 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6325 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6329 EXPORT_SYMBOL(free_reserved_area
);
6331 #ifdef CONFIG_HIGHMEM
6332 void free_highmem_page(struct page
*page
)
6334 __free_reserved_page(page
);
6336 page_zone(page
)->managed_pages
++;
6342 void __init
mem_init_print_info(const char *str
)
6344 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6345 unsigned long init_code_size
, init_data_size
;
6347 physpages
= get_num_physpages();
6348 codesize
= _etext
- _stext
;
6349 datasize
= _edata
- _sdata
;
6350 rosize
= __end_rodata
- __start_rodata
;
6351 bss_size
= __bss_stop
- __bss_start
;
6352 init_data_size
= __init_end
- __init_begin
;
6353 init_code_size
= _einittext
- _sinittext
;
6356 * Detect special cases and adjust section sizes accordingly:
6357 * 1) .init.* may be embedded into .data sections
6358 * 2) .init.text.* may be out of [__init_begin, __init_end],
6359 * please refer to arch/tile/kernel/vmlinux.lds.S.
6360 * 3) .rodata.* may be embedded into .text or .data sections.
6362 #define adj_init_size(start, end, size, pos, adj) \
6364 if (start <= pos && pos < end && size > adj) \
6368 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6369 _sinittext
, init_code_size
);
6370 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6371 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6372 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6373 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6375 #undef adj_init_size
6377 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6378 #ifdef CONFIG_HIGHMEM
6382 nr_free_pages() << (PAGE_SHIFT
- 10),
6383 physpages
<< (PAGE_SHIFT
- 10),
6384 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6385 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6386 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6387 totalcma_pages
<< (PAGE_SHIFT
- 10),
6388 #ifdef CONFIG_HIGHMEM
6389 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6391 str
? ", " : "", str
? str
: "");
6395 * set_dma_reserve - set the specified number of pages reserved in the first zone
6396 * @new_dma_reserve: The number of pages to mark reserved
6398 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6399 * In the DMA zone, a significant percentage may be consumed by kernel image
6400 * and other unfreeable allocations which can skew the watermarks badly. This
6401 * function may optionally be used to account for unfreeable pages in the
6402 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6403 * smaller per-cpu batchsize.
6405 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6407 dma_reserve
= new_dma_reserve
;
6410 void __init
free_area_init(unsigned long *zones_size
)
6412 free_area_init_node(0, zones_size
,
6413 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6416 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6417 unsigned long action
, void *hcpu
)
6419 int cpu
= (unsigned long)hcpu
;
6421 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6422 lru_add_drain_cpu(cpu
);
6426 * Spill the event counters of the dead processor
6427 * into the current processors event counters.
6428 * This artificially elevates the count of the current
6431 vm_events_fold_cpu(cpu
);
6434 * Zero the differential counters of the dead processor
6435 * so that the vm statistics are consistent.
6437 * This is only okay since the processor is dead and cannot
6438 * race with what we are doing.
6440 cpu_vm_stats_fold(cpu
);
6445 void __init
page_alloc_init(void)
6447 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6451 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6452 * or min_free_kbytes changes.
6454 static void calculate_totalreserve_pages(void)
6456 struct pglist_data
*pgdat
;
6457 unsigned long reserve_pages
= 0;
6458 enum zone_type i
, j
;
6460 for_each_online_pgdat(pgdat
) {
6461 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6462 struct zone
*zone
= pgdat
->node_zones
+ i
;
6465 /* Find valid and maximum lowmem_reserve in the zone */
6466 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6467 if (zone
->lowmem_reserve
[j
] > max
)
6468 max
= zone
->lowmem_reserve
[j
];
6471 /* we treat the high watermark as reserved pages. */
6472 max
+= high_wmark_pages(zone
);
6474 if (max
> zone
->managed_pages
)
6475 max
= zone
->managed_pages
;
6477 zone
->totalreserve_pages
= max
;
6479 reserve_pages
+= max
;
6482 totalreserve_pages
= reserve_pages
;
6486 * setup_per_zone_lowmem_reserve - called whenever
6487 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6488 * has a correct pages reserved value, so an adequate number of
6489 * pages are left in the zone after a successful __alloc_pages().
6491 static void setup_per_zone_lowmem_reserve(void)
6493 struct pglist_data
*pgdat
;
6494 enum zone_type j
, idx
;
6496 for_each_online_pgdat(pgdat
) {
6497 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6498 struct zone
*zone
= pgdat
->node_zones
+ j
;
6499 unsigned long managed_pages
= zone
->managed_pages
;
6501 zone
->lowmem_reserve
[j
] = 0;
6505 struct zone
*lower_zone
;
6509 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6510 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6512 lower_zone
= pgdat
->node_zones
+ idx
;
6513 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6514 sysctl_lowmem_reserve_ratio
[idx
];
6515 managed_pages
+= lower_zone
->managed_pages
;
6520 /* update totalreserve_pages */
6521 calculate_totalreserve_pages();
6524 static void __setup_per_zone_wmarks(void)
6526 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6527 unsigned long lowmem_pages
= 0;
6529 unsigned long flags
;
6531 /* Calculate total number of !ZONE_HIGHMEM pages */
6532 for_each_zone(zone
) {
6533 if (!is_highmem(zone
))
6534 lowmem_pages
+= zone
->managed_pages
;
6537 for_each_zone(zone
) {
6540 spin_lock_irqsave(&zone
->lock
, flags
);
6541 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6542 do_div(tmp
, lowmem_pages
);
6543 if (is_highmem(zone
)) {
6545 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6546 * need highmem pages, so cap pages_min to a small
6549 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6550 * deltas control asynch page reclaim, and so should
6551 * not be capped for highmem.
6553 unsigned long min_pages
;
6555 min_pages
= zone
->managed_pages
/ 1024;
6556 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6557 zone
->watermark
[WMARK_MIN
] = min_pages
;
6560 * If it's a lowmem zone, reserve a number of pages
6561 * proportionate to the zone's size.
6563 zone
->watermark
[WMARK_MIN
] = tmp
;
6567 * Set the kswapd watermarks distance according to the
6568 * scale factor in proportion to available memory, but
6569 * ensure a minimum size on small systems.
6571 tmp
= max_t(u64
, tmp
>> 2,
6572 mult_frac(zone
->managed_pages
,
6573 watermark_scale_factor
, 10000));
6575 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6576 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6578 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
6579 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
6580 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
6582 spin_unlock_irqrestore(&zone
->lock
, flags
);
6585 /* update totalreserve_pages */
6586 calculate_totalreserve_pages();
6590 * setup_per_zone_wmarks - called when min_free_kbytes changes
6591 * or when memory is hot-{added|removed}
6593 * Ensures that the watermark[min,low,high] values for each zone are set
6594 * correctly with respect to min_free_kbytes.
6596 void setup_per_zone_wmarks(void)
6598 mutex_lock(&zonelists_mutex
);
6599 __setup_per_zone_wmarks();
6600 mutex_unlock(&zonelists_mutex
);
6604 * The inactive anon list should be small enough that the VM never has to
6605 * do too much work, but large enough that each inactive page has a chance
6606 * to be referenced again before it is swapped out.
6608 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6609 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6610 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6611 * the anonymous pages are kept on the inactive list.
6614 * memory ratio inactive anon
6615 * -------------------------------------
6624 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
6626 unsigned int gb
, ratio
;
6628 /* Zone size in gigabytes */
6629 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
6631 ratio
= int_sqrt(10 * gb
);
6635 zone
->inactive_ratio
= ratio
;
6638 static void __meminit
setup_per_zone_inactive_ratio(void)
6643 calculate_zone_inactive_ratio(zone
);
6647 * Initialise min_free_kbytes.
6649 * For small machines we want it small (128k min). For large machines
6650 * we want it large (64MB max). But it is not linear, because network
6651 * bandwidth does not increase linearly with machine size. We use
6653 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6654 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6670 int __meminit
init_per_zone_wmark_min(void)
6672 unsigned long lowmem_kbytes
;
6673 int new_min_free_kbytes
;
6675 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6676 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6678 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6679 min_free_kbytes
= new_min_free_kbytes
;
6680 if (min_free_kbytes
< 128)
6681 min_free_kbytes
= 128;
6682 if (min_free_kbytes
> 65536)
6683 min_free_kbytes
= 65536;
6685 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6686 new_min_free_kbytes
, user_min_free_kbytes
);
6688 setup_per_zone_wmarks();
6689 refresh_zone_stat_thresholds();
6690 setup_per_zone_lowmem_reserve();
6691 setup_per_zone_inactive_ratio();
6694 core_initcall(init_per_zone_wmark_min
)
6697 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6698 * that we can call two helper functions whenever min_free_kbytes
6701 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6702 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6706 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6711 user_min_free_kbytes
= min_free_kbytes
;
6712 setup_per_zone_wmarks();
6717 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6718 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6722 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6727 setup_per_zone_wmarks();
6733 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6734 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6739 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6744 zone
->min_unmapped_pages
= (zone
->managed_pages
*
6745 sysctl_min_unmapped_ratio
) / 100;
6749 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6750 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6755 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6760 zone
->min_slab_pages
= (zone
->managed_pages
*
6761 sysctl_min_slab_ratio
) / 100;
6767 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6768 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6769 * whenever sysctl_lowmem_reserve_ratio changes.
6771 * The reserve ratio obviously has absolutely no relation with the
6772 * minimum watermarks. The lowmem reserve ratio can only make sense
6773 * if in function of the boot time zone sizes.
6775 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6776 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6778 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6779 setup_per_zone_lowmem_reserve();
6784 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6785 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6786 * pagelist can have before it gets flushed back to buddy allocator.
6788 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6789 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6792 int old_percpu_pagelist_fraction
;
6795 mutex_lock(&pcp_batch_high_lock
);
6796 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6798 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6799 if (!write
|| ret
< 0)
6802 /* Sanity checking to avoid pcp imbalance */
6803 if (percpu_pagelist_fraction
&&
6804 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6805 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6811 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6814 for_each_populated_zone(zone
) {
6817 for_each_possible_cpu(cpu
)
6818 pageset_set_high_and_batch(zone
,
6819 per_cpu_ptr(zone
->pageset
, cpu
));
6822 mutex_unlock(&pcp_batch_high_lock
);
6827 int hashdist
= HASHDIST_DEFAULT
;
6829 static int __init
set_hashdist(char *str
)
6833 hashdist
= simple_strtoul(str
, &str
, 0);
6836 __setup("hashdist=", set_hashdist
);
6840 * allocate a large system hash table from bootmem
6841 * - it is assumed that the hash table must contain an exact power-of-2
6842 * quantity of entries
6843 * - limit is the number of hash buckets, not the total allocation size
6845 void *__init
alloc_large_system_hash(const char *tablename
,
6846 unsigned long bucketsize
,
6847 unsigned long numentries
,
6850 unsigned int *_hash_shift
,
6851 unsigned int *_hash_mask
,
6852 unsigned long low_limit
,
6853 unsigned long high_limit
)
6855 unsigned long long max
= high_limit
;
6856 unsigned long log2qty
, size
;
6859 /* allow the kernel cmdline to have a say */
6861 /* round applicable memory size up to nearest megabyte */
6862 numentries
= nr_kernel_pages
;
6864 /* It isn't necessary when PAGE_SIZE >= 1MB */
6865 if (PAGE_SHIFT
< 20)
6866 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6868 /* limit to 1 bucket per 2^scale bytes of low memory */
6869 if (scale
> PAGE_SHIFT
)
6870 numentries
>>= (scale
- PAGE_SHIFT
);
6872 numentries
<<= (PAGE_SHIFT
- scale
);
6874 /* Make sure we've got at least a 0-order allocation.. */
6875 if (unlikely(flags
& HASH_SMALL
)) {
6876 /* Makes no sense without HASH_EARLY */
6877 WARN_ON(!(flags
& HASH_EARLY
));
6878 if (!(numentries
>> *_hash_shift
)) {
6879 numentries
= 1UL << *_hash_shift
;
6880 BUG_ON(!numentries
);
6882 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6883 numentries
= PAGE_SIZE
/ bucketsize
;
6885 numentries
= roundup_pow_of_two(numentries
);
6887 /* limit allocation size to 1/16 total memory by default */
6889 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6890 do_div(max
, bucketsize
);
6892 max
= min(max
, 0x80000000ULL
);
6894 if (numentries
< low_limit
)
6895 numentries
= low_limit
;
6896 if (numentries
> max
)
6899 log2qty
= ilog2(numentries
);
6902 size
= bucketsize
<< log2qty
;
6903 if (flags
& HASH_EARLY
)
6904 table
= memblock_virt_alloc_nopanic(size
, 0);
6906 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6909 * If bucketsize is not a power-of-two, we may free
6910 * some pages at the end of hash table which
6911 * alloc_pages_exact() automatically does
6913 if (get_order(size
) < MAX_ORDER
) {
6914 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6915 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6918 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6921 panic("Failed to allocate %s hash table\n", tablename
);
6923 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6924 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
6927 *_hash_shift
= log2qty
;
6929 *_hash_mask
= (1 << log2qty
) - 1;
6935 * This function checks whether pageblock includes unmovable pages or not.
6936 * If @count is not zero, it is okay to include less @count unmovable pages
6938 * PageLRU check without isolation or lru_lock could race so that
6939 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6940 * expect this function should be exact.
6942 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6943 bool skip_hwpoisoned_pages
)
6945 unsigned long pfn
, iter
, found
;
6949 * For avoiding noise data, lru_add_drain_all() should be called
6950 * If ZONE_MOVABLE, the zone never contains unmovable pages
6952 if (zone_idx(zone
) == ZONE_MOVABLE
)
6954 mt
= get_pageblock_migratetype(page
);
6955 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6958 pfn
= page_to_pfn(page
);
6959 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6960 unsigned long check
= pfn
+ iter
;
6962 if (!pfn_valid_within(check
))
6965 page
= pfn_to_page(check
);
6968 * Hugepages are not in LRU lists, but they're movable.
6969 * We need not scan over tail pages bacause we don't
6970 * handle each tail page individually in migration.
6972 if (PageHuge(page
)) {
6973 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6978 * We can't use page_count without pin a page
6979 * because another CPU can free compound page.
6980 * This check already skips compound tails of THP
6981 * because their page->_refcount is zero at all time.
6983 if (!page_ref_count(page
)) {
6984 if (PageBuddy(page
))
6985 iter
+= (1 << page_order(page
)) - 1;
6990 * The HWPoisoned page may be not in buddy system, and
6991 * page_count() is not 0.
6993 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6999 * If there are RECLAIMABLE pages, we need to check
7000 * it. But now, memory offline itself doesn't call
7001 * shrink_node_slabs() and it still to be fixed.
7004 * If the page is not RAM, page_count()should be 0.
7005 * we don't need more check. This is an _used_ not-movable page.
7007 * The problematic thing here is PG_reserved pages. PG_reserved
7008 * is set to both of a memory hole page and a _used_ kernel
7017 bool is_pageblock_removable_nolock(struct page
*page
)
7023 * We have to be careful here because we are iterating over memory
7024 * sections which are not zone aware so we might end up outside of
7025 * the zone but still within the section.
7026 * We have to take care about the node as well. If the node is offline
7027 * its NODE_DATA will be NULL - see page_zone.
7029 if (!node_online(page_to_nid(page
)))
7032 zone
= page_zone(page
);
7033 pfn
= page_to_pfn(page
);
7034 if (!zone_spans_pfn(zone
, pfn
))
7037 return !has_unmovable_pages(zone
, page
, 0, true);
7040 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7042 static unsigned long pfn_max_align_down(unsigned long pfn
)
7044 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7045 pageblock_nr_pages
) - 1);
7048 static unsigned long pfn_max_align_up(unsigned long pfn
)
7050 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7051 pageblock_nr_pages
));
7054 /* [start, end) must belong to a single zone. */
7055 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7056 unsigned long start
, unsigned long end
)
7058 /* This function is based on compact_zone() from compaction.c. */
7059 unsigned long nr_reclaimed
;
7060 unsigned long pfn
= start
;
7061 unsigned int tries
= 0;
7066 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7067 if (fatal_signal_pending(current
)) {
7072 if (list_empty(&cc
->migratepages
)) {
7073 cc
->nr_migratepages
= 0;
7074 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7080 } else if (++tries
== 5) {
7081 ret
= ret
< 0 ? ret
: -EBUSY
;
7085 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7087 cc
->nr_migratepages
-= nr_reclaimed
;
7089 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7090 NULL
, 0, cc
->mode
, MR_CMA
);
7093 putback_movable_pages(&cc
->migratepages
);
7100 * alloc_contig_range() -- tries to allocate given range of pages
7101 * @start: start PFN to allocate
7102 * @end: one-past-the-last PFN to allocate
7103 * @migratetype: migratetype of the underlaying pageblocks (either
7104 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7105 * in range must have the same migratetype and it must
7106 * be either of the two.
7108 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7109 * aligned, however it's the caller's responsibility to guarantee that
7110 * we are the only thread that changes migrate type of pageblocks the
7113 * The PFN range must belong to a single zone.
7115 * Returns zero on success or negative error code. On success all
7116 * pages which PFN is in [start, end) are allocated for the caller and
7117 * need to be freed with free_contig_range().
7119 int alloc_contig_range(unsigned long start
, unsigned long end
,
7120 unsigned migratetype
)
7122 unsigned long outer_start
, outer_end
;
7126 struct compact_control cc
= {
7127 .nr_migratepages
= 0,
7129 .zone
= page_zone(pfn_to_page(start
)),
7130 .mode
= MIGRATE_SYNC
,
7131 .ignore_skip_hint
= true,
7133 INIT_LIST_HEAD(&cc
.migratepages
);
7136 * What we do here is we mark all pageblocks in range as
7137 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7138 * have different sizes, and due to the way page allocator
7139 * work, we align the range to biggest of the two pages so
7140 * that page allocator won't try to merge buddies from
7141 * different pageblocks and change MIGRATE_ISOLATE to some
7142 * other migration type.
7144 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7145 * migrate the pages from an unaligned range (ie. pages that
7146 * we are interested in). This will put all the pages in
7147 * range back to page allocator as MIGRATE_ISOLATE.
7149 * When this is done, we take the pages in range from page
7150 * allocator removing them from the buddy system. This way
7151 * page allocator will never consider using them.
7153 * This lets us mark the pageblocks back as
7154 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7155 * aligned range but not in the unaligned, original range are
7156 * put back to page allocator so that buddy can use them.
7159 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7160 pfn_max_align_up(end
), migratetype
,
7166 * In case of -EBUSY, we'd like to know which page causes problem.
7167 * So, just fall through. We will check it in test_pages_isolated().
7169 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7170 if (ret
&& ret
!= -EBUSY
)
7174 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7175 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7176 * more, all pages in [start, end) are free in page allocator.
7177 * What we are going to do is to allocate all pages from
7178 * [start, end) (that is remove them from page allocator).
7180 * The only problem is that pages at the beginning and at the
7181 * end of interesting range may be not aligned with pages that
7182 * page allocator holds, ie. they can be part of higher order
7183 * pages. Because of this, we reserve the bigger range and
7184 * once this is done free the pages we are not interested in.
7186 * We don't have to hold zone->lock here because the pages are
7187 * isolated thus they won't get removed from buddy.
7190 lru_add_drain_all();
7191 drain_all_pages(cc
.zone
);
7194 outer_start
= start
;
7195 while (!PageBuddy(pfn_to_page(outer_start
))) {
7196 if (++order
>= MAX_ORDER
) {
7197 outer_start
= start
;
7200 outer_start
&= ~0UL << order
;
7203 if (outer_start
!= start
) {
7204 order
= page_order(pfn_to_page(outer_start
));
7207 * outer_start page could be small order buddy page and
7208 * it doesn't include start page. Adjust outer_start
7209 * in this case to report failed page properly
7210 * on tracepoint in test_pages_isolated()
7212 if (outer_start
+ (1UL << order
) <= start
)
7213 outer_start
= start
;
7216 /* Make sure the range is really isolated. */
7217 if (test_pages_isolated(outer_start
, end
, false)) {
7218 pr_info("%s: [%lx, %lx) PFNs busy\n",
7219 __func__
, outer_start
, end
);
7224 /* Grab isolated pages from freelists. */
7225 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7231 /* Free head and tail (if any) */
7232 if (start
!= outer_start
)
7233 free_contig_range(outer_start
, start
- outer_start
);
7234 if (end
!= outer_end
)
7235 free_contig_range(end
, outer_end
- end
);
7238 undo_isolate_page_range(pfn_max_align_down(start
),
7239 pfn_max_align_up(end
), migratetype
);
7243 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7245 unsigned int count
= 0;
7247 for (; nr_pages
--; pfn
++) {
7248 struct page
*page
= pfn_to_page(pfn
);
7250 count
+= page_count(page
) != 1;
7253 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7257 #ifdef CONFIG_MEMORY_HOTPLUG
7259 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7260 * page high values need to be recalulated.
7262 void __meminit
zone_pcp_update(struct zone
*zone
)
7265 mutex_lock(&pcp_batch_high_lock
);
7266 for_each_possible_cpu(cpu
)
7267 pageset_set_high_and_batch(zone
,
7268 per_cpu_ptr(zone
->pageset
, cpu
));
7269 mutex_unlock(&pcp_batch_high_lock
);
7273 void zone_pcp_reset(struct zone
*zone
)
7275 unsigned long flags
;
7277 struct per_cpu_pageset
*pset
;
7279 /* avoid races with drain_pages() */
7280 local_irq_save(flags
);
7281 if (zone
->pageset
!= &boot_pageset
) {
7282 for_each_online_cpu(cpu
) {
7283 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7284 drain_zonestat(zone
, pset
);
7286 free_percpu(zone
->pageset
);
7287 zone
->pageset
= &boot_pageset
;
7289 local_irq_restore(flags
);
7292 #ifdef CONFIG_MEMORY_HOTREMOVE
7294 * All pages in the range must be in a single zone and isolated
7295 * before calling this.
7298 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7302 unsigned int order
, i
;
7304 unsigned long flags
;
7305 /* find the first valid pfn */
7306 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7311 zone
= page_zone(pfn_to_page(pfn
));
7312 spin_lock_irqsave(&zone
->lock
, flags
);
7314 while (pfn
< end_pfn
) {
7315 if (!pfn_valid(pfn
)) {
7319 page
= pfn_to_page(pfn
);
7321 * The HWPoisoned page may be not in buddy system, and
7322 * page_count() is not 0.
7324 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7326 SetPageReserved(page
);
7330 BUG_ON(page_count(page
));
7331 BUG_ON(!PageBuddy(page
));
7332 order
= page_order(page
);
7333 #ifdef CONFIG_DEBUG_VM
7334 pr_info("remove from free list %lx %d %lx\n",
7335 pfn
, 1 << order
, end_pfn
);
7337 list_del(&page
->lru
);
7338 rmv_page_order(page
);
7339 zone
->free_area
[order
].nr_free
--;
7340 for (i
= 0; i
< (1 << order
); i
++)
7341 SetPageReserved((page
+i
));
7342 pfn
+= (1 << order
);
7344 spin_unlock_irqrestore(&zone
->lock
, flags
);
7348 bool is_free_buddy_page(struct page
*page
)
7350 struct zone
*zone
= page_zone(page
);
7351 unsigned long pfn
= page_to_pfn(page
);
7352 unsigned long flags
;
7355 spin_lock_irqsave(&zone
->lock
, flags
);
7356 for (order
= 0; order
< MAX_ORDER
; order
++) {
7357 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7359 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7362 spin_unlock_irqrestore(&zone
->lock
, flags
);
7364 return order
< MAX_ORDER
;