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
);
944 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
949 * We rely page->lru.next never has bit 0 set, unless the page
950 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
952 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
954 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
958 switch (page
- head_page
) {
960 /* the first tail page: ->mapping is compound_mapcount() */
961 if (unlikely(compound_mapcount(page
))) {
962 bad_page(page
, "nonzero compound_mapcount", 0);
968 * the second tail page: ->mapping is
969 * page_deferred_list().next -- ignore value.
973 if (page
->mapping
!= TAIL_MAPPING
) {
974 bad_page(page
, "corrupted mapping in tail page", 0);
979 if (unlikely(!PageTail(page
))) {
980 bad_page(page
, "PageTail not set", 0);
983 if (unlikely(compound_head(page
) != head_page
)) {
984 bad_page(page
, "compound_head not consistent", 0);
989 page
->mapping
= NULL
;
990 clear_compound_head(page
);
994 static bool free_pages_prepare(struct page
*page
, unsigned int order
);
996 #ifdef CONFIG_DEBUG_VM
997 static inline bool free_pcp_prepare(struct page
*page
)
999 return free_pages_prepare(page
, 0);
1002 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1007 static bool free_pcp_prepare(struct page
*page
)
1009 VM_BUG_ON_PAGE(PageTail(page
), page
);
1011 trace_mm_page_free(page
, 0);
1012 kmemcheck_free_shadow(page
, 0);
1013 kasan_free_pages(page
, 0);
1015 if (PageAnonHead(page
))
1016 page
->mapping
= NULL
;
1018 reset_page_owner(page
, 0);
1020 if (!PageHighMem(page
)) {
1021 debug_check_no_locks_freed(page_address(page
),
1023 debug_check_no_obj_freed(page_address(page
),
1026 arch_free_page(page
, 0);
1027 kernel_poison_pages(page
, 0, 0);
1028 kernel_map_pages(page
, 0, 0);
1030 page_cpupid_reset_last(page
);
1031 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1035 static bool bulkfree_pcp_prepare(struct page
*page
)
1037 return free_pages_check(page
);
1039 #endif /* CONFIG_DEBUG_VM */
1042 * Frees a number of pages from the PCP lists
1043 * Assumes all pages on list are in same zone, and of same order.
1044 * count is the number of pages to free.
1046 * If the zone was previously in an "all pages pinned" state then look to
1047 * see if this freeing clears that state.
1049 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1050 * pinned" detection logic.
1052 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1053 struct per_cpu_pages
*pcp
)
1055 int migratetype
= 0;
1057 unsigned long nr_scanned
;
1058 bool isolated_pageblocks
;
1060 spin_lock(&zone
->lock
);
1061 isolated_pageblocks
= has_isolate_pageblock(zone
);
1062 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
1064 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
1068 struct list_head
*list
;
1071 * Remove pages from lists in a round-robin fashion. A
1072 * batch_free count is maintained that is incremented when an
1073 * empty list is encountered. This is so more pages are freed
1074 * off fuller lists instead of spinning excessively around empty
1079 if (++migratetype
== MIGRATE_PCPTYPES
)
1081 list
= &pcp
->lists
[migratetype
];
1082 } while (list_empty(list
));
1084 /* This is the only non-empty list. Free them all. */
1085 if (batch_free
== MIGRATE_PCPTYPES
)
1089 int mt
; /* migratetype of the to-be-freed page */
1091 page
= list_last_entry(list
, struct page
, lru
);
1092 /* must delete as __free_one_page list manipulates */
1093 list_del(&page
->lru
);
1095 mt
= get_pcppage_migratetype(page
);
1096 /* MIGRATE_ISOLATE page should not go to pcplists */
1097 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1098 /* Pageblock could have been isolated meanwhile */
1099 if (unlikely(isolated_pageblocks
))
1100 mt
= get_pageblock_migratetype(page
);
1102 if (bulkfree_pcp_prepare(page
))
1105 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1106 trace_mm_page_pcpu_drain(page
, 0, mt
);
1107 } while (--count
&& --batch_free
&& !list_empty(list
));
1109 spin_unlock(&zone
->lock
);
1112 static void free_one_page(struct zone
*zone
,
1113 struct page
*page
, unsigned long pfn
,
1117 unsigned long nr_scanned
;
1118 spin_lock(&zone
->lock
);
1119 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
1121 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
1123 if (unlikely(has_isolate_pageblock(zone
) ||
1124 is_migrate_isolate(migratetype
))) {
1125 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1127 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1128 spin_unlock(&zone
->lock
);
1131 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1132 unsigned long zone
, int nid
)
1134 set_page_links(page
, zone
, nid
, pfn
);
1135 init_page_count(page
);
1136 page_mapcount_reset(page
);
1137 page_cpupid_reset_last(page
);
1139 INIT_LIST_HEAD(&page
->lru
);
1140 #ifdef WANT_PAGE_VIRTUAL
1141 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1142 if (!is_highmem_idx(zone
))
1143 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1147 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1150 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1153 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1154 static void init_reserved_page(unsigned long pfn
)
1159 if (!early_page_uninitialised(pfn
))
1162 nid
= early_pfn_to_nid(pfn
);
1163 pgdat
= NODE_DATA(nid
);
1165 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1166 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1168 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1171 __init_single_pfn(pfn
, zid
, nid
);
1174 static inline void init_reserved_page(unsigned long pfn
)
1177 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1180 * Initialised pages do not have PageReserved set. This function is
1181 * called for each range allocated by the bootmem allocator and
1182 * marks the pages PageReserved. The remaining valid pages are later
1183 * sent to the buddy page allocator.
1185 void __meminit
reserve_bootmem_region(unsigned long start
, unsigned long end
)
1187 unsigned long start_pfn
= PFN_DOWN(start
);
1188 unsigned long end_pfn
= PFN_UP(end
);
1190 for (; start_pfn
< end_pfn
; start_pfn
++) {
1191 if (pfn_valid(start_pfn
)) {
1192 struct page
*page
= pfn_to_page(start_pfn
);
1194 init_reserved_page(start_pfn
);
1196 /* Avoid false-positive PageTail() */
1197 INIT_LIST_HEAD(&page
->lru
);
1199 SetPageReserved(page
);
1204 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
1208 VM_BUG_ON_PAGE(PageTail(page
), page
);
1210 trace_mm_page_free(page
, order
);
1211 kmemcheck_free_shadow(page
, order
);
1212 kasan_free_pages(page
, order
);
1215 * Check tail pages before head page information is cleared to
1216 * avoid checking PageCompound for order-0 pages.
1218 if (unlikely(order
)) {
1219 bool compound
= PageCompound(page
);
1222 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1224 for (i
= 1; i
< (1 << order
); i
++) {
1226 bad
+= free_tail_pages_check(page
, page
+ i
);
1227 if (unlikely(free_pages_check(page
+ i
))) {
1231 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1234 if (PageAnonHead(page
))
1235 page
->mapping
= NULL
;
1236 bad
+= free_pages_check(page
);
1240 page_cpupid_reset_last(page
);
1241 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1242 reset_page_owner(page
, order
);
1244 if (!PageHighMem(page
)) {
1245 debug_check_no_locks_freed(page_address(page
),
1246 PAGE_SIZE
<< order
);
1247 debug_check_no_obj_freed(page_address(page
),
1248 PAGE_SIZE
<< order
);
1250 arch_free_page(page
, order
);
1251 kernel_poison_pages(page
, 1 << order
, 0);
1252 kernel_map_pages(page
, 1 << order
, 0);
1257 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1259 unsigned long flags
;
1261 unsigned long pfn
= page_to_pfn(page
);
1263 if (!free_pages_prepare(page
, order
))
1266 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1267 local_irq_save(flags
);
1268 __count_vm_events(PGFREE
, 1 << order
);
1269 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1270 local_irq_restore(flags
);
1273 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1275 unsigned int nr_pages
= 1 << order
;
1276 struct page
*p
= page
;
1280 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1282 __ClearPageReserved(p
);
1283 set_page_count(p
, 0);
1285 __ClearPageReserved(p
);
1286 set_page_count(p
, 0);
1288 page_zone(page
)->managed_pages
+= nr_pages
;
1289 set_page_refcounted(page
);
1290 __free_pages(page
, order
);
1293 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1294 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1296 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1298 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1300 static DEFINE_SPINLOCK(early_pfn_lock
);
1303 spin_lock(&early_pfn_lock
);
1304 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1307 spin_unlock(&early_pfn_lock
);
1313 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1314 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1315 struct mminit_pfnnid_cache
*state
)
1319 nid
= __early_pfn_to_nid(pfn
, state
);
1320 if (nid
>= 0 && nid
!= node
)
1325 /* Only safe to use early in boot when initialisation is single-threaded */
1326 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1328 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1333 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1337 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1338 struct mminit_pfnnid_cache
*state
)
1345 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1348 if (early_page_uninitialised(pfn
))
1350 return __free_pages_boot_core(page
, order
);
1354 * Check that the whole (or subset of) a pageblock given by the interval of
1355 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1356 * with the migration of free compaction scanner. The scanners then need to
1357 * use only pfn_valid_within() check for arches that allow holes within
1360 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1362 * It's possible on some configurations to have a setup like node0 node1 node0
1363 * i.e. it's possible that all pages within a zones range of pages do not
1364 * belong to a single zone. We assume that a border between node0 and node1
1365 * can occur within a single pageblock, but not a node0 node1 node0
1366 * interleaving within a single pageblock. It is therefore sufficient to check
1367 * the first and last page of a pageblock and avoid checking each individual
1368 * page in a pageblock.
1370 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1371 unsigned long end_pfn
, struct zone
*zone
)
1373 struct page
*start_page
;
1374 struct page
*end_page
;
1376 /* end_pfn is one past the range we are checking */
1379 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1382 start_page
= pfn_to_page(start_pfn
);
1384 if (page_zone(start_page
) != zone
)
1387 end_page
= pfn_to_page(end_pfn
);
1389 /* This gives a shorter code than deriving page_zone(end_page) */
1390 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1396 void set_zone_contiguous(struct zone
*zone
)
1398 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1399 unsigned long block_end_pfn
;
1401 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1402 for (; block_start_pfn
< zone_end_pfn(zone
);
1403 block_start_pfn
= block_end_pfn
,
1404 block_end_pfn
+= pageblock_nr_pages
) {
1406 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1408 if (!__pageblock_pfn_to_page(block_start_pfn
,
1409 block_end_pfn
, zone
))
1413 /* We confirm that there is no hole */
1414 zone
->contiguous
= true;
1417 void clear_zone_contiguous(struct zone
*zone
)
1419 zone
->contiguous
= false;
1422 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1423 static void __init
deferred_free_range(struct page
*page
,
1424 unsigned long pfn
, int nr_pages
)
1431 /* Free a large naturally-aligned chunk if possible */
1432 if (nr_pages
== MAX_ORDER_NR_PAGES
&&
1433 (pfn
& (MAX_ORDER_NR_PAGES
-1)) == 0) {
1434 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1435 __free_pages_boot_core(page
, MAX_ORDER
-1);
1439 for (i
= 0; i
< nr_pages
; i
++, page
++)
1440 __free_pages_boot_core(page
, 0);
1443 /* Completion tracking for deferred_init_memmap() threads */
1444 static atomic_t pgdat_init_n_undone __initdata
;
1445 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1447 static inline void __init
pgdat_init_report_one_done(void)
1449 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1450 complete(&pgdat_init_all_done_comp
);
1453 /* Initialise remaining memory on a node */
1454 static int __init
deferred_init_memmap(void *data
)
1456 pg_data_t
*pgdat
= data
;
1457 int nid
= pgdat
->node_id
;
1458 struct mminit_pfnnid_cache nid_init_state
= { };
1459 unsigned long start
= jiffies
;
1460 unsigned long nr_pages
= 0;
1461 unsigned long walk_start
, walk_end
;
1464 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1465 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1467 if (first_init_pfn
== ULONG_MAX
) {
1468 pgdat_init_report_one_done();
1472 /* Bind memory initialisation thread to a local node if possible */
1473 if (!cpumask_empty(cpumask
))
1474 set_cpus_allowed_ptr(current
, cpumask
);
1476 /* Sanity check boundaries */
1477 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1478 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1479 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1481 /* Only the highest zone is deferred so find it */
1482 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1483 zone
= pgdat
->node_zones
+ zid
;
1484 if (first_init_pfn
< zone_end_pfn(zone
))
1488 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1489 unsigned long pfn
, end_pfn
;
1490 struct page
*page
= NULL
;
1491 struct page
*free_base_page
= NULL
;
1492 unsigned long free_base_pfn
= 0;
1495 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1496 pfn
= first_init_pfn
;
1497 if (pfn
< walk_start
)
1499 if (pfn
< zone
->zone_start_pfn
)
1500 pfn
= zone
->zone_start_pfn
;
1502 for (; pfn
< end_pfn
; pfn
++) {
1503 if (!pfn_valid_within(pfn
))
1507 * Ensure pfn_valid is checked every
1508 * MAX_ORDER_NR_PAGES for memory holes
1510 if ((pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
1511 if (!pfn_valid(pfn
)) {
1517 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1522 /* Minimise pfn page lookups and scheduler checks */
1523 if (page
&& (pfn
& (MAX_ORDER_NR_PAGES
- 1)) != 0) {
1526 nr_pages
+= nr_to_free
;
1527 deferred_free_range(free_base_page
,
1528 free_base_pfn
, nr_to_free
);
1529 free_base_page
= NULL
;
1530 free_base_pfn
= nr_to_free
= 0;
1532 page
= pfn_to_page(pfn
);
1537 VM_BUG_ON(page_zone(page
) != zone
);
1541 __init_single_page(page
, pfn
, zid
, nid
);
1542 if (!free_base_page
) {
1543 free_base_page
= page
;
1544 free_base_pfn
= pfn
;
1549 /* Where possible, batch up pages for a single free */
1552 /* Free the current block of pages to allocator */
1553 nr_pages
+= nr_to_free
;
1554 deferred_free_range(free_base_page
, free_base_pfn
,
1556 free_base_page
= NULL
;
1557 free_base_pfn
= nr_to_free
= 0;
1560 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1563 /* Sanity check that the next zone really is unpopulated */
1564 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1566 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1567 jiffies_to_msecs(jiffies
- start
));
1569 pgdat_init_report_one_done();
1572 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1574 void __init
page_alloc_init_late(void)
1578 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1581 /* There will be num_node_state(N_MEMORY) threads */
1582 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1583 for_each_node_state(nid
, N_MEMORY
) {
1584 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1587 /* Block until all are initialised */
1588 wait_for_completion(&pgdat_init_all_done_comp
);
1590 /* Reinit limits that are based on free pages after the kernel is up */
1591 files_maxfiles_init();
1594 for_each_populated_zone(zone
)
1595 set_zone_contiguous(zone
);
1599 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1600 void __init
init_cma_reserved_pageblock(struct page
*page
)
1602 unsigned i
= pageblock_nr_pages
;
1603 struct page
*p
= page
;
1606 __ClearPageReserved(p
);
1607 set_page_count(p
, 0);
1610 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1612 if (pageblock_order
>= MAX_ORDER
) {
1613 i
= pageblock_nr_pages
;
1616 set_page_refcounted(p
);
1617 __free_pages(p
, MAX_ORDER
- 1);
1618 p
+= MAX_ORDER_NR_PAGES
;
1619 } while (i
-= MAX_ORDER_NR_PAGES
);
1621 set_page_refcounted(page
);
1622 __free_pages(page
, pageblock_order
);
1625 adjust_managed_page_count(page
, pageblock_nr_pages
);
1630 * The order of subdivision here is critical for the IO subsystem.
1631 * Please do not alter this order without good reasons and regression
1632 * testing. Specifically, as large blocks of memory are subdivided,
1633 * the order in which smaller blocks are delivered depends on the order
1634 * they're subdivided in this function. This is the primary factor
1635 * influencing the order in which pages are delivered to the IO
1636 * subsystem according to empirical testing, and this is also justified
1637 * by considering the behavior of a buddy system containing a single
1638 * large block of memory acted on by a series of small allocations.
1639 * This behavior is a critical factor in sglist merging's success.
1643 static inline void expand(struct zone
*zone
, struct page
*page
,
1644 int low
, int high
, struct free_area
*area
,
1647 unsigned long size
= 1 << high
;
1649 while (high
> low
) {
1653 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1655 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
1656 debug_guardpage_enabled() &&
1657 high
< debug_guardpage_minorder()) {
1659 * Mark as guard pages (or page), that will allow to
1660 * merge back to allocator when buddy will be freed.
1661 * Corresponding page table entries will not be touched,
1662 * pages will stay not present in virtual address space
1664 set_page_guard(zone
, &page
[size
], high
, migratetype
);
1667 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1669 set_page_order(&page
[size
], high
);
1674 * This page is about to be returned from the page allocator
1676 static inline int check_new_page(struct page
*page
)
1678 const char *bad_reason
;
1679 unsigned long bad_flags
;
1681 if (page_expected_state(page
, PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
))
1686 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1687 bad_reason
= "nonzero mapcount";
1688 if (unlikely(page
->mapping
!= NULL
))
1689 bad_reason
= "non-NULL mapping";
1690 if (unlikely(page_ref_count(page
) != 0))
1691 bad_reason
= "nonzero _count";
1692 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1693 bad_reason
= "HWPoisoned (hardware-corrupted)";
1694 bad_flags
= __PG_HWPOISON
;
1696 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1697 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1698 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1701 if (unlikely(page
->mem_cgroup
))
1702 bad_reason
= "page still charged to cgroup";
1704 if (unlikely(bad_reason
)) {
1705 bad_page(page
, bad_reason
, bad_flags
);
1711 static inline bool free_pages_prezeroed(bool poisoned
)
1713 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1714 page_poisoning_enabled() && poisoned
;
1717 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1718 unsigned int alloc_flags
)
1721 bool poisoned
= true;
1723 for (i
= 0; i
< (1 << order
); i
++) {
1724 struct page
*p
= page
+ i
;
1725 if (unlikely(check_new_page(p
)))
1728 poisoned
&= page_is_poisoned(p
);
1731 set_page_private(page
, 0);
1732 set_page_refcounted(page
);
1734 arch_alloc_page(page
, order
);
1735 kernel_map_pages(page
, 1 << order
, 1);
1736 kernel_poison_pages(page
, 1 << order
, 1);
1737 kasan_alloc_pages(page
, order
);
1739 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1740 for (i
= 0; i
< (1 << order
); i
++)
1741 clear_highpage(page
+ i
);
1743 if (order
&& (gfp_flags
& __GFP_COMP
))
1744 prep_compound_page(page
, order
);
1746 set_page_owner(page
, order
, gfp_flags
);
1749 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1750 * allocate the page. The expectation is that the caller is taking
1751 * steps that will free more memory. The caller should avoid the page
1752 * being used for !PFMEMALLOC purposes.
1754 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1755 set_page_pfmemalloc(page
);
1757 clear_page_pfmemalloc(page
);
1763 * Go through the free lists for the given migratetype and remove
1764 * the smallest available page from the freelists
1767 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1770 unsigned int current_order
;
1771 struct free_area
*area
;
1774 /* Find a page of the appropriate size in the preferred list */
1775 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1776 area
= &(zone
->free_area
[current_order
]);
1777 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1781 list_del(&page
->lru
);
1782 rmv_page_order(page
);
1784 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1785 set_pcppage_migratetype(page
, migratetype
);
1794 * This array describes the order lists are fallen back to when
1795 * the free lists for the desirable migrate type are depleted
1797 static int fallbacks
[MIGRATE_TYPES
][4] = {
1798 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1799 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1800 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1802 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1804 #ifdef CONFIG_MEMORY_ISOLATION
1805 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1810 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1813 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1816 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1817 unsigned int order
) { return NULL
; }
1821 * Move the free pages in a range to the free lists of the requested type.
1822 * Note that start_page and end_pages are not aligned on a pageblock
1823 * boundary. If alignment is required, use move_freepages_block()
1825 int move_freepages(struct zone
*zone
,
1826 struct page
*start_page
, struct page
*end_page
,
1831 int pages_moved
= 0;
1833 #ifndef CONFIG_HOLES_IN_ZONE
1835 * page_zone is not safe to call in this context when
1836 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1837 * anyway as we check zone boundaries in move_freepages_block().
1838 * Remove at a later date when no bug reports exist related to
1839 * grouping pages by mobility
1841 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1844 for (page
= start_page
; page
<= end_page
;) {
1845 /* Make sure we are not inadvertently changing nodes */
1846 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1848 if (!pfn_valid_within(page_to_pfn(page
))) {
1853 if (!PageBuddy(page
)) {
1858 order
= page_order(page
);
1859 list_move(&page
->lru
,
1860 &zone
->free_area
[order
].free_list
[migratetype
]);
1862 pages_moved
+= 1 << order
;
1868 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1871 unsigned long start_pfn
, end_pfn
;
1872 struct page
*start_page
, *end_page
;
1874 start_pfn
= page_to_pfn(page
);
1875 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1876 start_page
= pfn_to_page(start_pfn
);
1877 end_page
= start_page
+ pageblock_nr_pages
- 1;
1878 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1880 /* Do not cross zone boundaries */
1881 if (!zone_spans_pfn(zone
, start_pfn
))
1883 if (!zone_spans_pfn(zone
, end_pfn
))
1886 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1889 static void change_pageblock_range(struct page
*pageblock_page
,
1890 int start_order
, int migratetype
)
1892 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1894 while (nr_pageblocks
--) {
1895 set_pageblock_migratetype(pageblock_page
, migratetype
);
1896 pageblock_page
+= pageblock_nr_pages
;
1901 * When we are falling back to another migratetype during allocation, try to
1902 * steal extra free pages from the same pageblocks to satisfy further
1903 * allocations, instead of polluting multiple pageblocks.
1905 * If we are stealing a relatively large buddy page, it is likely there will
1906 * be more free pages in the pageblock, so try to steal them all. For
1907 * reclaimable and unmovable allocations, we steal regardless of page size,
1908 * as fragmentation caused by those allocations polluting movable pageblocks
1909 * is worse than movable allocations stealing from unmovable and reclaimable
1912 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1915 * Leaving this order check is intended, although there is
1916 * relaxed order check in next check. The reason is that
1917 * we can actually steal whole pageblock if this condition met,
1918 * but, below check doesn't guarantee it and that is just heuristic
1919 * so could be changed anytime.
1921 if (order
>= pageblock_order
)
1924 if (order
>= pageblock_order
/ 2 ||
1925 start_mt
== MIGRATE_RECLAIMABLE
||
1926 start_mt
== MIGRATE_UNMOVABLE
||
1927 page_group_by_mobility_disabled
)
1934 * This function implements actual steal behaviour. If order is large enough,
1935 * we can steal whole pageblock. If not, we first move freepages in this
1936 * pageblock and check whether half of pages are moved or not. If half of
1937 * pages are moved, we can change migratetype of pageblock and permanently
1938 * use it's pages as requested migratetype in the future.
1940 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1943 unsigned int current_order
= page_order(page
);
1946 /* Take ownership for orders >= pageblock_order */
1947 if (current_order
>= pageblock_order
) {
1948 change_pageblock_range(page
, current_order
, start_type
);
1952 pages
= move_freepages_block(zone
, page
, start_type
);
1954 /* Claim the whole block if over half of it is free */
1955 if (pages
>= (1 << (pageblock_order
-1)) ||
1956 page_group_by_mobility_disabled
)
1957 set_pageblock_migratetype(page
, start_type
);
1961 * Check whether there is a suitable fallback freepage with requested order.
1962 * If only_stealable is true, this function returns fallback_mt only if
1963 * we can steal other freepages all together. This would help to reduce
1964 * fragmentation due to mixed migratetype pages in one pageblock.
1966 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1967 int migratetype
, bool only_stealable
, bool *can_steal
)
1972 if (area
->nr_free
== 0)
1977 fallback_mt
= fallbacks
[migratetype
][i
];
1978 if (fallback_mt
== MIGRATE_TYPES
)
1981 if (list_empty(&area
->free_list
[fallback_mt
]))
1984 if (can_steal_fallback(order
, migratetype
))
1987 if (!only_stealable
)
1998 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1999 * there are no empty page blocks that contain a page with a suitable order
2001 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2002 unsigned int alloc_order
)
2005 unsigned long max_managed
, flags
;
2008 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2009 * Check is race-prone but harmless.
2011 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2012 if (zone
->nr_reserved_highatomic
>= max_managed
)
2015 spin_lock_irqsave(&zone
->lock
, flags
);
2017 /* Recheck the nr_reserved_highatomic limit under the lock */
2018 if (zone
->nr_reserved_highatomic
>= max_managed
)
2022 mt
= get_pageblock_migratetype(page
);
2023 if (mt
!= MIGRATE_HIGHATOMIC
&&
2024 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
2025 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2026 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2027 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
2031 spin_unlock_irqrestore(&zone
->lock
, flags
);
2035 * Used when an allocation is about to fail under memory pressure. This
2036 * potentially hurts the reliability of high-order allocations when under
2037 * intense memory pressure but failed atomic allocations should be easier
2038 * to recover from than an OOM.
2040 static void unreserve_highatomic_pageblock(const struct alloc_context
*ac
)
2042 struct zonelist
*zonelist
= ac
->zonelist
;
2043 unsigned long flags
;
2049 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2051 /* Preserve at least one pageblock */
2052 if (zone
->nr_reserved_highatomic
<= pageblock_nr_pages
)
2055 spin_lock_irqsave(&zone
->lock
, flags
);
2056 for (order
= 0; order
< MAX_ORDER
; order
++) {
2057 struct free_area
*area
= &(zone
->free_area
[order
]);
2059 page
= list_first_entry_or_null(
2060 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2066 * It should never happen but changes to locking could
2067 * inadvertently allow a per-cpu drain to add pages
2068 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2069 * and watch for underflows.
2071 zone
->nr_reserved_highatomic
-= min(pageblock_nr_pages
,
2072 zone
->nr_reserved_highatomic
);
2075 * Convert to ac->migratetype and avoid the normal
2076 * pageblock stealing heuristics. Minimally, the caller
2077 * is doing the work and needs the pages. More
2078 * importantly, if the block was always converted to
2079 * MIGRATE_UNMOVABLE or another type then the number
2080 * of pageblocks that cannot be completely freed
2083 set_pageblock_migratetype(page
, ac
->migratetype
);
2084 move_freepages_block(zone
, page
, ac
->migratetype
);
2085 spin_unlock_irqrestore(&zone
->lock
, flags
);
2088 spin_unlock_irqrestore(&zone
->lock
, flags
);
2092 /* Remove an element from the buddy allocator from the fallback list */
2093 static inline struct page
*
2094 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2096 struct free_area
*area
;
2097 unsigned int current_order
;
2102 /* Find the largest possible block of pages in the other list */
2103 for (current_order
= MAX_ORDER
-1;
2104 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2106 area
= &(zone
->free_area
[current_order
]);
2107 fallback_mt
= find_suitable_fallback(area
, current_order
,
2108 start_migratetype
, false, &can_steal
);
2109 if (fallback_mt
== -1)
2112 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2115 steal_suitable_fallback(zone
, page
, start_migratetype
);
2117 /* Remove the page from the freelists */
2119 list_del(&page
->lru
);
2120 rmv_page_order(page
);
2122 expand(zone
, page
, order
, current_order
, area
,
2125 * The pcppage_migratetype may differ from pageblock's
2126 * migratetype depending on the decisions in
2127 * find_suitable_fallback(). This is OK as long as it does not
2128 * differ for MIGRATE_CMA pageblocks. Those can be used as
2129 * fallback only via special __rmqueue_cma_fallback() function
2131 set_pcppage_migratetype(page
, start_migratetype
);
2133 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2134 start_migratetype
, fallback_mt
);
2143 * Do the hard work of removing an element from the buddy allocator.
2144 * Call me with the zone->lock already held.
2146 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2151 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2152 if (unlikely(!page
)) {
2153 if (migratetype
== MIGRATE_MOVABLE
)
2154 page
= __rmqueue_cma_fallback(zone
, order
);
2157 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2160 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2165 * Obtain a specified number of elements from the buddy allocator, all under
2166 * a single hold of the lock, for efficiency. Add them to the supplied list.
2167 * Returns the number of new pages which were placed at *list.
2169 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2170 unsigned long count
, struct list_head
*list
,
2171 int migratetype
, bool cold
)
2175 spin_lock(&zone
->lock
);
2176 for (i
= 0; i
< count
; ++i
) {
2177 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2178 if (unlikely(page
== NULL
))
2182 * Split buddy pages returned by expand() are received here
2183 * in physical page order. The page is added to the callers and
2184 * list and the list head then moves forward. From the callers
2185 * perspective, the linked list is ordered by page number in
2186 * some conditions. This is useful for IO devices that can
2187 * merge IO requests if the physical pages are ordered
2191 list_add(&page
->lru
, list
);
2193 list_add_tail(&page
->lru
, list
);
2195 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2196 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2199 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2200 spin_unlock(&zone
->lock
);
2206 * Called from the vmstat counter updater to drain pagesets of this
2207 * currently executing processor on remote nodes after they have
2210 * Note that this function must be called with the thread pinned to
2211 * a single processor.
2213 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2215 unsigned long flags
;
2216 int to_drain
, batch
;
2218 local_irq_save(flags
);
2219 batch
= READ_ONCE(pcp
->batch
);
2220 to_drain
= min(pcp
->count
, batch
);
2222 free_pcppages_bulk(zone
, to_drain
, pcp
);
2223 pcp
->count
-= to_drain
;
2225 local_irq_restore(flags
);
2230 * Drain pcplists of the indicated processor and zone.
2232 * The processor must either be the current processor and the
2233 * thread pinned to the current processor or a processor that
2236 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2238 unsigned long flags
;
2239 struct per_cpu_pageset
*pset
;
2240 struct per_cpu_pages
*pcp
;
2242 local_irq_save(flags
);
2243 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2247 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2250 local_irq_restore(flags
);
2254 * Drain pcplists of all zones on the indicated processor.
2256 * The processor must either be the current processor and the
2257 * thread pinned to the current processor or a processor that
2260 static void drain_pages(unsigned int cpu
)
2264 for_each_populated_zone(zone
) {
2265 drain_pages_zone(cpu
, zone
);
2270 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2272 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2273 * the single zone's pages.
2275 void drain_local_pages(struct zone
*zone
)
2277 int cpu
= smp_processor_id();
2280 drain_pages_zone(cpu
, zone
);
2286 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2288 * When zone parameter is non-NULL, spill just the single zone's pages.
2290 * Note that this code is protected against sending an IPI to an offline
2291 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2292 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2293 * nothing keeps CPUs from showing up after we populated the cpumask and
2294 * before the call to on_each_cpu_mask().
2296 void drain_all_pages(struct zone
*zone
)
2301 * Allocate in the BSS so we wont require allocation in
2302 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2304 static cpumask_t cpus_with_pcps
;
2307 * We don't care about racing with CPU hotplug event
2308 * as offline notification will cause the notified
2309 * cpu to drain that CPU pcps and on_each_cpu_mask
2310 * disables preemption as part of its processing
2312 for_each_online_cpu(cpu
) {
2313 struct per_cpu_pageset
*pcp
;
2315 bool has_pcps
= false;
2318 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2322 for_each_populated_zone(z
) {
2323 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2324 if (pcp
->pcp
.count
) {
2332 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2334 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2336 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
2340 #ifdef CONFIG_HIBERNATION
2342 void mark_free_pages(struct zone
*zone
)
2344 unsigned long pfn
, max_zone_pfn
;
2345 unsigned long flags
;
2346 unsigned int order
, t
;
2349 if (zone_is_empty(zone
))
2352 spin_lock_irqsave(&zone
->lock
, flags
);
2354 max_zone_pfn
= zone_end_pfn(zone
);
2355 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2356 if (pfn_valid(pfn
)) {
2357 page
= pfn_to_page(pfn
);
2359 if (page_zone(page
) != zone
)
2362 if (!swsusp_page_is_forbidden(page
))
2363 swsusp_unset_page_free(page
);
2366 for_each_migratetype_order(order
, t
) {
2367 list_for_each_entry(page
,
2368 &zone
->free_area
[order
].free_list
[t
], lru
) {
2371 pfn
= page_to_pfn(page
);
2372 for (i
= 0; i
< (1UL << order
); i
++)
2373 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2376 spin_unlock_irqrestore(&zone
->lock
, flags
);
2378 #endif /* CONFIG_PM */
2381 * Free a 0-order page
2382 * cold == true ? free a cold page : free a hot page
2384 void free_hot_cold_page(struct page
*page
, bool cold
)
2386 struct zone
*zone
= page_zone(page
);
2387 struct per_cpu_pages
*pcp
;
2388 unsigned long flags
;
2389 unsigned long pfn
= page_to_pfn(page
);
2392 if (!free_pcp_prepare(page
))
2395 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2396 set_pcppage_migratetype(page
, migratetype
);
2397 local_irq_save(flags
);
2398 __count_vm_event(PGFREE
);
2401 * We only track unmovable, reclaimable and movable on pcp lists.
2402 * Free ISOLATE pages back to the allocator because they are being
2403 * offlined but treat RESERVE as movable pages so we can get those
2404 * areas back if necessary. Otherwise, we may have to free
2405 * excessively into the page allocator
2407 if (migratetype
>= MIGRATE_PCPTYPES
) {
2408 if (unlikely(is_migrate_isolate(migratetype
))) {
2409 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2412 migratetype
= MIGRATE_MOVABLE
;
2415 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2417 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2419 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2421 if (pcp
->count
>= pcp
->high
) {
2422 unsigned long batch
= READ_ONCE(pcp
->batch
);
2423 free_pcppages_bulk(zone
, batch
, pcp
);
2424 pcp
->count
-= batch
;
2428 local_irq_restore(flags
);
2432 * Free a list of 0-order pages
2434 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2436 struct page
*page
, *next
;
2438 list_for_each_entry_safe(page
, next
, list
, lru
) {
2439 trace_mm_page_free_batched(page
, cold
);
2440 free_hot_cold_page(page
, cold
);
2445 * split_page takes a non-compound higher-order page, and splits it into
2446 * n (1<<order) sub-pages: page[0..n]
2447 * Each sub-page must be freed individually.
2449 * Note: this is probably too low level an operation for use in drivers.
2450 * Please consult with lkml before using this in your driver.
2452 void split_page(struct page
*page
, unsigned int order
)
2457 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2458 VM_BUG_ON_PAGE(!page_count(page
), page
);
2460 #ifdef CONFIG_KMEMCHECK
2462 * Split shadow pages too, because free(page[0]) would
2463 * otherwise free the whole shadow.
2465 if (kmemcheck_page_is_tracked(page
))
2466 split_page(virt_to_page(page
[0].shadow
), order
);
2469 gfp_mask
= get_page_owner_gfp(page
);
2470 set_page_owner(page
, 0, gfp_mask
);
2471 for (i
= 1; i
< (1 << order
); i
++) {
2472 set_page_refcounted(page
+ i
);
2473 set_page_owner(page
+ i
, 0, gfp_mask
);
2476 EXPORT_SYMBOL_GPL(split_page
);
2478 int __isolate_free_page(struct page
*page
, unsigned int order
)
2480 unsigned long watermark
;
2484 BUG_ON(!PageBuddy(page
));
2486 zone
= page_zone(page
);
2487 mt
= get_pageblock_migratetype(page
);
2489 if (!is_migrate_isolate(mt
)) {
2490 /* Obey watermarks as if the page was being allocated */
2491 watermark
= low_wmark_pages(zone
) + (1 << order
);
2492 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
2495 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2498 /* Remove page from free list */
2499 list_del(&page
->lru
);
2500 zone
->free_area
[order
].nr_free
--;
2501 rmv_page_order(page
);
2503 set_page_owner(page
, order
, __GFP_MOVABLE
);
2505 /* Set the pageblock if the isolated page is at least a pageblock */
2506 if (order
>= pageblock_order
- 1) {
2507 struct page
*endpage
= page
+ (1 << order
) - 1;
2508 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2509 int mt
= get_pageblock_migratetype(page
);
2510 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2511 set_pageblock_migratetype(page
,
2517 return 1UL << order
;
2521 * Similar to split_page except the page is already free. As this is only
2522 * being used for migration, the migratetype of the block also changes.
2523 * As this is called with interrupts disabled, the caller is responsible
2524 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2527 * Note: this is probably too low level an operation for use in drivers.
2528 * Please consult with lkml before using this in your driver.
2530 int split_free_page(struct page
*page
)
2535 order
= page_order(page
);
2537 nr_pages
= __isolate_free_page(page
, order
);
2541 /* Split into individual pages */
2542 set_page_refcounted(page
);
2543 split_page(page
, order
);
2548 * Update NUMA hit/miss statistics
2550 * Must be called with interrupts disabled.
2552 * When __GFP_OTHER_NODE is set assume the node of the preferred
2553 * zone is the local node. This is useful for daemons who allocate
2554 * memory on behalf of other processes.
2556 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2560 int local_nid
= numa_node_id();
2561 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2563 if (unlikely(flags
& __GFP_OTHER_NODE
)) {
2564 local_stat
= NUMA_OTHER
;
2565 local_nid
= preferred_zone
->node
;
2568 if (z
->node
== local_nid
) {
2569 __inc_zone_state(z
, NUMA_HIT
);
2570 __inc_zone_state(z
, local_stat
);
2572 __inc_zone_state(z
, NUMA_MISS
);
2573 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2579 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2582 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2583 struct zone
*zone
, unsigned int order
,
2584 gfp_t gfp_flags
, unsigned int alloc_flags
,
2587 unsigned long flags
;
2589 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2591 if (likely(order
== 0)) {
2592 struct per_cpu_pages
*pcp
;
2593 struct list_head
*list
;
2595 local_irq_save(flags
);
2596 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2597 list
= &pcp
->lists
[migratetype
];
2598 if (list_empty(list
)) {
2599 pcp
->count
+= rmqueue_bulk(zone
, 0,
2602 if (unlikely(list_empty(list
)))
2607 page
= list_last_entry(list
, struct page
, lru
);
2609 page
= list_first_entry(list
, struct page
, lru
);
2611 __dec_zone_state(zone
, NR_ALLOC_BATCH
);
2612 list_del(&page
->lru
);
2616 * We most definitely don't want callers attempting to
2617 * allocate greater than order-1 page units with __GFP_NOFAIL.
2619 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2620 spin_lock_irqsave(&zone
->lock
, flags
);
2623 if (alloc_flags
& ALLOC_HARDER
) {
2624 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2626 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2629 page
= __rmqueue(zone
, order
, migratetype
);
2630 spin_unlock(&zone
->lock
);
2633 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
2634 __mod_zone_freepage_state(zone
, -(1 << order
),
2635 get_pcppage_migratetype(page
));
2638 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
2639 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
2640 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2642 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
2643 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2644 local_irq_restore(flags
);
2646 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2650 local_irq_restore(flags
);
2654 #ifdef CONFIG_FAIL_PAGE_ALLOC
2657 struct fault_attr attr
;
2659 bool ignore_gfp_highmem
;
2660 bool ignore_gfp_reclaim
;
2662 } fail_page_alloc
= {
2663 .attr
= FAULT_ATTR_INITIALIZER
,
2664 .ignore_gfp_reclaim
= true,
2665 .ignore_gfp_highmem
= true,
2669 static int __init
setup_fail_page_alloc(char *str
)
2671 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2673 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2675 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2677 if (order
< fail_page_alloc
.min_order
)
2679 if (gfp_mask
& __GFP_NOFAIL
)
2681 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2683 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2684 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2687 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2690 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2692 static int __init
fail_page_alloc_debugfs(void)
2694 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2697 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2698 &fail_page_alloc
.attr
);
2700 return PTR_ERR(dir
);
2702 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2703 &fail_page_alloc
.ignore_gfp_reclaim
))
2705 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2706 &fail_page_alloc
.ignore_gfp_highmem
))
2708 if (!debugfs_create_u32("min-order", mode
, dir
,
2709 &fail_page_alloc
.min_order
))
2714 debugfs_remove_recursive(dir
);
2719 late_initcall(fail_page_alloc_debugfs
);
2721 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2723 #else /* CONFIG_FAIL_PAGE_ALLOC */
2725 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2730 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2733 * Return true if free base pages are above 'mark'. For high-order checks it
2734 * will return true of the order-0 watermark is reached and there is at least
2735 * one free page of a suitable size. Checking now avoids taking the zone lock
2736 * to check in the allocation paths if no pages are free.
2738 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
2739 unsigned long mark
, int classzone_idx
,
2740 unsigned int alloc_flags
,
2745 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2747 /* free_pages may go negative - that's OK */
2748 free_pages
-= (1 << order
) - 1;
2750 if (alloc_flags
& ALLOC_HIGH
)
2754 * If the caller does not have rights to ALLOC_HARDER then subtract
2755 * the high-atomic reserves. This will over-estimate the size of the
2756 * atomic reserve but it avoids a search.
2758 if (likely(!alloc_harder
))
2759 free_pages
-= z
->nr_reserved_highatomic
;
2764 /* If allocation can't use CMA areas don't use free CMA pages */
2765 if (!(alloc_flags
& ALLOC_CMA
))
2766 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2770 * Check watermarks for an order-0 allocation request. If these
2771 * are not met, then a high-order request also cannot go ahead
2772 * even if a suitable page happened to be free.
2774 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2777 /* If this is an order-0 request then the watermark is fine */
2781 /* For a high-order request, check at least one suitable page is free */
2782 for (o
= order
; o
< MAX_ORDER
; o
++) {
2783 struct free_area
*area
= &z
->free_area
[o
];
2792 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2793 if (!list_empty(&area
->free_list
[mt
]))
2798 if ((alloc_flags
& ALLOC_CMA
) &&
2799 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2807 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2808 int classzone_idx
, unsigned int alloc_flags
)
2810 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2811 zone_page_state(z
, NR_FREE_PAGES
));
2814 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2815 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2817 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2821 /* If allocation can't use CMA areas don't use free CMA pages */
2822 if (!(alloc_flags
& ALLOC_CMA
))
2823 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2827 * Fast check for order-0 only. If this fails then the reserves
2828 * need to be calculated. There is a corner case where the check
2829 * passes but only the high-order atomic reserve are free. If
2830 * the caller is !atomic then it'll uselessly search the free
2831 * list. That corner case is then slower but it is harmless.
2833 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2836 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2840 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2841 unsigned long mark
, int classzone_idx
)
2843 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2845 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2846 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2848 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2853 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2855 return local_zone
->node
== zone
->node
;
2858 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2860 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2863 #else /* CONFIG_NUMA */
2864 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2869 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2873 #endif /* CONFIG_NUMA */
2875 static void reset_alloc_batches(struct zone
*preferred_zone
)
2877 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2880 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2881 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2882 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2883 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2884 } while (zone
++ != preferred_zone
);
2888 * get_page_from_freelist goes through the zonelist trying to allocate
2891 static struct page
*
2892 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2893 const struct alloc_context
*ac
)
2895 struct zoneref
*z
= ac
->preferred_zoneref
;
2897 bool fair_skipped
= false;
2898 bool apply_fair
= (alloc_flags
& ALLOC_FAIR
);
2902 * Scan zonelist, looking for a zone with enough free.
2903 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2905 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2910 if (cpusets_enabled() &&
2911 (alloc_flags
& ALLOC_CPUSET
) &&
2912 !__cpuset_zone_allowed(zone
, gfp_mask
))
2915 * Distribute pages in proportion to the individual
2916 * zone size to ensure fair page aging. The zone a
2917 * page was allocated in should have no effect on the
2918 * time the page has in memory before being reclaimed.
2921 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2922 fair_skipped
= true;
2925 if (!zone_local(ac
->preferred_zoneref
->zone
, zone
)) {
2932 * When allocating a page cache page for writing, we
2933 * want to get it from a zone that is within its dirty
2934 * limit, such that no single zone holds more than its
2935 * proportional share of globally allowed dirty pages.
2936 * The dirty limits take into account the zone's
2937 * lowmem reserves and high watermark so that kswapd
2938 * should be able to balance it without having to
2939 * write pages from its LRU list.
2941 * This may look like it could increase pressure on
2942 * lower zones by failing allocations in higher zones
2943 * before they are full. But the pages that do spill
2944 * over are limited as the lower zones are protected
2945 * by this very same mechanism. It should not become
2946 * a practical burden to them.
2948 * XXX: For now, allow allocations to potentially
2949 * exceed the per-zone dirty limit in the slowpath
2950 * (spread_dirty_pages unset) before going into reclaim,
2951 * which is important when on a NUMA setup the allowed
2952 * zones are together not big enough to reach the
2953 * global limit. The proper fix for these situations
2954 * will require awareness of zones in the
2955 * dirty-throttling and the flusher threads.
2957 if (ac
->spread_dirty_pages
&& !zone_dirty_ok(zone
))
2960 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2961 if (!zone_watermark_fast(zone
, order
, mark
,
2962 ac_classzone_idx(ac
), alloc_flags
)) {
2965 /* Checked here to keep the fast path fast */
2966 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2967 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2970 if (zone_reclaim_mode
== 0 ||
2971 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
2974 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2976 case ZONE_RECLAIM_NOSCAN
:
2979 case ZONE_RECLAIM_FULL
:
2980 /* scanned but unreclaimable */
2983 /* did we reclaim enough */
2984 if (zone_watermark_ok(zone
, order
, mark
,
2985 ac_classzone_idx(ac
), alloc_flags
))
2993 page
= buffered_rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
2994 gfp_mask
, alloc_flags
, ac
->migratetype
);
2996 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
3000 * If this is a high-order atomic allocation then check
3001 * if the pageblock should be reserved for the future
3003 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3004 reserve_highatomic_pageblock(page
, zone
, order
);
3011 * The first pass makes sure allocations are spread fairly within the
3012 * local node. However, the local node might have free pages left
3013 * after the fairness batches are exhausted, and remote zones haven't
3014 * even been considered yet. Try once more without fairness, and
3015 * include remote zones now, before entering the slowpath and waking
3016 * kswapd: prefer spilling to a remote zone over swapping locally.
3021 fair_skipped
= false;
3022 reset_alloc_batches(ac
->preferred_zoneref
->zone
);
3030 * Large machines with many possible nodes should not always dump per-node
3031 * meminfo in irq context.
3033 static inline bool should_suppress_show_mem(void)
3038 ret
= in_interrupt();
3043 static DEFINE_RATELIMIT_STATE(nopage_rs
,
3044 DEFAULT_RATELIMIT_INTERVAL
,
3045 DEFAULT_RATELIMIT_BURST
);
3047 void warn_alloc_failed(gfp_t gfp_mask
, unsigned int order
, const char *fmt
, ...)
3049 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3051 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
3052 debug_guardpage_minorder() > 0)
3056 * This documents exceptions given to allocations in certain
3057 * contexts that are allowed to allocate outside current's set
3060 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3061 if (test_thread_flag(TIF_MEMDIE
) ||
3062 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3063 filter
&= ~SHOW_MEM_FILTER_NODES
;
3064 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3065 filter
&= ~SHOW_MEM_FILTER_NODES
;
3068 struct va_format vaf
;
3071 va_start(args
, fmt
);
3076 pr_warn("%pV", &vaf
);
3081 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3082 current
->comm
, order
, gfp_mask
, &gfp_mask
);
3084 if (!should_suppress_show_mem())
3088 static inline struct page
*
3089 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3090 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3092 struct oom_control oc
= {
3093 .zonelist
= ac
->zonelist
,
3094 .nodemask
= ac
->nodemask
,
3095 .gfp_mask
= gfp_mask
,
3100 *did_some_progress
= 0;
3103 * Acquire the oom lock. If that fails, somebody else is
3104 * making progress for us.
3106 if (!mutex_trylock(&oom_lock
)) {
3107 *did_some_progress
= 1;
3108 schedule_timeout_uninterruptible(1);
3113 * Go through the zonelist yet one more time, keep very high watermark
3114 * here, this is only to catch a parallel oom killing, we must fail if
3115 * we're still under heavy pressure.
3117 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3118 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3122 if (!(gfp_mask
& __GFP_NOFAIL
)) {
3123 /* Coredumps can quickly deplete all memory reserves */
3124 if (current
->flags
& PF_DUMPCORE
)
3126 /* The OOM killer will not help higher order allocs */
3127 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3129 /* The OOM killer does not needlessly kill tasks for lowmem */
3130 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3132 if (pm_suspended_storage())
3135 * XXX: GFP_NOFS allocations should rather fail than rely on
3136 * other request to make a forward progress.
3137 * We are in an unfortunate situation where out_of_memory cannot
3138 * do much for this context but let's try it to at least get
3139 * access to memory reserved if the current task is killed (see
3140 * out_of_memory). Once filesystems are ready to handle allocation
3141 * failures more gracefully we should just bail out here.
3144 /* The OOM killer may not free memory on a specific node */
3145 if (gfp_mask
& __GFP_THISNODE
)
3148 /* Exhausted what can be done so it's blamo time */
3149 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3150 *did_some_progress
= 1;
3152 if (gfp_mask
& __GFP_NOFAIL
) {
3153 page
= get_page_from_freelist(gfp_mask
, order
,
3154 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
3156 * fallback to ignore cpuset restriction if our nodes
3160 page
= get_page_from_freelist(gfp_mask
, order
,
3161 ALLOC_NO_WATERMARKS
, ac
);
3165 mutex_unlock(&oom_lock
);
3169 #ifdef CONFIG_COMPACTION
3170 /* Try memory compaction for high-order allocations before reclaim */
3171 static struct page
*
3172 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3173 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3174 enum migrate_mode mode
, int *contended_compaction
,
3175 bool *deferred_compaction
)
3177 unsigned long compact_result
;
3183 current
->flags
|= PF_MEMALLOC
;
3184 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3185 mode
, contended_compaction
);
3186 current
->flags
&= ~PF_MEMALLOC
;
3188 switch (compact_result
) {
3189 case COMPACT_DEFERRED
:
3190 *deferred_compaction
= true;
3192 case COMPACT_SKIPPED
:
3199 * At least in one zone compaction wasn't deferred or skipped, so let's
3200 * count a compaction stall
3202 count_vm_event(COMPACTSTALL
);
3204 page
= get_page_from_freelist(gfp_mask
, order
,
3205 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3208 struct zone
*zone
= page_zone(page
);
3210 zone
->compact_blockskip_flush
= false;
3211 compaction_defer_reset(zone
, order
, true);
3212 count_vm_event(COMPACTSUCCESS
);
3217 * It's bad if compaction run occurs and fails. The most likely reason
3218 * is that pages exist, but not enough to satisfy watermarks.
3220 count_vm_event(COMPACTFAIL
);
3227 static inline struct page
*
3228 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3229 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3230 enum migrate_mode mode
, int *contended_compaction
,
3231 bool *deferred_compaction
)
3235 #endif /* CONFIG_COMPACTION */
3237 /* Perform direct synchronous page reclaim */
3239 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3240 const struct alloc_context
*ac
)
3242 struct reclaim_state reclaim_state
;
3247 /* We now go into synchronous reclaim */
3248 cpuset_memory_pressure_bump();
3249 current
->flags
|= PF_MEMALLOC
;
3250 lockdep_set_current_reclaim_state(gfp_mask
);
3251 reclaim_state
.reclaimed_slab
= 0;
3252 current
->reclaim_state
= &reclaim_state
;
3254 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3257 current
->reclaim_state
= NULL
;
3258 lockdep_clear_current_reclaim_state();
3259 current
->flags
&= ~PF_MEMALLOC
;
3266 /* The really slow allocator path where we enter direct reclaim */
3267 static inline struct page
*
3268 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3269 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3270 unsigned long *did_some_progress
)
3272 struct page
*page
= NULL
;
3273 bool drained
= false;
3275 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3276 if (unlikely(!(*did_some_progress
)))
3280 page
= get_page_from_freelist(gfp_mask
, order
,
3281 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3284 * If an allocation failed after direct reclaim, it could be because
3285 * pages are pinned on the per-cpu lists or in high alloc reserves.
3286 * Shrink them them and try again
3288 if (!page
&& !drained
) {
3289 unreserve_highatomic_pageblock(ac
);
3290 drain_all_pages(NULL
);
3298 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3303 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3304 ac
->high_zoneidx
, ac
->nodemask
)
3305 wakeup_kswapd(zone
, order
, ac_classzone_idx(ac
));
3308 static inline unsigned int
3309 gfp_to_alloc_flags(gfp_t gfp_mask
)
3311 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3313 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3314 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3317 * The caller may dip into page reserves a bit more if the caller
3318 * cannot run direct reclaim, or if the caller has realtime scheduling
3319 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3320 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3322 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3324 if (gfp_mask
& __GFP_ATOMIC
) {
3326 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3327 * if it can't schedule.
3329 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3330 alloc_flags
|= ALLOC_HARDER
;
3332 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3333 * comment for __cpuset_node_allowed().
3335 alloc_flags
&= ~ALLOC_CPUSET
;
3336 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3337 alloc_flags
|= ALLOC_HARDER
;
3339 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
3340 if (gfp_mask
& __GFP_MEMALLOC
)
3341 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3342 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3343 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3344 else if (!in_interrupt() &&
3345 ((current
->flags
& PF_MEMALLOC
) ||
3346 unlikely(test_thread_flag(TIF_MEMDIE
))))
3347 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3350 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3351 alloc_flags
|= ALLOC_CMA
;
3356 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3358 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
3361 static inline bool is_thp_gfp_mask(gfp_t gfp_mask
)
3363 return (gfp_mask
& (GFP_TRANSHUGE
| __GFP_KSWAPD_RECLAIM
)) == GFP_TRANSHUGE
;
3366 static inline struct page
*
3367 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3368 struct alloc_context
*ac
)
3370 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3371 struct page
*page
= NULL
;
3372 unsigned int alloc_flags
;
3373 unsigned long pages_reclaimed
= 0;
3374 unsigned long did_some_progress
;
3375 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
3376 bool deferred_compaction
= false;
3377 int contended_compaction
= COMPACT_CONTENDED_NONE
;
3380 * In the slowpath, we sanity check order to avoid ever trying to
3381 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3382 * be using allocators in order of preference for an area that is
3385 if (order
>= MAX_ORDER
) {
3386 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3391 * We also sanity check to catch abuse of atomic reserves being used by
3392 * callers that are not in atomic context.
3394 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3395 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3396 gfp_mask
&= ~__GFP_ATOMIC
;
3399 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3400 wake_all_kswapds(order
, ac
);
3403 * OK, we're below the kswapd watermark and have kicked background
3404 * reclaim. Now things get more complex, so set up alloc_flags according
3405 * to how we want to proceed.
3407 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3409 /* This is the last chance, in general, before the goto nopage. */
3410 page
= get_page_from_freelist(gfp_mask
, order
,
3411 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3415 /* Allocate without watermarks if the context allows */
3416 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
3418 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3419 * the allocation is high priority and these type of
3420 * allocations are system rather than user orientated
3422 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3423 page
= get_page_from_freelist(gfp_mask
, order
,
3424 ALLOC_NO_WATERMARKS
, ac
);
3429 /* Caller is not willing to reclaim, we can't balance anything */
3430 if (!can_direct_reclaim
) {
3432 * All existing users of the __GFP_NOFAIL are blockable, so warn
3433 * of any new users that actually allow this type of allocation
3436 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3440 /* Avoid recursion of direct reclaim */
3441 if (current
->flags
& PF_MEMALLOC
) {
3443 * __GFP_NOFAIL request from this context is rather bizarre
3444 * because we cannot reclaim anything and only can loop waiting
3445 * for somebody to do a work for us.
3447 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3454 /* Avoid allocations with no watermarks from looping endlessly */
3455 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3459 * Try direct compaction. The first pass is asynchronous. Subsequent
3460 * attempts after direct reclaim are synchronous
3462 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3464 &contended_compaction
,
3465 &deferred_compaction
);
3469 /* Checks for THP-specific high-order allocations */
3470 if (is_thp_gfp_mask(gfp_mask
)) {
3472 * If compaction is deferred for high-order allocations, it is
3473 * because sync compaction recently failed. If this is the case
3474 * and the caller requested a THP allocation, we do not want
3475 * to heavily disrupt the system, so we fail the allocation
3476 * instead of entering direct reclaim.
3478 if (deferred_compaction
)
3482 * In all zones where compaction was attempted (and not
3483 * deferred or skipped), lock contention has been detected.
3484 * For THP allocation we do not want to disrupt the others
3485 * so we fallback to base pages instead.
3487 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
3491 * If compaction was aborted due to need_resched(), we do not
3492 * want to further increase allocation latency, unless it is
3493 * khugepaged trying to collapse.
3495 if (contended_compaction
== COMPACT_CONTENDED_SCHED
3496 && !(current
->flags
& PF_KTHREAD
))
3501 * It can become very expensive to allocate transparent hugepages at
3502 * fault, so use asynchronous memory compaction for THP unless it is
3503 * khugepaged trying to collapse.
3505 if (!is_thp_gfp_mask(gfp_mask
) || (current
->flags
& PF_KTHREAD
))
3506 migration_mode
= MIGRATE_SYNC_LIGHT
;
3508 /* Try direct reclaim and then allocating */
3509 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3510 &did_some_progress
);
3514 /* Do not loop if specifically requested */
3515 if (gfp_mask
& __GFP_NORETRY
)
3518 /* Keep reclaiming pages as long as there is reasonable progress */
3519 pages_reclaimed
+= did_some_progress
;
3520 if ((did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
) ||
3521 ((gfp_mask
& __GFP_REPEAT
) && pages_reclaimed
< (1 << order
))) {
3522 /* Wait for some write requests to complete then retry */
3523 wait_iff_congested(ac
->preferred_zoneref
->zone
, BLK_RW_ASYNC
, HZ
/50);
3527 /* Reclaim has failed us, start killing things */
3528 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3532 /* Retry as long as the OOM killer is making progress */
3533 if (did_some_progress
)
3538 * High-order allocations do not necessarily loop after
3539 * direct reclaim and reclaim/compaction depends on compaction
3540 * being called after reclaim so call directly if necessary
3542 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
,
3544 &contended_compaction
,
3545 &deferred_compaction
);
3549 warn_alloc_failed(gfp_mask
, order
, NULL
);
3555 * This is the 'heart' of the zoned buddy allocator.
3558 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3559 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3562 unsigned int cpuset_mems_cookie
;
3563 unsigned int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_FAIR
;
3564 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3565 struct alloc_context ac
= {
3566 .high_zoneidx
= gfp_zone(gfp_mask
),
3567 .zonelist
= zonelist
,
3568 .nodemask
= nodemask
,
3569 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3572 if (cpusets_enabled()) {
3573 alloc_mask
|= __GFP_HARDWALL
;
3574 alloc_flags
|= ALLOC_CPUSET
;
3576 ac
.nodemask
= &cpuset_current_mems_allowed
;
3579 gfp_mask
&= gfp_allowed_mask
;
3581 lockdep_trace_alloc(gfp_mask
);
3583 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3585 if (should_fail_alloc_page(gfp_mask
, order
))
3589 * Check the zones suitable for the gfp_mask contain at least one
3590 * valid zone. It's possible to have an empty zonelist as a result
3591 * of __GFP_THISNODE and a memoryless node
3593 if (unlikely(!zonelist
->_zonerefs
->zone
))
3596 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3597 alloc_flags
|= ALLOC_CMA
;
3600 cpuset_mems_cookie
= read_mems_allowed_begin();
3602 /* Dirty zone balancing only done in the fast path */
3603 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3605 /* The preferred zone is used for statistics later */
3606 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3607 ac
.high_zoneidx
, ac
.nodemask
);
3608 if (!ac
.preferred_zoneref
) {
3613 /* First allocation attempt */
3614 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3619 * Runtime PM, block IO and its error handling path can deadlock
3620 * because I/O on the device might not complete.
3622 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3623 ac
.spread_dirty_pages
= false;
3625 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3629 * When updating a task's mems_allowed, it is possible to race with
3630 * parallel threads in such a way that an allocation can fail while
3631 * the mask is being updated. If a page allocation is about to fail,
3632 * check if the cpuset changed during allocation and if so, retry.
3634 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3635 alloc_mask
= gfp_mask
;
3640 if (kmemcheck_enabled
&& page
)
3641 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3643 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3647 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3650 * Common helper functions.
3652 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3657 * __get_free_pages() returns a 32-bit address, which cannot represent
3660 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3662 page
= alloc_pages(gfp_mask
, order
);
3665 return (unsigned long) page_address(page
);
3667 EXPORT_SYMBOL(__get_free_pages
);
3669 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3671 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3673 EXPORT_SYMBOL(get_zeroed_page
);
3675 void __free_pages(struct page
*page
, unsigned int order
)
3677 if (put_page_testzero(page
)) {
3679 free_hot_cold_page(page
, false);
3681 __free_pages_ok(page
, order
);
3685 EXPORT_SYMBOL(__free_pages
);
3687 void free_pages(unsigned long addr
, unsigned int order
)
3690 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3691 __free_pages(virt_to_page((void *)addr
), order
);
3695 EXPORT_SYMBOL(free_pages
);
3699 * An arbitrary-length arbitrary-offset area of memory which resides
3700 * within a 0 or higher order page. Multiple fragments within that page
3701 * are individually refcounted, in the page's reference counter.
3703 * The page_frag functions below provide a simple allocation framework for
3704 * page fragments. This is used by the network stack and network device
3705 * drivers to provide a backing region of memory for use as either an
3706 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3708 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3711 struct page
*page
= NULL
;
3712 gfp_t gfp
= gfp_mask
;
3714 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3715 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3717 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3718 PAGE_FRAG_CACHE_MAX_ORDER
);
3719 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3721 if (unlikely(!page
))
3722 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3724 nc
->va
= page
? page_address(page
) : NULL
;
3729 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3730 unsigned int fragsz
, gfp_t gfp_mask
)
3732 unsigned int size
= PAGE_SIZE
;
3736 if (unlikely(!nc
->va
)) {
3738 page
= __page_frag_refill(nc
, gfp_mask
);
3742 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3743 /* if size can vary use size else just use PAGE_SIZE */
3746 /* Even if we own the page, we do not use atomic_set().
3747 * This would break get_page_unless_zero() users.
3749 page_ref_add(page
, size
- 1);
3751 /* reset page count bias and offset to start of new frag */
3752 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3753 nc
->pagecnt_bias
= size
;
3757 offset
= nc
->offset
- fragsz
;
3758 if (unlikely(offset
< 0)) {
3759 page
= virt_to_page(nc
->va
);
3761 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3764 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3765 /* if size can vary use size else just use PAGE_SIZE */
3768 /* OK, page count is 0, we can safely set it */
3769 set_page_count(page
, size
);
3771 /* reset page count bias and offset to start of new frag */
3772 nc
->pagecnt_bias
= size
;
3773 offset
= size
- fragsz
;
3777 nc
->offset
= offset
;
3779 return nc
->va
+ offset
;
3781 EXPORT_SYMBOL(__alloc_page_frag
);
3784 * Frees a page fragment allocated out of either a compound or order 0 page.
3786 void __free_page_frag(void *addr
)
3788 struct page
*page
= virt_to_head_page(addr
);
3790 if (unlikely(put_page_testzero(page
)))
3791 __free_pages_ok(page
, compound_order(page
));
3793 EXPORT_SYMBOL(__free_page_frag
);
3796 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3797 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3798 * equivalent to alloc_pages.
3800 * It should be used when the caller would like to use kmalloc, but since the
3801 * allocation is large, it has to fall back to the page allocator.
3803 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
3807 page
= alloc_pages(gfp_mask
, order
);
3808 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3809 __free_pages(page
, order
);
3815 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
3819 page
= alloc_pages_node(nid
, gfp_mask
, order
);
3820 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3821 __free_pages(page
, order
);
3828 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3831 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3833 memcg_kmem_uncharge(page
, order
);
3834 __free_pages(page
, order
);
3837 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3840 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3841 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3845 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3849 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3850 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3852 split_page(virt_to_page((void *)addr
), order
);
3853 while (used
< alloc_end
) {
3858 return (void *)addr
;
3862 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3863 * @size: the number of bytes to allocate
3864 * @gfp_mask: GFP flags for the allocation
3866 * This function is similar to alloc_pages(), except that it allocates the
3867 * minimum number of pages to satisfy the request. alloc_pages() can only
3868 * allocate memory in power-of-two pages.
3870 * This function is also limited by MAX_ORDER.
3872 * Memory allocated by this function must be released by free_pages_exact().
3874 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3876 unsigned int order
= get_order(size
);
3879 addr
= __get_free_pages(gfp_mask
, order
);
3880 return make_alloc_exact(addr
, order
, size
);
3882 EXPORT_SYMBOL(alloc_pages_exact
);
3885 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3887 * @nid: the preferred node ID where memory should be allocated
3888 * @size: the number of bytes to allocate
3889 * @gfp_mask: GFP flags for the allocation
3891 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3894 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3896 unsigned int order
= get_order(size
);
3897 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3900 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3904 * free_pages_exact - release memory allocated via alloc_pages_exact()
3905 * @virt: the value returned by alloc_pages_exact.
3906 * @size: size of allocation, same value as passed to alloc_pages_exact().
3908 * Release the memory allocated by a previous call to alloc_pages_exact.
3910 void free_pages_exact(void *virt
, size_t size
)
3912 unsigned long addr
= (unsigned long)virt
;
3913 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3915 while (addr
< end
) {
3920 EXPORT_SYMBOL(free_pages_exact
);
3923 * nr_free_zone_pages - count number of pages beyond high watermark
3924 * @offset: The zone index of the highest zone
3926 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3927 * high watermark within all zones at or below a given zone index. For each
3928 * zone, the number of pages is calculated as:
3929 * managed_pages - high_pages
3931 static unsigned long nr_free_zone_pages(int offset
)
3936 /* Just pick one node, since fallback list is circular */
3937 unsigned long sum
= 0;
3939 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3941 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3942 unsigned long size
= zone
->managed_pages
;
3943 unsigned long high
= high_wmark_pages(zone
);
3952 * nr_free_buffer_pages - count number of pages beyond high watermark
3954 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3955 * watermark within ZONE_DMA and ZONE_NORMAL.
3957 unsigned long nr_free_buffer_pages(void)
3959 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3961 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3964 * nr_free_pagecache_pages - count number of pages beyond high watermark
3966 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3967 * high watermark within all zones.
3969 unsigned long nr_free_pagecache_pages(void)
3971 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3974 static inline void show_node(struct zone
*zone
)
3976 if (IS_ENABLED(CONFIG_NUMA
))
3977 printk("Node %d ", zone_to_nid(zone
));
3980 long si_mem_available(void)
3983 unsigned long pagecache
;
3984 unsigned long wmark_low
= 0;
3985 unsigned long pages
[NR_LRU_LISTS
];
3989 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
3990 pages
[lru
] = global_page_state(NR_LRU_BASE
+ lru
);
3993 wmark_low
+= zone
->watermark
[WMARK_LOW
];
3996 * Estimate the amount of memory available for userspace allocations,
3997 * without causing swapping.
3999 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4002 * Not all the page cache can be freed, otherwise the system will
4003 * start swapping. Assume at least half of the page cache, or the
4004 * low watermark worth of cache, needs to stay.
4006 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4007 pagecache
-= min(pagecache
/ 2, wmark_low
);
4008 available
+= pagecache
;
4011 * Part of the reclaimable slab consists of items that are in use,
4012 * and cannot be freed. Cap this estimate at the low watermark.
4014 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4015 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4021 EXPORT_SYMBOL_GPL(si_mem_available
);
4023 void si_meminfo(struct sysinfo
*val
)
4025 val
->totalram
= totalram_pages
;
4026 val
->sharedram
= global_page_state(NR_SHMEM
);
4027 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4028 val
->bufferram
= nr_blockdev_pages();
4029 val
->totalhigh
= totalhigh_pages
;
4030 val
->freehigh
= nr_free_highpages();
4031 val
->mem_unit
= PAGE_SIZE
;
4034 EXPORT_SYMBOL(si_meminfo
);
4037 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4039 int zone_type
; /* needs to be signed */
4040 unsigned long managed_pages
= 0;
4041 unsigned long managed_highpages
= 0;
4042 unsigned long free_highpages
= 0;
4043 pg_data_t
*pgdat
= NODE_DATA(nid
);
4045 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4046 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4047 val
->totalram
= managed_pages
;
4048 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
4049 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
4050 #ifdef CONFIG_HIGHMEM
4051 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4052 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4054 if (is_highmem(zone
)) {
4055 managed_highpages
+= zone
->managed_pages
;
4056 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4059 val
->totalhigh
= managed_highpages
;
4060 val
->freehigh
= free_highpages
;
4062 val
->totalhigh
= managed_highpages
;
4063 val
->freehigh
= free_highpages
;
4065 val
->mem_unit
= PAGE_SIZE
;
4070 * Determine whether the node should be displayed or not, depending on whether
4071 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4073 bool skip_free_areas_node(unsigned int flags
, int nid
)
4076 unsigned int cpuset_mems_cookie
;
4078 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4082 cpuset_mems_cookie
= read_mems_allowed_begin();
4083 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
4084 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
4089 #define K(x) ((x) << (PAGE_SHIFT-10))
4091 static void show_migration_types(unsigned char type
)
4093 static const char types
[MIGRATE_TYPES
] = {
4094 [MIGRATE_UNMOVABLE
] = 'U',
4095 [MIGRATE_MOVABLE
] = 'M',
4096 [MIGRATE_RECLAIMABLE
] = 'E',
4097 [MIGRATE_HIGHATOMIC
] = 'H',
4099 [MIGRATE_CMA
] = 'C',
4101 #ifdef CONFIG_MEMORY_ISOLATION
4102 [MIGRATE_ISOLATE
] = 'I',
4105 char tmp
[MIGRATE_TYPES
+ 1];
4109 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4110 if (type
& (1 << i
))
4115 printk("(%s) ", tmp
);
4119 * Show free area list (used inside shift_scroll-lock stuff)
4120 * We also calculate the percentage fragmentation. We do this by counting the
4121 * memory on each free list with the exception of the first item on the list.
4124 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4127 void show_free_areas(unsigned int filter
)
4129 unsigned long free_pcp
= 0;
4133 for_each_populated_zone(zone
) {
4134 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4137 for_each_online_cpu(cpu
)
4138 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4141 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4142 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4143 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4144 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4145 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4146 " free:%lu free_pcp:%lu free_cma:%lu\n",
4147 global_page_state(NR_ACTIVE_ANON
),
4148 global_page_state(NR_INACTIVE_ANON
),
4149 global_page_state(NR_ISOLATED_ANON
),
4150 global_page_state(NR_ACTIVE_FILE
),
4151 global_page_state(NR_INACTIVE_FILE
),
4152 global_page_state(NR_ISOLATED_FILE
),
4153 global_page_state(NR_UNEVICTABLE
),
4154 global_page_state(NR_FILE_DIRTY
),
4155 global_page_state(NR_WRITEBACK
),
4156 global_page_state(NR_UNSTABLE_NFS
),
4157 global_page_state(NR_SLAB_RECLAIMABLE
),
4158 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4159 global_page_state(NR_FILE_MAPPED
),
4160 global_page_state(NR_SHMEM
),
4161 global_page_state(NR_PAGETABLE
),
4162 global_page_state(NR_BOUNCE
),
4163 global_page_state(NR_FREE_PAGES
),
4165 global_page_state(NR_FREE_CMA_PAGES
));
4167 for_each_populated_zone(zone
) {
4170 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4174 for_each_online_cpu(cpu
)
4175 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4183 " active_anon:%lukB"
4184 " inactive_anon:%lukB"
4185 " active_file:%lukB"
4186 " inactive_file:%lukB"
4187 " unevictable:%lukB"
4188 " isolated(anon):%lukB"
4189 " isolated(file):%lukB"
4197 " slab_reclaimable:%lukB"
4198 " slab_unreclaimable:%lukB"
4199 " kernel_stack:%lukB"
4206 " writeback_tmp:%lukB"
4207 " pages_scanned:%lu"
4208 " all_unreclaimable? %s"
4211 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4212 K(min_wmark_pages(zone
)),
4213 K(low_wmark_pages(zone
)),
4214 K(high_wmark_pages(zone
)),
4215 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
4216 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
4217 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
4218 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
4219 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
4220 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
4221 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
4222 K(zone
->present_pages
),
4223 K(zone
->managed_pages
),
4224 K(zone_page_state(zone
, NR_MLOCK
)),
4225 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
4226 K(zone_page_state(zone
, NR_WRITEBACK
)),
4227 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
4228 K(zone_page_state(zone
, NR_SHMEM
)),
4229 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4230 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4231 zone_page_state(zone
, NR_KERNEL_STACK
) *
4233 K(zone_page_state(zone
, NR_PAGETABLE
)),
4234 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
4235 K(zone_page_state(zone
, NR_BOUNCE
)),
4237 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4238 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
4239 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
4240 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
4241 (!zone_reclaimable(zone
) ? "yes" : "no")
4243 printk("lowmem_reserve[]:");
4244 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4245 printk(" %ld", zone
->lowmem_reserve
[i
]);
4249 for_each_populated_zone(zone
) {
4251 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4252 unsigned char types
[MAX_ORDER
];
4254 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4257 printk("%s: ", zone
->name
);
4259 spin_lock_irqsave(&zone
->lock
, flags
);
4260 for (order
= 0; order
< MAX_ORDER
; order
++) {
4261 struct free_area
*area
= &zone
->free_area
[order
];
4264 nr
[order
] = area
->nr_free
;
4265 total
+= nr
[order
] << order
;
4268 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4269 if (!list_empty(&area
->free_list
[type
]))
4270 types
[order
] |= 1 << type
;
4273 spin_unlock_irqrestore(&zone
->lock
, flags
);
4274 for (order
= 0; order
< MAX_ORDER
; order
++) {
4275 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
4277 show_migration_types(types
[order
]);
4279 printk("= %lukB\n", K(total
));
4282 hugetlb_show_meminfo();
4284 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
4286 show_swap_cache_info();
4289 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4291 zoneref
->zone
= zone
;
4292 zoneref
->zone_idx
= zone_idx(zone
);
4296 * Builds allocation fallback zone lists.
4298 * Add all populated zones of a node to the zonelist.
4300 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4304 enum zone_type zone_type
= MAX_NR_ZONES
;
4308 zone
= pgdat
->node_zones
+ zone_type
;
4309 if (populated_zone(zone
)) {
4310 zoneref_set_zone(zone
,
4311 &zonelist
->_zonerefs
[nr_zones
++]);
4312 check_highest_zone(zone_type
);
4314 } while (zone_type
);
4322 * 0 = automatic detection of better ordering.
4323 * 1 = order by ([node] distance, -zonetype)
4324 * 2 = order by (-zonetype, [node] distance)
4326 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4327 * the same zonelist. So only NUMA can configure this param.
4329 #define ZONELIST_ORDER_DEFAULT 0
4330 #define ZONELIST_ORDER_NODE 1
4331 #define ZONELIST_ORDER_ZONE 2
4333 /* zonelist order in the kernel.
4334 * set_zonelist_order() will set this to NODE or ZONE.
4336 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4337 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4341 /* The value user specified ....changed by config */
4342 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4343 /* string for sysctl */
4344 #define NUMA_ZONELIST_ORDER_LEN 16
4345 char numa_zonelist_order
[16] = "default";
4348 * interface for configure zonelist ordering.
4349 * command line option "numa_zonelist_order"
4350 * = "[dD]efault - default, automatic configuration.
4351 * = "[nN]ode - order by node locality, then by zone within node
4352 * = "[zZ]one - order by zone, then by locality within zone
4355 static int __parse_numa_zonelist_order(char *s
)
4357 if (*s
== 'd' || *s
== 'D') {
4358 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4359 } else if (*s
== 'n' || *s
== 'N') {
4360 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4361 } else if (*s
== 'z' || *s
== 'Z') {
4362 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4364 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4370 static __init
int setup_numa_zonelist_order(char *s
)
4377 ret
= __parse_numa_zonelist_order(s
);
4379 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4383 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4386 * sysctl handler for numa_zonelist_order
4388 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4389 void __user
*buffer
, size_t *length
,
4392 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4394 static DEFINE_MUTEX(zl_order_mutex
);
4396 mutex_lock(&zl_order_mutex
);
4398 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4402 strcpy(saved_string
, (char *)table
->data
);
4404 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4408 int oldval
= user_zonelist_order
;
4410 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4413 * bogus value. restore saved string
4415 strncpy((char *)table
->data
, saved_string
,
4416 NUMA_ZONELIST_ORDER_LEN
);
4417 user_zonelist_order
= oldval
;
4418 } else if (oldval
!= user_zonelist_order
) {
4419 mutex_lock(&zonelists_mutex
);
4420 build_all_zonelists(NULL
, NULL
);
4421 mutex_unlock(&zonelists_mutex
);
4425 mutex_unlock(&zl_order_mutex
);
4430 #define MAX_NODE_LOAD (nr_online_nodes)
4431 static int node_load
[MAX_NUMNODES
];
4434 * find_next_best_node - find the next node that should appear in a given node's fallback list
4435 * @node: node whose fallback list we're appending
4436 * @used_node_mask: nodemask_t of already used nodes
4438 * We use a number of factors to determine which is the next node that should
4439 * appear on a given node's fallback list. The node should not have appeared
4440 * already in @node's fallback list, and it should be the next closest node
4441 * according to the distance array (which contains arbitrary distance values
4442 * from each node to each node in the system), and should also prefer nodes
4443 * with no CPUs, since presumably they'll have very little allocation pressure
4444 * on them otherwise.
4445 * It returns -1 if no node is found.
4447 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4450 int min_val
= INT_MAX
;
4451 int best_node
= NUMA_NO_NODE
;
4452 const struct cpumask
*tmp
= cpumask_of_node(0);
4454 /* Use the local node if we haven't already */
4455 if (!node_isset(node
, *used_node_mask
)) {
4456 node_set(node
, *used_node_mask
);
4460 for_each_node_state(n
, N_MEMORY
) {
4462 /* Don't want a node to appear more than once */
4463 if (node_isset(n
, *used_node_mask
))
4466 /* Use the distance array to find the distance */
4467 val
= node_distance(node
, n
);
4469 /* Penalize nodes under us ("prefer the next node") */
4472 /* Give preference to headless and unused nodes */
4473 tmp
= cpumask_of_node(n
);
4474 if (!cpumask_empty(tmp
))
4475 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4477 /* Slight preference for less loaded node */
4478 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4479 val
+= node_load
[n
];
4481 if (val
< min_val
) {
4488 node_set(best_node
, *used_node_mask
);
4495 * Build zonelists ordered by node and zones within node.
4496 * This results in maximum locality--normal zone overflows into local
4497 * DMA zone, if any--but risks exhausting DMA zone.
4499 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4502 struct zonelist
*zonelist
;
4504 zonelist
= &pgdat
->node_zonelists
[0];
4505 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4507 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4508 zonelist
->_zonerefs
[j
].zone
= NULL
;
4509 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4513 * Build gfp_thisnode zonelists
4515 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4518 struct zonelist
*zonelist
;
4520 zonelist
= &pgdat
->node_zonelists
[1];
4521 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4522 zonelist
->_zonerefs
[j
].zone
= NULL
;
4523 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4527 * Build zonelists ordered by zone and nodes within zones.
4528 * This results in conserving DMA zone[s] until all Normal memory is
4529 * exhausted, but results in overflowing to remote node while memory
4530 * may still exist in local DMA zone.
4532 static int node_order
[MAX_NUMNODES
];
4534 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4537 int zone_type
; /* needs to be signed */
4539 struct zonelist
*zonelist
;
4541 zonelist
= &pgdat
->node_zonelists
[0];
4543 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4544 for (j
= 0; j
< nr_nodes
; j
++) {
4545 node
= node_order
[j
];
4546 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4547 if (populated_zone(z
)) {
4549 &zonelist
->_zonerefs
[pos
++]);
4550 check_highest_zone(zone_type
);
4554 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4555 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4558 #if defined(CONFIG_64BIT)
4560 * Devices that require DMA32/DMA are relatively rare and do not justify a
4561 * penalty to every machine in case the specialised case applies. Default
4562 * to Node-ordering on 64-bit NUMA machines
4564 static int default_zonelist_order(void)
4566 return ZONELIST_ORDER_NODE
;
4570 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4571 * by the kernel. If processes running on node 0 deplete the low memory zone
4572 * then reclaim will occur more frequency increasing stalls and potentially
4573 * be easier to OOM if a large percentage of the zone is under writeback or
4574 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4575 * Hence, default to zone ordering on 32-bit.
4577 static int default_zonelist_order(void)
4579 return ZONELIST_ORDER_ZONE
;
4581 #endif /* CONFIG_64BIT */
4583 static void set_zonelist_order(void)
4585 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4586 current_zonelist_order
= default_zonelist_order();
4588 current_zonelist_order
= user_zonelist_order
;
4591 static void build_zonelists(pg_data_t
*pgdat
)
4594 nodemask_t used_mask
;
4595 int local_node
, prev_node
;
4596 struct zonelist
*zonelist
;
4597 unsigned int order
= current_zonelist_order
;
4599 /* initialize zonelists */
4600 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4601 zonelist
= pgdat
->node_zonelists
+ i
;
4602 zonelist
->_zonerefs
[0].zone
= NULL
;
4603 zonelist
->_zonerefs
[0].zone_idx
= 0;
4606 /* NUMA-aware ordering of nodes */
4607 local_node
= pgdat
->node_id
;
4608 load
= nr_online_nodes
;
4609 prev_node
= local_node
;
4610 nodes_clear(used_mask
);
4612 memset(node_order
, 0, sizeof(node_order
));
4615 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4617 * We don't want to pressure a particular node.
4618 * So adding penalty to the first node in same
4619 * distance group to make it round-robin.
4621 if (node_distance(local_node
, node
) !=
4622 node_distance(local_node
, prev_node
))
4623 node_load
[node
] = load
;
4627 if (order
== ZONELIST_ORDER_NODE
)
4628 build_zonelists_in_node_order(pgdat
, node
);
4630 node_order
[i
++] = node
; /* remember order */
4633 if (order
== ZONELIST_ORDER_ZONE
) {
4634 /* calculate node order -- i.e., DMA last! */
4635 build_zonelists_in_zone_order(pgdat
, i
);
4638 build_thisnode_zonelists(pgdat
);
4641 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4643 * Return node id of node used for "local" allocations.
4644 * I.e., first node id of first zone in arg node's generic zonelist.
4645 * Used for initializing percpu 'numa_mem', which is used primarily
4646 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4648 int local_memory_node(int node
)
4652 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4653 gfp_zone(GFP_KERNEL
),
4655 return z
->zone
->node
;
4659 #else /* CONFIG_NUMA */
4661 static void set_zonelist_order(void)
4663 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4666 static void build_zonelists(pg_data_t
*pgdat
)
4668 int node
, local_node
;
4670 struct zonelist
*zonelist
;
4672 local_node
= pgdat
->node_id
;
4674 zonelist
= &pgdat
->node_zonelists
[0];
4675 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4678 * Now we build the zonelist so that it contains the zones
4679 * of all the other nodes.
4680 * We don't want to pressure a particular node, so when
4681 * building the zones for node N, we make sure that the
4682 * zones coming right after the local ones are those from
4683 * node N+1 (modulo N)
4685 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4686 if (!node_online(node
))
4688 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4690 for (node
= 0; node
< local_node
; node
++) {
4691 if (!node_online(node
))
4693 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4696 zonelist
->_zonerefs
[j
].zone
= NULL
;
4697 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4700 #endif /* CONFIG_NUMA */
4703 * Boot pageset table. One per cpu which is going to be used for all
4704 * zones and all nodes. The parameters will be set in such a way
4705 * that an item put on a list will immediately be handed over to
4706 * the buddy list. This is safe since pageset manipulation is done
4707 * with interrupts disabled.
4709 * The boot_pagesets must be kept even after bootup is complete for
4710 * unused processors and/or zones. They do play a role for bootstrapping
4711 * hotplugged processors.
4713 * zoneinfo_show() and maybe other functions do
4714 * not check if the processor is online before following the pageset pointer.
4715 * Other parts of the kernel may not check if the zone is available.
4717 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4718 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4719 static void setup_zone_pageset(struct zone
*zone
);
4722 * Global mutex to protect against size modification of zonelists
4723 * as well as to serialize pageset setup for the new populated zone.
4725 DEFINE_MUTEX(zonelists_mutex
);
4727 /* return values int ....just for stop_machine() */
4728 static int __build_all_zonelists(void *data
)
4732 pg_data_t
*self
= data
;
4735 memset(node_load
, 0, sizeof(node_load
));
4738 if (self
&& !node_online(self
->node_id
)) {
4739 build_zonelists(self
);
4742 for_each_online_node(nid
) {
4743 pg_data_t
*pgdat
= NODE_DATA(nid
);
4745 build_zonelists(pgdat
);
4749 * Initialize the boot_pagesets that are going to be used
4750 * for bootstrapping processors. The real pagesets for
4751 * each zone will be allocated later when the per cpu
4752 * allocator is available.
4754 * boot_pagesets are used also for bootstrapping offline
4755 * cpus if the system is already booted because the pagesets
4756 * are needed to initialize allocators on a specific cpu too.
4757 * F.e. the percpu allocator needs the page allocator which
4758 * needs the percpu allocator in order to allocate its pagesets
4759 * (a chicken-egg dilemma).
4761 for_each_possible_cpu(cpu
) {
4762 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4764 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4766 * We now know the "local memory node" for each node--
4767 * i.e., the node of the first zone in the generic zonelist.
4768 * Set up numa_mem percpu variable for on-line cpus. During
4769 * boot, only the boot cpu should be on-line; we'll init the
4770 * secondary cpus' numa_mem as they come on-line. During
4771 * node/memory hotplug, we'll fixup all on-line cpus.
4773 if (cpu_online(cpu
))
4774 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4781 static noinline
void __init
4782 build_all_zonelists_init(void)
4784 __build_all_zonelists(NULL
);
4785 mminit_verify_zonelist();
4786 cpuset_init_current_mems_allowed();
4790 * Called with zonelists_mutex held always
4791 * unless system_state == SYSTEM_BOOTING.
4793 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4794 * [we're only called with non-NULL zone through __meminit paths] and
4795 * (2) call of __init annotated helper build_all_zonelists_init
4796 * [protected by SYSTEM_BOOTING].
4798 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4800 set_zonelist_order();
4802 if (system_state
== SYSTEM_BOOTING
) {
4803 build_all_zonelists_init();
4805 #ifdef CONFIG_MEMORY_HOTPLUG
4807 setup_zone_pageset(zone
);
4809 /* we have to stop all cpus to guarantee there is no user
4811 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4812 /* cpuset refresh routine should be here */
4814 vm_total_pages
= nr_free_pagecache_pages();
4816 * Disable grouping by mobility if the number of pages in the
4817 * system is too low to allow the mechanism to work. It would be
4818 * more accurate, but expensive to check per-zone. This check is
4819 * made on memory-hotadd so a system can start with mobility
4820 * disabled and enable it later
4822 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4823 page_group_by_mobility_disabled
= 1;
4825 page_group_by_mobility_disabled
= 0;
4827 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4829 zonelist_order_name
[current_zonelist_order
],
4830 page_group_by_mobility_disabled
? "off" : "on",
4833 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4838 * Helper functions to size the waitqueue hash table.
4839 * Essentially these want to choose hash table sizes sufficiently
4840 * large so that collisions trying to wait on pages are rare.
4841 * But in fact, the number of active page waitqueues on typical
4842 * systems is ridiculously low, less than 200. So this is even
4843 * conservative, even though it seems large.
4845 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4846 * waitqueues, i.e. the size of the waitq table given the number of pages.
4848 #define PAGES_PER_WAITQUEUE 256
4850 #ifndef CONFIG_MEMORY_HOTPLUG
4851 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4853 unsigned long size
= 1;
4855 pages
/= PAGES_PER_WAITQUEUE
;
4857 while (size
< pages
)
4861 * Once we have dozens or even hundreds of threads sleeping
4862 * on IO we've got bigger problems than wait queue collision.
4863 * Limit the size of the wait table to a reasonable size.
4865 size
= min(size
, 4096UL);
4867 return max(size
, 4UL);
4871 * A zone's size might be changed by hot-add, so it is not possible to determine
4872 * a suitable size for its wait_table. So we use the maximum size now.
4874 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4876 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4877 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4878 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4880 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4881 * or more by the traditional way. (See above). It equals:
4883 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4884 * ia64(16K page size) : = ( 8G + 4M)byte.
4885 * powerpc (64K page size) : = (32G +16M)byte.
4887 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4894 * This is an integer logarithm so that shifts can be used later
4895 * to extract the more random high bits from the multiplicative
4896 * hash function before the remainder is taken.
4898 static inline unsigned long wait_table_bits(unsigned long size
)
4904 * Initially all pages are reserved - free ones are freed
4905 * up by free_all_bootmem() once the early boot process is
4906 * done. Non-atomic initialization, single-pass.
4908 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4909 unsigned long start_pfn
, enum memmap_context context
)
4911 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
4912 unsigned long end_pfn
= start_pfn
+ size
;
4913 pg_data_t
*pgdat
= NODE_DATA(nid
);
4915 unsigned long nr_initialised
= 0;
4916 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4917 struct memblock_region
*r
= NULL
, *tmp
;
4920 if (highest_memmap_pfn
< end_pfn
- 1)
4921 highest_memmap_pfn
= end_pfn
- 1;
4924 * Honor reservation requested by the driver for this ZONE_DEVICE
4927 if (altmap
&& start_pfn
== altmap
->base_pfn
)
4928 start_pfn
+= altmap
->reserve
;
4930 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4932 * There can be holes in boot-time mem_map[]s handed to this
4933 * function. They do not exist on hotplugged memory.
4935 if (context
!= MEMMAP_EARLY
)
4938 if (!early_pfn_valid(pfn
))
4940 if (!early_pfn_in_nid(pfn
, nid
))
4942 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
4945 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4947 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4948 * from zone_movable_pfn[nid] to end of each node should be
4949 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4951 if (!mirrored_kernelcore
&& zone_movable_pfn
[nid
])
4952 if (zone
== ZONE_NORMAL
&& pfn
>= zone_movable_pfn
[nid
])
4956 * Check given memblock attribute by firmware which can affect
4957 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4958 * mirrored, it's an overlapped memmap init. skip it.
4960 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
4961 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
4962 for_each_memblock(memory
, tmp
)
4963 if (pfn
< memblock_region_memory_end_pfn(tmp
))
4967 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
4968 memblock_is_mirror(r
)) {
4969 /* already initialized as NORMAL */
4970 pfn
= memblock_region_memory_end_pfn(r
);
4978 * Mark the block movable so that blocks are reserved for
4979 * movable at startup. This will force kernel allocations
4980 * to reserve their blocks rather than leaking throughout
4981 * the address space during boot when many long-lived
4982 * kernel allocations are made.
4984 * bitmap is created for zone's valid pfn range. but memmap
4985 * can be created for invalid pages (for alignment)
4986 * check here not to call set_pageblock_migratetype() against
4989 if (!(pfn
& (pageblock_nr_pages
- 1))) {
4990 struct page
*page
= pfn_to_page(pfn
);
4992 __init_single_page(page
, pfn
, zone
, nid
);
4993 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4995 __init_single_pfn(pfn
, zone
, nid
);
5000 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5002 unsigned int order
, t
;
5003 for_each_migratetype_order(order
, t
) {
5004 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5005 zone
->free_area
[order
].nr_free
= 0;
5009 #ifndef __HAVE_ARCH_MEMMAP_INIT
5010 #define memmap_init(size, nid, zone, start_pfn) \
5011 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5014 static int zone_batchsize(struct zone
*zone
)
5020 * The per-cpu-pages pools are set to around 1000th of the
5021 * size of the zone. But no more than 1/2 of a meg.
5023 * OK, so we don't know how big the cache is. So guess.
5025 batch
= zone
->managed_pages
/ 1024;
5026 if (batch
* PAGE_SIZE
> 512 * 1024)
5027 batch
= (512 * 1024) / PAGE_SIZE
;
5028 batch
/= 4; /* We effectively *= 4 below */
5033 * Clamp the batch to a 2^n - 1 value. Having a power
5034 * of 2 value was found to be more likely to have
5035 * suboptimal cache aliasing properties in some cases.
5037 * For example if 2 tasks are alternately allocating
5038 * batches of pages, one task can end up with a lot
5039 * of pages of one half of the possible page colors
5040 * and the other with pages of the other colors.
5042 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5047 /* The deferral and batching of frees should be suppressed under NOMMU
5050 * The problem is that NOMMU needs to be able to allocate large chunks
5051 * of contiguous memory as there's no hardware page translation to
5052 * assemble apparent contiguous memory from discontiguous pages.
5054 * Queueing large contiguous runs of pages for batching, however,
5055 * causes the pages to actually be freed in smaller chunks. As there
5056 * can be a significant delay between the individual batches being
5057 * recycled, this leads to the once large chunks of space being
5058 * fragmented and becoming unavailable for high-order allocations.
5065 * pcp->high and pcp->batch values are related and dependent on one another:
5066 * ->batch must never be higher then ->high.
5067 * The following function updates them in a safe manner without read side
5070 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5071 * those fields changing asynchronously (acording the the above rule).
5073 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5074 * outside of boot time (or some other assurance that no concurrent updaters
5077 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5078 unsigned long batch
)
5080 /* start with a fail safe value for batch */
5084 /* Update high, then batch, in order */
5091 /* a companion to pageset_set_high() */
5092 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5094 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5097 static void pageset_init(struct per_cpu_pageset
*p
)
5099 struct per_cpu_pages
*pcp
;
5102 memset(p
, 0, sizeof(*p
));
5106 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5107 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5110 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5113 pageset_set_batch(p
, batch
);
5117 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5118 * to the value high for the pageset p.
5120 static void pageset_set_high(struct per_cpu_pageset
*p
,
5123 unsigned long batch
= max(1UL, high
/ 4);
5124 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5125 batch
= PAGE_SHIFT
* 8;
5127 pageset_update(&p
->pcp
, high
, batch
);
5130 static void pageset_set_high_and_batch(struct zone
*zone
,
5131 struct per_cpu_pageset
*pcp
)
5133 if (percpu_pagelist_fraction
)
5134 pageset_set_high(pcp
,
5135 (zone
->managed_pages
/
5136 percpu_pagelist_fraction
));
5138 pageset_set_batch(pcp
, zone_batchsize(zone
));
5141 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5143 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5146 pageset_set_high_and_batch(zone
, pcp
);
5149 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5152 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5153 for_each_possible_cpu(cpu
)
5154 zone_pageset_init(zone
, cpu
);
5158 * Allocate per cpu pagesets and initialize them.
5159 * Before this call only boot pagesets were available.
5161 void __init
setup_per_cpu_pageset(void)
5165 for_each_populated_zone(zone
)
5166 setup_zone_pageset(zone
);
5169 static noinline __init_refok
5170 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
5176 * The per-page waitqueue mechanism uses hashed waitqueues
5179 zone
->wait_table_hash_nr_entries
=
5180 wait_table_hash_nr_entries(zone_size_pages
);
5181 zone
->wait_table_bits
=
5182 wait_table_bits(zone
->wait_table_hash_nr_entries
);
5183 alloc_size
= zone
->wait_table_hash_nr_entries
5184 * sizeof(wait_queue_head_t
);
5186 if (!slab_is_available()) {
5187 zone
->wait_table
= (wait_queue_head_t
*)
5188 memblock_virt_alloc_node_nopanic(
5189 alloc_size
, zone
->zone_pgdat
->node_id
);
5192 * This case means that a zone whose size was 0 gets new memory
5193 * via memory hot-add.
5194 * But it may be the case that a new node was hot-added. In
5195 * this case vmalloc() will not be able to use this new node's
5196 * memory - this wait_table must be initialized to use this new
5197 * node itself as well.
5198 * To use this new node's memory, further consideration will be
5201 zone
->wait_table
= vmalloc(alloc_size
);
5203 if (!zone
->wait_table
)
5206 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
5207 init_waitqueue_head(zone
->wait_table
+ i
);
5212 static __meminit
void zone_pcp_init(struct zone
*zone
)
5215 * per cpu subsystem is not up at this point. The following code
5216 * relies on the ability of the linker to provide the
5217 * offset of a (static) per cpu variable into the per cpu area.
5219 zone
->pageset
= &boot_pageset
;
5221 if (populated_zone(zone
))
5222 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5223 zone
->name
, zone
->present_pages
,
5224 zone_batchsize(zone
));
5227 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5228 unsigned long zone_start_pfn
,
5231 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5233 ret
= zone_wait_table_init(zone
, size
);
5236 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5238 zone
->zone_start_pfn
= zone_start_pfn
;
5240 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5241 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5243 (unsigned long)zone_idx(zone
),
5244 zone_start_pfn
, (zone_start_pfn
+ size
));
5246 zone_init_free_lists(zone
);
5251 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5252 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5255 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5257 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5258 struct mminit_pfnnid_cache
*state
)
5260 unsigned long start_pfn
, end_pfn
;
5263 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5264 return state
->last_nid
;
5266 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5268 state
->last_start
= start_pfn
;
5269 state
->last_end
= end_pfn
;
5270 state
->last_nid
= nid
;
5275 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5278 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5279 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5280 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5282 * If an architecture guarantees that all ranges registered contain no holes
5283 * and may be freed, this this function may be used instead of calling
5284 * memblock_free_early_nid() manually.
5286 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5288 unsigned long start_pfn
, end_pfn
;
5291 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5292 start_pfn
= min(start_pfn
, max_low_pfn
);
5293 end_pfn
= min(end_pfn
, max_low_pfn
);
5295 if (start_pfn
< end_pfn
)
5296 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5297 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5303 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5304 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5306 * If an architecture guarantees that all ranges registered contain no holes and may
5307 * be freed, this function may be used instead of calling memory_present() manually.
5309 void __init
sparse_memory_present_with_active_regions(int nid
)
5311 unsigned long start_pfn
, end_pfn
;
5314 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5315 memory_present(this_nid
, start_pfn
, end_pfn
);
5319 * get_pfn_range_for_nid - Return the start and end page frames for a node
5320 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5321 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5322 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5324 * It returns the start and end page frame of a node based on information
5325 * provided by memblock_set_node(). If called for a node
5326 * with no available memory, a warning is printed and the start and end
5329 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5330 unsigned long *start_pfn
, unsigned long *end_pfn
)
5332 unsigned long this_start_pfn
, this_end_pfn
;
5338 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5339 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5340 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5343 if (*start_pfn
== -1UL)
5348 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5349 * assumption is made that zones within a node are ordered in monotonic
5350 * increasing memory addresses so that the "highest" populated zone is used
5352 static void __init
find_usable_zone_for_movable(void)
5355 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5356 if (zone_index
== ZONE_MOVABLE
)
5359 if (arch_zone_highest_possible_pfn
[zone_index
] >
5360 arch_zone_lowest_possible_pfn
[zone_index
])
5364 VM_BUG_ON(zone_index
== -1);
5365 movable_zone
= zone_index
;
5369 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5370 * because it is sized independent of architecture. Unlike the other zones,
5371 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5372 * in each node depending on the size of each node and how evenly kernelcore
5373 * is distributed. This helper function adjusts the zone ranges
5374 * provided by the architecture for a given node by using the end of the
5375 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5376 * zones within a node are in order of monotonic increases memory addresses
5378 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5379 unsigned long zone_type
,
5380 unsigned long node_start_pfn
,
5381 unsigned long node_end_pfn
,
5382 unsigned long *zone_start_pfn
,
5383 unsigned long *zone_end_pfn
)
5385 /* Only adjust if ZONE_MOVABLE is on this node */
5386 if (zone_movable_pfn
[nid
]) {
5387 /* Size ZONE_MOVABLE */
5388 if (zone_type
== ZONE_MOVABLE
) {
5389 *zone_start_pfn
= zone_movable_pfn
[nid
];
5390 *zone_end_pfn
= min(node_end_pfn
,
5391 arch_zone_highest_possible_pfn
[movable_zone
]);
5393 /* Check if this whole range is within ZONE_MOVABLE */
5394 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5395 *zone_start_pfn
= *zone_end_pfn
;
5400 * Return the number of pages a zone spans in a node, including holes
5401 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5403 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5404 unsigned long zone_type
,
5405 unsigned long node_start_pfn
,
5406 unsigned long node_end_pfn
,
5407 unsigned long *zone_start_pfn
,
5408 unsigned long *zone_end_pfn
,
5409 unsigned long *ignored
)
5411 /* When hotadd a new node from cpu_up(), the node should be empty */
5412 if (!node_start_pfn
&& !node_end_pfn
)
5415 /* Get the start and end of the zone */
5416 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5417 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5418 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5419 node_start_pfn
, node_end_pfn
,
5420 zone_start_pfn
, zone_end_pfn
);
5422 /* Check that this node has pages within the zone's required range */
5423 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5426 /* Move the zone boundaries inside the node if necessary */
5427 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5428 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5430 /* Return the spanned pages */
5431 return *zone_end_pfn
- *zone_start_pfn
;
5435 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5436 * then all holes in the requested range will be accounted for.
5438 unsigned long __meminit
__absent_pages_in_range(int nid
,
5439 unsigned long range_start_pfn
,
5440 unsigned long range_end_pfn
)
5442 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5443 unsigned long start_pfn
, end_pfn
;
5446 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5447 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5448 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5449 nr_absent
-= end_pfn
- start_pfn
;
5455 * absent_pages_in_range - Return number of page frames in holes within a range
5456 * @start_pfn: The start PFN to start searching for holes
5457 * @end_pfn: The end PFN to stop searching for holes
5459 * It returns the number of pages frames in memory holes within a range.
5461 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5462 unsigned long end_pfn
)
5464 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5467 /* Return the number of page frames in holes in a zone on a node */
5468 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5469 unsigned long zone_type
,
5470 unsigned long node_start_pfn
,
5471 unsigned long node_end_pfn
,
5472 unsigned long *ignored
)
5474 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5475 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5476 unsigned long zone_start_pfn
, zone_end_pfn
;
5477 unsigned long nr_absent
;
5479 /* When hotadd a new node from cpu_up(), the node should be empty */
5480 if (!node_start_pfn
&& !node_end_pfn
)
5483 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5484 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5486 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5487 node_start_pfn
, node_end_pfn
,
5488 &zone_start_pfn
, &zone_end_pfn
);
5489 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5492 * ZONE_MOVABLE handling.
5493 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5496 if (zone_movable_pfn
[nid
]) {
5497 if (mirrored_kernelcore
) {
5498 unsigned long start_pfn
, end_pfn
;
5499 struct memblock_region
*r
;
5501 for_each_memblock(memory
, r
) {
5502 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5503 zone_start_pfn
, zone_end_pfn
);
5504 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5505 zone_start_pfn
, zone_end_pfn
);
5507 if (zone_type
== ZONE_MOVABLE
&&
5508 memblock_is_mirror(r
))
5509 nr_absent
+= end_pfn
- start_pfn
;
5511 if (zone_type
== ZONE_NORMAL
&&
5512 !memblock_is_mirror(r
))
5513 nr_absent
+= end_pfn
- start_pfn
;
5516 if (zone_type
== ZONE_NORMAL
)
5517 nr_absent
+= node_end_pfn
- zone_movable_pfn
[nid
];
5524 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5525 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5526 unsigned long zone_type
,
5527 unsigned long node_start_pfn
,
5528 unsigned long node_end_pfn
,
5529 unsigned long *zone_start_pfn
,
5530 unsigned long *zone_end_pfn
,
5531 unsigned long *zones_size
)
5535 *zone_start_pfn
= node_start_pfn
;
5536 for (zone
= 0; zone
< zone_type
; zone
++)
5537 *zone_start_pfn
+= zones_size
[zone
];
5539 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5541 return zones_size
[zone_type
];
5544 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5545 unsigned long zone_type
,
5546 unsigned long node_start_pfn
,
5547 unsigned long node_end_pfn
,
5548 unsigned long *zholes_size
)
5553 return zholes_size
[zone_type
];
5556 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5558 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5559 unsigned long node_start_pfn
,
5560 unsigned long node_end_pfn
,
5561 unsigned long *zones_size
,
5562 unsigned long *zholes_size
)
5564 unsigned long realtotalpages
= 0, totalpages
= 0;
5567 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5568 struct zone
*zone
= pgdat
->node_zones
+ i
;
5569 unsigned long zone_start_pfn
, zone_end_pfn
;
5570 unsigned long size
, real_size
;
5572 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5578 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5579 node_start_pfn
, node_end_pfn
,
5582 zone
->zone_start_pfn
= zone_start_pfn
;
5584 zone
->zone_start_pfn
= 0;
5585 zone
->spanned_pages
= size
;
5586 zone
->present_pages
= real_size
;
5589 realtotalpages
+= real_size
;
5592 pgdat
->node_spanned_pages
= totalpages
;
5593 pgdat
->node_present_pages
= realtotalpages
;
5594 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5598 #ifndef CONFIG_SPARSEMEM
5600 * Calculate the size of the zone->blockflags rounded to an unsigned long
5601 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5602 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5603 * round what is now in bits to nearest long in bits, then return it in
5606 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5608 unsigned long usemapsize
;
5610 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5611 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5612 usemapsize
= usemapsize
>> pageblock_order
;
5613 usemapsize
*= NR_PAGEBLOCK_BITS
;
5614 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5616 return usemapsize
/ 8;
5619 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5621 unsigned long zone_start_pfn
,
5622 unsigned long zonesize
)
5624 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5625 zone
->pageblock_flags
= NULL
;
5627 zone
->pageblock_flags
=
5628 memblock_virt_alloc_node_nopanic(usemapsize
,
5632 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5633 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5634 #endif /* CONFIG_SPARSEMEM */
5636 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5638 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5639 void __paginginit
set_pageblock_order(void)
5643 /* Check that pageblock_nr_pages has not already been setup */
5644 if (pageblock_order
)
5647 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5648 order
= HUGETLB_PAGE_ORDER
;
5650 order
= MAX_ORDER
- 1;
5653 * Assume the largest contiguous order of interest is a huge page.
5654 * This value may be variable depending on boot parameters on IA64 and
5657 pageblock_order
= order
;
5659 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5662 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5663 * is unused as pageblock_order is set at compile-time. See
5664 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5667 void __paginginit
set_pageblock_order(void)
5671 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5673 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5674 unsigned long present_pages
)
5676 unsigned long pages
= spanned_pages
;
5679 * Provide a more accurate estimation if there are holes within
5680 * the zone and SPARSEMEM is in use. If there are holes within the
5681 * zone, each populated memory region may cost us one or two extra
5682 * memmap pages due to alignment because memmap pages for each
5683 * populated regions may not naturally algined on page boundary.
5684 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5686 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5687 IS_ENABLED(CONFIG_SPARSEMEM
))
5688 pages
= present_pages
;
5690 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5694 * Set up the zone data structures:
5695 * - mark all pages reserved
5696 * - mark all memory queues empty
5697 * - clear the memory bitmaps
5699 * NOTE: pgdat should get zeroed by caller.
5701 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5704 int nid
= pgdat
->node_id
;
5707 pgdat_resize_init(pgdat
);
5708 #ifdef CONFIG_NUMA_BALANCING
5709 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5710 pgdat
->numabalancing_migrate_nr_pages
= 0;
5711 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5713 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5714 spin_lock_init(&pgdat
->split_queue_lock
);
5715 INIT_LIST_HEAD(&pgdat
->split_queue
);
5716 pgdat
->split_queue_len
= 0;
5718 init_waitqueue_head(&pgdat
->kswapd_wait
);
5719 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5720 #ifdef CONFIG_COMPACTION
5721 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5723 pgdat_page_ext_init(pgdat
);
5725 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5726 struct zone
*zone
= pgdat
->node_zones
+ j
;
5727 unsigned long size
, realsize
, freesize
, memmap_pages
;
5728 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5730 size
= zone
->spanned_pages
;
5731 realsize
= freesize
= zone
->present_pages
;
5734 * Adjust freesize so that it accounts for how much memory
5735 * is used by this zone for memmap. This affects the watermark
5736 * and per-cpu initialisations
5738 memmap_pages
= calc_memmap_size(size
, realsize
);
5739 if (!is_highmem_idx(j
)) {
5740 if (freesize
>= memmap_pages
) {
5741 freesize
-= memmap_pages
;
5744 " %s zone: %lu pages used for memmap\n",
5745 zone_names
[j
], memmap_pages
);
5747 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5748 zone_names
[j
], memmap_pages
, freesize
);
5751 /* Account for reserved pages */
5752 if (j
== 0 && freesize
> dma_reserve
) {
5753 freesize
-= dma_reserve
;
5754 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5755 zone_names
[0], dma_reserve
);
5758 if (!is_highmem_idx(j
))
5759 nr_kernel_pages
+= freesize
;
5760 /* Charge for highmem memmap if there are enough kernel pages */
5761 else if (nr_kernel_pages
> memmap_pages
* 2)
5762 nr_kernel_pages
-= memmap_pages
;
5763 nr_all_pages
+= freesize
;
5766 * Set an approximate value for lowmem here, it will be adjusted
5767 * when the bootmem allocator frees pages into the buddy system.
5768 * And all highmem pages will be managed by the buddy system.
5770 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5773 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
5775 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
5777 zone
->name
= zone_names
[j
];
5778 spin_lock_init(&zone
->lock
);
5779 spin_lock_init(&zone
->lru_lock
);
5780 zone_seqlock_init(zone
);
5781 zone
->zone_pgdat
= pgdat
;
5782 zone_pcp_init(zone
);
5784 /* For bootup, initialized properly in watermark setup */
5785 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
5787 lruvec_init(&zone
->lruvec
);
5791 set_pageblock_order();
5792 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5793 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5795 memmap_init(size
, nid
, j
, zone_start_pfn
);
5799 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
5801 unsigned long __maybe_unused start
= 0;
5802 unsigned long __maybe_unused offset
= 0;
5804 /* Skip empty nodes */
5805 if (!pgdat
->node_spanned_pages
)
5808 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5809 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5810 offset
= pgdat
->node_start_pfn
- start
;
5811 /* ia64 gets its own node_mem_map, before this, without bootmem */
5812 if (!pgdat
->node_mem_map
) {
5813 unsigned long size
, end
;
5817 * The zone's endpoints aren't required to be MAX_ORDER
5818 * aligned but the node_mem_map endpoints must be in order
5819 * for the buddy allocator to function correctly.
5821 end
= pgdat_end_pfn(pgdat
);
5822 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5823 size
= (end
- start
) * sizeof(struct page
);
5824 map
= alloc_remap(pgdat
->node_id
, size
);
5826 map
= memblock_virt_alloc_node_nopanic(size
,
5828 pgdat
->node_mem_map
= map
+ offset
;
5830 #ifndef CONFIG_NEED_MULTIPLE_NODES
5832 * With no DISCONTIG, the global mem_map is just set as node 0's
5834 if (pgdat
== NODE_DATA(0)) {
5835 mem_map
= NODE_DATA(0)->node_mem_map
;
5836 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5837 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5839 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5842 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5845 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5846 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5848 pg_data_t
*pgdat
= NODE_DATA(nid
);
5849 unsigned long start_pfn
= 0;
5850 unsigned long end_pfn
= 0;
5852 /* pg_data_t should be reset to zero when it's allocated */
5853 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5855 reset_deferred_meminit(pgdat
);
5856 pgdat
->node_id
= nid
;
5857 pgdat
->node_start_pfn
= node_start_pfn
;
5858 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5859 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5860 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5861 (u64
)start_pfn
<< PAGE_SHIFT
,
5862 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5864 start_pfn
= node_start_pfn
;
5866 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5867 zones_size
, zholes_size
);
5869 alloc_node_mem_map(pgdat
);
5870 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5871 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5872 nid
, (unsigned long)pgdat
,
5873 (unsigned long)pgdat
->node_mem_map
);
5876 free_area_init_core(pgdat
);
5879 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5881 #if MAX_NUMNODES > 1
5883 * Figure out the number of possible node ids.
5885 void __init
setup_nr_node_ids(void)
5887 unsigned int highest
;
5889 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5890 nr_node_ids
= highest
+ 1;
5895 * node_map_pfn_alignment - determine the maximum internode alignment
5897 * This function should be called after node map is populated and sorted.
5898 * It calculates the maximum power of two alignment which can distinguish
5901 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5902 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5903 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5904 * shifted, 1GiB is enough and this function will indicate so.
5906 * This is used to test whether pfn -> nid mapping of the chosen memory
5907 * model has fine enough granularity to avoid incorrect mapping for the
5908 * populated node map.
5910 * Returns the determined alignment in pfn's. 0 if there is no alignment
5911 * requirement (single node).
5913 unsigned long __init
node_map_pfn_alignment(void)
5915 unsigned long accl_mask
= 0, last_end
= 0;
5916 unsigned long start
, end
, mask
;
5920 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5921 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5928 * Start with a mask granular enough to pin-point to the
5929 * start pfn and tick off bits one-by-one until it becomes
5930 * too coarse to separate the current node from the last.
5932 mask
= ~((1 << __ffs(start
)) - 1);
5933 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5936 /* accumulate all internode masks */
5940 /* convert mask to number of pages */
5941 return ~accl_mask
+ 1;
5944 /* Find the lowest pfn for a node */
5945 static unsigned long __init
find_min_pfn_for_node(int nid
)
5947 unsigned long min_pfn
= ULONG_MAX
;
5948 unsigned long start_pfn
;
5951 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5952 min_pfn
= min(min_pfn
, start_pfn
);
5954 if (min_pfn
== ULONG_MAX
) {
5955 pr_warn("Could not find start_pfn for node %d\n", nid
);
5963 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5965 * It returns the minimum PFN based on information provided via
5966 * memblock_set_node().
5968 unsigned long __init
find_min_pfn_with_active_regions(void)
5970 return find_min_pfn_for_node(MAX_NUMNODES
);
5974 * early_calculate_totalpages()
5975 * Sum pages in active regions for movable zone.
5976 * Populate N_MEMORY for calculating usable_nodes.
5978 static unsigned long __init
early_calculate_totalpages(void)
5980 unsigned long totalpages
= 0;
5981 unsigned long start_pfn
, end_pfn
;
5984 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5985 unsigned long pages
= end_pfn
- start_pfn
;
5987 totalpages
+= pages
;
5989 node_set_state(nid
, N_MEMORY
);
5995 * Find the PFN the Movable zone begins in each node. Kernel memory
5996 * is spread evenly between nodes as long as the nodes have enough
5997 * memory. When they don't, some nodes will have more kernelcore than
6000 static void __init
find_zone_movable_pfns_for_nodes(void)
6003 unsigned long usable_startpfn
;
6004 unsigned long kernelcore_node
, kernelcore_remaining
;
6005 /* save the state before borrow the nodemask */
6006 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6007 unsigned long totalpages
= early_calculate_totalpages();
6008 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6009 struct memblock_region
*r
;
6011 /* Need to find movable_zone earlier when movable_node is specified. */
6012 find_usable_zone_for_movable();
6015 * If movable_node is specified, ignore kernelcore and movablecore
6018 if (movable_node_is_enabled()) {
6019 for_each_memblock(memory
, r
) {
6020 if (!memblock_is_hotpluggable(r
))
6025 usable_startpfn
= PFN_DOWN(r
->base
);
6026 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6027 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6035 * If kernelcore=mirror is specified, ignore movablecore option
6037 if (mirrored_kernelcore
) {
6038 bool mem_below_4gb_not_mirrored
= false;
6040 for_each_memblock(memory
, r
) {
6041 if (memblock_is_mirror(r
))
6046 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6048 if (usable_startpfn
< 0x100000) {
6049 mem_below_4gb_not_mirrored
= true;
6053 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6054 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6058 if (mem_below_4gb_not_mirrored
)
6059 pr_warn("This configuration results in unmirrored kernel memory.");
6065 * If movablecore=nn[KMG] was specified, calculate what size of
6066 * kernelcore that corresponds so that memory usable for
6067 * any allocation type is evenly spread. If both kernelcore
6068 * and movablecore are specified, then the value of kernelcore
6069 * will be used for required_kernelcore if it's greater than
6070 * what movablecore would have allowed.
6072 if (required_movablecore
) {
6073 unsigned long corepages
;
6076 * Round-up so that ZONE_MOVABLE is at least as large as what
6077 * was requested by the user
6079 required_movablecore
=
6080 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6081 required_movablecore
= min(totalpages
, required_movablecore
);
6082 corepages
= totalpages
- required_movablecore
;
6084 required_kernelcore
= max(required_kernelcore
, corepages
);
6088 * If kernelcore was not specified or kernelcore size is larger
6089 * than totalpages, there is no ZONE_MOVABLE.
6091 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6094 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6095 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6098 /* Spread kernelcore memory as evenly as possible throughout nodes */
6099 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6100 for_each_node_state(nid
, N_MEMORY
) {
6101 unsigned long start_pfn
, end_pfn
;
6104 * Recalculate kernelcore_node if the division per node
6105 * now exceeds what is necessary to satisfy the requested
6106 * amount of memory for the kernel
6108 if (required_kernelcore
< kernelcore_node
)
6109 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6112 * As the map is walked, we track how much memory is usable
6113 * by the kernel using kernelcore_remaining. When it is
6114 * 0, the rest of the node is usable by ZONE_MOVABLE
6116 kernelcore_remaining
= kernelcore_node
;
6118 /* Go through each range of PFNs within this node */
6119 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6120 unsigned long size_pages
;
6122 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6123 if (start_pfn
>= end_pfn
)
6126 /* Account for what is only usable for kernelcore */
6127 if (start_pfn
< usable_startpfn
) {
6128 unsigned long kernel_pages
;
6129 kernel_pages
= min(end_pfn
, usable_startpfn
)
6132 kernelcore_remaining
-= min(kernel_pages
,
6133 kernelcore_remaining
);
6134 required_kernelcore
-= min(kernel_pages
,
6135 required_kernelcore
);
6137 /* Continue if range is now fully accounted */
6138 if (end_pfn
<= usable_startpfn
) {
6141 * Push zone_movable_pfn to the end so
6142 * that if we have to rebalance
6143 * kernelcore across nodes, we will
6144 * not double account here
6146 zone_movable_pfn
[nid
] = end_pfn
;
6149 start_pfn
= usable_startpfn
;
6153 * The usable PFN range for ZONE_MOVABLE is from
6154 * start_pfn->end_pfn. Calculate size_pages as the
6155 * number of pages used as kernelcore
6157 size_pages
= end_pfn
- start_pfn
;
6158 if (size_pages
> kernelcore_remaining
)
6159 size_pages
= kernelcore_remaining
;
6160 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6163 * Some kernelcore has been met, update counts and
6164 * break if the kernelcore for this node has been
6167 required_kernelcore
-= min(required_kernelcore
,
6169 kernelcore_remaining
-= size_pages
;
6170 if (!kernelcore_remaining
)
6176 * If there is still required_kernelcore, we do another pass with one
6177 * less node in the count. This will push zone_movable_pfn[nid] further
6178 * along on the nodes that still have memory until kernelcore is
6182 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6186 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6187 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6188 zone_movable_pfn
[nid
] =
6189 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6192 /* restore the node_state */
6193 node_states
[N_MEMORY
] = saved_node_state
;
6196 /* Any regular or high memory on that node ? */
6197 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6199 enum zone_type zone_type
;
6201 if (N_MEMORY
== N_NORMAL_MEMORY
)
6204 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6205 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6206 if (populated_zone(zone
)) {
6207 node_set_state(nid
, N_HIGH_MEMORY
);
6208 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6209 zone_type
<= ZONE_NORMAL
)
6210 node_set_state(nid
, N_NORMAL_MEMORY
);
6217 * free_area_init_nodes - Initialise all pg_data_t and zone data
6218 * @max_zone_pfn: an array of max PFNs for each zone
6220 * This will call free_area_init_node() for each active node in the system.
6221 * Using the page ranges provided by memblock_set_node(), the size of each
6222 * zone in each node and their holes is calculated. If the maximum PFN
6223 * between two adjacent zones match, it is assumed that the zone is empty.
6224 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6225 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6226 * starts where the previous one ended. For example, ZONE_DMA32 starts
6227 * at arch_max_dma_pfn.
6229 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6231 unsigned long start_pfn
, end_pfn
;
6234 /* Record where the zone boundaries are */
6235 memset(arch_zone_lowest_possible_pfn
, 0,
6236 sizeof(arch_zone_lowest_possible_pfn
));
6237 memset(arch_zone_highest_possible_pfn
, 0,
6238 sizeof(arch_zone_highest_possible_pfn
));
6239 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
6240 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
6241 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
6242 if (i
== ZONE_MOVABLE
)
6244 arch_zone_lowest_possible_pfn
[i
] =
6245 arch_zone_highest_possible_pfn
[i
-1];
6246 arch_zone_highest_possible_pfn
[i
] =
6247 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
6249 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6250 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6252 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6253 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6254 find_zone_movable_pfns_for_nodes();
6256 /* Print out the zone ranges */
6257 pr_info("Zone ranges:\n");
6258 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6259 if (i
== ZONE_MOVABLE
)
6261 pr_info(" %-8s ", zone_names
[i
]);
6262 if (arch_zone_lowest_possible_pfn
[i
] ==
6263 arch_zone_highest_possible_pfn
[i
])
6266 pr_cont("[mem %#018Lx-%#018Lx]\n",
6267 (u64
)arch_zone_lowest_possible_pfn
[i
]
6269 ((u64
)arch_zone_highest_possible_pfn
[i
]
6270 << PAGE_SHIFT
) - 1);
6273 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6274 pr_info("Movable zone start for each node\n");
6275 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6276 if (zone_movable_pfn
[i
])
6277 pr_info(" Node %d: %#018Lx\n", i
,
6278 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6281 /* Print out the early node map */
6282 pr_info("Early memory node ranges\n");
6283 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6284 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6285 (u64
)start_pfn
<< PAGE_SHIFT
,
6286 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6288 /* Initialise every node */
6289 mminit_verify_pageflags_layout();
6290 setup_nr_node_ids();
6291 for_each_online_node(nid
) {
6292 pg_data_t
*pgdat
= NODE_DATA(nid
);
6293 free_area_init_node(nid
, NULL
,
6294 find_min_pfn_for_node(nid
), NULL
);
6296 /* Any memory on that node */
6297 if (pgdat
->node_present_pages
)
6298 node_set_state(nid
, N_MEMORY
);
6299 check_for_memory(pgdat
, nid
);
6303 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6305 unsigned long long coremem
;
6309 coremem
= memparse(p
, &p
);
6310 *core
= coremem
>> PAGE_SHIFT
;
6312 /* Paranoid check that UL is enough for the coremem value */
6313 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6319 * kernelcore=size sets the amount of memory for use for allocations that
6320 * cannot be reclaimed or migrated.
6322 static int __init
cmdline_parse_kernelcore(char *p
)
6324 /* parse kernelcore=mirror */
6325 if (parse_option_str(p
, "mirror")) {
6326 mirrored_kernelcore
= true;
6330 return cmdline_parse_core(p
, &required_kernelcore
);
6334 * movablecore=size sets the amount of memory for use for allocations that
6335 * can be reclaimed or migrated.
6337 static int __init
cmdline_parse_movablecore(char *p
)
6339 return cmdline_parse_core(p
, &required_movablecore
);
6342 early_param("kernelcore", cmdline_parse_kernelcore
);
6343 early_param("movablecore", cmdline_parse_movablecore
);
6345 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6347 void adjust_managed_page_count(struct page
*page
, long count
)
6349 spin_lock(&managed_page_count_lock
);
6350 page_zone(page
)->managed_pages
+= count
;
6351 totalram_pages
+= count
;
6352 #ifdef CONFIG_HIGHMEM
6353 if (PageHighMem(page
))
6354 totalhigh_pages
+= count
;
6356 spin_unlock(&managed_page_count_lock
);
6358 EXPORT_SYMBOL(adjust_managed_page_count
);
6360 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6363 unsigned long pages
= 0;
6365 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6366 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6367 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6368 if ((unsigned int)poison
<= 0xFF)
6369 memset(pos
, poison
, PAGE_SIZE
);
6370 free_reserved_page(virt_to_page(pos
));
6374 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6375 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6379 EXPORT_SYMBOL(free_reserved_area
);
6381 #ifdef CONFIG_HIGHMEM
6382 void free_highmem_page(struct page
*page
)
6384 __free_reserved_page(page
);
6386 page_zone(page
)->managed_pages
++;
6392 void __init
mem_init_print_info(const char *str
)
6394 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6395 unsigned long init_code_size
, init_data_size
;
6397 physpages
= get_num_physpages();
6398 codesize
= _etext
- _stext
;
6399 datasize
= _edata
- _sdata
;
6400 rosize
= __end_rodata
- __start_rodata
;
6401 bss_size
= __bss_stop
- __bss_start
;
6402 init_data_size
= __init_end
- __init_begin
;
6403 init_code_size
= _einittext
- _sinittext
;
6406 * Detect special cases and adjust section sizes accordingly:
6407 * 1) .init.* may be embedded into .data sections
6408 * 2) .init.text.* may be out of [__init_begin, __init_end],
6409 * please refer to arch/tile/kernel/vmlinux.lds.S.
6410 * 3) .rodata.* may be embedded into .text or .data sections.
6412 #define adj_init_size(start, end, size, pos, adj) \
6414 if (start <= pos && pos < end && size > adj) \
6418 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6419 _sinittext
, init_code_size
);
6420 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6421 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6422 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6423 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6425 #undef adj_init_size
6427 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6428 #ifdef CONFIG_HIGHMEM
6432 nr_free_pages() << (PAGE_SHIFT
- 10),
6433 physpages
<< (PAGE_SHIFT
- 10),
6434 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6435 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6436 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6437 totalcma_pages
<< (PAGE_SHIFT
- 10),
6438 #ifdef CONFIG_HIGHMEM
6439 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6441 str
? ", " : "", str
? str
: "");
6445 * set_dma_reserve - set the specified number of pages reserved in the first zone
6446 * @new_dma_reserve: The number of pages to mark reserved
6448 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6449 * In the DMA zone, a significant percentage may be consumed by kernel image
6450 * and other unfreeable allocations which can skew the watermarks badly. This
6451 * function may optionally be used to account for unfreeable pages in the
6452 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6453 * smaller per-cpu batchsize.
6455 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6457 dma_reserve
= new_dma_reserve
;
6460 void __init
free_area_init(unsigned long *zones_size
)
6462 free_area_init_node(0, zones_size
,
6463 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6466 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6467 unsigned long action
, void *hcpu
)
6469 int cpu
= (unsigned long)hcpu
;
6471 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6472 lru_add_drain_cpu(cpu
);
6476 * Spill the event counters of the dead processor
6477 * into the current processors event counters.
6478 * This artificially elevates the count of the current
6481 vm_events_fold_cpu(cpu
);
6484 * Zero the differential counters of the dead processor
6485 * so that the vm statistics are consistent.
6487 * This is only okay since the processor is dead and cannot
6488 * race with what we are doing.
6490 cpu_vm_stats_fold(cpu
);
6495 void __init
page_alloc_init(void)
6497 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6501 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6502 * or min_free_kbytes changes.
6504 static void calculate_totalreserve_pages(void)
6506 struct pglist_data
*pgdat
;
6507 unsigned long reserve_pages
= 0;
6508 enum zone_type i
, j
;
6510 for_each_online_pgdat(pgdat
) {
6511 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6512 struct zone
*zone
= pgdat
->node_zones
+ i
;
6515 /* Find valid and maximum lowmem_reserve in the zone */
6516 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6517 if (zone
->lowmem_reserve
[j
] > max
)
6518 max
= zone
->lowmem_reserve
[j
];
6521 /* we treat the high watermark as reserved pages. */
6522 max
+= high_wmark_pages(zone
);
6524 if (max
> zone
->managed_pages
)
6525 max
= zone
->managed_pages
;
6527 zone
->totalreserve_pages
= max
;
6529 reserve_pages
+= max
;
6532 totalreserve_pages
= reserve_pages
;
6536 * setup_per_zone_lowmem_reserve - called whenever
6537 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6538 * has a correct pages reserved value, so an adequate number of
6539 * pages are left in the zone after a successful __alloc_pages().
6541 static void setup_per_zone_lowmem_reserve(void)
6543 struct pglist_data
*pgdat
;
6544 enum zone_type j
, idx
;
6546 for_each_online_pgdat(pgdat
) {
6547 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6548 struct zone
*zone
= pgdat
->node_zones
+ j
;
6549 unsigned long managed_pages
= zone
->managed_pages
;
6551 zone
->lowmem_reserve
[j
] = 0;
6555 struct zone
*lower_zone
;
6559 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6560 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6562 lower_zone
= pgdat
->node_zones
+ idx
;
6563 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6564 sysctl_lowmem_reserve_ratio
[idx
];
6565 managed_pages
+= lower_zone
->managed_pages
;
6570 /* update totalreserve_pages */
6571 calculate_totalreserve_pages();
6574 static void __setup_per_zone_wmarks(void)
6576 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6577 unsigned long lowmem_pages
= 0;
6579 unsigned long flags
;
6581 /* Calculate total number of !ZONE_HIGHMEM pages */
6582 for_each_zone(zone
) {
6583 if (!is_highmem(zone
))
6584 lowmem_pages
+= zone
->managed_pages
;
6587 for_each_zone(zone
) {
6590 spin_lock_irqsave(&zone
->lock
, flags
);
6591 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6592 do_div(tmp
, lowmem_pages
);
6593 if (is_highmem(zone
)) {
6595 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6596 * need highmem pages, so cap pages_min to a small
6599 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6600 * deltas control asynch page reclaim, and so should
6601 * not be capped for highmem.
6603 unsigned long min_pages
;
6605 min_pages
= zone
->managed_pages
/ 1024;
6606 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6607 zone
->watermark
[WMARK_MIN
] = min_pages
;
6610 * If it's a lowmem zone, reserve a number of pages
6611 * proportionate to the zone's size.
6613 zone
->watermark
[WMARK_MIN
] = tmp
;
6617 * Set the kswapd watermarks distance according to the
6618 * scale factor in proportion to available memory, but
6619 * ensure a minimum size on small systems.
6621 tmp
= max_t(u64
, tmp
>> 2,
6622 mult_frac(zone
->managed_pages
,
6623 watermark_scale_factor
, 10000));
6625 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6626 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6628 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
6629 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
6630 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
6632 spin_unlock_irqrestore(&zone
->lock
, flags
);
6635 /* update totalreserve_pages */
6636 calculate_totalreserve_pages();
6640 * setup_per_zone_wmarks - called when min_free_kbytes changes
6641 * or when memory is hot-{added|removed}
6643 * Ensures that the watermark[min,low,high] values for each zone are set
6644 * correctly with respect to min_free_kbytes.
6646 void setup_per_zone_wmarks(void)
6648 mutex_lock(&zonelists_mutex
);
6649 __setup_per_zone_wmarks();
6650 mutex_unlock(&zonelists_mutex
);
6654 * The inactive anon list should be small enough that the VM never has to
6655 * do too much work, but large enough that each inactive page has a chance
6656 * to be referenced again before it is swapped out.
6658 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6659 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6660 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6661 * the anonymous pages are kept on the inactive list.
6664 * memory ratio inactive anon
6665 * -------------------------------------
6674 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
6676 unsigned int gb
, ratio
;
6678 /* Zone size in gigabytes */
6679 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
6681 ratio
= int_sqrt(10 * gb
);
6685 zone
->inactive_ratio
= ratio
;
6688 static void __meminit
setup_per_zone_inactive_ratio(void)
6693 calculate_zone_inactive_ratio(zone
);
6697 * Initialise min_free_kbytes.
6699 * For small machines we want it small (128k min). For large machines
6700 * we want it large (64MB max). But it is not linear, because network
6701 * bandwidth does not increase linearly with machine size. We use
6703 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6704 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6720 int __meminit
init_per_zone_wmark_min(void)
6722 unsigned long lowmem_kbytes
;
6723 int new_min_free_kbytes
;
6725 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6726 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6728 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6729 min_free_kbytes
= new_min_free_kbytes
;
6730 if (min_free_kbytes
< 128)
6731 min_free_kbytes
= 128;
6732 if (min_free_kbytes
> 65536)
6733 min_free_kbytes
= 65536;
6735 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6736 new_min_free_kbytes
, user_min_free_kbytes
);
6738 setup_per_zone_wmarks();
6739 refresh_zone_stat_thresholds();
6740 setup_per_zone_lowmem_reserve();
6741 setup_per_zone_inactive_ratio();
6744 core_initcall(init_per_zone_wmark_min
)
6747 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6748 * that we can call two helper functions whenever min_free_kbytes
6751 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6752 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6756 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6761 user_min_free_kbytes
= min_free_kbytes
;
6762 setup_per_zone_wmarks();
6767 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6768 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6772 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6777 setup_per_zone_wmarks();
6783 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6784 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6789 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6794 zone
->min_unmapped_pages
= (zone
->managed_pages
*
6795 sysctl_min_unmapped_ratio
) / 100;
6799 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6800 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6805 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6810 zone
->min_slab_pages
= (zone
->managed_pages
*
6811 sysctl_min_slab_ratio
) / 100;
6817 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6818 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6819 * whenever sysctl_lowmem_reserve_ratio changes.
6821 * The reserve ratio obviously has absolutely no relation with the
6822 * minimum watermarks. The lowmem reserve ratio can only make sense
6823 * if in function of the boot time zone sizes.
6825 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6826 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6828 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6829 setup_per_zone_lowmem_reserve();
6834 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6835 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6836 * pagelist can have before it gets flushed back to buddy allocator.
6838 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6839 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6842 int old_percpu_pagelist_fraction
;
6845 mutex_lock(&pcp_batch_high_lock
);
6846 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6848 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6849 if (!write
|| ret
< 0)
6852 /* Sanity checking to avoid pcp imbalance */
6853 if (percpu_pagelist_fraction
&&
6854 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6855 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6861 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6864 for_each_populated_zone(zone
) {
6867 for_each_possible_cpu(cpu
)
6868 pageset_set_high_and_batch(zone
,
6869 per_cpu_ptr(zone
->pageset
, cpu
));
6872 mutex_unlock(&pcp_batch_high_lock
);
6877 int hashdist
= HASHDIST_DEFAULT
;
6879 static int __init
set_hashdist(char *str
)
6883 hashdist
= simple_strtoul(str
, &str
, 0);
6886 __setup("hashdist=", set_hashdist
);
6890 * allocate a large system hash table from bootmem
6891 * - it is assumed that the hash table must contain an exact power-of-2
6892 * quantity of entries
6893 * - limit is the number of hash buckets, not the total allocation size
6895 void *__init
alloc_large_system_hash(const char *tablename
,
6896 unsigned long bucketsize
,
6897 unsigned long numentries
,
6900 unsigned int *_hash_shift
,
6901 unsigned int *_hash_mask
,
6902 unsigned long low_limit
,
6903 unsigned long high_limit
)
6905 unsigned long long max
= high_limit
;
6906 unsigned long log2qty
, size
;
6909 /* allow the kernel cmdline to have a say */
6911 /* round applicable memory size up to nearest megabyte */
6912 numentries
= nr_kernel_pages
;
6914 /* It isn't necessary when PAGE_SIZE >= 1MB */
6915 if (PAGE_SHIFT
< 20)
6916 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6918 /* limit to 1 bucket per 2^scale bytes of low memory */
6919 if (scale
> PAGE_SHIFT
)
6920 numentries
>>= (scale
- PAGE_SHIFT
);
6922 numentries
<<= (PAGE_SHIFT
- scale
);
6924 /* Make sure we've got at least a 0-order allocation.. */
6925 if (unlikely(flags
& HASH_SMALL
)) {
6926 /* Makes no sense without HASH_EARLY */
6927 WARN_ON(!(flags
& HASH_EARLY
));
6928 if (!(numentries
>> *_hash_shift
)) {
6929 numentries
= 1UL << *_hash_shift
;
6930 BUG_ON(!numentries
);
6932 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6933 numentries
= PAGE_SIZE
/ bucketsize
;
6935 numentries
= roundup_pow_of_two(numentries
);
6937 /* limit allocation size to 1/16 total memory by default */
6939 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6940 do_div(max
, bucketsize
);
6942 max
= min(max
, 0x80000000ULL
);
6944 if (numentries
< low_limit
)
6945 numentries
= low_limit
;
6946 if (numentries
> max
)
6949 log2qty
= ilog2(numentries
);
6952 size
= bucketsize
<< log2qty
;
6953 if (flags
& HASH_EARLY
)
6954 table
= memblock_virt_alloc_nopanic(size
, 0);
6956 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6959 * If bucketsize is not a power-of-two, we may free
6960 * some pages at the end of hash table which
6961 * alloc_pages_exact() automatically does
6963 if (get_order(size
) < MAX_ORDER
) {
6964 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6965 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6968 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6971 panic("Failed to allocate %s hash table\n", tablename
);
6973 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6974 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
6977 *_hash_shift
= log2qty
;
6979 *_hash_mask
= (1 << log2qty
) - 1;
6985 * This function checks whether pageblock includes unmovable pages or not.
6986 * If @count is not zero, it is okay to include less @count unmovable pages
6988 * PageLRU check without isolation or lru_lock could race so that
6989 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6990 * expect this function should be exact.
6992 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6993 bool skip_hwpoisoned_pages
)
6995 unsigned long pfn
, iter
, found
;
6999 * For avoiding noise data, lru_add_drain_all() should be called
7000 * If ZONE_MOVABLE, the zone never contains unmovable pages
7002 if (zone_idx(zone
) == ZONE_MOVABLE
)
7004 mt
= get_pageblock_migratetype(page
);
7005 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7008 pfn
= page_to_pfn(page
);
7009 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7010 unsigned long check
= pfn
+ iter
;
7012 if (!pfn_valid_within(check
))
7015 page
= pfn_to_page(check
);
7018 * Hugepages are not in LRU lists, but they're movable.
7019 * We need not scan over tail pages bacause we don't
7020 * handle each tail page individually in migration.
7022 if (PageHuge(page
)) {
7023 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7028 * We can't use page_count without pin a page
7029 * because another CPU can free compound page.
7030 * This check already skips compound tails of THP
7031 * because their page->_refcount is zero at all time.
7033 if (!page_ref_count(page
)) {
7034 if (PageBuddy(page
))
7035 iter
+= (1 << page_order(page
)) - 1;
7040 * The HWPoisoned page may be not in buddy system, and
7041 * page_count() is not 0.
7043 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7049 * If there are RECLAIMABLE pages, we need to check
7050 * it. But now, memory offline itself doesn't call
7051 * shrink_node_slabs() and it still to be fixed.
7054 * If the page is not RAM, page_count()should be 0.
7055 * we don't need more check. This is an _used_ not-movable page.
7057 * The problematic thing here is PG_reserved pages. PG_reserved
7058 * is set to both of a memory hole page and a _used_ kernel
7067 bool is_pageblock_removable_nolock(struct page
*page
)
7073 * We have to be careful here because we are iterating over memory
7074 * sections which are not zone aware so we might end up outside of
7075 * the zone but still within the section.
7076 * We have to take care about the node as well. If the node is offline
7077 * its NODE_DATA will be NULL - see page_zone.
7079 if (!node_online(page_to_nid(page
)))
7082 zone
= page_zone(page
);
7083 pfn
= page_to_pfn(page
);
7084 if (!zone_spans_pfn(zone
, pfn
))
7087 return !has_unmovable_pages(zone
, page
, 0, true);
7090 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7092 static unsigned long pfn_max_align_down(unsigned long pfn
)
7094 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7095 pageblock_nr_pages
) - 1);
7098 static unsigned long pfn_max_align_up(unsigned long pfn
)
7100 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7101 pageblock_nr_pages
));
7104 /* [start, end) must belong to a single zone. */
7105 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7106 unsigned long start
, unsigned long end
)
7108 /* This function is based on compact_zone() from compaction.c. */
7109 unsigned long nr_reclaimed
;
7110 unsigned long pfn
= start
;
7111 unsigned int tries
= 0;
7116 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7117 if (fatal_signal_pending(current
)) {
7122 if (list_empty(&cc
->migratepages
)) {
7123 cc
->nr_migratepages
= 0;
7124 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7130 } else if (++tries
== 5) {
7131 ret
= ret
< 0 ? ret
: -EBUSY
;
7135 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7137 cc
->nr_migratepages
-= nr_reclaimed
;
7139 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7140 NULL
, 0, cc
->mode
, MR_CMA
);
7143 putback_movable_pages(&cc
->migratepages
);
7150 * alloc_contig_range() -- tries to allocate given range of pages
7151 * @start: start PFN to allocate
7152 * @end: one-past-the-last PFN to allocate
7153 * @migratetype: migratetype of the underlaying pageblocks (either
7154 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7155 * in range must have the same migratetype and it must
7156 * be either of the two.
7158 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7159 * aligned, however it's the caller's responsibility to guarantee that
7160 * we are the only thread that changes migrate type of pageblocks the
7163 * The PFN range must belong to a single zone.
7165 * Returns zero on success or negative error code. On success all
7166 * pages which PFN is in [start, end) are allocated for the caller and
7167 * need to be freed with free_contig_range().
7169 int alloc_contig_range(unsigned long start
, unsigned long end
,
7170 unsigned migratetype
)
7172 unsigned long outer_start
, outer_end
;
7176 struct compact_control cc
= {
7177 .nr_migratepages
= 0,
7179 .zone
= page_zone(pfn_to_page(start
)),
7180 .mode
= MIGRATE_SYNC
,
7181 .ignore_skip_hint
= true,
7183 INIT_LIST_HEAD(&cc
.migratepages
);
7186 * What we do here is we mark all pageblocks in range as
7187 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7188 * have different sizes, and due to the way page allocator
7189 * work, we align the range to biggest of the two pages so
7190 * that page allocator won't try to merge buddies from
7191 * different pageblocks and change MIGRATE_ISOLATE to some
7192 * other migration type.
7194 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7195 * migrate the pages from an unaligned range (ie. pages that
7196 * we are interested in). This will put all the pages in
7197 * range back to page allocator as MIGRATE_ISOLATE.
7199 * When this is done, we take the pages in range from page
7200 * allocator removing them from the buddy system. This way
7201 * page allocator will never consider using them.
7203 * This lets us mark the pageblocks back as
7204 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7205 * aligned range but not in the unaligned, original range are
7206 * put back to page allocator so that buddy can use them.
7209 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7210 pfn_max_align_up(end
), migratetype
,
7216 * In case of -EBUSY, we'd like to know which page causes problem.
7217 * So, just fall through. We will check it in test_pages_isolated().
7219 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7220 if (ret
&& ret
!= -EBUSY
)
7224 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7225 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7226 * more, all pages in [start, end) are free in page allocator.
7227 * What we are going to do is to allocate all pages from
7228 * [start, end) (that is remove them from page allocator).
7230 * The only problem is that pages at the beginning and at the
7231 * end of interesting range may be not aligned with pages that
7232 * page allocator holds, ie. they can be part of higher order
7233 * pages. Because of this, we reserve the bigger range and
7234 * once this is done free the pages we are not interested in.
7236 * We don't have to hold zone->lock here because the pages are
7237 * isolated thus they won't get removed from buddy.
7240 lru_add_drain_all();
7241 drain_all_pages(cc
.zone
);
7244 outer_start
= start
;
7245 while (!PageBuddy(pfn_to_page(outer_start
))) {
7246 if (++order
>= MAX_ORDER
) {
7247 outer_start
= start
;
7250 outer_start
&= ~0UL << order
;
7253 if (outer_start
!= start
) {
7254 order
= page_order(pfn_to_page(outer_start
));
7257 * outer_start page could be small order buddy page and
7258 * it doesn't include start page. Adjust outer_start
7259 * in this case to report failed page properly
7260 * on tracepoint in test_pages_isolated()
7262 if (outer_start
+ (1UL << order
) <= start
)
7263 outer_start
= start
;
7266 /* Make sure the range is really isolated. */
7267 if (test_pages_isolated(outer_start
, end
, false)) {
7268 pr_info("%s: [%lx, %lx) PFNs busy\n",
7269 __func__
, outer_start
, end
);
7274 /* Grab isolated pages from freelists. */
7275 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7281 /* Free head and tail (if any) */
7282 if (start
!= outer_start
)
7283 free_contig_range(outer_start
, start
- outer_start
);
7284 if (end
!= outer_end
)
7285 free_contig_range(end
, outer_end
- end
);
7288 undo_isolate_page_range(pfn_max_align_down(start
),
7289 pfn_max_align_up(end
), migratetype
);
7293 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7295 unsigned int count
= 0;
7297 for (; nr_pages
--; pfn
++) {
7298 struct page
*page
= pfn_to_page(pfn
);
7300 count
+= page_count(page
) != 1;
7303 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7307 #ifdef CONFIG_MEMORY_HOTPLUG
7309 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7310 * page high values need to be recalulated.
7312 void __meminit
zone_pcp_update(struct zone
*zone
)
7315 mutex_lock(&pcp_batch_high_lock
);
7316 for_each_possible_cpu(cpu
)
7317 pageset_set_high_and_batch(zone
,
7318 per_cpu_ptr(zone
->pageset
, cpu
));
7319 mutex_unlock(&pcp_batch_high_lock
);
7323 void zone_pcp_reset(struct zone
*zone
)
7325 unsigned long flags
;
7327 struct per_cpu_pageset
*pset
;
7329 /* avoid races with drain_pages() */
7330 local_irq_save(flags
);
7331 if (zone
->pageset
!= &boot_pageset
) {
7332 for_each_online_cpu(cpu
) {
7333 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7334 drain_zonestat(zone
, pset
);
7336 free_percpu(zone
->pageset
);
7337 zone
->pageset
= &boot_pageset
;
7339 local_irq_restore(flags
);
7342 #ifdef CONFIG_MEMORY_HOTREMOVE
7344 * All pages in the range must be in a single zone and isolated
7345 * before calling this.
7348 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7352 unsigned int order
, i
;
7354 unsigned long flags
;
7355 /* find the first valid pfn */
7356 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7361 zone
= page_zone(pfn_to_page(pfn
));
7362 spin_lock_irqsave(&zone
->lock
, flags
);
7364 while (pfn
< end_pfn
) {
7365 if (!pfn_valid(pfn
)) {
7369 page
= pfn_to_page(pfn
);
7371 * The HWPoisoned page may be not in buddy system, and
7372 * page_count() is not 0.
7374 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7376 SetPageReserved(page
);
7380 BUG_ON(page_count(page
));
7381 BUG_ON(!PageBuddy(page
));
7382 order
= page_order(page
);
7383 #ifdef CONFIG_DEBUG_VM
7384 pr_info("remove from free list %lx %d %lx\n",
7385 pfn
, 1 << order
, end_pfn
);
7387 list_del(&page
->lru
);
7388 rmv_page_order(page
);
7389 zone
->free_area
[order
].nr_free
--;
7390 for (i
= 0; i
< (1 << order
); i
++)
7391 SetPageReserved((page
+i
));
7392 pfn
+= (1 << order
);
7394 spin_unlock_irqrestore(&zone
->lock
, flags
);
7398 bool is_free_buddy_page(struct page
*page
)
7400 struct zone
*zone
= page_zone(page
);
7401 unsigned long pfn
= page_to_pfn(page
);
7402 unsigned long flags
;
7405 spin_lock_irqsave(&zone
->lock
, flags
);
7406 for (order
= 0; order
< MAX_ORDER
; order
++) {
7407 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7409 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7412 spin_unlock_irqrestore(&zone
->lock
, flags
);
7414 return order
< MAX_ORDER
;