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>
66 #include <linux/memcontrol.h>
68 #include <asm/sections.h>
69 #include <asm/tlbflush.h>
70 #include <asm/div64.h>
73 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74 static DEFINE_MUTEX(pcp_batch_high_lock
);
75 #define MIN_PERCPU_PAGELIST_FRACTION (8)
77 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78 DEFINE_PER_CPU(int, numa_node
);
79 EXPORT_PER_CPU_SYMBOL(numa_node
);
82 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
89 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
90 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
91 int _node_numa_mem_
[MAX_NUMNODES
];
94 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
95 volatile u64 latent_entropy __latent_entropy
;
96 EXPORT_SYMBOL(latent_entropy
);
100 * Array of node states.
102 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
103 [N_POSSIBLE
] = NODE_MASK_ALL
,
104 [N_ONLINE
] = { { [0] = 1UL } },
106 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
107 #ifdef CONFIG_HIGHMEM
108 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
110 #ifdef CONFIG_MOVABLE_NODE
111 [N_MEMORY
] = { { [0] = 1UL } },
113 [N_CPU
] = { { [0] = 1UL } },
116 EXPORT_SYMBOL(node_states
);
118 /* Protect totalram_pages and zone->managed_pages */
119 static DEFINE_SPINLOCK(managed_page_count_lock
);
121 unsigned long totalram_pages __read_mostly
;
122 unsigned long totalreserve_pages __read_mostly
;
123 unsigned long totalcma_pages __read_mostly
;
125 int percpu_pagelist_fraction
;
126 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page
*page
)
141 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
143 page
->index
= migratetype
;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask
;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex
));
161 if (saved_gfp_mask
) {
162 gfp_allowed_mask
= saved_gfp_mask
;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex
));
170 WARN_ON(saved_gfp_mask
);
171 saved_gfp_mask
= gfp_allowed_mask
;
172 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly
;
187 static void __free_pages_ok(struct page
*page
, unsigned int order
);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages
);
215 static char * const zone_names
[MAX_NR_ZONES
] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 char * const migratetype_names
[MIGRATE_TYPES
] = {
240 #ifdef CONFIG_MEMORY_ISOLATION
245 compound_page_dtor
* const compound_page_dtors
[] = {
248 #ifdef CONFIG_HUGETLB_PAGE
251 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
256 int min_free_kbytes
= 1024;
257 int user_min_free_kbytes
= -1;
258 int watermark_scale_factor
= 10;
260 static unsigned long __meminitdata nr_kernel_pages
;
261 static unsigned long __meminitdata nr_all_pages
;
262 static unsigned long __meminitdata dma_reserve
;
264 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
265 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
266 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
267 static unsigned long __initdata required_kernelcore
;
268 static unsigned long __initdata required_movablecore
;
269 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
270 static bool mirrored_kernelcore
;
272 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
274 EXPORT_SYMBOL(movable_zone
);
275 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
278 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
279 int nr_online_nodes __read_mostly
= 1;
280 EXPORT_SYMBOL(nr_node_ids
);
281 EXPORT_SYMBOL(nr_online_nodes
);
284 int page_group_by_mobility_disabled __read_mostly
;
286 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
287 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
289 pgdat
->first_deferred_pfn
= ULONG_MAX
;
292 /* Returns true if the struct page for the pfn is uninitialised */
293 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
295 int nid
= early_pfn_to_nid(pfn
);
297 if (node_online(nid
) && 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 update_defer_init(pg_data_t
*pgdat
,
343 unsigned long pfn
, unsigned long zone_end
,
344 unsigned long *nr_initialised
)
350 /* Return a pointer to the bitmap storing bits affecting a block of pages */
351 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
354 #ifdef CONFIG_SPARSEMEM
355 return __pfn_to_section(pfn
)->pageblock_flags
;
357 return page_zone(page
)->pageblock_flags
;
358 #endif /* CONFIG_SPARSEMEM */
361 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
363 #ifdef CONFIG_SPARSEMEM
364 pfn
&= (PAGES_PER_SECTION
-1);
365 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
367 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
368 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
369 #endif /* CONFIG_SPARSEMEM */
373 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
374 * @page: The page within the block of interest
375 * @pfn: The target page frame number
376 * @end_bitidx: The last bit of interest to retrieve
377 * @mask: mask of bits that the caller is interested in
379 * Return: pageblock_bits flags
381 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
383 unsigned long end_bitidx
,
386 unsigned long *bitmap
;
387 unsigned long bitidx
, word_bitidx
;
390 bitmap
= get_pageblock_bitmap(page
, pfn
);
391 bitidx
= pfn_to_bitidx(page
, pfn
);
392 word_bitidx
= bitidx
/ BITS_PER_LONG
;
393 bitidx
&= (BITS_PER_LONG
-1);
395 word
= bitmap
[word_bitidx
];
396 bitidx
+= end_bitidx
;
397 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
400 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
401 unsigned long end_bitidx
,
404 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
407 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
409 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
413 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
414 * @page: The page within the block of interest
415 * @flags: The flags to set
416 * @pfn: The target page frame number
417 * @end_bitidx: The last bit of interest
418 * @mask: mask of bits that the caller is interested in
420 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
422 unsigned long end_bitidx
,
425 unsigned long *bitmap
;
426 unsigned long bitidx
, word_bitidx
;
427 unsigned long old_word
, word
;
429 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
431 bitmap
= get_pageblock_bitmap(page
, pfn
);
432 bitidx
= pfn_to_bitidx(page
, pfn
);
433 word_bitidx
= bitidx
/ BITS_PER_LONG
;
434 bitidx
&= (BITS_PER_LONG
-1);
436 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
438 bitidx
+= end_bitidx
;
439 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
440 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
442 word
= READ_ONCE(bitmap
[word_bitidx
]);
444 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
445 if (word
== old_word
)
451 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
453 if (unlikely(page_group_by_mobility_disabled
&&
454 migratetype
< MIGRATE_PCPTYPES
))
455 migratetype
= MIGRATE_UNMOVABLE
;
457 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
458 PB_migrate
, PB_migrate_end
);
461 #ifdef CONFIG_DEBUG_VM
462 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
466 unsigned long pfn
= page_to_pfn(page
);
467 unsigned long sp
, start_pfn
;
470 seq
= zone_span_seqbegin(zone
);
471 start_pfn
= zone
->zone_start_pfn
;
472 sp
= zone
->spanned_pages
;
473 if (!zone_spans_pfn(zone
, pfn
))
475 } while (zone_span_seqretry(zone
, seq
));
478 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
479 pfn
, zone_to_nid(zone
), zone
->name
,
480 start_pfn
, start_pfn
+ sp
);
485 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
487 if (!pfn_valid_within(page_to_pfn(page
)))
489 if (zone
!= page_zone(page
))
495 * Temporary debugging check for pages not lying within a given zone.
497 static int bad_range(struct zone
*zone
, struct page
*page
)
499 if (page_outside_zone_boundaries(zone
, page
))
501 if (!page_is_consistent(zone
, page
))
507 static inline int bad_range(struct zone
*zone
, struct page
*page
)
513 static void bad_page(struct page
*page
, const char *reason
,
514 unsigned long bad_flags
)
516 static unsigned long resume
;
517 static unsigned long nr_shown
;
518 static unsigned long nr_unshown
;
521 * Allow a burst of 60 reports, then keep quiet for that minute;
522 * or allow a steady drip of one report per second.
524 if (nr_shown
== 60) {
525 if (time_before(jiffies
, resume
)) {
531 "BUG: Bad page state: %lu messages suppressed\n",
538 resume
= jiffies
+ 60 * HZ
;
540 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
541 current
->comm
, page_to_pfn(page
));
542 __dump_page(page
, reason
);
543 bad_flags
&= page
->flags
;
545 pr_alert("bad because of flags: %#lx(%pGp)\n",
546 bad_flags
, &bad_flags
);
547 dump_page_owner(page
);
552 /* Leave bad fields for debug, except PageBuddy could make trouble */
553 page_mapcount_reset(page
); /* remove PageBuddy */
554 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
558 * Higher-order pages are called "compound pages". They are structured thusly:
560 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
562 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
563 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
565 * The first tail page's ->compound_dtor holds the offset in array of compound
566 * page destructors. See compound_page_dtors.
568 * The first tail page's ->compound_order holds the order of allocation.
569 * This usage means that zero-order pages may not be compound.
572 void free_compound_page(struct page
*page
)
574 __free_pages_ok(page
, compound_order(page
));
577 void prep_compound_page(struct page
*page
, unsigned int order
)
580 int nr_pages
= 1 << order
;
582 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
583 set_compound_order(page
, order
);
585 for (i
= 1; i
< nr_pages
; i
++) {
586 struct page
*p
= page
+ i
;
587 set_page_count(p
, 0);
588 p
->mapping
= TAIL_MAPPING
;
589 set_compound_head(p
, page
);
591 atomic_set(compound_mapcount_ptr(page
), -1);
594 #ifdef CONFIG_DEBUG_PAGEALLOC
595 unsigned int _debug_guardpage_minorder
;
596 bool _debug_pagealloc_enabled __read_mostly
597 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
598 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
599 bool _debug_guardpage_enabled __read_mostly
;
601 static int __init
early_debug_pagealloc(char *buf
)
605 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
607 early_param("debug_pagealloc", early_debug_pagealloc
);
609 static bool need_debug_guardpage(void)
611 /* If we don't use debug_pagealloc, we don't need guard page */
612 if (!debug_pagealloc_enabled())
615 if (!debug_guardpage_minorder())
621 static void init_debug_guardpage(void)
623 if (!debug_pagealloc_enabled())
626 if (!debug_guardpage_minorder())
629 _debug_guardpage_enabled
= true;
632 struct page_ext_operations debug_guardpage_ops
= {
633 .need
= need_debug_guardpage
,
634 .init
= init_debug_guardpage
,
637 static int __init
debug_guardpage_minorder_setup(char *buf
)
641 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
642 pr_err("Bad debug_guardpage_minorder value\n");
645 _debug_guardpage_minorder
= res
;
646 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
649 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
651 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
652 unsigned int order
, int migratetype
)
654 struct page_ext
*page_ext
;
656 if (!debug_guardpage_enabled())
659 if (order
>= debug_guardpage_minorder())
662 page_ext
= lookup_page_ext(page
);
663 if (unlikely(!page_ext
))
666 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
668 INIT_LIST_HEAD(&page
->lru
);
669 set_page_private(page
, order
);
670 /* Guard pages are not available for any usage */
671 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
676 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
677 unsigned int order
, int migratetype
)
679 struct page_ext
*page_ext
;
681 if (!debug_guardpage_enabled())
684 page_ext
= lookup_page_ext(page
);
685 if (unlikely(!page_ext
))
688 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
690 set_page_private(page
, 0);
691 if (!is_migrate_isolate(migratetype
))
692 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
695 struct page_ext_operations debug_guardpage_ops
;
696 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
697 unsigned int order
, int migratetype
) { return false; }
698 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
699 unsigned int order
, int migratetype
) {}
702 static inline void set_page_order(struct page
*page
, unsigned int order
)
704 set_page_private(page
, order
);
705 __SetPageBuddy(page
);
708 static inline void rmv_page_order(struct page
*page
)
710 __ClearPageBuddy(page
);
711 set_page_private(page
, 0);
715 * This function checks whether a page is free && is the buddy
716 * we can do coalesce a page and its buddy if
717 * (a) the buddy is not in a hole &&
718 * (b) the buddy is in the buddy system &&
719 * (c) a page and its buddy have the same order &&
720 * (d) a page and its buddy are in the same zone.
722 * For recording whether a page is in the buddy system, we set ->_mapcount
723 * PAGE_BUDDY_MAPCOUNT_VALUE.
724 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
725 * serialized by zone->lock.
727 * For recording page's order, we use page_private(page).
729 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
732 if (!pfn_valid_within(page_to_pfn(buddy
)))
735 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
736 if (page_zone_id(page
) != page_zone_id(buddy
))
739 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
744 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
746 * zone check is done late to avoid uselessly
747 * calculating zone/node ids for pages that could
750 if (page_zone_id(page
) != page_zone_id(buddy
))
753 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
761 * Freeing function for a buddy system allocator.
763 * The concept of a buddy system is to maintain direct-mapped table
764 * (containing bit values) for memory blocks of various "orders".
765 * The bottom level table contains the map for the smallest allocatable
766 * units of memory (here, pages), and each level above it describes
767 * pairs of units from the levels below, hence, "buddies".
768 * At a high level, all that happens here is marking the table entry
769 * at the bottom level available, and propagating the changes upward
770 * as necessary, plus some accounting needed to play nicely with other
771 * parts of the VM system.
772 * At each level, we keep a list of pages, which are heads of continuous
773 * free pages of length of (1 << order) and marked with _mapcount
774 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
776 * So when we are allocating or freeing one, we can derive the state of the
777 * other. That is, if we allocate a small block, and both were
778 * free, the remainder of the region must be split into blocks.
779 * If a block is freed, and its buddy is also free, then this
780 * triggers coalescing into a block of larger size.
785 static inline void __free_one_page(struct page
*page
,
787 struct zone
*zone
, unsigned int order
,
790 unsigned long page_idx
;
791 unsigned long combined_idx
;
792 unsigned long uninitialized_var(buddy_idx
);
794 unsigned int max_order
;
796 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
798 VM_BUG_ON(!zone_is_initialized(zone
));
799 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
801 VM_BUG_ON(migratetype
== -1);
802 if (likely(!is_migrate_isolate(migratetype
)))
803 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
805 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
807 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
808 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
811 while (order
< max_order
- 1) {
812 buddy_idx
= __find_buddy_index(page_idx
, order
);
813 buddy
= page
+ (buddy_idx
- page_idx
);
814 if (!page_is_buddy(page
, buddy
, order
))
817 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
818 * merge with it and move up one order.
820 if (page_is_guard(buddy
)) {
821 clear_page_guard(zone
, buddy
, order
, migratetype
);
823 list_del(&buddy
->lru
);
824 zone
->free_area
[order
].nr_free
--;
825 rmv_page_order(buddy
);
827 combined_idx
= buddy_idx
& page_idx
;
828 page
= page
+ (combined_idx
- page_idx
);
829 page_idx
= combined_idx
;
832 if (max_order
< MAX_ORDER
) {
833 /* If we are here, it means order is >= pageblock_order.
834 * We want to prevent merge between freepages on isolate
835 * pageblock and normal pageblock. Without this, pageblock
836 * isolation could cause incorrect freepage or CMA accounting.
838 * We don't want to hit this code for the more frequent
841 if (unlikely(has_isolate_pageblock(zone
))) {
844 buddy_idx
= __find_buddy_index(page_idx
, order
);
845 buddy
= page
+ (buddy_idx
- page_idx
);
846 buddy_mt
= get_pageblock_migratetype(buddy
);
848 if (migratetype
!= buddy_mt
849 && (is_migrate_isolate(migratetype
) ||
850 is_migrate_isolate(buddy_mt
)))
854 goto continue_merging
;
858 set_page_order(page
, order
);
861 * If this is not the largest possible page, check if the buddy
862 * of the next-highest order is free. If it is, it's possible
863 * that pages are being freed that will coalesce soon. In case,
864 * that is happening, add the free page to the tail of the list
865 * so it's less likely to be used soon and more likely to be merged
866 * as a higher order page
868 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
869 struct page
*higher_page
, *higher_buddy
;
870 combined_idx
= buddy_idx
& page_idx
;
871 higher_page
= page
+ (combined_idx
- page_idx
);
872 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
873 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
874 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
875 list_add_tail(&page
->lru
,
876 &zone
->free_area
[order
].free_list
[migratetype
]);
881 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
883 zone
->free_area
[order
].nr_free
++;
887 * A bad page could be due to a number of fields. Instead of multiple branches,
888 * try and check multiple fields with one check. The caller must do a detailed
889 * check if necessary.
891 static inline bool page_expected_state(struct page
*page
,
892 unsigned long check_flags
)
894 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
897 if (unlikely((unsigned long)page
->mapping
|
898 page_ref_count(page
) |
900 (unsigned long)page
->mem_cgroup
|
902 (page
->flags
& check_flags
)))
908 static void free_pages_check_bad(struct page
*page
)
910 const char *bad_reason
;
911 unsigned long bad_flags
;
916 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
917 bad_reason
= "nonzero mapcount";
918 if (unlikely(page
->mapping
!= NULL
))
919 bad_reason
= "non-NULL mapping";
920 if (unlikely(page_ref_count(page
) != 0))
921 bad_reason
= "nonzero _refcount";
922 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
923 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
924 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
927 if (unlikely(page
->mem_cgroup
))
928 bad_reason
= "page still charged to cgroup";
930 bad_page(page
, bad_reason
, bad_flags
);
933 static inline int free_pages_check(struct page
*page
)
935 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
938 /* Something has gone sideways, find it */
939 free_pages_check_bad(page
);
943 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
948 * We rely page->lru.next never has bit 0 set, unless the page
949 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
951 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
953 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
957 switch (page
- head_page
) {
959 /* the first tail page: ->mapping is compound_mapcount() */
960 if (unlikely(compound_mapcount(page
))) {
961 bad_page(page
, "nonzero compound_mapcount", 0);
967 * the second tail page: ->mapping is
968 * page_deferred_list().next -- ignore value.
972 if (page
->mapping
!= TAIL_MAPPING
) {
973 bad_page(page
, "corrupted mapping in tail page", 0);
978 if (unlikely(!PageTail(page
))) {
979 bad_page(page
, "PageTail not set", 0);
982 if (unlikely(compound_head(page
) != head_page
)) {
983 bad_page(page
, "compound_head not consistent", 0);
988 page
->mapping
= NULL
;
989 clear_compound_head(page
);
993 static __always_inline
bool free_pages_prepare(struct page
*page
,
994 unsigned int order
, bool check_free
)
998 VM_BUG_ON_PAGE(PageTail(page
), page
);
1000 trace_mm_page_free(page
, order
);
1001 kmemcheck_free_shadow(page
, order
);
1004 * Check tail pages before head page information is cleared to
1005 * avoid checking PageCompound for order-0 pages.
1007 if (unlikely(order
)) {
1008 bool compound
= PageCompound(page
);
1011 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1014 ClearPageDoubleMap(page
);
1015 for (i
= 1; i
< (1 << order
); i
++) {
1017 bad
+= free_tail_pages_check(page
, page
+ i
);
1018 if (unlikely(free_pages_check(page
+ i
))) {
1022 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1025 if (PageMappingFlags(page
))
1026 page
->mapping
= NULL
;
1027 if (memcg_kmem_enabled() && PageKmemcg(page
))
1028 memcg_kmem_uncharge(page
, order
);
1030 bad
+= free_pages_check(page
);
1034 page_cpupid_reset_last(page
);
1035 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1036 reset_page_owner(page
, order
);
1038 if (!PageHighMem(page
)) {
1039 debug_check_no_locks_freed(page_address(page
),
1040 PAGE_SIZE
<< order
);
1041 debug_check_no_obj_freed(page_address(page
),
1042 PAGE_SIZE
<< order
);
1044 arch_free_page(page
, order
);
1045 kernel_poison_pages(page
, 1 << order
, 0);
1046 kernel_map_pages(page
, 1 << order
, 0);
1047 kasan_free_pages(page
, order
);
1052 #ifdef CONFIG_DEBUG_VM
1053 static inline bool free_pcp_prepare(struct page
*page
)
1055 return free_pages_prepare(page
, 0, true);
1058 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1063 static bool free_pcp_prepare(struct page
*page
)
1065 return free_pages_prepare(page
, 0, false);
1068 static bool bulkfree_pcp_prepare(struct page
*page
)
1070 return free_pages_check(page
);
1072 #endif /* CONFIG_DEBUG_VM */
1075 * Frees a number of pages from the PCP lists
1076 * Assumes all pages on list are in same zone, and of same order.
1077 * count is the number of pages to free.
1079 * If the zone was previously in an "all pages pinned" state then look to
1080 * see if this freeing clears that state.
1082 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1083 * pinned" detection logic.
1085 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1086 struct per_cpu_pages
*pcp
)
1088 int migratetype
= 0;
1090 unsigned long nr_scanned
;
1091 bool isolated_pageblocks
;
1093 spin_lock(&zone
->lock
);
1094 isolated_pageblocks
= has_isolate_pageblock(zone
);
1095 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1097 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1101 struct list_head
*list
;
1104 * Remove pages from lists in a round-robin fashion. A
1105 * batch_free count is maintained that is incremented when an
1106 * empty list is encountered. This is so more pages are freed
1107 * off fuller lists instead of spinning excessively around empty
1112 if (++migratetype
== MIGRATE_PCPTYPES
)
1114 list
= &pcp
->lists
[migratetype
];
1115 } while (list_empty(list
));
1117 /* This is the only non-empty list. Free them all. */
1118 if (batch_free
== MIGRATE_PCPTYPES
)
1122 int mt
; /* migratetype of the to-be-freed page */
1124 page
= list_last_entry(list
, struct page
, lru
);
1125 /* must delete as __free_one_page list manipulates */
1126 list_del(&page
->lru
);
1128 mt
= get_pcppage_migratetype(page
);
1129 /* MIGRATE_ISOLATE page should not go to pcplists */
1130 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1131 /* Pageblock could have been isolated meanwhile */
1132 if (unlikely(isolated_pageblocks
))
1133 mt
= get_pageblock_migratetype(page
);
1135 if (bulkfree_pcp_prepare(page
))
1138 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1139 trace_mm_page_pcpu_drain(page
, 0, mt
);
1140 } while (--count
&& --batch_free
&& !list_empty(list
));
1142 spin_unlock(&zone
->lock
);
1145 static void free_one_page(struct zone
*zone
,
1146 struct page
*page
, unsigned long pfn
,
1150 unsigned long nr_scanned
;
1151 spin_lock(&zone
->lock
);
1152 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1154 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1156 if (unlikely(has_isolate_pageblock(zone
) ||
1157 is_migrate_isolate(migratetype
))) {
1158 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1160 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1161 spin_unlock(&zone
->lock
);
1164 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1165 unsigned long zone
, int nid
)
1167 set_page_links(page
, zone
, nid
, pfn
);
1168 init_page_count(page
);
1169 page_mapcount_reset(page
);
1170 page_cpupid_reset_last(page
);
1172 INIT_LIST_HEAD(&page
->lru
);
1173 #ifdef WANT_PAGE_VIRTUAL
1174 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1175 if (!is_highmem_idx(zone
))
1176 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1180 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1183 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1186 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1187 static void init_reserved_page(unsigned long pfn
)
1192 if (!early_page_uninitialised(pfn
))
1195 nid
= early_pfn_to_nid(pfn
);
1196 pgdat
= NODE_DATA(nid
);
1198 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1199 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1201 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1204 __init_single_pfn(pfn
, zid
, nid
);
1207 static inline void init_reserved_page(unsigned long pfn
)
1210 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1213 * Initialised pages do not have PageReserved set. This function is
1214 * called for each range allocated by the bootmem allocator and
1215 * marks the pages PageReserved. The remaining valid pages are later
1216 * sent to the buddy page allocator.
1218 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1220 unsigned long start_pfn
= PFN_DOWN(start
);
1221 unsigned long end_pfn
= PFN_UP(end
);
1223 for (; start_pfn
< end_pfn
; start_pfn
++) {
1224 if (pfn_valid(start_pfn
)) {
1225 struct page
*page
= pfn_to_page(start_pfn
);
1227 init_reserved_page(start_pfn
);
1229 /* Avoid false-positive PageTail() */
1230 INIT_LIST_HEAD(&page
->lru
);
1232 SetPageReserved(page
);
1237 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1239 unsigned long flags
;
1241 unsigned long pfn
= page_to_pfn(page
);
1243 if (!free_pages_prepare(page
, order
, true))
1246 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1247 local_irq_save(flags
);
1248 __count_vm_events(PGFREE
, 1 << order
);
1249 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1250 local_irq_restore(flags
);
1253 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1255 unsigned int nr_pages
= 1 << order
;
1256 struct page
*p
= page
;
1260 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1262 __ClearPageReserved(p
);
1263 set_page_count(p
, 0);
1265 __ClearPageReserved(p
);
1266 set_page_count(p
, 0);
1268 page_zone(page
)->managed_pages
+= nr_pages
;
1269 set_page_refcounted(page
);
1270 __free_pages(page
, order
);
1273 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1274 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1276 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1278 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1280 static DEFINE_SPINLOCK(early_pfn_lock
);
1283 spin_lock(&early_pfn_lock
);
1284 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1286 nid
= first_online_node
;
1287 spin_unlock(&early_pfn_lock
);
1293 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1294 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1295 struct mminit_pfnnid_cache
*state
)
1299 nid
= __early_pfn_to_nid(pfn
, state
);
1300 if (nid
>= 0 && nid
!= node
)
1305 /* Only safe to use early in boot when initialisation is single-threaded */
1306 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1308 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1313 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1317 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1318 struct mminit_pfnnid_cache
*state
)
1325 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1328 if (early_page_uninitialised(pfn
))
1330 return __free_pages_boot_core(page
, order
);
1334 * Check that the whole (or subset of) a pageblock given by the interval of
1335 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1336 * with the migration of free compaction scanner. The scanners then need to
1337 * use only pfn_valid_within() check for arches that allow holes within
1340 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1342 * It's possible on some configurations to have a setup like node0 node1 node0
1343 * i.e. it's possible that all pages within a zones range of pages do not
1344 * belong to a single zone. We assume that a border between node0 and node1
1345 * can occur within a single pageblock, but not a node0 node1 node0
1346 * interleaving within a single pageblock. It is therefore sufficient to check
1347 * the first and last page of a pageblock and avoid checking each individual
1348 * page in a pageblock.
1350 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1351 unsigned long end_pfn
, struct zone
*zone
)
1353 struct page
*start_page
;
1354 struct page
*end_page
;
1356 /* end_pfn is one past the range we are checking */
1359 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1362 start_page
= pfn_to_page(start_pfn
);
1364 if (page_zone(start_page
) != zone
)
1367 end_page
= pfn_to_page(end_pfn
);
1369 /* This gives a shorter code than deriving page_zone(end_page) */
1370 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1376 void set_zone_contiguous(struct zone
*zone
)
1378 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1379 unsigned long block_end_pfn
;
1381 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1382 for (; block_start_pfn
< zone_end_pfn(zone
);
1383 block_start_pfn
= block_end_pfn
,
1384 block_end_pfn
+= pageblock_nr_pages
) {
1386 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1388 if (!__pageblock_pfn_to_page(block_start_pfn
,
1389 block_end_pfn
, zone
))
1393 /* We confirm that there is no hole */
1394 zone
->contiguous
= true;
1397 void clear_zone_contiguous(struct zone
*zone
)
1399 zone
->contiguous
= false;
1402 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1403 static void __init
deferred_free_range(struct page
*page
,
1404 unsigned long pfn
, int nr_pages
)
1411 /* Free a large naturally-aligned chunk if possible */
1412 if (nr_pages
== pageblock_nr_pages
&&
1413 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1414 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1415 __free_pages_boot_core(page
, pageblock_order
);
1419 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1420 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1421 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1422 __free_pages_boot_core(page
, 0);
1426 /* Completion tracking for deferred_init_memmap() threads */
1427 static atomic_t pgdat_init_n_undone __initdata
;
1428 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1430 static inline void __init
pgdat_init_report_one_done(void)
1432 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1433 complete(&pgdat_init_all_done_comp
);
1436 /* Initialise remaining memory on a node */
1437 static int __init
deferred_init_memmap(void *data
)
1439 pg_data_t
*pgdat
= data
;
1440 int nid
= pgdat
->node_id
;
1441 struct mminit_pfnnid_cache nid_init_state
= { };
1442 unsigned long start
= jiffies
;
1443 unsigned long nr_pages
= 0;
1444 unsigned long walk_start
, walk_end
;
1447 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1448 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1450 if (first_init_pfn
== ULONG_MAX
) {
1451 pgdat_init_report_one_done();
1455 /* Bind memory initialisation thread to a local node if possible */
1456 if (!cpumask_empty(cpumask
))
1457 set_cpus_allowed_ptr(current
, cpumask
);
1459 /* Sanity check boundaries */
1460 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1461 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1462 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1464 /* Only the highest zone is deferred so find it */
1465 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1466 zone
= pgdat
->node_zones
+ zid
;
1467 if (first_init_pfn
< zone_end_pfn(zone
))
1471 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1472 unsigned long pfn
, end_pfn
;
1473 struct page
*page
= NULL
;
1474 struct page
*free_base_page
= NULL
;
1475 unsigned long free_base_pfn
= 0;
1478 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1479 pfn
= first_init_pfn
;
1480 if (pfn
< walk_start
)
1482 if (pfn
< zone
->zone_start_pfn
)
1483 pfn
= zone
->zone_start_pfn
;
1485 for (; pfn
< end_pfn
; pfn
++) {
1486 if (!pfn_valid_within(pfn
))
1490 * Ensure pfn_valid is checked every
1491 * pageblock_nr_pages for memory holes
1493 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1494 if (!pfn_valid(pfn
)) {
1500 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1505 /* Minimise pfn page lookups and scheduler checks */
1506 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1509 nr_pages
+= nr_to_free
;
1510 deferred_free_range(free_base_page
,
1511 free_base_pfn
, nr_to_free
);
1512 free_base_page
= NULL
;
1513 free_base_pfn
= nr_to_free
= 0;
1515 page
= pfn_to_page(pfn
);
1520 VM_BUG_ON(page_zone(page
) != zone
);
1524 __init_single_page(page
, pfn
, zid
, nid
);
1525 if (!free_base_page
) {
1526 free_base_page
= page
;
1527 free_base_pfn
= pfn
;
1532 /* Where possible, batch up pages for a single free */
1535 /* Free the current block of pages to allocator */
1536 nr_pages
+= nr_to_free
;
1537 deferred_free_range(free_base_page
, free_base_pfn
,
1539 free_base_page
= NULL
;
1540 free_base_pfn
= nr_to_free
= 0;
1542 /* Free the last block of pages to allocator */
1543 nr_pages
+= nr_to_free
;
1544 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1546 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1549 /* Sanity check that the next zone really is unpopulated */
1550 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1552 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1553 jiffies_to_msecs(jiffies
- start
));
1555 pgdat_init_report_one_done();
1558 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1560 void __init
page_alloc_init_late(void)
1564 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1567 /* There will be num_node_state(N_MEMORY) threads */
1568 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1569 for_each_node_state(nid
, N_MEMORY
) {
1570 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1573 /* Block until all are initialised */
1574 wait_for_completion(&pgdat_init_all_done_comp
);
1576 /* Reinit limits that are based on free pages after the kernel is up */
1577 files_maxfiles_init();
1580 for_each_populated_zone(zone
)
1581 set_zone_contiguous(zone
);
1585 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1586 void __init
init_cma_reserved_pageblock(struct page
*page
)
1588 unsigned i
= pageblock_nr_pages
;
1589 struct page
*p
= page
;
1592 __ClearPageReserved(p
);
1593 set_page_count(p
, 0);
1596 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1598 if (pageblock_order
>= MAX_ORDER
) {
1599 i
= pageblock_nr_pages
;
1602 set_page_refcounted(p
);
1603 __free_pages(p
, MAX_ORDER
- 1);
1604 p
+= MAX_ORDER_NR_PAGES
;
1605 } while (i
-= MAX_ORDER_NR_PAGES
);
1607 set_page_refcounted(page
);
1608 __free_pages(page
, pageblock_order
);
1611 adjust_managed_page_count(page
, pageblock_nr_pages
);
1616 * The order of subdivision here is critical for the IO subsystem.
1617 * Please do not alter this order without good reasons and regression
1618 * testing. Specifically, as large blocks of memory are subdivided,
1619 * the order in which smaller blocks are delivered depends on the order
1620 * they're subdivided in this function. This is the primary factor
1621 * influencing the order in which pages are delivered to the IO
1622 * subsystem according to empirical testing, and this is also justified
1623 * by considering the behavior of a buddy system containing a single
1624 * large block of memory acted on by a series of small allocations.
1625 * This behavior is a critical factor in sglist merging's success.
1629 static inline void expand(struct zone
*zone
, struct page
*page
,
1630 int low
, int high
, struct free_area
*area
,
1633 unsigned long size
= 1 << high
;
1635 while (high
> low
) {
1639 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1642 * Mark as guard pages (or page), that will allow to
1643 * merge back to allocator when buddy will be freed.
1644 * Corresponding page table entries will not be touched,
1645 * pages will stay not present in virtual address space
1647 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1650 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1652 set_page_order(&page
[size
], high
);
1656 static void check_new_page_bad(struct page
*page
)
1658 const char *bad_reason
= NULL
;
1659 unsigned long bad_flags
= 0;
1661 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1662 bad_reason
= "nonzero mapcount";
1663 if (unlikely(page
->mapping
!= NULL
))
1664 bad_reason
= "non-NULL mapping";
1665 if (unlikely(page_ref_count(page
) != 0))
1666 bad_reason
= "nonzero _count";
1667 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1668 bad_reason
= "HWPoisoned (hardware-corrupted)";
1669 bad_flags
= __PG_HWPOISON
;
1670 /* Don't complain about hwpoisoned pages */
1671 page_mapcount_reset(page
); /* remove PageBuddy */
1674 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1675 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1676 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1679 if (unlikely(page
->mem_cgroup
))
1680 bad_reason
= "page still charged to cgroup";
1682 bad_page(page
, bad_reason
, bad_flags
);
1686 * This page is about to be returned from the page allocator
1688 static inline int check_new_page(struct page
*page
)
1690 if (likely(page_expected_state(page
,
1691 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1694 check_new_page_bad(page
);
1698 static inline bool free_pages_prezeroed(bool poisoned
)
1700 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1701 page_poisoning_enabled() && poisoned
;
1704 #ifdef CONFIG_DEBUG_VM
1705 static bool check_pcp_refill(struct page
*page
)
1710 static bool check_new_pcp(struct page
*page
)
1712 return check_new_page(page
);
1715 static bool check_pcp_refill(struct page
*page
)
1717 return check_new_page(page
);
1719 static bool check_new_pcp(struct page
*page
)
1723 #endif /* CONFIG_DEBUG_VM */
1725 static bool check_new_pages(struct page
*page
, unsigned int order
)
1728 for (i
= 0; i
< (1 << order
); i
++) {
1729 struct page
*p
= page
+ i
;
1731 if (unlikely(check_new_page(p
)))
1738 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1741 set_page_private(page
, 0);
1742 set_page_refcounted(page
);
1744 arch_alloc_page(page
, order
);
1745 kernel_map_pages(page
, 1 << order
, 1);
1746 kernel_poison_pages(page
, 1 << order
, 1);
1747 kasan_alloc_pages(page
, order
);
1748 set_page_owner(page
, order
, gfp_flags
);
1751 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1752 unsigned int alloc_flags
)
1755 bool poisoned
= true;
1757 for (i
= 0; i
< (1 << order
); i
++) {
1758 struct page
*p
= page
+ i
;
1760 poisoned
&= page_is_poisoned(p
);
1763 post_alloc_hook(page
, order
, gfp_flags
);
1765 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1766 for (i
= 0; i
< (1 << order
); i
++)
1767 clear_highpage(page
+ i
);
1769 if (order
&& (gfp_flags
& __GFP_COMP
))
1770 prep_compound_page(page
, order
);
1773 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1774 * allocate the page. The expectation is that the caller is taking
1775 * steps that will free more memory. The caller should avoid the page
1776 * being used for !PFMEMALLOC purposes.
1778 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1779 set_page_pfmemalloc(page
);
1781 clear_page_pfmemalloc(page
);
1785 * Go through the free lists for the given migratetype and remove
1786 * the smallest available page from the freelists
1789 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1792 unsigned int current_order
;
1793 struct free_area
*area
;
1796 /* Find a page of the appropriate size in the preferred list */
1797 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1798 area
= &(zone
->free_area
[current_order
]);
1799 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1803 list_del(&page
->lru
);
1804 rmv_page_order(page
);
1806 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1807 set_pcppage_migratetype(page
, migratetype
);
1816 * This array describes the order lists are fallen back to when
1817 * the free lists for the desirable migrate type are depleted
1819 static int fallbacks
[MIGRATE_TYPES
][4] = {
1820 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1821 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1822 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1824 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1826 #ifdef CONFIG_MEMORY_ISOLATION
1827 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1832 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1835 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1838 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1839 unsigned int order
) { return NULL
; }
1843 * Move the free pages in a range to the free lists of the requested type.
1844 * Note that start_page and end_pages are not aligned on a pageblock
1845 * boundary. If alignment is required, use move_freepages_block()
1847 int move_freepages(struct zone
*zone
,
1848 struct page
*start_page
, struct page
*end_page
,
1853 int pages_moved
= 0;
1855 #ifndef CONFIG_HOLES_IN_ZONE
1857 * page_zone is not safe to call in this context when
1858 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1859 * anyway as we check zone boundaries in move_freepages_block().
1860 * Remove at a later date when no bug reports exist related to
1861 * grouping pages by mobility
1863 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1866 for (page
= start_page
; page
<= end_page
;) {
1867 /* Make sure we are not inadvertently changing nodes */
1868 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1870 if (!pfn_valid_within(page_to_pfn(page
))) {
1875 if (!PageBuddy(page
)) {
1880 order
= page_order(page
);
1881 list_move(&page
->lru
,
1882 &zone
->free_area
[order
].free_list
[migratetype
]);
1884 pages_moved
+= 1 << order
;
1890 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1893 unsigned long start_pfn
, end_pfn
;
1894 struct page
*start_page
, *end_page
;
1896 start_pfn
= page_to_pfn(page
);
1897 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1898 start_page
= pfn_to_page(start_pfn
);
1899 end_page
= start_page
+ pageblock_nr_pages
- 1;
1900 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1902 /* Do not cross zone boundaries */
1903 if (!zone_spans_pfn(zone
, start_pfn
))
1905 if (!zone_spans_pfn(zone
, end_pfn
))
1908 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1911 static void change_pageblock_range(struct page
*pageblock_page
,
1912 int start_order
, int migratetype
)
1914 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1916 while (nr_pageblocks
--) {
1917 set_pageblock_migratetype(pageblock_page
, migratetype
);
1918 pageblock_page
+= pageblock_nr_pages
;
1923 * When we are falling back to another migratetype during allocation, try to
1924 * steal extra free pages from the same pageblocks to satisfy further
1925 * allocations, instead of polluting multiple pageblocks.
1927 * If we are stealing a relatively large buddy page, it is likely there will
1928 * be more free pages in the pageblock, so try to steal them all. For
1929 * reclaimable and unmovable allocations, we steal regardless of page size,
1930 * as fragmentation caused by those allocations polluting movable pageblocks
1931 * is worse than movable allocations stealing from unmovable and reclaimable
1934 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1937 * Leaving this order check is intended, although there is
1938 * relaxed order check in next check. The reason is that
1939 * we can actually steal whole pageblock if this condition met,
1940 * but, below check doesn't guarantee it and that is just heuristic
1941 * so could be changed anytime.
1943 if (order
>= pageblock_order
)
1946 if (order
>= pageblock_order
/ 2 ||
1947 start_mt
== MIGRATE_RECLAIMABLE
||
1948 start_mt
== MIGRATE_UNMOVABLE
||
1949 page_group_by_mobility_disabled
)
1956 * This function implements actual steal behaviour. If order is large enough,
1957 * we can steal whole pageblock. If not, we first move freepages in this
1958 * pageblock and check whether half of pages are moved or not. If half of
1959 * pages are moved, we can change migratetype of pageblock and permanently
1960 * use it's pages as requested migratetype in the future.
1962 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1965 unsigned int current_order
= page_order(page
);
1968 /* Take ownership for orders >= pageblock_order */
1969 if (current_order
>= pageblock_order
) {
1970 change_pageblock_range(page
, current_order
, start_type
);
1974 pages
= move_freepages_block(zone
, page
, start_type
);
1976 /* Claim the whole block if over half of it is free */
1977 if (pages
>= (1 << (pageblock_order
-1)) ||
1978 page_group_by_mobility_disabled
)
1979 set_pageblock_migratetype(page
, start_type
);
1983 * Check whether there is a suitable fallback freepage with requested order.
1984 * If only_stealable is true, this function returns fallback_mt only if
1985 * we can steal other freepages all together. This would help to reduce
1986 * fragmentation due to mixed migratetype pages in one pageblock.
1988 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1989 int migratetype
, bool only_stealable
, bool *can_steal
)
1994 if (area
->nr_free
== 0)
1999 fallback_mt
= fallbacks
[migratetype
][i
];
2000 if (fallback_mt
== MIGRATE_TYPES
)
2003 if (list_empty(&area
->free_list
[fallback_mt
]))
2006 if (can_steal_fallback(order
, migratetype
))
2009 if (!only_stealable
)
2020 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2021 * there are no empty page blocks that contain a page with a suitable order
2023 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2024 unsigned int alloc_order
)
2027 unsigned long max_managed
, flags
;
2030 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2031 * Check is race-prone but harmless.
2033 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2034 if (zone
->nr_reserved_highatomic
>= max_managed
)
2037 spin_lock_irqsave(&zone
->lock
, flags
);
2039 /* Recheck the nr_reserved_highatomic limit under the lock */
2040 if (zone
->nr_reserved_highatomic
>= max_managed
)
2044 mt
= get_pageblock_migratetype(page
);
2045 if (mt
!= MIGRATE_HIGHATOMIC
&&
2046 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
2047 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2048 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2049 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
2053 spin_unlock_irqrestore(&zone
->lock
, flags
);
2057 * Used when an allocation is about to fail under memory pressure. This
2058 * potentially hurts the reliability of high-order allocations when under
2059 * intense memory pressure but failed atomic allocations should be easier
2060 * to recover from than an OOM.
2062 static void unreserve_highatomic_pageblock(const struct alloc_context
*ac
)
2064 struct zonelist
*zonelist
= ac
->zonelist
;
2065 unsigned long flags
;
2071 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2073 /* Preserve at least one pageblock */
2074 if (zone
->nr_reserved_highatomic
<= pageblock_nr_pages
)
2077 spin_lock_irqsave(&zone
->lock
, flags
);
2078 for (order
= 0; order
< MAX_ORDER
; order
++) {
2079 struct free_area
*area
= &(zone
->free_area
[order
]);
2081 page
= list_first_entry_or_null(
2082 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2088 * It should never happen but changes to locking could
2089 * inadvertently allow a per-cpu drain to add pages
2090 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2091 * and watch for underflows.
2093 zone
->nr_reserved_highatomic
-= min(pageblock_nr_pages
,
2094 zone
->nr_reserved_highatomic
);
2097 * Convert to ac->migratetype and avoid the normal
2098 * pageblock stealing heuristics. Minimally, the caller
2099 * is doing the work and needs the pages. More
2100 * importantly, if the block was always converted to
2101 * MIGRATE_UNMOVABLE or another type then the number
2102 * of pageblocks that cannot be completely freed
2105 set_pageblock_migratetype(page
, ac
->migratetype
);
2106 move_freepages_block(zone
, page
, ac
->migratetype
);
2107 spin_unlock_irqrestore(&zone
->lock
, flags
);
2110 spin_unlock_irqrestore(&zone
->lock
, flags
);
2114 /* Remove an element from the buddy allocator from the fallback list */
2115 static inline struct page
*
2116 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2118 struct free_area
*area
;
2119 unsigned int current_order
;
2124 /* Find the largest possible block of pages in the other list */
2125 for (current_order
= MAX_ORDER
-1;
2126 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2128 area
= &(zone
->free_area
[current_order
]);
2129 fallback_mt
= find_suitable_fallback(area
, current_order
,
2130 start_migratetype
, false, &can_steal
);
2131 if (fallback_mt
== -1)
2134 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2137 steal_suitable_fallback(zone
, page
, start_migratetype
);
2139 /* Remove the page from the freelists */
2141 list_del(&page
->lru
);
2142 rmv_page_order(page
);
2144 expand(zone
, page
, order
, current_order
, area
,
2147 * The pcppage_migratetype may differ from pageblock's
2148 * migratetype depending on the decisions in
2149 * find_suitable_fallback(). This is OK as long as it does not
2150 * differ for MIGRATE_CMA pageblocks. Those can be used as
2151 * fallback only via special __rmqueue_cma_fallback() function
2153 set_pcppage_migratetype(page
, start_migratetype
);
2155 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2156 start_migratetype
, fallback_mt
);
2165 * Do the hard work of removing an element from the buddy allocator.
2166 * Call me with the zone->lock already held.
2168 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2173 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2174 if (unlikely(!page
)) {
2175 if (migratetype
== MIGRATE_MOVABLE
)
2176 page
= __rmqueue_cma_fallback(zone
, order
);
2179 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2182 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2187 * Obtain a specified number of elements from the buddy allocator, all under
2188 * a single hold of the lock, for efficiency. Add them to the supplied list.
2189 * Returns the number of new pages which were placed at *list.
2191 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2192 unsigned long count
, struct list_head
*list
,
2193 int migratetype
, bool cold
)
2197 spin_lock(&zone
->lock
);
2198 for (i
= 0; i
< count
; ++i
) {
2199 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2200 if (unlikely(page
== NULL
))
2203 if (unlikely(check_pcp_refill(page
)))
2207 * Split buddy pages returned by expand() are received here
2208 * in physical page order. The page is added to the callers and
2209 * list and the list head then moves forward. From the callers
2210 * perspective, the linked list is ordered by page number in
2211 * some conditions. This is useful for IO devices that can
2212 * merge IO requests if the physical pages are ordered
2216 list_add(&page
->lru
, list
);
2218 list_add_tail(&page
->lru
, list
);
2220 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2221 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2224 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2225 spin_unlock(&zone
->lock
);
2231 * Called from the vmstat counter updater to drain pagesets of this
2232 * currently executing processor on remote nodes after they have
2235 * Note that this function must be called with the thread pinned to
2236 * a single processor.
2238 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2240 unsigned long flags
;
2241 int to_drain
, batch
;
2243 local_irq_save(flags
);
2244 batch
= READ_ONCE(pcp
->batch
);
2245 to_drain
= min(pcp
->count
, batch
);
2247 free_pcppages_bulk(zone
, to_drain
, pcp
);
2248 pcp
->count
-= to_drain
;
2250 local_irq_restore(flags
);
2255 * Drain pcplists of the indicated processor and zone.
2257 * The processor must either be the current processor and the
2258 * thread pinned to the current processor or a processor that
2261 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2263 unsigned long flags
;
2264 struct per_cpu_pageset
*pset
;
2265 struct per_cpu_pages
*pcp
;
2267 local_irq_save(flags
);
2268 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2272 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2275 local_irq_restore(flags
);
2279 * Drain pcplists of all zones on the indicated processor.
2281 * The processor must either be the current processor and the
2282 * thread pinned to the current processor or a processor that
2285 static void drain_pages(unsigned int cpu
)
2289 for_each_populated_zone(zone
) {
2290 drain_pages_zone(cpu
, zone
);
2295 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2297 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2298 * the single zone's pages.
2300 void drain_local_pages(struct zone
*zone
)
2302 int cpu
= smp_processor_id();
2305 drain_pages_zone(cpu
, zone
);
2311 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2313 * When zone parameter is non-NULL, spill just the single zone's pages.
2315 * Note that this code is protected against sending an IPI to an offline
2316 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2317 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2318 * nothing keeps CPUs from showing up after we populated the cpumask and
2319 * before the call to on_each_cpu_mask().
2321 void drain_all_pages(struct zone
*zone
)
2326 * Allocate in the BSS so we wont require allocation in
2327 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2329 static cpumask_t cpus_with_pcps
;
2332 * We don't care about racing with CPU hotplug event
2333 * as offline notification will cause the notified
2334 * cpu to drain that CPU pcps and on_each_cpu_mask
2335 * disables preemption as part of its processing
2337 for_each_online_cpu(cpu
) {
2338 struct per_cpu_pageset
*pcp
;
2340 bool has_pcps
= false;
2343 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2347 for_each_populated_zone(z
) {
2348 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2349 if (pcp
->pcp
.count
) {
2357 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2359 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2361 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
2365 #ifdef CONFIG_HIBERNATION
2367 void mark_free_pages(struct zone
*zone
)
2369 unsigned long pfn
, max_zone_pfn
;
2370 unsigned long flags
;
2371 unsigned int order
, t
;
2374 if (zone_is_empty(zone
))
2377 spin_lock_irqsave(&zone
->lock
, flags
);
2379 max_zone_pfn
= zone_end_pfn(zone
);
2380 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2381 if (pfn_valid(pfn
)) {
2382 page
= pfn_to_page(pfn
);
2384 if (page_zone(page
) != zone
)
2387 if (!swsusp_page_is_forbidden(page
))
2388 swsusp_unset_page_free(page
);
2391 for_each_migratetype_order(order
, t
) {
2392 list_for_each_entry(page
,
2393 &zone
->free_area
[order
].free_list
[t
], lru
) {
2396 pfn
= page_to_pfn(page
);
2397 for (i
= 0; i
< (1UL << order
); i
++)
2398 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2401 spin_unlock_irqrestore(&zone
->lock
, flags
);
2403 #endif /* CONFIG_PM */
2406 * Free a 0-order page
2407 * cold == true ? free a cold page : free a hot page
2409 void free_hot_cold_page(struct page
*page
, bool cold
)
2411 struct zone
*zone
= page_zone(page
);
2412 struct per_cpu_pages
*pcp
;
2413 unsigned long flags
;
2414 unsigned long pfn
= page_to_pfn(page
);
2417 if (!free_pcp_prepare(page
))
2420 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2421 set_pcppage_migratetype(page
, migratetype
);
2422 local_irq_save(flags
);
2423 __count_vm_event(PGFREE
);
2426 * We only track unmovable, reclaimable and movable on pcp lists.
2427 * Free ISOLATE pages back to the allocator because they are being
2428 * offlined but treat RESERVE as movable pages so we can get those
2429 * areas back if necessary. Otherwise, we may have to free
2430 * excessively into the page allocator
2432 if (migratetype
>= MIGRATE_PCPTYPES
) {
2433 if (unlikely(is_migrate_isolate(migratetype
))) {
2434 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2437 migratetype
= MIGRATE_MOVABLE
;
2440 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2442 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2444 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2446 if (pcp
->count
>= pcp
->high
) {
2447 unsigned long batch
= READ_ONCE(pcp
->batch
);
2448 free_pcppages_bulk(zone
, batch
, pcp
);
2449 pcp
->count
-= batch
;
2453 local_irq_restore(flags
);
2457 * Free a list of 0-order pages
2459 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2461 struct page
*page
, *next
;
2463 list_for_each_entry_safe(page
, next
, list
, lru
) {
2464 trace_mm_page_free_batched(page
, cold
);
2465 free_hot_cold_page(page
, cold
);
2470 * split_page takes a non-compound higher-order page, and splits it into
2471 * n (1<<order) sub-pages: page[0..n]
2472 * Each sub-page must be freed individually.
2474 * Note: this is probably too low level an operation for use in drivers.
2475 * Please consult with lkml before using this in your driver.
2477 void split_page(struct page
*page
, unsigned int order
)
2481 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2482 VM_BUG_ON_PAGE(!page_count(page
), page
);
2484 #ifdef CONFIG_KMEMCHECK
2486 * Split shadow pages too, because free(page[0]) would
2487 * otherwise free the whole shadow.
2489 if (kmemcheck_page_is_tracked(page
))
2490 split_page(virt_to_page(page
[0].shadow
), order
);
2493 for (i
= 1; i
< (1 << order
); i
++)
2494 set_page_refcounted(page
+ i
);
2495 split_page_owner(page
, order
);
2497 EXPORT_SYMBOL_GPL(split_page
);
2499 int __isolate_free_page(struct page
*page
, unsigned int order
)
2501 unsigned long watermark
;
2505 BUG_ON(!PageBuddy(page
));
2507 zone
= page_zone(page
);
2508 mt
= get_pageblock_migratetype(page
);
2510 if (!is_migrate_isolate(mt
)) {
2512 * Obey watermarks as if the page was being allocated. We can
2513 * emulate a high-order watermark check with a raised order-0
2514 * watermark, because we already know our high-order page
2517 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2518 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2521 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2524 /* Remove page from free list */
2525 list_del(&page
->lru
);
2526 zone
->free_area
[order
].nr_free
--;
2527 rmv_page_order(page
);
2530 * Set the pageblock if the isolated page is at least half of a
2533 if (order
>= pageblock_order
- 1) {
2534 struct page
*endpage
= page
+ (1 << order
) - 1;
2535 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2536 int mt
= get_pageblock_migratetype(page
);
2537 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2538 set_pageblock_migratetype(page
,
2544 return 1UL << 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
);
2597 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2598 list
= &pcp
->lists
[migratetype
];
2599 if (list_empty(list
)) {
2600 pcp
->count
+= rmqueue_bulk(zone
, 0,
2603 if (unlikely(list_empty(list
)))
2608 page
= list_last_entry(list
, struct page
, lru
);
2610 page
= list_first_entry(list
, struct page
, lru
);
2612 list_del(&page
->lru
);
2615 } while (check_new_pcp(page
));
2618 * We most definitely don't want callers attempting to
2619 * allocate greater than order-1 page units with __GFP_NOFAIL.
2621 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2622 spin_lock_irqsave(&zone
->lock
, flags
);
2626 if (alloc_flags
& ALLOC_HARDER
) {
2627 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2629 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2632 page
= __rmqueue(zone
, order
, migratetype
);
2633 } while (page
&& check_new_pages(page
, order
));
2634 spin_unlock(&zone
->lock
);
2637 __mod_zone_freepage_state(zone
, -(1 << order
),
2638 get_pcppage_migratetype(page
));
2641 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2642 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2643 local_irq_restore(flags
);
2645 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2649 local_irq_restore(flags
);
2653 #ifdef CONFIG_FAIL_PAGE_ALLOC
2656 struct fault_attr attr
;
2658 bool ignore_gfp_highmem
;
2659 bool ignore_gfp_reclaim
;
2661 } fail_page_alloc
= {
2662 .attr
= FAULT_ATTR_INITIALIZER
,
2663 .ignore_gfp_reclaim
= true,
2664 .ignore_gfp_highmem
= true,
2668 static int __init
setup_fail_page_alloc(char *str
)
2670 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2672 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2674 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2676 if (order
< fail_page_alloc
.min_order
)
2678 if (gfp_mask
& __GFP_NOFAIL
)
2680 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2682 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2683 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2686 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2689 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2691 static int __init
fail_page_alloc_debugfs(void)
2693 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2696 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2697 &fail_page_alloc
.attr
);
2699 return PTR_ERR(dir
);
2701 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2702 &fail_page_alloc
.ignore_gfp_reclaim
))
2704 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2705 &fail_page_alloc
.ignore_gfp_highmem
))
2707 if (!debugfs_create_u32("min-order", mode
, dir
,
2708 &fail_page_alloc
.min_order
))
2713 debugfs_remove_recursive(dir
);
2718 late_initcall(fail_page_alloc_debugfs
);
2720 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2722 #else /* CONFIG_FAIL_PAGE_ALLOC */
2724 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2729 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2732 * Return true if free base pages are above 'mark'. For high-order checks it
2733 * will return true of the order-0 watermark is reached and there is at least
2734 * one free page of a suitable size. Checking now avoids taking the zone lock
2735 * to check in the allocation paths if no pages are free.
2737 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2738 int classzone_idx
, unsigned int alloc_flags
,
2743 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2745 /* free_pages may go negative - that's OK */
2746 free_pages
-= (1 << order
) - 1;
2748 if (alloc_flags
& ALLOC_HIGH
)
2752 * If the caller does not have rights to ALLOC_HARDER then subtract
2753 * the high-atomic reserves. This will over-estimate the size of the
2754 * atomic reserve but it avoids a search.
2756 if (likely(!alloc_harder
))
2757 free_pages
-= z
->nr_reserved_highatomic
;
2762 /* If allocation can't use CMA areas don't use free CMA pages */
2763 if (!(alloc_flags
& ALLOC_CMA
))
2764 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2768 * Check watermarks for an order-0 allocation request. If these
2769 * are not met, then a high-order request also cannot go ahead
2770 * even if a suitable page happened to be free.
2772 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2775 /* If this is an order-0 request then the watermark is fine */
2779 /* For a high-order request, check at least one suitable page is free */
2780 for (o
= order
; o
< MAX_ORDER
; o
++) {
2781 struct free_area
*area
= &z
->free_area
[o
];
2790 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2791 if (!list_empty(&area
->free_list
[mt
]))
2796 if ((alloc_flags
& ALLOC_CMA
) &&
2797 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2805 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2806 int classzone_idx
, unsigned int alloc_flags
)
2808 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2809 zone_page_state(z
, NR_FREE_PAGES
));
2812 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2813 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2815 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2819 /* If allocation can't use CMA areas don't use free CMA pages */
2820 if (!(alloc_flags
& ALLOC_CMA
))
2821 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2825 * Fast check for order-0 only. If this fails then the reserves
2826 * need to be calculated. There is a corner case where the check
2827 * passes but only the high-order atomic reserve are free. If
2828 * the caller is !atomic then it'll uselessly search the free
2829 * list. That corner case is then slower but it is harmless.
2831 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2834 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2838 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2839 unsigned long mark
, int classzone_idx
)
2841 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2843 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2844 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2846 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2851 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2853 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2856 #else /* CONFIG_NUMA */
2857 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2861 #endif /* CONFIG_NUMA */
2864 * get_page_from_freelist goes through the zonelist trying to allocate
2867 static struct page
*
2868 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2869 const struct alloc_context
*ac
)
2871 struct zoneref
*z
= ac
->preferred_zoneref
;
2873 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
2876 * Scan zonelist, looking for a zone with enough free.
2877 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2879 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2884 if (cpusets_enabled() &&
2885 (alloc_flags
& ALLOC_CPUSET
) &&
2886 !__cpuset_zone_allowed(zone
, gfp_mask
))
2889 * When allocating a page cache page for writing, we
2890 * want to get it from a node that is within its dirty
2891 * limit, such that no single node holds more than its
2892 * proportional share of globally allowed dirty pages.
2893 * The dirty limits take into account the node's
2894 * lowmem reserves and high watermark so that kswapd
2895 * should be able to balance it without having to
2896 * write pages from its LRU list.
2898 * XXX: For now, allow allocations to potentially
2899 * exceed the per-node dirty limit in the slowpath
2900 * (spread_dirty_pages unset) before going into reclaim,
2901 * which is important when on a NUMA setup the allowed
2902 * nodes are together not big enough to reach the
2903 * global limit. The proper fix for these situations
2904 * will require awareness of nodes in the
2905 * dirty-throttling and the flusher threads.
2907 if (ac
->spread_dirty_pages
) {
2908 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
2911 if (!node_dirty_ok(zone
->zone_pgdat
)) {
2912 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
2917 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2918 if (!zone_watermark_fast(zone
, order
, mark
,
2919 ac_classzone_idx(ac
), alloc_flags
)) {
2922 /* Checked here to keep the fast path fast */
2923 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2924 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2927 if (node_reclaim_mode
== 0 ||
2928 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
2931 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
2933 case NODE_RECLAIM_NOSCAN
:
2936 case NODE_RECLAIM_FULL
:
2937 /* scanned but unreclaimable */
2940 /* did we reclaim enough */
2941 if (zone_watermark_ok(zone
, order
, mark
,
2942 ac_classzone_idx(ac
), alloc_flags
))
2950 page
= buffered_rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
2951 gfp_mask
, alloc_flags
, ac
->migratetype
);
2953 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
2956 * If this is a high-order atomic allocation then check
2957 * if the pageblock should be reserved for the future
2959 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
2960 reserve_highatomic_pageblock(page
, zone
, order
);
2970 * Large machines with many possible nodes should not always dump per-node
2971 * meminfo in irq context.
2973 static inline bool should_suppress_show_mem(void)
2978 ret
= in_interrupt();
2983 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2984 DEFAULT_RATELIMIT_INTERVAL
,
2985 DEFAULT_RATELIMIT_BURST
);
2987 void warn_alloc(gfp_t gfp_mask
, const char *fmt
, ...)
2989 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2990 struct va_format vaf
;
2993 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2994 debug_guardpage_minorder() > 0)
2998 * This documents exceptions given to allocations in certain
2999 * contexts that are allowed to allocate outside current's set
3002 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3003 if (test_thread_flag(TIF_MEMDIE
) ||
3004 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3005 filter
&= ~SHOW_MEM_FILTER_NODES
;
3006 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3007 filter
&= ~SHOW_MEM_FILTER_NODES
;
3009 pr_warn("%s: ", current
->comm
);
3011 va_start(args
, fmt
);
3014 pr_cont("%pV", &vaf
);
3017 pr_cont(", mode:%#x(%pGg)\n", gfp_mask
, &gfp_mask
);
3020 if (!should_suppress_show_mem())
3024 static inline struct page
*
3025 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3026 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3028 struct oom_control oc
= {
3029 .zonelist
= ac
->zonelist
,
3030 .nodemask
= ac
->nodemask
,
3032 .gfp_mask
= gfp_mask
,
3037 *did_some_progress
= 0;
3040 * Acquire the oom lock. If that fails, somebody else is
3041 * making progress for us.
3043 if (!mutex_trylock(&oom_lock
)) {
3044 *did_some_progress
= 1;
3045 schedule_timeout_uninterruptible(1);
3050 * Go through the zonelist yet one more time, keep very high watermark
3051 * here, this is only to catch a parallel oom killing, we must fail if
3052 * we're still under heavy pressure.
3054 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3055 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3059 if (!(gfp_mask
& __GFP_NOFAIL
)) {
3060 /* Coredumps can quickly deplete all memory reserves */
3061 if (current
->flags
& PF_DUMPCORE
)
3063 /* The OOM killer will not help higher order allocs */
3064 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3066 /* The OOM killer does not needlessly kill tasks for lowmem */
3067 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3069 if (pm_suspended_storage())
3072 * XXX: GFP_NOFS allocations should rather fail than rely on
3073 * other request to make a forward progress.
3074 * We are in an unfortunate situation where out_of_memory cannot
3075 * do much for this context but let's try it to at least get
3076 * access to memory reserved if the current task is killed (see
3077 * out_of_memory). Once filesystems are ready to handle allocation
3078 * failures more gracefully we should just bail out here.
3081 /* The OOM killer may not free memory on a specific node */
3082 if (gfp_mask
& __GFP_THISNODE
)
3085 /* Exhausted what can be done so it's blamo time */
3086 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3087 *did_some_progress
= 1;
3089 if (gfp_mask
& __GFP_NOFAIL
) {
3090 page
= get_page_from_freelist(gfp_mask
, order
,
3091 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
3093 * fallback to ignore cpuset restriction if our nodes
3097 page
= get_page_from_freelist(gfp_mask
, order
,
3098 ALLOC_NO_WATERMARKS
, ac
);
3102 mutex_unlock(&oom_lock
);
3107 * Maximum number of compaction retries wit a progress before OOM
3108 * killer is consider as the only way to move forward.
3110 #define MAX_COMPACT_RETRIES 16
3112 #ifdef CONFIG_COMPACTION
3113 /* Try memory compaction for high-order allocations before reclaim */
3114 static struct page
*
3115 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3116 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3117 enum compact_priority prio
, enum compact_result
*compact_result
)
3124 current
->flags
|= PF_MEMALLOC
;
3125 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3127 current
->flags
&= ~PF_MEMALLOC
;
3129 if (*compact_result
<= COMPACT_INACTIVE
)
3133 * At least in one zone compaction wasn't deferred or skipped, so let's
3134 * count a compaction stall
3136 count_vm_event(COMPACTSTALL
);
3138 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3141 struct zone
*zone
= page_zone(page
);
3143 zone
->compact_blockskip_flush
= false;
3144 compaction_defer_reset(zone
, order
, true);
3145 count_vm_event(COMPACTSUCCESS
);
3150 * It's bad if compaction run occurs and fails. The most likely reason
3151 * is that pages exist, but not enough to satisfy watermarks.
3153 count_vm_event(COMPACTFAIL
);
3161 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3162 enum compact_result compact_result
,
3163 enum compact_priority
*compact_priority
,
3164 int *compaction_retries
)
3166 int max_retries
= MAX_COMPACT_RETRIES
;
3172 if (compaction_made_progress(compact_result
))
3173 (*compaction_retries
)++;
3176 * compaction considers all the zone as desperately out of memory
3177 * so it doesn't really make much sense to retry except when the
3178 * failure could be caused by insufficient priority
3180 if (compaction_failed(compact_result
))
3181 goto check_priority
;
3184 * make sure the compaction wasn't deferred or didn't bail out early
3185 * due to locks contention before we declare that we should give up.
3186 * But do not retry if the given zonelist is not suitable for
3189 if (compaction_withdrawn(compact_result
))
3190 return compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3193 * !costly requests are much more important than __GFP_REPEAT
3194 * costly ones because they are de facto nofail and invoke OOM
3195 * killer to move on while costly can fail and users are ready
3196 * to cope with that. 1/4 retries is rather arbitrary but we
3197 * would need much more detailed feedback from compaction to
3198 * make a better decision.
3200 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3202 if (*compaction_retries
<= max_retries
)
3206 * Make sure there are attempts at the highest priority if we exhausted
3207 * all retries or failed at the lower priorities.
3210 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3211 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3212 if (*compact_priority
> min_priority
) {
3213 (*compact_priority
)--;
3214 *compaction_retries
= 0;
3220 static inline struct page
*
3221 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3222 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3223 enum compact_priority prio
, enum compact_result
*compact_result
)
3225 *compact_result
= COMPACT_SKIPPED
;
3230 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3231 enum compact_result compact_result
,
3232 enum compact_priority
*compact_priority
,
3233 int *compaction_retries
)
3238 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3242 * There are setups with compaction disabled which would prefer to loop
3243 * inside the allocator rather than hit the oom killer prematurely.
3244 * Let's give them a good hope and keep retrying while the order-0
3245 * watermarks are OK.
3247 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3249 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3250 ac_classzone_idx(ac
), alloc_flags
))
3255 #endif /* CONFIG_COMPACTION */
3257 /* Perform direct synchronous page reclaim */
3259 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3260 const struct alloc_context
*ac
)
3262 struct reclaim_state reclaim_state
;
3267 /* We now go into synchronous reclaim */
3268 cpuset_memory_pressure_bump();
3269 current
->flags
|= PF_MEMALLOC
;
3270 lockdep_set_current_reclaim_state(gfp_mask
);
3271 reclaim_state
.reclaimed_slab
= 0;
3272 current
->reclaim_state
= &reclaim_state
;
3274 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3277 current
->reclaim_state
= NULL
;
3278 lockdep_clear_current_reclaim_state();
3279 current
->flags
&= ~PF_MEMALLOC
;
3286 /* The really slow allocator path where we enter direct reclaim */
3287 static inline struct page
*
3288 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3289 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3290 unsigned long *did_some_progress
)
3292 struct page
*page
= NULL
;
3293 bool drained
= false;
3295 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3296 if (unlikely(!(*did_some_progress
)))
3300 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3303 * If an allocation failed after direct reclaim, it could be because
3304 * pages are pinned on the per-cpu lists or in high alloc reserves.
3305 * Shrink them them and try again
3307 if (!page
&& !drained
) {
3308 unreserve_highatomic_pageblock(ac
);
3309 drain_all_pages(NULL
);
3317 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3321 pg_data_t
*last_pgdat
= NULL
;
3323 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3324 ac
->high_zoneidx
, ac
->nodemask
) {
3325 if (last_pgdat
!= zone
->zone_pgdat
)
3326 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3327 last_pgdat
= zone
->zone_pgdat
;
3331 static inline unsigned int
3332 gfp_to_alloc_flags(gfp_t gfp_mask
)
3334 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3336 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3337 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3340 * The caller may dip into page reserves a bit more if the caller
3341 * cannot run direct reclaim, or if the caller has realtime scheduling
3342 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3343 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3345 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3347 if (gfp_mask
& __GFP_ATOMIC
) {
3349 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3350 * if it can't schedule.
3352 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3353 alloc_flags
|= ALLOC_HARDER
;
3355 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3356 * comment for __cpuset_node_allowed().
3358 alloc_flags
&= ~ALLOC_CPUSET
;
3359 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3360 alloc_flags
|= ALLOC_HARDER
;
3363 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3364 alloc_flags
|= ALLOC_CMA
;
3369 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3371 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3374 if (gfp_mask
& __GFP_MEMALLOC
)
3376 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3378 if (!in_interrupt() &&
3379 ((current
->flags
& PF_MEMALLOC
) ||
3380 unlikely(test_thread_flag(TIF_MEMDIE
))))
3387 * Maximum number of reclaim retries without any progress before OOM killer
3388 * is consider as the only way to move forward.
3390 #define MAX_RECLAIM_RETRIES 16
3393 * Checks whether it makes sense to retry the reclaim to make a forward progress
3394 * for the given allocation request.
3395 * The reclaim feedback represented by did_some_progress (any progress during
3396 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3397 * any progress in a row) is considered as well as the reclaimable pages on the
3398 * applicable zone list (with a backoff mechanism which is a function of
3399 * no_progress_loops).
3401 * Returns true if a retry is viable or false to enter the oom path.
3404 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3405 struct alloc_context
*ac
, int alloc_flags
,
3406 bool did_some_progress
, int *no_progress_loops
)
3412 * Costly allocations might have made a progress but this doesn't mean
3413 * their order will become available due to high fragmentation so
3414 * always increment the no progress counter for them
3416 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3417 *no_progress_loops
= 0;
3419 (*no_progress_loops
)++;
3422 * Make sure we converge to OOM if we cannot make any progress
3423 * several times in the row.
3425 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
)
3429 * Keep reclaiming pages while there is a chance this will lead
3430 * somewhere. If none of the target zones can satisfy our allocation
3431 * request even if all reclaimable pages are considered then we are
3432 * screwed and have to go OOM.
3434 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3436 unsigned long available
;
3437 unsigned long reclaimable
;
3439 available
= reclaimable
= zone_reclaimable_pages(zone
);
3440 available
-= DIV_ROUND_UP((*no_progress_loops
) * available
,
3441 MAX_RECLAIM_RETRIES
);
3442 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3445 * Would the allocation succeed if we reclaimed the whole
3448 if (__zone_watermark_ok(zone
, order
, min_wmark_pages(zone
),
3449 ac_classzone_idx(ac
), alloc_flags
, available
)) {
3451 * If we didn't make any progress and have a lot of
3452 * dirty + writeback pages then we should wait for
3453 * an IO to complete to slow down the reclaim and
3454 * prevent from pre mature OOM
3456 if (!did_some_progress
) {
3457 unsigned long write_pending
;
3459 write_pending
= zone_page_state_snapshot(zone
,
3460 NR_ZONE_WRITE_PENDING
);
3462 if (2 * write_pending
> reclaimable
) {
3463 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3469 * Memory allocation/reclaim might be called from a WQ
3470 * context and the current implementation of the WQ
3471 * concurrency control doesn't recognize that
3472 * a particular WQ is congested if the worker thread is
3473 * looping without ever sleeping. Therefore we have to
3474 * do a short sleep here rather than calling
3477 if (current
->flags
& PF_WQ_WORKER
)
3478 schedule_timeout_uninterruptible(1);
3489 static inline struct page
*
3490 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3491 struct alloc_context
*ac
)
3493 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3494 struct page
*page
= NULL
;
3495 unsigned int alloc_flags
;
3496 unsigned long did_some_progress
;
3497 enum compact_priority compact_priority
= DEF_COMPACT_PRIORITY
;
3498 enum compact_result compact_result
;
3499 int compaction_retries
= 0;
3500 int no_progress_loops
= 0;
3501 unsigned long alloc_start
= jiffies
;
3502 unsigned int stall_timeout
= 10 * HZ
;
3505 * In the slowpath, we sanity check order to avoid ever trying to
3506 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3507 * be using allocators in order of preference for an area that is
3510 if (order
>= MAX_ORDER
) {
3511 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3516 * We also sanity check to catch abuse of atomic reserves being used by
3517 * callers that are not in atomic context.
3519 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3520 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3521 gfp_mask
&= ~__GFP_ATOMIC
;
3524 * The fast path uses conservative alloc_flags to succeed only until
3525 * kswapd needs to be woken up, and to avoid the cost of setting up
3526 * alloc_flags precisely. So we do that now.
3528 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3530 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3531 wake_all_kswapds(order
, ac
);
3534 * The adjusted alloc_flags might result in immediate success, so try
3537 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3542 * For costly allocations, try direct compaction first, as it's likely
3543 * that we have enough base pages and don't need to reclaim. Don't try
3544 * that for allocations that are allowed to ignore watermarks, as the
3545 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3547 if (can_direct_reclaim
&& order
> PAGE_ALLOC_COSTLY_ORDER
&&
3548 !gfp_pfmemalloc_allowed(gfp_mask
)) {
3549 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3551 INIT_COMPACT_PRIORITY
,
3557 * Checks for costly allocations with __GFP_NORETRY, which
3558 * includes THP page fault allocations
3560 if (gfp_mask
& __GFP_NORETRY
) {
3562 * If compaction is deferred for high-order allocations,
3563 * it is because sync compaction recently failed. If
3564 * this is the case and the caller requested a THP
3565 * allocation, we do not want to heavily disrupt the
3566 * system, so we fail the allocation instead of entering
3569 if (compact_result
== COMPACT_DEFERRED
)
3573 * Looks like reclaim/compaction is worth trying, but
3574 * sync compaction could be very expensive, so keep
3575 * using async compaction.
3577 compact_priority
= INIT_COMPACT_PRIORITY
;
3582 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3583 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3584 wake_all_kswapds(order
, ac
);
3586 if (gfp_pfmemalloc_allowed(gfp_mask
))
3587 alloc_flags
= ALLOC_NO_WATERMARKS
;
3590 * Reset the zonelist iterators if memory policies can be ignored.
3591 * These allocations are high priority and system rather than user
3594 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3595 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3596 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3597 ac
->high_zoneidx
, ac
->nodemask
);
3600 /* Attempt with potentially adjusted zonelist and alloc_flags */
3601 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3605 /* Caller is not willing to reclaim, we can't balance anything */
3606 if (!can_direct_reclaim
) {
3608 * All existing users of the __GFP_NOFAIL are blockable, so warn
3609 * of any new users that actually allow this type of allocation
3612 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3616 /* Avoid recursion of direct reclaim */
3617 if (current
->flags
& PF_MEMALLOC
) {
3619 * __GFP_NOFAIL request from this context is rather bizarre
3620 * because we cannot reclaim anything and only can loop waiting
3621 * for somebody to do a work for us.
3623 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3630 /* Avoid allocations with no watermarks from looping endlessly */
3631 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3635 /* Try direct reclaim and then allocating */
3636 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3637 &did_some_progress
);
3641 /* Try direct compaction and then allocating */
3642 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3643 compact_priority
, &compact_result
);
3647 /* Do not loop if specifically requested */
3648 if (gfp_mask
& __GFP_NORETRY
)
3652 * Do not retry costly high order allocations unless they are
3655 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_REPEAT
))
3658 /* Make sure we know about allocations which stall for too long */
3659 if (time_after(jiffies
, alloc_start
+ stall_timeout
)) {
3660 warn_alloc(gfp_mask
,
3661 "page alloction stalls for %ums, order:%u\n",
3662 jiffies_to_msecs(jiffies
-alloc_start
), order
);
3663 stall_timeout
+= 10 * HZ
;
3666 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3667 did_some_progress
> 0, &no_progress_loops
))
3671 * It doesn't make any sense to retry for the compaction if the order-0
3672 * reclaim is not able to make any progress because the current
3673 * implementation of the compaction depends on the sufficient amount
3674 * of free memory (see __compaction_suitable)
3676 if (did_some_progress
> 0 &&
3677 should_compact_retry(ac
, order
, alloc_flags
,
3678 compact_result
, &compact_priority
,
3679 &compaction_retries
))
3682 /* Reclaim has failed us, start killing things */
3683 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3687 /* Retry as long as the OOM killer is making progress */
3688 if (did_some_progress
) {
3689 no_progress_loops
= 0;
3694 warn_alloc(gfp_mask
,
3695 "page allocation failure: order:%u", order
);
3701 * This is the 'heart' of the zoned buddy allocator.
3704 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3705 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3708 unsigned int cpuset_mems_cookie
;
3709 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
3710 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3711 struct alloc_context ac
= {
3712 .high_zoneidx
= gfp_zone(gfp_mask
),
3713 .zonelist
= zonelist
,
3714 .nodemask
= nodemask
,
3715 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3718 if (cpusets_enabled()) {
3719 alloc_mask
|= __GFP_HARDWALL
;
3720 alloc_flags
|= ALLOC_CPUSET
;
3722 ac
.nodemask
= &cpuset_current_mems_allowed
;
3725 gfp_mask
&= gfp_allowed_mask
;
3727 lockdep_trace_alloc(gfp_mask
);
3729 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3731 if (should_fail_alloc_page(gfp_mask
, order
))
3735 * Check the zones suitable for the gfp_mask contain at least one
3736 * valid zone. It's possible to have an empty zonelist as a result
3737 * of __GFP_THISNODE and a memoryless node
3739 if (unlikely(!zonelist
->_zonerefs
->zone
))
3742 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3743 alloc_flags
|= ALLOC_CMA
;
3746 cpuset_mems_cookie
= read_mems_allowed_begin();
3748 /* Dirty zone balancing only done in the fast path */
3749 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3752 * The preferred zone is used for statistics but crucially it is
3753 * also used as the starting point for the zonelist iterator. It
3754 * may get reset for allocations that ignore memory policies.
3756 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3757 ac
.high_zoneidx
, ac
.nodemask
);
3758 if (!ac
.preferred_zoneref
) {
3763 /* First allocation attempt */
3764 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3769 * Runtime PM, block IO and its error handling path can deadlock
3770 * because I/O on the device might not complete.
3772 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3773 ac
.spread_dirty_pages
= false;
3776 * Restore the original nodemask if it was potentially replaced with
3777 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3779 if (cpusets_enabled())
3780 ac
.nodemask
= nodemask
;
3781 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3785 * When updating a task's mems_allowed, it is possible to race with
3786 * parallel threads in such a way that an allocation can fail while
3787 * the mask is being updated. If a page allocation is about to fail,
3788 * check if the cpuset changed during allocation and if so, retry.
3790 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3791 alloc_mask
= gfp_mask
;
3796 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
3797 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
3798 __free_pages(page
, order
);
3802 if (kmemcheck_enabled
&& page
)
3803 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3805 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3809 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3812 * Common helper functions.
3814 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3819 * __get_free_pages() returns a 32-bit address, which cannot represent
3822 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3824 page
= alloc_pages(gfp_mask
, order
);
3827 return (unsigned long) page_address(page
);
3829 EXPORT_SYMBOL(__get_free_pages
);
3831 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3833 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3835 EXPORT_SYMBOL(get_zeroed_page
);
3837 void __free_pages(struct page
*page
, unsigned int order
)
3839 if (put_page_testzero(page
)) {
3841 free_hot_cold_page(page
, false);
3843 __free_pages_ok(page
, order
);
3847 EXPORT_SYMBOL(__free_pages
);
3849 void free_pages(unsigned long addr
, unsigned int order
)
3852 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3853 __free_pages(virt_to_page((void *)addr
), order
);
3857 EXPORT_SYMBOL(free_pages
);
3861 * An arbitrary-length arbitrary-offset area of memory which resides
3862 * within a 0 or higher order page. Multiple fragments within that page
3863 * are individually refcounted, in the page's reference counter.
3865 * The page_frag functions below provide a simple allocation framework for
3866 * page fragments. This is used by the network stack and network device
3867 * drivers to provide a backing region of memory for use as either an
3868 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3870 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3873 struct page
*page
= NULL
;
3874 gfp_t gfp
= gfp_mask
;
3876 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3877 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3879 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3880 PAGE_FRAG_CACHE_MAX_ORDER
);
3881 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3883 if (unlikely(!page
))
3884 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3886 nc
->va
= page
? page_address(page
) : NULL
;
3891 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3892 unsigned int fragsz
, gfp_t gfp_mask
)
3894 unsigned int size
= PAGE_SIZE
;
3898 if (unlikely(!nc
->va
)) {
3900 page
= __page_frag_refill(nc
, gfp_mask
);
3904 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3905 /* if size can vary use size else just use PAGE_SIZE */
3908 /* Even if we own the page, we do not use atomic_set().
3909 * This would break get_page_unless_zero() users.
3911 page_ref_add(page
, size
- 1);
3913 /* reset page count bias and offset to start of new frag */
3914 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3915 nc
->pagecnt_bias
= size
;
3919 offset
= nc
->offset
- fragsz
;
3920 if (unlikely(offset
< 0)) {
3921 page
= virt_to_page(nc
->va
);
3923 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3926 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3927 /* if size can vary use size else just use PAGE_SIZE */
3930 /* OK, page count is 0, we can safely set it */
3931 set_page_count(page
, size
);
3933 /* reset page count bias and offset to start of new frag */
3934 nc
->pagecnt_bias
= size
;
3935 offset
= size
- fragsz
;
3939 nc
->offset
= offset
;
3941 return nc
->va
+ offset
;
3943 EXPORT_SYMBOL(__alloc_page_frag
);
3946 * Frees a page fragment allocated out of either a compound or order 0 page.
3948 void __free_page_frag(void *addr
)
3950 struct page
*page
= virt_to_head_page(addr
);
3952 if (unlikely(put_page_testzero(page
)))
3953 __free_pages_ok(page
, compound_order(page
));
3955 EXPORT_SYMBOL(__free_page_frag
);
3957 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3961 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3962 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3964 split_page(virt_to_page((void *)addr
), order
);
3965 while (used
< alloc_end
) {
3970 return (void *)addr
;
3974 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3975 * @size: the number of bytes to allocate
3976 * @gfp_mask: GFP flags for the allocation
3978 * This function is similar to alloc_pages(), except that it allocates the
3979 * minimum number of pages to satisfy the request. alloc_pages() can only
3980 * allocate memory in power-of-two pages.
3982 * This function is also limited by MAX_ORDER.
3984 * Memory allocated by this function must be released by free_pages_exact().
3986 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3988 unsigned int order
= get_order(size
);
3991 addr
= __get_free_pages(gfp_mask
, order
);
3992 return make_alloc_exact(addr
, order
, size
);
3994 EXPORT_SYMBOL(alloc_pages_exact
);
3997 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3999 * @nid: the preferred node ID where memory should be allocated
4000 * @size: the number of bytes to allocate
4001 * @gfp_mask: GFP flags for the allocation
4003 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4006 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4008 unsigned int order
= get_order(size
);
4009 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4012 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4016 * free_pages_exact - release memory allocated via alloc_pages_exact()
4017 * @virt: the value returned by alloc_pages_exact.
4018 * @size: size of allocation, same value as passed to alloc_pages_exact().
4020 * Release the memory allocated by a previous call to alloc_pages_exact.
4022 void free_pages_exact(void *virt
, size_t size
)
4024 unsigned long addr
= (unsigned long)virt
;
4025 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4027 while (addr
< end
) {
4032 EXPORT_SYMBOL(free_pages_exact
);
4035 * nr_free_zone_pages - count number of pages beyond high watermark
4036 * @offset: The zone index of the highest zone
4038 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4039 * high watermark within all zones at or below a given zone index. For each
4040 * zone, the number of pages is calculated as:
4041 * managed_pages - high_pages
4043 static unsigned long nr_free_zone_pages(int offset
)
4048 /* Just pick one node, since fallback list is circular */
4049 unsigned long sum
= 0;
4051 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4053 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4054 unsigned long size
= zone
->managed_pages
;
4055 unsigned long high
= high_wmark_pages(zone
);
4064 * nr_free_buffer_pages - count number of pages beyond high watermark
4066 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4067 * watermark within ZONE_DMA and ZONE_NORMAL.
4069 unsigned long nr_free_buffer_pages(void)
4071 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4073 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4076 * nr_free_pagecache_pages - count number of pages beyond high watermark
4078 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4079 * high watermark within all zones.
4081 unsigned long nr_free_pagecache_pages(void)
4083 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4086 static inline void show_node(struct zone
*zone
)
4088 if (IS_ENABLED(CONFIG_NUMA
))
4089 printk("Node %d ", zone_to_nid(zone
));
4092 long si_mem_available(void)
4095 unsigned long pagecache
;
4096 unsigned long wmark_low
= 0;
4097 unsigned long pages
[NR_LRU_LISTS
];
4101 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4102 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4105 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4108 * Estimate the amount of memory available for userspace allocations,
4109 * without causing swapping.
4111 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4114 * Not all the page cache can be freed, otherwise the system will
4115 * start swapping. Assume at least half of the page cache, or the
4116 * low watermark worth of cache, needs to stay.
4118 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4119 pagecache
-= min(pagecache
/ 2, wmark_low
);
4120 available
+= pagecache
;
4123 * Part of the reclaimable slab consists of items that are in use,
4124 * and cannot be freed. Cap this estimate at the low watermark.
4126 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4127 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4133 EXPORT_SYMBOL_GPL(si_mem_available
);
4135 void si_meminfo(struct sysinfo
*val
)
4137 val
->totalram
= totalram_pages
;
4138 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4139 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4140 val
->bufferram
= nr_blockdev_pages();
4141 val
->totalhigh
= totalhigh_pages
;
4142 val
->freehigh
= nr_free_highpages();
4143 val
->mem_unit
= PAGE_SIZE
;
4146 EXPORT_SYMBOL(si_meminfo
);
4149 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4151 int zone_type
; /* needs to be signed */
4152 unsigned long managed_pages
= 0;
4153 unsigned long managed_highpages
= 0;
4154 unsigned long free_highpages
= 0;
4155 pg_data_t
*pgdat
= NODE_DATA(nid
);
4157 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4158 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4159 val
->totalram
= managed_pages
;
4160 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4161 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4162 #ifdef CONFIG_HIGHMEM
4163 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4164 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4166 if (is_highmem(zone
)) {
4167 managed_highpages
+= zone
->managed_pages
;
4168 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4171 val
->totalhigh
= managed_highpages
;
4172 val
->freehigh
= free_highpages
;
4174 val
->totalhigh
= managed_highpages
;
4175 val
->freehigh
= free_highpages
;
4177 val
->mem_unit
= PAGE_SIZE
;
4182 * Determine whether the node should be displayed or not, depending on whether
4183 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4185 bool skip_free_areas_node(unsigned int flags
, int nid
)
4188 unsigned int cpuset_mems_cookie
;
4190 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4194 cpuset_mems_cookie
= read_mems_allowed_begin();
4195 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
4196 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
4201 #define K(x) ((x) << (PAGE_SHIFT-10))
4203 static void show_migration_types(unsigned char type
)
4205 static const char types
[MIGRATE_TYPES
] = {
4206 [MIGRATE_UNMOVABLE
] = 'U',
4207 [MIGRATE_MOVABLE
] = 'M',
4208 [MIGRATE_RECLAIMABLE
] = 'E',
4209 [MIGRATE_HIGHATOMIC
] = 'H',
4211 [MIGRATE_CMA
] = 'C',
4213 #ifdef CONFIG_MEMORY_ISOLATION
4214 [MIGRATE_ISOLATE
] = 'I',
4217 char tmp
[MIGRATE_TYPES
+ 1];
4221 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4222 if (type
& (1 << i
))
4227 printk("(%s) ", tmp
);
4231 * Show free area list (used inside shift_scroll-lock stuff)
4232 * We also calculate the percentage fragmentation. We do this by counting the
4233 * memory on each free list with the exception of the first item on the list.
4236 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4239 void show_free_areas(unsigned int filter
)
4241 unsigned long free_pcp
= 0;
4246 for_each_populated_zone(zone
) {
4247 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4250 for_each_online_cpu(cpu
)
4251 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4254 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4255 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4256 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4257 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4258 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4259 " free:%lu free_pcp:%lu free_cma:%lu\n",
4260 global_node_page_state(NR_ACTIVE_ANON
),
4261 global_node_page_state(NR_INACTIVE_ANON
),
4262 global_node_page_state(NR_ISOLATED_ANON
),
4263 global_node_page_state(NR_ACTIVE_FILE
),
4264 global_node_page_state(NR_INACTIVE_FILE
),
4265 global_node_page_state(NR_ISOLATED_FILE
),
4266 global_node_page_state(NR_UNEVICTABLE
),
4267 global_node_page_state(NR_FILE_DIRTY
),
4268 global_node_page_state(NR_WRITEBACK
),
4269 global_node_page_state(NR_UNSTABLE_NFS
),
4270 global_page_state(NR_SLAB_RECLAIMABLE
),
4271 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4272 global_node_page_state(NR_FILE_MAPPED
),
4273 global_node_page_state(NR_SHMEM
),
4274 global_page_state(NR_PAGETABLE
),
4275 global_page_state(NR_BOUNCE
),
4276 global_page_state(NR_FREE_PAGES
),
4278 global_page_state(NR_FREE_CMA_PAGES
));
4280 for_each_online_pgdat(pgdat
) {
4282 " active_anon:%lukB"
4283 " inactive_anon:%lukB"
4284 " active_file:%lukB"
4285 " inactive_file:%lukB"
4286 " unevictable:%lukB"
4287 " isolated(anon):%lukB"
4288 " isolated(file):%lukB"
4293 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4295 " shmem_pmdmapped: %lukB"
4298 " writeback_tmp:%lukB"
4300 " pages_scanned:%lu"
4301 " all_unreclaimable? %s"
4304 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4305 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4306 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4307 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4308 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4309 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4310 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4311 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4312 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4313 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4314 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4315 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4316 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4318 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4320 K(node_page_state(pgdat
, NR_SHMEM
)),
4321 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4322 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4323 node_page_state(pgdat
, NR_PAGES_SCANNED
),
4324 !pgdat_reclaimable(pgdat
) ? "yes" : "no");
4327 for_each_populated_zone(zone
) {
4330 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4334 for_each_online_cpu(cpu
)
4335 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4343 " active_anon:%lukB"
4344 " inactive_anon:%lukB"
4345 " active_file:%lukB"
4346 " inactive_file:%lukB"
4347 " unevictable:%lukB"
4348 " writepending:%lukB"
4352 " slab_reclaimable:%lukB"
4353 " slab_unreclaimable:%lukB"
4354 " kernel_stack:%lukB"
4362 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4363 K(min_wmark_pages(zone
)),
4364 K(low_wmark_pages(zone
)),
4365 K(high_wmark_pages(zone
)),
4366 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4367 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4368 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4369 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4370 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4371 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4372 K(zone
->present_pages
),
4373 K(zone
->managed_pages
),
4374 K(zone_page_state(zone
, NR_MLOCK
)),
4375 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4376 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4377 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4378 K(zone_page_state(zone
, NR_PAGETABLE
)),
4379 K(zone_page_state(zone
, NR_BOUNCE
)),
4381 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4382 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4383 printk("lowmem_reserve[]:");
4384 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4385 printk(" %ld", zone
->lowmem_reserve
[i
]);
4389 for_each_populated_zone(zone
) {
4391 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4392 unsigned char types
[MAX_ORDER
];
4394 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4397 printk("%s: ", zone
->name
);
4399 spin_lock_irqsave(&zone
->lock
, flags
);
4400 for (order
= 0; order
< MAX_ORDER
; order
++) {
4401 struct free_area
*area
= &zone
->free_area
[order
];
4404 nr
[order
] = area
->nr_free
;
4405 total
+= nr
[order
] << order
;
4408 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4409 if (!list_empty(&area
->free_list
[type
]))
4410 types
[order
] |= 1 << type
;
4413 spin_unlock_irqrestore(&zone
->lock
, flags
);
4414 for (order
= 0; order
< MAX_ORDER
; order
++) {
4415 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
4417 show_migration_types(types
[order
]);
4419 printk("= %lukB\n", K(total
));
4422 hugetlb_show_meminfo();
4424 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4426 show_swap_cache_info();
4429 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4431 zoneref
->zone
= zone
;
4432 zoneref
->zone_idx
= zone_idx(zone
);
4436 * Builds allocation fallback zone lists.
4438 * Add all populated zones of a node to the zonelist.
4440 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4444 enum zone_type zone_type
= MAX_NR_ZONES
;
4448 zone
= pgdat
->node_zones
+ zone_type
;
4449 if (managed_zone(zone
)) {
4450 zoneref_set_zone(zone
,
4451 &zonelist
->_zonerefs
[nr_zones
++]);
4452 check_highest_zone(zone_type
);
4454 } while (zone_type
);
4462 * 0 = automatic detection of better ordering.
4463 * 1 = order by ([node] distance, -zonetype)
4464 * 2 = order by (-zonetype, [node] distance)
4466 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4467 * the same zonelist. So only NUMA can configure this param.
4469 #define ZONELIST_ORDER_DEFAULT 0
4470 #define ZONELIST_ORDER_NODE 1
4471 #define ZONELIST_ORDER_ZONE 2
4473 /* zonelist order in the kernel.
4474 * set_zonelist_order() will set this to NODE or ZONE.
4476 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4477 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4481 /* The value user specified ....changed by config */
4482 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4483 /* string for sysctl */
4484 #define NUMA_ZONELIST_ORDER_LEN 16
4485 char numa_zonelist_order
[16] = "default";
4488 * interface for configure zonelist ordering.
4489 * command line option "numa_zonelist_order"
4490 * = "[dD]efault - default, automatic configuration.
4491 * = "[nN]ode - order by node locality, then by zone within node
4492 * = "[zZ]one - order by zone, then by locality within zone
4495 static int __parse_numa_zonelist_order(char *s
)
4497 if (*s
== 'd' || *s
== 'D') {
4498 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4499 } else if (*s
== 'n' || *s
== 'N') {
4500 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4501 } else if (*s
== 'z' || *s
== 'Z') {
4502 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4504 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4510 static __init
int setup_numa_zonelist_order(char *s
)
4517 ret
= __parse_numa_zonelist_order(s
);
4519 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4523 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4526 * sysctl handler for numa_zonelist_order
4528 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4529 void __user
*buffer
, size_t *length
,
4532 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4534 static DEFINE_MUTEX(zl_order_mutex
);
4536 mutex_lock(&zl_order_mutex
);
4538 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4542 strcpy(saved_string
, (char *)table
->data
);
4544 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4548 int oldval
= user_zonelist_order
;
4550 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4553 * bogus value. restore saved string
4555 strncpy((char *)table
->data
, saved_string
,
4556 NUMA_ZONELIST_ORDER_LEN
);
4557 user_zonelist_order
= oldval
;
4558 } else if (oldval
!= user_zonelist_order
) {
4559 mutex_lock(&zonelists_mutex
);
4560 build_all_zonelists(NULL
, NULL
);
4561 mutex_unlock(&zonelists_mutex
);
4565 mutex_unlock(&zl_order_mutex
);
4570 #define MAX_NODE_LOAD (nr_online_nodes)
4571 static int node_load
[MAX_NUMNODES
];
4574 * find_next_best_node - find the next node that should appear in a given node's fallback list
4575 * @node: node whose fallback list we're appending
4576 * @used_node_mask: nodemask_t of already used nodes
4578 * We use a number of factors to determine which is the next node that should
4579 * appear on a given node's fallback list. The node should not have appeared
4580 * already in @node's fallback list, and it should be the next closest node
4581 * according to the distance array (which contains arbitrary distance values
4582 * from each node to each node in the system), and should also prefer nodes
4583 * with no CPUs, since presumably they'll have very little allocation pressure
4584 * on them otherwise.
4585 * It returns -1 if no node is found.
4587 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4590 int min_val
= INT_MAX
;
4591 int best_node
= NUMA_NO_NODE
;
4592 const struct cpumask
*tmp
= cpumask_of_node(0);
4594 /* Use the local node if we haven't already */
4595 if (!node_isset(node
, *used_node_mask
)) {
4596 node_set(node
, *used_node_mask
);
4600 for_each_node_state(n
, N_MEMORY
) {
4602 /* Don't want a node to appear more than once */
4603 if (node_isset(n
, *used_node_mask
))
4606 /* Use the distance array to find the distance */
4607 val
= node_distance(node
, n
);
4609 /* Penalize nodes under us ("prefer the next node") */
4612 /* Give preference to headless and unused nodes */
4613 tmp
= cpumask_of_node(n
);
4614 if (!cpumask_empty(tmp
))
4615 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4617 /* Slight preference for less loaded node */
4618 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4619 val
+= node_load
[n
];
4621 if (val
< min_val
) {
4628 node_set(best_node
, *used_node_mask
);
4635 * Build zonelists ordered by node and zones within node.
4636 * This results in maximum locality--normal zone overflows into local
4637 * DMA zone, if any--but risks exhausting DMA zone.
4639 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4642 struct zonelist
*zonelist
;
4644 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4645 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4647 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4648 zonelist
->_zonerefs
[j
].zone
= NULL
;
4649 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4653 * Build gfp_thisnode zonelists
4655 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4658 struct zonelist
*zonelist
;
4660 zonelist
= &pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
];
4661 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4662 zonelist
->_zonerefs
[j
].zone
= NULL
;
4663 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4667 * Build zonelists ordered by zone and nodes within zones.
4668 * This results in conserving DMA zone[s] until all Normal memory is
4669 * exhausted, but results in overflowing to remote node while memory
4670 * may still exist in local DMA zone.
4672 static int node_order
[MAX_NUMNODES
];
4674 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4677 int zone_type
; /* needs to be signed */
4679 struct zonelist
*zonelist
;
4681 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4683 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4684 for (j
= 0; j
< nr_nodes
; j
++) {
4685 node
= node_order
[j
];
4686 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4687 if (managed_zone(z
)) {
4689 &zonelist
->_zonerefs
[pos
++]);
4690 check_highest_zone(zone_type
);
4694 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4695 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4698 #if defined(CONFIG_64BIT)
4700 * Devices that require DMA32/DMA are relatively rare and do not justify a
4701 * penalty to every machine in case the specialised case applies. Default
4702 * to Node-ordering on 64-bit NUMA machines
4704 static int default_zonelist_order(void)
4706 return ZONELIST_ORDER_NODE
;
4710 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4711 * by the kernel. If processes running on node 0 deplete the low memory zone
4712 * then reclaim will occur more frequency increasing stalls and potentially
4713 * be easier to OOM if a large percentage of the zone is under writeback or
4714 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4715 * Hence, default to zone ordering on 32-bit.
4717 static int default_zonelist_order(void)
4719 return ZONELIST_ORDER_ZONE
;
4721 #endif /* CONFIG_64BIT */
4723 static void set_zonelist_order(void)
4725 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4726 current_zonelist_order
= default_zonelist_order();
4728 current_zonelist_order
= user_zonelist_order
;
4731 static void build_zonelists(pg_data_t
*pgdat
)
4734 nodemask_t used_mask
;
4735 int local_node
, prev_node
;
4736 struct zonelist
*zonelist
;
4737 unsigned int order
= current_zonelist_order
;
4739 /* initialize zonelists */
4740 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4741 zonelist
= pgdat
->node_zonelists
+ i
;
4742 zonelist
->_zonerefs
[0].zone
= NULL
;
4743 zonelist
->_zonerefs
[0].zone_idx
= 0;
4746 /* NUMA-aware ordering of nodes */
4747 local_node
= pgdat
->node_id
;
4748 load
= nr_online_nodes
;
4749 prev_node
= local_node
;
4750 nodes_clear(used_mask
);
4752 memset(node_order
, 0, sizeof(node_order
));
4755 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4757 * We don't want to pressure a particular node.
4758 * So adding penalty to the first node in same
4759 * distance group to make it round-robin.
4761 if (node_distance(local_node
, node
) !=
4762 node_distance(local_node
, prev_node
))
4763 node_load
[node
] = load
;
4767 if (order
== ZONELIST_ORDER_NODE
)
4768 build_zonelists_in_node_order(pgdat
, node
);
4770 node_order
[i
++] = node
; /* remember order */
4773 if (order
== ZONELIST_ORDER_ZONE
) {
4774 /* calculate node order -- i.e., DMA last! */
4775 build_zonelists_in_zone_order(pgdat
, i
);
4778 build_thisnode_zonelists(pgdat
);
4781 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4783 * Return node id of node used for "local" allocations.
4784 * I.e., first node id of first zone in arg node's generic zonelist.
4785 * Used for initializing percpu 'numa_mem', which is used primarily
4786 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4788 int local_memory_node(int node
)
4792 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4793 gfp_zone(GFP_KERNEL
),
4795 return z
->zone
->node
;
4799 static void setup_min_unmapped_ratio(void);
4800 static void setup_min_slab_ratio(void);
4801 #else /* CONFIG_NUMA */
4803 static void set_zonelist_order(void)
4805 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4808 static void build_zonelists(pg_data_t
*pgdat
)
4810 int node
, local_node
;
4812 struct zonelist
*zonelist
;
4814 local_node
= pgdat
->node_id
;
4816 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4817 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4820 * Now we build the zonelist so that it contains the zones
4821 * of all the other nodes.
4822 * We don't want to pressure a particular node, so when
4823 * building the zones for node N, we make sure that the
4824 * zones coming right after the local ones are those from
4825 * node N+1 (modulo N)
4827 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4828 if (!node_online(node
))
4830 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4832 for (node
= 0; node
< local_node
; node
++) {
4833 if (!node_online(node
))
4835 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4838 zonelist
->_zonerefs
[j
].zone
= NULL
;
4839 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4842 #endif /* CONFIG_NUMA */
4845 * Boot pageset table. One per cpu which is going to be used for all
4846 * zones and all nodes. The parameters will be set in such a way
4847 * that an item put on a list will immediately be handed over to
4848 * the buddy list. This is safe since pageset manipulation is done
4849 * with interrupts disabled.
4851 * The boot_pagesets must be kept even after bootup is complete for
4852 * unused processors and/or zones. They do play a role for bootstrapping
4853 * hotplugged processors.
4855 * zoneinfo_show() and maybe other functions do
4856 * not check if the processor is online before following the pageset pointer.
4857 * Other parts of the kernel may not check if the zone is available.
4859 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4860 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4861 static void setup_zone_pageset(struct zone
*zone
);
4864 * Global mutex to protect against size modification of zonelists
4865 * as well as to serialize pageset setup for the new populated zone.
4867 DEFINE_MUTEX(zonelists_mutex
);
4869 /* return values int ....just for stop_machine() */
4870 static int __build_all_zonelists(void *data
)
4874 pg_data_t
*self
= data
;
4877 memset(node_load
, 0, sizeof(node_load
));
4880 if (self
&& !node_online(self
->node_id
)) {
4881 build_zonelists(self
);
4884 for_each_online_node(nid
) {
4885 pg_data_t
*pgdat
= NODE_DATA(nid
);
4887 build_zonelists(pgdat
);
4891 * Initialize the boot_pagesets that are going to be used
4892 * for bootstrapping processors. The real pagesets for
4893 * each zone will be allocated later when the per cpu
4894 * allocator is available.
4896 * boot_pagesets are used also for bootstrapping offline
4897 * cpus if the system is already booted because the pagesets
4898 * are needed to initialize allocators on a specific cpu too.
4899 * F.e. the percpu allocator needs the page allocator which
4900 * needs the percpu allocator in order to allocate its pagesets
4901 * (a chicken-egg dilemma).
4903 for_each_possible_cpu(cpu
) {
4904 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4906 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4908 * We now know the "local memory node" for each node--
4909 * i.e., the node of the first zone in the generic zonelist.
4910 * Set up numa_mem percpu variable for on-line cpus. During
4911 * boot, only the boot cpu should be on-line; we'll init the
4912 * secondary cpus' numa_mem as they come on-line. During
4913 * node/memory hotplug, we'll fixup all on-line cpus.
4915 if (cpu_online(cpu
))
4916 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4923 static noinline
void __init
4924 build_all_zonelists_init(void)
4926 __build_all_zonelists(NULL
);
4927 mminit_verify_zonelist();
4928 cpuset_init_current_mems_allowed();
4932 * Called with zonelists_mutex held always
4933 * unless system_state == SYSTEM_BOOTING.
4935 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4936 * [we're only called with non-NULL zone through __meminit paths] and
4937 * (2) call of __init annotated helper build_all_zonelists_init
4938 * [protected by SYSTEM_BOOTING].
4940 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4942 set_zonelist_order();
4944 if (system_state
== SYSTEM_BOOTING
) {
4945 build_all_zonelists_init();
4947 #ifdef CONFIG_MEMORY_HOTPLUG
4949 setup_zone_pageset(zone
);
4951 /* we have to stop all cpus to guarantee there is no user
4953 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4954 /* cpuset refresh routine should be here */
4956 vm_total_pages
= nr_free_pagecache_pages();
4958 * Disable grouping by mobility if the number of pages in the
4959 * system is too low to allow the mechanism to work. It would be
4960 * more accurate, but expensive to check per-zone. This check is
4961 * made on memory-hotadd so a system can start with mobility
4962 * disabled and enable it later
4964 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4965 page_group_by_mobility_disabled
= 1;
4967 page_group_by_mobility_disabled
= 0;
4969 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4971 zonelist_order_name
[current_zonelist_order
],
4972 page_group_by_mobility_disabled
? "off" : "on",
4975 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4980 * Initially all pages are reserved - free ones are freed
4981 * up by free_all_bootmem() once the early boot process is
4982 * done. Non-atomic initialization, single-pass.
4984 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4985 unsigned long start_pfn
, enum memmap_context context
)
4987 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
4988 unsigned long end_pfn
= start_pfn
+ size
;
4989 pg_data_t
*pgdat
= NODE_DATA(nid
);
4991 unsigned long nr_initialised
= 0;
4992 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4993 struct memblock_region
*r
= NULL
, *tmp
;
4996 if (highest_memmap_pfn
< end_pfn
- 1)
4997 highest_memmap_pfn
= end_pfn
- 1;
5000 * Honor reservation requested by the driver for this ZONE_DEVICE
5003 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5004 start_pfn
+= altmap
->reserve
;
5006 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5008 * There can be holes in boot-time mem_map[]s handed to this
5009 * function. They do not exist on hotplugged memory.
5011 if (context
!= MEMMAP_EARLY
)
5014 if (!early_pfn_valid(pfn
))
5016 if (!early_pfn_in_nid(pfn
, nid
))
5018 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5021 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5023 * Check given memblock attribute by firmware which can affect
5024 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5025 * mirrored, it's an overlapped memmap init. skip it.
5027 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5028 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5029 for_each_memblock(memory
, tmp
)
5030 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5034 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5035 memblock_is_mirror(r
)) {
5036 /* already initialized as NORMAL */
5037 pfn
= memblock_region_memory_end_pfn(r
);
5045 * Mark the block movable so that blocks are reserved for
5046 * movable at startup. This will force kernel allocations
5047 * to reserve their blocks rather than leaking throughout
5048 * the address space during boot when many long-lived
5049 * kernel allocations are made.
5051 * bitmap is created for zone's valid pfn range. but memmap
5052 * can be created for invalid pages (for alignment)
5053 * check here not to call set_pageblock_migratetype() against
5056 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5057 struct page
*page
= pfn_to_page(pfn
);
5059 __init_single_page(page
, pfn
, zone
, nid
);
5060 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5062 __init_single_pfn(pfn
, zone
, nid
);
5067 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5069 unsigned int order
, t
;
5070 for_each_migratetype_order(order
, t
) {
5071 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5072 zone
->free_area
[order
].nr_free
= 0;
5076 #ifndef __HAVE_ARCH_MEMMAP_INIT
5077 #define memmap_init(size, nid, zone, start_pfn) \
5078 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5081 static int zone_batchsize(struct zone
*zone
)
5087 * The per-cpu-pages pools are set to around 1000th of the
5088 * size of the zone. But no more than 1/2 of a meg.
5090 * OK, so we don't know how big the cache is. So guess.
5092 batch
= zone
->managed_pages
/ 1024;
5093 if (batch
* PAGE_SIZE
> 512 * 1024)
5094 batch
= (512 * 1024) / PAGE_SIZE
;
5095 batch
/= 4; /* We effectively *= 4 below */
5100 * Clamp the batch to a 2^n - 1 value. Having a power
5101 * of 2 value was found to be more likely to have
5102 * suboptimal cache aliasing properties in some cases.
5104 * For example if 2 tasks are alternately allocating
5105 * batches of pages, one task can end up with a lot
5106 * of pages of one half of the possible page colors
5107 * and the other with pages of the other colors.
5109 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5114 /* The deferral and batching of frees should be suppressed under NOMMU
5117 * The problem is that NOMMU needs to be able to allocate large chunks
5118 * of contiguous memory as there's no hardware page translation to
5119 * assemble apparent contiguous memory from discontiguous pages.
5121 * Queueing large contiguous runs of pages for batching, however,
5122 * causes the pages to actually be freed in smaller chunks. As there
5123 * can be a significant delay between the individual batches being
5124 * recycled, this leads to the once large chunks of space being
5125 * fragmented and becoming unavailable for high-order allocations.
5132 * pcp->high and pcp->batch values are related and dependent on one another:
5133 * ->batch must never be higher then ->high.
5134 * The following function updates them in a safe manner without read side
5137 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5138 * those fields changing asynchronously (acording the the above rule).
5140 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5141 * outside of boot time (or some other assurance that no concurrent updaters
5144 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5145 unsigned long batch
)
5147 /* start with a fail safe value for batch */
5151 /* Update high, then batch, in order */
5158 /* a companion to pageset_set_high() */
5159 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5161 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5164 static void pageset_init(struct per_cpu_pageset
*p
)
5166 struct per_cpu_pages
*pcp
;
5169 memset(p
, 0, sizeof(*p
));
5173 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5174 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5177 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5180 pageset_set_batch(p
, batch
);
5184 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5185 * to the value high for the pageset p.
5187 static void pageset_set_high(struct per_cpu_pageset
*p
,
5190 unsigned long batch
= max(1UL, high
/ 4);
5191 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5192 batch
= PAGE_SHIFT
* 8;
5194 pageset_update(&p
->pcp
, high
, batch
);
5197 static void pageset_set_high_and_batch(struct zone
*zone
,
5198 struct per_cpu_pageset
*pcp
)
5200 if (percpu_pagelist_fraction
)
5201 pageset_set_high(pcp
,
5202 (zone
->managed_pages
/
5203 percpu_pagelist_fraction
));
5205 pageset_set_batch(pcp
, zone_batchsize(zone
));
5208 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5210 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5213 pageset_set_high_and_batch(zone
, pcp
);
5216 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5219 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5220 for_each_possible_cpu(cpu
)
5221 zone_pageset_init(zone
, cpu
);
5225 * Allocate per cpu pagesets and initialize them.
5226 * Before this call only boot pagesets were available.
5228 void __init
setup_per_cpu_pageset(void)
5230 struct pglist_data
*pgdat
;
5233 for_each_populated_zone(zone
)
5234 setup_zone_pageset(zone
);
5236 for_each_online_pgdat(pgdat
)
5237 pgdat
->per_cpu_nodestats
=
5238 alloc_percpu(struct per_cpu_nodestat
);
5241 static __meminit
void zone_pcp_init(struct zone
*zone
)
5244 * per cpu subsystem is not up at this point. The following code
5245 * relies on the ability of the linker to provide the
5246 * offset of a (static) per cpu variable into the per cpu area.
5248 zone
->pageset
= &boot_pageset
;
5250 if (populated_zone(zone
))
5251 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5252 zone
->name
, zone
->present_pages
,
5253 zone_batchsize(zone
));
5256 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5257 unsigned long zone_start_pfn
,
5260 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5262 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5264 zone
->zone_start_pfn
= zone_start_pfn
;
5266 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5267 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5269 (unsigned long)zone_idx(zone
),
5270 zone_start_pfn
, (zone_start_pfn
+ size
));
5272 zone_init_free_lists(zone
);
5273 zone
->initialized
= 1;
5278 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5279 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5282 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5284 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5285 struct mminit_pfnnid_cache
*state
)
5287 unsigned long start_pfn
, end_pfn
;
5290 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5291 return state
->last_nid
;
5293 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5295 state
->last_start
= start_pfn
;
5296 state
->last_end
= end_pfn
;
5297 state
->last_nid
= nid
;
5302 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5305 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5306 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5307 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5309 * If an architecture guarantees that all ranges registered contain no holes
5310 * and may be freed, this this function may be used instead of calling
5311 * memblock_free_early_nid() manually.
5313 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5315 unsigned long start_pfn
, end_pfn
;
5318 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5319 start_pfn
= min(start_pfn
, max_low_pfn
);
5320 end_pfn
= min(end_pfn
, max_low_pfn
);
5322 if (start_pfn
< end_pfn
)
5323 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5324 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5330 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5331 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5333 * If an architecture guarantees that all ranges registered contain no holes and may
5334 * be freed, this function may be used instead of calling memory_present() manually.
5336 void __init
sparse_memory_present_with_active_regions(int nid
)
5338 unsigned long start_pfn
, end_pfn
;
5341 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5342 memory_present(this_nid
, start_pfn
, end_pfn
);
5346 * get_pfn_range_for_nid - Return the start and end page frames for a node
5347 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5348 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5349 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5351 * It returns the start and end page frame of a node based on information
5352 * provided by memblock_set_node(). If called for a node
5353 * with no available memory, a warning is printed and the start and end
5356 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5357 unsigned long *start_pfn
, unsigned long *end_pfn
)
5359 unsigned long this_start_pfn
, this_end_pfn
;
5365 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5366 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5367 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5370 if (*start_pfn
== -1UL)
5375 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5376 * assumption is made that zones within a node are ordered in monotonic
5377 * increasing memory addresses so that the "highest" populated zone is used
5379 static void __init
find_usable_zone_for_movable(void)
5382 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5383 if (zone_index
== ZONE_MOVABLE
)
5386 if (arch_zone_highest_possible_pfn
[zone_index
] >
5387 arch_zone_lowest_possible_pfn
[zone_index
])
5391 VM_BUG_ON(zone_index
== -1);
5392 movable_zone
= zone_index
;
5396 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5397 * because it is sized independent of architecture. Unlike the other zones,
5398 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5399 * in each node depending on the size of each node and how evenly kernelcore
5400 * is distributed. This helper function adjusts the zone ranges
5401 * provided by the architecture for a given node by using the end of the
5402 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5403 * zones within a node are in order of monotonic increases memory addresses
5405 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5406 unsigned long zone_type
,
5407 unsigned long node_start_pfn
,
5408 unsigned long node_end_pfn
,
5409 unsigned long *zone_start_pfn
,
5410 unsigned long *zone_end_pfn
)
5412 /* Only adjust if ZONE_MOVABLE is on this node */
5413 if (zone_movable_pfn
[nid
]) {
5414 /* Size ZONE_MOVABLE */
5415 if (zone_type
== ZONE_MOVABLE
) {
5416 *zone_start_pfn
= zone_movable_pfn
[nid
];
5417 *zone_end_pfn
= min(node_end_pfn
,
5418 arch_zone_highest_possible_pfn
[movable_zone
]);
5420 /* Adjust for ZONE_MOVABLE starting within this range */
5421 } else if (!mirrored_kernelcore
&&
5422 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5423 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5424 *zone_end_pfn
= zone_movable_pfn
[nid
];
5426 /* Check if this whole range is within ZONE_MOVABLE */
5427 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5428 *zone_start_pfn
= *zone_end_pfn
;
5433 * Return the number of pages a zone spans in a node, including holes
5434 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5436 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5437 unsigned long zone_type
,
5438 unsigned long node_start_pfn
,
5439 unsigned long node_end_pfn
,
5440 unsigned long *zone_start_pfn
,
5441 unsigned long *zone_end_pfn
,
5442 unsigned long *ignored
)
5444 /* When hotadd a new node from cpu_up(), the node should be empty */
5445 if (!node_start_pfn
&& !node_end_pfn
)
5448 /* Get the start and end of the zone */
5449 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5450 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5451 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5452 node_start_pfn
, node_end_pfn
,
5453 zone_start_pfn
, zone_end_pfn
);
5455 /* Check that this node has pages within the zone's required range */
5456 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5459 /* Move the zone boundaries inside the node if necessary */
5460 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5461 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5463 /* Return the spanned pages */
5464 return *zone_end_pfn
- *zone_start_pfn
;
5468 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5469 * then all holes in the requested range will be accounted for.
5471 unsigned long __meminit
__absent_pages_in_range(int nid
,
5472 unsigned long range_start_pfn
,
5473 unsigned long range_end_pfn
)
5475 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5476 unsigned long start_pfn
, end_pfn
;
5479 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5480 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5481 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5482 nr_absent
-= end_pfn
- start_pfn
;
5488 * absent_pages_in_range - Return number of page frames in holes within a range
5489 * @start_pfn: The start PFN to start searching for holes
5490 * @end_pfn: The end PFN to stop searching for holes
5492 * It returns the number of pages frames in memory holes within a range.
5494 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5495 unsigned long end_pfn
)
5497 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5500 /* Return the number of page frames in holes in a zone on a node */
5501 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5502 unsigned long zone_type
,
5503 unsigned long node_start_pfn
,
5504 unsigned long node_end_pfn
,
5505 unsigned long *ignored
)
5507 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5508 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5509 unsigned long zone_start_pfn
, zone_end_pfn
;
5510 unsigned long nr_absent
;
5512 /* When hotadd a new node from cpu_up(), the node should be empty */
5513 if (!node_start_pfn
&& !node_end_pfn
)
5516 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5517 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5519 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5520 node_start_pfn
, node_end_pfn
,
5521 &zone_start_pfn
, &zone_end_pfn
);
5522 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5525 * ZONE_MOVABLE handling.
5526 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5529 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5530 unsigned long start_pfn
, end_pfn
;
5531 struct memblock_region
*r
;
5533 for_each_memblock(memory
, r
) {
5534 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5535 zone_start_pfn
, zone_end_pfn
);
5536 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5537 zone_start_pfn
, zone_end_pfn
);
5539 if (zone_type
== ZONE_MOVABLE
&&
5540 memblock_is_mirror(r
))
5541 nr_absent
+= end_pfn
- start_pfn
;
5543 if (zone_type
== ZONE_NORMAL
&&
5544 !memblock_is_mirror(r
))
5545 nr_absent
+= end_pfn
- start_pfn
;
5552 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5553 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5554 unsigned long zone_type
,
5555 unsigned long node_start_pfn
,
5556 unsigned long node_end_pfn
,
5557 unsigned long *zone_start_pfn
,
5558 unsigned long *zone_end_pfn
,
5559 unsigned long *zones_size
)
5563 *zone_start_pfn
= node_start_pfn
;
5564 for (zone
= 0; zone
< zone_type
; zone
++)
5565 *zone_start_pfn
+= zones_size
[zone
];
5567 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5569 return zones_size
[zone_type
];
5572 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5573 unsigned long zone_type
,
5574 unsigned long node_start_pfn
,
5575 unsigned long node_end_pfn
,
5576 unsigned long *zholes_size
)
5581 return zholes_size
[zone_type
];
5584 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5586 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5587 unsigned long node_start_pfn
,
5588 unsigned long node_end_pfn
,
5589 unsigned long *zones_size
,
5590 unsigned long *zholes_size
)
5592 unsigned long realtotalpages
= 0, totalpages
= 0;
5595 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5596 struct zone
*zone
= pgdat
->node_zones
+ i
;
5597 unsigned long zone_start_pfn
, zone_end_pfn
;
5598 unsigned long size
, real_size
;
5600 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5606 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5607 node_start_pfn
, node_end_pfn
,
5610 zone
->zone_start_pfn
= zone_start_pfn
;
5612 zone
->zone_start_pfn
= 0;
5613 zone
->spanned_pages
= size
;
5614 zone
->present_pages
= real_size
;
5617 realtotalpages
+= real_size
;
5620 pgdat
->node_spanned_pages
= totalpages
;
5621 pgdat
->node_present_pages
= realtotalpages
;
5622 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5626 #ifndef CONFIG_SPARSEMEM
5628 * Calculate the size of the zone->blockflags rounded to an unsigned long
5629 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5630 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5631 * round what is now in bits to nearest long in bits, then return it in
5634 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5636 unsigned long usemapsize
;
5638 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5639 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5640 usemapsize
= usemapsize
>> pageblock_order
;
5641 usemapsize
*= NR_PAGEBLOCK_BITS
;
5642 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5644 return usemapsize
/ 8;
5647 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5649 unsigned long zone_start_pfn
,
5650 unsigned long zonesize
)
5652 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5653 zone
->pageblock_flags
= NULL
;
5655 zone
->pageblock_flags
=
5656 memblock_virt_alloc_node_nopanic(usemapsize
,
5660 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5661 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5662 #endif /* CONFIG_SPARSEMEM */
5664 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5666 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5667 void __paginginit
set_pageblock_order(void)
5671 /* Check that pageblock_nr_pages has not already been setup */
5672 if (pageblock_order
)
5675 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5676 order
= HUGETLB_PAGE_ORDER
;
5678 order
= MAX_ORDER
- 1;
5681 * Assume the largest contiguous order of interest is a huge page.
5682 * This value may be variable depending on boot parameters on IA64 and
5685 pageblock_order
= order
;
5687 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5690 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5691 * is unused as pageblock_order is set at compile-time. See
5692 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5695 void __paginginit
set_pageblock_order(void)
5699 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5701 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5702 unsigned long present_pages
)
5704 unsigned long pages
= spanned_pages
;
5707 * Provide a more accurate estimation if there are holes within
5708 * the zone and SPARSEMEM is in use. If there are holes within the
5709 * zone, each populated memory region may cost us one or two extra
5710 * memmap pages due to alignment because memmap pages for each
5711 * populated regions may not naturally algined on page boundary.
5712 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5714 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5715 IS_ENABLED(CONFIG_SPARSEMEM
))
5716 pages
= present_pages
;
5718 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5722 * Set up the zone data structures:
5723 * - mark all pages reserved
5724 * - mark all memory queues empty
5725 * - clear the memory bitmaps
5727 * NOTE: pgdat should get zeroed by caller.
5729 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5732 int nid
= pgdat
->node_id
;
5735 pgdat_resize_init(pgdat
);
5736 #ifdef CONFIG_NUMA_BALANCING
5737 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5738 pgdat
->numabalancing_migrate_nr_pages
= 0;
5739 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5741 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5742 spin_lock_init(&pgdat
->split_queue_lock
);
5743 INIT_LIST_HEAD(&pgdat
->split_queue
);
5744 pgdat
->split_queue_len
= 0;
5746 init_waitqueue_head(&pgdat
->kswapd_wait
);
5747 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5748 #ifdef CONFIG_COMPACTION
5749 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5751 pgdat_page_ext_init(pgdat
);
5752 spin_lock_init(&pgdat
->lru_lock
);
5753 lruvec_init(node_lruvec(pgdat
));
5755 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5756 struct zone
*zone
= pgdat
->node_zones
+ j
;
5757 unsigned long size
, realsize
, freesize
, memmap_pages
;
5758 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5760 size
= zone
->spanned_pages
;
5761 realsize
= freesize
= zone
->present_pages
;
5764 * Adjust freesize so that it accounts for how much memory
5765 * is used by this zone for memmap. This affects the watermark
5766 * and per-cpu initialisations
5768 memmap_pages
= calc_memmap_size(size
, realsize
);
5769 if (!is_highmem_idx(j
)) {
5770 if (freesize
>= memmap_pages
) {
5771 freesize
-= memmap_pages
;
5774 " %s zone: %lu pages used for memmap\n",
5775 zone_names
[j
], memmap_pages
);
5777 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5778 zone_names
[j
], memmap_pages
, freesize
);
5781 /* Account for reserved pages */
5782 if (j
== 0 && freesize
> dma_reserve
) {
5783 freesize
-= dma_reserve
;
5784 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5785 zone_names
[0], dma_reserve
);
5788 if (!is_highmem_idx(j
))
5789 nr_kernel_pages
+= freesize
;
5790 /* Charge for highmem memmap if there are enough kernel pages */
5791 else if (nr_kernel_pages
> memmap_pages
* 2)
5792 nr_kernel_pages
-= memmap_pages
;
5793 nr_all_pages
+= freesize
;
5796 * Set an approximate value for lowmem here, it will be adjusted
5797 * when the bootmem allocator frees pages into the buddy system.
5798 * And all highmem pages will be managed by the buddy system.
5800 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5804 zone
->name
= zone_names
[j
];
5805 zone
->zone_pgdat
= pgdat
;
5806 spin_lock_init(&zone
->lock
);
5807 zone_seqlock_init(zone
);
5808 zone_pcp_init(zone
);
5813 set_pageblock_order();
5814 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5815 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5817 memmap_init(size
, nid
, j
, zone_start_pfn
);
5821 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
5823 unsigned long __maybe_unused start
= 0;
5824 unsigned long __maybe_unused offset
= 0;
5826 /* Skip empty nodes */
5827 if (!pgdat
->node_spanned_pages
)
5830 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5831 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5832 offset
= pgdat
->node_start_pfn
- start
;
5833 /* ia64 gets its own node_mem_map, before this, without bootmem */
5834 if (!pgdat
->node_mem_map
) {
5835 unsigned long size
, end
;
5839 * The zone's endpoints aren't required to be MAX_ORDER
5840 * aligned but the node_mem_map endpoints must be in order
5841 * for the buddy allocator to function correctly.
5843 end
= pgdat_end_pfn(pgdat
);
5844 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5845 size
= (end
- start
) * sizeof(struct page
);
5846 map
= alloc_remap(pgdat
->node_id
, size
);
5848 map
= memblock_virt_alloc_node_nopanic(size
,
5850 pgdat
->node_mem_map
= map
+ offset
;
5852 #ifndef CONFIG_NEED_MULTIPLE_NODES
5854 * With no DISCONTIG, the global mem_map is just set as node 0's
5856 if (pgdat
== NODE_DATA(0)) {
5857 mem_map
= NODE_DATA(0)->node_mem_map
;
5858 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5859 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5861 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5864 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5867 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5868 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5870 pg_data_t
*pgdat
= NODE_DATA(nid
);
5871 unsigned long start_pfn
= 0;
5872 unsigned long end_pfn
= 0;
5874 /* pg_data_t should be reset to zero when it's allocated */
5875 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
5877 reset_deferred_meminit(pgdat
);
5878 pgdat
->node_id
= nid
;
5879 pgdat
->node_start_pfn
= node_start_pfn
;
5880 pgdat
->per_cpu_nodestats
= NULL
;
5881 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5882 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5883 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5884 (u64
)start_pfn
<< PAGE_SHIFT
,
5885 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5887 start_pfn
= node_start_pfn
;
5889 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5890 zones_size
, zholes_size
);
5892 alloc_node_mem_map(pgdat
);
5893 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5894 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5895 nid
, (unsigned long)pgdat
,
5896 (unsigned long)pgdat
->node_mem_map
);
5899 free_area_init_core(pgdat
);
5902 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5904 #if MAX_NUMNODES > 1
5906 * Figure out the number of possible node ids.
5908 void __init
setup_nr_node_ids(void)
5910 unsigned int highest
;
5912 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5913 nr_node_ids
= highest
+ 1;
5918 * node_map_pfn_alignment - determine the maximum internode alignment
5920 * This function should be called after node map is populated and sorted.
5921 * It calculates the maximum power of two alignment which can distinguish
5924 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5925 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5926 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5927 * shifted, 1GiB is enough and this function will indicate so.
5929 * This is used to test whether pfn -> nid mapping of the chosen memory
5930 * model has fine enough granularity to avoid incorrect mapping for the
5931 * populated node map.
5933 * Returns the determined alignment in pfn's. 0 if there is no alignment
5934 * requirement (single node).
5936 unsigned long __init
node_map_pfn_alignment(void)
5938 unsigned long accl_mask
= 0, last_end
= 0;
5939 unsigned long start
, end
, mask
;
5943 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5944 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5951 * Start with a mask granular enough to pin-point to the
5952 * start pfn and tick off bits one-by-one until it becomes
5953 * too coarse to separate the current node from the last.
5955 mask
= ~((1 << __ffs(start
)) - 1);
5956 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5959 /* accumulate all internode masks */
5963 /* convert mask to number of pages */
5964 return ~accl_mask
+ 1;
5967 /* Find the lowest pfn for a node */
5968 static unsigned long __init
find_min_pfn_for_node(int nid
)
5970 unsigned long min_pfn
= ULONG_MAX
;
5971 unsigned long start_pfn
;
5974 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5975 min_pfn
= min(min_pfn
, start_pfn
);
5977 if (min_pfn
== ULONG_MAX
) {
5978 pr_warn("Could not find start_pfn for node %d\n", nid
);
5986 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5988 * It returns the minimum PFN based on information provided via
5989 * memblock_set_node().
5991 unsigned long __init
find_min_pfn_with_active_regions(void)
5993 return find_min_pfn_for_node(MAX_NUMNODES
);
5997 * early_calculate_totalpages()
5998 * Sum pages in active regions for movable zone.
5999 * Populate N_MEMORY for calculating usable_nodes.
6001 static unsigned long __init
early_calculate_totalpages(void)
6003 unsigned long totalpages
= 0;
6004 unsigned long start_pfn
, end_pfn
;
6007 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6008 unsigned long pages
= end_pfn
- start_pfn
;
6010 totalpages
+= pages
;
6012 node_set_state(nid
, N_MEMORY
);
6018 * Find the PFN the Movable zone begins in each node. Kernel memory
6019 * is spread evenly between nodes as long as the nodes have enough
6020 * memory. When they don't, some nodes will have more kernelcore than
6023 static void __init
find_zone_movable_pfns_for_nodes(void)
6026 unsigned long usable_startpfn
;
6027 unsigned long kernelcore_node
, kernelcore_remaining
;
6028 /* save the state before borrow the nodemask */
6029 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6030 unsigned long totalpages
= early_calculate_totalpages();
6031 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6032 struct memblock_region
*r
;
6034 /* Need to find movable_zone earlier when movable_node is specified. */
6035 find_usable_zone_for_movable();
6038 * If movable_node is specified, ignore kernelcore and movablecore
6041 if (movable_node_is_enabled()) {
6042 for_each_memblock(memory
, r
) {
6043 if (!memblock_is_hotpluggable(r
))
6048 usable_startpfn
= PFN_DOWN(r
->base
);
6049 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6050 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6058 * If kernelcore=mirror is specified, ignore movablecore option
6060 if (mirrored_kernelcore
) {
6061 bool mem_below_4gb_not_mirrored
= false;
6063 for_each_memblock(memory
, r
) {
6064 if (memblock_is_mirror(r
))
6069 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6071 if (usable_startpfn
< 0x100000) {
6072 mem_below_4gb_not_mirrored
= true;
6076 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6077 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6081 if (mem_below_4gb_not_mirrored
)
6082 pr_warn("This configuration results in unmirrored kernel memory.");
6088 * If movablecore=nn[KMG] was specified, calculate what size of
6089 * kernelcore that corresponds so that memory usable for
6090 * any allocation type is evenly spread. If both kernelcore
6091 * and movablecore are specified, then the value of kernelcore
6092 * will be used for required_kernelcore if it's greater than
6093 * what movablecore would have allowed.
6095 if (required_movablecore
) {
6096 unsigned long corepages
;
6099 * Round-up so that ZONE_MOVABLE is at least as large as what
6100 * was requested by the user
6102 required_movablecore
=
6103 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6104 required_movablecore
= min(totalpages
, required_movablecore
);
6105 corepages
= totalpages
- required_movablecore
;
6107 required_kernelcore
= max(required_kernelcore
, corepages
);
6111 * If kernelcore was not specified or kernelcore size is larger
6112 * than totalpages, there is no ZONE_MOVABLE.
6114 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6117 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6118 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6121 /* Spread kernelcore memory as evenly as possible throughout nodes */
6122 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6123 for_each_node_state(nid
, N_MEMORY
) {
6124 unsigned long start_pfn
, end_pfn
;
6127 * Recalculate kernelcore_node if the division per node
6128 * now exceeds what is necessary to satisfy the requested
6129 * amount of memory for the kernel
6131 if (required_kernelcore
< kernelcore_node
)
6132 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6135 * As the map is walked, we track how much memory is usable
6136 * by the kernel using kernelcore_remaining. When it is
6137 * 0, the rest of the node is usable by ZONE_MOVABLE
6139 kernelcore_remaining
= kernelcore_node
;
6141 /* Go through each range of PFNs within this node */
6142 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6143 unsigned long size_pages
;
6145 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6146 if (start_pfn
>= end_pfn
)
6149 /* Account for what is only usable for kernelcore */
6150 if (start_pfn
< usable_startpfn
) {
6151 unsigned long kernel_pages
;
6152 kernel_pages
= min(end_pfn
, usable_startpfn
)
6155 kernelcore_remaining
-= min(kernel_pages
,
6156 kernelcore_remaining
);
6157 required_kernelcore
-= min(kernel_pages
,
6158 required_kernelcore
);
6160 /* Continue if range is now fully accounted */
6161 if (end_pfn
<= usable_startpfn
) {
6164 * Push zone_movable_pfn to the end so
6165 * that if we have to rebalance
6166 * kernelcore across nodes, we will
6167 * not double account here
6169 zone_movable_pfn
[nid
] = end_pfn
;
6172 start_pfn
= usable_startpfn
;
6176 * The usable PFN range for ZONE_MOVABLE is from
6177 * start_pfn->end_pfn. Calculate size_pages as the
6178 * number of pages used as kernelcore
6180 size_pages
= end_pfn
- start_pfn
;
6181 if (size_pages
> kernelcore_remaining
)
6182 size_pages
= kernelcore_remaining
;
6183 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6186 * Some kernelcore has been met, update counts and
6187 * break if the kernelcore for this node has been
6190 required_kernelcore
-= min(required_kernelcore
,
6192 kernelcore_remaining
-= size_pages
;
6193 if (!kernelcore_remaining
)
6199 * If there is still required_kernelcore, we do another pass with one
6200 * less node in the count. This will push zone_movable_pfn[nid] further
6201 * along on the nodes that still have memory until kernelcore is
6205 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6209 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6210 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6211 zone_movable_pfn
[nid
] =
6212 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6215 /* restore the node_state */
6216 node_states
[N_MEMORY
] = saved_node_state
;
6219 /* Any regular or high memory on that node ? */
6220 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6222 enum zone_type zone_type
;
6224 if (N_MEMORY
== N_NORMAL_MEMORY
)
6227 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6228 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6229 if (populated_zone(zone
)) {
6230 node_set_state(nid
, N_HIGH_MEMORY
);
6231 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6232 zone_type
<= ZONE_NORMAL
)
6233 node_set_state(nid
, N_NORMAL_MEMORY
);
6240 * free_area_init_nodes - Initialise all pg_data_t and zone data
6241 * @max_zone_pfn: an array of max PFNs for each zone
6243 * This will call free_area_init_node() for each active node in the system.
6244 * Using the page ranges provided by memblock_set_node(), the size of each
6245 * zone in each node and their holes is calculated. If the maximum PFN
6246 * between two adjacent zones match, it is assumed that the zone is empty.
6247 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6248 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6249 * starts where the previous one ended. For example, ZONE_DMA32 starts
6250 * at arch_max_dma_pfn.
6252 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6254 unsigned long start_pfn
, end_pfn
;
6257 /* Record where the zone boundaries are */
6258 memset(arch_zone_lowest_possible_pfn
, 0,
6259 sizeof(arch_zone_lowest_possible_pfn
));
6260 memset(arch_zone_highest_possible_pfn
, 0,
6261 sizeof(arch_zone_highest_possible_pfn
));
6263 start_pfn
= find_min_pfn_with_active_regions();
6265 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6266 if (i
== ZONE_MOVABLE
)
6269 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6270 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6271 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6273 start_pfn
= end_pfn
;
6275 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6276 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6278 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6279 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6280 find_zone_movable_pfns_for_nodes();
6282 /* Print out the zone ranges */
6283 pr_info("Zone ranges:\n");
6284 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6285 if (i
== ZONE_MOVABLE
)
6287 pr_info(" %-8s ", zone_names
[i
]);
6288 if (arch_zone_lowest_possible_pfn
[i
] ==
6289 arch_zone_highest_possible_pfn
[i
])
6292 pr_cont("[mem %#018Lx-%#018Lx]\n",
6293 (u64
)arch_zone_lowest_possible_pfn
[i
]
6295 ((u64
)arch_zone_highest_possible_pfn
[i
]
6296 << PAGE_SHIFT
) - 1);
6299 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6300 pr_info("Movable zone start for each node\n");
6301 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6302 if (zone_movable_pfn
[i
])
6303 pr_info(" Node %d: %#018Lx\n", i
,
6304 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6307 /* Print out the early node map */
6308 pr_info("Early memory node ranges\n");
6309 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6310 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6311 (u64
)start_pfn
<< PAGE_SHIFT
,
6312 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6314 /* Initialise every node */
6315 mminit_verify_pageflags_layout();
6316 setup_nr_node_ids();
6317 for_each_online_node(nid
) {
6318 pg_data_t
*pgdat
= NODE_DATA(nid
);
6319 free_area_init_node(nid
, NULL
,
6320 find_min_pfn_for_node(nid
), NULL
);
6322 /* Any memory on that node */
6323 if (pgdat
->node_present_pages
)
6324 node_set_state(nid
, N_MEMORY
);
6325 check_for_memory(pgdat
, nid
);
6329 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6331 unsigned long long coremem
;
6335 coremem
= memparse(p
, &p
);
6336 *core
= coremem
>> PAGE_SHIFT
;
6338 /* Paranoid check that UL is enough for the coremem value */
6339 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6345 * kernelcore=size sets the amount of memory for use for allocations that
6346 * cannot be reclaimed or migrated.
6348 static int __init
cmdline_parse_kernelcore(char *p
)
6350 /* parse kernelcore=mirror */
6351 if (parse_option_str(p
, "mirror")) {
6352 mirrored_kernelcore
= true;
6356 return cmdline_parse_core(p
, &required_kernelcore
);
6360 * movablecore=size sets the amount of memory for use for allocations that
6361 * can be reclaimed or migrated.
6363 static int __init
cmdline_parse_movablecore(char *p
)
6365 return cmdline_parse_core(p
, &required_movablecore
);
6368 early_param("kernelcore", cmdline_parse_kernelcore
);
6369 early_param("movablecore", cmdline_parse_movablecore
);
6371 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6373 void adjust_managed_page_count(struct page
*page
, long count
)
6375 spin_lock(&managed_page_count_lock
);
6376 page_zone(page
)->managed_pages
+= count
;
6377 totalram_pages
+= count
;
6378 #ifdef CONFIG_HIGHMEM
6379 if (PageHighMem(page
))
6380 totalhigh_pages
+= count
;
6382 spin_unlock(&managed_page_count_lock
);
6384 EXPORT_SYMBOL(adjust_managed_page_count
);
6386 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6389 unsigned long pages
= 0;
6391 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6392 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6393 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6394 if ((unsigned int)poison
<= 0xFF)
6395 memset(pos
, poison
, PAGE_SIZE
);
6396 free_reserved_page(virt_to_page(pos
));
6400 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6401 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6405 EXPORT_SYMBOL(free_reserved_area
);
6407 #ifdef CONFIG_HIGHMEM
6408 void free_highmem_page(struct page
*page
)
6410 __free_reserved_page(page
);
6412 page_zone(page
)->managed_pages
++;
6418 void __init
mem_init_print_info(const char *str
)
6420 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6421 unsigned long init_code_size
, init_data_size
;
6423 physpages
= get_num_physpages();
6424 codesize
= _etext
- _stext
;
6425 datasize
= _edata
- _sdata
;
6426 rosize
= __end_rodata
- __start_rodata
;
6427 bss_size
= __bss_stop
- __bss_start
;
6428 init_data_size
= __init_end
- __init_begin
;
6429 init_code_size
= _einittext
- _sinittext
;
6432 * Detect special cases and adjust section sizes accordingly:
6433 * 1) .init.* may be embedded into .data sections
6434 * 2) .init.text.* may be out of [__init_begin, __init_end],
6435 * please refer to arch/tile/kernel/vmlinux.lds.S.
6436 * 3) .rodata.* may be embedded into .text or .data sections.
6438 #define adj_init_size(start, end, size, pos, adj) \
6440 if (start <= pos && pos < end && size > adj) \
6444 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6445 _sinittext
, init_code_size
);
6446 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6447 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6448 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6449 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6451 #undef adj_init_size
6453 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6454 #ifdef CONFIG_HIGHMEM
6458 nr_free_pages() << (PAGE_SHIFT
- 10),
6459 physpages
<< (PAGE_SHIFT
- 10),
6460 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6461 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6462 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6463 totalcma_pages
<< (PAGE_SHIFT
- 10),
6464 #ifdef CONFIG_HIGHMEM
6465 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6467 str
? ", " : "", str
? str
: "");
6471 * set_dma_reserve - set the specified number of pages reserved in the first zone
6472 * @new_dma_reserve: The number of pages to mark reserved
6474 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6475 * In the DMA zone, a significant percentage may be consumed by kernel image
6476 * and other unfreeable allocations which can skew the watermarks badly. This
6477 * function may optionally be used to account for unfreeable pages in the
6478 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6479 * smaller per-cpu batchsize.
6481 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6483 dma_reserve
= new_dma_reserve
;
6486 void __init
free_area_init(unsigned long *zones_size
)
6488 free_area_init_node(0, zones_size
,
6489 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6492 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6493 unsigned long action
, void *hcpu
)
6495 int cpu
= (unsigned long)hcpu
;
6497 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6498 lru_add_drain_cpu(cpu
);
6502 * Spill the event counters of the dead processor
6503 * into the current processors event counters.
6504 * This artificially elevates the count of the current
6507 vm_events_fold_cpu(cpu
);
6510 * Zero the differential counters of the dead processor
6511 * so that the vm statistics are consistent.
6513 * This is only okay since the processor is dead and cannot
6514 * race with what we are doing.
6516 cpu_vm_stats_fold(cpu
);
6521 void __init
page_alloc_init(void)
6523 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6527 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6528 * or min_free_kbytes changes.
6530 static void calculate_totalreserve_pages(void)
6532 struct pglist_data
*pgdat
;
6533 unsigned long reserve_pages
= 0;
6534 enum zone_type i
, j
;
6536 for_each_online_pgdat(pgdat
) {
6538 pgdat
->totalreserve_pages
= 0;
6540 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6541 struct zone
*zone
= pgdat
->node_zones
+ i
;
6544 /* Find valid and maximum lowmem_reserve in the zone */
6545 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6546 if (zone
->lowmem_reserve
[j
] > max
)
6547 max
= zone
->lowmem_reserve
[j
];
6550 /* we treat the high watermark as reserved pages. */
6551 max
+= high_wmark_pages(zone
);
6553 if (max
> zone
->managed_pages
)
6554 max
= zone
->managed_pages
;
6556 pgdat
->totalreserve_pages
+= max
;
6558 reserve_pages
+= max
;
6561 totalreserve_pages
= reserve_pages
;
6565 * setup_per_zone_lowmem_reserve - called whenever
6566 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6567 * has a correct pages reserved value, so an adequate number of
6568 * pages are left in the zone after a successful __alloc_pages().
6570 static void setup_per_zone_lowmem_reserve(void)
6572 struct pglist_data
*pgdat
;
6573 enum zone_type j
, idx
;
6575 for_each_online_pgdat(pgdat
) {
6576 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6577 struct zone
*zone
= pgdat
->node_zones
+ j
;
6578 unsigned long managed_pages
= zone
->managed_pages
;
6580 zone
->lowmem_reserve
[j
] = 0;
6584 struct zone
*lower_zone
;
6588 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6589 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6591 lower_zone
= pgdat
->node_zones
+ idx
;
6592 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6593 sysctl_lowmem_reserve_ratio
[idx
];
6594 managed_pages
+= lower_zone
->managed_pages
;
6599 /* update totalreserve_pages */
6600 calculate_totalreserve_pages();
6603 static void __setup_per_zone_wmarks(void)
6605 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6606 unsigned long lowmem_pages
= 0;
6608 unsigned long flags
;
6610 /* Calculate total number of !ZONE_HIGHMEM pages */
6611 for_each_zone(zone
) {
6612 if (!is_highmem(zone
))
6613 lowmem_pages
+= zone
->managed_pages
;
6616 for_each_zone(zone
) {
6619 spin_lock_irqsave(&zone
->lock
, flags
);
6620 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6621 do_div(tmp
, lowmem_pages
);
6622 if (is_highmem(zone
)) {
6624 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6625 * need highmem pages, so cap pages_min to a small
6628 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6629 * deltas control asynch page reclaim, and so should
6630 * not be capped for highmem.
6632 unsigned long min_pages
;
6634 min_pages
= zone
->managed_pages
/ 1024;
6635 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6636 zone
->watermark
[WMARK_MIN
] = min_pages
;
6639 * If it's a lowmem zone, reserve a number of pages
6640 * proportionate to the zone's size.
6642 zone
->watermark
[WMARK_MIN
] = tmp
;
6646 * Set the kswapd watermarks distance according to the
6647 * scale factor in proportion to available memory, but
6648 * ensure a minimum size on small systems.
6650 tmp
= max_t(u64
, tmp
>> 2,
6651 mult_frac(zone
->managed_pages
,
6652 watermark_scale_factor
, 10000));
6654 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6655 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6657 spin_unlock_irqrestore(&zone
->lock
, flags
);
6660 /* update totalreserve_pages */
6661 calculate_totalreserve_pages();
6665 * setup_per_zone_wmarks - called when min_free_kbytes changes
6666 * or when memory is hot-{added|removed}
6668 * Ensures that the watermark[min,low,high] values for each zone are set
6669 * correctly with respect to min_free_kbytes.
6671 void setup_per_zone_wmarks(void)
6673 mutex_lock(&zonelists_mutex
);
6674 __setup_per_zone_wmarks();
6675 mutex_unlock(&zonelists_mutex
);
6679 * Initialise min_free_kbytes.
6681 * For small machines we want it small (128k min). For large machines
6682 * we want it large (64MB max). But it is not linear, because network
6683 * bandwidth does not increase linearly with machine size. We use
6685 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6686 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6702 int __meminit
init_per_zone_wmark_min(void)
6704 unsigned long lowmem_kbytes
;
6705 int new_min_free_kbytes
;
6707 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6708 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6710 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6711 min_free_kbytes
= new_min_free_kbytes
;
6712 if (min_free_kbytes
< 128)
6713 min_free_kbytes
= 128;
6714 if (min_free_kbytes
> 65536)
6715 min_free_kbytes
= 65536;
6717 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6718 new_min_free_kbytes
, user_min_free_kbytes
);
6720 setup_per_zone_wmarks();
6721 refresh_zone_stat_thresholds();
6722 setup_per_zone_lowmem_reserve();
6725 setup_min_unmapped_ratio();
6726 setup_min_slab_ratio();
6731 core_initcall(init_per_zone_wmark_min
)
6734 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6735 * that we can call two helper functions whenever min_free_kbytes
6738 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6739 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6743 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6748 user_min_free_kbytes
= min_free_kbytes
;
6749 setup_per_zone_wmarks();
6754 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6755 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6759 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6764 setup_per_zone_wmarks();
6770 static void setup_min_unmapped_ratio(void)
6775 for_each_online_pgdat(pgdat
)
6776 pgdat
->min_unmapped_pages
= 0;
6779 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
6780 sysctl_min_unmapped_ratio
) / 100;
6784 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6785 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6789 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6793 setup_min_unmapped_ratio();
6798 static void setup_min_slab_ratio(void)
6803 for_each_online_pgdat(pgdat
)
6804 pgdat
->min_slab_pages
= 0;
6807 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
6808 sysctl_min_slab_ratio
) / 100;
6811 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6812 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6816 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6820 setup_min_slab_ratio();
6827 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6828 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6829 * whenever sysctl_lowmem_reserve_ratio changes.
6831 * The reserve ratio obviously has absolutely no relation with the
6832 * minimum watermarks. The lowmem reserve ratio can only make sense
6833 * if in function of the boot time zone sizes.
6835 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6836 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6838 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6839 setup_per_zone_lowmem_reserve();
6844 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6845 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6846 * pagelist can have before it gets flushed back to buddy allocator.
6848 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6849 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6852 int old_percpu_pagelist_fraction
;
6855 mutex_lock(&pcp_batch_high_lock
);
6856 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6858 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6859 if (!write
|| ret
< 0)
6862 /* Sanity checking to avoid pcp imbalance */
6863 if (percpu_pagelist_fraction
&&
6864 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6865 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6871 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6874 for_each_populated_zone(zone
) {
6877 for_each_possible_cpu(cpu
)
6878 pageset_set_high_and_batch(zone
,
6879 per_cpu_ptr(zone
->pageset
, cpu
));
6882 mutex_unlock(&pcp_batch_high_lock
);
6887 int hashdist
= HASHDIST_DEFAULT
;
6889 static int __init
set_hashdist(char *str
)
6893 hashdist
= simple_strtoul(str
, &str
, 0);
6896 __setup("hashdist=", set_hashdist
);
6899 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6901 * Returns the number of pages that arch has reserved but
6902 * is not known to alloc_large_system_hash().
6904 static unsigned long __init
arch_reserved_kernel_pages(void)
6911 * allocate a large system hash table from bootmem
6912 * - it is assumed that the hash table must contain an exact power-of-2
6913 * quantity of entries
6914 * - limit is the number of hash buckets, not the total allocation size
6916 void *__init
alloc_large_system_hash(const char *tablename
,
6917 unsigned long bucketsize
,
6918 unsigned long numentries
,
6921 unsigned int *_hash_shift
,
6922 unsigned int *_hash_mask
,
6923 unsigned long low_limit
,
6924 unsigned long high_limit
)
6926 unsigned long long max
= high_limit
;
6927 unsigned long log2qty
, size
;
6930 /* allow the kernel cmdline to have a say */
6932 /* round applicable memory size up to nearest megabyte */
6933 numentries
= nr_kernel_pages
;
6934 numentries
-= arch_reserved_kernel_pages();
6936 /* It isn't necessary when PAGE_SIZE >= 1MB */
6937 if (PAGE_SHIFT
< 20)
6938 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6940 /* limit to 1 bucket per 2^scale bytes of low memory */
6941 if (scale
> PAGE_SHIFT
)
6942 numentries
>>= (scale
- PAGE_SHIFT
);
6944 numentries
<<= (PAGE_SHIFT
- scale
);
6946 /* Make sure we've got at least a 0-order allocation.. */
6947 if (unlikely(flags
& HASH_SMALL
)) {
6948 /* Makes no sense without HASH_EARLY */
6949 WARN_ON(!(flags
& HASH_EARLY
));
6950 if (!(numentries
>> *_hash_shift
)) {
6951 numentries
= 1UL << *_hash_shift
;
6952 BUG_ON(!numentries
);
6954 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6955 numentries
= PAGE_SIZE
/ bucketsize
;
6957 numentries
= roundup_pow_of_two(numentries
);
6959 /* limit allocation size to 1/16 total memory by default */
6961 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6962 do_div(max
, bucketsize
);
6964 max
= min(max
, 0x80000000ULL
);
6966 if (numentries
< low_limit
)
6967 numentries
= low_limit
;
6968 if (numentries
> max
)
6971 log2qty
= ilog2(numentries
);
6974 size
= bucketsize
<< log2qty
;
6975 if (flags
& HASH_EARLY
)
6976 table
= memblock_virt_alloc_nopanic(size
, 0);
6978 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6981 * If bucketsize is not a power-of-two, we may free
6982 * some pages at the end of hash table which
6983 * alloc_pages_exact() automatically does
6985 if (get_order(size
) < MAX_ORDER
) {
6986 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6987 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6990 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6993 panic("Failed to allocate %s hash table\n", tablename
);
6995 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6996 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
6999 *_hash_shift
= log2qty
;
7001 *_hash_mask
= (1 << log2qty
) - 1;
7007 * This function checks whether pageblock includes unmovable pages or not.
7008 * If @count is not zero, it is okay to include less @count unmovable pages
7010 * PageLRU check without isolation or lru_lock could race so that
7011 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7012 * expect this function should be exact.
7014 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7015 bool skip_hwpoisoned_pages
)
7017 unsigned long pfn
, iter
, found
;
7021 * For avoiding noise data, lru_add_drain_all() should be called
7022 * If ZONE_MOVABLE, the zone never contains unmovable pages
7024 if (zone_idx(zone
) == ZONE_MOVABLE
)
7026 mt
= get_pageblock_migratetype(page
);
7027 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7030 pfn
= page_to_pfn(page
);
7031 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7032 unsigned long check
= pfn
+ iter
;
7034 if (!pfn_valid_within(check
))
7037 page
= pfn_to_page(check
);
7040 * Hugepages are not in LRU lists, but they're movable.
7041 * We need not scan over tail pages bacause we don't
7042 * handle each tail page individually in migration.
7044 if (PageHuge(page
)) {
7045 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7050 * We can't use page_count without pin a page
7051 * because another CPU can free compound page.
7052 * This check already skips compound tails of THP
7053 * because their page->_refcount is zero at all time.
7055 if (!page_ref_count(page
)) {
7056 if (PageBuddy(page
))
7057 iter
+= (1 << page_order(page
)) - 1;
7062 * The HWPoisoned page may be not in buddy system, and
7063 * page_count() is not 0.
7065 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7071 * If there are RECLAIMABLE pages, we need to check
7072 * it. But now, memory offline itself doesn't call
7073 * shrink_node_slabs() and it still to be fixed.
7076 * If the page is not RAM, page_count()should be 0.
7077 * we don't need more check. This is an _used_ not-movable page.
7079 * The problematic thing here is PG_reserved pages. PG_reserved
7080 * is set to both of a memory hole page and a _used_ kernel
7089 bool is_pageblock_removable_nolock(struct page
*page
)
7095 * We have to be careful here because we are iterating over memory
7096 * sections which are not zone aware so we might end up outside of
7097 * the zone but still within the section.
7098 * We have to take care about the node as well. If the node is offline
7099 * its NODE_DATA will be NULL - see page_zone.
7101 if (!node_online(page_to_nid(page
)))
7104 zone
= page_zone(page
);
7105 pfn
= page_to_pfn(page
);
7106 if (!zone_spans_pfn(zone
, pfn
))
7109 return !has_unmovable_pages(zone
, page
, 0, true);
7112 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7114 static unsigned long pfn_max_align_down(unsigned long pfn
)
7116 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7117 pageblock_nr_pages
) - 1);
7120 static unsigned long pfn_max_align_up(unsigned long pfn
)
7122 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7123 pageblock_nr_pages
));
7126 /* [start, end) must belong to a single zone. */
7127 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7128 unsigned long start
, unsigned long end
)
7130 /* This function is based on compact_zone() from compaction.c. */
7131 unsigned long nr_reclaimed
;
7132 unsigned long pfn
= start
;
7133 unsigned int tries
= 0;
7138 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7139 if (fatal_signal_pending(current
)) {
7144 if (list_empty(&cc
->migratepages
)) {
7145 cc
->nr_migratepages
= 0;
7146 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7152 } else if (++tries
== 5) {
7153 ret
= ret
< 0 ? ret
: -EBUSY
;
7157 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7159 cc
->nr_migratepages
-= nr_reclaimed
;
7161 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7162 NULL
, 0, cc
->mode
, MR_CMA
);
7165 putback_movable_pages(&cc
->migratepages
);
7172 * alloc_contig_range() -- tries to allocate given range of pages
7173 * @start: start PFN to allocate
7174 * @end: one-past-the-last PFN to allocate
7175 * @migratetype: migratetype of the underlaying pageblocks (either
7176 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7177 * in range must have the same migratetype and it must
7178 * be either of the two.
7180 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7181 * aligned, however it's the caller's responsibility to guarantee that
7182 * we are the only thread that changes migrate type of pageblocks the
7185 * The PFN range must belong to a single zone.
7187 * Returns zero on success or negative error code. On success all
7188 * pages which PFN is in [start, end) are allocated for the caller and
7189 * need to be freed with free_contig_range().
7191 int alloc_contig_range(unsigned long start
, unsigned long end
,
7192 unsigned migratetype
)
7194 unsigned long outer_start
, outer_end
;
7198 struct compact_control cc
= {
7199 .nr_migratepages
= 0,
7201 .zone
= page_zone(pfn_to_page(start
)),
7202 .mode
= MIGRATE_SYNC
,
7203 .ignore_skip_hint
= true,
7205 INIT_LIST_HEAD(&cc
.migratepages
);
7208 * What we do here is we mark all pageblocks in range as
7209 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7210 * have different sizes, and due to the way page allocator
7211 * work, we align the range to biggest of the two pages so
7212 * that page allocator won't try to merge buddies from
7213 * different pageblocks and change MIGRATE_ISOLATE to some
7214 * other migration type.
7216 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7217 * migrate the pages from an unaligned range (ie. pages that
7218 * we are interested in). This will put all the pages in
7219 * range back to page allocator as MIGRATE_ISOLATE.
7221 * When this is done, we take the pages in range from page
7222 * allocator removing them from the buddy system. This way
7223 * page allocator will never consider using them.
7225 * This lets us mark the pageblocks back as
7226 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7227 * aligned range but not in the unaligned, original range are
7228 * put back to page allocator so that buddy can use them.
7231 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7232 pfn_max_align_up(end
), migratetype
,
7238 * In case of -EBUSY, we'd like to know which page causes problem.
7239 * So, just fall through. We will check it in test_pages_isolated().
7241 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7242 if (ret
&& ret
!= -EBUSY
)
7246 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7247 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7248 * more, all pages in [start, end) are free in page allocator.
7249 * What we are going to do is to allocate all pages from
7250 * [start, end) (that is remove them from page allocator).
7252 * The only problem is that pages at the beginning and at the
7253 * end of interesting range may be not aligned with pages that
7254 * page allocator holds, ie. they can be part of higher order
7255 * pages. Because of this, we reserve the bigger range and
7256 * once this is done free the pages we are not interested in.
7258 * We don't have to hold zone->lock here because the pages are
7259 * isolated thus they won't get removed from buddy.
7262 lru_add_drain_all();
7263 drain_all_pages(cc
.zone
);
7266 outer_start
= start
;
7267 while (!PageBuddy(pfn_to_page(outer_start
))) {
7268 if (++order
>= MAX_ORDER
) {
7269 outer_start
= start
;
7272 outer_start
&= ~0UL << order
;
7275 if (outer_start
!= start
) {
7276 order
= page_order(pfn_to_page(outer_start
));
7279 * outer_start page could be small order buddy page and
7280 * it doesn't include start page. Adjust outer_start
7281 * in this case to report failed page properly
7282 * on tracepoint in test_pages_isolated()
7284 if (outer_start
+ (1UL << order
) <= start
)
7285 outer_start
= start
;
7288 /* Make sure the range is really isolated. */
7289 if (test_pages_isolated(outer_start
, end
, false)) {
7290 pr_info("%s: [%lx, %lx) PFNs busy\n",
7291 __func__
, outer_start
, end
);
7296 /* Grab isolated pages from freelists. */
7297 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7303 /* Free head and tail (if any) */
7304 if (start
!= outer_start
)
7305 free_contig_range(outer_start
, start
- outer_start
);
7306 if (end
!= outer_end
)
7307 free_contig_range(end
, outer_end
- end
);
7310 undo_isolate_page_range(pfn_max_align_down(start
),
7311 pfn_max_align_up(end
), migratetype
);
7315 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7317 unsigned int count
= 0;
7319 for (; nr_pages
--; pfn
++) {
7320 struct page
*page
= pfn_to_page(pfn
);
7322 count
+= page_count(page
) != 1;
7325 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7329 #ifdef CONFIG_MEMORY_HOTPLUG
7331 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7332 * page high values need to be recalulated.
7334 void __meminit
zone_pcp_update(struct zone
*zone
)
7337 mutex_lock(&pcp_batch_high_lock
);
7338 for_each_possible_cpu(cpu
)
7339 pageset_set_high_and_batch(zone
,
7340 per_cpu_ptr(zone
->pageset
, cpu
));
7341 mutex_unlock(&pcp_batch_high_lock
);
7345 void zone_pcp_reset(struct zone
*zone
)
7347 unsigned long flags
;
7349 struct per_cpu_pageset
*pset
;
7351 /* avoid races with drain_pages() */
7352 local_irq_save(flags
);
7353 if (zone
->pageset
!= &boot_pageset
) {
7354 for_each_online_cpu(cpu
) {
7355 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7356 drain_zonestat(zone
, pset
);
7358 free_percpu(zone
->pageset
);
7359 zone
->pageset
= &boot_pageset
;
7361 local_irq_restore(flags
);
7364 #ifdef CONFIG_MEMORY_HOTREMOVE
7366 * All pages in the range must be in a single zone and isolated
7367 * before calling this.
7370 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7374 unsigned int order
, i
;
7376 unsigned long flags
;
7377 /* find the first valid pfn */
7378 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7383 zone
= page_zone(pfn_to_page(pfn
));
7384 spin_lock_irqsave(&zone
->lock
, flags
);
7386 while (pfn
< end_pfn
) {
7387 if (!pfn_valid(pfn
)) {
7391 page
= pfn_to_page(pfn
);
7393 * The HWPoisoned page may be not in buddy system, and
7394 * page_count() is not 0.
7396 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7398 SetPageReserved(page
);
7402 BUG_ON(page_count(page
));
7403 BUG_ON(!PageBuddy(page
));
7404 order
= page_order(page
);
7405 #ifdef CONFIG_DEBUG_VM
7406 pr_info("remove from free list %lx %d %lx\n",
7407 pfn
, 1 << order
, end_pfn
);
7409 list_del(&page
->lru
);
7410 rmv_page_order(page
);
7411 zone
->free_area
[order
].nr_free
--;
7412 for (i
= 0; i
< (1 << order
); i
++)
7413 SetPageReserved((page
+i
));
7414 pfn
+= (1 << order
);
7416 spin_unlock_irqrestore(&zone
->lock
, flags
);
7420 bool is_free_buddy_page(struct page
*page
)
7422 struct zone
*zone
= page_zone(page
);
7423 unsigned long pfn
= page_to_pfn(page
);
7424 unsigned long flags
;
7427 spin_lock_irqsave(&zone
->lock
, flags
);
7428 for (order
= 0; order
< MAX_ORDER
; order
++) {
7429 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7431 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7434 spin_unlock_irqrestore(&zone
->lock
, flags
);
7436 return order
< MAX_ORDER
;