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
];
95 * Array of node states.
97 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
98 [N_POSSIBLE
] = NODE_MASK_ALL
,
99 [N_ONLINE
] = { { [0] = 1UL } },
101 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
102 #ifdef CONFIG_HIGHMEM
103 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
105 #ifdef CONFIG_MOVABLE_NODE
106 [N_MEMORY
] = { { [0] = 1UL } },
108 [N_CPU
] = { { [0] = 1UL } },
111 EXPORT_SYMBOL(node_states
);
113 /* Protect totalram_pages and zone->managed_pages */
114 static DEFINE_SPINLOCK(managed_page_count_lock
);
116 unsigned long totalram_pages __read_mostly
;
117 unsigned long totalreserve_pages __read_mostly
;
118 unsigned long totalcma_pages __read_mostly
;
120 int percpu_pagelist_fraction
;
121 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 * A cached value of the page's pageblock's migratetype, used when the page is
125 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
126 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
127 * Also the migratetype set in the page does not necessarily match the pcplist
128 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
129 * other index - this ensures that it will be put on the correct CMA freelist.
131 static inline int get_pcppage_migratetype(struct page
*page
)
136 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
138 page
->index
= migratetype
;
141 #ifdef CONFIG_PM_SLEEP
143 * The following functions are used by the suspend/hibernate code to temporarily
144 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
145 * while devices are suspended. To avoid races with the suspend/hibernate code,
146 * they should always be called with pm_mutex held (gfp_allowed_mask also should
147 * only be modified with pm_mutex held, unless the suspend/hibernate code is
148 * guaranteed not to run in parallel with that modification).
151 static gfp_t saved_gfp_mask
;
153 void pm_restore_gfp_mask(void)
155 WARN_ON(!mutex_is_locked(&pm_mutex
));
156 if (saved_gfp_mask
) {
157 gfp_allowed_mask
= saved_gfp_mask
;
162 void pm_restrict_gfp_mask(void)
164 WARN_ON(!mutex_is_locked(&pm_mutex
));
165 WARN_ON(saved_gfp_mask
);
166 saved_gfp_mask
= gfp_allowed_mask
;
167 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
170 bool pm_suspended_storage(void)
172 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
176 #endif /* CONFIG_PM_SLEEP */
178 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
179 unsigned int pageblock_order __read_mostly
;
182 static void __free_pages_ok(struct page
*page
, unsigned int order
);
185 * results with 256, 32 in the lowmem_reserve sysctl:
186 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
187 * 1G machine -> (16M dma, 784M normal, 224M high)
188 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
189 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
190 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
192 * TBD: should special case ZONE_DMA32 machines here - in those we normally
193 * don't need any ZONE_NORMAL reservation
195 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
196 #ifdef CONFIG_ZONE_DMA
199 #ifdef CONFIG_ZONE_DMA32
202 #ifdef CONFIG_HIGHMEM
208 EXPORT_SYMBOL(totalram_pages
);
210 static char * const zone_names
[MAX_NR_ZONES
] = {
211 #ifdef CONFIG_ZONE_DMA
214 #ifdef CONFIG_ZONE_DMA32
218 #ifdef CONFIG_HIGHMEM
222 #ifdef CONFIG_ZONE_DEVICE
227 char * const migratetype_names
[MIGRATE_TYPES
] = {
235 #ifdef CONFIG_MEMORY_ISOLATION
240 compound_page_dtor
* const compound_page_dtors
[] = {
243 #ifdef CONFIG_HUGETLB_PAGE
246 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
251 int min_free_kbytes
= 1024;
252 int user_min_free_kbytes
= -1;
253 int watermark_scale_factor
= 10;
255 static unsigned long __meminitdata nr_kernel_pages
;
256 static unsigned long __meminitdata nr_all_pages
;
257 static unsigned long __meminitdata dma_reserve
;
259 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
260 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
261 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
262 static unsigned long __initdata required_kernelcore
;
263 static unsigned long __initdata required_movablecore
;
264 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
265 static bool mirrored_kernelcore
;
267 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
269 EXPORT_SYMBOL(movable_zone
);
270 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
273 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
274 int nr_online_nodes __read_mostly
= 1;
275 EXPORT_SYMBOL(nr_node_ids
);
276 EXPORT_SYMBOL(nr_online_nodes
);
279 int page_group_by_mobility_disabled __read_mostly
;
281 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
282 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
284 pgdat
->first_deferred_pfn
= ULONG_MAX
;
287 /* Returns true if the struct page for the pfn is uninitialised */
288 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
290 int nid
= early_pfn_to_nid(pfn
);
292 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
299 * Returns false when the remaining initialisation should be deferred until
300 * later in the boot cycle when it can be parallelised.
302 static inline bool update_defer_init(pg_data_t
*pgdat
,
303 unsigned long pfn
, unsigned long zone_end
,
304 unsigned long *nr_initialised
)
306 unsigned long max_initialise
;
308 /* Always populate low zones for address-contrained allocations */
309 if (zone_end
< pgdat_end_pfn(pgdat
))
312 * Initialise at least 2G of a node but also take into account that
313 * two large system hashes that can take up 1GB for 0.25TB/node.
315 max_initialise
= max(2UL << (30 - PAGE_SHIFT
),
316 (pgdat
->node_spanned_pages
>> 8));
319 if ((*nr_initialised
> max_initialise
) &&
320 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
321 pgdat
->first_deferred_pfn
= pfn
;
328 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
332 static inline bool early_page_uninitialised(unsigned long pfn
)
337 static inline bool update_defer_init(pg_data_t
*pgdat
,
338 unsigned long pfn
, unsigned long zone_end
,
339 unsigned long *nr_initialised
)
345 /* Return a pointer to the bitmap storing bits affecting a block of pages */
346 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
349 #ifdef CONFIG_SPARSEMEM
350 return __pfn_to_section(pfn
)->pageblock_flags
;
352 return page_zone(page
)->pageblock_flags
;
353 #endif /* CONFIG_SPARSEMEM */
356 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
358 #ifdef CONFIG_SPARSEMEM
359 pfn
&= (PAGES_PER_SECTION
-1);
360 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
362 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
363 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
364 #endif /* CONFIG_SPARSEMEM */
368 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
369 * @page: The page within the block of interest
370 * @pfn: The target page frame number
371 * @end_bitidx: The last bit of interest to retrieve
372 * @mask: mask of bits that the caller is interested in
374 * Return: pageblock_bits flags
376 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
378 unsigned long end_bitidx
,
381 unsigned long *bitmap
;
382 unsigned long bitidx
, word_bitidx
;
385 bitmap
= get_pageblock_bitmap(page
, pfn
);
386 bitidx
= pfn_to_bitidx(page
, pfn
);
387 word_bitidx
= bitidx
/ BITS_PER_LONG
;
388 bitidx
&= (BITS_PER_LONG
-1);
390 word
= bitmap
[word_bitidx
];
391 bitidx
+= end_bitidx
;
392 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
395 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
396 unsigned long end_bitidx
,
399 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
402 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
404 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
408 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
409 * @page: The page within the block of interest
410 * @flags: The flags to set
411 * @pfn: The target page frame number
412 * @end_bitidx: The last bit of interest
413 * @mask: mask of bits that the caller is interested in
415 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
417 unsigned long end_bitidx
,
420 unsigned long *bitmap
;
421 unsigned long bitidx
, word_bitidx
;
422 unsigned long old_word
, word
;
424 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
426 bitmap
= get_pageblock_bitmap(page
, pfn
);
427 bitidx
= pfn_to_bitidx(page
, pfn
);
428 word_bitidx
= bitidx
/ BITS_PER_LONG
;
429 bitidx
&= (BITS_PER_LONG
-1);
431 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
433 bitidx
+= end_bitidx
;
434 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
435 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
437 word
= READ_ONCE(bitmap
[word_bitidx
]);
439 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
440 if (word
== old_word
)
446 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
448 if (unlikely(page_group_by_mobility_disabled
&&
449 migratetype
< MIGRATE_PCPTYPES
))
450 migratetype
= MIGRATE_UNMOVABLE
;
452 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
453 PB_migrate
, PB_migrate_end
);
456 #ifdef CONFIG_DEBUG_VM
457 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
461 unsigned long pfn
= page_to_pfn(page
);
462 unsigned long sp
, start_pfn
;
465 seq
= zone_span_seqbegin(zone
);
466 start_pfn
= zone
->zone_start_pfn
;
467 sp
= zone
->spanned_pages
;
468 if (!zone_spans_pfn(zone
, pfn
))
470 } while (zone_span_seqretry(zone
, seq
));
473 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
474 pfn
, zone_to_nid(zone
), zone
->name
,
475 start_pfn
, start_pfn
+ sp
);
480 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
482 if (!pfn_valid_within(page_to_pfn(page
)))
484 if (zone
!= page_zone(page
))
490 * Temporary debugging check for pages not lying within a given zone.
492 static int bad_range(struct zone
*zone
, struct page
*page
)
494 if (page_outside_zone_boundaries(zone
, page
))
496 if (!page_is_consistent(zone
, page
))
502 static inline int bad_range(struct zone
*zone
, struct page
*page
)
508 static void bad_page(struct page
*page
, const char *reason
,
509 unsigned long bad_flags
)
511 static unsigned long resume
;
512 static unsigned long nr_shown
;
513 static unsigned long nr_unshown
;
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
519 if (nr_shown
== 60) {
520 if (time_before(jiffies
, resume
)) {
526 "BUG: Bad page state: %lu messages suppressed\n",
533 resume
= jiffies
+ 60 * HZ
;
535 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
536 current
->comm
, page_to_pfn(page
));
537 __dump_page(page
, reason
);
538 bad_flags
&= page
->flags
;
540 pr_alert("bad because of flags: %#lx(%pGp)\n",
541 bad_flags
, &bad_flags
);
542 dump_page_owner(page
);
547 /* Leave bad fields for debug, except PageBuddy could make trouble */
548 page_mapcount_reset(page
); /* remove PageBuddy */
549 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
553 * Higher-order pages are called "compound pages". They are structured thusly:
555 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
557 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
558 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
560 * The first tail page's ->compound_dtor holds the offset in array of compound
561 * page destructors. See compound_page_dtors.
563 * The first tail page's ->compound_order holds the order of allocation.
564 * This usage means that zero-order pages may not be compound.
567 void free_compound_page(struct page
*page
)
569 __free_pages_ok(page
, compound_order(page
));
572 void prep_compound_page(struct page
*page
, unsigned int order
)
575 int nr_pages
= 1 << order
;
577 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
578 set_compound_order(page
, order
);
580 for (i
= 1; i
< nr_pages
; i
++) {
581 struct page
*p
= page
+ i
;
582 set_page_count(p
, 0);
583 p
->mapping
= TAIL_MAPPING
;
584 set_compound_head(p
, page
);
586 atomic_set(compound_mapcount_ptr(page
), -1);
589 #ifdef CONFIG_DEBUG_PAGEALLOC
590 unsigned int _debug_guardpage_minorder
;
591 bool _debug_pagealloc_enabled __read_mostly
592 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
593 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
594 bool _debug_guardpage_enabled __read_mostly
;
596 static int __init
early_debug_pagealloc(char *buf
)
600 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
602 early_param("debug_pagealloc", early_debug_pagealloc
);
604 static bool need_debug_guardpage(void)
606 /* If we don't use debug_pagealloc, we don't need guard page */
607 if (!debug_pagealloc_enabled())
613 static void init_debug_guardpage(void)
615 if (!debug_pagealloc_enabled())
618 _debug_guardpage_enabled
= true;
621 struct page_ext_operations debug_guardpage_ops
= {
622 .need
= need_debug_guardpage
,
623 .init
= init_debug_guardpage
,
626 static int __init
debug_guardpage_minorder_setup(char *buf
)
630 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
631 pr_err("Bad debug_guardpage_minorder value\n");
634 _debug_guardpage_minorder
= res
;
635 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
638 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
640 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
641 unsigned int order
, int migratetype
)
643 struct page_ext
*page_ext
;
645 if (!debug_guardpage_enabled())
648 page_ext
= lookup_page_ext(page
);
649 if (unlikely(!page_ext
))
652 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
654 INIT_LIST_HEAD(&page
->lru
);
655 set_page_private(page
, order
);
656 /* Guard pages are not available for any usage */
657 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
660 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
661 unsigned int order
, int migratetype
)
663 struct page_ext
*page_ext
;
665 if (!debug_guardpage_enabled())
668 page_ext
= lookup_page_ext(page
);
669 if (unlikely(!page_ext
))
672 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
674 set_page_private(page
, 0);
675 if (!is_migrate_isolate(migratetype
))
676 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
679 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
680 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
681 unsigned int order
, int migratetype
) {}
682 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
683 unsigned int order
, int migratetype
) {}
686 static inline void set_page_order(struct page
*page
, unsigned int order
)
688 set_page_private(page
, order
);
689 __SetPageBuddy(page
);
692 static inline void rmv_page_order(struct page
*page
)
694 __ClearPageBuddy(page
);
695 set_page_private(page
, 0);
699 * This function checks whether a page is free && is the buddy
700 * we can do coalesce a page and its buddy if
701 * (a) the buddy is not in a hole &&
702 * (b) the buddy is in the buddy system &&
703 * (c) a page and its buddy have the same order &&
704 * (d) a page and its buddy are in the same zone.
706 * For recording whether a page is in the buddy system, we set ->_mapcount
707 * PAGE_BUDDY_MAPCOUNT_VALUE.
708 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
709 * serialized by zone->lock.
711 * For recording page's order, we use page_private(page).
713 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
716 if (!pfn_valid_within(page_to_pfn(buddy
)))
719 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
720 if (page_zone_id(page
) != page_zone_id(buddy
))
723 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
728 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
730 * zone check is done late to avoid uselessly
731 * calculating zone/node ids for pages that could
734 if (page_zone_id(page
) != page_zone_id(buddy
))
737 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
745 * Freeing function for a buddy system allocator.
747 * The concept of a buddy system is to maintain direct-mapped table
748 * (containing bit values) for memory blocks of various "orders".
749 * The bottom level table contains the map for the smallest allocatable
750 * units of memory (here, pages), and each level above it describes
751 * pairs of units from the levels below, hence, "buddies".
752 * At a high level, all that happens here is marking the table entry
753 * at the bottom level available, and propagating the changes upward
754 * as necessary, plus some accounting needed to play nicely with other
755 * parts of the VM system.
756 * At each level, we keep a list of pages, which are heads of continuous
757 * free pages of length of (1 << order) and marked with _mapcount
758 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
760 * So when we are allocating or freeing one, we can derive the state of the
761 * other. That is, if we allocate a small block, and both were
762 * free, the remainder of the region must be split into blocks.
763 * If a block is freed, and its buddy is also free, then this
764 * triggers coalescing into a block of larger size.
769 static inline void __free_one_page(struct page
*page
,
771 struct zone
*zone
, unsigned int order
,
774 unsigned long page_idx
;
775 unsigned long combined_idx
;
776 unsigned long uninitialized_var(buddy_idx
);
778 unsigned int max_order
;
780 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
782 VM_BUG_ON(!zone_is_initialized(zone
));
783 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
785 VM_BUG_ON(migratetype
== -1);
786 if (likely(!is_migrate_isolate(migratetype
)))
787 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
789 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
791 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
792 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
795 while (order
< max_order
- 1) {
796 buddy_idx
= __find_buddy_index(page_idx
, order
);
797 buddy
= page
+ (buddy_idx
- page_idx
);
798 if (!page_is_buddy(page
, buddy
, order
))
801 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
802 * merge with it and move up one order.
804 if (page_is_guard(buddy
)) {
805 clear_page_guard(zone
, buddy
, order
, migratetype
);
807 list_del(&buddy
->lru
);
808 zone
->free_area
[order
].nr_free
--;
809 rmv_page_order(buddy
);
811 combined_idx
= buddy_idx
& page_idx
;
812 page
= page
+ (combined_idx
- page_idx
);
813 page_idx
= combined_idx
;
816 if (max_order
< MAX_ORDER
) {
817 /* If we are here, it means order is >= pageblock_order.
818 * We want to prevent merge between freepages on isolate
819 * pageblock and normal pageblock. Without this, pageblock
820 * isolation could cause incorrect freepage or CMA accounting.
822 * We don't want to hit this code for the more frequent
825 if (unlikely(has_isolate_pageblock(zone
))) {
828 buddy_idx
= __find_buddy_index(page_idx
, order
);
829 buddy
= page
+ (buddy_idx
- page_idx
);
830 buddy_mt
= get_pageblock_migratetype(buddy
);
832 if (migratetype
!= buddy_mt
833 && (is_migrate_isolate(migratetype
) ||
834 is_migrate_isolate(buddy_mt
)))
838 goto continue_merging
;
842 set_page_order(page
, order
);
845 * If this is not the largest possible page, check if the buddy
846 * of the next-highest order is free. If it is, it's possible
847 * that pages are being freed that will coalesce soon. In case,
848 * that is happening, add the free page to the tail of the list
849 * so it's less likely to be used soon and more likely to be merged
850 * as a higher order page
852 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
853 struct page
*higher_page
, *higher_buddy
;
854 combined_idx
= buddy_idx
& page_idx
;
855 higher_page
= page
+ (combined_idx
- page_idx
);
856 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
857 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
858 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
859 list_add_tail(&page
->lru
,
860 &zone
->free_area
[order
].free_list
[migratetype
]);
865 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
867 zone
->free_area
[order
].nr_free
++;
871 * A bad page could be due to a number of fields. Instead of multiple branches,
872 * try and check multiple fields with one check. The caller must do a detailed
873 * check if necessary.
875 static inline bool page_expected_state(struct page
*page
,
876 unsigned long check_flags
)
878 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
881 if (unlikely((unsigned long)page
->mapping
|
882 page_ref_count(page
) |
884 (unsigned long)page
->mem_cgroup
|
886 (page
->flags
& check_flags
)))
892 static void free_pages_check_bad(struct page
*page
)
894 const char *bad_reason
;
895 unsigned long bad_flags
;
900 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
901 bad_reason
= "nonzero mapcount";
902 if (unlikely(page
->mapping
!= NULL
))
903 bad_reason
= "non-NULL mapping";
904 if (unlikely(page_ref_count(page
) != 0))
905 bad_reason
= "nonzero _refcount";
906 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
907 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
908 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
911 if (unlikely(page
->mem_cgroup
))
912 bad_reason
= "page still charged to cgroup";
914 bad_page(page
, bad_reason
, bad_flags
);
917 static inline int free_pages_check(struct page
*page
)
919 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
922 /* Something has gone sideways, find it */
923 free_pages_check_bad(page
);
927 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
932 * We rely page->lru.next never has bit 0 set, unless the page
933 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
935 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
937 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
941 switch (page
- head_page
) {
943 /* the first tail page: ->mapping is compound_mapcount() */
944 if (unlikely(compound_mapcount(page
))) {
945 bad_page(page
, "nonzero compound_mapcount", 0);
951 * the second tail page: ->mapping is
952 * page_deferred_list().next -- ignore value.
956 if (page
->mapping
!= TAIL_MAPPING
) {
957 bad_page(page
, "corrupted mapping in tail page", 0);
962 if (unlikely(!PageTail(page
))) {
963 bad_page(page
, "PageTail not set", 0);
966 if (unlikely(compound_head(page
) != head_page
)) {
967 bad_page(page
, "compound_head not consistent", 0);
972 page
->mapping
= NULL
;
973 clear_compound_head(page
);
977 static __always_inline
bool free_pages_prepare(struct page
*page
,
978 unsigned int order
, bool check_free
)
982 VM_BUG_ON_PAGE(PageTail(page
), page
);
984 trace_mm_page_free(page
, order
);
985 kmemcheck_free_shadow(page
, order
);
988 * Check tail pages before head page information is cleared to
989 * avoid checking PageCompound for order-0 pages.
991 if (unlikely(order
)) {
992 bool compound
= PageCompound(page
);
995 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
998 ClearPageDoubleMap(page
);
999 for (i
= 1; i
< (1 << order
); i
++) {
1001 bad
+= free_tail_pages_check(page
, page
+ i
);
1002 if (unlikely(free_pages_check(page
+ i
))) {
1006 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1009 if (PageMappingFlags(page
))
1010 page
->mapping
= NULL
;
1011 if (memcg_kmem_enabled() && PageKmemcg(page
))
1012 memcg_kmem_uncharge(page
, order
);
1014 bad
+= free_pages_check(page
);
1018 page_cpupid_reset_last(page
);
1019 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1020 reset_page_owner(page
, order
);
1022 if (!PageHighMem(page
)) {
1023 debug_check_no_locks_freed(page_address(page
),
1024 PAGE_SIZE
<< order
);
1025 debug_check_no_obj_freed(page_address(page
),
1026 PAGE_SIZE
<< order
);
1028 arch_free_page(page
, order
);
1029 kernel_poison_pages(page
, 1 << order
, 0);
1030 kernel_map_pages(page
, 1 << order
, 0);
1031 kasan_free_pages(page
, order
);
1036 #ifdef CONFIG_DEBUG_VM
1037 static inline bool free_pcp_prepare(struct page
*page
)
1039 return free_pages_prepare(page
, 0, true);
1042 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1047 static bool free_pcp_prepare(struct page
*page
)
1049 return free_pages_prepare(page
, 0, false);
1052 static bool bulkfree_pcp_prepare(struct page
*page
)
1054 return free_pages_check(page
);
1056 #endif /* CONFIG_DEBUG_VM */
1059 * Frees a number of pages from the PCP lists
1060 * Assumes all pages on list are in same zone, and of same order.
1061 * count is the number of pages to free.
1063 * If the zone was previously in an "all pages pinned" state then look to
1064 * see if this freeing clears that state.
1066 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1067 * pinned" detection logic.
1069 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1070 struct per_cpu_pages
*pcp
)
1072 int migratetype
= 0;
1074 unsigned long nr_scanned
;
1075 bool isolated_pageblocks
;
1077 spin_lock(&zone
->lock
);
1078 isolated_pageblocks
= has_isolate_pageblock(zone
);
1079 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1081 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1085 struct list_head
*list
;
1088 * Remove pages from lists in a round-robin fashion. A
1089 * batch_free count is maintained that is incremented when an
1090 * empty list is encountered. This is so more pages are freed
1091 * off fuller lists instead of spinning excessively around empty
1096 if (++migratetype
== MIGRATE_PCPTYPES
)
1098 list
= &pcp
->lists
[migratetype
];
1099 } while (list_empty(list
));
1101 /* This is the only non-empty list. Free them all. */
1102 if (batch_free
== MIGRATE_PCPTYPES
)
1106 int mt
; /* migratetype of the to-be-freed page */
1108 page
= list_last_entry(list
, struct page
, lru
);
1109 /* must delete as __free_one_page list manipulates */
1110 list_del(&page
->lru
);
1112 mt
= get_pcppage_migratetype(page
);
1113 /* MIGRATE_ISOLATE page should not go to pcplists */
1114 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1115 /* Pageblock could have been isolated meanwhile */
1116 if (unlikely(isolated_pageblocks
))
1117 mt
= get_pageblock_migratetype(page
);
1119 if (bulkfree_pcp_prepare(page
))
1122 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1123 trace_mm_page_pcpu_drain(page
, 0, mt
);
1124 } while (--count
&& --batch_free
&& !list_empty(list
));
1126 spin_unlock(&zone
->lock
);
1129 static void free_one_page(struct zone
*zone
,
1130 struct page
*page
, unsigned long pfn
,
1134 unsigned long nr_scanned
;
1135 spin_lock(&zone
->lock
);
1136 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1138 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1140 if (unlikely(has_isolate_pageblock(zone
) ||
1141 is_migrate_isolate(migratetype
))) {
1142 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1144 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1145 spin_unlock(&zone
->lock
);
1148 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1149 unsigned long zone
, int nid
)
1151 set_page_links(page
, zone
, nid
, pfn
);
1152 init_page_count(page
);
1153 page_mapcount_reset(page
);
1154 page_cpupid_reset_last(page
);
1156 INIT_LIST_HEAD(&page
->lru
);
1157 #ifdef WANT_PAGE_VIRTUAL
1158 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1159 if (!is_highmem_idx(zone
))
1160 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1164 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1167 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1170 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1171 static void init_reserved_page(unsigned long pfn
)
1176 if (!early_page_uninitialised(pfn
))
1179 nid
= early_pfn_to_nid(pfn
);
1180 pgdat
= NODE_DATA(nid
);
1182 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1183 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1185 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1188 __init_single_pfn(pfn
, zid
, nid
);
1191 static inline void init_reserved_page(unsigned long pfn
)
1194 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1197 * Initialised pages do not have PageReserved set. This function is
1198 * called for each range allocated by the bootmem allocator and
1199 * marks the pages PageReserved. The remaining valid pages are later
1200 * sent to the buddy page allocator.
1202 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1204 unsigned long start_pfn
= PFN_DOWN(start
);
1205 unsigned long end_pfn
= PFN_UP(end
);
1207 for (; start_pfn
< end_pfn
; start_pfn
++) {
1208 if (pfn_valid(start_pfn
)) {
1209 struct page
*page
= pfn_to_page(start_pfn
);
1211 init_reserved_page(start_pfn
);
1213 /* Avoid false-positive PageTail() */
1214 INIT_LIST_HEAD(&page
->lru
);
1216 SetPageReserved(page
);
1221 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1223 unsigned long flags
;
1225 unsigned long pfn
= page_to_pfn(page
);
1227 if (!free_pages_prepare(page
, order
, true))
1230 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1231 local_irq_save(flags
);
1232 __count_vm_events(PGFREE
, 1 << order
);
1233 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1234 local_irq_restore(flags
);
1237 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1239 unsigned int nr_pages
= 1 << order
;
1240 struct page
*p
= page
;
1244 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1246 __ClearPageReserved(p
);
1247 set_page_count(p
, 0);
1249 __ClearPageReserved(p
);
1250 set_page_count(p
, 0);
1252 page_zone(page
)->managed_pages
+= nr_pages
;
1253 set_page_refcounted(page
);
1254 __free_pages(page
, order
);
1257 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1258 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1260 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1262 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1264 static DEFINE_SPINLOCK(early_pfn_lock
);
1267 spin_lock(&early_pfn_lock
);
1268 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1270 nid
= first_online_node
;
1271 spin_unlock(&early_pfn_lock
);
1277 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1278 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1279 struct mminit_pfnnid_cache
*state
)
1283 nid
= __early_pfn_to_nid(pfn
, state
);
1284 if (nid
>= 0 && nid
!= node
)
1289 /* Only safe to use early in boot when initialisation is single-threaded */
1290 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1292 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1297 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1301 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1302 struct mminit_pfnnid_cache
*state
)
1309 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1312 if (early_page_uninitialised(pfn
))
1314 return __free_pages_boot_core(page
, order
);
1318 * Check that the whole (or subset of) a pageblock given by the interval of
1319 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1320 * with the migration of free compaction scanner. The scanners then need to
1321 * use only pfn_valid_within() check for arches that allow holes within
1324 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1326 * It's possible on some configurations to have a setup like node0 node1 node0
1327 * i.e. it's possible that all pages within a zones range of pages do not
1328 * belong to a single zone. We assume that a border between node0 and node1
1329 * can occur within a single pageblock, but not a node0 node1 node0
1330 * interleaving within a single pageblock. It is therefore sufficient to check
1331 * the first and last page of a pageblock and avoid checking each individual
1332 * page in a pageblock.
1334 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1335 unsigned long end_pfn
, struct zone
*zone
)
1337 struct page
*start_page
;
1338 struct page
*end_page
;
1340 /* end_pfn is one past the range we are checking */
1343 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1346 start_page
= pfn_to_page(start_pfn
);
1348 if (page_zone(start_page
) != zone
)
1351 end_page
= pfn_to_page(end_pfn
);
1353 /* This gives a shorter code than deriving page_zone(end_page) */
1354 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1360 void set_zone_contiguous(struct zone
*zone
)
1362 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1363 unsigned long block_end_pfn
;
1365 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1366 for (; block_start_pfn
< zone_end_pfn(zone
);
1367 block_start_pfn
= block_end_pfn
,
1368 block_end_pfn
+= pageblock_nr_pages
) {
1370 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1372 if (!__pageblock_pfn_to_page(block_start_pfn
,
1373 block_end_pfn
, zone
))
1377 /* We confirm that there is no hole */
1378 zone
->contiguous
= true;
1381 void clear_zone_contiguous(struct zone
*zone
)
1383 zone
->contiguous
= false;
1386 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1387 static void __init
deferred_free_range(struct page
*page
,
1388 unsigned long pfn
, int nr_pages
)
1395 /* Free a large naturally-aligned chunk if possible */
1396 if (nr_pages
== MAX_ORDER_NR_PAGES
&&
1397 (pfn
& (MAX_ORDER_NR_PAGES
-1)) == 0) {
1398 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1399 __free_pages_boot_core(page
, MAX_ORDER
-1);
1403 for (i
= 0; i
< nr_pages
; i
++, page
++)
1404 __free_pages_boot_core(page
, 0);
1407 /* Completion tracking for deferred_init_memmap() threads */
1408 static atomic_t pgdat_init_n_undone __initdata
;
1409 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1411 static inline void __init
pgdat_init_report_one_done(void)
1413 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1414 complete(&pgdat_init_all_done_comp
);
1417 /* Initialise remaining memory on a node */
1418 static int __init
deferred_init_memmap(void *data
)
1420 pg_data_t
*pgdat
= data
;
1421 int nid
= pgdat
->node_id
;
1422 struct mminit_pfnnid_cache nid_init_state
= { };
1423 unsigned long start
= jiffies
;
1424 unsigned long nr_pages
= 0;
1425 unsigned long walk_start
, walk_end
;
1428 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1429 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1431 if (first_init_pfn
== ULONG_MAX
) {
1432 pgdat_init_report_one_done();
1436 /* Bind memory initialisation thread to a local node if possible */
1437 if (!cpumask_empty(cpumask
))
1438 set_cpus_allowed_ptr(current
, cpumask
);
1440 /* Sanity check boundaries */
1441 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1442 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1443 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1445 /* Only the highest zone is deferred so find it */
1446 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1447 zone
= pgdat
->node_zones
+ zid
;
1448 if (first_init_pfn
< zone_end_pfn(zone
))
1452 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1453 unsigned long pfn
, end_pfn
;
1454 struct page
*page
= NULL
;
1455 struct page
*free_base_page
= NULL
;
1456 unsigned long free_base_pfn
= 0;
1459 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1460 pfn
= first_init_pfn
;
1461 if (pfn
< walk_start
)
1463 if (pfn
< zone
->zone_start_pfn
)
1464 pfn
= zone
->zone_start_pfn
;
1466 for (; pfn
< end_pfn
; pfn
++) {
1467 if (!pfn_valid_within(pfn
))
1471 * Ensure pfn_valid is checked every
1472 * MAX_ORDER_NR_PAGES for memory holes
1474 if ((pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
1475 if (!pfn_valid(pfn
)) {
1481 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1486 /* Minimise pfn page lookups and scheduler checks */
1487 if (page
&& (pfn
& (MAX_ORDER_NR_PAGES
- 1)) != 0) {
1490 nr_pages
+= nr_to_free
;
1491 deferred_free_range(free_base_page
,
1492 free_base_pfn
, nr_to_free
);
1493 free_base_page
= NULL
;
1494 free_base_pfn
= nr_to_free
= 0;
1496 page
= pfn_to_page(pfn
);
1501 VM_BUG_ON(page_zone(page
) != zone
);
1505 __init_single_page(page
, pfn
, zid
, nid
);
1506 if (!free_base_page
) {
1507 free_base_page
= page
;
1508 free_base_pfn
= pfn
;
1513 /* Where possible, batch up pages for a single free */
1516 /* Free the current block of pages to allocator */
1517 nr_pages
+= nr_to_free
;
1518 deferred_free_range(free_base_page
, free_base_pfn
,
1520 free_base_page
= NULL
;
1521 free_base_pfn
= nr_to_free
= 0;
1524 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1527 /* Sanity check that the next zone really is unpopulated */
1528 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1530 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1531 jiffies_to_msecs(jiffies
- start
));
1533 pgdat_init_report_one_done();
1536 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1538 void __init
page_alloc_init_late(void)
1542 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1545 /* There will be num_node_state(N_MEMORY) threads */
1546 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1547 for_each_node_state(nid
, N_MEMORY
) {
1548 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1551 /* Block until all are initialised */
1552 wait_for_completion(&pgdat_init_all_done_comp
);
1554 /* Reinit limits that are based on free pages after the kernel is up */
1555 files_maxfiles_init();
1558 for_each_populated_zone(zone
)
1559 set_zone_contiguous(zone
);
1563 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1564 void __init
init_cma_reserved_pageblock(struct page
*page
)
1566 unsigned i
= pageblock_nr_pages
;
1567 struct page
*p
= page
;
1570 __ClearPageReserved(p
);
1571 set_page_count(p
, 0);
1574 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1576 if (pageblock_order
>= MAX_ORDER
) {
1577 i
= pageblock_nr_pages
;
1580 set_page_refcounted(p
);
1581 __free_pages(p
, MAX_ORDER
- 1);
1582 p
+= MAX_ORDER_NR_PAGES
;
1583 } while (i
-= MAX_ORDER_NR_PAGES
);
1585 set_page_refcounted(page
);
1586 __free_pages(page
, pageblock_order
);
1589 adjust_managed_page_count(page
, pageblock_nr_pages
);
1594 * The order of subdivision here is critical for the IO subsystem.
1595 * Please do not alter this order without good reasons and regression
1596 * testing. Specifically, as large blocks of memory are subdivided,
1597 * the order in which smaller blocks are delivered depends on the order
1598 * they're subdivided in this function. This is the primary factor
1599 * influencing the order in which pages are delivered to the IO
1600 * subsystem according to empirical testing, and this is also justified
1601 * by considering the behavior of a buddy system containing a single
1602 * large block of memory acted on by a series of small allocations.
1603 * This behavior is a critical factor in sglist merging's success.
1607 static inline void expand(struct zone
*zone
, struct page
*page
,
1608 int low
, int high
, struct free_area
*area
,
1611 unsigned long size
= 1 << high
;
1613 while (high
> low
) {
1617 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1619 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
1620 debug_guardpage_enabled() &&
1621 high
< debug_guardpage_minorder()) {
1623 * Mark as guard pages (or page), that will allow to
1624 * merge back to allocator when buddy will be freed.
1625 * Corresponding page table entries will not be touched,
1626 * pages will stay not present in virtual address space
1628 set_page_guard(zone
, &page
[size
], high
, migratetype
);
1631 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1633 set_page_order(&page
[size
], high
);
1637 static void check_new_page_bad(struct page
*page
)
1639 const char *bad_reason
= NULL
;
1640 unsigned long bad_flags
= 0;
1642 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1643 bad_reason
= "nonzero mapcount";
1644 if (unlikely(page
->mapping
!= NULL
))
1645 bad_reason
= "non-NULL mapping";
1646 if (unlikely(page_ref_count(page
) != 0))
1647 bad_reason
= "nonzero _count";
1648 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1649 bad_reason
= "HWPoisoned (hardware-corrupted)";
1650 bad_flags
= __PG_HWPOISON
;
1651 /* Don't complain about hwpoisoned pages */
1652 page_mapcount_reset(page
); /* remove PageBuddy */
1655 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1656 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1657 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1660 if (unlikely(page
->mem_cgroup
))
1661 bad_reason
= "page still charged to cgroup";
1663 bad_page(page
, bad_reason
, bad_flags
);
1667 * This page is about to be returned from the page allocator
1669 static inline int check_new_page(struct page
*page
)
1671 if (likely(page_expected_state(page
,
1672 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1675 check_new_page_bad(page
);
1679 static inline bool free_pages_prezeroed(bool poisoned
)
1681 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1682 page_poisoning_enabled() && poisoned
;
1685 #ifdef CONFIG_DEBUG_VM
1686 static bool check_pcp_refill(struct page
*page
)
1691 static bool check_new_pcp(struct page
*page
)
1693 return check_new_page(page
);
1696 static bool check_pcp_refill(struct page
*page
)
1698 return check_new_page(page
);
1700 static bool check_new_pcp(struct page
*page
)
1704 #endif /* CONFIG_DEBUG_VM */
1706 static bool check_new_pages(struct page
*page
, unsigned int order
)
1709 for (i
= 0; i
< (1 << order
); i
++) {
1710 struct page
*p
= page
+ i
;
1712 if (unlikely(check_new_page(p
)))
1719 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1722 set_page_private(page
, 0);
1723 set_page_refcounted(page
);
1725 arch_alloc_page(page
, order
);
1726 kernel_map_pages(page
, 1 << order
, 1);
1727 kernel_poison_pages(page
, 1 << order
, 1);
1728 kasan_alloc_pages(page
, order
);
1729 set_page_owner(page
, order
, gfp_flags
);
1732 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1733 unsigned int alloc_flags
)
1736 bool poisoned
= true;
1738 for (i
= 0; i
< (1 << order
); i
++) {
1739 struct page
*p
= page
+ i
;
1741 poisoned
&= page_is_poisoned(p
);
1744 post_alloc_hook(page
, order
, gfp_flags
);
1746 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1747 for (i
= 0; i
< (1 << order
); i
++)
1748 clear_highpage(page
+ i
);
1750 if (order
&& (gfp_flags
& __GFP_COMP
))
1751 prep_compound_page(page
, order
);
1754 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1755 * allocate the page. The expectation is that the caller is taking
1756 * steps that will free more memory. The caller should avoid the page
1757 * being used for !PFMEMALLOC purposes.
1759 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1760 set_page_pfmemalloc(page
);
1762 clear_page_pfmemalloc(page
);
1766 * Go through the free lists for the given migratetype and remove
1767 * the smallest available page from the freelists
1770 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1773 unsigned int current_order
;
1774 struct free_area
*area
;
1777 /* Find a page of the appropriate size in the preferred list */
1778 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1779 area
= &(zone
->free_area
[current_order
]);
1780 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1784 list_del(&page
->lru
);
1785 rmv_page_order(page
);
1787 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1788 set_pcppage_migratetype(page
, migratetype
);
1797 * This array describes the order lists are fallen back to when
1798 * the free lists for the desirable migrate type are depleted
1800 static int fallbacks
[MIGRATE_TYPES
][4] = {
1801 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1802 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1803 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1805 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1807 #ifdef CONFIG_MEMORY_ISOLATION
1808 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1813 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1816 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1819 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1820 unsigned int order
) { return NULL
; }
1824 * Move the free pages in a range to the free lists of the requested type.
1825 * Note that start_page and end_pages are not aligned on a pageblock
1826 * boundary. If alignment is required, use move_freepages_block()
1828 int move_freepages(struct zone
*zone
,
1829 struct page
*start_page
, struct page
*end_page
,
1834 int pages_moved
= 0;
1836 #ifndef CONFIG_HOLES_IN_ZONE
1838 * page_zone is not safe to call in this context when
1839 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1840 * anyway as we check zone boundaries in move_freepages_block().
1841 * Remove at a later date when no bug reports exist related to
1842 * grouping pages by mobility
1844 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1847 for (page
= start_page
; page
<= end_page
;) {
1848 /* Make sure we are not inadvertently changing nodes */
1849 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1851 if (!pfn_valid_within(page_to_pfn(page
))) {
1856 if (!PageBuddy(page
)) {
1861 order
= page_order(page
);
1862 list_move(&page
->lru
,
1863 &zone
->free_area
[order
].free_list
[migratetype
]);
1865 pages_moved
+= 1 << order
;
1871 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1874 unsigned long start_pfn
, end_pfn
;
1875 struct page
*start_page
, *end_page
;
1877 start_pfn
= page_to_pfn(page
);
1878 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1879 start_page
= pfn_to_page(start_pfn
);
1880 end_page
= start_page
+ pageblock_nr_pages
- 1;
1881 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1883 /* Do not cross zone boundaries */
1884 if (!zone_spans_pfn(zone
, start_pfn
))
1886 if (!zone_spans_pfn(zone
, end_pfn
))
1889 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1892 static void change_pageblock_range(struct page
*pageblock_page
,
1893 int start_order
, int migratetype
)
1895 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1897 while (nr_pageblocks
--) {
1898 set_pageblock_migratetype(pageblock_page
, migratetype
);
1899 pageblock_page
+= pageblock_nr_pages
;
1904 * When we are falling back to another migratetype during allocation, try to
1905 * steal extra free pages from the same pageblocks to satisfy further
1906 * allocations, instead of polluting multiple pageblocks.
1908 * If we are stealing a relatively large buddy page, it is likely there will
1909 * be more free pages in the pageblock, so try to steal them all. For
1910 * reclaimable and unmovable allocations, we steal regardless of page size,
1911 * as fragmentation caused by those allocations polluting movable pageblocks
1912 * is worse than movable allocations stealing from unmovable and reclaimable
1915 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1918 * Leaving this order check is intended, although there is
1919 * relaxed order check in next check. The reason is that
1920 * we can actually steal whole pageblock if this condition met,
1921 * but, below check doesn't guarantee it and that is just heuristic
1922 * so could be changed anytime.
1924 if (order
>= pageblock_order
)
1927 if (order
>= pageblock_order
/ 2 ||
1928 start_mt
== MIGRATE_RECLAIMABLE
||
1929 start_mt
== MIGRATE_UNMOVABLE
||
1930 page_group_by_mobility_disabled
)
1937 * This function implements actual steal behaviour. If order is large enough,
1938 * we can steal whole pageblock. If not, we first move freepages in this
1939 * pageblock and check whether half of pages are moved or not. If half of
1940 * pages are moved, we can change migratetype of pageblock and permanently
1941 * use it's pages as requested migratetype in the future.
1943 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1946 unsigned int current_order
= page_order(page
);
1949 /* Take ownership for orders >= pageblock_order */
1950 if (current_order
>= pageblock_order
) {
1951 change_pageblock_range(page
, current_order
, start_type
);
1955 pages
= move_freepages_block(zone
, page
, start_type
);
1957 /* Claim the whole block if over half of it is free */
1958 if (pages
>= (1 << (pageblock_order
-1)) ||
1959 page_group_by_mobility_disabled
)
1960 set_pageblock_migratetype(page
, start_type
);
1964 * Check whether there is a suitable fallback freepage with requested order.
1965 * If only_stealable is true, this function returns fallback_mt only if
1966 * we can steal other freepages all together. This would help to reduce
1967 * fragmentation due to mixed migratetype pages in one pageblock.
1969 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1970 int migratetype
, bool only_stealable
, bool *can_steal
)
1975 if (area
->nr_free
== 0)
1980 fallback_mt
= fallbacks
[migratetype
][i
];
1981 if (fallback_mt
== MIGRATE_TYPES
)
1984 if (list_empty(&area
->free_list
[fallback_mt
]))
1987 if (can_steal_fallback(order
, migratetype
))
1990 if (!only_stealable
)
2001 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2002 * there are no empty page blocks that contain a page with a suitable order
2004 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2005 unsigned int alloc_order
)
2008 unsigned long max_managed
, flags
;
2011 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2012 * Check is race-prone but harmless.
2014 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2015 if (zone
->nr_reserved_highatomic
>= max_managed
)
2018 spin_lock_irqsave(&zone
->lock
, flags
);
2020 /* Recheck the nr_reserved_highatomic limit under the lock */
2021 if (zone
->nr_reserved_highatomic
>= max_managed
)
2025 mt
= get_pageblock_migratetype(page
);
2026 if (mt
!= MIGRATE_HIGHATOMIC
&&
2027 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
2028 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2029 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2030 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
2034 spin_unlock_irqrestore(&zone
->lock
, flags
);
2038 * Used when an allocation is about to fail under memory pressure. This
2039 * potentially hurts the reliability of high-order allocations when under
2040 * intense memory pressure but failed atomic allocations should be easier
2041 * to recover from than an OOM.
2043 static void unreserve_highatomic_pageblock(const struct alloc_context
*ac
)
2045 struct zonelist
*zonelist
= ac
->zonelist
;
2046 unsigned long flags
;
2052 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2054 /* Preserve at least one pageblock */
2055 if (zone
->nr_reserved_highatomic
<= pageblock_nr_pages
)
2058 spin_lock_irqsave(&zone
->lock
, flags
);
2059 for (order
= 0; order
< MAX_ORDER
; order
++) {
2060 struct free_area
*area
= &(zone
->free_area
[order
]);
2062 page
= list_first_entry_or_null(
2063 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2069 * It should never happen but changes to locking could
2070 * inadvertently allow a per-cpu drain to add pages
2071 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2072 * and watch for underflows.
2074 zone
->nr_reserved_highatomic
-= min(pageblock_nr_pages
,
2075 zone
->nr_reserved_highatomic
);
2078 * Convert to ac->migratetype and avoid the normal
2079 * pageblock stealing heuristics. Minimally, the caller
2080 * is doing the work and needs the pages. More
2081 * importantly, if the block was always converted to
2082 * MIGRATE_UNMOVABLE or another type then the number
2083 * of pageblocks that cannot be completely freed
2086 set_pageblock_migratetype(page
, ac
->migratetype
);
2087 move_freepages_block(zone
, page
, ac
->migratetype
);
2088 spin_unlock_irqrestore(&zone
->lock
, flags
);
2091 spin_unlock_irqrestore(&zone
->lock
, flags
);
2095 /* Remove an element from the buddy allocator from the fallback list */
2096 static inline struct page
*
2097 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2099 struct free_area
*area
;
2100 unsigned int current_order
;
2105 /* Find the largest possible block of pages in the other list */
2106 for (current_order
= MAX_ORDER
-1;
2107 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2109 area
= &(zone
->free_area
[current_order
]);
2110 fallback_mt
= find_suitable_fallback(area
, current_order
,
2111 start_migratetype
, false, &can_steal
);
2112 if (fallback_mt
== -1)
2115 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2118 steal_suitable_fallback(zone
, page
, start_migratetype
);
2120 /* Remove the page from the freelists */
2122 list_del(&page
->lru
);
2123 rmv_page_order(page
);
2125 expand(zone
, page
, order
, current_order
, area
,
2128 * The pcppage_migratetype may differ from pageblock's
2129 * migratetype depending on the decisions in
2130 * find_suitable_fallback(). This is OK as long as it does not
2131 * differ for MIGRATE_CMA pageblocks. Those can be used as
2132 * fallback only via special __rmqueue_cma_fallback() function
2134 set_pcppage_migratetype(page
, start_migratetype
);
2136 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2137 start_migratetype
, fallback_mt
);
2146 * Do the hard work of removing an element from the buddy allocator.
2147 * Call me with the zone->lock already held.
2149 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2154 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2155 if (unlikely(!page
)) {
2156 if (migratetype
== MIGRATE_MOVABLE
)
2157 page
= __rmqueue_cma_fallback(zone
, order
);
2160 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2163 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2168 * Obtain a specified number of elements from the buddy allocator, all under
2169 * a single hold of the lock, for efficiency. Add them to the supplied list.
2170 * Returns the number of new pages which were placed at *list.
2172 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2173 unsigned long count
, struct list_head
*list
,
2174 int migratetype
, bool cold
)
2178 spin_lock(&zone
->lock
);
2179 for (i
= 0; i
< count
; ++i
) {
2180 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2181 if (unlikely(page
== NULL
))
2184 if (unlikely(check_pcp_refill(page
)))
2188 * Split buddy pages returned by expand() are received here
2189 * in physical page order. The page is added to the callers and
2190 * list and the list head then moves forward. From the callers
2191 * perspective, the linked list is ordered by page number in
2192 * some conditions. This is useful for IO devices that can
2193 * merge IO requests if the physical pages are ordered
2197 list_add(&page
->lru
, list
);
2199 list_add_tail(&page
->lru
, list
);
2201 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2202 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2205 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2206 spin_unlock(&zone
->lock
);
2212 * Called from the vmstat counter updater to drain pagesets of this
2213 * currently executing processor on remote nodes after they have
2216 * Note that this function must be called with the thread pinned to
2217 * a single processor.
2219 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2221 unsigned long flags
;
2222 int to_drain
, batch
;
2224 local_irq_save(flags
);
2225 batch
= READ_ONCE(pcp
->batch
);
2226 to_drain
= min(pcp
->count
, batch
);
2228 free_pcppages_bulk(zone
, to_drain
, pcp
);
2229 pcp
->count
-= to_drain
;
2231 local_irq_restore(flags
);
2236 * Drain pcplists of the indicated processor and zone.
2238 * The processor must either be the current processor and the
2239 * thread pinned to the current processor or a processor that
2242 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2244 unsigned long flags
;
2245 struct per_cpu_pageset
*pset
;
2246 struct per_cpu_pages
*pcp
;
2248 local_irq_save(flags
);
2249 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2253 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2256 local_irq_restore(flags
);
2260 * Drain pcplists of all zones on the indicated processor.
2262 * The processor must either be the current processor and the
2263 * thread pinned to the current processor or a processor that
2266 static void drain_pages(unsigned int cpu
)
2270 for_each_populated_zone(zone
) {
2271 drain_pages_zone(cpu
, zone
);
2276 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2278 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2279 * the single zone's pages.
2281 void drain_local_pages(struct zone
*zone
)
2283 int cpu
= smp_processor_id();
2286 drain_pages_zone(cpu
, zone
);
2292 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2294 * When zone parameter is non-NULL, spill just the single zone's pages.
2296 * Note that this code is protected against sending an IPI to an offline
2297 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2298 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2299 * nothing keeps CPUs from showing up after we populated the cpumask and
2300 * before the call to on_each_cpu_mask().
2302 void drain_all_pages(struct zone
*zone
)
2307 * Allocate in the BSS so we wont require allocation in
2308 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2310 static cpumask_t cpus_with_pcps
;
2313 * We don't care about racing with CPU hotplug event
2314 * as offline notification will cause the notified
2315 * cpu to drain that CPU pcps and on_each_cpu_mask
2316 * disables preemption as part of its processing
2318 for_each_online_cpu(cpu
) {
2319 struct per_cpu_pageset
*pcp
;
2321 bool has_pcps
= false;
2324 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2328 for_each_populated_zone(z
) {
2329 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2330 if (pcp
->pcp
.count
) {
2338 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2340 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2342 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
2346 #ifdef CONFIG_HIBERNATION
2348 void mark_free_pages(struct zone
*zone
)
2350 unsigned long pfn
, max_zone_pfn
;
2351 unsigned long flags
;
2352 unsigned int order
, t
;
2355 if (zone_is_empty(zone
))
2358 spin_lock_irqsave(&zone
->lock
, flags
);
2360 max_zone_pfn
= zone_end_pfn(zone
);
2361 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2362 if (pfn_valid(pfn
)) {
2363 page
= pfn_to_page(pfn
);
2365 if (page_zone(page
) != zone
)
2368 if (!swsusp_page_is_forbidden(page
))
2369 swsusp_unset_page_free(page
);
2372 for_each_migratetype_order(order
, t
) {
2373 list_for_each_entry(page
,
2374 &zone
->free_area
[order
].free_list
[t
], lru
) {
2377 pfn
= page_to_pfn(page
);
2378 for (i
= 0; i
< (1UL << order
); i
++)
2379 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2382 spin_unlock_irqrestore(&zone
->lock
, flags
);
2384 #endif /* CONFIG_PM */
2387 * Free a 0-order page
2388 * cold == true ? free a cold page : free a hot page
2390 void free_hot_cold_page(struct page
*page
, bool cold
)
2392 struct zone
*zone
= page_zone(page
);
2393 struct per_cpu_pages
*pcp
;
2394 unsigned long flags
;
2395 unsigned long pfn
= page_to_pfn(page
);
2398 if (!free_pcp_prepare(page
))
2401 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2402 set_pcppage_migratetype(page
, migratetype
);
2403 local_irq_save(flags
);
2404 __count_vm_event(PGFREE
);
2407 * We only track unmovable, reclaimable and movable on pcp lists.
2408 * Free ISOLATE pages back to the allocator because they are being
2409 * offlined but treat RESERVE as movable pages so we can get those
2410 * areas back if necessary. Otherwise, we may have to free
2411 * excessively into the page allocator
2413 if (migratetype
>= MIGRATE_PCPTYPES
) {
2414 if (unlikely(is_migrate_isolate(migratetype
))) {
2415 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2418 migratetype
= MIGRATE_MOVABLE
;
2421 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2423 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2425 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2427 if (pcp
->count
>= pcp
->high
) {
2428 unsigned long batch
= READ_ONCE(pcp
->batch
);
2429 free_pcppages_bulk(zone
, batch
, pcp
);
2430 pcp
->count
-= batch
;
2434 local_irq_restore(flags
);
2438 * Free a list of 0-order pages
2440 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2442 struct page
*page
, *next
;
2444 list_for_each_entry_safe(page
, next
, list
, lru
) {
2445 trace_mm_page_free_batched(page
, cold
);
2446 free_hot_cold_page(page
, cold
);
2451 * split_page takes a non-compound higher-order page, and splits it into
2452 * n (1<<order) sub-pages: page[0..n]
2453 * Each sub-page must be freed individually.
2455 * Note: this is probably too low level an operation for use in drivers.
2456 * Please consult with lkml before using this in your driver.
2458 void split_page(struct page
*page
, unsigned int order
)
2462 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2463 VM_BUG_ON_PAGE(!page_count(page
), page
);
2465 #ifdef CONFIG_KMEMCHECK
2467 * Split shadow pages too, because free(page[0]) would
2468 * otherwise free the whole shadow.
2470 if (kmemcheck_page_is_tracked(page
))
2471 split_page(virt_to_page(page
[0].shadow
), order
);
2474 for (i
= 1; i
< (1 << order
); i
++)
2475 set_page_refcounted(page
+ i
);
2476 split_page_owner(page
, order
);
2478 EXPORT_SYMBOL_GPL(split_page
);
2480 int __isolate_free_page(struct page
*page
, unsigned int order
)
2482 unsigned long watermark
;
2486 BUG_ON(!PageBuddy(page
));
2488 zone
= page_zone(page
);
2489 mt
= get_pageblock_migratetype(page
);
2491 if (!is_migrate_isolate(mt
)) {
2492 /* Obey watermarks as if the page was being allocated */
2493 watermark
= low_wmark_pages(zone
) + (1 << order
);
2494 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
2497 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2500 /* Remove page from free list */
2501 list_del(&page
->lru
);
2502 zone
->free_area
[order
].nr_free
--;
2503 rmv_page_order(page
);
2506 * Set the pageblock if the isolated page is at least half of a
2509 if (order
>= pageblock_order
- 1) {
2510 struct page
*endpage
= page
+ (1 << order
) - 1;
2511 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2512 int mt
= get_pageblock_migratetype(page
);
2513 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2514 set_pageblock_migratetype(page
,
2520 return 1UL << order
;
2524 * Update NUMA hit/miss statistics
2526 * Must be called with interrupts disabled.
2528 * When __GFP_OTHER_NODE is set assume the node of the preferred
2529 * zone is the local node. This is useful for daemons who allocate
2530 * memory on behalf of other processes.
2532 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2536 int local_nid
= numa_node_id();
2537 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2539 if (unlikely(flags
& __GFP_OTHER_NODE
)) {
2540 local_stat
= NUMA_OTHER
;
2541 local_nid
= preferred_zone
->node
;
2544 if (z
->node
== local_nid
) {
2545 __inc_zone_state(z
, NUMA_HIT
);
2546 __inc_zone_state(z
, local_stat
);
2548 __inc_zone_state(z
, NUMA_MISS
);
2549 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2555 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2558 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2559 struct zone
*zone
, unsigned int order
,
2560 gfp_t gfp_flags
, unsigned int alloc_flags
,
2563 unsigned long flags
;
2565 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2567 if (likely(order
== 0)) {
2568 struct per_cpu_pages
*pcp
;
2569 struct list_head
*list
;
2571 local_irq_save(flags
);
2573 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2574 list
= &pcp
->lists
[migratetype
];
2575 if (list_empty(list
)) {
2576 pcp
->count
+= rmqueue_bulk(zone
, 0,
2579 if (unlikely(list_empty(list
)))
2584 page
= list_last_entry(list
, struct page
, lru
);
2586 page
= list_first_entry(list
, struct page
, lru
);
2588 list_del(&page
->lru
);
2591 } while (check_new_pcp(page
));
2594 * We most definitely don't want callers attempting to
2595 * allocate greater than order-1 page units with __GFP_NOFAIL.
2597 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2598 spin_lock_irqsave(&zone
->lock
, flags
);
2602 if (alloc_flags
& ALLOC_HARDER
) {
2603 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2605 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2608 page
= __rmqueue(zone
, order
, migratetype
);
2609 } while (page
&& check_new_pages(page
, order
));
2610 spin_unlock(&zone
->lock
);
2613 __mod_zone_freepage_state(zone
, -(1 << order
),
2614 get_pcppage_migratetype(page
));
2617 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2618 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2619 local_irq_restore(flags
);
2621 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2625 local_irq_restore(flags
);
2629 #ifdef CONFIG_FAIL_PAGE_ALLOC
2632 struct fault_attr attr
;
2634 bool ignore_gfp_highmem
;
2635 bool ignore_gfp_reclaim
;
2637 } fail_page_alloc
= {
2638 .attr
= FAULT_ATTR_INITIALIZER
,
2639 .ignore_gfp_reclaim
= true,
2640 .ignore_gfp_highmem
= true,
2644 static int __init
setup_fail_page_alloc(char *str
)
2646 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2648 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2650 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2652 if (order
< fail_page_alloc
.min_order
)
2654 if (gfp_mask
& __GFP_NOFAIL
)
2656 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2658 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2659 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2662 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2665 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2667 static int __init
fail_page_alloc_debugfs(void)
2669 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2672 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2673 &fail_page_alloc
.attr
);
2675 return PTR_ERR(dir
);
2677 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2678 &fail_page_alloc
.ignore_gfp_reclaim
))
2680 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2681 &fail_page_alloc
.ignore_gfp_highmem
))
2683 if (!debugfs_create_u32("min-order", mode
, dir
,
2684 &fail_page_alloc
.min_order
))
2689 debugfs_remove_recursive(dir
);
2694 late_initcall(fail_page_alloc_debugfs
);
2696 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2698 #else /* CONFIG_FAIL_PAGE_ALLOC */
2700 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2705 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2708 * Return true if free base pages are above 'mark'. For high-order checks it
2709 * will return true of the order-0 watermark is reached and there is at least
2710 * one free page of a suitable size. Checking now avoids taking the zone lock
2711 * to check in the allocation paths if no pages are free.
2713 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2714 int classzone_idx
, unsigned int alloc_flags
,
2719 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2721 /* free_pages may go negative - that's OK */
2722 free_pages
-= (1 << order
) - 1;
2724 if (alloc_flags
& ALLOC_HIGH
)
2728 * If the caller does not have rights to ALLOC_HARDER then subtract
2729 * the high-atomic reserves. This will over-estimate the size of the
2730 * atomic reserve but it avoids a search.
2732 if (likely(!alloc_harder
))
2733 free_pages
-= z
->nr_reserved_highatomic
;
2738 /* If allocation can't use CMA areas don't use free CMA pages */
2739 if (!(alloc_flags
& ALLOC_CMA
))
2740 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2744 * Check watermarks for an order-0 allocation request. If these
2745 * are not met, then a high-order request also cannot go ahead
2746 * even if a suitable page happened to be free.
2748 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2751 /* If this is an order-0 request then the watermark is fine */
2755 /* For a high-order request, check at least one suitable page is free */
2756 for (o
= order
; o
< MAX_ORDER
; o
++) {
2757 struct free_area
*area
= &z
->free_area
[o
];
2766 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2767 if (!list_empty(&area
->free_list
[mt
]))
2772 if ((alloc_flags
& ALLOC_CMA
) &&
2773 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2781 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2782 int classzone_idx
, unsigned int alloc_flags
)
2784 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2785 zone_page_state(z
, NR_FREE_PAGES
));
2788 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2789 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2791 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2795 /* If allocation can't use CMA areas don't use free CMA pages */
2796 if (!(alloc_flags
& ALLOC_CMA
))
2797 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2801 * Fast check for order-0 only. If this fails then the reserves
2802 * need to be calculated. There is a corner case where the check
2803 * passes but only the high-order atomic reserve are free. If
2804 * the caller is !atomic then it'll uselessly search the free
2805 * list. That corner case is then slower but it is harmless.
2807 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2810 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2814 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2815 unsigned long mark
, int classzone_idx
)
2817 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2819 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2820 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2822 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2827 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2829 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2832 #else /* CONFIG_NUMA */
2833 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2837 #endif /* CONFIG_NUMA */
2840 * get_page_from_freelist goes through the zonelist trying to allocate
2843 static struct page
*
2844 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2845 const struct alloc_context
*ac
)
2847 struct zoneref
*z
= ac
->preferred_zoneref
;
2849 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
2852 * Scan zonelist, looking for a zone with enough free.
2853 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2855 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2860 if (cpusets_enabled() &&
2861 (alloc_flags
& ALLOC_CPUSET
) &&
2862 !__cpuset_zone_allowed(zone
, gfp_mask
))
2865 * When allocating a page cache page for writing, we
2866 * want to get it from a node that is within its dirty
2867 * limit, such that no single node holds more than its
2868 * proportional share of globally allowed dirty pages.
2869 * The dirty limits take into account the node's
2870 * lowmem reserves and high watermark so that kswapd
2871 * should be able to balance it without having to
2872 * write pages from its LRU list.
2874 * XXX: For now, allow allocations to potentially
2875 * exceed the per-node dirty limit in the slowpath
2876 * (spread_dirty_pages unset) before going into reclaim,
2877 * which is important when on a NUMA setup the allowed
2878 * nodes are together not big enough to reach the
2879 * global limit. The proper fix for these situations
2880 * will require awareness of nodes in the
2881 * dirty-throttling and the flusher threads.
2883 if (ac
->spread_dirty_pages
) {
2884 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
2887 if (!node_dirty_ok(zone
->zone_pgdat
)) {
2888 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
2893 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2894 if (!zone_watermark_fast(zone
, order
, mark
,
2895 ac_classzone_idx(ac
), alloc_flags
)) {
2898 /* Checked here to keep the fast path fast */
2899 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2900 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2903 if (node_reclaim_mode
== 0 ||
2904 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
2907 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
2909 case NODE_RECLAIM_NOSCAN
:
2912 case NODE_RECLAIM_FULL
:
2913 /* scanned but unreclaimable */
2916 /* did we reclaim enough */
2917 if (zone_watermark_ok(zone
, order
, mark
,
2918 ac_classzone_idx(ac
), alloc_flags
))
2926 page
= buffered_rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
2927 gfp_mask
, alloc_flags
, ac
->migratetype
);
2929 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
2932 * If this is a high-order atomic allocation then check
2933 * if the pageblock should be reserved for the future
2935 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
2936 reserve_highatomic_pageblock(page
, zone
, order
);
2946 * Large machines with many possible nodes should not always dump per-node
2947 * meminfo in irq context.
2949 static inline bool should_suppress_show_mem(void)
2954 ret
= in_interrupt();
2959 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2960 DEFAULT_RATELIMIT_INTERVAL
,
2961 DEFAULT_RATELIMIT_BURST
);
2963 void warn_alloc_failed(gfp_t gfp_mask
, unsigned int order
, const char *fmt
, ...)
2965 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2967 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2968 debug_guardpage_minorder() > 0)
2972 * This documents exceptions given to allocations in certain
2973 * contexts that are allowed to allocate outside current's set
2976 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2977 if (test_thread_flag(TIF_MEMDIE
) ||
2978 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2979 filter
&= ~SHOW_MEM_FILTER_NODES
;
2980 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
2981 filter
&= ~SHOW_MEM_FILTER_NODES
;
2984 struct va_format vaf
;
2987 va_start(args
, fmt
);
2992 pr_warn("%pV", &vaf
);
2997 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2998 current
->comm
, order
, gfp_mask
, &gfp_mask
);
3000 if (!should_suppress_show_mem())
3004 static inline struct page
*
3005 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3006 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3008 struct oom_control oc
= {
3009 .zonelist
= ac
->zonelist
,
3010 .nodemask
= ac
->nodemask
,
3012 .gfp_mask
= gfp_mask
,
3017 *did_some_progress
= 0;
3020 * Acquire the oom lock. If that fails, somebody else is
3021 * making progress for us.
3023 if (!mutex_trylock(&oom_lock
)) {
3024 *did_some_progress
= 1;
3025 schedule_timeout_uninterruptible(1);
3030 * Go through the zonelist yet one more time, keep very high watermark
3031 * here, this is only to catch a parallel oom killing, we must fail if
3032 * we're still under heavy pressure.
3034 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3035 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3039 if (!(gfp_mask
& __GFP_NOFAIL
)) {
3040 /* Coredumps can quickly deplete all memory reserves */
3041 if (current
->flags
& PF_DUMPCORE
)
3043 /* The OOM killer will not help higher order allocs */
3044 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3046 /* The OOM killer does not needlessly kill tasks for lowmem */
3047 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3049 if (pm_suspended_storage())
3052 * XXX: GFP_NOFS allocations should rather fail than rely on
3053 * other request to make a forward progress.
3054 * We are in an unfortunate situation where out_of_memory cannot
3055 * do much for this context but let's try it to at least get
3056 * access to memory reserved if the current task is killed (see
3057 * out_of_memory). Once filesystems are ready to handle allocation
3058 * failures more gracefully we should just bail out here.
3061 /* The OOM killer may not free memory on a specific node */
3062 if (gfp_mask
& __GFP_THISNODE
)
3065 /* Exhausted what can be done so it's blamo time */
3066 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3067 *did_some_progress
= 1;
3069 if (gfp_mask
& __GFP_NOFAIL
) {
3070 page
= get_page_from_freelist(gfp_mask
, order
,
3071 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
3073 * fallback to ignore cpuset restriction if our nodes
3077 page
= get_page_from_freelist(gfp_mask
, order
,
3078 ALLOC_NO_WATERMARKS
, ac
);
3082 mutex_unlock(&oom_lock
);
3087 * Maximum number of compaction retries wit a progress before OOM
3088 * killer is consider as the only way to move forward.
3090 #define MAX_COMPACT_RETRIES 16
3092 #ifdef CONFIG_COMPACTION
3093 /* Try memory compaction for high-order allocations before reclaim */
3094 static struct page
*
3095 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3096 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3097 enum compact_priority prio
, enum compact_result
*compact_result
)
3104 current
->flags
|= PF_MEMALLOC
;
3105 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3107 current
->flags
&= ~PF_MEMALLOC
;
3109 if (*compact_result
<= COMPACT_INACTIVE
)
3113 * At least in one zone compaction wasn't deferred or skipped, so let's
3114 * count a compaction stall
3116 count_vm_event(COMPACTSTALL
);
3118 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3121 struct zone
*zone
= page_zone(page
);
3123 zone
->compact_blockskip_flush
= false;
3124 compaction_defer_reset(zone
, order
, true);
3125 count_vm_event(COMPACTSUCCESS
);
3130 * It's bad if compaction run occurs and fails. The most likely reason
3131 * is that pages exist, but not enough to satisfy watermarks.
3133 count_vm_event(COMPACTFAIL
);
3141 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3142 enum compact_result compact_result
,
3143 enum compact_priority
*compact_priority
,
3144 int compaction_retries
)
3146 int max_retries
= MAX_COMPACT_RETRIES
;
3152 * compaction considers all the zone as desperately out of memory
3153 * so it doesn't really make much sense to retry except when the
3154 * failure could be caused by insufficient priority
3156 if (compaction_failed(compact_result
)) {
3157 if (*compact_priority
> MIN_COMPACT_PRIORITY
) {
3158 (*compact_priority
)--;
3165 * make sure the compaction wasn't deferred or didn't bail out early
3166 * due to locks contention before we declare that we should give up.
3167 * But do not retry if the given zonelist is not suitable for
3170 if (compaction_withdrawn(compact_result
))
3171 return compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3174 * !costly requests are much more important than __GFP_REPEAT
3175 * costly ones because they are de facto nofail and invoke OOM
3176 * killer to move on while costly can fail and users are ready
3177 * to cope with that. 1/4 retries is rather arbitrary but we
3178 * would need much more detailed feedback from compaction to
3179 * make a better decision.
3181 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3183 if (compaction_retries
<= max_retries
)
3189 static inline struct page
*
3190 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3191 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3192 enum compact_priority prio
, enum compact_result
*compact_result
)
3194 *compact_result
= COMPACT_SKIPPED
;
3199 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3200 enum compact_result compact_result
,
3201 enum compact_priority
*compact_priority
,
3202 int compaction_retries
)
3207 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3211 * There are setups with compaction disabled which would prefer to loop
3212 * inside the allocator rather than hit the oom killer prematurely.
3213 * Let's give them a good hope and keep retrying while the order-0
3214 * watermarks are OK.
3216 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3218 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3219 ac_classzone_idx(ac
), alloc_flags
))
3224 #endif /* CONFIG_COMPACTION */
3226 /* Perform direct synchronous page reclaim */
3228 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3229 const struct alloc_context
*ac
)
3231 struct reclaim_state reclaim_state
;
3236 /* We now go into synchronous reclaim */
3237 cpuset_memory_pressure_bump();
3238 current
->flags
|= PF_MEMALLOC
;
3239 lockdep_set_current_reclaim_state(gfp_mask
);
3240 reclaim_state
.reclaimed_slab
= 0;
3241 current
->reclaim_state
= &reclaim_state
;
3243 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3246 current
->reclaim_state
= NULL
;
3247 lockdep_clear_current_reclaim_state();
3248 current
->flags
&= ~PF_MEMALLOC
;
3255 /* The really slow allocator path where we enter direct reclaim */
3256 static inline struct page
*
3257 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3258 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3259 unsigned long *did_some_progress
)
3261 struct page
*page
= NULL
;
3262 bool drained
= false;
3264 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3265 if (unlikely(!(*did_some_progress
)))
3269 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3272 * If an allocation failed after direct reclaim, it could be because
3273 * pages are pinned on the per-cpu lists or in high alloc reserves.
3274 * Shrink them them and try again
3276 if (!page
&& !drained
) {
3277 unreserve_highatomic_pageblock(ac
);
3278 drain_all_pages(NULL
);
3286 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3290 pg_data_t
*last_pgdat
= NULL
;
3292 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3293 ac
->high_zoneidx
, ac
->nodemask
) {
3294 if (last_pgdat
!= zone
->zone_pgdat
)
3295 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3296 last_pgdat
= zone
->zone_pgdat
;
3300 static inline unsigned int
3301 gfp_to_alloc_flags(gfp_t gfp_mask
)
3303 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3305 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3306 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3309 * The caller may dip into page reserves a bit more if the caller
3310 * cannot run direct reclaim, or if the caller has realtime scheduling
3311 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3312 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3314 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3316 if (gfp_mask
& __GFP_ATOMIC
) {
3318 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3319 * if it can't schedule.
3321 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3322 alloc_flags
|= ALLOC_HARDER
;
3324 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3325 * comment for __cpuset_node_allowed().
3327 alloc_flags
&= ~ALLOC_CPUSET
;
3328 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3329 alloc_flags
|= ALLOC_HARDER
;
3332 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3333 alloc_flags
|= ALLOC_CMA
;
3338 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3340 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3343 if (gfp_mask
& __GFP_MEMALLOC
)
3345 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3347 if (!in_interrupt() &&
3348 ((current
->flags
& PF_MEMALLOC
) ||
3349 unlikely(test_thread_flag(TIF_MEMDIE
))))
3356 * Maximum number of reclaim retries without any progress before OOM killer
3357 * is consider as the only way to move forward.
3359 #define MAX_RECLAIM_RETRIES 16
3362 * Checks whether it makes sense to retry the reclaim to make a forward progress
3363 * for the given allocation request.
3364 * The reclaim feedback represented by did_some_progress (any progress during
3365 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3366 * any progress in a row) is considered as well as the reclaimable pages on the
3367 * applicable zone list (with a backoff mechanism which is a function of
3368 * no_progress_loops).
3370 * Returns true if a retry is viable or false to enter the oom path.
3373 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3374 struct alloc_context
*ac
, int alloc_flags
,
3375 bool did_some_progress
, int no_progress_loops
)
3381 * Make sure we converge to OOM if we cannot make any progress
3382 * several times in the row.
3384 if (no_progress_loops
> MAX_RECLAIM_RETRIES
)
3388 * Keep reclaiming pages while there is a chance this will lead
3389 * somewhere. If none of the target zones can satisfy our allocation
3390 * request even if all reclaimable pages are considered then we are
3391 * screwed and have to go OOM.
3393 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3395 unsigned long available
;
3396 unsigned long reclaimable
;
3398 available
= reclaimable
= zone_reclaimable_pages(zone
);
3399 available
-= DIV_ROUND_UP(no_progress_loops
* available
,
3400 MAX_RECLAIM_RETRIES
);
3401 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3404 * Would the allocation succeed if we reclaimed the whole
3407 if (__zone_watermark_ok(zone
, order
, min_wmark_pages(zone
),
3408 ac_classzone_idx(ac
), alloc_flags
, available
)) {
3410 * If we didn't make any progress and have a lot of
3411 * dirty + writeback pages then we should wait for
3412 * an IO to complete to slow down the reclaim and
3413 * prevent from pre mature OOM
3415 if (!did_some_progress
) {
3416 unsigned long write_pending
;
3418 write_pending
= zone_page_state_snapshot(zone
,
3419 NR_ZONE_WRITE_PENDING
);
3421 if (2 * write_pending
> reclaimable
) {
3422 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3428 * Memory allocation/reclaim might be called from a WQ
3429 * context and the current implementation of the WQ
3430 * concurrency control doesn't recognize that
3431 * a particular WQ is congested if the worker thread is
3432 * looping without ever sleeping. Therefore we have to
3433 * do a short sleep here rather than calling
3436 if (current
->flags
& PF_WQ_WORKER
)
3437 schedule_timeout_uninterruptible(1);
3448 static inline struct page
*
3449 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3450 struct alloc_context
*ac
)
3452 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3453 struct page
*page
= NULL
;
3454 unsigned int alloc_flags
;
3455 unsigned long did_some_progress
;
3456 enum compact_priority compact_priority
= DEF_COMPACT_PRIORITY
;
3457 enum compact_result compact_result
;
3458 int compaction_retries
= 0;
3459 int no_progress_loops
= 0;
3462 * In the slowpath, we sanity check order to avoid ever trying to
3463 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3464 * be using allocators in order of preference for an area that is
3467 if (order
>= MAX_ORDER
) {
3468 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3473 * We also sanity check to catch abuse of atomic reserves being used by
3474 * callers that are not in atomic context.
3476 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3477 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3478 gfp_mask
&= ~__GFP_ATOMIC
;
3481 * The fast path uses conservative alloc_flags to succeed only until
3482 * kswapd needs to be woken up, and to avoid the cost of setting up
3483 * alloc_flags precisely. So we do that now.
3485 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3487 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3488 wake_all_kswapds(order
, ac
);
3491 * The adjusted alloc_flags might result in immediate success, so try
3494 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3499 * For costly allocations, try direct compaction first, as it's likely
3500 * that we have enough base pages and don't need to reclaim. Don't try
3501 * that for allocations that are allowed to ignore watermarks, as the
3502 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3504 if (can_direct_reclaim
&& order
> PAGE_ALLOC_COSTLY_ORDER
&&
3505 !gfp_pfmemalloc_allowed(gfp_mask
)) {
3506 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3508 INIT_COMPACT_PRIORITY
,
3514 * Checks for costly allocations with __GFP_NORETRY, which
3515 * includes THP page fault allocations
3517 if (gfp_mask
& __GFP_NORETRY
) {
3519 * If compaction is deferred for high-order allocations,
3520 * it is because sync compaction recently failed. If
3521 * this is the case and the caller requested a THP
3522 * allocation, we do not want to heavily disrupt the
3523 * system, so we fail the allocation instead of entering
3526 if (compact_result
== COMPACT_DEFERRED
)
3530 * Looks like reclaim/compaction is worth trying, but
3531 * sync compaction could be very expensive, so keep
3532 * using async compaction.
3534 compact_priority
= INIT_COMPACT_PRIORITY
;
3539 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3540 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3541 wake_all_kswapds(order
, ac
);
3543 if (gfp_pfmemalloc_allowed(gfp_mask
))
3544 alloc_flags
= ALLOC_NO_WATERMARKS
;
3547 * Reset the zonelist iterators if memory policies can be ignored.
3548 * These allocations are high priority and system rather than user
3551 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3552 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3553 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3554 ac
->high_zoneidx
, ac
->nodemask
);
3557 /* Attempt with potentially adjusted zonelist and alloc_flags */
3558 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3562 /* Caller is not willing to reclaim, we can't balance anything */
3563 if (!can_direct_reclaim
) {
3565 * All existing users of the __GFP_NOFAIL are blockable, so warn
3566 * of any new users that actually allow this type of allocation
3569 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3573 /* Avoid recursion of direct reclaim */
3574 if (current
->flags
& PF_MEMALLOC
) {
3576 * __GFP_NOFAIL request from this context is rather bizarre
3577 * because we cannot reclaim anything and only can loop waiting
3578 * for somebody to do a work for us.
3580 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3587 /* Avoid allocations with no watermarks from looping endlessly */
3588 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3592 /* Try direct reclaim and then allocating */
3593 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3594 &did_some_progress
);
3598 /* Try direct compaction and then allocating */
3599 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3600 compact_priority
, &compact_result
);
3604 if (order
&& compaction_made_progress(compact_result
))
3605 compaction_retries
++;
3607 /* Do not loop if specifically requested */
3608 if (gfp_mask
& __GFP_NORETRY
)
3612 * Do not retry costly high order allocations unless they are
3615 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_REPEAT
))
3619 * Costly allocations might have made a progress but this doesn't mean
3620 * their order will become available due to high fragmentation so
3621 * always increment the no progress counter for them
3623 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3624 no_progress_loops
= 0;
3626 no_progress_loops
++;
3628 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3629 did_some_progress
> 0, no_progress_loops
))
3633 * It doesn't make any sense to retry for the compaction if the order-0
3634 * reclaim is not able to make any progress because the current
3635 * implementation of the compaction depends on the sufficient amount
3636 * of free memory (see __compaction_suitable)
3638 if (did_some_progress
> 0 &&
3639 should_compact_retry(ac
, order
, alloc_flags
,
3640 compact_result
, &compact_priority
,
3641 compaction_retries
))
3644 /* Reclaim has failed us, start killing things */
3645 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3649 /* Retry as long as the OOM killer is making progress */
3650 if (did_some_progress
) {
3651 no_progress_loops
= 0;
3656 warn_alloc_failed(gfp_mask
, order
, NULL
);
3662 * This is the 'heart' of the zoned buddy allocator.
3665 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3666 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3669 unsigned int cpuset_mems_cookie
;
3670 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
3671 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3672 struct alloc_context ac
= {
3673 .high_zoneidx
= gfp_zone(gfp_mask
),
3674 .zonelist
= zonelist
,
3675 .nodemask
= nodemask
,
3676 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3679 if (cpusets_enabled()) {
3680 alloc_mask
|= __GFP_HARDWALL
;
3681 alloc_flags
|= ALLOC_CPUSET
;
3683 ac
.nodemask
= &cpuset_current_mems_allowed
;
3686 gfp_mask
&= gfp_allowed_mask
;
3688 lockdep_trace_alloc(gfp_mask
);
3690 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3692 if (should_fail_alloc_page(gfp_mask
, order
))
3696 * Check the zones suitable for the gfp_mask contain at least one
3697 * valid zone. It's possible to have an empty zonelist as a result
3698 * of __GFP_THISNODE and a memoryless node
3700 if (unlikely(!zonelist
->_zonerefs
->zone
))
3703 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3704 alloc_flags
|= ALLOC_CMA
;
3707 cpuset_mems_cookie
= read_mems_allowed_begin();
3709 /* Dirty zone balancing only done in the fast path */
3710 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3713 * The preferred zone is used for statistics but crucially it is
3714 * also used as the starting point for the zonelist iterator. It
3715 * may get reset for allocations that ignore memory policies.
3717 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3718 ac
.high_zoneidx
, ac
.nodemask
);
3719 if (!ac
.preferred_zoneref
) {
3724 /* First allocation attempt */
3725 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3730 * Runtime PM, block IO and its error handling path can deadlock
3731 * because I/O on the device might not complete.
3733 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3734 ac
.spread_dirty_pages
= false;
3737 * Restore the original nodemask if it was potentially replaced with
3738 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3740 if (cpusets_enabled())
3741 ac
.nodemask
= nodemask
;
3742 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3746 * When updating a task's mems_allowed, it is possible to race with
3747 * parallel threads in such a way that an allocation can fail while
3748 * the mask is being updated. If a page allocation is about to fail,
3749 * check if the cpuset changed during allocation and if so, retry.
3751 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3752 alloc_mask
= gfp_mask
;
3757 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
3758 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
3759 __free_pages(page
, order
);
3763 if (kmemcheck_enabled
&& page
)
3764 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3766 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3770 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3773 * Common helper functions.
3775 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3780 * __get_free_pages() returns a 32-bit address, which cannot represent
3783 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3785 page
= alloc_pages(gfp_mask
, order
);
3788 return (unsigned long) page_address(page
);
3790 EXPORT_SYMBOL(__get_free_pages
);
3792 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3794 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3796 EXPORT_SYMBOL(get_zeroed_page
);
3798 void __free_pages(struct page
*page
, unsigned int order
)
3800 if (put_page_testzero(page
)) {
3802 free_hot_cold_page(page
, false);
3804 __free_pages_ok(page
, order
);
3808 EXPORT_SYMBOL(__free_pages
);
3810 void free_pages(unsigned long addr
, unsigned int order
)
3813 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3814 __free_pages(virt_to_page((void *)addr
), order
);
3818 EXPORT_SYMBOL(free_pages
);
3822 * An arbitrary-length arbitrary-offset area of memory which resides
3823 * within a 0 or higher order page. Multiple fragments within that page
3824 * are individually refcounted, in the page's reference counter.
3826 * The page_frag functions below provide a simple allocation framework for
3827 * page fragments. This is used by the network stack and network device
3828 * drivers to provide a backing region of memory for use as either an
3829 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3831 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3834 struct page
*page
= NULL
;
3835 gfp_t gfp
= gfp_mask
;
3837 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3838 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3840 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3841 PAGE_FRAG_CACHE_MAX_ORDER
);
3842 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3844 if (unlikely(!page
))
3845 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3847 nc
->va
= page
? page_address(page
) : NULL
;
3852 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3853 unsigned int fragsz
, gfp_t gfp_mask
)
3855 unsigned int size
= PAGE_SIZE
;
3859 if (unlikely(!nc
->va
)) {
3861 page
= __page_frag_refill(nc
, gfp_mask
);
3865 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3866 /* if size can vary use size else just use PAGE_SIZE */
3869 /* Even if we own the page, we do not use atomic_set().
3870 * This would break get_page_unless_zero() users.
3872 page_ref_add(page
, size
- 1);
3874 /* reset page count bias and offset to start of new frag */
3875 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3876 nc
->pagecnt_bias
= size
;
3880 offset
= nc
->offset
- fragsz
;
3881 if (unlikely(offset
< 0)) {
3882 page
= virt_to_page(nc
->va
);
3884 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3887 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3888 /* if size can vary use size else just use PAGE_SIZE */
3891 /* OK, page count is 0, we can safely set it */
3892 set_page_count(page
, size
);
3894 /* reset page count bias and offset to start of new frag */
3895 nc
->pagecnt_bias
= size
;
3896 offset
= size
- fragsz
;
3900 nc
->offset
= offset
;
3902 return nc
->va
+ offset
;
3904 EXPORT_SYMBOL(__alloc_page_frag
);
3907 * Frees a page fragment allocated out of either a compound or order 0 page.
3909 void __free_page_frag(void *addr
)
3911 struct page
*page
= virt_to_head_page(addr
);
3913 if (unlikely(put_page_testzero(page
)))
3914 __free_pages_ok(page
, compound_order(page
));
3916 EXPORT_SYMBOL(__free_page_frag
);
3918 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3922 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3923 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3925 split_page(virt_to_page((void *)addr
), order
);
3926 while (used
< alloc_end
) {
3931 return (void *)addr
;
3935 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3936 * @size: the number of bytes to allocate
3937 * @gfp_mask: GFP flags for the allocation
3939 * This function is similar to alloc_pages(), except that it allocates the
3940 * minimum number of pages to satisfy the request. alloc_pages() can only
3941 * allocate memory in power-of-two pages.
3943 * This function is also limited by MAX_ORDER.
3945 * Memory allocated by this function must be released by free_pages_exact().
3947 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3949 unsigned int order
= get_order(size
);
3952 addr
= __get_free_pages(gfp_mask
, order
);
3953 return make_alloc_exact(addr
, order
, size
);
3955 EXPORT_SYMBOL(alloc_pages_exact
);
3958 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3960 * @nid: the preferred node ID where memory should be allocated
3961 * @size: the number of bytes to allocate
3962 * @gfp_mask: GFP flags for the allocation
3964 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3967 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3969 unsigned int order
= get_order(size
);
3970 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3973 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3977 * free_pages_exact - release memory allocated via alloc_pages_exact()
3978 * @virt: the value returned by alloc_pages_exact.
3979 * @size: size of allocation, same value as passed to alloc_pages_exact().
3981 * Release the memory allocated by a previous call to alloc_pages_exact.
3983 void free_pages_exact(void *virt
, size_t size
)
3985 unsigned long addr
= (unsigned long)virt
;
3986 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3988 while (addr
< end
) {
3993 EXPORT_SYMBOL(free_pages_exact
);
3996 * nr_free_zone_pages - count number of pages beyond high watermark
3997 * @offset: The zone index of the highest zone
3999 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4000 * high watermark within all zones at or below a given zone index. For each
4001 * zone, the number of pages is calculated as:
4002 * managed_pages - high_pages
4004 static unsigned long nr_free_zone_pages(int offset
)
4009 /* Just pick one node, since fallback list is circular */
4010 unsigned long sum
= 0;
4012 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4014 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4015 unsigned long size
= zone
->managed_pages
;
4016 unsigned long high
= high_wmark_pages(zone
);
4025 * nr_free_buffer_pages - count number of pages beyond high watermark
4027 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4028 * watermark within ZONE_DMA and ZONE_NORMAL.
4030 unsigned long nr_free_buffer_pages(void)
4032 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4034 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4037 * nr_free_pagecache_pages - count number of pages beyond high watermark
4039 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4040 * high watermark within all zones.
4042 unsigned long nr_free_pagecache_pages(void)
4044 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4047 static inline void show_node(struct zone
*zone
)
4049 if (IS_ENABLED(CONFIG_NUMA
))
4050 printk("Node %d ", zone_to_nid(zone
));
4053 long si_mem_available(void)
4056 unsigned long pagecache
;
4057 unsigned long wmark_low
= 0;
4058 unsigned long pages
[NR_LRU_LISTS
];
4062 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4063 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4066 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4069 * Estimate the amount of memory available for userspace allocations,
4070 * without causing swapping.
4072 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4075 * Not all the page cache can be freed, otherwise the system will
4076 * start swapping. Assume at least half of the page cache, or the
4077 * low watermark worth of cache, needs to stay.
4079 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4080 pagecache
-= min(pagecache
/ 2, wmark_low
);
4081 available
+= pagecache
;
4084 * Part of the reclaimable slab consists of items that are in use,
4085 * and cannot be freed. Cap this estimate at the low watermark.
4087 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4088 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4094 EXPORT_SYMBOL_GPL(si_mem_available
);
4096 void si_meminfo(struct sysinfo
*val
)
4098 val
->totalram
= totalram_pages
;
4099 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4100 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4101 val
->bufferram
= nr_blockdev_pages();
4102 val
->totalhigh
= totalhigh_pages
;
4103 val
->freehigh
= nr_free_highpages();
4104 val
->mem_unit
= PAGE_SIZE
;
4107 EXPORT_SYMBOL(si_meminfo
);
4110 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4112 int zone_type
; /* needs to be signed */
4113 unsigned long managed_pages
= 0;
4114 unsigned long managed_highpages
= 0;
4115 unsigned long free_highpages
= 0;
4116 pg_data_t
*pgdat
= NODE_DATA(nid
);
4118 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4119 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4120 val
->totalram
= managed_pages
;
4121 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4122 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4123 #ifdef CONFIG_HIGHMEM
4124 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4125 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4127 if (is_highmem(zone
)) {
4128 managed_highpages
+= zone
->managed_pages
;
4129 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4132 val
->totalhigh
= managed_highpages
;
4133 val
->freehigh
= free_highpages
;
4135 val
->totalhigh
= managed_highpages
;
4136 val
->freehigh
= free_highpages
;
4138 val
->mem_unit
= PAGE_SIZE
;
4143 * Determine whether the node should be displayed or not, depending on whether
4144 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4146 bool skip_free_areas_node(unsigned int flags
, int nid
)
4149 unsigned int cpuset_mems_cookie
;
4151 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4155 cpuset_mems_cookie
= read_mems_allowed_begin();
4156 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
4157 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
4162 #define K(x) ((x) << (PAGE_SHIFT-10))
4164 static void show_migration_types(unsigned char type
)
4166 static const char types
[MIGRATE_TYPES
] = {
4167 [MIGRATE_UNMOVABLE
] = 'U',
4168 [MIGRATE_MOVABLE
] = 'M',
4169 [MIGRATE_RECLAIMABLE
] = 'E',
4170 [MIGRATE_HIGHATOMIC
] = 'H',
4172 [MIGRATE_CMA
] = 'C',
4174 #ifdef CONFIG_MEMORY_ISOLATION
4175 [MIGRATE_ISOLATE
] = 'I',
4178 char tmp
[MIGRATE_TYPES
+ 1];
4182 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4183 if (type
& (1 << i
))
4188 printk("(%s) ", tmp
);
4192 * Show free area list (used inside shift_scroll-lock stuff)
4193 * We also calculate the percentage fragmentation. We do this by counting the
4194 * memory on each free list with the exception of the first item on the list.
4197 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4200 void show_free_areas(unsigned int filter
)
4202 unsigned long free_pcp
= 0;
4207 for_each_populated_zone(zone
) {
4208 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4211 for_each_online_cpu(cpu
)
4212 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4215 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4216 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4217 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4218 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4219 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4220 " free:%lu free_pcp:%lu free_cma:%lu\n",
4221 global_node_page_state(NR_ACTIVE_ANON
),
4222 global_node_page_state(NR_INACTIVE_ANON
),
4223 global_node_page_state(NR_ISOLATED_ANON
),
4224 global_node_page_state(NR_ACTIVE_FILE
),
4225 global_node_page_state(NR_INACTIVE_FILE
),
4226 global_node_page_state(NR_ISOLATED_FILE
),
4227 global_node_page_state(NR_UNEVICTABLE
),
4228 global_node_page_state(NR_FILE_DIRTY
),
4229 global_node_page_state(NR_WRITEBACK
),
4230 global_node_page_state(NR_UNSTABLE_NFS
),
4231 global_page_state(NR_SLAB_RECLAIMABLE
),
4232 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4233 global_node_page_state(NR_FILE_MAPPED
),
4234 global_node_page_state(NR_SHMEM
),
4235 global_page_state(NR_PAGETABLE
),
4236 global_page_state(NR_BOUNCE
),
4237 global_page_state(NR_FREE_PAGES
),
4239 global_page_state(NR_FREE_CMA_PAGES
));
4241 for_each_online_pgdat(pgdat
) {
4243 " active_anon:%lukB"
4244 " inactive_anon:%lukB"
4245 " active_file:%lukB"
4246 " inactive_file:%lukB"
4247 " unevictable:%lukB"
4248 " isolated(anon):%lukB"
4249 " isolated(file):%lukB"
4254 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4256 " shmem_pmdmapped: %lukB"
4259 " writeback_tmp:%lukB"
4261 " pages_scanned:%lu"
4262 " all_unreclaimable? %s"
4265 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4266 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4267 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4268 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4269 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4270 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4271 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4272 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4273 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4274 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4275 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4276 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4277 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4279 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4281 K(node_page_state(pgdat
, NR_SHMEM
)),
4282 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4283 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4284 node_page_state(pgdat
, NR_PAGES_SCANNED
),
4285 !pgdat_reclaimable(pgdat
) ? "yes" : "no");
4288 for_each_populated_zone(zone
) {
4291 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4295 for_each_online_cpu(cpu
)
4296 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4304 " active_anon:%lukB"
4305 " inactive_anon:%lukB"
4306 " active_file:%lukB"
4307 " inactive_file:%lukB"
4308 " unevictable:%lukB"
4309 " writepending:%lukB"
4313 " slab_reclaimable:%lukB"
4314 " slab_unreclaimable:%lukB"
4315 " kernel_stack:%lukB"
4323 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4324 K(min_wmark_pages(zone
)),
4325 K(low_wmark_pages(zone
)),
4326 K(high_wmark_pages(zone
)),
4327 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4328 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4329 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4330 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4331 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4332 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4333 K(zone
->present_pages
),
4334 K(zone
->managed_pages
),
4335 K(zone_page_state(zone
, NR_MLOCK
)),
4336 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4337 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4338 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4339 K(zone_page_state(zone
, NR_PAGETABLE
)),
4340 K(zone_page_state(zone
, NR_BOUNCE
)),
4342 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4343 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4344 printk("lowmem_reserve[]:");
4345 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4346 printk(" %ld", zone
->lowmem_reserve
[i
]);
4350 for_each_populated_zone(zone
) {
4352 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4353 unsigned char types
[MAX_ORDER
];
4355 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4358 printk("%s: ", zone
->name
);
4360 spin_lock_irqsave(&zone
->lock
, flags
);
4361 for (order
= 0; order
< MAX_ORDER
; order
++) {
4362 struct free_area
*area
= &zone
->free_area
[order
];
4365 nr
[order
] = area
->nr_free
;
4366 total
+= nr
[order
] << order
;
4369 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4370 if (!list_empty(&area
->free_list
[type
]))
4371 types
[order
] |= 1 << type
;
4374 spin_unlock_irqrestore(&zone
->lock
, flags
);
4375 for (order
= 0; order
< MAX_ORDER
; order
++) {
4376 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
4378 show_migration_types(types
[order
]);
4380 printk("= %lukB\n", K(total
));
4383 hugetlb_show_meminfo();
4385 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4387 show_swap_cache_info();
4390 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4392 zoneref
->zone
= zone
;
4393 zoneref
->zone_idx
= zone_idx(zone
);
4397 * Builds allocation fallback zone lists.
4399 * Add all populated zones of a node to the zonelist.
4401 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4405 enum zone_type zone_type
= MAX_NR_ZONES
;
4409 zone
= pgdat
->node_zones
+ zone_type
;
4410 if (populated_zone(zone
)) {
4411 zoneref_set_zone(zone
,
4412 &zonelist
->_zonerefs
[nr_zones
++]);
4413 check_highest_zone(zone_type
);
4415 } while (zone_type
);
4423 * 0 = automatic detection of better ordering.
4424 * 1 = order by ([node] distance, -zonetype)
4425 * 2 = order by (-zonetype, [node] distance)
4427 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4428 * the same zonelist. So only NUMA can configure this param.
4430 #define ZONELIST_ORDER_DEFAULT 0
4431 #define ZONELIST_ORDER_NODE 1
4432 #define ZONELIST_ORDER_ZONE 2
4434 /* zonelist order in the kernel.
4435 * set_zonelist_order() will set this to NODE or ZONE.
4437 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4438 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4442 /* The value user specified ....changed by config */
4443 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4444 /* string for sysctl */
4445 #define NUMA_ZONELIST_ORDER_LEN 16
4446 char numa_zonelist_order
[16] = "default";
4449 * interface for configure zonelist ordering.
4450 * command line option "numa_zonelist_order"
4451 * = "[dD]efault - default, automatic configuration.
4452 * = "[nN]ode - order by node locality, then by zone within node
4453 * = "[zZ]one - order by zone, then by locality within zone
4456 static int __parse_numa_zonelist_order(char *s
)
4458 if (*s
== 'd' || *s
== 'D') {
4459 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4460 } else if (*s
== 'n' || *s
== 'N') {
4461 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4462 } else if (*s
== 'z' || *s
== 'Z') {
4463 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4465 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4471 static __init
int setup_numa_zonelist_order(char *s
)
4478 ret
= __parse_numa_zonelist_order(s
);
4480 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4484 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4487 * sysctl handler for numa_zonelist_order
4489 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4490 void __user
*buffer
, size_t *length
,
4493 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4495 static DEFINE_MUTEX(zl_order_mutex
);
4497 mutex_lock(&zl_order_mutex
);
4499 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4503 strcpy(saved_string
, (char *)table
->data
);
4505 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4509 int oldval
= user_zonelist_order
;
4511 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4514 * bogus value. restore saved string
4516 strncpy((char *)table
->data
, saved_string
,
4517 NUMA_ZONELIST_ORDER_LEN
);
4518 user_zonelist_order
= oldval
;
4519 } else if (oldval
!= user_zonelist_order
) {
4520 mutex_lock(&zonelists_mutex
);
4521 build_all_zonelists(NULL
, NULL
);
4522 mutex_unlock(&zonelists_mutex
);
4526 mutex_unlock(&zl_order_mutex
);
4531 #define MAX_NODE_LOAD (nr_online_nodes)
4532 static int node_load
[MAX_NUMNODES
];
4535 * find_next_best_node - find the next node that should appear in a given node's fallback list
4536 * @node: node whose fallback list we're appending
4537 * @used_node_mask: nodemask_t of already used nodes
4539 * We use a number of factors to determine which is the next node that should
4540 * appear on a given node's fallback list. The node should not have appeared
4541 * already in @node's fallback list, and it should be the next closest node
4542 * according to the distance array (which contains arbitrary distance values
4543 * from each node to each node in the system), and should also prefer nodes
4544 * with no CPUs, since presumably they'll have very little allocation pressure
4545 * on them otherwise.
4546 * It returns -1 if no node is found.
4548 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4551 int min_val
= INT_MAX
;
4552 int best_node
= NUMA_NO_NODE
;
4553 const struct cpumask
*tmp
= cpumask_of_node(0);
4555 /* Use the local node if we haven't already */
4556 if (!node_isset(node
, *used_node_mask
)) {
4557 node_set(node
, *used_node_mask
);
4561 for_each_node_state(n
, N_MEMORY
) {
4563 /* Don't want a node to appear more than once */
4564 if (node_isset(n
, *used_node_mask
))
4567 /* Use the distance array to find the distance */
4568 val
= node_distance(node
, n
);
4570 /* Penalize nodes under us ("prefer the next node") */
4573 /* Give preference to headless and unused nodes */
4574 tmp
= cpumask_of_node(n
);
4575 if (!cpumask_empty(tmp
))
4576 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4578 /* Slight preference for less loaded node */
4579 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4580 val
+= node_load
[n
];
4582 if (val
< min_val
) {
4589 node_set(best_node
, *used_node_mask
);
4596 * Build zonelists ordered by node and zones within node.
4597 * This results in maximum locality--normal zone overflows into local
4598 * DMA zone, if any--but risks exhausting DMA zone.
4600 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4603 struct zonelist
*zonelist
;
4605 zonelist
= &pgdat
->node_zonelists
[0];
4606 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4608 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4609 zonelist
->_zonerefs
[j
].zone
= NULL
;
4610 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4614 * Build gfp_thisnode zonelists
4616 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4619 struct zonelist
*zonelist
;
4621 zonelist
= &pgdat
->node_zonelists
[1];
4622 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4623 zonelist
->_zonerefs
[j
].zone
= NULL
;
4624 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4628 * Build zonelists ordered by zone and nodes within zones.
4629 * This results in conserving DMA zone[s] until all Normal memory is
4630 * exhausted, but results in overflowing to remote node while memory
4631 * may still exist in local DMA zone.
4633 static int node_order
[MAX_NUMNODES
];
4635 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4638 int zone_type
; /* needs to be signed */
4640 struct zonelist
*zonelist
;
4642 zonelist
= &pgdat
->node_zonelists
[0];
4644 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4645 for (j
= 0; j
< nr_nodes
; j
++) {
4646 node
= node_order
[j
];
4647 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4648 if (populated_zone(z
)) {
4650 &zonelist
->_zonerefs
[pos
++]);
4651 check_highest_zone(zone_type
);
4655 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4656 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4659 #if defined(CONFIG_64BIT)
4661 * Devices that require DMA32/DMA are relatively rare and do not justify a
4662 * penalty to every machine in case the specialised case applies. Default
4663 * to Node-ordering on 64-bit NUMA machines
4665 static int default_zonelist_order(void)
4667 return ZONELIST_ORDER_NODE
;
4671 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4672 * by the kernel. If processes running on node 0 deplete the low memory zone
4673 * then reclaim will occur more frequency increasing stalls and potentially
4674 * be easier to OOM if a large percentage of the zone is under writeback or
4675 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4676 * Hence, default to zone ordering on 32-bit.
4678 static int default_zonelist_order(void)
4680 return ZONELIST_ORDER_ZONE
;
4682 #endif /* CONFIG_64BIT */
4684 static void set_zonelist_order(void)
4686 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4687 current_zonelist_order
= default_zonelist_order();
4689 current_zonelist_order
= user_zonelist_order
;
4692 static void build_zonelists(pg_data_t
*pgdat
)
4695 nodemask_t used_mask
;
4696 int local_node
, prev_node
;
4697 struct zonelist
*zonelist
;
4698 unsigned int order
= current_zonelist_order
;
4700 /* initialize zonelists */
4701 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4702 zonelist
= pgdat
->node_zonelists
+ i
;
4703 zonelist
->_zonerefs
[0].zone
= NULL
;
4704 zonelist
->_zonerefs
[0].zone_idx
= 0;
4707 /* NUMA-aware ordering of nodes */
4708 local_node
= pgdat
->node_id
;
4709 load
= nr_online_nodes
;
4710 prev_node
= local_node
;
4711 nodes_clear(used_mask
);
4713 memset(node_order
, 0, sizeof(node_order
));
4716 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4718 * We don't want to pressure a particular node.
4719 * So adding penalty to the first node in same
4720 * distance group to make it round-robin.
4722 if (node_distance(local_node
, node
) !=
4723 node_distance(local_node
, prev_node
))
4724 node_load
[node
] = load
;
4728 if (order
== ZONELIST_ORDER_NODE
)
4729 build_zonelists_in_node_order(pgdat
, node
);
4731 node_order
[i
++] = node
; /* remember order */
4734 if (order
== ZONELIST_ORDER_ZONE
) {
4735 /* calculate node order -- i.e., DMA last! */
4736 build_zonelists_in_zone_order(pgdat
, i
);
4739 build_thisnode_zonelists(pgdat
);
4742 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4744 * Return node id of node used for "local" allocations.
4745 * I.e., first node id of first zone in arg node's generic zonelist.
4746 * Used for initializing percpu 'numa_mem', which is used primarily
4747 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4749 int local_memory_node(int node
)
4753 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4754 gfp_zone(GFP_KERNEL
),
4756 return z
->zone
->node
;
4760 static void setup_min_unmapped_ratio(void);
4761 static void setup_min_slab_ratio(void);
4762 #else /* CONFIG_NUMA */
4764 static void set_zonelist_order(void)
4766 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4769 static void build_zonelists(pg_data_t
*pgdat
)
4771 int node
, local_node
;
4773 struct zonelist
*zonelist
;
4775 local_node
= pgdat
->node_id
;
4777 zonelist
= &pgdat
->node_zonelists
[0];
4778 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4781 * Now we build the zonelist so that it contains the zones
4782 * of all the other nodes.
4783 * We don't want to pressure a particular node, so when
4784 * building the zones for node N, we make sure that the
4785 * zones coming right after the local ones are those from
4786 * node N+1 (modulo N)
4788 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4789 if (!node_online(node
))
4791 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4793 for (node
= 0; node
< local_node
; node
++) {
4794 if (!node_online(node
))
4796 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4799 zonelist
->_zonerefs
[j
].zone
= NULL
;
4800 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4803 #endif /* CONFIG_NUMA */
4806 * Boot pageset table. One per cpu which is going to be used for all
4807 * zones and all nodes. The parameters will be set in such a way
4808 * that an item put on a list will immediately be handed over to
4809 * the buddy list. This is safe since pageset manipulation is done
4810 * with interrupts disabled.
4812 * The boot_pagesets must be kept even after bootup is complete for
4813 * unused processors and/or zones. They do play a role for bootstrapping
4814 * hotplugged processors.
4816 * zoneinfo_show() and maybe other functions do
4817 * not check if the processor is online before following the pageset pointer.
4818 * Other parts of the kernel may not check if the zone is available.
4820 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4821 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4822 static void setup_zone_pageset(struct zone
*zone
);
4825 * Global mutex to protect against size modification of zonelists
4826 * as well as to serialize pageset setup for the new populated zone.
4828 DEFINE_MUTEX(zonelists_mutex
);
4830 /* return values int ....just for stop_machine() */
4831 static int __build_all_zonelists(void *data
)
4835 pg_data_t
*self
= data
;
4838 memset(node_load
, 0, sizeof(node_load
));
4841 if (self
&& !node_online(self
->node_id
)) {
4842 build_zonelists(self
);
4845 for_each_online_node(nid
) {
4846 pg_data_t
*pgdat
= NODE_DATA(nid
);
4848 build_zonelists(pgdat
);
4852 * Initialize the boot_pagesets that are going to be used
4853 * for bootstrapping processors. The real pagesets for
4854 * each zone will be allocated later when the per cpu
4855 * allocator is available.
4857 * boot_pagesets are used also for bootstrapping offline
4858 * cpus if the system is already booted because the pagesets
4859 * are needed to initialize allocators on a specific cpu too.
4860 * F.e. the percpu allocator needs the page allocator which
4861 * needs the percpu allocator in order to allocate its pagesets
4862 * (a chicken-egg dilemma).
4864 for_each_possible_cpu(cpu
) {
4865 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4867 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4869 * We now know the "local memory node" for each node--
4870 * i.e., the node of the first zone in the generic zonelist.
4871 * Set up numa_mem percpu variable for on-line cpus. During
4872 * boot, only the boot cpu should be on-line; we'll init the
4873 * secondary cpus' numa_mem as they come on-line. During
4874 * node/memory hotplug, we'll fixup all on-line cpus.
4876 if (cpu_online(cpu
))
4877 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4884 static noinline
void __init
4885 build_all_zonelists_init(void)
4887 __build_all_zonelists(NULL
);
4888 mminit_verify_zonelist();
4889 cpuset_init_current_mems_allowed();
4893 * Called with zonelists_mutex held always
4894 * unless system_state == SYSTEM_BOOTING.
4896 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4897 * [we're only called with non-NULL zone through __meminit paths] and
4898 * (2) call of __init annotated helper build_all_zonelists_init
4899 * [protected by SYSTEM_BOOTING].
4901 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4903 set_zonelist_order();
4905 if (system_state
== SYSTEM_BOOTING
) {
4906 build_all_zonelists_init();
4908 #ifdef CONFIG_MEMORY_HOTPLUG
4910 setup_zone_pageset(zone
);
4912 /* we have to stop all cpus to guarantee there is no user
4914 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4915 /* cpuset refresh routine should be here */
4917 vm_total_pages
= nr_free_pagecache_pages();
4919 * Disable grouping by mobility if the number of pages in the
4920 * system is too low to allow the mechanism to work. It would be
4921 * more accurate, but expensive to check per-zone. This check is
4922 * made on memory-hotadd so a system can start with mobility
4923 * disabled and enable it later
4925 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4926 page_group_by_mobility_disabled
= 1;
4928 page_group_by_mobility_disabled
= 0;
4930 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4932 zonelist_order_name
[current_zonelist_order
],
4933 page_group_by_mobility_disabled
? "off" : "on",
4936 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4941 * Helper functions to size the waitqueue hash table.
4942 * Essentially these want to choose hash table sizes sufficiently
4943 * large so that collisions trying to wait on pages are rare.
4944 * But in fact, the number of active page waitqueues on typical
4945 * systems is ridiculously low, less than 200. So this is even
4946 * conservative, even though it seems large.
4948 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4949 * waitqueues, i.e. the size of the waitq table given the number of pages.
4951 #define PAGES_PER_WAITQUEUE 256
4953 #ifndef CONFIG_MEMORY_HOTPLUG
4954 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4956 unsigned long size
= 1;
4958 pages
/= PAGES_PER_WAITQUEUE
;
4960 while (size
< pages
)
4964 * Once we have dozens or even hundreds of threads sleeping
4965 * on IO we've got bigger problems than wait queue collision.
4966 * Limit the size of the wait table to a reasonable size.
4968 size
= min(size
, 4096UL);
4970 return max(size
, 4UL);
4974 * A zone's size might be changed by hot-add, so it is not possible to determine
4975 * a suitable size for its wait_table. So we use the maximum size now.
4977 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4979 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4980 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4981 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4983 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4984 * or more by the traditional way. (See above). It equals:
4986 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4987 * ia64(16K page size) : = ( 8G + 4M)byte.
4988 * powerpc (64K page size) : = (32G +16M)byte.
4990 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4997 * This is an integer logarithm so that shifts can be used later
4998 * to extract the more random high bits from the multiplicative
4999 * hash function before the remainder is taken.
5001 static inline unsigned long wait_table_bits(unsigned long size
)
5007 * Initially all pages are reserved - free ones are freed
5008 * up by free_all_bootmem() once the early boot process is
5009 * done. Non-atomic initialization, single-pass.
5011 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5012 unsigned long start_pfn
, enum memmap_context context
)
5014 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5015 unsigned long end_pfn
= start_pfn
+ size
;
5016 pg_data_t
*pgdat
= NODE_DATA(nid
);
5018 unsigned long nr_initialised
= 0;
5019 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5020 struct memblock_region
*r
= NULL
, *tmp
;
5023 if (highest_memmap_pfn
< end_pfn
- 1)
5024 highest_memmap_pfn
= end_pfn
- 1;
5027 * Honor reservation requested by the driver for this ZONE_DEVICE
5030 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5031 start_pfn
+= altmap
->reserve
;
5033 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5035 * There can be holes in boot-time mem_map[]s handed to this
5036 * function. They do not exist on hotplugged memory.
5038 if (context
!= MEMMAP_EARLY
)
5041 if (!early_pfn_valid(pfn
))
5043 if (!early_pfn_in_nid(pfn
, nid
))
5045 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5048 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5050 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5051 * from zone_movable_pfn[nid] to end of each node should be
5052 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5054 if (!mirrored_kernelcore
&& zone_movable_pfn
[nid
])
5055 if (zone
== ZONE_NORMAL
&& pfn
>= zone_movable_pfn
[nid
])
5059 * Check given memblock attribute by firmware which can affect
5060 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5061 * mirrored, it's an overlapped memmap init. skip it.
5063 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5064 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5065 for_each_memblock(memory
, tmp
)
5066 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5070 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5071 memblock_is_mirror(r
)) {
5072 /* already initialized as NORMAL */
5073 pfn
= memblock_region_memory_end_pfn(r
);
5081 * Mark the block movable so that blocks are reserved for
5082 * movable at startup. This will force kernel allocations
5083 * to reserve their blocks rather than leaking throughout
5084 * the address space during boot when many long-lived
5085 * kernel allocations are made.
5087 * bitmap is created for zone's valid pfn range. but memmap
5088 * can be created for invalid pages (for alignment)
5089 * check here not to call set_pageblock_migratetype() against
5092 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5093 struct page
*page
= pfn_to_page(pfn
);
5095 __init_single_page(page
, pfn
, zone
, nid
);
5096 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5098 __init_single_pfn(pfn
, zone
, nid
);
5103 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5105 unsigned int order
, t
;
5106 for_each_migratetype_order(order
, t
) {
5107 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5108 zone
->free_area
[order
].nr_free
= 0;
5112 #ifndef __HAVE_ARCH_MEMMAP_INIT
5113 #define memmap_init(size, nid, zone, start_pfn) \
5114 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5117 static int zone_batchsize(struct zone
*zone
)
5123 * The per-cpu-pages pools are set to around 1000th of the
5124 * size of the zone. But no more than 1/2 of a meg.
5126 * OK, so we don't know how big the cache is. So guess.
5128 batch
= zone
->managed_pages
/ 1024;
5129 if (batch
* PAGE_SIZE
> 512 * 1024)
5130 batch
= (512 * 1024) / PAGE_SIZE
;
5131 batch
/= 4; /* We effectively *= 4 below */
5136 * Clamp the batch to a 2^n - 1 value. Having a power
5137 * of 2 value was found to be more likely to have
5138 * suboptimal cache aliasing properties in some cases.
5140 * For example if 2 tasks are alternately allocating
5141 * batches of pages, one task can end up with a lot
5142 * of pages of one half of the possible page colors
5143 * and the other with pages of the other colors.
5145 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5150 /* The deferral and batching of frees should be suppressed under NOMMU
5153 * The problem is that NOMMU needs to be able to allocate large chunks
5154 * of contiguous memory as there's no hardware page translation to
5155 * assemble apparent contiguous memory from discontiguous pages.
5157 * Queueing large contiguous runs of pages for batching, however,
5158 * causes the pages to actually be freed in smaller chunks. As there
5159 * can be a significant delay between the individual batches being
5160 * recycled, this leads to the once large chunks of space being
5161 * fragmented and becoming unavailable for high-order allocations.
5168 * pcp->high and pcp->batch values are related and dependent on one another:
5169 * ->batch must never be higher then ->high.
5170 * The following function updates them in a safe manner without read side
5173 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5174 * those fields changing asynchronously (acording the the above rule).
5176 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5177 * outside of boot time (or some other assurance that no concurrent updaters
5180 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5181 unsigned long batch
)
5183 /* start with a fail safe value for batch */
5187 /* Update high, then batch, in order */
5194 /* a companion to pageset_set_high() */
5195 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5197 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5200 static void pageset_init(struct per_cpu_pageset
*p
)
5202 struct per_cpu_pages
*pcp
;
5205 memset(p
, 0, sizeof(*p
));
5209 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5210 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5213 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5216 pageset_set_batch(p
, batch
);
5220 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5221 * to the value high for the pageset p.
5223 static void pageset_set_high(struct per_cpu_pageset
*p
,
5226 unsigned long batch
= max(1UL, high
/ 4);
5227 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5228 batch
= PAGE_SHIFT
* 8;
5230 pageset_update(&p
->pcp
, high
, batch
);
5233 static void pageset_set_high_and_batch(struct zone
*zone
,
5234 struct per_cpu_pageset
*pcp
)
5236 if (percpu_pagelist_fraction
)
5237 pageset_set_high(pcp
,
5238 (zone
->managed_pages
/
5239 percpu_pagelist_fraction
));
5241 pageset_set_batch(pcp
, zone_batchsize(zone
));
5244 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5246 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5249 pageset_set_high_and_batch(zone
, pcp
);
5252 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5255 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5256 for_each_possible_cpu(cpu
)
5257 zone_pageset_init(zone
, cpu
);
5261 * Allocate per cpu pagesets and initialize them.
5262 * Before this call only boot pagesets were available.
5264 void __init
setup_per_cpu_pageset(void)
5266 struct pglist_data
*pgdat
;
5269 for_each_populated_zone(zone
)
5270 setup_zone_pageset(zone
);
5272 for_each_online_pgdat(pgdat
)
5273 pgdat
->per_cpu_nodestats
=
5274 alloc_percpu(struct per_cpu_nodestat
);
5277 static noinline __ref
5278 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
5284 * The per-page waitqueue mechanism uses hashed waitqueues
5287 zone
->wait_table_hash_nr_entries
=
5288 wait_table_hash_nr_entries(zone_size_pages
);
5289 zone
->wait_table_bits
=
5290 wait_table_bits(zone
->wait_table_hash_nr_entries
);
5291 alloc_size
= zone
->wait_table_hash_nr_entries
5292 * sizeof(wait_queue_head_t
);
5294 if (!slab_is_available()) {
5295 zone
->wait_table
= (wait_queue_head_t
*)
5296 memblock_virt_alloc_node_nopanic(
5297 alloc_size
, zone
->zone_pgdat
->node_id
);
5300 * This case means that a zone whose size was 0 gets new memory
5301 * via memory hot-add.
5302 * But it may be the case that a new node was hot-added. In
5303 * this case vmalloc() will not be able to use this new node's
5304 * memory - this wait_table must be initialized to use this new
5305 * node itself as well.
5306 * To use this new node's memory, further consideration will be
5309 zone
->wait_table
= vmalloc(alloc_size
);
5311 if (!zone
->wait_table
)
5314 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
5315 init_waitqueue_head(zone
->wait_table
+ i
);
5320 static __meminit
void zone_pcp_init(struct zone
*zone
)
5323 * per cpu subsystem is not up at this point. The following code
5324 * relies on the ability of the linker to provide the
5325 * offset of a (static) per cpu variable into the per cpu area.
5327 zone
->pageset
= &boot_pageset
;
5329 if (populated_zone(zone
))
5330 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5331 zone
->name
, zone
->present_pages
,
5332 zone_batchsize(zone
));
5335 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5336 unsigned long zone_start_pfn
,
5339 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5341 ret
= zone_wait_table_init(zone
, size
);
5344 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5346 zone
->zone_start_pfn
= zone_start_pfn
;
5348 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5349 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5351 (unsigned long)zone_idx(zone
),
5352 zone_start_pfn
, (zone_start_pfn
+ size
));
5354 zone_init_free_lists(zone
);
5359 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5360 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5363 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5365 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5366 struct mminit_pfnnid_cache
*state
)
5368 unsigned long start_pfn
, end_pfn
;
5371 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5372 return state
->last_nid
;
5374 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5376 state
->last_start
= start_pfn
;
5377 state
->last_end
= end_pfn
;
5378 state
->last_nid
= nid
;
5383 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5386 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5387 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5388 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5390 * If an architecture guarantees that all ranges registered contain no holes
5391 * and may be freed, this this function may be used instead of calling
5392 * memblock_free_early_nid() manually.
5394 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5396 unsigned long start_pfn
, end_pfn
;
5399 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5400 start_pfn
= min(start_pfn
, max_low_pfn
);
5401 end_pfn
= min(end_pfn
, max_low_pfn
);
5403 if (start_pfn
< end_pfn
)
5404 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5405 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5411 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5412 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5414 * If an architecture guarantees that all ranges registered contain no holes and may
5415 * be freed, this function may be used instead of calling memory_present() manually.
5417 void __init
sparse_memory_present_with_active_regions(int nid
)
5419 unsigned long start_pfn
, end_pfn
;
5422 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5423 memory_present(this_nid
, start_pfn
, end_pfn
);
5427 * get_pfn_range_for_nid - Return the start and end page frames for a node
5428 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5429 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5430 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5432 * It returns the start and end page frame of a node based on information
5433 * provided by memblock_set_node(). If called for a node
5434 * with no available memory, a warning is printed and the start and end
5437 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5438 unsigned long *start_pfn
, unsigned long *end_pfn
)
5440 unsigned long this_start_pfn
, this_end_pfn
;
5446 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5447 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5448 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5451 if (*start_pfn
== -1UL)
5456 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5457 * assumption is made that zones within a node are ordered in monotonic
5458 * increasing memory addresses so that the "highest" populated zone is used
5460 static void __init
find_usable_zone_for_movable(void)
5463 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5464 if (zone_index
== ZONE_MOVABLE
)
5467 if (arch_zone_highest_possible_pfn
[zone_index
] >
5468 arch_zone_lowest_possible_pfn
[zone_index
])
5472 VM_BUG_ON(zone_index
== -1);
5473 movable_zone
= zone_index
;
5477 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5478 * because it is sized independent of architecture. Unlike the other zones,
5479 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5480 * in each node depending on the size of each node and how evenly kernelcore
5481 * is distributed. This helper function adjusts the zone ranges
5482 * provided by the architecture for a given node by using the end of the
5483 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5484 * zones within a node are in order of monotonic increases memory addresses
5486 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5487 unsigned long zone_type
,
5488 unsigned long node_start_pfn
,
5489 unsigned long node_end_pfn
,
5490 unsigned long *zone_start_pfn
,
5491 unsigned long *zone_end_pfn
)
5493 /* Only adjust if ZONE_MOVABLE is on this node */
5494 if (zone_movable_pfn
[nid
]) {
5495 /* Size ZONE_MOVABLE */
5496 if (zone_type
== ZONE_MOVABLE
) {
5497 *zone_start_pfn
= zone_movable_pfn
[nid
];
5498 *zone_end_pfn
= min(node_end_pfn
,
5499 arch_zone_highest_possible_pfn
[movable_zone
]);
5501 /* Check if this whole range is within ZONE_MOVABLE */
5502 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5503 *zone_start_pfn
= *zone_end_pfn
;
5508 * Return the number of pages a zone spans in a node, including holes
5509 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5511 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5512 unsigned long zone_type
,
5513 unsigned long node_start_pfn
,
5514 unsigned long node_end_pfn
,
5515 unsigned long *zone_start_pfn
,
5516 unsigned long *zone_end_pfn
,
5517 unsigned long *ignored
)
5519 /* When hotadd a new node from cpu_up(), the node should be empty */
5520 if (!node_start_pfn
&& !node_end_pfn
)
5523 /* Get the start and end of the zone */
5524 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5525 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5526 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5527 node_start_pfn
, node_end_pfn
,
5528 zone_start_pfn
, zone_end_pfn
);
5530 /* Check that this node has pages within the zone's required range */
5531 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5534 /* Move the zone boundaries inside the node if necessary */
5535 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5536 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5538 /* Return the spanned pages */
5539 return *zone_end_pfn
- *zone_start_pfn
;
5543 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5544 * then all holes in the requested range will be accounted for.
5546 unsigned long __meminit
__absent_pages_in_range(int nid
,
5547 unsigned long range_start_pfn
,
5548 unsigned long range_end_pfn
)
5550 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5551 unsigned long start_pfn
, end_pfn
;
5554 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5555 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5556 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5557 nr_absent
-= end_pfn
- start_pfn
;
5563 * absent_pages_in_range - Return number of page frames in holes within a range
5564 * @start_pfn: The start PFN to start searching for holes
5565 * @end_pfn: The end PFN to stop searching for holes
5567 * It returns the number of pages frames in memory holes within a range.
5569 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5570 unsigned long end_pfn
)
5572 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5575 /* Return the number of page frames in holes in a zone on a node */
5576 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5577 unsigned long zone_type
,
5578 unsigned long node_start_pfn
,
5579 unsigned long node_end_pfn
,
5580 unsigned long *ignored
)
5582 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5583 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5584 unsigned long zone_start_pfn
, zone_end_pfn
;
5585 unsigned long nr_absent
;
5587 /* When hotadd a new node from cpu_up(), the node should be empty */
5588 if (!node_start_pfn
&& !node_end_pfn
)
5591 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5592 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5594 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5595 node_start_pfn
, node_end_pfn
,
5596 &zone_start_pfn
, &zone_end_pfn
);
5597 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5600 * ZONE_MOVABLE handling.
5601 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5604 if (zone_movable_pfn
[nid
]) {
5605 if (mirrored_kernelcore
) {
5606 unsigned long start_pfn
, end_pfn
;
5607 struct memblock_region
*r
;
5609 for_each_memblock(memory
, r
) {
5610 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5611 zone_start_pfn
, zone_end_pfn
);
5612 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5613 zone_start_pfn
, zone_end_pfn
);
5615 if (zone_type
== ZONE_MOVABLE
&&
5616 memblock_is_mirror(r
))
5617 nr_absent
+= end_pfn
- start_pfn
;
5619 if (zone_type
== ZONE_NORMAL
&&
5620 !memblock_is_mirror(r
))
5621 nr_absent
+= end_pfn
- start_pfn
;
5624 if (zone_type
== ZONE_NORMAL
)
5625 nr_absent
+= node_end_pfn
- zone_movable_pfn
[nid
];
5632 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5633 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5634 unsigned long zone_type
,
5635 unsigned long node_start_pfn
,
5636 unsigned long node_end_pfn
,
5637 unsigned long *zone_start_pfn
,
5638 unsigned long *zone_end_pfn
,
5639 unsigned long *zones_size
)
5643 *zone_start_pfn
= node_start_pfn
;
5644 for (zone
= 0; zone
< zone_type
; zone
++)
5645 *zone_start_pfn
+= zones_size
[zone
];
5647 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5649 return zones_size
[zone_type
];
5652 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5653 unsigned long zone_type
,
5654 unsigned long node_start_pfn
,
5655 unsigned long node_end_pfn
,
5656 unsigned long *zholes_size
)
5661 return zholes_size
[zone_type
];
5664 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5666 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5667 unsigned long node_start_pfn
,
5668 unsigned long node_end_pfn
,
5669 unsigned long *zones_size
,
5670 unsigned long *zholes_size
)
5672 unsigned long realtotalpages
= 0, totalpages
= 0;
5675 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5676 struct zone
*zone
= pgdat
->node_zones
+ i
;
5677 unsigned long zone_start_pfn
, zone_end_pfn
;
5678 unsigned long size
, real_size
;
5680 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5686 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5687 node_start_pfn
, node_end_pfn
,
5690 zone
->zone_start_pfn
= zone_start_pfn
;
5692 zone
->zone_start_pfn
= 0;
5693 zone
->spanned_pages
= size
;
5694 zone
->present_pages
= real_size
;
5697 realtotalpages
+= real_size
;
5700 pgdat
->node_spanned_pages
= totalpages
;
5701 pgdat
->node_present_pages
= realtotalpages
;
5702 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5706 #ifndef CONFIG_SPARSEMEM
5708 * Calculate the size of the zone->blockflags rounded to an unsigned long
5709 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5710 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5711 * round what is now in bits to nearest long in bits, then return it in
5714 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5716 unsigned long usemapsize
;
5718 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5719 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5720 usemapsize
= usemapsize
>> pageblock_order
;
5721 usemapsize
*= NR_PAGEBLOCK_BITS
;
5722 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5724 return usemapsize
/ 8;
5727 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5729 unsigned long zone_start_pfn
,
5730 unsigned long zonesize
)
5732 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5733 zone
->pageblock_flags
= NULL
;
5735 zone
->pageblock_flags
=
5736 memblock_virt_alloc_node_nopanic(usemapsize
,
5740 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5741 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5742 #endif /* CONFIG_SPARSEMEM */
5744 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5746 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5747 void __paginginit
set_pageblock_order(void)
5751 /* Check that pageblock_nr_pages has not already been setup */
5752 if (pageblock_order
)
5755 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5756 order
= HUGETLB_PAGE_ORDER
;
5758 order
= MAX_ORDER
- 1;
5761 * Assume the largest contiguous order of interest is a huge page.
5762 * This value may be variable depending on boot parameters on IA64 and
5765 pageblock_order
= order
;
5767 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5770 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5771 * is unused as pageblock_order is set at compile-time. See
5772 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5775 void __paginginit
set_pageblock_order(void)
5779 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5781 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5782 unsigned long present_pages
)
5784 unsigned long pages
= spanned_pages
;
5787 * Provide a more accurate estimation if there are holes within
5788 * the zone and SPARSEMEM is in use. If there are holes within the
5789 * zone, each populated memory region may cost us one or two extra
5790 * memmap pages due to alignment because memmap pages for each
5791 * populated regions may not naturally algined on page boundary.
5792 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5794 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5795 IS_ENABLED(CONFIG_SPARSEMEM
))
5796 pages
= present_pages
;
5798 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5802 * Set up the zone data structures:
5803 * - mark all pages reserved
5804 * - mark all memory queues empty
5805 * - clear the memory bitmaps
5807 * NOTE: pgdat should get zeroed by caller.
5809 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5812 int nid
= pgdat
->node_id
;
5815 pgdat_resize_init(pgdat
);
5816 #ifdef CONFIG_NUMA_BALANCING
5817 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5818 pgdat
->numabalancing_migrate_nr_pages
= 0;
5819 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5821 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5822 spin_lock_init(&pgdat
->split_queue_lock
);
5823 INIT_LIST_HEAD(&pgdat
->split_queue
);
5824 pgdat
->split_queue_len
= 0;
5826 init_waitqueue_head(&pgdat
->kswapd_wait
);
5827 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5828 #ifdef CONFIG_COMPACTION
5829 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5831 pgdat_page_ext_init(pgdat
);
5832 spin_lock_init(&pgdat
->lru_lock
);
5833 lruvec_init(node_lruvec(pgdat
));
5835 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5836 struct zone
*zone
= pgdat
->node_zones
+ j
;
5837 unsigned long size
, realsize
, freesize
, memmap_pages
;
5838 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5840 size
= zone
->spanned_pages
;
5841 realsize
= freesize
= zone
->present_pages
;
5844 * Adjust freesize so that it accounts for how much memory
5845 * is used by this zone for memmap. This affects the watermark
5846 * and per-cpu initialisations
5848 memmap_pages
= calc_memmap_size(size
, realsize
);
5849 if (!is_highmem_idx(j
)) {
5850 if (freesize
>= memmap_pages
) {
5851 freesize
-= memmap_pages
;
5854 " %s zone: %lu pages used for memmap\n",
5855 zone_names
[j
], memmap_pages
);
5857 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5858 zone_names
[j
], memmap_pages
, freesize
);
5861 /* Account for reserved pages */
5862 if (j
== 0 && freesize
> dma_reserve
) {
5863 freesize
-= dma_reserve
;
5864 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5865 zone_names
[0], dma_reserve
);
5868 if (!is_highmem_idx(j
))
5869 nr_kernel_pages
+= freesize
;
5870 /* Charge for highmem memmap if there are enough kernel pages */
5871 else if (nr_kernel_pages
> memmap_pages
* 2)
5872 nr_kernel_pages
-= memmap_pages
;
5873 nr_all_pages
+= freesize
;
5876 * Set an approximate value for lowmem here, it will be adjusted
5877 * when the bootmem allocator frees pages into the buddy system.
5878 * And all highmem pages will be managed by the buddy system.
5880 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5884 zone
->name
= zone_names
[j
];
5885 zone
->zone_pgdat
= pgdat
;
5886 spin_lock_init(&zone
->lock
);
5887 zone_seqlock_init(zone
);
5888 zone_pcp_init(zone
);
5893 set_pageblock_order();
5894 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5895 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5897 memmap_init(size
, nid
, j
, zone_start_pfn
);
5901 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
5903 unsigned long __maybe_unused start
= 0;
5904 unsigned long __maybe_unused offset
= 0;
5906 /* Skip empty nodes */
5907 if (!pgdat
->node_spanned_pages
)
5910 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5911 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5912 offset
= pgdat
->node_start_pfn
- start
;
5913 /* ia64 gets its own node_mem_map, before this, without bootmem */
5914 if (!pgdat
->node_mem_map
) {
5915 unsigned long size
, end
;
5919 * The zone's endpoints aren't required to be MAX_ORDER
5920 * aligned but the node_mem_map endpoints must be in order
5921 * for the buddy allocator to function correctly.
5923 end
= pgdat_end_pfn(pgdat
);
5924 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5925 size
= (end
- start
) * sizeof(struct page
);
5926 map
= alloc_remap(pgdat
->node_id
, size
);
5928 map
= memblock_virt_alloc_node_nopanic(size
,
5930 pgdat
->node_mem_map
= map
+ offset
;
5932 #ifndef CONFIG_NEED_MULTIPLE_NODES
5934 * With no DISCONTIG, the global mem_map is just set as node 0's
5936 if (pgdat
== NODE_DATA(0)) {
5937 mem_map
= NODE_DATA(0)->node_mem_map
;
5938 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5939 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5941 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5944 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5947 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5948 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5950 pg_data_t
*pgdat
= NODE_DATA(nid
);
5951 unsigned long start_pfn
= 0;
5952 unsigned long end_pfn
= 0;
5954 /* pg_data_t should be reset to zero when it's allocated */
5955 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
5957 reset_deferred_meminit(pgdat
);
5958 pgdat
->node_id
= nid
;
5959 pgdat
->node_start_pfn
= node_start_pfn
;
5960 pgdat
->per_cpu_nodestats
= NULL
;
5961 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5962 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5963 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5964 (u64
)start_pfn
<< PAGE_SHIFT
,
5965 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5967 start_pfn
= node_start_pfn
;
5969 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5970 zones_size
, zholes_size
);
5972 alloc_node_mem_map(pgdat
);
5973 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5974 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5975 nid
, (unsigned long)pgdat
,
5976 (unsigned long)pgdat
->node_mem_map
);
5979 free_area_init_core(pgdat
);
5982 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5984 #if MAX_NUMNODES > 1
5986 * Figure out the number of possible node ids.
5988 void __init
setup_nr_node_ids(void)
5990 unsigned int highest
;
5992 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5993 nr_node_ids
= highest
+ 1;
5998 * node_map_pfn_alignment - determine the maximum internode alignment
6000 * This function should be called after node map is populated and sorted.
6001 * It calculates the maximum power of two alignment which can distinguish
6004 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6005 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6006 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6007 * shifted, 1GiB is enough and this function will indicate so.
6009 * This is used to test whether pfn -> nid mapping of the chosen memory
6010 * model has fine enough granularity to avoid incorrect mapping for the
6011 * populated node map.
6013 * Returns the determined alignment in pfn's. 0 if there is no alignment
6014 * requirement (single node).
6016 unsigned long __init
node_map_pfn_alignment(void)
6018 unsigned long accl_mask
= 0, last_end
= 0;
6019 unsigned long start
, end
, mask
;
6023 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6024 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6031 * Start with a mask granular enough to pin-point to the
6032 * start pfn and tick off bits one-by-one until it becomes
6033 * too coarse to separate the current node from the last.
6035 mask
= ~((1 << __ffs(start
)) - 1);
6036 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6039 /* accumulate all internode masks */
6043 /* convert mask to number of pages */
6044 return ~accl_mask
+ 1;
6047 /* Find the lowest pfn for a node */
6048 static unsigned long __init
find_min_pfn_for_node(int nid
)
6050 unsigned long min_pfn
= ULONG_MAX
;
6051 unsigned long start_pfn
;
6054 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6055 min_pfn
= min(min_pfn
, start_pfn
);
6057 if (min_pfn
== ULONG_MAX
) {
6058 pr_warn("Could not find start_pfn for node %d\n", nid
);
6066 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6068 * It returns the minimum PFN based on information provided via
6069 * memblock_set_node().
6071 unsigned long __init
find_min_pfn_with_active_regions(void)
6073 return find_min_pfn_for_node(MAX_NUMNODES
);
6077 * early_calculate_totalpages()
6078 * Sum pages in active regions for movable zone.
6079 * Populate N_MEMORY for calculating usable_nodes.
6081 static unsigned long __init
early_calculate_totalpages(void)
6083 unsigned long totalpages
= 0;
6084 unsigned long start_pfn
, end_pfn
;
6087 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6088 unsigned long pages
= end_pfn
- start_pfn
;
6090 totalpages
+= pages
;
6092 node_set_state(nid
, N_MEMORY
);
6098 * Find the PFN the Movable zone begins in each node. Kernel memory
6099 * is spread evenly between nodes as long as the nodes have enough
6100 * memory. When they don't, some nodes will have more kernelcore than
6103 static void __init
find_zone_movable_pfns_for_nodes(void)
6106 unsigned long usable_startpfn
;
6107 unsigned long kernelcore_node
, kernelcore_remaining
;
6108 /* save the state before borrow the nodemask */
6109 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6110 unsigned long totalpages
= early_calculate_totalpages();
6111 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6112 struct memblock_region
*r
;
6114 /* Need to find movable_zone earlier when movable_node is specified. */
6115 find_usable_zone_for_movable();
6118 * If movable_node is specified, ignore kernelcore and movablecore
6121 if (movable_node_is_enabled()) {
6122 for_each_memblock(memory
, r
) {
6123 if (!memblock_is_hotpluggable(r
))
6128 usable_startpfn
= PFN_DOWN(r
->base
);
6129 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6130 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6138 * If kernelcore=mirror is specified, ignore movablecore option
6140 if (mirrored_kernelcore
) {
6141 bool mem_below_4gb_not_mirrored
= false;
6143 for_each_memblock(memory
, r
) {
6144 if (memblock_is_mirror(r
))
6149 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6151 if (usable_startpfn
< 0x100000) {
6152 mem_below_4gb_not_mirrored
= true;
6156 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6157 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6161 if (mem_below_4gb_not_mirrored
)
6162 pr_warn("This configuration results in unmirrored kernel memory.");
6168 * If movablecore=nn[KMG] was specified, calculate what size of
6169 * kernelcore that corresponds so that memory usable for
6170 * any allocation type is evenly spread. If both kernelcore
6171 * and movablecore are specified, then the value of kernelcore
6172 * will be used for required_kernelcore if it's greater than
6173 * what movablecore would have allowed.
6175 if (required_movablecore
) {
6176 unsigned long corepages
;
6179 * Round-up so that ZONE_MOVABLE is at least as large as what
6180 * was requested by the user
6182 required_movablecore
=
6183 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6184 required_movablecore
= min(totalpages
, required_movablecore
);
6185 corepages
= totalpages
- required_movablecore
;
6187 required_kernelcore
= max(required_kernelcore
, corepages
);
6191 * If kernelcore was not specified or kernelcore size is larger
6192 * than totalpages, there is no ZONE_MOVABLE.
6194 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6197 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6198 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6201 /* Spread kernelcore memory as evenly as possible throughout nodes */
6202 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6203 for_each_node_state(nid
, N_MEMORY
) {
6204 unsigned long start_pfn
, end_pfn
;
6207 * Recalculate kernelcore_node if the division per node
6208 * now exceeds what is necessary to satisfy the requested
6209 * amount of memory for the kernel
6211 if (required_kernelcore
< kernelcore_node
)
6212 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6215 * As the map is walked, we track how much memory is usable
6216 * by the kernel using kernelcore_remaining. When it is
6217 * 0, the rest of the node is usable by ZONE_MOVABLE
6219 kernelcore_remaining
= kernelcore_node
;
6221 /* Go through each range of PFNs within this node */
6222 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6223 unsigned long size_pages
;
6225 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6226 if (start_pfn
>= end_pfn
)
6229 /* Account for what is only usable for kernelcore */
6230 if (start_pfn
< usable_startpfn
) {
6231 unsigned long kernel_pages
;
6232 kernel_pages
= min(end_pfn
, usable_startpfn
)
6235 kernelcore_remaining
-= min(kernel_pages
,
6236 kernelcore_remaining
);
6237 required_kernelcore
-= min(kernel_pages
,
6238 required_kernelcore
);
6240 /* Continue if range is now fully accounted */
6241 if (end_pfn
<= usable_startpfn
) {
6244 * Push zone_movable_pfn to the end so
6245 * that if we have to rebalance
6246 * kernelcore across nodes, we will
6247 * not double account here
6249 zone_movable_pfn
[nid
] = end_pfn
;
6252 start_pfn
= usable_startpfn
;
6256 * The usable PFN range for ZONE_MOVABLE is from
6257 * start_pfn->end_pfn. Calculate size_pages as the
6258 * number of pages used as kernelcore
6260 size_pages
= end_pfn
- start_pfn
;
6261 if (size_pages
> kernelcore_remaining
)
6262 size_pages
= kernelcore_remaining
;
6263 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6266 * Some kernelcore has been met, update counts and
6267 * break if the kernelcore for this node has been
6270 required_kernelcore
-= min(required_kernelcore
,
6272 kernelcore_remaining
-= size_pages
;
6273 if (!kernelcore_remaining
)
6279 * If there is still required_kernelcore, we do another pass with one
6280 * less node in the count. This will push zone_movable_pfn[nid] further
6281 * along on the nodes that still have memory until kernelcore is
6285 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6289 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6290 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6291 zone_movable_pfn
[nid
] =
6292 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6295 /* restore the node_state */
6296 node_states
[N_MEMORY
] = saved_node_state
;
6299 /* Any regular or high memory on that node ? */
6300 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6302 enum zone_type zone_type
;
6304 if (N_MEMORY
== N_NORMAL_MEMORY
)
6307 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6308 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6309 if (populated_zone(zone
)) {
6310 node_set_state(nid
, N_HIGH_MEMORY
);
6311 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6312 zone_type
<= ZONE_NORMAL
)
6313 node_set_state(nid
, N_NORMAL_MEMORY
);
6320 * free_area_init_nodes - Initialise all pg_data_t and zone data
6321 * @max_zone_pfn: an array of max PFNs for each zone
6323 * This will call free_area_init_node() for each active node in the system.
6324 * Using the page ranges provided by memblock_set_node(), the size of each
6325 * zone in each node and their holes is calculated. If the maximum PFN
6326 * between two adjacent zones match, it is assumed that the zone is empty.
6327 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6328 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6329 * starts where the previous one ended. For example, ZONE_DMA32 starts
6330 * at arch_max_dma_pfn.
6332 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6334 unsigned long start_pfn
, end_pfn
;
6337 /* Record where the zone boundaries are */
6338 memset(arch_zone_lowest_possible_pfn
, 0,
6339 sizeof(arch_zone_lowest_possible_pfn
));
6340 memset(arch_zone_highest_possible_pfn
, 0,
6341 sizeof(arch_zone_highest_possible_pfn
));
6343 start_pfn
= find_min_pfn_with_active_regions();
6345 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6346 if (i
== ZONE_MOVABLE
)
6349 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6350 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6351 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6353 start_pfn
= end_pfn
;
6355 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6356 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6358 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6359 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6360 find_zone_movable_pfns_for_nodes();
6362 /* Print out the zone ranges */
6363 pr_info("Zone ranges:\n");
6364 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6365 if (i
== ZONE_MOVABLE
)
6367 pr_info(" %-8s ", zone_names
[i
]);
6368 if (arch_zone_lowest_possible_pfn
[i
] ==
6369 arch_zone_highest_possible_pfn
[i
])
6372 pr_cont("[mem %#018Lx-%#018Lx]\n",
6373 (u64
)arch_zone_lowest_possible_pfn
[i
]
6375 ((u64
)arch_zone_highest_possible_pfn
[i
]
6376 << PAGE_SHIFT
) - 1);
6379 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6380 pr_info("Movable zone start for each node\n");
6381 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6382 if (zone_movable_pfn
[i
])
6383 pr_info(" Node %d: %#018Lx\n", i
,
6384 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6387 /* Print out the early node map */
6388 pr_info("Early memory node ranges\n");
6389 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6390 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6391 (u64
)start_pfn
<< PAGE_SHIFT
,
6392 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6394 /* Initialise every node */
6395 mminit_verify_pageflags_layout();
6396 setup_nr_node_ids();
6397 for_each_online_node(nid
) {
6398 pg_data_t
*pgdat
= NODE_DATA(nid
);
6399 free_area_init_node(nid
, NULL
,
6400 find_min_pfn_for_node(nid
), NULL
);
6402 /* Any memory on that node */
6403 if (pgdat
->node_present_pages
)
6404 node_set_state(nid
, N_MEMORY
);
6405 check_for_memory(pgdat
, nid
);
6409 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6411 unsigned long long coremem
;
6415 coremem
= memparse(p
, &p
);
6416 *core
= coremem
>> PAGE_SHIFT
;
6418 /* Paranoid check that UL is enough for the coremem value */
6419 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6425 * kernelcore=size sets the amount of memory for use for allocations that
6426 * cannot be reclaimed or migrated.
6428 static int __init
cmdline_parse_kernelcore(char *p
)
6430 /* parse kernelcore=mirror */
6431 if (parse_option_str(p
, "mirror")) {
6432 mirrored_kernelcore
= true;
6436 return cmdline_parse_core(p
, &required_kernelcore
);
6440 * movablecore=size sets the amount of memory for use for allocations that
6441 * can be reclaimed or migrated.
6443 static int __init
cmdline_parse_movablecore(char *p
)
6445 return cmdline_parse_core(p
, &required_movablecore
);
6448 early_param("kernelcore", cmdline_parse_kernelcore
);
6449 early_param("movablecore", cmdline_parse_movablecore
);
6451 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6453 void adjust_managed_page_count(struct page
*page
, long count
)
6455 spin_lock(&managed_page_count_lock
);
6456 page_zone(page
)->managed_pages
+= count
;
6457 totalram_pages
+= count
;
6458 #ifdef CONFIG_HIGHMEM
6459 if (PageHighMem(page
))
6460 totalhigh_pages
+= count
;
6462 spin_unlock(&managed_page_count_lock
);
6464 EXPORT_SYMBOL(adjust_managed_page_count
);
6466 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6469 unsigned long pages
= 0;
6471 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6472 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6473 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6474 if ((unsigned int)poison
<= 0xFF)
6475 memset(pos
, poison
, PAGE_SIZE
);
6476 free_reserved_page(virt_to_page(pos
));
6480 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6481 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6485 EXPORT_SYMBOL(free_reserved_area
);
6487 #ifdef CONFIG_HIGHMEM
6488 void free_highmem_page(struct page
*page
)
6490 __free_reserved_page(page
);
6492 page_zone(page
)->managed_pages
++;
6498 void __init
mem_init_print_info(const char *str
)
6500 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6501 unsigned long init_code_size
, init_data_size
;
6503 physpages
= get_num_physpages();
6504 codesize
= _etext
- _stext
;
6505 datasize
= _edata
- _sdata
;
6506 rosize
= __end_rodata
- __start_rodata
;
6507 bss_size
= __bss_stop
- __bss_start
;
6508 init_data_size
= __init_end
- __init_begin
;
6509 init_code_size
= _einittext
- _sinittext
;
6512 * Detect special cases and adjust section sizes accordingly:
6513 * 1) .init.* may be embedded into .data sections
6514 * 2) .init.text.* may be out of [__init_begin, __init_end],
6515 * please refer to arch/tile/kernel/vmlinux.lds.S.
6516 * 3) .rodata.* may be embedded into .text or .data sections.
6518 #define adj_init_size(start, end, size, pos, adj) \
6520 if (start <= pos && pos < end && size > adj) \
6524 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6525 _sinittext
, init_code_size
);
6526 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6527 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6528 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6529 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6531 #undef adj_init_size
6533 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6534 #ifdef CONFIG_HIGHMEM
6538 nr_free_pages() << (PAGE_SHIFT
- 10),
6539 physpages
<< (PAGE_SHIFT
- 10),
6540 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6541 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6542 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6543 totalcma_pages
<< (PAGE_SHIFT
- 10),
6544 #ifdef CONFIG_HIGHMEM
6545 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6547 str
? ", " : "", str
? str
: "");
6551 * set_dma_reserve - set the specified number of pages reserved in the first zone
6552 * @new_dma_reserve: The number of pages to mark reserved
6554 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6555 * In the DMA zone, a significant percentage may be consumed by kernel image
6556 * and other unfreeable allocations which can skew the watermarks badly. This
6557 * function may optionally be used to account for unfreeable pages in the
6558 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6559 * smaller per-cpu batchsize.
6561 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6563 dma_reserve
= new_dma_reserve
;
6566 void __init
free_area_init(unsigned long *zones_size
)
6568 free_area_init_node(0, zones_size
,
6569 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6572 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6573 unsigned long action
, void *hcpu
)
6575 int cpu
= (unsigned long)hcpu
;
6577 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6578 lru_add_drain_cpu(cpu
);
6582 * Spill the event counters of the dead processor
6583 * into the current processors event counters.
6584 * This artificially elevates the count of the current
6587 vm_events_fold_cpu(cpu
);
6590 * Zero the differential counters of the dead processor
6591 * so that the vm statistics are consistent.
6593 * This is only okay since the processor is dead and cannot
6594 * race with what we are doing.
6596 cpu_vm_stats_fold(cpu
);
6601 void __init
page_alloc_init(void)
6603 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6607 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6608 * or min_free_kbytes changes.
6610 static void calculate_totalreserve_pages(void)
6612 struct pglist_data
*pgdat
;
6613 unsigned long reserve_pages
= 0;
6614 enum zone_type i
, j
;
6616 for_each_online_pgdat(pgdat
) {
6618 pgdat
->totalreserve_pages
= 0;
6620 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6621 struct zone
*zone
= pgdat
->node_zones
+ i
;
6624 /* Find valid and maximum lowmem_reserve in the zone */
6625 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6626 if (zone
->lowmem_reserve
[j
] > max
)
6627 max
= zone
->lowmem_reserve
[j
];
6630 /* we treat the high watermark as reserved pages. */
6631 max
+= high_wmark_pages(zone
);
6633 if (max
> zone
->managed_pages
)
6634 max
= zone
->managed_pages
;
6636 pgdat
->totalreserve_pages
+= max
;
6638 reserve_pages
+= max
;
6641 totalreserve_pages
= reserve_pages
;
6645 * setup_per_zone_lowmem_reserve - called whenever
6646 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6647 * has a correct pages reserved value, so an adequate number of
6648 * pages are left in the zone after a successful __alloc_pages().
6650 static void setup_per_zone_lowmem_reserve(void)
6652 struct pglist_data
*pgdat
;
6653 enum zone_type j
, idx
;
6655 for_each_online_pgdat(pgdat
) {
6656 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6657 struct zone
*zone
= pgdat
->node_zones
+ j
;
6658 unsigned long managed_pages
= zone
->managed_pages
;
6660 zone
->lowmem_reserve
[j
] = 0;
6664 struct zone
*lower_zone
;
6668 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6669 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6671 lower_zone
= pgdat
->node_zones
+ idx
;
6672 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6673 sysctl_lowmem_reserve_ratio
[idx
];
6674 managed_pages
+= lower_zone
->managed_pages
;
6679 /* update totalreserve_pages */
6680 calculate_totalreserve_pages();
6683 static void __setup_per_zone_wmarks(void)
6685 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6686 unsigned long lowmem_pages
= 0;
6688 unsigned long flags
;
6690 /* Calculate total number of !ZONE_HIGHMEM pages */
6691 for_each_zone(zone
) {
6692 if (!is_highmem(zone
))
6693 lowmem_pages
+= zone
->managed_pages
;
6696 for_each_zone(zone
) {
6699 spin_lock_irqsave(&zone
->lock
, flags
);
6700 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6701 do_div(tmp
, lowmem_pages
);
6702 if (is_highmem(zone
)) {
6704 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6705 * need highmem pages, so cap pages_min to a small
6708 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6709 * deltas control asynch page reclaim, and so should
6710 * not be capped for highmem.
6712 unsigned long min_pages
;
6714 min_pages
= zone
->managed_pages
/ 1024;
6715 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6716 zone
->watermark
[WMARK_MIN
] = min_pages
;
6719 * If it's a lowmem zone, reserve a number of pages
6720 * proportionate to the zone's size.
6722 zone
->watermark
[WMARK_MIN
] = tmp
;
6726 * Set the kswapd watermarks distance according to the
6727 * scale factor in proportion to available memory, but
6728 * ensure a minimum size on small systems.
6730 tmp
= max_t(u64
, tmp
>> 2,
6731 mult_frac(zone
->managed_pages
,
6732 watermark_scale_factor
, 10000));
6734 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6735 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6737 spin_unlock_irqrestore(&zone
->lock
, flags
);
6740 /* update totalreserve_pages */
6741 calculate_totalreserve_pages();
6745 * setup_per_zone_wmarks - called when min_free_kbytes changes
6746 * or when memory is hot-{added|removed}
6748 * Ensures that the watermark[min,low,high] values for each zone are set
6749 * correctly with respect to min_free_kbytes.
6751 void setup_per_zone_wmarks(void)
6753 mutex_lock(&zonelists_mutex
);
6754 __setup_per_zone_wmarks();
6755 mutex_unlock(&zonelists_mutex
);
6759 * Initialise min_free_kbytes.
6761 * For small machines we want it small (128k min). For large machines
6762 * we want it large (64MB max). But it is not linear, because network
6763 * bandwidth does not increase linearly with machine size. We use
6765 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6766 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6782 int __meminit
init_per_zone_wmark_min(void)
6784 unsigned long lowmem_kbytes
;
6785 int new_min_free_kbytes
;
6787 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6788 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6790 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6791 min_free_kbytes
= new_min_free_kbytes
;
6792 if (min_free_kbytes
< 128)
6793 min_free_kbytes
= 128;
6794 if (min_free_kbytes
> 65536)
6795 min_free_kbytes
= 65536;
6797 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6798 new_min_free_kbytes
, user_min_free_kbytes
);
6800 setup_per_zone_wmarks();
6801 refresh_zone_stat_thresholds();
6802 setup_per_zone_lowmem_reserve();
6805 setup_min_unmapped_ratio();
6806 setup_min_slab_ratio();
6811 core_initcall(init_per_zone_wmark_min
)
6814 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6815 * that we can call two helper functions whenever min_free_kbytes
6818 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6819 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6823 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6828 user_min_free_kbytes
= min_free_kbytes
;
6829 setup_per_zone_wmarks();
6834 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6835 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6839 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6844 setup_per_zone_wmarks();
6850 static void setup_min_unmapped_ratio(void)
6855 for_each_online_pgdat(pgdat
)
6856 pgdat
->min_unmapped_pages
= 0;
6859 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
6860 sysctl_min_unmapped_ratio
) / 100;
6864 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6865 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6869 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6873 setup_min_unmapped_ratio();
6878 static void setup_min_slab_ratio(void)
6883 for_each_online_pgdat(pgdat
)
6884 pgdat
->min_slab_pages
= 0;
6887 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
6888 sysctl_min_slab_ratio
) / 100;
6891 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6892 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6896 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6900 setup_min_slab_ratio();
6907 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6908 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6909 * whenever sysctl_lowmem_reserve_ratio changes.
6911 * The reserve ratio obviously has absolutely no relation with the
6912 * minimum watermarks. The lowmem reserve ratio can only make sense
6913 * if in function of the boot time zone sizes.
6915 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6916 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6918 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6919 setup_per_zone_lowmem_reserve();
6924 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6925 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6926 * pagelist can have before it gets flushed back to buddy allocator.
6928 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6929 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6932 int old_percpu_pagelist_fraction
;
6935 mutex_lock(&pcp_batch_high_lock
);
6936 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6938 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6939 if (!write
|| ret
< 0)
6942 /* Sanity checking to avoid pcp imbalance */
6943 if (percpu_pagelist_fraction
&&
6944 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6945 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6951 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6954 for_each_populated_zone(zone
) {
6957 for_each_possible_cpu(cpu
)
6958 pageset_set_high_and_batch(zone
,
6959 per_cpu_ptr(zone
->pageset
, cpu
));
6962 mutex_unlock(&pcp_batch_high_lock
);
6967 int hashdist
= HASHDIST_DEFAULT
;
6969 static int __init
set_hashdist(char *str
)
6973 hashdist
= simple_strtoul(str
, &str
, 0);
6976 __setup("hashdist=", set_hashdist
);
6980 * allocate a large system hash table from bootmem
6981 * - it is assumed that the hash table must contain an exact power-of-2
6982 * quantity of entries
6983 * - limit is the number of hash buckets, not the total allocation size
6985 void *__init
alloc_large_system_hash(const char *tablename
,
6986 unsigned long bucketsize
,
6987 unsigned long numentries
,
6990 unsigned int *_hash_shift
,
6991 unsigned int *_hash_mask
,
6992 unsigned long low_limit
,
6993 unsigned long high_limit
)
6995 unsigned long long max
= high_limit
;
6996 unsigned long log2qty
, size
;
6999 /* allow the kernel cmdline to have a say */
7001 /* round applicable memory size up to nearest megabyte */
7002 numentries
= nr_kernel_pages
;
7004 /* It isn't necessary when PAGE_SIZE >= 1MB */
7005 if (PAGE_SHIFT
< 20)
7006 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7008 /* limit to 1 bucket per 2^scale bytes of low memory */
7009 if (scale
> PAGE_SHIFT
)
7010 numentries
>>= (scale
- PAGE_SHIFT
);
7012 numentries
<<= (PAGE_SHIFT
- scale
);
7014 /* Make sure we've got at least a 0-order allocation.. */
7015 if (unlikely(flags
& HASH_SMALL
)) {
7016 /* Makes no sense without HASH_EARLY */
7017 WARN_ON(!(flags
& HASH_EARLY
));
7018 if (!(numentries
>> *_hash_shift
)) {
7019 numentries
= 1UL << *_hash_shift
;
7020 BUG_ON(!numentries
);
7022 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7023 numentries
= PAGE_SIZE
/ bucketsize
;
7025 numentries
= roundup_pow_of_two(numentries
);
7027 /* limit allocation size to 1/16 total memory by default */
7029 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7030 do_div(max
, bucketsize
);
7032 max
= min(max
, 0x80000000ULL
);
7034 if (numentries
< low_limit
)
7035 numentries
= low_limit
;
7036 if (numentries
> max
)
7039 log2qty
= ilog2(numentries
);
7042 size
= bucketsize
<< log2qty
;
7043 if (flags
& HASH_EARLY
)
7044 table
= memblock_virt_alloc_nopanic(size
, 0);
7046 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
7049 * If bucketsize is not a power-of-two, we may free
7050 * some pages at the end of hash table which
7051 * alloc_pages_exact() automatically does
7053 if (get_order(size
) < MAX_ORDER
) {
7054 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
7055 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
7058 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7061 panic("Failed to allocate %s hash table\n", tablename
);
7063 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7064 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7067 *_hash_shift
= log2qty
;
7069 *_hash_mask
= (1 << log2qty
) - 1;
7075 * This function checks whether pageblock includes unmovable pages or not.
7076 * If @count is not zero, it is okay to include less @count unmovable pages
7078 * PageLRU check without isolation or lru_lock could race so that
7079 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7080 * expect this function should be exact.
7082 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7083 bool skip_hwpoisoned_pages
)
7085 unsigned long pfn
, iter
, found
;
7089 * For avoiding noise data, lru_add_drain_all() should be called
7090 * If ZONE_MOVABLE, the zone never contains unmovable pages
7092 if (zone_idx(zone
) == ZONE_MOVABLE
)
7094 mt
= get_pageblock_migratetype(page
);
7095 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7098 pfn
= page_to_pfn(page
);
7099 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7100 unsigned long check
= pfn
+ iter
;
7102 if (!pfn_valid_within(check
))
7105 page
= pfn_to_page(check
);
7108 * Hugepages are not in LRU lists, but they're movable.
7109 * We need not scan over tail pages bacause we don't
7110 * handle each tail page individually in migration.
7112 if (PageHuge(page
)) {
7113 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7118 * We can't use page_count without pin a page
7119 * because another CPU can free compound page.
7120 * This check already skips compound tails of THP
7121 * because their page->_refcount is zero at all time.
7123 if (!page_ref_count(page
)) {
7124 if (PageBuddy(page
))
7125 iter
+= (1 << page_order(page
)) - 1;
7130 * The HWPoisoned page may be not in buddy system, and
7131 * page_count() is not 0.
7133 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7139 * If there are RECLAIMABLE pages, we need to check
7140 * it. But now, memory offline itself doesn't call
7141 * shrink_node_slabs() and it still to be fixed.
7144 * If the page is not RAM, page_count()should be 0.
7145 * we don't need more check. This is an _used_ not-movable page.
7147 * The problematic thing here is PG_reserved pages. PG_reserved
7148 * is set to both of a memory hole page and a _used_ kernel
7157 bool is_pageblock_removable_nolock(struct page
*page
)
7163 * We have to be careful here because we are iterating over memory
7164 * sections which are not zone aware so we might end up outside of
7165 * the zone but still within the section.
7166 * We have to take care about the node as well. If the node is offline
7167 * its NODE_DATA will be NULL - see page_zone.
7169 if (!node_online(page_to_nid(page
)))
7172 zone
= page_zone(page
);
7173 pfn
= page_to_pfn(page
);
7174 if (!zone_spans_pfn(zone
, pfn
))
7177 return !has_unmovable_pages(zone
, page
, 0, true);
7180 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7182 static unsigned long pfn_max_align_down(unsigned long pfn
)
7184 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7185 pageblock_nr_pages
) - 1);
7188 static unsigned long pfn_max_align_up(unsigned long pfn
)
7190 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7191 pageblock_nr_pages
));
7194 /* [start, end) must belong to a single zone. */
7195 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7196 unsigned long start
, unsigned long end
)
7198 /* This function is based on compact_zone() from compaction.c. */
7199 unsigned long nr_reclaimed
;
7200 unsigned long pfn
= start
;
7201 unsigned int tries
= 0;
7206 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7207 if (fatal_signal_pending(current
)) {
7212 if (list_empty(&cc
->migratepages
)) {
7213 cc
->nr_migratepages
= 0;
7214 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7220 } else if (++tries
== 5) {
7221 ret
= ret
< 0 ? ret
: -EBUSY
;
7225 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7227 cc
->nr_migratepages
-= nr_reclaimed
;
7229 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7230 NULL
, 0, cc
->mode
, MR_CMA
);
7233 putback_movable_pages(&cc
->migratepages
);
7240 * alloc_contig_range() -- tries to allocate given range of pages
7241 * @start: start PFN to allocate
7242 * @end: one-past-the-last PFN to allocate
7243 * @migratetype: migratetype of the underlaying pageblocks (either
7244 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7245 * in range must have the same migratetype and it must
7246 * be either of the two.
7248 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7249 * aligned, however it's the caller's responsibility to guarantee that
7250 * we are the only thread that changes migrate type of pageblocks the
7253 * The PFN range must belong to a single zone.
7255 * Returns zero on success or negative error code. On success all
7256 * pages which PFN is in [start, end) are allocated for the caller and
7257 * need to be freed with free_contig_range().
7259 int alloc_contig_range(unsigned long start
, unsigned long end
,
7260 unsigned migratetype
)
7262 unsigned long outer_start
, outer_end
;
7266 struct compact_control cc
= {
7267 .nr_migratepages
= 0,
7269 .zone
= page_zone(pfn_to_page(start
)),
7270 .mode
= MIGRATE_SYNC
,
7271 .ignore_skip_hint
= true,
7273 INIT_LIST_HEAD(&cc
.migratepages
);
7276 * What we do here is we mark all pageblocks in range as
7277 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7278 * have different sizes, and due to the way page allocator
7279 * work, we align the range to biggest of the two pages so
7280 * that page allocator won't try to merge buddies from
7281 * different pageblocks and change MIGRATE_ISOLATE to some
7282 * other migration type.
7284 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7285 * migrate the pages from an unaligned range (ie. pages that
7286 * we are interested in). This will put all the pages in
7287 * range back to page allocator as MIGRATE_ISOLATE.
7289 * When this is done, we take the pages in range from page
7290 * allocator removing them from the buddy system. This way
7291 * page allocator will never consider using them.
7293 * This lets us mark the pageblocks back as
7294 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7295 * aligned range but not in the unaligned, original range are
7296 * put back to page allocator so that buddy can use them.
7299 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7300 pfn_max_align_up(end
), migratetype
,
7306 * In case of -EBUSY, we'd like to know which page causes problem.
7307 * So, just fall through. We will check it in test_pages_isolated().
7309 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7310 if (ret
&& ret
!= -EBUSY
)
7314 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7315 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7316 * more, all pages in [start, end) are free in page allocator.
7317 * What we are going to do is to allocate all pages from
7318 * [start, end) (that is remove them from page allocator).
7320 * The only problem is that pages at the beginning and at the
7321 * end of interesting range may be not aligned with pages that
7322 * page allocator holds, ie. they can be part of higher order
7323 * pages. Because of this, we reserve the bigger range and
7324 * once this is done free the pages we are not interested in.
7326 * We don't have to hold zone->lock here because the pages are
7327 * isolated thus they won't get removed from buddy.
7330 lru_add_drain_all();
7331 drain_all_pages(cc
.zone
);
7334 outer_start
= start
;
7335 while (!PageBuddy(pfn_to_page(outer_start
))) {
7336 if (++order
>= MAX_ORDER
) {
7337 outer_start
= start
;
7340 outer_start
&= ~0UL << order
;
7343 if (outer_start
!= start
) {
7344 order
= page_order(pfn_to_page(outer_start
));
7347 * outer_start page could be small order buddy page and
7348 * it doesn't include start page. Adjust outer_start
7349 * in this case to report failed page properly
7350 * on tracepoint in test_pages_isolated()
7352 if (outer_start
+ (1UL << order
) <= start
)
7353 outer_start
= start
;
7356 /* Make sure the range is really isolated. */
7357 if (test_pages_isolated(outer_start
, end
, false)) {
7358 pr_info("%s: [%lx, %lx) PFNs busy\n",
7359 __func__
, outer_start
, end
);
7364 /* Grab isolated pages from freelists. */
7365 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7371 /* Free head and tail (if any) */
7372 if (start
!= outer_start
)
7373 free_contig_range(outer_start
, start
- outer_start
);
7374 if (end
!= outer_end
)
7375 free_contig_range(end
, outer_end
- end
);
7378 undo_isolate_page_range(pfn_max_align_down(start
),
7379 pfn_max_align_up(end
), migratetype
);
7383 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7385 unsigned int count
= 0;
7387 for (; nr_pages
--; pfn
++) {
7388 struct page
*page
= pfn_to_page(pfn
);
7390 count
+= page_count(page
) != 1;
7393 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7397 #ifdef CONFIG_MEMORY_HOTPLUG
7399 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7400 * page high values need to be recalulated.
7402 void __meminit
zone_pcp_update(struct zone
*zone
)
7405 mutex_lock(&pcp_batch_high_lock
);
7406 for_each_possible_cpu(cpu
)
7407 pageset_set_high_and_batch(zone
,
7408 per_cpu_ptr(zone
->pageset
, cpu
));
7409 mutex_unlock(&pcp_batch_high_lock
);
7413 void zone_pcp_reset(struct zone
*zone
)
7415 unsigned long flags
;
7417 struct per_cpu_pageset
*pset
;
7419 /* avoid races with drain_pages() */
7420 local_irq_save(flags
);
7421 if (zone
->pageset
!= &boot_pageset
) {
7422 for_each_online_cpu(cpu
) {
7423 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7424 drain_zonestat(zone
, pset
);
7426 free_percpu(zone
->pageset
);
7427 zone
->pageset
= &boot_pageset
;
7429 local_irq_restore(flags
);
7432 #ifdef CONFIG_MEMORY_HOTREMOVE
7434 * All pages in the range must be in a single zone and isolated
7435 * before calling this.
7438 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7442 unsigned int order
, i
;
7444 unsigned long flags
;
7445 /* find the first valid pfn */
7446 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7451 zone
= page_zone(pfn_to_page(pfn
));
7452 spin_lock_irqsave(&zone
->lock
, flags
);
7454 while (pfn
< end_pfn
) {
7455 if (!pfn_valid(pfn
)) {
7459 page
= pfn_to_page(pfn
);
7461 * The HWPoisoned page may be not in buddy system, and
7462 * page_count() is not 0.
7464 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7466 SetPageReserved(page
);
7470 BUG_ON(page_count(page
));
7471 BUG_ON(!PageBuddy(page
));
7472 order
= page_order(page
);
7473 #ifdef CONFIG_DEBUG_VM
7474 pr_info("remove from free list %lx %d %lx\n",
7475 pfn
, 1 << order
, end_pfn
);
7477 list_del(&page
->lru
);
7478 rmv_page_order(page
);
7479 zone
->free_area
[order
].nr_free
--;
7480 for (i
= 0; i
< (1 << order
); i
++)
7481 SetPageReserved((page
+i
));
7482 pfn
+= (1 << order
);
7484 spin_unlock_irqrestore(&zone
->lock
, flags
);
7488 bool is_free_buddy_page(struct page
*page
)
7490 struct zone
*zone
= page_zone(page
);
7491 unsigned long pfn
= page_to_pfn(page
);
7492 unsigned long flags
;
7495 spin_lock_irqsave(&zone
->lock
, flags
);
7496 for (order
= 0; order
< MAX_ORDER
; order
++) {
7497 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7499 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7502 spin_unlock_irqrestore(&zone
->lock
, flags
);
7504 return order
< MAX_ORDER
;