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 nr_memory_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())
610 if (!debug_guardpage_minorder())
616 static void init_debug_guardpage(void)
618 if (!debug_pagealloc_enabled())
621 if (!debug_guardpage_minorder())
624 _debug_guardpage_enabled
= true;
627 struct page_ext_operations debug_guardpage_ops
= {
628 .need
= need_debug_guardpage
,
629 .init
= init_debug_guardpage
,
632 static int __init
debug_guardpage_minorder_setup(char *buf
)
636 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
637 pr_err("Bad debug_guardpage_minorder value\n");
640 _debug_guardpage_minorder
= res
;
641 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
644 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
646 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
647 unsigned int order
, int migratetype
)
649 struct page_ext
*page_ext
;
651 if (!debug_guardpage_enabled())
654 if (order
>= debug_guardpage_minorder())
657 page_ext
= lookup_page_ext(page
);
658 if (unlikely(!page_ext
))
661 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
663 INIT_LIST_HEAD(&page
->lru
);
664 set_page_private(page
, order
);
665 /* Guard pages are not available for any usage */
666 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
671 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
672 unsigned int order
, int migratetype
)
674 struct page_ext
*page_ext
;
676 if (!debug_guardpage_enabled())
679 page_ext
= lookup_page_ext(page
);
680 if (unlikely(!page_ext
))
683 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
685 set_page_private(page
, 0);
686 if (!is_migrate_isolate(migratetype
))
687 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
690 struct page_ext_operations debug_guardpage_ops
;
691 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
692 unsigned int order
, int migratetype
) { return false; }
693 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
694 unsigned int order
, int migratetype
) {}
697 static inline void set_page_order(struct page
*page
, unsigned int order
)
699 set_page_private(page
, order
);
700 __SetPageBuddy(page
);
703 static inline void rmv_page_order(struct page
*page
)
705 __ClearPageBuddy(page
);
706 set_page_private(page
, 0);
710 * This function checks whether a page is free && is the buddy
711 * we can do coalesce a page and its buddy if
712 * (a) the buddy is not in a hole &&
713 * (b) the buddy is in the buddy system &&
714 * (c) a page and its buddy have the same order &&
715 * (d) a page and its buddy are in the same zone.
717 * For recording whether a page is in the buddy system, we set ->_mapcount
718 * PAGE_BUDDY_MAPCOUNT_VALUE.
719 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
720 * serialized by zone->lock.
722 * For recording page's order, we use page_private(page).
724 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
727 if (!pfn_valid_within(page_to_pfn(buddy
)))
730 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
731 if (page_zone_id(page
) != page_zone_id(buddy
))
734 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
739 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
741 * zone check is done late to avoid uselessly
742 * calculating zone/node ids for pages that could
745 if (page_zone_id(page
) != page_zone_id(buddy
))
748 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
756 * Freeing function for a buddy system allocator.
758 * The concept of a buddy system is to maintain direct-mapped table
759 * (containing bit values) for memory blocks of various "orders".
760 * The bottom level table contains the map for the smallest allocatable
761 * units of memory (here, pages), and each level above it describes
762 * pairs of units from the levels below, hence, "buddies".
763 * At a high level, all that happens here is marking the table entry
764 * at the bottom level available, and propagating the changes upward
765 * as necessary, plus some accounting needed to play nicely with other
766 * parts of the VM system.
767 * At each level, we keep a list of pages, which are heads of continuous
768 * free pages of length of (1 << order) and marked with _mapcount
769 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
771 * So when we are allocating or freeing one, we can derive the state of the
772 * other. That is, if we allocate a small block, and both were
773 * free, the remainder of the region must be split into blocks.
774 * If a block is freed, and its buddy is also free, then this
775 * triggers coalescing into a block of larger size.
780 static inline void __free_one_page(struct page
*page
,
782 struct zone
*zone
, unsigned int order
,
785 unsigned long page_idx
;
786 unsigned long combined_idx
;
787 unsigned long uninitialized_var(buddy_idx
);
789 unsigned int max_order
;
791 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
793 VM_BUG_ON(!zone_is_initialized(zone
));
794 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
796 VM_BUG_ON(migratetype
== -1);
797 if (likely(!is_migrate_isolate(migratetype
)))
798 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
800 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
802 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
803 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
806 while (order
< max_order
- 1) {
807 buddy_idx
= __find_buddy_index(page_idx
, order
);
808 buddy
= page
+ (buddy_idx
- page_idx
);
809 if (!page_is_buddy(page
, buddy
, order
))
812 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
813 * merge with it and move up one order.
815 if (page_is_guard(buddy
)) {
816 clear_page_guard(zone
, buddy
, order
, migratetype
);
818 list_del(&buddy
->lru
);
819 zone
->free_area
[order
].nr_free
--;
820 rmv_page_order(buddy
);
822 combined_idx
= buddy_idx
& page_idx
;
823 page
= page
+ (combined_idx
- page_idx
);
824 page_idx
= combined_idx
;
827 if (max_order
< MAX_ORDER
) {
828 /* If we are here, it means order is >= pageblock_order.
829 * We want to prevent merge between freepages on isolate
830 * pageblock and normal pageblock. Without this, pageblock
831 * isolation could cause incorrect freepage or CMA accounting.
833 * We don't want to hit this code for the more frequent
836 if (unlikely(has_isolate_pageblock(zone
))) {
839 buddy_idx
= __find_buddy_index(page_idx
, order
);
840 buddy
= page
+ (buddy_idx
- page_idx
);
841 buddy_mt
= get_pageblock_migratetype(buddy
);
843 if (migratetype
!= buddy_mt
844 && (is_migrate_isolate(migratetype
) ||
845 is_migrate_isolate(buddy_mt
)))
849 goto continue_merging
;
853 set_page_order(page
, order
);
856 * If this is not the largest possible page, check if the buddy
857 * of the next-highest order is free. If it is, it's possible
858 * that pages are being freed that will coalesce soon. In case,
859 * that is happening, add the free page to the tail of the list
860 * so it's less likely to be used soon and more likely to be merged
861 * as a higher order page
863 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
864 struct page
*higher_page
, *higher_buddy
;
865 combined_idx
= buddy_idx
& page_idx
;
866 higher_page
= page
+ (combined_idx
- page_idx
);
867 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
868 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
869 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
870 list_add_tail(&page
->lru
,
871 &zone
->free_area
[order
].free_list
[migratetype
]);
876 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
878 zone
->free_area
[order
].nr_free
++;
882 * A bad page could be due to a number of fields. Instead of multiple branches,
883 * try and check multiple fields with one check. The caller must do a detailed
884 * check if necessary.
886 static inline bool page_expected_state(struct page
*page
,
887 unsigned long check_flags
)
889 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
892 if (unlikely((unsigned long)page
->mapping
|
893 page_ref_count(page
) |
895 (unsigned long)page
->mem_cgroup
|
897 (page
->flags
& check_flags
)))
903 static void free_pages_check_bad(struct page
*page
)
905 const char *bad_reason
;
906 unsigned long bad_flags
;
911 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
912 bad_reason
= "nonzero mapcount";
913 if (unlikely(page
->mapping
!= NULL
))
914 bad_reason
= "non-NULL mapping";
915 if (unlikely(page_ref_count(page
) != 0))
916 bad_reason
= "nonzero _refcount";
917 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
918 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
919 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
922 if (unlikely(page
->mem_cgroup
))
923 bad_reason
= "page still charged to cgroup";
925 bad_page(page
, bad_reason
, bad_flags
);
928 static inline int free_pages_check(struct page
*page
)
930 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
933 /* Something has gone sideways, find it */
934 free_pages_check_bad(page
);
938 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
943 * We rely page->lru.next never has bit 0 set, unless the page
944 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
946 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
948 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
952 switch (page
- head_page
) {
954 /* the first tail page: ->mapping is compound_mapcount() */
955 if (unlikely(compound_mapcount(page
))) {
956 bad_page(page
, "nonzero compound_mapcount", 0);
962 * the second tail page: ->mapping is
963 * page_deferred_list().next -- ignore value.
967 if (page
->mapping
!= TAIL_MAPPING
) {
968 bad_page(page
, "corrupted mapping in tail page", 0);
973 if (unlikely(!PageTail(page
))) {
974 bad_page(page
, "PageTail not set", 0);
977 if (unlikely(compound_head(page
) != head_page
)) {
978 bad_page(page
, "compound_head not consistent", 0);
983 page
->mapping
= NULL
;
984 clear_compound_head(page
);
988 static __always_inline
bool free_pages_prepare(struct page
*page
,
989 unsigned int order
, bool check_free
)
993 VM_BUG_ON_PAGE(PageTail(page
), page
);
995 trace_mm_page_free(page
, order
);
996 kmemcheck_free_shadow(page
, order
);
999 * Check tail pages before head page information is cleared to
1000 * avoid checking PageCompound for order-0 pages.
1002 if (unlikely(order
)) {
1003 bool compound
= PageCompound(page
);
1006 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1009 ClearPageDoubleMap(page
);
1010 for (i
= 1; i
< (1 << order
); i
++) {
1012 bad
+= free_tail_pages_check(page
, page
+ i
);
1013 if (unlikely(free_pages_check(page
+ i
))) {
1017 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1020 if (PageMappingFlags(page
))
1021 page
->mapping
= NULL
;
1022 if (memcg_kmem_enabled() && PageKmemcg(page
))
1023 memcg_kmem_uncharge(page
, order
);
1025 bad
+= free_pages_check(page
);
1029 page_cpupid_reset_last(page
);
1030 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1031 reset_page_owner(page
, order
);
1033 if (!PageHighMem(page
)) {
1034 debug_check_no_locks_freed(page_address(page
),
1035 PAGE_SIZE
<< order
);
1036 debug_check_no_obj_freed(page_address(page
),
1037 PAGE_SIZE
<< order
);
1039 arch_free_page(page
, order
);
1040 kernel_poison_pages(page
, 1 << order
, 0);
1041 kernel_map_pages(page
, 1 << order
, 0);
1042 kasan_free_pages(page
, order
);
1047 #ifdef CONFIG_DEBUG_VM
1048 static inline bool free_pcp_prepare(struct page
*page
)
1050 return free_pages_prepare(page
, 0, true);
1053 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1058 static bool free_pcp_prepare(struct page
*page
)
1060 return free_pages_prepare(page
, 0, false);
1063 static bool bulkfree_pcp_prepare(struct page
*page
)
1065 return free_pages_check(page
);
1067 #endif /* CONFIG_DEBUG_VM */
1070 * Frees a number of pages from the PCP lists
1071 * Assumes all pages on list are in same zone, and of same order.
1072 * count is the number of pages to free.
1074 * If the zone was previously in an "all pages pinned" state then look to
1075 * see if this freeing clears that state.
1077 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1078 * pinned" detection logic.
1080 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1081 struct per_cpu_pages
*pcp
)
1083 int migratetype
= 0;
1085 unsigned long nr_scanned
;
1086 bool isolated_pageblocks
;
1088 spin_lock(&zone
->lock
);
1089 isolated_pageblocks
= has_isolate_pageblock(zone
);
1090 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1092 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1096 struct list_head
*list
;
1099 * Remove pages from lists in a round-robin fashion. A
1100 * batch_free count is maintained that is incremented when an
1101 * empty list is encountered. This is so more pages are freed
1102 * off fuller lists instead of spinning excessively around empty
1107 if (++migratetype
== MIGRATE_PCPTYPES
)
1109 list
= &pcp
->lists
[migratetype
];
1110 } while (list_empty(list
));
1112 /* This is the only non-empty list. Free them all. */
1113 if (batch_free
== MIGRATE_PCPTYPES
)
1117 int mt
; /* migratetype of the to-be-freed page */
1119 page
= list_last_entry(list
, struct page
, lru
);
1120 /* must delete as __free_one_page list manipulates */
1121 list_del(&page
->lru
);
1123 mt
= get_pcppage_migratetype(page
);
1124 /* MIGRATE_ISOLATE page should not go to pcplists */
1125 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1126 /* Pageblock could have been isolated meanwhile */
1127 if (unlikely(isolated_pageblocks
))
1128 mt
= get_pageblock_migratetype(page
);
1130 if (bulkfree_pcp_prepare(page
))
1133 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1134 trace_mm_page_pcpu_drain(page
, 0, mt
);
1135 } while (--count
&& --batch_free
&& !list_empty(list
));
1137 spin_unlock(&zone
->lock
);
1140 static void free_one_page(struct zone
*zone
,
1141 struct page
*page
, unsigned long pfn
,
1145 unsigned long nr_scanned
;
1146 spin_lock(&zone
->lock
);
1147 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1149 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1151 if (unlikely(has_isolate_pageblock(zone
) ||
1152 is_migrate_isolate(migratetype
))) {
1153 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1155 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1156 spin_unlock(&zone
->lock
);
1159 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1160 unsigned long zone
, int nid
)
1162 set_page_links(page
, zone
, nid
, pfn
);
1163 init_page_count(page
);
1164 page_mapcount_reset(page
);
1165 page_cpupid_reset_last(page
);
1167 INIT_LIST_HEAD(&page
->lru
);
1168 #ifdef WANT_PAGE_VIRTUAL
1169 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1170 if (!is_highmem_idx(zone
))
1171 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1175 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1178 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1181 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1182 static void init_reserved_page(unsigned long pfn
)
1187 if (!early_page_uninitialised(pfn
))
1190 nid
= early_pfn_to_nid(pfn
);
1191 pgdat
= NODE_DATA(nid
);
1193 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1194 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1196 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1199 __init_single_pfn(pfn
, zid
, nid
);
1202 static inline void init_reserved_page(unsigned long pfn
)
1205 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1208 * Initialised pages do not have PageReserved set. This function is
1209 * called for each range allocated by the bootmem allocator and
1210 * marks the pages PageReserved. The remaining valid pages are later
1211 * sent to the buddy page allocator.
1213 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1215 unsigned long start_pfn
= PFN_DOWN(start
);
1216 unsigned long end_pfn
= PFN_UP(end
);
1218 for (; start_pfn
< end_pfn
; start_pfn
++) {
1219 if (pfn_valid(start_pfn
)) {
1220 struct page
*page
= pfn_to_page(start_pfn
);
1222 init_reserved_page(start_pfn
);
1224 /* Avoid false-positive PageTail() */
1225 INIT_LIST_HEAD(&page
->lru
);
1227 SetPageReserved(page
);
1232 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1234 unsigned long flags
;
1236 unsigned long pfn
= page_to_pfn(page
);
1238 if (!free_pages_prepare(page
, order
, true))
1241 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1242 local_irq_save(flags
);
1243 __count_vm_events(PGFREE
, 1 << order
);
1244 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1245 local_irq_restore(flags
);
1248 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1250 unsigned int nr_pages
= 1 << order
;
1251 struct page
*p
= page
;
1255 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1257 __ClearPageReserved(p
);
1258 set_page_count(p
, 0);
1260 __ClearPageReserved(p
);
1261 set_page_count(p
, 0);
1263 page_zone(page
)->managed_pages
+= nr_pages
;
1264 set_page_refcounted(page
);
1265 __free_pages(page
, order
);
1268 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1269 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1271 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1273 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1275 static DEFINE_SPINLOCK(early_pfn_lock
);
1278 spin_lock(&early_pfn_lock
);
1279 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1281 nid
= first_online_node
;
1282 spin_unlock(&early_pfn_lock
);
1288 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1289 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1290 struct mminit_pfnnid_cache
*state
)
1294 nid
= __early_pfn_to_nid(pfn
, state
);
1295 if (nid
>= 0 && nid
!= node
)
1300 /* Only safe to use early in boot when initialisation is single-threaded */
1301 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1303 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1308 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1312 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1313 struct mminit_pfnnid_cache
*state
)
1320 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1323 if (early_page_uninitialised(pfn
))
1325 return __free_pages_boot_core(page
, order
);
1329 * Check that the whole (or subset of) a pageblock given by the interval of
1330 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1331 * with the migration of free compaction scanner. The scanners then need to
1332 * use only pfn_valid_within() check for arches that allow holes within
1335 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1337 * It's possible on some configurations to have a setup like node0 node1 node0
1338 * i.e. it's possible that all pages within a zones range of pages do not
1339 * belong to a single zone. We assume that a border between node0 and node1
1340 * can occur within a single pageblock, but not a node0 node1 node0
1341 * interleaving within a single pageblock. It is therefore sufficient to check
1342 * the first and last page of a pageblock and avoid checking each individual
1343 * page in a pageblock.
1345 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1346 unsigned long end_pfn
, struct zone
*zone
)
1348 struct page
*start_page
;
1349 struct page
*end_page
;
1351 /* end_pfn is one past the range we are checking */
1354 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1357 start_page
= pfn_to_page(start_pfn
);
1359 if (page_zone(start_page
) != zone
)
1362 end_page
= pfn_to_page(end_pfn
);
1364 /* This gives a shorter code than deriving page_zone(end_page) */
1365 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1371 void set_zone_contiguous(struct zone
*zone
)
1373 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1374 unsigned long block_end_pfn
;
1376 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1377 for (; block_start_pfn
< zone_end_pfn(zone
);
1378 block_start_pfn
= block_end_pfn
,
1379 block_end_pfn
+= pageblock_nr_pages
) {
1381 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1383 if (!__pageblock_pfn_to_page(block_start_pfn
,
1384 block_end_pfn
, zone
))
1388 /* We confirm that there is no hole */
1389 zone
->contiguous
= true;
1392 void clear_zone_contiguous(struct zone
*zone
)
1394 zone
->contiguous
= false;
1397 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1398 static void __init
deferred_free_range(struct page
*page
,
1399 unsigned long pfn
, int nr_pages
)
1406 /* Free a large naturally-aligned chunk if possible */
1407 if (nr_pages
== pageblock_nr_pages
&&
1408 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1409 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1410 __free_pages_boot_core(page
, pageblock_order
);
1414 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1415 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1416 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1417 __free_pages_boot_core(page
, 0);
1421 /* Completion tracking for deferred_init_memmap() threads */
1422 static atomic_t pgdat_init_n_undone __initdata
;
1423 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1425 static inline void __init
pgdat_init_report_one_done(void)
1427 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1428 complete(&pgdat_init_all_done_comp
);
1431 /* Initialise remaining memory on a node */
1432 static int __init
deferred_init_memmap(void *data
)
1434 pg_data_t
*pgdat
= data
;
1435 int nid
= pgdat
->node_id
;
1436 struct mminit_pfnnid_cache nid_init_state
= { };
1437 unsigned long start
= jiffies
;
1438 unsigned long nr_pages
= 0;
1439 unsigned long walk_start
, walk_end
;
1442 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1443 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1445 if (first_init_pfn
== ULONG_MAX
) {
1446 pgdat_init_report_one_done();
1450 /* Bind memory initialisation thread to a local node if possible */
1451 if (!cpumask_empty(cpumask
))
1452 set_cpus_allowed_ptr(current
, cpumask
);
1454 /* Sanity check boundaries */
1455 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1456 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1457 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1459 /* Only the highest zone is deferred so find it */
1460 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1461 zone
= pgdat
->node_zones
+ zid
;
1462 if (first_init_pfn
< zone_end_pfn(zone
))
1466 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1467 unsigned long pfn
, end_pfn
;
1468 struct page
*page
= NULL
;
1469 struct page
*free_base_page
= NULL
;
1470 unsigned long free_base_pfn
= 0;
1473 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1474 pfn
= first_init_pfn
;
1475 if (pfn
< walk_start
)
1477 if (pfn
< zone
->zone_start_pfn
)
1478 pfn
= zone
->zone_start_pfn
;
1480 for (; pfn
< end_pfn
; pfn
++) {
1481 if (!pfn_valid_within(pfn
))
1485 * Ensure pfn_valid is checked every
1486 * pageblock_nr_pages for memory holes
1488 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1489 if (!pfn_valid(pfn
)) {
1495 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1500 /* Minimise pfn page lookups and scheduler checks */
1501 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1504 nr_pages
+= nr_to_free
;
1505 deferred_free_range(free_base_page
,
1506 free_base_pfn
, nr_to_free
);
1507 free_base_page
= NULL
;
1508 free_base_pfn
= nr_to_free
= 0;
1510 page
= pfn_to_page(pfn
);
1515 VM_BUG_ON(page_zone(page
) != zone
);
1519 __init_single_page(page
, pfn
, zid
, nid
);
1520 if (!free_base_page
) {
1521 free_base_page
= page
;
1522 free_base_pfn
= pfn
;
1527 /* Where possible, batch up pages for a single free */
1530 /* Free the current block of pages to allocator */
1531 nr_pages
+= nr_to_free
;
1532 deferred_free_range(free_base_page
, free_base_pfn
,
1534 free_base_page
= NULL
;
1535 free_base_pfn
= nr_to_free
= 0;
1537 /* Free the last block of pages to allocator */
1538 nr_pages
+= nr_to_free
;
1539 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1541 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1544 /* Sanity check that the next zone really is unpopulated */
1545 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1547 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1548 jiffies_to_msecs(jiffies
- start
));
1550 pgdat_init_report_one_done();
1553 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1555 void __init
page_alloc_init_late(void)
1559 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1562 /* There will be num_node_state(N_MEMORY) threads */
1563 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1564 for_each_node_state(nid
, N_MEMORY
) {
1565 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1568 /* Block until all are initialised */
1569 wait_for_completion(&pgdat_init_all_done_comp
);
1571 /* Reinit limits that are based on free pages after the kernel is up */
1572 files_maxfiles_init();
1575 for_each_populated_zone(zone
)
1576 set_zone_contiguous(zone
);
1580 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1581 void __init
init_cma_reserved_pageblock(struct page
*page
)
1583 unsigned i
= pageblock_nr_pages
;
1584 struct page
*p
= page
;
1587 __ClearPageReserved(p
);
1588 set_page_count(p
, 0);
1591 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1593 if (pageblock_order
>= MAX_ORDER
) {
1594 i
= pageblock_nr_pages
;
1597 set_page_refcounted(p
);
1598 __free_pages(p
, MAX_ORDER
- 1);
1599 p
+= MAX_ORDER_NR_PAGES
;
1600 } while (i
-= MAX_ORDER_NR_PAGES
);
1602 set_page_refcounted(page
);
1603 __free_pages(page
, pageblock_order
);
1606 adjust_managed_page_count(page
, pageblock_nr_pages
);
1611 * The order of subdivision here is critical for the IO subsystem.
1612 * Please do not alter this order without good reasons and regression
1613 * testing. Specifically, as large blocks of memory are subdivided,
1614 * the order in which smaller blocks are delivered depends on the order
1615 * they're subdivided in this function. This is the primary factor
1616 * influencing the order in which pages are delivered to the IO
1617 * subsystem according to empirical testing, and this is also justified
1618 * by considering the behavior of a buddy system containing a single
1619 * large block of memory acted on by a series of small allocations.
1620 * This behavior is a critical factor in sglist merging's success.
1624 static inline void expand(struct zone
*zone
, struct page
*page
,
1625 int low
, int high
, struct free_area
*area
,
1628 unsigned long size
= 1 << high
;
1630 while (high
> low
) {
1634 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1637 * Mark as guard pages (or page), that will allow to
1638 * merge back to allocator when buddy will be freed.
1639 * Corresponding page table entries will not be touched,
1640 * pages will stay not present in virtual address space
1642 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1645 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1647 set_page_order(&page
[size
], high
);
1651 static void check_new_page_bad(struct page
*page
)
1653 const char *bad_reason
= NULL
;
1654 unsigned long bad_flags
= 0;
1656 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1657 bad_reason
= "nonzero mapcount";
1658 if (unlikely(page
->mapping
!= NULL
))
1659 bad_reason
= "non-NULL mapping";
1660 if (unlikely(page_ref_count(page
) != 0))
1661 bad_reason
= "nonzero _count";
1662 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1663 bad_reason
= "HWPoisoned (hardware-corrupted)";
1664 bad_flags
= __PG_HWPOISON
;
1665 /* Don't complain about hwpoisoned pages */
1666 page_mapcount_reset(page
); /* remove PageBuddy */
1669 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1670 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1671 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1674 if (unlikely(page
->mem_cgroup
))
1675 bad_reason
= "page still charged to cgroup";
1677 bad_page(page
, bad_reason
, bad_flags
);
1681 * This page is about to be returned from the page allocator
1683 static inline int check_new_page(struct page
*page
)
1685 if (likely(page_expected_state(page
,
1686 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1689 check_new_page_bad(page
);
1693 static inline bool free_pages_prezeroed(bool poisoned
)
1695 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1696 page_poisoning_enabled() && poisoned
;
1699 #ifdef CONFIG_DEBUG_VM
1700 static bool check_pcp_refill(struct page
*page
)
1705 static bool check_new_pcp(struct page
*page
)
1707 return check_new_page(page
);
1710 static bool check_pcp_refill(struct page
*page
)
1712 return check_new_page(page
);
1714 static bool check_new_pcp(struct page
*page
)
1718 #endif /* CONFIG_DEBUG_VM */
1720 static bool check_new_pages(struct page
*page
, unsigned int order
)
1723 for (i
= 0; i
< (1 << order
); i
++) {
1724 struct page
*p
= page
+ i
;
1726 if (unlikely(check_new_page(p
)))
1733 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1736 set_page_private(page
, 0);
1737 set_page_refcounted(page
);
1739 arch_alloc_page(page
, order
);
1740 kernel_map_pages(page
, 1 << order
, 1);
1741 kernel_poison_pages(page
, 1 << order
, 1);
1742 kasan_alloc_pages(page
, order
);
1743 set_page_owner(page
, order
, gfp_flags
);
1746 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1747 unsigned int alloc_flags
)
1750 bool poisoned
= true;
1752 for (i
= 0; i
< (1 << order
); i
++) {
1753 struct page
*p
= page
+ i
;
1755 poisoned
&= page_is_poisoned(p
);
1758 post_alloc_hook(page
, order
, gfp_flags
);
1760 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1761 for (i
= 0; i
< (1 << order
); i
++)
1762 clear_highpage(page
+ i
);
1764 if (order
&& (gfp_flags
& __GFP_COMP
))
1765 prep_compound_page(page
, order
);
1768 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1769 * allocate the page. The expectation is that the caller is taking
1770 * steps that will free more memory. The caller should avoid the page
1771 * being used for !PFMEMALLOC purposes.
1773 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1774 set_page_pfmemalloc(page
);
1776 clear_page_pfmemalloc(page
);
1780 * Go through the free lists for the given migratetype and remove
1781 * the smallest available page from the freelists
1784 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1787 unsigned int current_order
;
1788 struct free_area
*area
;
1791 /* Find a page of the appropriate size in the preferred list */
1792 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1793 area
= &(zone
->free_area
[current_order
]);
1794 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1798 list_del(&page
->lru
);
1799 rmv_page_order(page
);
1801 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1802 set_pcppage_migratetype(page
, migratetype
);
1811 * This array describes the order lists are fallen back to when
1812 * the free lists for the desirable migrate type are depleted
1814 static int fallbacks
[MIGRATE_TYPES
][4] = {
1815 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1816 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1817 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1819 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1821 #ifdef CONFIG_MEMORY_ISOLATION
1822 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1827 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1830 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1833 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1834 unsigned int order
) { return NULL
; }
1838 * Move the free pages in a range to the free lists of the requested type.
1839 * Note that start_page and end_pages are not aligned on a pageblock
1840 * boundary. If alignment is required, use move_freepages_block()
1842 int move_freepages(struct zone
*zone
,
1843 struct page
*start_page
, struct page
*end_page
,
1848 int pages_moved
= 0;
1850 #ifndef CONFIG_HOLES_IN_ZONE
1852 * page_zone is not safe to call in this context when
1853 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1854 * anyway as we check zone boundaries in move_freepages_block().
1855 * Remove at a later date when no bug reports exist related to
1856 * grouping pages by mobility
1858 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1861 for (page
= start_page
; page
<= end_page
;) {
1862 /* Make sure we are not inadvertently changing nodes */
1863 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1865 if (!pfn_valid_within(page_to_pfn(page
))) {
1870 if (!PageBuddy(page
)) {
1875 order
= page_order(page
);
1876 list_move(&page
->lru
,
1877 &zone
->free_area
[order
].free_list
[migratetype
]);
1879 pages_moved
+= 1 << order
;
1885 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1888 unsigned long start_pfn
, end_pfn
;
1889 struct page
*start_page
, *end_page
;
1891 start_pfn
= page_to_pfn(page
);
1892 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1893 start_page
= pfn_to_page(start_pfn
);
1894 end_page
= start_page
+ pageblock_nr_pages
- 1;
1895 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1897 /* Do not cross zone boundaries */
1898 if (!zone_spans_pfn(zone
, start_pfn
))
1900 if (!zone_spans_pfn(zone
, end_pfn
))
1903 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1906 static void change_pageblock_range(struct page
*pageblock_page
,
1907 int start_order
, int migratetype
)
1909 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1911 while (nr_pageblocks
--) {
1912 set_pageblock_migratetype(pageblock_page
, migratetype
);
1913 pageblock_page
+= pageblock_nr_pages
;
1918 * When we are falling back to another migratetype during allocation, try to
1919 * steal extra free pages from the same pageblocks to satisfy further
1920 * allocations, instead of polluting multiple pageblocks.
1922 * If we are stealing a relatively large buddy page, it is likely there will
1923 * be more free pages in the pageblock, so try to steal them all. For
1924 * reclaimable and unmovable allocations, we steal regardless of page size,
1925 * as fragmentation caused by those allocations polluting movable pageblocks
1926 * is worse than movable allocations stealing from unmovable and reclaimable
1929 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1932 * Leaving this order check is intended, although there is
1933 * relaxed order check in next check. The reason is that
1934 * we can actually steal whole pageblock if this condition met,
1935 * but, below check doesn't guarantee it and that is just heuristic
1936 * so could be changed anytime.
1938 if (order
>= pageblock_order
)
1941 if (order
>= pageblock_order
/ 2 ||
1942 start_mt
== MIGRATE_RECLAIMABLE
||
1943 start_mt
== MIGRATE_UNMOVABLE
||
1944 page_group_by_mobility_disabled
)
1951 * This function implements actual steal behaviour. If order is large enough,
1952 * we can steal whole pageblock. If not, we first move freepages in this
1953 * pageblock and check whether half of pages are moved or not. If half of
1954 * pages are moved, we can change migratetype of pageblock and permanently
1955 * use it's pages as requested migratetype in the future.
1957 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1960 unsigned int current_order
= page_order(page
);
1963 /* Take ownership for orders >= pageblock_order */
1964 if (current_order
>= pageblock_order
) {
1965 change_pageblock_range(page
, current_order
, start_type
);
1969 pages
= move_freepages_block(zone
, page
, start_type
);
1971 /* Claim the whole block if over half of it is free */
1972 if (pages
>= (1 << (pageblock_order
-1)) ||
1973 page_group_by_mobility_disabled
)
1974 set_pageblock_migratetype(page
, start_type
);
1978 * Check whether there is a suitable fallback freepage with requested order.
1979 * If only_stealable is true, this function returns fallback_mt only if
1980 * we can steal other freepages all together. This would help to reduce
1981 * fragmentation due to mixed migratetype pages in one pageblock.
1983 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1984 int migratetype
, bool only_stealable
, bool *can_steal
)
1989 if (area
->nr_free
== 0)
1994 fallback_mt
= fallbacks
[migratetype
][i
];
1995 if (fallback_mt
== MIGRATE_TYPES
)
1998 if (list_empty(&area
->free_list
[fallback_mt
]))
2001 if (can_steal_fallback(order
, migratetype
))
2004 if (!only_stealable
)
2015 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2016 * there are no empty page blocks that contain a page with a suitable order
2018 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2019 unsigned int alloc_order
)
2022 unsigned long max_managed
, flags
;
2025 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2026 * Check is race-prone but harmless.
2028 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2029 if (zone
->nr_reserved_highatomic
>= max_managed
)
2032 spin_lock_irqsave(&zone
->lock
, flags
);
2034 /* Recheck the nr_reserved_highatomic limit under the lock */
2035 if (zone
->nr_reserved_highatomic
>= max_managed
)
2039 mt
= get_pageblock_migratetype(page
);
2040 if (mt
!= MIGRATE_HIGHATOMIC
&&
2041 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
2042 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2043 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2044 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
2048 spin_unlock_irqrestore(&zone
->lock
, flags
);
2052 * Used when an allocation is about to fail under memory pressure. This
2053 * potentially hurts the reliability of high-order allocations when under
2054 * intense memory pressure but failed atomic allocations should be easier
2055 * to recover from than an OOM.
2057 static void unreserve_highatomic_pageblock(const struct alloc_context
*ac
)
2059 struct zonelist
*zonelist
= ac
->zonelist
;
2060 unsigned long flags
;
2066 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2068 /* Preserve at least one pageblock */
2069 if (zone
->nr_reserved_highatomic
<= pageblock_nr_pages
)
2072 spin_lock_irqsave(&zone
->lock
, flags
);
2073 for (order
= 0; order
< MAX_ORDER
; order
++) {
2074 struct free_area
*area
= &(zone
->free_area
[order
]);
2076 page
= list_first_entry_or_null(
2077 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2083 * It should never happen but changes to locking could
2084 * inadvertently allow a per-cpu drain to add pages
2085 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2086 * and watch for underflows.
2088 zone
->nr_reserved_highatomic
-= min(pageblock_nr_pages
,
2089 zone
->nr_reserved_highatomic
);
2092 * Convert to ac->migratetype and avoid the normal
2093 * pageblock stealing heuristics. Minimally, the caller
2094 * is doing the work and needs the pages. More
2095 * importantly, if the block was always converted to
2096 * MIGRATE_UNMOVABLE or another type then the number
2097 * of pageblocks that cannot be completely freed
2100 set_pageblock_migratetype(page
, ac
->migratetype
);
2101 move_freepages_block(zone
, page
, ac
->migratetype
);
2102 spin_unlock_irqrestore(&zone
->lock
, flags
);
2105 spin_unlock_irqrestore(&zone
->lock
, flags
);
2109 /* Remove an element from the buddy allocator from the fallback list */
2110 static inline struct page
*
2111 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2113 struct free_area
*area
;
2114 unsigned int current_order
;
2119 /* Find the largest possible block of pages in the other list */
2120 for (current_order
= MAX_ORDER
-1;
2121 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2123 area
= &(zone
->free_area
[current_order
]);
2124 fallback_mt
= find_suitable_fallback(area
, current_order
,
2125 start_migratetype
, false, &can_steal
);
2126 if (fallback_mt
== -1)
2129 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2132 steal_suitable_fallback(zone
, page
, start_migratetype
);
2134 /* Remove the page from the freelists */
2136 list_del(&page
->lru
);
2137 rmv_page_order(page
);
2139 expand(zone
, page
, order
, current_order
, area
,
2142 * The pcppage_migratetype may differ from pageblock's
2143 * migratetype depending on the decisions in
2144 * find_suitable_fallback(). This is OK as long as it does not
2145 * differ for MIGRATE_CMA pageblocks. Those can be used as
2146 * fallback only via special __rmqueue_cma_fallback() function
2148 set_pcppage_migratetype(page
, start_migratetype
);
2150 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2151 start_migratetype
, fallback_mt
);
2160 * Do the hard work of removing an element from the buddy allocator.
2161 * Call me with the zone->lock already held.
2163 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2168 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2169 if (unlikely(!page
)) {
2170 if (migratetype
== MIGRATE_MOVABLE
)
2171 page
= __rmqueue_cma_fallback(zone
, order
);
2174 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2177 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2182 * Obtain a specified number of elements from the buddy allocator, all under
2183 * a single hold of the lock, for efficiency. Add them to the supplied list.
2184 * Returns the number of new pages which were placed at *list.
2186 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2187 unsigned long count
, struct list_head
*list
,
2188 int migratetype
, bool cold
)
2192 spin_lock(&zone
->lock
);
2193 for (i
= 0; i
< count
; ++i
) {
2194 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2195 if (unlikely(page
== NULL
))
2198 if (unlikely(check_pcp_refill(page
)))
2202 * Split buddy pages returned by expand() are received here
2203 * in physical page order. The page is added to the callers and
2204 * list and the list head then moves forward. From the callers
2205 * perspective, the linked list is ordered by page number in
2206 * some conditions. This is useful for IO devices that can
2207 * merge IO requests if the physical pages are ordered
2211 list_add(&page
->lru
, list
);
2213 list_add_tail(&page
->lru
, list
);
2215 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2216 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2219 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2220 spin_unlock(&zone
->lock
);
2226 * Called from the vmstat counter updater to drain pagesets of this
2227 * currently executing processor on remote nodes after they have
2230 * Note that this function must be called with the thread pinned to
2231 * a single processor.
2233 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2235 unsigned long flags
;
2236 int to_drain
, batch
;
2238 local_irq_save(flags
);
2239 batch
= READ_ONCE(pcp
->batch
);
2240 to_drain
= min(pcp
->count
, batch
);
2242 free_pcppages_bulk(zone
, to_drain
, pcp
);
2243 pcp
->count
-= to_drain
;
2245 local_irq_restore(flags
);
2250 * Drain pcplists of the indicated processor and zone.
2252 * The processor must either be the current processor and the
2253 * thread pinned to the current processor or a processor that
2256 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2258 unsigned long flags
;
2259 struct per_cpu_pageset
*pset
;
2260 struct per_cpu_pages
*pcp
;
2262 local_irq_save(flags
);
2263 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2267 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2270 local_irq_restore(flags
);
2274 * Drain pcplists of all zones on the indicated processor.
2276 * The processor must either be the current processor and the
2277 * thread pinned to the current processor or a processor that
2280 static void drain_pages(unsigned int cpu
)
2284 for_each_populated_zone(zone
) {
2285 drain_pages_zone(cpu
, zone
);
2290 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2292 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2293 * the single zone's pages.
2295 void drain_local_pages(struct zone
*zone
)
2297 int cpu
= smp_processor_id();
2300 drain_pages_zone(cpu
, zone
);
2306 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2308 * When zone parameter is non-NULL, spill just the single zone's pages.
2310 * Note that this code is protected against sending an IPI to an offline
2311 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2312 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2313 * nothing keeps CPUs from showing up after we populated the cpumask and
2314 * before the call to on_each_cpu_mask().
2316 void drain_all_pages(struct zone
*zone
)
2321 * Allocate in the BSS so we wont require allocation in
2322 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2324 static cpumask_t cpus_with_pcps
;
2327 * We don't care about racing with CPU hotplug event
2328 * as offline notification will cause the notified
2329 * cpu to drain that CPU pcps and on_each_cpu_mask
2330 * disables preemption as part of its processing
2332 for_each_online_cpu(cpu
) {
2333 struct per_cpu_pageset
*pcp
;
2335 bool has_pcps
= false;
2338 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2342 for_each_populated_zone(z
) {
2343 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2344 if (pcp
->pcp
.count
) {
2352 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2354 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2356 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
2360 #ifdef CONFIG_HIBERNATION
2362 void mark_free_pages(struct zone
*zone
)
2364 unsigned long pfn
, max_zone_pfn
;
2365 unsigned long flags
;
2366 unsigned int order
, t
;
2369 if (zone_is_empty(zone
))
2372 spin_lock_irqsave(&zone
->lock
, flags
);
2374 max_zone_pfn
= zone_end_pfn(zone
);
2375 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2376 if (pfn_valid(pfn
)) {
2377 page
= pfn_to_page(pfn
);
2379 if (page_zone(page
) != zone
)
2382 if (!swsusp_page_is_forbidden(page
))
2383 swsusp_unset_page_free(page
);
2386 for_each_migratetype_order(order
, t
) {
2387 list_for_each_entry(page
,
2388 &zone
->free_area
[order
].free_list
[t
], lru
) {
2391 pfn
= page_to_pfn(page
);
2392 for (i
= 0; i
< (1UL << order
); i
++)
2393 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2396 spin_unlock_irqrestore(&zone
->lock
, flags
);
2398 #endif /* CONFIG_PM */
2401 * Free a 0-order page
2402 * cold == true ? free a cold page : free a hot page
2404 void free_hot_cold_page(struct page
*page
, bool cold
)
2406 struct zone
*zone
= page_zone(page
);
2407 struct per_cpu_pages
*pcp
;
2408 unsigned long flags
;
2409 unsigned long pfn
= page_to_pfn(page
);
2412 if (!free_pcp_prepare(page
))
2415 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2416 set_pcppage_migratetype(page
, migratetype
);
2417 local_irq_save(flags
);
2418 __count_vm_event(PGFREE
);
2421 * We only track unmovable, reclaimable and movable on pcp lists.
2422 * Free ISOLATE pages back to the allocator because they are being
2423 * offlined but treat RESERVE as movable pages so we can get those
2424 * areas back if necessary. Otherwise, we may have to free
2425 * excessively into the page allocator
2427 if (migratetype
>= MIGRATE_PCPTYPES
) {
2428 if (unlikely(is_migrate_isolate(migratetype
))) {
2429 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2432 migratetype
= MIGRATE_MOVABLE
;
2435 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2437 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2439 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2441 if (pcp
->count
>= pcp
->high
) {
2442 unsigned long batch
= READ_ONCE(pcp
->batch
);
2443 free_pcppages_bulk(zone
, batch
, pcp
);
2444 pcp
->count
-= batch
;
2448 local_irq_restore(flags
);
2452 * Free a list of 0-order pages
2454 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2456 struct page
*page
, *next
;
2458 list_for_each_entry_safe(page
, next
, list
, lru
) {
2459 trace_mm_page_free_batched(page
, cold
);
2460 free_hot_cold_page(page
, cold
);
2465 * split_page takes a non-compound higher-order page, and splits it into
2466 * n (1<<order) sub-pages: page[0..n]
2467 * Each sub-page must be freed individually.
2469 * Note: this is probably too low level an operation for use in drivers.
2470 * Please consult with lkml before using this in your driver.
2472 void split_page(struct page
*page
, unsigned int order
)
2476 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2477 VM_BUG_ON_PAGE(!page_count(page
), page
);
2479 #ifdef CONFIG_KMEMCHECK
2481 * Split shadow pages too, because free(page[0]) would
2482 * otherwise free the whole shadow.
2484 if (kmemcheck_page_is_tracked(page
))
2485 split_page(virt_to_page(page
[0].shadow
), order
);
2488 for (i
= 1; i
< (1 << order
); i
++)
2489 set_page_refcounted(page
+ i
);
2490 split_page_owner(page
, order
);
2492 EXPORT_SYMBOL_GPL(split_page
);
2494 int __isolate_free_page(struct page
*page
, unsigned int order
)
2496 unsigned long watermark
;
2500 BUG_ON(!PageBuddy(page
));
2502 zone
= page_zone(page
);
2503 mt
= get_pageblock_migratetype(page
);
2505 if (!is_migrate_isolate(mt
)) {
2507 * Obey watermarks as if the page was being allocated. We can
2508 * emulate a high-order watermark check with a raised order-0
2509 * watermark, because we already know our high-order page
2512 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2513 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2516 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2519 /* Remove page from free list */
2520 list_del(&page
->lru
);
2521 zone
->free_area
[order
].nr_free
--;
2522 rmv_page_order(page
);
2525 * Set the pageblock if the isolated page is at least half of a
2528 if (order
>= pageblock_order
- 1) {
2529 struct page
*endpage
= page
+ (1 << order
) - 1;
2530 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2531 int mt
= get_pageblock_migratetype(page
);
2532 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2533 set_pageblock_migratetype(page
,
2539 return 1UL << order
;
2543 * Update NUMA hit/miss statistics
2545 * Must be called with interrupts disabled.
2547 * When __GFP_OTHER_NODE is set assume the node of the preferred
2548 * zone is the local node. This is useful for daemons who allocate
2549 * memory on behalf of other processes.
2551 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2555 int local_nid
= numa_node_id();
2556 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2558 if (unlikely(flags
& __GFP_OTHER_NODE
)) {
2559 local_stat
= NUMA_OTHER
;
2560 local_nid
= preferred_zone
->node
;
2563 if (z
->node
== local_nid
) {
2564 __inc_zone_state(z
, NUMA_HIT
);
2565 __inc_zone_state(z
, local_stat
);
2567 __inc_zone_state(z
, NUMA_MISS
);
2568 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2574 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2577 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2578 struct zone
*zone
, unsigned int order
,
2579 gfp_t gfp_flags
, unsigned int alloc_flags
,
2582 unsigned long flags
;
2584 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2586 if (likely(order
== 0)) {
2587 struct per_cpu_pages
*pcp
;
2588 struct list_head
*list
;
2590 local_irq_save(flags
);
2592 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2593 list
= &pcp
->lists
[migratetype
];
2594 if (list_empty(list
)) {
2595 pcp
->count
+= rmqueue_bulk(zone
, 0,
2598 if (unlikely(list_empty(list
)))
2603 page
= list_last_entry(list
, struct page
, lru
);
2605 page
= list_first_entry(list
, struct page
, lru
);
2607 list_del(&page
->lru
);
2610 } while (check_new_pcp(page
));
2613 * We most definitely don't want callers attempting to
2614 * allocate greater than order-1 page units with __GFP_NOFAIL.
2616 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2617 spin_lock_irqsave(&zone
->lock
, flags
);
2621 if (alloc_flags
& ALLOC_HARDER
) {
2622 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2624 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2627 page
= __rmqueue(zone
, order
, migratetype
);
2628 } while (page
&& check_new_pages(page
, order
));
2629 spin_unlock(&zone
->lock
);
2632 __mod_zone_freepage_state(zone
, -(1 << order
),
2633 get_pcppage_migratetype(page
));
2636 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2637 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2638 local_irq_restore(flags
);
2640 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2644 local_irq_restore(flags
);
2648 #ifdef CONFIG_FAIL_PAGE_ALLOC
2651 struct fault_attr attr
;
2653 bool ignore_gfp_highmem
;
2654 bool ignore_gfp_reclaim
;
2656 } fail_page_alloc
= {
2657 .attr
= FAULT_ATTR_INITIALIZER
,
2658 .ignore_gfp_reclaim
= true,
2659 .ignore_gfp_highmem
= true,
2663 static int __init
setup_fail_page_alloc(char *str
)
2665 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2667 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2669 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2671 if (order
< fail_page_alloc
.min_order
)
2673 if (gfp_mask
& __GFP_NOFAIL
)
2675 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2677 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2678 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2681 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2684 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2686 static int __init
fail_page_alloc_debugfs(void)
2688 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2691 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2692 &fail_page_alloc
.attr
);
2694 return PTR_ERR(dir
);
2696 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2697 &fail_page_alloc
.ignore_gfp_reclaim
))
2699 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2700 &fail_page_alloc
.ignore_gfp_highmem
))
2702 if (!debugfs_create_u32("min-order", mode
, dir
,
2703 &fail_page_alloc
.min_order
))
2708 debugfs_remove_recursive(dir
);
2713 late_initcall(fail_page_alloc_debugfs
);
2715 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2717 #else /* CONFIG_FAIL_PAGE_ALLOC */
2719 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2724 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2727 * Return true if free base pages are above 'mark'. For high-order checks it
2728 * will return true of the order-0 watermark is reached and there is at least
2729 * one free page of a suitable size. Checking now avoids taking the zone lock
2730 * to check in the allocation paths if no pages are free.
2732 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2733 int classzone_idx
, unsigned int alloc_flags
,
2738 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2740 /* free_pages may go negative - that's OK */
2741 free_pages
-= (1 << order
) - 1;
2743 if (alloc_flags
& ALLOC_HIGH
)
2747 * If the caller does not have rights to ALLOC_HARDER then subtract
2748 * the high-atomic reserves. This will over-estimate the size of the
2749 * atomic reserve but it avoids a search.
2751 if (likely(!alloc_harder
))
2752 free_pages
-= z
->nr_reserved_highatomic
;
2757 /* If allocation can't use CMA areas don't use free CMA pages */
2758 if (!(alloc_flags
& ALLOC_CMA
))
2759 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2763 * Check watermarks for an order-0 allocation request. If these
2764 * are not met, then a high-order request also cannot go ahead
2765 * even if a suitable page happened to be free.
2767 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2770 /* If this is an order-0 request then the watermark is fine */
2774 /* For a high-order request, check at least one suitable page is free */
2775 for (o
= order
; o
< MAX_ORDER
; o
++) {
2776 struct free_area
*area
= &z
->free_area
[o
];
2785 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2786 if (!list_empty(&area
->free_list
[mt
]))
2791 if ((alloc_flags
& ALLOC_CMA
) &&
2792 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2800 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2801 int classzone_idx
, unsigned int alloc_flags
)
2803 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2804 zone_page_state(z
, NR_FREE_PAGES
));
2807 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2808 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2810 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2814 /* If allocation can't use CMA areas don't use free CMA pages */
2815 if (!(alloc_flags
& ALLOC_CMA
))
2816 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2820 * Fast check for order-0 only. If this fails then the reserves
2821 * need to be calculated. There is a corner case where the check
2822 * passes but only the high-order atomic reserve are free. If
2823 * the caller is !atomic then it'll uselessly search the free
2824 * list. That corner case is then slower but it is harmless.
2826 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2829 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2833 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2834 unsigned long mark
, int classzone_idx
)
2836 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2838 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2839 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2841 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2846 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2848 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2851 #else /* CONFIG_NUMA */
2852 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2856 #endif /* CONFIG_NUMA */
2859 * get_page_from_freelist goes through the zonelist trying to allocate
2862 static struct page
*
2863 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2864 const struct alloc_context
*ac
)
2866 struct zoneref
*z
= ac
->preferred_zoneref
;
2868 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
2871 * Scan zonelist, looking for a zone with enough free.
2872 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2874 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2879 if (cpusets_enabled() &&
2880 (alloc_flags
& ALLOC_CPUSET
) &&
2881 !__cpuset_zone_allowed(zone
, gfp_mask
))
2884 * When allocating a page cache page for writing, we
2885 * want to get it from a node that is within its dirty
2886 * limit, such that no single node holds more than its
2887 * proportional share of globally allowed dirty pages.
2888 * The dirty limits take into account the node's
2889 * lowmem reserves and high watermark so that kswapd
2890 * should be able to balance it without having to
2891 * write pages from its LRU list.
2893 * XXX: For now, allow allocations to potentially
2894 * exceed the per-node dirty limit in the slowpath
2895 * (spread_dirty_pages unset) before going into reclaim,
2896 * which is important when on a NUMA setup the allowed
2897 * nodes are together not big enough to reach the
2898 * global limit. The proper fix for these situations
2899 * will require awareness of nodes in the
2900 * dirty-throttling and the flusher threads.
2902 if (ac
->spread_dirty_pages
) {
2903 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
2906 if (!node_dirty_ok(zone
->zone_pgdat
)) {
2907 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
2912 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2913 if (!zone_watermark_fast(zone
, order
, mark
,
2914 ac_classzone_idx(ac
), alloc_flags
)) {
2917 /* Checked here to keep the fast path fast */
2918 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2919 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2922 if (node_reclaim_mode
== 0 ||
2923 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
2926 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
2928 case NODE_RECLAIM_NOSCAN
:
2931 case NODE_RECLAIM_FULL
:
2932 /* scanned but unreclaimable */
2935 /* did we reclaim enough */
2936 if (zone_watermark_ok(zone
, order
, mark
,
2937 ac_classzone_idx(ac
), alloc_flags
))
2945 page
= buffered_rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
2946 gfp_mask
, alloc_flags
, ac
->migratetype
);
2948 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
2951 * If this is a high-order atomic allocation then check
2952 * if the pageblock should be reserved for the future
2954 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
2955 reserve_highatomic_pageblock(page
, zone
, order
);
2965 * Large machines with many possible nodes should not always dump per-node
2966 * meminfo in irq context.
2968 static inline bool should_suppress_show_mem(void)
2973 ret
= in_interrupt();
2978 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2979 DEFAULT_RATELIMIT_INTERVAL
,
2980 DEFAULT_RATELIMIT_BURST
);
2982 void warn_alloc_failed(gfp_t gfp_mask
, unsigned int order
, const char *fmt
, ...)
2984 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2986 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2987 debug_guardpage_minorder() > 0)
2991 * This documents exceptions given to allocations in certain
2992 * contexts that are allowed to allocate outside current's set
2995 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2996 if (test_thread_flag(TIF_MEMDIE
) ||
2997 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2998 filter
&= ~SHOW_MEM_FILTER_NODES
;
2999 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3000 filter
&= ~SHOW_MEM_FILTER_NODES
;
3003 struct va_format vaf
;
3006 va_start(args
, fmt
);
3011 pr_warn("%pV", &vaf
);
3016 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3017 current
->comm
, order
, gfp_mask
, &gfp_mask
);
3019 if (!should_suppress_show_mem())
3023 static inline struct page
*
3024 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3025 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3027 struct oom_control oc
= {
3028 .zonelist
= ac
->zonelist
,
3029 .nodemask
= ac
->nodemask
,
3031 .gfp_mask
= gfp_mask
,
3036 *did_some_progress
= 0;
3039 * Acquire the oom lock. If that fails, somebody else is
3040 * making progress for us.
3042 if (!mutex_trylock(&oom_lock
)) {
3043 *did_some_progress
= 1;
3044 schedule_timeout_uninterruptible(1);
3049 * Go through the zonelist yet one more time, keep very high watermark
3050 * here, this is only to catch a parallel oom killing, we must fail if
3051 * we're still under heavy pressure.
3053 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3054 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3058 if (!(gfp_mask
& __GFP_NOFAIL
)) {
3059 /* Coredumps can quickly deplete all memory reserves */
3060 if (current
->flags
& PF_DUMPCORE
)
3062 /* The OOM killer will not help higher order allocs */
3063 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3065 /* The OOM killer does not needlessly kill tasks for lowmem */
3066 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3068 if (pm_suspended_storage())
3071 * XXX: GFP_NOFS allocations should rather fail than rely on
3072 * other request to make a forward progress.
3073 * We are in an unfortunate situation where out_of_memory cannot
3074 * do much for this context but let's try it to at least get
3075 * access to memory reserved if the current task is killed (see
3076 * out_of_memory). Once filesystems are ready to handle allocation
3077 * failures more gracefully we should just bail out here.
3080 /* The OOM killer may not free memory on a specific node */
3081 if (gfp_mask
& __GFP_THISNODE
)
3084 /* Exhausted what can be done so it's blamo time */
3085 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3086 *did_some_progress
= 1;
3088 if (gfp_mask
& __GFP_NOFAIL
) {
3089 page
= get_page_from_freelist(gfp_mask
, order
,
3090 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
3092 * fallback to ignore cpuset restriction if our nodes
3096 page
= get_page_from_freelist(gfp_mask
, order
,
3097 ALLOC_NO_WATERMARKS
, ac
);
3101 mutex_unlock(&oom_lock
);
3106 * Maximum number of compaction retries wit a progress before OOM
3107 * killer is consider as the only way to move forward.
3109 #define MAX_COMPACT_RETRIES 16
3111 #ifdef CONFIG_COMPACTION
3112 /* Try memory compaction for high-order allocations before reclaim */
3113 static struct page
*
3114 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3115 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3116 enum compact_priority prio
, enum compact_result
*compact_result
)
3123 current
->flags
|= PF_MEMALLOC
;
3124 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3126 current
->flags
&= ~PF_MEMALLOC
;
3128 if (*compact_result
<= COMPACT_INACTIVE
)
3132 * At least in one zone compaction wasn't deferred or skipped, so let's
3133 * count a compaction stall
3135 count_vm_event(COMPACTSTALL
);
3137 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3140 struct zone
*zone
= page_zone(page
);
3142 zone
->compact_blockskip_flush
= false;
3143 compaction_defer_reset(zone
, order
, true);
3144 count_vm_event(COMPACTSUCCESS
);
3149 * It's bad if compaction run occurs and fails. The most likely reason
3150 * is that pages exist, but not enough to satisfy watermarks.
3152 count_vm_event(COMPACTFAIL
);
3160 static inline struct page
*
3161 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3162 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3163 enum compact_priority prio
, enum compact_result
*compact_result
)
3165 *compact_result
= COMPACT_SKIPPED
;
3169 #endif /* CONFIG_COMPACTION */
3172 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3173 enum compact_result compact_result
,
3174 enum compact_priority
*compact_priority
,
3175 int compaction_retries
)
3180 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3184 * There are setups with compaction disabled which would prefer to loop
3185 * inside the allocator rather than hit the oom killer prematurely.
3186 * Let's give them a good hope and keep retrying while the order-0
3187 * watermarks are OK.
3189 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3191 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3192 ac_classzone_idx(ac
), alloc_flags
))
3198 /* Perform direct synchronous page reclaim */
3200 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3201 const struct alloc_context
*ac
)
3203 struct reclaim_state reclaim_state
;
3208 /* We now go into synchronous reclaim */
3209 cpuset_memory_pressure_bump();
3210 current
->flags
|= PF_MEMALLOC
;
3211 lockdep_set_current_reclaim_state(gfp_mask
);
3212 reclaim_state
.reclaimed_slab
= 0;
3213 current
->reclaim_state
= &reclaim_state
;
3215 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3218 current
->reclaim_state
= NULL
;
3219 lockdep_clear_current_reclaim_state();
3220 current
->flags
&= ~PF_MEMALLOC
;
3227 /* The really slow allocator path where we enter direct reclaim */
3228 static inline struct page
*
3229 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3230 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3231 unsigned long *did_some_progress
)
3233 struct page
*page
= NULL
;
3234 bool drained
= false;
3236 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3237 if (unlikely(!(*did_some_progress
)))
3241 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3244 * If an allocation failed after direct reclaim, it could be because
3245 * pages are pinned on the per-cpu lists or in high alloc reserves.
3246 * Shrink them them and try again
3248 if (!page
&& !drained
) {
3249 unreserve_highatomic_pageblock(ac
);
3250 drain_all_pages(NULL
);
3258 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3262 pg_data_t
*last_pgdat
= NULL
;
3264 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3265 ac
->high_zoneidx
, ac
->nodemask
) {
3266 if (last_pgdat
!= zone
->zone_pgdat
)
3267 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3268 last_pgdat
= zone
->zone_pgdat
;
3272 static inline unsigned int
3273 gfp_to_alloc_flags(gfp_t gfp_mask
)
3275 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3277 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3278 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3281 * The caller may dip into page reserves a bit more if the caller
3282 * cannot run direct reclaim, or if the caller has realtime scheduling
3283 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3284 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3286 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3288 if (gfp_mask
& __GFP_ATOMIC
) {
3290 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3291 * if it can't schedule.
3293 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3294 alloc_flags
|= ALLOC_HARDER
;
3296 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3297 * comment for __cpuset_node_allowed().
3299 alloc_flags
&= ~ALLOC_CPUSET
;
3300 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3301 alloc_flags
|= ALLOC_HARDER
;
3304 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3305 alloc_flags
|= ALLOC_CMA
;
3310 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3312 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3315 if (gfp_mask
& __GFP_MEMALLOC
)
3317 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3319 if (!in_interrupt() &&
3320 ((current
->flags
& PF_MEMALLOC
) ||
3321 unlikely(test_thread_flag(TIF_MEMDIE
))))
3328 * Maximum number of reclaim retries without any progress before OOM killer
3329 * is consider as the only way to move forward.
3331 #define MAX_RECLAIM_RETRIES 16
3334 * Checks whether it makes sense to retry the reclaim to make a forward progress
3335 * for the given allocation request.
3336 * The reclaim feedback represented by did_some_progress (any progress during
3337 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3338 * any progress in a row) is considered as well as the reclaimable pages on the
3339 * applicable zone list (with a backoff mechanism which is a function of
3340 * no_progress_loops).
3342 * Returns true if a retry is viable or false to enter the oom path.
3345 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3346 struct alloc_context
*ac
, int alloc_flags
,
3347 bool did_some_progress
, int no_progress_loops
)
3353 * Make sure we converge to OOM if we cannot make any progress
3354 * several times in the row.
3356 if (no_progress_loops
> MAX_RECLAIM_RETRIES
)
3360 * Keep reclaiming pages while there is a chance this will lead
3361 * somewhere. If none of the target zones can satisfy our allocation
3362 * request even if all reclaimable pages are considered then we are
3363 * screwed and have to go OOM.
3365 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3367 unsigned long available
;
3368 unsigned long reclaimable
;
3370 available
= reclaimable
= zone_reclaimable_pages(zone
);
3371 available
-= DIV_ROUND_UP(no_progress_loops
* available
,
3372 MAX_RECLAIM_RETRIES
);
3373 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3376 * Would the allocation succeed if we reclaimed the whole
3379 if (__zone_watermark_ok(zone
, order
, min_wmark_pages(zone
),
3380 ac_classzone_idx(ac
), alloc_flags
, available
)) {
3382 * If we didn't make any progress and have a lot of
3383 * dirty + writeback pages then we should wait for
3384 * an IO to complete to slow down the reclaim and
3385 * prevent from pre mature OOM
3387 if (!did_some_progress
) {
3388 unsigned long write_pending
;
3390 write_pending
= zone_page_state_snapshot(zone
,
3391 NR_ZONE_WRITE_PENDING
);
3393 if (2 * write_pending
> reclaimable
) {
3394 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3400 * Memory allocation/reclaim might be called from a WQ
3401 * context and the current implementation of the WQ
3402 * concurrency control doesn't recognize that
3403 * a particular WQ is congested if the worker thread is
3404 * looping without ever sleeping. Therefore we have to
3405 * do a short sleep here rather than calling
3408 if (current
->flags
& PF_WQ_WORKER
)
3409 schedule_timeout_uninterruptible(1);
3420 static inline struct page
*
3421 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3422 struct alloc_context
*ac
)
3424 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3425 struct page
*page
= NULL
;
3426 unsigned int alloc_flags
;
3427 unsigned long did_some_progress
;
3428 enum compact_priority compact_priority
= DEF_COMPACT_PRIORITY
;
3429 enum compact_result compact_result
;
3430 int compaction_retries
= 0;
3431 int no_progress_loops
= 0;
3434 * In the slowpath, we sanity check order to avoid ever trying to
3435 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3436 * be using allocators in order of preference for an area that is
3439 if (order
>= MAX_ORDER
) {
3440 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3445 * We also sanity check to catch abuse of atomic reserves being used by
3446 * callers that are not in atomic context.
3448 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3449 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3450 gfp_mask
&= ~__GFP_ATOMIC
;
3453 * The fast path uses conservative alloc_flags to succeed only until
3454 * kswapd needs to be woken up, and to avoid the cost of setting up
3455 * alloc_flags precisely. So we do that now.
3457 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3459 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3460 wake_all_kswapds(order
, ac
);
3463 * The adjusted alloc_flags might result in immediate success, so try
3466 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3471 * For costly allocations, try direct compaction first, as it's likely
3472 * that we have enough base pages and don't need to reclaim. Don't try
3473 * that for allocations that are allowed to ignore watermarks, as the
3474 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3476 if (can_direct_reclaim
&& order
> PAGE_ALLOC_COSTLY_ORDER
&&
3477 !gfp_pfmemalloc_allowed(gfp_mask
)) {
3478 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3480 INIT_COMPACT_PRIORITY
,
3486 * Checks for costly allocations with __GFP_NORETRY, which
3487 * includes THP page fault allocations
3489 if (gfp_mask
& __GFP_NORETRY
) {
3491 * If compaction is deferred for high-order allocations,
3492 * it is because sync compaction recently failed. If
3493 * this is the case and the caller requested a THP
3494 * allocation, we do not want to heavily disrupt the
3495 * system, so we fail the allocation instead of entering
3498 if (compact_result
== COMPACT_DEFERRED
)
3502 * Looks like reclaim/compaction is worth trying, but
3503 * sync compaction could be very expensive, so keep
3504 * using async compaction.
3506 compact_priority
= INIT_COMPACT_PRIORITY
;
3511 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3512 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3513 wake_all_kswapds(order
, ac
);
3515 if (gfp_pfmemalloc_allowed(gfp_mask
))
3516 alloc_flags
= ALLOC_NO_WATERMARKS
;
3519 * Reset the zonelist iterators if memory policies can be ignored.
3520 * These allocations are high priority and system rather than user
3523 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3524 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3525 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3526 ac
->high_zoneidx
, ac
->nodemask
);
3529 /* Attempt with potentially adjusted zonelist and alloc_flags */
3530 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3534 /* Caller is not willing to reclaim, we can't balance anything */
3535 if (!can_direct_reclaim
) {
3537 * All existing users of the __GFP_NOFAIL are blockable, so warn
3538 * of any new users that actually allow this type of allocation
3541 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3545 /* Avoid recursion of direct reclaim */
3546 if (current
->flags
& PF_MEMALLOC
) {
3548 * __GFP_NOFAIL request from this context is rather bizarre
3549 * because we cannot reclaim anything and only can loop waiting
3550 * for somebody to do a work for us.
3552 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3559 /* Avoid allocations with no watermarks from looping endlessly */
3560 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3564 /* Try direct reclaim and then allocating */
3565 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3566 &did_some_progress
);
3570 /* Try direct compaction and then allocating */
3571 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3572 compact_priority
, &compact_result
);
3576 if (order
&& compaction_made_progress(compact_result
))
3577 compaction_retries
++;
3579 /* Do not loop if specifically requested */
3580 if (gfp_mask
& __GFP_NORETRY
)
3584 * Do not retry costly high order allocations unless they are
3587 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_REPEAT
))
3591 * Costly allocations might have made a progress but this doesn't mean
3592 * their order will become available due to high fragmentation so
3593 * always increment the no progress counter for them
3595 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3596 no_progress_loops
= 0;
3598 no_progress_loops
++;
3600 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3601 did_some_progress
> 0, no_progress_loops
))
3605 * It doesn't make any sense to retry for the compaction if the order-0
3606 * reclaim is not able to make any progress because the current
3607 * implementation of the compaction depends on the sufficient amount
3608 * of free memory (see __compaction_suitable)
3610 if (did_some_progress
> 0 &&
3611 should_compact_retry(ac
, order
, alloc_flags
,
3612 compact_result
, &compact_priority
,
3613 compaction_retries
))
3616 /* Reclaim has failed us, start killing things */
3617 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3621 /* Retry as long as the OOM killer is making progress */
3622 if (did_some_progress
) {
3623 no_progress_loops
= 0;
3628 warn_alloc_failed(gfp_mask
, order
, NULL
);
3634 * This is the 'heart' of the zoned buddy allocator.
3637 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3638 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3641 unsigned int cpuset_mems_cookie
;
3642 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
3643 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3644 struct alloc_context ac
= {
3645 .high_zoneidx
= gfp_zone(gfp_mask
),
3646 .zonelist
= zonelist
,
3647 .nodemask
= nodemask
,
3648 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3651 if (cpusets_enabled()) {
3652 alloc_mask
|= __GFP_HARDWALL
;
3653 alloc_flags
|= ALLOC_CPUSET
;
3655 ac
.nodemask
= &cpuset_current_mems_allowed
;
3658 gfp_mask
&= gfp_allowed_mask
;
3660 lockdep_trace_alloc(gfp_mask
);
3662 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3664 if (should_fail_alloc_page(gfp_mask
, order
))
3668 * Check the zones suitable for the gfp_mask contain at least one
3669 * valid zone. It's possible to have an empty zonelist as a result
3670 * of __GFP_THISNODE and a memoryless node
3672 if (unlikely(!zonelist
->_zonerefs
->zone
))
3675 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3676 alloc_flags
|= ALLOC_CMA
;
3679 cpuset_mems_cookie
= read_mems_allowed_begin();
3681 /* Dirty zone balancing only done in the fast path */
3682 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3685 * The preferred zone is used for statistics but crucially it is
3686 * also used as the starting point for the zonelist iterator. It
3687 * may get reset for allocations that ignore memory policies.
3689 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3690 ac
.high_zoneidx
, ac
.nodemask
);
3691 if (!ac
.preferred_zoneref
) {
3696 /* First allocation attempt */
3697 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3702 * Runtime PM, block IO and its error handling path can deadlock
3703 * because I/O on the device might not complete.
3705 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3706 ac
.spread_dirty_pages
= false;
3709 * Restore the original nodemask if it was potentially replaced with
3710 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3712 if (cpusets_enabled())
3713 ac
.nodemask
= nodemask
;
3714 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3718 * When updating a task's mems_allowed, it is possible to race with
3719 * parallel threads in such a way that an allocation can fail while
3720 * the mask is being updated. If a page allocation is about to fail,
3721 * check if the cpuset changed during allocation and if so, retry.
3723 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3724 alloc_mask
= gfp_mask
;
3729 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
3730 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
3731 __free_pages(page
, order
);
3735 if (kmemcheck_enabled
&& page
)
3736 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3738 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3742 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3745 * Common helper functions.
3747 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3752 * __get_free_pages() returns a 32-bit address, which cannot represent
3755 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3757 page
= alloc_pages(gfp_mask
, order
);
3760 return (unsigned long) page_address(page
);
3762 EXPORT_SYMBOL(__get_free_pages
);
3764 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3766 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3768 EXPORT_SYMBOL(get_zeroed_page
);
3770 void __free_pages(struct page
*page
, unsigned int order
)
3772 if (put_page_testzero(page
)) {
3774 free_hot_cold_page(page
, false);
3776 __free_pages_ok(page
, order
);
3780 EXPORT_SYMBOL(__free_pages
);
3782 void free_pages(unsigned long addr
, unsigned int order
)
3785 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3786 __free_pages(virt_to_page((void *)addr
), order
);
3790 EXPORT_SYMBOL(free_pages
);
3794 * An arbitrary-length arbitrary-offset area of memory which resides
3795 * within a 0 or higher order page. Multiple fragments within that page
3796 * are individually refcounted, in the page's reference counter.
3798 * The page_frag functions below provide a simple allocation framework for
3799 * page fragments. This is used by the network stack and network device
3800 * drivers to provide a backing region of memory for use as either an
3801 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3803 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3806 struct page
*page
= NULL
;
3807 gfp_t gfp
= gfp_mask
;
3809 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3810 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3812 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3813 PAGE_FRAG_CACHE_MAX_ORDER
);
3814 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3816 if (unlikely(!page
))
3817 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3819 nc
->va
= page
? page_address(page
) : NULL
;
3824 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3825 unsigned int fragsz
, gfp_t gfp_mask
)
3827 unsigned int size
= PAGE_SIZE
;
3831 if (unlikely(!nc
->va
)) {
3833 page
= __page_frag_refill(nc
, gfp_mask
);
3837 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3838 /* if size can vary use size else just use PAGE_SIZE */
3841 /* Even if we own the page, we do not use atomic_set().
3842 * This would break get_page_unless_zero() users.
3844 page_ref_add(page
, size
- 1);
3846 /* reset page count bias and offset to start of new frag */
3847 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3848 nc
->pagecnt_bias
= size
;
3852 offset
= nc
->offset
- fragsz
;
3853 if (unlikely(offset
< 0)) {
3854 page
= virt_to_page(nc
->va
);
3856 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3859 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3860 /* if size can vary use size else just use PAGE_SIZE */
3863 /* OK, page count is 0, we can safely set it */
3864 set_page_count(page
, size
);
3866 /* reset page count bias and offset to start of new frag */
3867 nc
->pagecnt_bias
= size
;
3868 offset
= size
- fragsz
;
3872 nc
->offset
= offset
;
3874 return nc
->va
+ offset
;
3876 EXPORT_SYMBOL(__alloc_page_frag
);
3879 * Frees a page fragment allocated out of either a compound or order 0 page.
3881 void __free_page_frag(void *addr
)
3883 struct page
*page
= virt_to_head_page(addr
);
3885 if (unlikely(put_page_testzero(page
)))
3886 __free_pages_ok(page
, compound_order(page
));
3888 EXPORT_SYMBOL(__free_page_frag
);
3890 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3894 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3895 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3897 split_page(virt_to_page((void *)addr
), order
);
3898 while (used
< alloc_end
) {
3903 return (void *)addr
;
3907 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3908 * @size: the number of bytes to allocate
3909 * @gfp_mask: GFP flags for the allocation
3911 * This function is similar to alloc_pages(), except that it allocates the
3912 * minimum number of pages to satisfy the request. alloc_pages() can only
3913 * allocate memory in power-of-two pages.
3915 * This function is also limited by MAX_ORDER.
3917 * Memory allocated by this function must be released by free_pages_exact().
3919 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3921 unsigned int order
= get_order(size
);
3924 addr
= __get_free_pages(gfp_mask
, order
);
3925 return make_alloc_exact(addr
, order
, size
);
3927 EXPORT_SYMBOL(alloc_pages_exact
);
3930 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3932 * @nid: the preferred node ID where memory should be allocated
3933 * @size: the number of bytes to allocate
3934 * @gfp_mask: GFP flags for the allocation
3936 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3939 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3941 unsigned int order
= get_order(size
);
3942 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3945 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3949 * free_pages_exact - release memory allocated via alloc_pages_exact()
3950 * @virt: the value returned by alloc_pages_exact.
3951 * @size: size of allocation, same value as passed to alloc_pages_exact().
3953 * Release the memory allocated by a previous call to alloc_pages_exact.
3955 void free_pages_exact(void *virt
, size_t size
)
3957 unsigned long addr
= (unsigned long)virt
;
3958 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3960 while (addr
< end
) {
3965 EXPORT_SYMBOL(free_pages_exact
);
3968 * nr_free_zone_pages - count number of pages beyond high watermark
3969 * @offset: The zone index of the highest zone
3971 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3972 * high watermark within all zones at or below a given zone index. For each
3973 * zone, the number of pages is calculated as:
3974 * managed_pages - high_pages
3976 static unsigned long nr_free_zone_pages(int offset
)
3981 /* Just pick one node, since fallback list is circular */
3982 unsigned long sum
= 0;
3984 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3986 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3987 unsigned long size
= zone
->managed_pages
;
3988 unsigned long high
= high_wmark_pages(zone
);
3997 * nr_free_buffer_pages - count number of pages beyond high watermark
3999 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4000 * watermark within ZONE_DMA and ZONE_NORMAL.
4002 unsigned long nr_free_buffer_pages(void)
4004 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4006 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4009 * nr_free_pagecache_pages - count number of pages beyond high watermark
4011 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4012 * high watermark within all zones.
4014 unsigned long nr_free_pagecache_pages(void)
4016 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4019 static inline void show_node(struct zone
*zone
)
4021 if (IS_ENABLED(CONFIG_NUMA
))
4022 printk("Node %d ", zone_to_nid(zone
));
4025 long si_mem_available(void)
4028 unsigned long pagecache
;
4029 unsigned long wmark_low
= 0;
4030 unsigned long pages
[NR_LRU_LISTS
];
4034 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4035 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4038 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4041 * Estimate the amount of memory available for userspace allocations,
4042 * without causing swapping.
4044 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4047 * Not all the page cache can be freed, otherwise the system will
4048 * start swapping. Assume at least half of the page cache, or the
4049 * low watermark worth of cache, needs to stay.
4051 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4052 pagecache
-= min(pagecache
/ 2, wmark_low
);
4053 available
+= pagecache
;
4056 * Part of the reclaimable slab consists of items that are in use,
4057 * and cannot be freed. Cap this estimate at the low watermark.
4059 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4060 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4066 EXPORT_SYMBOL_GPL(si_mem_available
);
4068 void si_meminfo(struct sysinfo
*val
)
4070 val
->totalram
= totalram_pages
;
4071 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4072 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4073 val
->bufferram
= nr_blockdev_pages();
4074 val
->totalhigh
= totalhigh_pages
;
4075 val
->freehigh
= nr_free_highpages();
4076 val
->mem_unit
= PAGE_SIZE
;
4079 EXPORT_SYMBOL(si_meminfo
);
4082 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4084 int zone_type
; /* needs to be signed */
4085 unsigned long managed_pages
= 0;
4086 unsigned long managed_highpages
= 0;
4087 unsigned long free_highpages
= 0;
4088 pg_data_t
*pgdat
= NODE_DATA(nid
);
4090 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4091 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4092 val
->totalram
= managed_pages
;
4093 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4094 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4095 #ifdef CONFIG_HIGHMEM
4096 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4097 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4099 if (is_highmem(zone
)) {
4100 managed_highpages
+= zone
->managed_pages
;
4101 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4104 val
->totalhigh
= managed_highpages
;
4105 val
->freehigh
= free_highpages
;
4107 val
->totalhigh
= managed_highpages
;
4108 val
->freehigh
= free_highpages
;
4110 val
->mem_unit
= PAGE_SIZE
;
4115 * Determine whether the node should be displayed or not, depending on whether
4116 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4118 bool skip_free_areas_node(unsigned int flags
, int nid
)
4121 unsigned int cpuset_mems_cookie
;
4123 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4127 cpuset_mems_cookie
= read_mems_allowed_begin();
4128 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
4129 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
4134 #define K(x) ((x) << (PAGE_SHIFT-10))
4136 static void show_migration_types(unsigned char type
)
4138 static const char types
[MIGRATE_TYPES
] = {
4139 [MIGRATE_UNMOVABLE
] = 'U',
4140 [MIGRATE_MOVABLE
] = 'M',
4141 [MIGRATE_RECLAIMABLE
] = 'E',
4142 [MIGRATE_HIGHATOMIC
] = 'H',
4144 [MIGRATE_CMA
] = 'C',
4146 #ifdef CONFIG_MEMORY_ISOLATION
4147 [MIGRATE_ISOLATE
] = 'I',
4150 char tmp
[MIGRATE_TYPES
+ 1];
4154 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4155 if (type
& (1 << i
))
4160 printk("(%s) ", tmp
);
4164 * Show free area list (used inside shift_scroll-lock stuff)
4165 * We also calculate the percentage fragmentation. We do this by counting the
4166 * memory on each free list with the exception of the first item on the list.
4169 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4172 void show_free_areas(unsigned int filter
)
4174 unsigned long free_pcp
= 0;
4179 for_each_populated_zone(zone
) {
4180 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4183 for_each_online_cpu(cpu
)
4184 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4187 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4188 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4189 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4190 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4191 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4192 " free:%lu free_pcp:%lu free_cma:%lu\n",
4193 global_node_page_state(NR_ACTIVE_ANON
),
4194 global_node_page_state(NR_INACTIVE_ANON
),
4195 global_node_page_state(NR_ISOLATED_ANON
),
4196 global_node_page_state(NR_ACTIVE_FILE
),
4197 global_node_page_state(NR_INACTIVE_FILE
),
4198 global_node_page_state(NR_ISOLATED_FILE
),
4199 global_node_page_state(NR_UNEVICTABLE
),
4200 global_node_page_state(NR_FILE_DIRTY
),
4201 global_node_page_state(NR_WRITEBACK
),
4202 global_node_page_state(NR_UNSTABLE_NFS
),
4203 global_page_state(NR_SLAB_RECLAIMABLE
),
4204 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4205 global_node_page_state(NR_FILE_MAPPED
),
4206 global_node_page_state(NR_SHMEM
),
4207 global_page_state(NR_PAGETABLE
),
4208 global_page_state(NR_BOUNCE
),
4209 global_page_state(NR_FREE_PAGES
),
4211 global_page_state(NR_FREE_CMA_PAGES
));
4213 for_each_online_pgdat(pgdat
) {
4215 " active_anon:%lukB"
4216 " inactive_anon:%lukB"
4217 " active_file:%lukB"
4218 " inactive_file:%lukB"
4219 " unevictable:%lukB"
4220 " isolated(anon):%lukB"
4221 " isolated(file):%lukB"
4226 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4228 " shmem_pmdmapped: %lukB"
4231 " writeback_tmp:%lukB"
4233 " pages_scanned:%lu"
4234 " all_unreclaimable? %s"
4237 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4238 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4239 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4240 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4241 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4242 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4243 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4244 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4245 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4246 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4247 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4248 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4249 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4251 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4253 K(node_page_state(pgdat
, NR_SHMEM
)),
4254 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4255 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4256 node_page_state(pgdat
, NR_PAGES_SCANNED
),
4257 !pgdat_reclaimable(pgdat
) ? "yes" : "no");
4260 for_each_populated_zone(zone
) {
4263 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4267 for_each_online_cpu(cpu
)
4268 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4276 " active_anon:%lukB"
4277 " inactive_anon:%lukB"
4278 " active_file:%lukB"
4279 " inactive_file:%lukB"
4280 " unevictable:%lukB"
4281 " writepending:%lukB"
4285 " slab_reclaimable:%lukB"
4286 " slab_unreclaimable:%lukB"
4287 " kernel_stack:%lukB"
4295 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4296 K(min_wmark_pages(zone
)),
4297 K(low_wmark_pages(zone
)),
4298 K(high_wmark_pages(zone
)),
4299 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4300 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4301 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4302 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4303 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4304 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4305 K(zone
->present_pages
),
4306 K(zone
->managed_pages
),
4307 K(zone_page_state(zone
, NR_MLOCK
)),
4308 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4309 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4310 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4311 K(zone_page_state(zone
, NR_PAGETABLE
)),
4312 K(zone_page_state(zone
, NR_BOUNCE
)),
4314 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4315 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4316 printk("lowmem_reserve[]:");
4317 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4318 printk(" %ld", zone
->lowmem_reserve
[i
]);
4322 for_each_populated_zone(zone
) {
4324 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4325 unsigned char types
[MAX_ORDER
];
4327 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4330 printk("%s: ", zone
->name
);
4332 spin_lock_irqsave(&zone
->lock
, flags
);
4333 for (order
= 0; order
< MAX_ORDER
; order
++) {
4334 struct free_area
*area
= &zone
->free_area
[order
];
4337 nr
[order
] = area
->nr_free
;
4338 total
+= nr
[order
] << order
;
4341 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4342 if (!list_empty(&area
->free_list
[type
]))
4343 types
[order
] |= 1 << type
;
4346 spin_unlock_irqrestore(&zone
->lock
, flags
);
4347 for (order
= 0; order
< MAX_ORDER
; order
++) {
4348 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
4350 show_migration_types(types
[order
]);
4352 printk("= %lukB\n", K(total
));
4355 hugetlb_show_meminfo();
4357 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4359 show_swap_cache_info();
4362 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4364 zoneref
->zone
= zone
;
4365 zoneref
->zone_idx
= zone_idx(zone
);
4369 * Builds allocation fallback zone lists.
4371 * Add all populated zones of a node to the zonelist.
4373 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4377 enum zone_type zone_type
= MAX_NR_ZONES
;
4381 zone
= pgdat
->node_zones
+ zone_type
;
4382 if (managed_zone(zone
)) {
4383 zoneref_set_zone(zone
,
4384 &zonelist
->_zonerefs
[nr_zones
++]);
4385 check_highest_zone(zone_type
);
4387 } while (zone_type
);
4395 * 0 = automatic detection of better ordering.
4396 * 1 = order by ([node] distance, -zonetype)
4397 * 2 = order by (-zonetype, [node] distance)
4399 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4400 * the same zonelist. So only NUMA can configure this param.
4402 #define ZONELIST_ORDER_DEFAULT 0
4403 #define ZONELIST_ORDER_NODE 1
4404 #define ZONELIST_ORDER_ZONE 2
4406 /* zonelist order in the kernel.
4407 * set_zonelist_order() will set this to NODE or ZONE.
4409 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4410 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4414 /* The value user specified ....changed by config */
4415 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4416 /* string for sysctl */
4417 #define NUMA_ZONELIST_ORDER_LEN 16
4418 char numa_zonelist_order
[16] = "default";
4421 * interface for configure zonelist ordering.
4422 * command line option "numa_zonelist_order"
4423 * = "[dD]efault - default, automatic configuration.
4424 * = "[nN]ode - order by node locality, then by zone within node
4425 * = "[zZ]one - order by zone, then by locality within zone
4428 static int __parse_numa_zonelist_order(char *s
)
4430 if (*s
== 'd' || *s
== 'D') {
4431 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4432 } else if (*s
== 'n' || *s
== 'N') {
4433 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4434 } else if (*s
== 'z' || *s
== 'Z') {
4435 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4437 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4443 static __init
int setup_numa_zonelist_order(char *s
)
4450 ret
= __parse_numa_zonelist_order(s
);
4452 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4456 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4459 * sysctl handler for numa_zonelist_order
4461 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4462 void __user
*buffer
, size_t *length
,
4465 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4467 static DEFINE_MUTEX(zl_order_mutex
);
4469 mutex_lock(&zl_order_mutex
);
4471 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4475 strcpy(saved_string
, (char *)table
->data
);
4477 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4481 int oldval
= user_zonelist_order
;
4483 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4486 * bogus value. restore saved string
4488 strncpy((char *)table
->data
, saved_string
,
4489 NUMA_ZONELIST_ORDER_LEN
);
4490 user_zonelist_order
= oldval
;
4491 } else if (oldval
!= user_zonelist_order
) {
4492 mutex_lock(&zonelists_mutex
);
4493 build_all_zonelists(NULL
, NULL
);
4494 mutex_unlock(&zonelists_mutex
);
4498 mutex_unlock(&zl_order_mutex
);
4503 #define MAX_NODE_LOAD (nr_online_nodes)
4504 static int node_load
[MAX_NUMNODES
];
4507 * find_next_best_node - find the next node that should appear in a given node's fallback list
4508 * @node: node whose fallback list we're appending
4509 * @used_node_mask: nodemask_t of already used nodes
4511 * We use a number of factors to determine which is the next node that should
4512 * appear on a given node's fallback list. The node should not have appeared
4513 * already in @node's fallback list, and it should be the next closest node
4514 * according to the distance array (which contains arbitrary distance values
4515 * from each node to each node in the system), and should also prefer nodes
4516 * with no CPUs, since presumably they'll have very little allocation pressure
4517 * on them otherwise.
4518 * It returns -1 if no node is found.
4520 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4523 int min_val
= INT_MAX
;
4524 int best_node
= NUMA_NO_NODE
;
4525 const struct cpumask
*tmp
= cpumask_of_node(0);
4527 /* Use the local node if we haven't already */
4528 if (!node_isset(node
, *used_node_mask
)) {
4529 node_set(node
, *used_node_mask
);
4533 for_each_node_state(n
, N_MEMORY
) {
4535 /* Don't want a node to appear more than once */
4536 if (node_isset(n
, *used_node_mask
))
4539 /* Use the distance array to find the distance */
4540 val
= node_distance(node
, n
);
4542 /* Penalize nodes under us ("prefer the next node") */
4545 /* Give preference to headless and unused nodes */
4546 tmp
= cpumask_of_node(n
);
4547 if (!cpumask_empty(tmp
))
4548 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4550 /* Slight preference for less loaded node */
4551 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4552 val
+= node_load
[n
];
4554 if (val
< min_val
) {
4561 node_set(best_node
, *used_node_mask
);
4568 * Build zonelists ordered by node and zones within node.
4569 * This results in maximum locality--normal zone overflows into local
4570 * DMA zone, if any--but risks exhausting DMA zone.
4572 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4575 struct zonelist
*zonelist
;
4577 zonelist
= &pgdat
->node_zonelists
[0];
4578 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4580 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4581 zonelist
->_zonerefs
[j
].zone
= NULL
;
4582 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4586 * Build gfp_thisnode zonelists
4588 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4591 struct zonelist
*zonelist
;
4593 zonelist
= &pgdat
->node_zonelists
[1];
4594 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4595 zonelist
->_zonerefs
[j
].zone
= NULL
;
4596 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4600 * Build zonelists ordered by zone and nodes within zones.
4601 * This results in conserving DMA zone[s] until all Normal memory is
4602 * exhausted, but results in overflowing to remote node while memory
4603 * may still exist in local DMA zone.
4605 static int node_order
[MAX_NUMNODES
];
4607 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4610 int zone_type
; /* needs to be signed */
4612 struct zonelist
*zonelist
;
4614 zonelist
= &pgdat
->node_zonelists
[0];
4616 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4617 for (j
= 0; j
< nr_nodes
; j
++) {
4618 node
= node_order
[j
];
4619 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4620 if (managed_zone(z
)) {
4622 &zonelist
->_zonerefs
[pos
++]);
4623 check_highest_zone(zone_type
);
4627 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4628 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4631 #if defined(CONFIG_64BIT)
4633 * Devices that require DMA32/DMA are relatively rare and do not justify a
4634 * penalty to every machine in case the specialised case applies. Default
4635 * to Node-ordering on 64-bit NUMA machines
4637 static int default_zonelist_order(void)
4639 return ZONELIST_ORDER_NODE
;
4643 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4644 * by the kernel. If processes running on node 0 deplete the low memory zone
4645 * then reclaim will occur more frequency increasing stalls and potentially
4646 * be easier to OOM if a large percentage of the zone is under writeback or
4647 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4648 * Hence, default to zone ordering on 32-bit.
4650 static int default_zonelist_order(void)
4652 return ZONELIST_ORDER_ZONE
;
4654 #endif /* CONFIG_64BIT */
4656 static void set_zonelist_order(void)
4658 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4659 current_zonelist_order
= default_zonelist_order();
4661 current_zonelist_order
= user_zonelist_order
;
4664 static void build_zonelists(pg_data_t
*pgdat
)
4667 nodemask_t used_mask
;
4668 int local_node
, prev_node
;
4669 struct zonelist
*zonelist
;
4670 unsigned int order
= current_zonelist_order
;
4672 /* initialize zonelists */
4673 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4674 zonelist
= pgdat
->node_zonelists
+ i
;
4675 zonelist
->_zonerefs
[0].zone
= NULL
;
4676 zonelist
->_zonerefs
[0].zone_idx
= 0;
4679 /* NUMA-aware ordering of nodes */
4680 local_node
= pgdat
->node_id
;
4681 load
= nr_online_nodes
;
4682 prev_node
= local_node
;
4683 nodes_clear(used_mask
);
4685 memset(node_order
, 0, sizeof(node_order
));
4688 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4690 * We don't want to pressure a particular node.
4691 * So adding penalty to the first node in same
4692 * distance group to make it round-robin.
4694 if (node_distance(local_node
, node
) !=
4695 node_distance(local_node
, prev_node
))
4696 node_load
[node
] = load
;
4700 if (order
== ZONELIST_ORDER_NODE
)
4701 build_zonelists_in_node_order(pgdat
, node
);
4703 node_order
[i
++] = node
; /* remember order */
4706 if (order
== ZONELIST_ORDER_ZONE
) {
4707 /* calculate node order -- i.e., DMA last! */
4708 build_zonelists_in_zone_order(pgdat
, i
);
4711 build_thisnode_zonelists(pgdat
);
4714 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4716 * Return node id of node used for "local" allocations.
4717 * I.e., first node id of first zone in arg node's generic zonelist.
4718 * Used for initializing percpu 'numa_mem', which is used primarily
4719 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4721 int local_memory_node(int node
)
4725 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4726 gfp_zone(GFP_KERNEL
),
4728 return z
->zone
->node
;
4732 static void setup_min_unmapped_ratio(void);
4733 static void setup_min_slab_ratio(void);
4734 #else /* CONFIG_NUMA */
4736 static void set_zonelist_order(void)
4738 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4741 static void build_zonelists(pg_data_t
*pgdat
)
4743 int node
, local_node
;
4745 struct zonelist
*zonelist
;
4747 local_node
= pgdat
->node_id
;
4749 zonelist
= &pgdat
->node_zonelists
[0];
4750 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4753 * Now we build the zonelist so that it contains the zones
4754 * of all the other nodes.
4755 * We don't want to pressure a particular node, so when
4756 * building the zones for node N, we make sure that the
4757 * zones coming right after the local ones are those from
4758 * node N+1 (modulo N)
4760 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4761 if (!node_online(node
))
4763 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4765 for (node
= 0; node
< local_node
; node
++) {
4766 if (!node_online(node
))
4768 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4771 zonelist
->_zonerefs
[j
].zone
= NULL
;
4772 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4775 #endif /* CONFIG_NUMA */
4778 * Boot pageset table. One per cpu which is going to be used for all
4779 * zones and all nodes. The parameters will be set in such a way
4780 * that an item put on a list will immediately be handed over to
4781 * the buddy list. This is safe since pageset manipulation is done
4782 * with interrupts disabled.
4784 * The boot_pagesets must be kept even after bootup is complete for
4785 * unused processors and/or zones. They do play a role for bootstrapping
4786 * hotplugged processors.
4788 * zoneinfo_show() and maybe other functions do
4789 * not check if the processor is online before following the pageset pointer.
4790 * Other parts of the kernel may not check if the zone is available.
4792 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4793 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4794 static void setup_zone_pageset(struct zone
*zone
);
4797 * Global mutex to protect against size modification of zonelists
4798 * as well as to serialize pageset setup for the new populated zone.
4800 DEFINE_MUTEX(zonelists_mutex
);
4802 /* return values int ....just for stop_machine() */
4803 static int __build_all_zonelists(void *data
)
4807 pg_data_t
*self
= data
;
4810 memset(node_load
, 0, sizeof(node_load
));
4813 if (self
&& !node_online(self
->node_id
)) {
4814 build_zonelists(self
);
4817 for_each_online_node(nid
) {
4818 pg_data_t
*pgdat
= NODE_DATA(nid
);
4820 build_zonelists(pgdat
);
4824 * Initialize the boot_pagesets that are going to be used
4825 * for bootstrapping processors. The real pagesets for
4826 * each zone will be allocated later when the per cpu
4827 * allocator is available.
4829 * boot_pagesets are used also for bootstrapping offline
4830 * cpus if the system is already booted because the pagesets
4831 * are needed to initialize allocators on a specific cpu too.
4832 * F.e. the percpu allocator needs the page allocator which
4833 * needs the percpu allocator in order to allocate its pagesets
4834 * (a chicken-egg dilemma).
4836 for_each_possible_cpu(cpu
) {
4837 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4839 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4841 * We now know the "local memory node" for each node--
4842 * i.e., the node of the first zone in the generic zonelist.
4843 * Set up numa_mem percpu variable for on-line cpus. During
4844 * boot, only the boot cpu should be on-line; we'll init the
4845 * secondary cpus' numa_mem as they come on-line. During
4846 * node/memory hotplug, we'll fixup all on-line cpus.
4848 if (cpu_online(cpu
))
4849 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4856 static noinline
void __init
4857 build_all_zonelists_init(void)
4859 __build_all_zonelists(NULL
);
4860 mminit_verify_zonelist();
4861 cpuset_init_current_mems_allowed();
4865 * Called with zonelists_mutex held always
4866 * unless system_state == SYSTEM_BOOTING.
4868 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4869 * [we're only called with non-NULL zone through __meminit paths] and
4870 * (2) call of __init annotated helper build_all_zonelists_init
4871 * [protected by SYSTEM_BOOTING].
4873 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4875 set_zonelist_order();
4877 if (system_state
== SYSTEM_BOOTING
) {
4878 build_all_zonelists_init();
4880 #ifdef CONFIG_MEMORY_HOTPLUG
4882 setup_zone_pageset(zone
);
4884 /* we have to stop all cpus to guarantee there is no user
4886 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4887 /* cpuset refresh routine should be here */
4889 vm_total_pages
= nr_free_pagecache_pages();
4891 * Disable grouping by mobility if the number of pages in the
4892 * system is too low to allow the mechanism to work. It would be
4893 * more accurate, but expensive to check per-zone. This check is
4894 * made on memory-hotadd so a system can start with mobility
4895 * disabled and enable it later
4897 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4898 page_group_by_mobility_disabled
= 1;
4900 page_group_by_mobility_disabled
= 0;
4902 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4904 zonelist_order_name
[current_zonelist_order
],
4905 page_group_by_mobility_disabled
? "off" : "on",
4908 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4913 * Helper functions to size the waitqueue hash table.
4914 * Essentially these want to choose hash table sizes sufficiently
4915 * large so that collisions trying to wait on pages are rare.
4916 * But in fact, the number of active page waitqueues on typical
4917 * systems is ridiculously low, less than 200. So this is even
4918 * conservative, even though it seems large.
4920 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4921 * waitqueues, i.e. the size of the waitq table given the number of pages.
4923 #define PAGES_PER_WAITQUEUE 256
4925 #ifndef CONFIG_MEMORY_HOTPLUG
4926 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4928 unsigned long size
= 1;
4930 pages
/= PAGES_PER_WAITQUEUE
;
4932 while (size
< pages
)
4936 * Once we have dozens or even hundreds of threads sleeping
4937 * on IO we've got bigger problems than wait queue collision.
4938 * Limit the size of the wait table to a reasonable size.
4940 size
= min(size
, 4096UL);
4942 return max(size
, 4UL);
4946 * A zone's size might be changed by hot-add, so it is not possible to determine
4947 * a suitable size for its wait_table. So we use the maximum size now.
4949 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4951 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4952 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4953 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4955 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4956 * or more by the traditional way. (See above). It equals:
4958 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4959 * ia64(16K page size) : = ( 8G + 4M)byte.
4960 * powerpc (64K page size) : = (32G +16M)byte.
4962 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4969 * This is an integer logarithm so that shifts can be used later
4970 * to extract the more random high bits from the multiplicative
4971 * hash function before the remainder is taken.
4973 static inline unsigned long wait_table_bits(unsigned long size
)
4979 * Initially all pages are reserved - free ones are freed
4980 * up by free_all_bootmem() once the early boot process is
4981 * done. Non-atomic initialization, single-pass.
4983 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4984 unsigned long start_pfn
, enum memmap_context context
)
4986 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
4987 unsigned long end_pfn
= start_pfn
+ size
;
4988 pg_data_t
*pgdat
= NODE_DATA(nid
);
4990 unsigned long nr_initialised
= 0;
4991 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4992 struct memblock_region
*r
= NULL
, *tmp
;
4995 if (highest_memmap_pfn
< end_pfn
- 1)
4996 highest_memmap_pfn
= end_pfn
- 1;
4999 * Honor reservation requested by the driver for this ZONE_DEVICE
5002 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5003 start_pfn
+= altmap
->reserve
;
5005 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5007 * There can be holes in boot-time mem_map[]s handed to this
5008 * function. They do not exist on hotplugged memory.
5010 if (context
!= MEMMAP_EARLY
)
5013 if (!early_pfn_valid(pfn
))
5015 if (!early_pfn_in_nid(pfn
, nid
))
5017 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5020 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5022 * Check given memblock attribute by firmware which can affect
5023 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5024 * mirrored, it's an overlapped memmap init. skip it.
5026 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5027 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5028 for_each_memblock(memory
, tmp
)
5029 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5033 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5034 memblock_is_mirror(r
)) {
5035 /* already initialized as NORMAL */
5036 pfn
= memblock_region_memory_end_pfn(r
);
5044 * Mark the block movable so that blocks are reserved for
5045 * movable at startup. This will force kernel allocations
5046 * to reserve their blocks rather than leaking throughout
5047 * the address space during boot when many long-lived
5048 * kernel allocations are made.
5050 * bitmap is created for zone's valid pfn range. but memmap
5051 * can be created for invalid pages (for alignment)
5052 * check here not to call set_pageblock_migratetype() against
5055 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5056 struct page
*page
= pfn_to_page(pfn
);
5058 __init_single_page(page
, pfn
, zone
, nid
);
5059 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5061 __init_single_pfn(pfn
, zone
, nid
);
5066 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5068 unsigned int order
, t
;
5069 for_each_migratetype_order(order
, t
) {
5070 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5071 zone
->free_area
[order
].nr_free
= 0;
5075 #ifndef __HAVE_ARCH_MEMMAP_INIT
5076 #define memmap_init(size, nid, zone, start_pfn) \
5077 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5080 static int zone_batchsize(struct zone
*zone
)
5086 * The per-cpu-pages pools are set to around 1000th of the
5087 * size of the zone. But no more than 1/2 of a meg.
5089 * OK, so we don't know how big the cache is. So guess.
5091 batch
= zone
->managed_pages
/ 1024;
5092 if (batch
* PAGE_SIZE
> 512 * 1024)
5093 batch
= (512 * 1024) / PAGE_SIZE
;
5094 batch
/= 4; /* We effectively *= 4 below */
5099 * Clamp the batch to a 2^n - 1 value. Having a power
5100 * of 2 value was found to be more likely to have
5101 * suboptimal cache aliasing properties in some cases.
5103 * For example if 2 tasks are alternately allocating
5104 * batches of pages, one task can end up with a lot
5105 * of pages of one half of the possible page colors
5106 * and the other with pages of the other colors.
5108 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5113 /* The deferral and batching of frees should be suppressed under NOMMU
5116 * The problem is that NOMMU needs to be able to allocate large chunks
5117 * of contiguous memory as there's no hardware page translation to
5118 * assemble apparent contiguous memory from discontiguous pages.
5120 * Queueing large contiguous runs of pages for batching, however,
5121 * causes the pages to actually be freed in smaller chunks. As there
5122 * can be a significant delay between the individual batches being
5123 * recycled, this leads to the once large chunks of space being
5124 * fragmented and becoming unavailable for high-order allocations.
5131 * pcp->high and pcp->batch values are related and dependent on one another:
5132 * ->batch must never be higher then ->high.
5133 * The following function updates them in a safe manner without read side
5136 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5137 * those fields changing asynchronously (acording the the above rule).
5139 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5140 * outside of boot time (or some other assurance that no concurrent updaters
5143 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5144 unsigned long batch
)
5146 /* start with a fail safe value for batch */
5150 /* Update high, then batch, in order */
5157 /* a companion to pageset_set_high() */
5158 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5160 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5163 static void pageset_init(struct per_cpu_pageset
*p
)
5165 struct per_cpu_pages
*pcp
;
5168 memset(p
, 0, sizeof(*p
));
5172 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5173 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5176 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5179 pageset_set_batch(p
, batch
);
5183 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5184 * to the value high for the pageset p.
5186 static void pageset_set_high(struct per_cpu_pageset
*p
,
5189 unsigned long batch
= max(1UL, high
/ 4);
5190 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5191 batch
= PAGE_SHIFT
* 8;
5193 pageset_update(&p
->pcp
, high
, batch
);
5196 static void pageset_set_high_and_batch(struct zone
*zone
,
5197 struct per_cpu_pageset
*pcp
)
5199 if (percpu_pagelist_fraction
)
5200 pageset_set_high(pcp
,
5201 (zone
->managed_pages
/
5202 percpu_pagelist_fraction
));
5204 pageset_set_batch(pcp
, zone_batchsize(zone
));
5207 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5209 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5212 pageset_set_high_and_batch(zone
, pcp
);
5215 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5218 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5219 for_each_possible_cpu(cpu
)
5220 zone_pageset_init(zone
, cpu
);
5224 * Allocate per cpu pagesets and initialize them.
5225 * Before this call only boot pagesets were available.
5227 void __init
setup_per_cpu_pageset(void)
5229 struct pglist_data
*pgdat
;
5232 for_each_populated_zone(zone
)
5233 setup_zone_pageset(zone
);
5235 for_each_online_pgdat(pgdat
)
5236 pgdat
->per_cpu_nodestats
=
5237 alloc_percpu(struct per_cpu_nodestat
);
5240 static noinline __ref
5241 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
5247 * The per-page waitqueue mechanism uses hashed waitqueues
5250 zone
->wait_table_hash_nr_entries
=
5251 wait_table_hash_nr_entries(zone_size_pages
);
5252 zone
->wait_table_bits
=
5253 wait_table_bits(zone
->wait_table_hash_nr_entries
);
5254 alloc_size
= zone
->wait_table_hash_nr_entries
5255 * sizeof(wait_queue_head_t
);
5257 if (!slab_is_available()) {
5258 zone
->wait_table
= (wait_queue_head_t
*)
5259 memblock_virt_alloc_node_nopanic(
5260 alloc_size
, zone
->zone_pgdat
->node_id
);
5263 * This case means that a zone whose size was 0 gets new memory
5264 * via memory hot-add.
5265 * But it may be the case that a new node was hot-added. In
5266 * this case vmalloc() will not be able to use this new node's
5267 * memory - this wait_table must be initialized to use this new
5268 * node itself as well.
5269 * To use this new node's memory, further consideration will be
5272 zone
->wait_table
= vmalloc(alloc_size
);
5274 if (!zone
->wait_table
)
5277 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
5278 init_waitqueue_head(zone
->wait_table
+ i
);
5283 static __meminit
void zone_pcp_init(struct zone
*zone
)
5286 * per cpu subsystem is not up at this point. The following code
5287 * relies on the ability of the linker to provide the
5288 * offset of a (static) per cpu variable into the per cpu area.
5290 zone
->pageset
= &boot_pageset
;
5292 if (populated_zone(zone
))
5293 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5294 zone
->name
, zone
->present_pages
,
5295 zone_batchsize(zone
));
5298 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5299 unsigned long zone_start_pfn
,
5302 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5304 ret
= zone_wait_table_init(zone
, size
);
5307 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5309 zone
->zone_start_pfn
= zone_start_pfn
;
5311 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5312 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5314 (unsigned long)zone_idx(zone
),
5315 zone_start_pfn
, (zone_start_pfn
+ size
));
5317 zone_init_free_lists(zone
);
5322 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5323 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5326 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5328 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5329 struct mminit_pfnnid_cache
*state
)
5331 unsigned long start_pfn
, end_pfn
;
5334 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5335 return state
->last_nid
;
5337 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5339 state
->last_start
= start_pfn
;
5340 state
->last_end
= end_pfn
;
5341 state
->last_nid
= nid
;
5346 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5349 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5350 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5351 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5353 * If an architecture guarantees that all ranges registered contain no holes
5354 * and may be freed, this this function may be used instead of calling
5355 * memblock_free_early_nid() manually.
5357 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5359 unsigned long start_pfn
, end_pfn
;
5362 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5363 start_pfn
= min(start_pfn
, max_low_pfn
);
5364 end_pfn
= min(end_pfn
, max_low_pfn
);
5366 if (start_pfn
< end_pfn
)
5367 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5368 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5374 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5375 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5377 * If an architecture guarantees that all ranges registered contain no holes and may
5378 * be freed, this function may be used instead of calling memory_present() manually.
5380 void __init
sparse_memory_present_with_active_regions(int nid
)
5382 unsigned long start_pfn
, end_pfn
;
5385 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5386 memory_present(this_nid
, start_pfn
, end_pfn
);
5390 * get_pfn_range_for_nid - Return the start and end page frames for a node
5391 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5392 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5393 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5395 * It returns the start and end page frame of a node based on information
5396 * provided by memblock_set_node(). If called for a node
5397 * with no available memory, a warning is printed and the start and end
5400 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5401 unsigned long *start_pfn
, unsigned long *end_pfn
)
5403 unsigned long this_start_pfn
, this_end_pfn
;
5409 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5410 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5411 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5414 if (*start_pfn
== -1UL)
5419 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5420 * assumption is made that zones within a node are ordered in monotonic
5421 * increasing memory addresses so that the "highest" populated zone is used
5423 static void __init
find_usable_zone_for_movable(void)
5426 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5427 if (zone_index
== ZONE_MOVABLE
)
5430 if (arch_zone_highest_possible_pfn
[zone_index
] >
5431 arch_zone_lowest_possible_pfn
[zone_index
])
5435 VM_BUG_ON(zone_index
== -1);
5436 movable_zone
= zone_index
;
5440 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5441 * because it is sized independent of architecture. Unlike the other zones,
5442 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5443 * in each node depending on the size of each node and how evenly kernelcore
5444 * is distributed. This helper function adjusts the zone ranges
5445 * provided by the architecture for a given node by using the end of the
5446 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5447 * zones within a node are in order of monotonic increases memory addresses
5449 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5450 unsigned long zone_type
,
5451 unsigned long node_start_pfn
,
5452 unsigned long node_end_pfn
,
5453 unsigned long *zone_start_pfn
,
5454 unsigned long *zone_end_pfn
)
5456 /* Only adjust if ZONE_MOVABLE is on this node */
5457 if (zone_movable_pfn
[nid
]) {
5458 /* Size ZONE_MOVABLE */
5459 if (zone_type
== ZONE_MOVABLE
) {
5460 *zone_start_pfn
= zone_movable_pfn
[nid
];
5461 *zone_end_pfn
= min(node_end_pfn
,
5462 arch_zone_highest_possible_pfn
[movable_zone
]);
5464 /* Adjust for ZONE_MOVABLE starting within this range */
5465 } else if (!mirrored_kernelcore
&&
5466 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5467 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5468 *zone_end_pfn
= zone_movable_pfn
[nid
];
5470 /* Check if this whole range is within ZONE_MOVABLE */
5471 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5472 *zone_start_pfn
= *zone_end_pfn
;
5477 * Return the number of pages a zone spans in a node, including holes
5478 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5480 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5481 unsigned long zone_type
,
5482 unsigned long node_start_pfn
,
5483 unsigned long node_end_pfn
,
5484 unsigned long *zone_start_pfn
,
5485 unsigned long *zone_end_pfn
,
5486 unsigned long *ignored
)
5488 /* When hotadd a new node from cpu_up(), the node should be empty */
5489 if (!node_start_pfn
&& !node_end_pfn
)
5492 /* Get the start and end of the zone */
5493 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5494 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5495 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5496 node_start_pfn
, node_end_pfn
,
5497 zone_start_pfn
, zone_end_pfn
);
5499 /* Check that this node has pages within the zone's required range */
5500 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5503 /* Move the zone boundaries inside the node if necessary */
5504 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5505 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5507 /* Return the spanned pages */
5508 return *zone_end_pfn
- *zone_start_pfn
;
5512 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5513 * then all holes in the requested range will be accounted for.
5515 unsigned long __meminit
__absent_pages_in_range(int nid
,
5516 unsigned long range_start_pfn
,
5517 unsigned long range_end_pfn
)
5519 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5520 unsigned long start_pfn
, end_pfn
;
5523 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5524 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5525 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5526 nr_absent
-= end_pfn
- start_pfn
;
5532 * absent_pages_in_range - Return number of page frames in holes within a range
5533 * @start_pfn: The start PFN to start searching for holes
5534 * @end_pfn: The end PFN to stop searching for holes
5536 * It returns the number of pages frames in memory holes within a range.
5538 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5539 unsigned long end_pfn
)
5541 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5544 /* Return the number of page frames in holes in a zone on a node */
5545 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5546 unsigned long zone_type
,
5547 unsigned long node_start_pfn
,
5548 unsigned long node_end_pfn
,
5549 unsigned long *ignored
)
5551 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5552 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5553 unsigned long zone_start_pfn
, zone_end_pfn
;
5554 unsigned long nr_absent
;
5556 /* When hotadd a new node from cpu_up(), the node should be empty */
5557 if (!node_start_pfn
&& !node_end_pfn
)
5560 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5561 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5563 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5564 node_start_pfn
, node_end_pfn
,
5565 &zone_start_pfn
, &zone_end_pfn
);
5566 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5569 * ZONE_MOVABLE handling.
5570 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5573 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5574 unsigned long start_pfn
, end_pfn
;
5575 struct memblock_region
*r
;
5577 for_each_memblock(memory
, r
) {
5578 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5579 zone_start_pfn
, zone_end_pfn
);
5580 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5581 zone_start_pfn
, zone_end_pfn
);
5583 if (zone_type
== ZONE_MOVABLE
&&
5584 memblock_is_mirror(r
))
5585 nr_absent
+= end_pfn
- start_pfn
;
5587 if (zone_type
== ZONE_NORMAL
&&
5588 !memblock_is_mirror(r
))
5589 nr_absent
+= end_pfn
- start_pfn
;
5596 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5597 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5598 unsigned long zone_type
,
5599 unsigned long node_start_pfn
,
5600 unsigned long node_end_pfn
,
5601 unsigned long *zone_start_pfn
,
5602 unsigned long *zone_end_pfn
,
5603 unsigned long *zones_size
)
5607 *zone_start_pfn
= node_start_pfn
;
5608 for (zone
= 0; zone
< zone_type
; zone
++)
5609 *zone_start_pfn
+= zones_size
[zone
];
5611 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5613 return zones_size
[zone_type
];
5616 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5617 unsigned long zone_type
,
5618 unsigned long node_start_pfn
,
5619 unsigned long node_end_pfn
,
5620 unsigned long *zholes_size
)
5625 return zholes_size
[zone_type
];
5628 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5630 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5631 unsigned long node_start_pfn
,
5632 unsigned long node_end_pfn
,
5633 unsigned long *zones_size
,
5634 unsigned long *zholes_size
)
5636 unsigned long realtotalpages
= 0, totalpages
= 0;
5639 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5640 struct zone
*zone
= pgdat
->node_zones
+ i
;
5641 unsigned long zone_start_pfn
, zone_end_pfn
;
5642 unsigned long size
, real_size
;
5644 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5650 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5651 node_start_pfn
, node_end_pfn
,
5654 zone
->zone_start_pfn
= zone_start_pfn
;
5656 zone
->zone_start_pfn
= 0;
5657 zone
->spanned_pages
= size
;
5658 zone
->present_pages
= real_size
;
5661 realtotalpages
+= real_size
;
5664 pgdat
->node_spanned_pages
= totalpages
;
5665 pgdat
->node_present_pages
= realtotalpages
;
5666 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5670 #ifndef CONFIG_SPARSEMEM
5672 * Calculate the size of the zone->blockflags rounded to an unsigned long
5673 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5674 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5675 * round what is now in bits to nearest long in bits, then return it in
5678 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5680 unsigned long usemapsize
;
5682 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5683 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5684 usemapsize
= usemapsize
>> pageblock_order
;
5685 usemapsize
*= NR_PAGEBLOCK_BITS
;
5686 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5688 return usemapsize
/ 8;
5691 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5693 unsigned long zone_start_pfn
,
5694 unsigned long zonesize
)
5696 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5697 zone
->pageblock_flags
= NULL
;
5699 zone
->pageblock_flags
=
5700 memblock_virt_alloc_node_nopanic(usemapsize
,
5704 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5705 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5706 #endif /* CONFIG_SPARSEMEM */
5708 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5710 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5711 void __paginginit
set_pageblock_order(void)
5715 /* Check that pageblock_nr_pages has not already been setup */
5716 if (pageblock_order
)
5719 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5720 order
= HUGETLB_PAGE_ORDER
;
5722 order
= MAX_ORDER
- 1;
5725 * Assume the largest contiguous order of interest is a huge page.
5726 * This value may be variable depending on boot parameters on IA64 and
5729 pageblock_order
= order
;
5731 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5734 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5735 * is unused as pageblock_order is set at compile-time. See
5736 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5739 void __paginginit
set_pageblock_order(void)
5743 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5745 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5746 unsigned long present_pages
)
5748 unsigned long pages
= spanned_pages
;
5751 * Provide a more accurate estimation if there are holes within
5752 * the zone and SPARSEMEM is in use. If there are holes within the
5753 * zone, each populated memory region may cost us one or two extra
5754 * memmap pages due to alignment because memmap pages for each
5755 * populated regions may not naturally algined on page boundary.
5756 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5758 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5759 IS_ENABLED(CONFIG_SPARSEMEM
))
5760 pages
= present_pages
;
5762 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5766 * Set up the zone data structures:
5767 * - mark all pages reserved
5768 * - mark all memory queues empty
5769 * - clear the memory bitmaps
5771 * NOTE: pgdat should get zeroed by caller.
5773 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5776 int nid
= pgdat
->node_id
;
5779 pgdat_resize_init(pgdat
);
5780 #ifdef CONFIG_NUMA_BALANCING
5781 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5782 pgdat
->numabalancing_migrate_nr_pages
= 0;
5783 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5785 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5786 spin_lock_init(&pgdat
->split_queue_lock
);
5787 INIT_LIST_HEAD(&pgdat
->split_queue
);
5788 pgdat
->split_queue_len
= 0;
5790 init_waitqueue_head(&pgdat
->kswapd_wait
);
5791 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5792 #ifdef CONFIG_COMPACTION
5793 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5795 pgdat_page_ext_init(pgdat
);
5796 spin_lock_init(&pgdat
->lru_lock
);
5797 lruvec_init(node_lruvec(pgdat
));
5799 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5800 struct zone
*zone
= pgdat
->node_zones
+ j
;
5801 unsigned long size
, realsize
, freesize
, memmap_pages
;
5802 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5804 size
= zone
->spanned_pages
;
5805 realsize
= freesize
= zone
->present_pages
;
5808 * Adjust freesize so that it accounts for how much memory
5809 * is used by this zone for memmap. This affects the watermark
5810 * and per-cpu initialisations
5812 memmap_pages
= calc_memmap_size(size
, realsize
);
5813 if (!is_highmem_idx(j
)) {
5814 if (freesize
>= memmap_pages
) {
5815 freesize
-= memmap_pages
;
5818 " %s zone: %lu pages used for memmap\n",
5819 zone_names
[j
], memmap_pages
);
5821 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5822 zone_names
[j
], memmap_pages
, freesize
);
5825 /* Account for reserved pages */
5826 if (j
== 0 && freesize
> nr_memory_reserve
) {
5827 freesize
-= nr_memory_reserve
;
5828 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5829 zone_names
[0], nr_memory_reserve
);
5832 if (!is_highmem_idx(j
))
5833 nr_kernel_pages
+= freesize
;
5834 /* Charge for highmem memmap if there are enough kernel pages */
5835 else if (nr_kernel_pages
> memmap_pages
* 2)
5836 nr_kernel_pages
-= memmap_pages
;
5837 nr_all_pages
+= freesize
;
5840 * Set an approximate value for lowmem here, it will be adjusted
5841 * when the bootmem allocator frees pages into the buddy system.
5842 * And all highmem pages will be managed by the buddy system.
5844 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5848 zone
->name
= zone_names
[j
];
5849 zone
->zone_pgdat
= pgdat
;
5850 spin_lock_init(&zone
->lock
);
5851 zone_seqlock_init(zone
);
5852 zone_pcp_init(zone
);
5857 set_pageblock_order();
5858 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5859 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5861 memmap_init(size
, nid
, j
, zone_start_pfn
);
5865 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
5867 unsigned long __maybe_unused start
= 0;
5868 unsigned long __maybe_unused offset
= 0;
5870 /* Skip empty nodes */
5871 if (!pgdat
->node_spanned_pages
)
5874 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5875 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5876 offset
= pgdat
->node_start_pfn
- start
;
5877 /* ia64 gets its own node_mem_map, before this, without bootmem */
5878 if (!pgdat
->node_mem_map
) {
5879 unsigned long size
, end
;
5883 * The zone's endpoints aren't required to be MAX_ORDER
5884 * aligned but the node_mem_map endpoints must be in order
5885 * for the buddy allocator to function correctly.
5887 end
= pgdat_end_pfn(pgdat
);
5888 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5889 size
= (end
- start
) * sizeof(struct page
);
5890 map
= alloc_remap(pgdat
->node_id
, size
);
5892 map
= memblock_virt_alloc_node_nopanic(size
,
5894 pgdat
->node_mem_map
= map
+ offset
;
5896 #ifndef CONFIG_NEED_MULTIPLE_NODES
5898 * With no DISCONTIG, the global mem_map is just set as node 0's
5900 if (pgdat
== NODE_DATA(0)) {
5901 mem_map
= NODE_DATA(0)->node_mem_map
;
5902 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5903 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5905 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5908 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5911 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5912 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5914 pg_data_t
*pgdat
= NODE_DATA(nid
);
5915 unsigned long start_pfn
= 0;
5916 unsigned long end_pfn
= 0;
5918 /* pg_data_t should be reset to zero when it's allocated */
5919 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
5921 reset_deferred_meminit(pgdat
);
5922 pgdat
->node_id
= nid
;
5923 pgdat
->node_start_pfn
= node_start_pfn
;
5924 pgdat
->per_cpu_nodestats
= NULL
;
5925 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5926 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5927 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5928 (u64
)start_pfn
<< PAGE_SHIFT
,
5929 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5931 start_pfn
= node_start_pfn
;
5933 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5934 zones_size
, zholes_size
);
5936 alloc_node_mem_map(pgdat
);
5937 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5938 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5939 nid
, (unsigned long)pgdat
,
5940 (unsigned long)pgdat
->node_mem_map
);
5943 free_area_init_core(pgdat
);
5946 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5948 #if MAX_NUMNODES > 1
5950 * Figure out the number of possible node ids.
5952 void __init
setup_nr_node_ids(void)
5954 unsigned int highest
;
5956 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5957 nr_node_ids
= highest
+ 1;
5962 * node_map_pfn_alignment - determine the maximum internode alignment
5964 * This function should be called after node map is populated and sorted.
5965 * It calculates the maximum power of two alignment which can distinguish
5968 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5969 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5970 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5971 * shifted, 1GiB is enough and this function will indicate so.
5973 * This is used to test whether pfn -> nid mapping of the chosen memory
5974 * model has fine enough granularity to avoid incorrect mapping for the
5975 * populated node map.
5977 * Returns the determined alignment in pfn's. 0 if there is no alignment
5978 * requirement (single node).
5980 unsigned long __init
node_map_pfn_alignment(void)
5982 unsigned long accl_mask
= 0, last_end
= 0;
5983 unsigned long start
, end
, mask
;
5987 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5988 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5995 * Start with a mask granular enough to pin-point to the
5996 * start pfn and tick off bits one-by-one until it becomes
5997 * too coarse to separate the current node from the last.
5999 mask
= ~((1 << __ffs(start
)) - 1);
6000 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6003 /* accumulate all internode masks */
6007 /* convert mask to number of pages */
6008 return ~accl_mask
+ 1;
6011 /* Find the lowest pfn for a node */
6012 static unsigned long __init
find_min_pfn_for_node(int nid
)
6014 unsigned long min_pfn
= ULONG_MAX
;
6015 unsigned long start_pfn
;
6018 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6019 min_pfn
= min(min_pfn
, start_pfn
);
6021 if (min_pfn
== ULONG_MAX
) {
6022 pr_warn("Could not find start_pfn for node %d\n", nid
);
6030 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6032 * It returns the minimum PFN based on information provided via
6033 * memblock_set_node().
6035 unsigned long __init
find_min_pfn_with_active_regions(void)
6037 return find_min_pfn_for_node(MAX_NUMNODES
);
6041 * early_calculate_totalpages()
6042 * Sum pages in active regions for movable zone.
6043 * Populate N_MEMORY for calculating usable_nodes.
6045 static unsigned long __init
early_calculate_totalpages(void)
6047 unsigned long totalpages
= 0;
6048 unsigned long start_pfn
, end_pfn
;
6051 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6052 unsigned long pages
= end_pfn
- start_pfn
;
6054 totalpages
+= pages
;
6056 node_set_state(nid
, N_MEMORY
);
6062 * Find the PFN the Movable zone begins in each node. Kernel memory
6063 * is spread evenly between nodes as long as the nodes have enough
6064 * memory. When they don't, some nodes will have more kernelcore than
6067 static void __init
find_zone_movable_pfns_for_nodes(void)
6070 unsigned long usable_startpfn
;
6071 unsigned long kernelcore_node
, kernelcore_remaining
;
6072 /* save the state before borrow the nodemask */
6073 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6074 unsigned long totalpages
= early_calculate_totalpages();
6075 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6076 struct memblock_region
*r
;
6078 /* Need to find movable_zone earlier when movable_node is specified. */
6079 find_usable_zone_for_movable();
6082 * If movable_node is specified, ignore kernelcore and movablecore
6085 if (movable_node_is_enabled()) {
6086 for_each_memblock(memory
, r
) {
6087 if (!memblock_is_hotpluggable(r
))
6092 usable_startpfn
= PFN_DOWN(r
->base
);
6093 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6094 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6102 * If kernelcore=mirror is specified, ignore movablecore option
6104 if (mirrored_kernelcore
) {
6105 bool mem_below_4gb_not_mirrored
= false;
6107 for_each_memblock(memory
, r
) {
6108 if (memblock_is_mirror(r
))
6113 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6115 if (usable_startpfn
< 0x100000) {
6116 mem_below_4gb_not_mirrored
= true;
6120 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6121 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6125 if (mem_below_4gb_not_mirrored
)
6126 pr_warn("This configuration results in unmirrored kernel memory.");
6132 * If movablecore=nn[KMG] was specified, calculate what size of
6133 * kernelcore that corresponds so that memory usable for
6134 * any allocation type is evenly spread. If both kernelcore
6135 * and movablecore are specified, then the value of kernelcore
6136 * will be used for required_kernelcore if it's greater than
6137 * what movablecore would have allowed.
6139 if (required_movablecore
) {
6140 unsigned long corepages
;
6143 * Round-up so that ZONE_MOVABLE is at least as large as what
6144 * was requested by the user
6146 required_movablecore
=
6147 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6148 required_movablecore
= min(totalpages
, required_movablecore
);
6149 corepages
= totalpages
- required_movablecore
;
6151 required_kernelcore
= max(required_kernelcore
, corepages
);
6155 * If kernelcore was not specified or kernelcore size is larger
6156 * than totalpages, there is no ZONE_MOVABLE.
6158 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6161 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6162 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6165 /* Spread kernelcore memory as evenly as possible throughout nodes */
6166 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6167 for_each_node_state(nid
, N_MEMORY
) {
6168 unsigned long start_pfn
, end_pfn
;
6171 * Recalculate kernelcore_node if the division per node
6172 * now exceeds what is necessary to satisfy the requested
6173 * amount of memory for the kernel
6175 if (required_kernelcore
< kernelcore_node
)
6176 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6179 * As the map is walked, we track how much memory is usable
6180 * by the kernel using kernelcore_remaining. When it is
6181 * 0, the rest of the node is usable by ZONE_MOVABLE
6183 kernelcore_remaining
= kernelcore_node
;
6185 /* Go through each range of PFNs within this node */
6186 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6187 unsigned long size_pages
;
6189 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6190 if (start_pfn
>= end_pfn
)
6193 /* Account for what is only usable for kernelcore */
6194 if (start_pfn
< usable_startpfn
) {
6195 unsigned long kernel_pages
;
6196 kernel_pages
= min(end_pfn
, usable_startpfn
)
6199 kernelcore_remaining
-= min(kernel_pages
,
6200 kernelcore_remaining
);
6201 required_kernelcore
-= min(kernel_pages
,
6202 required_kernelcore
);
6204 /* Continue if range is now fully accounted */
6205 if (end_pfn
<= usable_startpfn
) {
6208 * Push zone_movable_pfn to the end so
6209 * that if we have to rebalance
6210 * kernelcore across nodes, we will
6211 * not double account here
6213 zone_movable_pfn
[nid
] = end_pfn
;
6216 start_pfn
= usable_startpfn
;
6220 * The usable PFN range for ZONE_MOVABLE is from
6221 * start_pfn->end_pfn. Calculate size_pages as the
6222 * number of pages used as kernelcore
6224 size_pages
= end_pfn
- start_pfn
;
6225 if (size_pages
> kernelcore_remaining
)
6226 size_pages
= kernelcore_remaining
;
6227 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6230 * Some kernelcore has been met, update counts and
6231 * break if the kernelcore for this node has been
6234 required_kernelcore
-= min(required_kernelcore
,
6236 kernelcore_remaining
-= size_pages
;
6237 if (!kernelcore_remaining
)
6243 * If there is still required_kernelcore, we do another pass with one
6244 * less node in the count. This will push zone_movable_pfn[nid] further
6245 * along on the nodes that still have memory until kernelcore is
6249 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6253 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6254 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6255 zone_movable_pfn
[nid
] =
6256 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6259 /* restore the node_state */
6260 node_states
[N_MEMORY
] = saved_node_state
;
6263 /* Any regular or high memory on that node ? */
6264 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6266 enum zone_type zone_type
;
6268 if (N_MEMORY
== N_NORMAL_MEMORY
)
6271 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6272 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6273 if (populated_zone(zone
)) {
6274 node_set_state(nid
, N_HIGH_MEMORY
);
6275 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6276 zone_type
<= ZONE_NORMAL
)
6277 node_set_state(nid
, N_NORMAL_MEMORY
);
6284 * free_area_init_nodes - Initialise all pg_data_t and zone data
6285 * @max_zone_pfn: an array of max PFNs for each zone
6287 * This will call free_area_init_node() for each active node in the system.
6288 * Using the page ranges provided by memblock_set_node(), the size of each
6289 * zone in each node and their holes is calculated. If the maximum PFN
6290 * between two adjacent zones match, it is assumed that the zone is empty.
6291 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6292 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6293 * starts where the previous one ended. For example, ZONE_DMA32 starts
6294 * at arch_max_dma_pfn.
6296 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6298 unsigned long start_pfn
, end_pfn
;
6301 /* Record where the zone boundaries are */
6302 memset(arch_zone_lowest_possible_pfn
, 0,
6303 sizeof(arch_zone_lowest_possible_pfn
));
6304 memset(arch_zone_highest_possible_pfn
, 0,
6305 sizeof(arch_zone_highest_possible_pfn
));
6307 start_pfn
= find_min_pfn_with_active_regions();
6309 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6310 if (i
== ZONE_MOVABLE
)
6313 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6314 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6315 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6317 start_pfn
= end_pfn
;
6319 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6320 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6322 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6323 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6324 find_zone_movable_pfns_for_nodes();
6326 /* Print out the zone ranges */
6327 pr_info("Zone ranges:\n");
6328 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6329 if (i
== ZONE_MOVABLE
)
6331 pr_info(" %-8s ", zone_names
[i
]);
6332 if (arch_zone_lowest_possible_pfn
[i
] ==
6333 arch_zone_highest_possible_pfn
[i
])
6336 pr_cont("[mem %#018Lx-%#018Lx]\n",
6337 (u64
)arch_zone_lowest_possible_pfn
[i
]
6339 ((u64
)arch_zone_highest_possible_pfn
[i
]
6340 << PAGE_SHIFT
) - 1);
6343 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6344 pr_info("Movable zone start for each node\n");
6345 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6346 if (zone_movable_pfn
[i
])
6347 pr_info(" Node %d: %#018Lx\n", i
,
6348 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6351 /* Print out the early node map */
6352 pr_info("Early memory node ranges\n");
6353 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6354 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6355 (u64
)start_pfn
<< PAGE_SHIFT
,
6356 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6358 /* Initialise every node */
6359 mminit_verify_pageflags_layout();
6360 setup_nr_node_ids();
6361 for_each_online_node(nid
) {
6362 pg_data_t
*pgdat
= NODE_DATA(nid
);
6363 free_area_init_node(nid
, NULL
,
6364 find_min_pfn_for_node(nid
), NULL
);
6366 /* Any memory on that node */
6367 if (pgdat
->node_present_pages
)
6368 node_set_state(nid
, N_MEMORY
);
6369 check_for_memory(pgdat
, nid
);
6373 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6375 unsigned long long coremem
;
6379 coremem
= memparse(p
, &p
);
6380 *core
= coremem
>> PAGE_SHIFT
;
6382 /* Paranoid check that UL is enough for the coremem value */
6383 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6389 * kernelcore=size sets the amount of memory for use for allocations that
6390 * cannot be reclaimed or migrated.
6392 static int __init
cmdline_parse_kernelcore(char *p
)
6394 /* parse kernelcore=mirror */
6395 if (parse_option_str(p
, "mirror")) {
6396 mirrored_kernelcore
= true;
6400 return cmdline_parse_core(p
, &required_kernelcore
);
6404 * movablecore=size sets the amount of memory for use for allocations that
6405 * can be reclaimed or migrated.
6407 static int __init
cmdline_parse_movablecore(char *p
)
6409 return cmdline_parse_core(p
, &required_movablecore
);
6412 early_param("kernelcore", cmdline_parse_kernelcore
);
6413 early_param("movablecore", cmdline_parse_movablecore
);
6415 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6417 void adjust_managed_page_count(struct page
*page
, long count
)
6419 spin_lock(&managed_page_count_lock
);
6420 page_zone(page
)->managed_pages
+= count
;
6421 totalram_pages
+= count
;
6422 #ifdef CONFIG_HIGHMEM
6423 if (PageHighMem(page
))
6424 totalhigh_pages
+= count
;
6426 spin_unlock(&managed_page_count_lock
);
6428 EXPORT_SYMBOL(adjust_managed_page_count
);
6430 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6433 unsigned long pages
= 0;
6435 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6436 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6437 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6438 if ((unsigned int)poison
<= 0xFF)
6439 memset(pos
, poison
, PAGE_SIZE
);
6440 free_reserved_page(virt_to_page(pos
));
6444 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6445 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6449 EXPORT_SYMBOL(free_reserved_area
);
6451 #ifdef CONFIG_HIGHMEM
6452 void free_highmem_page(struct page
*page
)
6454 __free_reserved_page(page
);
6456 page_zone(page
)->managed_pages
++;
6462 void __init
mem_init_print_info(const char *str
)
6464 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6465 unsigned long init_code_size
, init_data_size
;
6467 physpages
= get_num_physpages();
6468 codesize
= _etext
- _stext
;
6469 datasize
= _edata
- _sdata
;
6470 rosize
= __end_rodata
- __start_rodata
;
6471 bss_size
= __bss_stop
- __bss_start
;
6472 init_data_size
= __init_end
- __init_begin
;
6473 init_code_size
= _einittext
- _sinittext
;
6476 * Detect special cases and adjust section sizes accordingly:
6477 * 1) .init.* may be embedded into .data sections
6478 * 2) .init.text.* may be out of [__init_begin, __init_end],
6479 * please refer to arch/tile/kernel/vmlinux.lds.S.
6480 * 3) .rodata.* may be embedded into .text or .data sections.
6482 #define adj_init_size(start, end, size, pos, adj) \
6484 if (start <= pos && pos < end && size > adj) \
6488 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6489 _sinittext
, init_code_size
);
6490 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6491 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6492 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6493 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6495 #undef adj_init_size
6497 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6498 #ifdef CONFIG_HIGHMEM
6502 nr_free_pages() << (PAGE_SHIFT
- 10),
6503 physpages
<< (PAGE_SHIFT
- 10),
6504 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6505 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6506 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6507 totalcma_pages
<< (PAGE_SHIFT
- 10),
6508 #ifdef CONFIG_HIGHMEM
6509 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6511 str
? ", " : "", str
? str
: "");
6515 * set_memory_reserve - set number of pages reserved in the first zone
6516 * @nr_reserve: The number of pages to mark reserved
6517 * @inc: true increment to existing value; false set new value.
6519 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6520 * In the DMA zone, a significant percentage may be consumed by kernel image
6521 * and other unfreeable allocations which can skew the watermarks badly. This
6522 * function may optionally be used to account for unfreeable pages in the
6523 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6524 * smaller per-cpu batchsize.
6526 void __init
set_memory_reserve(unsigned long nr_reserve
, bool inc
)
6529 nr_memory_reserve
+= nr_reserve
;
6531 nr_memory_reserve
= nr_reserve
;
6534 void __init
free_area_init(unsigned long *zones_size
)
6536 free_area_init_node(0, zones_size
,
6537 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6540 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6541 unsigned long action
, void *hcpu
)
6543 int cpu
= (unsigned long)hcpu
;
6545 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6546 lru_add_drain_cpu(cpu
);
6550 * Spill the event counters of the dead processor
6551 * into the current processors event counters.
6552 * This artificially elevates the count of the current
6555 vm_events_fold_cpu(cpu
);
6558 * Zero the differential counters of the dead processor
6559 * so that the vm statistics are consistent.
6561 * This is only okay since the processor is dead and cannot
6562 * race with what we are doing.
6564 cpu_vm_stats_fold(cpu
);
6569 void __init
page_alloc_init(void)
6571 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6575 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6576 * or min_free_kbytes changes.
6578 static void calculate_totalreserve_pages(void)
6580 struct pglist_data
*pgdat
;
6581 unsigned long reserve_pages
= 0;
6582 enum zone_type i
, j
;
6584 for_each_online_pgdat(pgdat
) {
6586 pgdat
->totalreserve_pages
= 0;
6588 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6589 struct zone
*zone
= pgdat
->node_zones
+ i
;
6592 /* Find valid and maximum lowmem_reserve in the zone */
6593 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6594 if (zone
->lowmem_reserve
[j
] > max
)
6595 max
= zone
->lowmem_reserve
[j
];
6598 /* we treat the high watermark as reserved pages. */
6599 max
+= high_wmark_pages(zone
);
6601 if (max
> zone
->managed_pages
)
6602 max
= zone
->managed_pages
;
6604 pgdat
->totalreserve_pages
+= max
;
6606 reserve_pages
+= max
;
6609 totalreserve_pages
= reserve_pages
;
6613 * setup_per_zone_lowmem_reserve - called whenever
6614 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6615 * has a correct pages reserved value, so an adequate number of
6616 * pages are left in the zone after a successful __alloc_pages().
6618 static void setup_per_zone_lowmem_reserve(void)
6620 struct pglist_data
*pgdat
;
6621 enum zone_type j
, idx
;
6623 for_each_online_pgdat(pgdat
) {
6624 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6625 struct zone
*zone
= pgdat
->node_zones
+ j
;
6626 unsigned long managed_pages
= zone
->managed_pages
;
6628 zone
->lowmem_reserve
[j
] = 0;
6632 struct zone
*lower_zone
;
6636 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6637 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6639 lower_zone
= pgdat
->node_zones
+ idx
;
6640 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6641 sysctl_lowmem_reserve_ratio
[idx
];
6642 managed_pages
+= lower_zone
->managed_pages
;
6647 /* update totalreserve_pages */
6648 calculate_totalreserve_pages();
6651 static void __setup_per_zone_wmarks(void)
6653 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6654 unsigned long lowmem_pages
= 0;
6656 unsigned long flags
;
6658 /* Calculate total number of !ZONE_HIGHMEM pages */
6659 for_each_zone(zone
) {
6660 if (!is_highmem(zone
))
6661 lowmem_pages
+= zone
->managed_pages
;
6664 for_each_zone(zone
) {
6667 spin_lock_irqsave(&zone
->lock
, flags
);
6668 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6669 do_div(tmp
, lowmem_pages
);
6670 if (is_highmem(zone
)) {
6672 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6673 * need highmem pages, so cap pages_min to a small
6676 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6677 * deltas control asynch page reclaim, and so should
6678 * not be capped for highmem.
6680 unsigned long min_pages
;
6682 min_pages
= zone
->managed_pages
/ 1024;
6683 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6684 zone
->watermark
[WMARK_MIN
] = min_pages
;
6687 * If it's a lowmem zone, reserve a number of pages
6688 * proportionate to the zone's size.
6690 zone
->watermark
[WMARK_MIN
] = tmp
;
6694 * Set the kswapd watermarks distance according to the
6695 * scale factor in proportion to available memory, but
6696 * ensure a minimum size on small systems.
6698 tmp
= max_t(u64
, tmp
>> 2,
6699 mult_frac(zone
->managed_pages
,
6700 watermark_scale_factor
, 10000));
6702 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6703 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6705 spin_unlock_irqrestore(&zone
->lock
, flags
);
6708 /* update totalreserve_pages */
6709 calculate_totalreserve_pages();
6713 * setup_per_zone_wmarks - called when min_free_kbytes changes
6714 * or when memory is hot-{added|removed}
6716 * Ensures that the watermark[min,low,high] values for each zone are set
6717 * correctly with respect to min_free_kbytes.
6719 void setup_per_zone_wmarks(void)
6721 mutex_lock(&zonelists_mutex
);
6722 __setup_per_zone_wmarks();
6723 mutex_unlock(&zonelists_mutex
);
6727 * Initialise min_free_kbytes.
6729 * For small machines we want it small (128k min). For large machines
6730 * we want it large (64MB max). But it is not linear, because network
6731 * bandwidth does not increase linearly with machine size. We use
6733 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6734 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6750 int __meminit
init_per_zone_wmark_min(void)
6752 unsigned long lowmem_kbytes
;
6753 int new_min_free_kbytes
;
6755 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6756 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6758 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6759 min_free_kbytes
= new_min_free_kbytes
;
6760 if (min_free_kbytes
< 128)
6761 min_free_kbytes
= 128;
6762 if (min_free_kbytes
> 65536)
6763 min_free_kbytes
= 65536;
6765 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6766 new_min_free_kbytes
, user_min_free_kbytes
);
6768 setup_per_zone_wmarks();
6769 refresh_zone_stat_thresholds();
6770 setup_per_zone_lowmem_reserve();
6773 setup_min_unmapped_ratio();
6774 setup_min_slab_ratio();
6779 core_initcall(init_per_zone_wmark_min
)
6782 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6783 * that we can call two helper functions whenever min_free_kbytes
6786 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6787 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6791 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6796 user_min_free_kbytes
= min_free_kbytes
;
6797 setup_per_zone_wmarks();
6802 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6803 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6807 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6812 setup_per_zone_wmarks();
6818 static void setup_min_unmapped_ratio(void)
6823 for_each_online_pgdat(pgdat
)
6824 pgdat
->min_unmapped_pages
= 0;
6827 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
6828 sysctl_min_unmapped_ratio
) / 100;
6832 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6833 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6837 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6841 setup_min_unmapped_ratio();
6846 static void setup_min_slab_ratio(void)
6851 for_each_online_pgdat(pgdat
)
6852 pgdat
->min_slab_pages
= 0;
6855 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
6856 sysctl_min_slab_ratio
) / 100;
6859 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6860 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6864 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6868 setup_min_slab_ratio();
6875 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6876 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6877 * whenever sysctl_lowmem_reserve_ratio changes.
6879 * The reserve ratio obviously has absolutely no relation with the
6880 * minimum watermarks. The lowmem reserve ratio can only make sense
6881 * if in function of the boot time zone sizes.
6883 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6884 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6886 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6887 setup_per_zone_lowmem_reserve();
6892 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6893 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6894 * pagelist can have before it gets flushed back to buddy allocator.
6896 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6897 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6900 int old_percpu_pagelist_fraction
;
6903 mutex_lock(&pcp_batch_high_lock
);
6904 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6906 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6907 if (!write
|| ret
< 0)
6910 /* Sanity checking to avoid pcp imbalance */
6911 if (percpu_pagelist_fraction
&&
6912 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6913 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6919 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6922 for_each_populated_zone(zone
) {
6925 for_each_possible_cpu(cpu
)
6926 pageset_set_high_and_batch(zone
,
6927 per_cpu_ptr(zone
->pageset
, cpu
));
6930 mutex_unlock(&pcp_batch_high_lock
);
6935 int hashdist
= HASHDIST_DEFAULT
;
6937 static int __init
set_hashdist(char *str
)
6941 hashdist
= simple_strtoul(str
, &str
, 0);
6944 __setup("hashdist=", set_hashdist
);
6948 * allocate a large system hash table from bootmem
6949 * - it is assumed that the hash table must contain an exact power-of-2
6950 * quantity of entries
6951 * - limit is the number of hash buckets, not the total allocation size
6953 void *__init
alloc_large_system_hash(const char *tablename
,
6954 unsigned long bucketsize
,
6955 unsigned long numentries
,
6958 unsigned int *_hash_shift
,
6959 unsigned int *_hash_mask
,
6960 unsigned long low_limit
,
6961 unsigned long high_limit
)
6963 unsigned long long max
= high_limit
;
6964 unsigned long log2qty
, size
;
6967 /* allow the kernel cmdline to have a say */
6969 /* round applicable memory size up to nearest megabyte */
6970 numentries
= nr_kernel_pages
;
6972 /* It isn't necessary when PAGE_SIZE >= 1MB */
6973 if (PAGE_SHIFT
< 20)
6974 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6976 /* limit to 1 bucket per 2^scale bytes of low memory */
6977 if (scale
> PAGE_SHIFT
)
6978 numentries
>>= (scale
- PAGE_SHIFT
);
6980 numentries
<<= (PAGE_SHIFT
- scale
);
6982 /* Make sure we've got at least a 0-order allocation.. */
6983 if (unlikely(flags
& HASH_SMALL
)) {
6984 /* Makes no sense without HASH_EARLY */
6985 WARN_ON(!(flags
& HASH_EARLY
));
6986 if (!(numentries
>> *_hash_shift
)) {
6987 numentries
= 1UL << *_hash_shift
;
6988 BUG_ON(!numentries
);
6990 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6991 numentries
= PAGE_SIZE
/ bucketsize
;
6993 numentries
= roundup_pow_of_two(numentries
);
6995 /* limit allocation size to 1/16 total memory by default */
6997 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6998 do_div(max
, bucketsize
);
7000 max
= min(max
, 0x80000000ULL
);
7002 if (numentries
< low_limit
)
7003 numentries
= low_limit
;
7004 if (numentries
> max
)
7007 log2qty
= ilog2(numentries
);
7010 size
= bucketsize
<< log2qty
;
7011 if (flags
& HASH_EARLY
)
7012 table
= memblock_virt_alloc_nopanic(size
, 0);
7014 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
7017 * If bucketsize is not a power-of-two, we may free
7018 * some pages at the end of hash table which
7019 * alloc_pages_exact() automatically does
7021 if (get_order(size
) < MAX_ORDER
) {
7022 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
7023 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
7026 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7029 panic("Failed to allocate %s hash table\n", tablename
);
7031 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7032 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7035 *_hash_shift
= log2qty
;
7037 *_hash_mask
= (1 << log2qty
) - 1;
7043 * This function checks whether pageblock includes unmovable pages or not.
7044 * If @count is not zero, it is okay to include less @count unmovable pages
7046 * PageLRU check without isolation or lru_lock could race so that
7047 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7048 * expect this function should be exact.
7050 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7051 bool skip_hwpoisoned_pages
)
7053 unsigned long pfn
, iter
, found
;
7057 * For avoiding noise data, lru_add_drain_all() should be called
7058 * If ZONE_MOVABLE, the zone never contains unmovable pages
7060 if (zone_idx(zone
) == ZONE_MOVABLE
)
7062 mt
= get_pageblock_migratetype(page
);
7063 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7066 pfn
= page_to_pfn(page
);
7067 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7068 unsigned long check
= pfn
+ iter
;
7070 if (!pfn_valid_within(check
))
7073 page
= pfn_to_page(check
);
7076 * Hugepages are not in LRU lists, but they're movable.
7077 * We need not scan over tail pages bacause we don't
7078 * handle each tail page individually in migration.
7080 if (PageHuge(page
)) {
7081 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7086 * We can't use page_count without pin a page
7087 * because another CPU can free compound page.
7088 * This check already skips compound tails of THP
7089 * because their page->_refcount is zero at all time.
7091 if (!page_ref_count(page
)) {
7092 if (PageBuddy(page
))
7093 iter
+= (1 << page_order(page
)) - 1;
7098 * The HWPoisoned page may be not in buddy system, and
7099 * page_count() is not 0.
7101 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7107 * If there are RECLAIMABLE pages, we need to check
7108 * it. But now, memory offline itself doesn't call
7109 * shrink_node_slabs() and it still to be fixed.
7112 * If the page is not RAM, page_count()should be 0.
7113 * we don't need more check. This is an _used_ not-movable page.
7115 * The problematic thing here is PG_reserved pages. PG_reserved
7116 * is set to both of a memory hole page and a _used_ kernel
7125 bool is_pageblock_removable_nolock(struct page
*page
)
7131 * We have to be careful here because we are iterating over memory
7132 * sections which are not zone aware so we might end up outside of
7133 * the zone but still within the section.
7134 * We have to take care about the node as well. If the node is offline
7135 * its NODE_DATA will be NULL - see page_zone.
7137 if (!node_online(page_to_nid(page
)))
7140 zone
= page_zone(page
);
7141 pfn
= page_to_pfn(page
);
7142 if (!zone_spans_pfn(zone
, pfn
))
7145 return !has_unmovable_pages(zone
, page
, 0, true);
7148 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7150 static unsigned long pfn_max_align_down(unsigned long pfn
)
7152 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7153 pageblock_nr_pages
) - 1);
7156 static unsigned long pfn_max_align_up(unsigned long pfn
)
7158 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7159 pageblock_nr_pages
));
7162 /* [start, end) must belong to a single zone. */
7163 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7164 unsigned long start
, unsigned long end
)
7166 /* This function is based on compact_zone() from compaction.c. */
7167 unsigned long nr_reclaimed
;
7168 unsigned long pfn
= start
;
7169 unsigned int tries
= 0;
7174 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7175 if (fatal_signal_pending(current
)) {
7180 if (list_empty(&cc
->migratepages
)) {
7181 cc
->nr_migratepages
= 0;
7182 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7188 } else if (++tries
== 5) {
7189 ret
= ret
< 0 ? ret
: -EBUSY
;
7193 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7195 cc
->nr_migratepages
-= nr_reclaimed
;
7197 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7198 NULL
, 0, cc
->mode
, MR_CMA
);
7201 putback_movable_pages(&cc
->migratepages
);
7208 * alloc_contig_range() -- tries to allocate given range of pages
7209 * @start: start PFN to allocate
7210 * @end: one-past-the-last PFN to allocate
7211 * @migratetype: migratetype of the underlaying pageblocks (either
7212 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7213 * in range must have the same migratetype and it must
7214 * be either of the two.
7216 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7217 * aligned, however it's the caller's responsibility to guarantee that
7218 * we are the only thread that changes migrate type of pageblocks the
7221 * The PFN range must belong to a single zone.
7223 * Returns zero on success or negative error code. On success all
7224 * pages which PFN is in [start, end) are allocated for the caller and
7225 * need to be freed with free_contig_range().
7227 int alloc_contig_range(unsigned long start
, unsigned long end
,
7228 unsigned migratetype
)
7230 unsigned long outer_start
, outer_end
;
7234 struct compact_control cc
= {
7235 .nr_migratepages
= 0,
7237 .zone
= page_zone(pfn_to_page(start
)),
7238 .mode
= MIGRATE_SYNC
,
7239 .ignore_skip_hint
= true,
7241 INIT_LIST_HEAD(&cc
.migratepages
);
7244 * What we do here is we mark all pageblocks in range as
7245 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7246 * have different sizes, and due to the way page allocator
7247 * work, we align the range to biggest of the two pages so
7248 * that page allocator won't try to merge buddies from
7249 * different pageblocks and change MIGRATE_ISOLATE to some
7250 * other migration type.
7252 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7253 * migrate the pages from an unaligned range (ie. pages that
7254 * we are interested in). This will put all the pages in
7255 * range back to page allocator as MIGRATE_ISOLATE.
7257 * When this is done, we take the pages in range from page
7258 * allocator removing them from the buddy system. This way
7259 * page allocator will never consider using them.
7261 * This lets us mark the pageblocks back as
7262 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7263 * aligned range but not in the unaligned, original range are
7264 * put back to page allocator so that buddy can use them.
7267 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7268 pfn_max_align_up(end
), migratetype
,
7274 * In case of -EBUSY, we'd like to know which page causes problem.
7275 * So, just fall through. We will check it in test_pages_isolated().
7277 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7278 if (ret
&& ret
!= -EBUSY
)
7282 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7283 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7284 * more, all pages in [start, end) are free in page allocator.
7285 * What we are going to do is to allocate all pages from
7286 * [start, end) (that is remove them from page allocator).
7288 * The only problem is that pages at the beginning and at the
7289 * end of interesting range may be not aligned with pages that
7290 * page allocator holds, ie. they can be part of higher order
7291 * pages. Because of this, we reserve the bigger range and
7292 * once this is done free the pages we are not interested in.
7294 * We don't have to hold zone->lock here because the pages are
7295 * isolated thus they won't get removed from buddy.
7298 lru_add_drain_all();
7299 drain_all_pages(cc
.zone
);
7302 outer_start
= start
;
7303 while (!PageBuddy(pfn_to_page(outer_start
))) {
7304 if (++order
>= MAX_ORDER
) {
7305 outer_start
= start
;
7308 outer_start
&= ~0UL << order
;
7311 if (outer_start
!= start
) {
7312 order
= page_order(pfn_to_page(outer_start
));
7315 * outer_start page could be small order buddy page and
7316 * it doesn't include start page. Adjust outer_start
7317 * in this case to report failed page properly
7318 * on tracepoint in test_pages_isolated()
7320 if (outer_start
+ (1UL << order
) <= start
)
7321 outer_start
= start
;
7324 /* Make sure the range is really isolated. */
7325 if (test_pages_isolated(outer_start
, end
, false)) {
7326 pr_info("%s: [%lx, %lx) PFNs busy\n",
7327 __func__
, outer_start
, end
);
7332 /* Grab isolated pages from freelists. */
7333 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7339 /* Free head and tail (if any) */
7340 if (start
!= outer_start
)
7341 free_contig_range(outer_start
, start
- outer_start
);
7342 if (end
!= outer_end
)
7343 free_contig_range(end
, outer_end
- end
);
7346 undo_isolate_page_range(pfn_max_align_down(start
),
7347 pfn_max_align_up(end
), migratetype
);
7351 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7353 unsigned int count
= 0;
7355 for (; nr_pages
--; pfn
++) {
7356 struct page
*page
= pfn_to_page(pfn
);
7358 count
+= page_count(page
) != 1;
7361 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7365 #ifdef CONFIG_MEMORY_HOTPLUG
7367 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7368 * page high values need to be recalulated.
7370 void __meminit
zone_pcp_update(struct zone
*zone
)
7373 mutex_lock(&pcp_batch_high_lock
);
7374 for_each_possible_cpu(cpu
)
7375 pageset_set_high_and_batch(zone
,
7376 per_cpu_ptr(zone
->pageset
, cpu
));
7377 mutex_unlock(&pcp_batch_high_lock
);
7381 void zone_pcp_reset(struct zone
*zone
)
7383 unsigned long flags
;
7385 struct per_cpu_pageset
*pset
;
7387 /* avoid races with drain_pages() */
7388 local_irq_save(flags
);
7389 if (zone
->pageset
!= &boot_pageset
) {
7390 for_each_online_cpu(cpu
) {
7391 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7392 drain_zonestat(zone
, pset
);
7394 free_percpu(zone
->pageset
);
7395 zone
->pageset
= &boot_pageset
;
7397 local_irq_restore(flags
);
7400 #ifdef CONFIG_MEMORY_HOTREMOVE
7402 * All pages in the range must be in a single zone and isolated
7403 * before calling this.
7406 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7410 unsigned int order
, i
;
7412 unsigned long flags
;
7413 /* find the first valid pfn */
7414 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7419 zone
= page_zone(pfn_to_page(pfn
));
7420 spin_lock_irqsave(&zone
->lock
, flags
);
7422 while (pfn
< end_pfn
) {
7423 if (!pfn_valid(pfn
)) {
7427 page
= pfn_to_page(pfn
);
7429 * The HWPoisoned page may be not in buddy system, and
7430 * page_count() is not 0.
7432 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7434 SetPageReserved(page
);
7438 BUG_ON(page_count(page
));
7439 BUG_ON(!PageBuddy(page
));
7440 order
= page_order(page
);
7441 #ifdef CONFIG_DEBUG_VM
7442 pr_info("remove from free list %lx %d %lx\n",
7443 pfn
, 1 << order
, end_pfn
);
7445 list_del(&page
->lru
);
7446 rmv_page_order(page
);
7447 zone
->free_area
[order
].nr_free
--;
7448 for (i
= 0; i
< (1 << order
); i
++)
7449 SetPageReserved((page
+i
));
7450 pfn
+= (1 << order
);
7452 spin_unlock_irqrestore(&zone
->lock
, flags
);
7456 bool is_free_buddy_page(struct page
*page
)
7458 struct zone
*zone
= page_zone(page
);
7459 unsigned long pfn
= page_to_pfn(page
);
7460 unsigned long flags
;
7463 spin_lock_irqsave(&zone
->lock
, flags
);
7464 for (order
= 0; order
< MAX_ORDER
; order
++) {
7465 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7467 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7470 spin_unlock_irqrestore(&zone
->lock
, flags
);
7472 return order
< MAX_ORDER
;