2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
68 #include <asm/sections.h>
69 #include <asm/tlbflush.h>
70 #include <asm/div64.h>
73 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74 static DEFINE_MUTEX(pcp_batch_high_lock
);
75 #define MIN_PERCPU_PAGELIST_FRACTION (8)
77 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78 DEFINE_PER_CPU(int, numa_node
);
79 EXPORT_PER_CPU_SYMBOL(numa_node
);
82 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
89 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
90 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
91 int _node_numa_mem_
[MAX_NUMNODES
];
95 * Array of node states.
97 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
98 [N_POSSIBLE
] = NODE_MASK_ALL
,
99 [N_ONLINE
] = { { [0] = 1UL } },
101 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
102 #ifdef CONFIG_HIGHMEM
103 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
105 #ifdef CONFIG_MOVABLE_NODE
106 [N_MEMORY
] = { { [0] = 1UL } },
108 [N_CPU
] = { { [0] = 1UL } },
111 EXPORT_SYMBOL(node_states
);
113 /* Protect totalram_pages and zone->managed_pages */
114 static DEFINE_SPINLOCK(managed_page_count_lock
);
116 unsigned long totalram_pages __read_mostly
;
117 unsigned long totalreserve_pages __read_mostly
;
118 unsigned long totalcma_pages __read_mostly
;
120 int percpu_pagelist_fraction
;
121 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 * A cached value of the page's pageblock's migratetype, used when the page is
125 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
126 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
127 * Also the migratetype set in the page does not necessarily match the pcplist
128 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
129 * other index - this ensures that it will be put on the correct CMA freelist.
131 static inline int get_pcppage_migratetype(struct page
*page
)
136 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
138 page
->index
= migratetype
;
141 #ifdef CONFIG_PM_SLEEP
143 * The following functions are used by the suspend/hibernate code to temporarily
144 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
145 * while devices are suspended. To avoid races with the suspend/hibernate code,
146 * they should always be called with pm_mutex held (gfp_allowed_mask also should
147 * only be modified with pm_mutex held, unless the suspend/hibernate code is
148 * guaranteed not to run in parallel with that modification).
151 static gfp_t saved_gfp_mask
;
153 void pm_restore_gfp_mask(void)
155 WARN_ON(!mutex_is_locked(&pm_mutex
));
156 if (saved_gfp_mask
) {
157 gfp_allowed_mask
= saved_gfp_mask
;
162 void pm_restrict_gfp_mask(void)
164 WARN_ON(!mutex_is_locked(&pm_mutex
));
165 WARN_ON(saved_gfp_mask
);
166 saved_gfp_mask
= gfp_allowed_mask
;
167 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
170 bool pm_suspended_storage(void)
172 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
176 #endif /* CONFIG_PM_SLEEP */
178 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
179 unsigned int pageblock_order __read_mostly
;
182 static void __free_pages_ok(struct page
*page
, unsigned int order
);
185 * results with 256, 32 in the lowmem_reserve sysctl:
186 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
187 * 1G machine -> (16M dma, 784M normal, 224M high)
188 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
189 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
190 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
192 * TBD: should special case ZONE_DMA32 machines here - in those we normally
193 * don't need any ZONE_NORMAL reservation
195 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
196 #ifdef CONFIG_ZONE_DMA
199 #ifdef CONFIG_ZONE_DMA32
202 #ifdef CONFIG_HIGHMEM
208 EXPORT_SYMBOL(totalram_pages
);
210 static char * const zone_names
[MAX_NR_ZONES
] = {
211 #ifdef CONFIG_ZONE_DMA
214 #ifdef CONFIG_ZONE_DMA32
218 #ifdef CONFIG_HIGHMEM
222 #ifdef CONFIG_ZONE_DEVICE
227 char * const migratetype_names
[MIGRATE_TYPES
] = {
235 #ifdef CONFIG_MEMORY_ISOLATION
240 compound_page_dtor
* const compound_page_dtors
[] = {
243 #ifdef CONFIG_HUGETLB_PAGE
246 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
251 int min_free_kbytes
= 1024;
252 int user_min_free_kbytes
= -1;
253 int watermark_scale_factor
= 10;
255 static unsigned long __meminitdata nr_kernel_pages
;
256 static unsigned long __meminitdata nr_all_pages
;
257 static unsigned long __meminitdata dma_reserve
;
259 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
260 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
261 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
262 static unsigned long __initdata required_kernelcore
;
263 static unsigned long __initdata required_movablecore
;
264 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
265 static bool mirrored_kernelcore
;
267 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
269 EXPORT_SYMBOL(movable_zone
);
270 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
273 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
274 int nr_online_nodes __read_mostly
= 1;
275 EXPORT_SYMBOL(nr_node_ids
);
276 EXPORT_SYMBOL(nr_online_nodes
);
279 int page_group_by_mobility_disabled __read_mostly
;
281 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
282 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
284 pgdat
->first_deferred_pfn
= ULONG_MAX
;
287 /* Returns true if the struct page for the pfn is uninitialised */
288 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
290 int nid
= early_pfn_to_nid(pfn
);
292 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
299 * Returns false when the remaining initialisation should be deferred until
300 * later in the boot cycle when it can be parallelised.
302 static inline bool update_defer_init(pg_data_t
*pgdat
,
303 unsigned long pfn
, unsigned long zone_end
,
304 unsigned long *nr_initialised
)
306 unsigned long max_initialise
;
308 /* Always populate low zones for address-contrained allocations */
309 if (zone_end
< pgdat_end_pfn(pgdat
))
312 * Initialise at least 2G of a node but also take into account that
313 * two large system hashes that can take up 1GB for 0.25TB/node.
315 max_initialise
= max(2UL << (30 - PAGE_SHIFT
),
316 (pgdat
->node_spanned_pages
>> 8));
319 if ((*nr_initialised
> max_initialise
) &&
320 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
321 pgdat
->first_deferred_pfn
= pfn
;
328 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
332 static inline bool early_page_uninitialised(unsigned long pfn
)
337 static inline bool update_defer_init(pg_data_t
*pgdat
,
338 unsigned long pfn
, unsigned long zone_end
,
339 unsigned long *nr_initialised
)
345 /* Return a pointer to the bitmap storing bits affecting a block of pages */
346 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
349 #ifdef CONFIG_SPARSEMEM
350 return __pfn_to_section(pfn
)->pageblock_flags
;
352 return page_zone(page
)->pageblock_flags
;
353 #endif /* CONFIG_SPARSEMEM */
356 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
358 #ifdef CONFIG_SPARSEMEM
359 pfn
&= (PAGES_PER_SECTION
-1);
360 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
362 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
363 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
364 #endif /* CONFIG_SPARSEMEM */
368 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
369 * @page: The page within the block of interest
370 * @pfn: The target page frame number
371 * @end_bitidx: The last bit of interest to retrieve
372 * @mask: mask of bits that the caller is interested in
374 * Return: pageblock_bits flags
376 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
378 unsigned long end_bitidx
,
381 unsigned long *bitmap
;
382 unsigned long bitidx
, word_bitidx
;
385 bitmap
= get_pageblock_bitmap(page
, pfn
);
386 bitidx
= pfn_to_bitidx(page
, pfn
);
387 word_bitidx
= bitidx
/ BITS_PER_LONG
;
388 bitidx
&= (BITS_PER_LONG
-1);
390 word
= bitmap
[word_bitidx
];
391 bitidx
+= end_bitidx
;
392 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
395 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
396 unsigned long end_bitidx
,
399 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
402 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
404 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
408 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
409 * @page: The page within the block of interest
410 * @flags: The flags to set
411 * @pfn: The target page frame number
412 * @end_bitidx: The last bit of interest
413 * @mask: mask of bits that the caller is interested in
415 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
417 unsigned long end_bitidx
,
420 unsigned long *bitmap
;
421 unsigned long bitidx
, word_bitidx
;
422 unsigned long old_word
, word
;
424 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
426 bitmap
= get_pageblock_bitmap(page
, pfn
);
427 bitidx
= pfn_to_bitidx(page
, pfn
);
428 word_bitidx
= bitidx
/ BITS_PER_LONG
;
429 bitidx
&= (BITS_PER_LONG
-1);
431 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
433 bitidx
+= end_bitidx
;
434 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
435 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
437 word
= READ_ONCE(bitmap
[word_bitidx
]);
439 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
440 if (word
== old_word
)
446 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
448 if (unlikely(page_group_by_mobility_disabled
&&
449 migratetype
< MIGRATE_PCPTYPES
))
450 migratetype
= MIGRATE_UNMOVABLE
;
452 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
453 PB_migrate
, PB_migrate_end
);
456 #ifdef CONFIG_DEBUG_VM
457 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
461 unsigned long pfn
= page_to_pfn(page
);
462 unsigned long sp
, start_pfn
;
465 seq
= zone_span_seqbegin(zone
);
466 start_pfn
= zone
->zone_start_pfn
;
467 sp
= zone
->spanned_pages
;
468 if (!zone_spans_pfn(zone
, pfn
))
470 } while (zone_span_seqretry(zone
, seq
));
473 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
474 pfn
, zone_to_nid(zone
), zone
->name
,
475 start_pfn
, start_pfn
+ sp
);
480 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
482 if (!pfn_valid_within(page_to_pfn(page
)))
484 if (zone
!= page_zone(page
))
490 * Temporary debugging check for pages not lying within a given zone.
492 static int bad_range(struct zone
*zone
, struct page
*page
)
494 if (page_outside_zone_boundaries(zone
, page
))
496 if (!page_is_consistent(zone
, page
))
502 static inline int bad_range(struct zone
*zone
, struct page
*page
)
508 static void bad_page(struct page
*page
, const char *reason
,
509 unsigned long bad_flags
)
511 static unsigned long resume
;
512 static unsigned long nr_shown
;
513 static unsigned long nr_unshown
;
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
519 if (nr_shown
== 60) {
520 if (time_before(jiffies
, resume
)) {
526 "BUG: Bad page state: %lu messages suppressed\n",
533 resume
= jiffies
+ 60 * HZ
;
535 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
536 current
->comm
, page_to_pfn(page
));
537 __dump_page(page
, reason
);
538 bad_flags
&= page
->flags
;
540 pr_alert("bad because of flags: %#lx(%pGp)\n",
541 bad_flags
, &bad_flags
);
542 dump_page_owner(page
);
547 /* Leave bad fields for debug, except PageBuddy could make trouble */
548 page_mapcount_reset(page
); /* remove PageBuddy */
549 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
553 * Higher-order pages are called "compound pages". They are structured thusly:
555 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
557 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
558 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
560 * The first tail page's ->compound_dtor holds the offset in array of compound
561 * page destructors. See compound_page_dtors.
563 * The first tail page's ->compound_order holds the order of allocation.
564 * This usage means that zero-order pages may not be compound.
567 void free_compound_page(struct page
*page
)
569 __free_pages_ok(page
, compound_order(page
));
572 void prep_compound_page(struct page
*page
, unsigned int order
)
575 int nr_pages
= 1 << order
;
577 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
578 set_compound_order(page
, order
);
580 for (i
= 1; i
< nr_pages
; i
++) {
581 struct page
*p
= page
+ i
;
582 set_page_count(p
, 0);
583 p
->mapping
= TAIL_MAPPING
;
584 set_compound_head(p
, page
);
586 atomic_set(compound_mapcount_ptr(page
), -1);
589 #ifdef CONFIG_DEBUG_PAGEALLOC
590 unsigned int _debug_guardpage_minorder
;
591 bool _debug_pagealloc_enabled __read_mostly
592 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
593 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
594 bool _debug_guardpage_enabled __read_mostly
;
596 static int __init
early_debug_pagealloc(char *buf
)
600 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
602 early_param("debug_pagealloc", early_debug_pagealloc
);
604 static bool need_debug_guardpage(void)
606 /* If we don't use debug_pagealloc, we don't need guard page */
607 if (!debug_pagealloc_enabled())
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(gfp_t gfp_mask
, const char *fmt
, ...)
2984 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2985 struct va_format vaf
;
2988 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2989 debug_guardpage_minorder() > 0)
2993 * This documents exceptions given to allocations in certain
2994 * contexts that are allowed to allocate outside current's set
2997 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2998 if (test_thread_flag(TIF_MEMDIE
) ||
2999 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3000 filter
&= ~SHOW_MEM_FILTER_NODES
;
3001 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3002 filter
&= ~SHOW_MEM_FILTER_NODES
;
3004 pr_warn("%s: ", current
->comm
);
3006 va_start(args
, fmt
);
3009 pr_cont("%pV", &vaf
);
3012 pr_cont(", mode:%#x(%pGg)\n", gfp_mask
, &gfp_mask
);
3015 if (!should_suppress_show_mem())
3019 static inline struct page
*
3020 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3021 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3023 struct oom_control oc
= {
3024 .zonelist
= ac
->zonelist
,
3025 .nodemask
= ac
->nodemask
,
3027 .gfp_mask
= gfp_mask
,
3032 *did_some_progress
= 0;
3035 * Acquire the oom lock. If that fails, somebody else is
3036 * making progress for us.
3038 if (!mutex_trylock(&oom_lock
)) {
3039 *did_some_progress
= 1;
3040 schedule_timeout_uninterruptible(1);
3045 * Go through the zonelist yet one more time, keep very high watermark
3046 * here, this is only to catch a parallel oom killing, we must fail if
3047 * we're still under heavy pressure.
3049 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3050 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3054 if (!(gfp_mask
& __GFP_NOFAIL
)) {
3055 /* Coredumps can quickly deplete all memory reserves */
3056 if (current
->flags
& PF_DUMPCORE
)
3058 /* The OOM killer will not help higher order allocs */
3059 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3061 /* The OOM killer does not needlessly kill tasks for lowmem */
3062 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3064 if (pm_suspended_storage())
3067 * XXX: GFP_NOFS allocations should rather fail than rely on
3068 * other request to make a forward progress.
3069 * We are in an unfortunate situation where out_of_memory cannot
3070 * do much for this context but let's try it to at least get
3071 * access to memory reserved if the current task is killed (see
3072 * out_of_memory). Once filesystems are ready to handle allocation
3073 * failures more gracefully we should just bail out here.
3076 /* The OOM killer may not free memory on a specific node */
3077 if (gfp_mask
& __GFP_THISNODE
)
3080 /* Exhausted what can be done so it's blamo time */
3081 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3082 *did_some_progress
= 1;
3084 if (gfp_mask
& __GFP_NOFAIL
) {
3085 page
= get_page_from_freelist(gfp_mask
, order
,
3086 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
3088 * fallback to ignore cpuset restriction if our nodes
3092 page
= get_page_from_freelist(gfp_mask
, order
,
3093 ALLOC_NO_WATERMARKS
, ac
);
3097 mutex_unlock(&oom_lock
);
3102 * Maximum number of compaction retries wit a progress before OOM
3103 * killer is consider as the only way to move forward.
3105 #define MAX_COMPACT_RETRIES 16
3107 #ifdef CONFIG_COMPACTION
3108 /* Try memory compaction for high-order allocations before reclaim */
3109 static struct page
*
3110 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3111 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3112 enum compact_priority prio
, enum compact_result
*compact_result
)
3119 current
->flags
|= PF_MEMALLOC
;
3120 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3122 current
->flags
&= ~PF_MEMALLOC
;
3124 if (*compact_result
<= COMPACT_INACTIVE
)
3128 * At least in one zone compaction wasn't deferred or skipped, so let's
3129 * count a compaction stall
3131 count_vm_event(COMPACTSTALL
);
3133 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3136 struct zone
*zone
= page_zone(page
);
3138 zone
->compact_blockskip_flush
= false;
3139 compaction_defer_reset(zone
, order
, true);
3140 count_vm_event(COMPACTSUCCESS
);
3145 * It's bad if compaction run occurs and fails. The most likely reason
3146 * is that pages exist, but not enough to satisfy watermarks.
3148 count_vm_event(COMPACTFAIL
);
3156 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3157 enum compact_result compact_result
,
3158 enum compact_priority
*compact_priority
,
3159 int *compaction_retries
)
3161 int max_retries
= MAX_COMPACT_RETRIES
;
3167 if (compaction_made_progress(compact_result
))
3168 (*compaction_retries
)++;
3171 * compaction considers all the zone as desperately out of memory
3172 * so it doesn't really make much sense to retry except when the
3173 * failure could be caused by insufficient priority
3175 if (compaction_failed(compact_result
))
3176 goto check_priority
;
3179 * make sure the compaction wasn't deferred or didn't bail out early
3180 * due to locks contention before we declare that we should give up.
3181 * But do not retry if the given zonelist is not suitable for
3184 if (compaction_withdrawn(compact_result
))
3185 return compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3188 * !costly requests are much more important than __GFP_REPEAT
3189 * costly ones because they are de facto nofail and invoke OOM
3190 * killer to move on while costly can fail and users are ready
3191 * to cope with that. 1/4 retries is rather arbitrary but we
3192 * would need much more detailed feedback from compaction to
3193 * make a better decision.
3195 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3197 if (*compaction_retries
<= max_retries
)
3201 * Make sure there are attempts at the highest priority if we exhausted
3202 * all retries or failed at the lower priorities.
3205 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3206 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3207 if (*compact_priority
> min_priority
) {
3208 (*compact_priority
)--;
3209 *compaction_retries
= 0;
3215 static inline struct page
*
3216 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3217 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3218 enum compact_priority prio
, enum compact_result
*compact_result
)
3220 *compact_result
= COMPACT_SKIPPED
;
3225 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3226 enum compact_result compact_result
,
3227 enum compact_priority
*compact_priority
,
3228 int *compaction_retries
)
3233 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3237 * There are setups with compaction disabled which would prefer to loop
3238 * inside the allocator rather than hit the oom killer prematurely.
3239 * Let's give them a good hope and keep retrying while the order-0
3240 * watermarks are OK.
3242 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3244 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3245 ac_classzone_idx(ac
), alloc_flags
))
3250 #endif /* CONFIG_COMPACTION */
3252 /* Perform direct synchronous page reclaim */
3254 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3255 const struct alloc_context
*ac
)
3257 struct reclaim_state reclaim_state
;
3262 /* We now go into synchronous reclaim */
3263 cpuset_memory_pressure_bump();
3264 current
->flags
|= PF_MEMALLOC
;
3265 lockdep_set_current_reclaim_state(gfp_mask
);
3266 reclaim_state
.reclaimed_slab
= 0;
3267 current
->reclaim_state
= &reclaim_state
;
3269 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3272 current
->reclaim_state
= NULL
;
3273 lockdep_clear_current_reclaim_state();
3274 current
->flags
&= ~PF_MEMALLOC
;
3281 /* The really slow allocator path where we enter direct reclaim */
3282 static inline struct page
*
3283 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3284 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3285 unsigned long *did_some_progress
)
3287 struct page
*page
= NULL
;
3288 bool drained
= false;
3290 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3291 if (unlikely(!(*did_some_progress
)))
3295 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3298 * If an allocation failed after direct reclaim, it could be because
3299 * pages are pinned on the per-cpu lists or in high alloc reserves.
3300 * Shrink them them and try again
3302 if (!page
&& !drained
) {
3303 unreserve_highatomic_pageblock(ac
);
3304 drain_all_pages(NULL
);
3312 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3316 pg_data_t
*last_pgdat
= NULL
;
3318 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3319 ac
->high_zoneidx
, ac
->nodemask
) {
3320 if (last_pgdat
!= zone
->zone_pgdat
)
3321 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3322 last_pgdat
= zone
->zone_pgdat
;
3326 static inline unsigned int
3327 gfp_to_alloc_flags(gfp_t gfp_mask
)
3329 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3331 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3332 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3335 * The caller may dip into page reserves a bit more if the caller
3336 * cannot run direct reclaim, or if the caller has realtime scheduling
3337 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3338 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3340 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3342 if (gfp_mask
& __GFP_ATOMIC
) {
3344 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3345 * if it can't schedule.
3347 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3348 alloc_flags
|= ALLOC_HARDER
;
3350 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3351 * comment for __cpuset_node_allowed().
3353 alloc_flags
&= ~ALLOC_CPUSET
;
3354 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3355 alloc_flags
|= ALLOC_HARDER
;
3358 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3359 alloc_flags
|= ALLOC_CMA
;
3364 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3366 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3369 if (gfp_mask
& __GFP_MEMALLOC
)
3371 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3373 if (!in_interrupt() &&
3374 ((current
->flags
& PF_MEMALLOC
) ||
3375 unlikely(test_thread_flag(TIF_MEMDIE
))))
3382 * Maximum number of reclaim retries without any progress before OOM killer
3383 * is consider as the only way to move forward.
3385 #define MAX_RECLAIM_RETRIES 16
3388 * Checks whether it makes sense to retry the reclaim to make a forward progress
3389 * for the given allocation request.
3390 * The reclaim feedback represented by did_some_progress (any progress during
3391 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3392 * any progress in a row) is considered as well as the reclaimable pages on the
3393 * applicable zone list (with a backoff mechanism which is a function of
3394 * no_progress_loops).
3396 * Returns true if a retry is viable or false to enter the oom path.
3399 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3400 struct alloc_context
*ac
, int alloc_flags
,
3401 bool did_some_progress
, int *no_progress_loops
)
3407 * Costly allocations might have made a progress but this doesn't mean
3408 * their order will become available due to high fragmentation so
3409 * always increment the no progress counter for them
3411 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3412 *no_progress_loops
= 0;
3414 (*no_progress_loops
)++;
3417 * Make sure we converge to OOM if we cannot make any progress
3418 * several times in the row.
3420 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
)
3424 * Keep reclaiming pages while there is a chance this will lead
3425 * somewhere. If none of the target zones can satisfy our allocation
3426 * request even if all reclaimable pages are considered then we are
3427 * screwed and have to go OOM.
3429 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3431 unsigned long available
;
3432 unsigned long reclaimable
;
3434 available
= reclaimable
= zone_reclaimable_pages(zone
);
3435 available
-= DIV_ROUND_UP((*no_progress_loops
) * available
,
3436 MAX_RECLAIM_RETRIES
);
3437 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3440 * Would the allocation succeed if we reclaimed the whole
3443 if (__zone_watermark_ok(zone
, order
, min_wmark_pages(zone
),
3444 ac_classzone_idx(ac
), alloc_flags
, available
)) {
3446 * If we didn't make any progress and have a lot of
3447 * dirty + writeback pages then we should wait for
3448 * an IO to complete to slow down the reclaim and
3449 * prevent from pre mature OOM
3451 if (!did_some_progress
) {
3452 unsigned long write_pending
;
3454 write_pending
= zone_page_state_snapshot(zone
,
3455 NR_ZONE_WRITE_PENDING
);
3457 if (2 * write_pending
> reclaimable
) {
3458 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3464 * Memory allocation/reclaim might be called from a WQ
3465 * context and the current implementation of the WQ
3466 * concurrency control doesn't recognize that
3467 * a particular WQ is congested if the worker thread is
3468 * looping without ever sleeping. Therefore we have to
3469 * do a short sleep here rather than calling
3472 if (current
->flags
& PF_WQ_WORKER
)
3473 schedule_timeout_uninterruptible(1);
3484 static inline struct page
*
3485 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3486 struct alloc_context
*ac
)
3488 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3489 struct page
*page
= NULL
;
3490 unsigned int alloc_flags
;
3491 unsigned long did_some_progress
;
3492 enum compact_priority compact_priority
= DEF_COMPACT_PRIORITY
;
3493 enum compact_result compact_result
;
3494 int compaction_retries
= 0;
3495 int no_progress_loops
= 0;
3496 unsigned long alloc_start
= jiffies
;
3497 unsigned int stall_timeout
= 10 * HZ
;
3500 * In the slowpath, we sanity check order to avoid ever trying to
3501 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3502 * be using allocators in order of preference for an area that is
3505 if (order
>= MAX_ORDER
) {
3506 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3511 * We also sanity check to catch abuse of atomic reserves being used by
3512 * callers that are not in atomic context.
3514 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3515 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3516 gfp_mask
&= ~__GFP_ATOMIC
;
3519 * The fast path uses conservative alloc_flags to succeed only until
3520 * kswapd needs to be woken up, and to avoid the cost of setting up
3521 * alloc_flags precisely. So we do that now.
3523 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3525 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3526 wake_all_kswapds(order
, ac
);
3529 * The adjusted alloc_flags might result in immediate success, so try
3532 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3537 * For costly allocations, try direct compaction first, as it's likely
3538 * that we have enough base pages and don't need to reclaim. Don't try
3539 * that for allocations that are allowed to ignore watermarks, as the
3540 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3542 if (can_direct_reclaim
&& order
> PAGE_ALLOC_COSTLY_ORDER
&&
3543 !gfp_pfmemalloc_allowed(gfp_mask
)) {
3544 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3546 INIT_COMPACT_PRIORITY
,
3552 * Checks for costly allocations with __GFP_NORETRY, which
3553 * includes THP page fault allocations
3555 if (gfp_mask
& __GFP_NORETRY
) {
3557 * If compaction is deferred for high-order allocations,
3558 * it is because sync compaction recently failed. If
3559 * this is the case and the caller requested a THP
3560 * allocation, we do not want to heavily disrupt the
3561 * system, so we fail the allocation instead of entering
3564 if (compact_result
== COMPACT_DEFERRED
)
3568 * Looks like reclaim/compaction is worth trying, but
3569 * sync compaction could be very expensive, so keep
3570 * using async compaction.
3572 compact_priority
= INIT_COMPACT_PRIORITY
;
3577 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3578 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3579 wake_all_kswapds(order
, ac
);
3581 if (gfp_pfmemalloc_allowed(gfp_mask
))
3582 alloc_flags
= ALLOC_NO_WATERMARKS
;
3585 * Reset the zonelist iterators if memory policies can be ignored.
3586 * These allocations are high priority and system rather than user
3589 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3590 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3591 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3592 ac
->high_zoneidx
, ac
->nodemask
);
3595 /* Attempt with potentially adjusted zonelist and alloc_flags */
3596 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3600 /* Caller is not willing to reclaim, we can't balance anything */
3601 if (!can_direct_reclaim
) {
3603 * All existing users of the __GFP_NOFAIL are blockable, so warn
3604 * of any new users that actually allow this type of allocation
3607 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3611 /* Avoid recursion of direct reclaim */
3612 if (current
->flags
& PF_MEMALLOC
) {
3614 * __GFP_NOFAIL request from this context is rather bizarre
3615 * because we cannot reclaim anything and only can loop waiting
3616 * for somebody to do a work for us.
3618 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3625 /* Avoid allocations with no watermarks from looping endlessly */
3626 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3630 /* Try direct reclaim and then allocating */
3631 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3632 &did_some_progress
);
3636 /* Try direct compaction and then allocating */
3637 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3638 compact_priority
, &compact_result
);
3642 /* Do not loop if specifically requested */
3643 if (gfp_mask
& __GFP_NORETRY
)
3647 * Do not retry costly high order allocations unless they are
3650 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_REPEAT
))
3653 /* Make sure we know about allocations which stall for too long */
3654 if (time_after(jiffies
, alloc_start
+ stall_timeout
)) {
3655 warn_alloc(gfp_mask
,
3656 "page alloction stalls for %ums, order:%u\n",
3657 jiffies_to_msecs(jiffies
-alloc_start
), order
);
3658 stall_timeout
+= 10 * HZ
;
3661 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3662 did_some_progress
> 0, &no_progress_loops
))
3666 * It doesn't make any sense to retry for the compaction if the order-0
3667 * reclaim is not able to make any progress because the current
3668 * implementation of the compaction depends on the sufficient amount
3669 * of free memory (see __compaction_suitable)
3671 if (did_some_progress
> 0 &&
3672 should_compact_retry(ac
, order
, alloc_flags
,
3673 compact_result
, &compact_priority
,
3674 &compaction_retries
))
3677 /* Reclaim has failed us, start killing things */
3678 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3682 /* Retry as long as the OOM killer is making progress */
3683 if (did_some_progress
) {
3684 no_progress_loops
= 0;
3689 warn_alloc(gfp_mask
,
3690 "page allocation failure: order:%u", order
);
3696 * This is the 'heart' of the zoned buddy allocator.
3699 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3700 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3703 unsigned int cpuset_mems_cookie
;
3704 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
3705 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3706 struct alloc_context ac
= {
3707 .high_zoneidx
= gfp_zone(gfp_mask
),
3708 .zonelist
= zonelist
,
3709 .nodemask
= nodemask
,
3710 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3713 if (cpusets_enabled()) {
3714 alloc_mask
|= __GFP_HARDWALL
;
3715 alloc_flags
|= ALLOC_CPUSET
;
3717 ac
.nodemask
= &cpuset_current_mems_allowed
;
3720 gfp_mask
&= gfp_allowed_mask
;
3722 lockdep_trace_alloc(gfp_mask
);
3724 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3726 if (should_fail_alloc_page(gfp_mask
, order
))
3730 * Check the zones suitable for the gfp_mask contain at least one
3731 * valid zone. It's possible to have an empty zonelist as a result
3732 * of __GFP_THISNODE and a memoryless node
3734 if (unlikely(!zonelist
->_zonerefs
->zone
))
3737 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3738 alloc_flags
|= ALLOC_CMA
;
3741 cpuset_mems_cookie
= read_mems_allowed_begin();
3743 /* Dirty zone balancing only done in the fast path */
3744 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3747 * The preferred zone is used for statistics but crucially it is
3748 * also used as the starting point for the zonelist iterator. It
3749 * may get reset for allocations that ignore memory policies.
3751 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3752 ac
.high_zoneidx
, ac
.nodemask
);
3753 if (!ac
.preferred_zoneref
) {
3758 /* First allocation attempt */
3759 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3764 * Runtime PM, block IO and its error handling path can deadlock
3765 * because I/O on the device might not complete.
3767 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3768 ac
.spread_dirty_pages
= false;
3771 * Restore the original nodemask if it was potentially replaced with
3772 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3774 if (cpusets_enabled())
3775 ac
.nodemask
= nodemask
;
3776 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3780 * When updating a task's mems_allowed, it is possible to race with
3781 * parallel threads in such a way that an allocation can fail while
3782 * the mask is being updated. If a page allocation is about to fail,
3783 * check if the cpuset changed during allocation and if so, retry.
3785 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3786 alloc_mask
= gfp_mask
;
3791 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
3792 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
3793 __free_pages(page
, order
);
3797 if (kmemcheck_enabled
&& page
)
3798 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3800 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3804 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3807 * Common helper functions.
3809 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3814 * __get_free_pages() returns a 32-bit address, which cannot represent
3817 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3819 page
= alloc_pages(gfp_mask
, order
);
3822 return (unsigned long) page_address(page
);
3824 EXPORT_SYMBOL(__get_free_pages
);
3826 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3828 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3830 EXPORT_SYMBOL(get_zeroed_page
);
3832 void __free_pages(struct page
*page
, unsigned int order
)
3834 if (put_page_testzero(page
)) {
3836 free_hot_cold_page(page
, false);
3838 __free_pages_ok(page
, order
);
3842 EXPORT_SYMBOL(__free_pages
);
3844 void free_pages(unsigned long addr
, unsigned int order
)
3847 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3848 __free_pages(virt_to_page((void *)addr
), order
);
3852 EXPORT_SYMBOL(free_pages
);
3856 * An arbitrary-length arbitrary-offset area of memory which resides
3857 * within a 0 or higher order page. Multiple fragments within that page
3858 * are individually refcounted, in the page's reference counter.
3860 * The page_frag functions below provide a simple allocation framework for
3861 * page fragments. This is used by the network stack and network device
3862 * drivers to provide a backing region of memory for use as either an
3863 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3865 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3868 struct page
*page
= NULL
;
3869 gfp_t gfp
= gfp_mask
;
3871 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3872 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3874 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3875 PAGE_FRAG_CACHE_MAX_ORDER
);
3876 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3878 if (unlikely(!page
))
3879 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3881 nc
->va
= page
? page_address(page
) : NULL
;
3886 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3887 unsigned int fragsz
, gfp_t gfp_mask
)
3889 unsigned int size
= PAGE_SIZE
;
3893 if (unlikely(!nc
->va
)) {
3895 page
= __page_frag_refill(nc
, gfp_mask
);
3899 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3900 /* if size can vary use size else just use PAGE_SIZE */
3903 /* Even if we own the page, we do not use atomic_set().
3904 * This would break get_page_unless_zero() users.
3906 page_ref_add(page
, size
- 1);
3908 /* reset page count bias and offset to start of new frag */
3909 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3910 nc
->pagecnt_bias
= size
;
3914 offset
= nc
->offset
- fragsz
;
3915 if (unlikely(offset
< 0)) {
3916 page
= virt_to_page(nc
->va
);
3918 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3921 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3922 /* if size can vary use size else just use PAGE_SIZE */
3925 /* OK, page count is 0, we can safely set it */
3926 set_page_count(page
, size
);
3928 /* reset page count bias and offset to start of new frag */
3929 nc
->pagecnt_bias
= size
;
3930 offset
= size
- fragsz
;
3934 nc
->offset
= offset
;
3936 return nc
->va
+ offset
;
3938 EXPORT_SYMBOL(__alloc_page_frag
);
3941 * Frees a page fragment allocated out of either a compound or order 0 page.
3943 void __free_page_frag(void *addr
)
3945 struct page
*page
= virt_to_head_page(addr
);
3947 if (unlikely(put_page_testzero(page
)))
3948 __free_pages_ok(page
, compound_order(page
));
3950 EXPORT_SYMBOL(__free_page_frag
);
3952 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3956 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3957 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3959 split_page(virt_to_page((void *)addr
), order
);
3960 while (used
< alloc_end
) {
3965 return (void *)addr
;
3969 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3970 * @size: the number of bytes to allocate
3971 * @gfp_mask: GFP flags for the allocation
3973 * This function is similar to alloc_pages(), except that it allocates the
3974 * minimum number of pages to satisfy the request. alloc_pages() can only
3975 * allocate memory in power-of-two pages.
3977 * This function is also limited by MAX_ORDER.
3979 * Memory allocated by this function must be released by free_pages_exact().
3981 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3983 unsigned int order
= get_order(size
);
3986 addr
= __get_free_pages(gfp_mask
, order
);
3987 return make_alloc_exact(addr
, order
, size
);
3989 EXPORT_SYMBOL(alloc_pages_exact
);
3992 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3994 * @nid: the preferred node ID where memory should be allocated
3995 * @size: the number of bytes to allocate
3996 * @gfp_mask: GFP flags for the allocation
3998 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4001 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4003 unsigned int order
= get_order(size
);
4004 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4007 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4011 * free_pages_exact - release memory allocated via alloc_pages_exact()
4012 * @virt: the value returned by alloc_pages_exact.
4013 * @size: size of allocation, same value as passed to alloc_pages_exact().
4015 * Release the memory allocated by a previous call to alloc_pages_exact.
4017 void free_pages_exact(void *virt
, size_t size
)
4019 unsigned long addr
= (unsigned long)virt
;
4020 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4022 while (addr
< end
) {
4027 EXPORT_SYMBOL(free_pages_exact
);
4030 * nr_free_zone_pages - count number of pages beyond high watermark
4031 * @offset: The zone index of the highest zone
4033 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4034 * high watermark within all zones at or below a given zone index. For each
4035 * zone, the number of pages is calculated as:
4036 * managed_pages - high_pages
4038 static unsigned long nr_free_zone_pages(int offset
)
4043 /* Just pick one node, since fallback list is circular */
4044 unsigned long sum
= 0;
4046 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4048 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4049 unsigned long size
= zone
->managed_pages
;
4050 unsigned long high
= high_wmark_pages(zone
);
4059 * nr_free_buffer_pages - count number of pages beyond high watermark
4061 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4062 * watermark within ZONE_DMA and ZONE_NORMAL.
4064 unsigned long nr_free_buffer_pages(void)
4066 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4068 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4071 * nr_free_pagecache_pages - count number of pages beyond high watermark
4073 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4074 * high watermark within all zones.
4076 unsigned long nr_free_pagecache_pages(void)
4078 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4081 static inline void show_node(struct zone
*zone
)
4083 if (IS_ENABLED(CONFIG_NUMA
))
4084 printk("Node %d ", zone_to_nid(zone
));
4087 long si_mem_available(void)
4090 unsigned long pagecache
;
4091 unsigned long wmark_low
= 0;
4092 unsigned long pages
[NR_LRU_LISTS
];
4096 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4097 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4100 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4103 * Estimate the amount of memory available for userspace allocations,
4104 * without causing swapping.
4106 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4109 * Not all the page cache can be freed, otherwise the system will
4110 * start swapping. Assume at least half of the page cache, or the
4111 * low watermark worth of cache, needs to stay.
4113 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4114 pagecache
-= min(pagecache
/ 2, wmark_low
);
4115 available
+= pagecache
;
4118 * Part of the reclaimable slab consists of items that are in use,
4119 * and cannot be freed. Cap this estimate at the low watermark.
4121 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4122 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4128 EXPORT_SYMBOL_GPL(si_mem_available
);
4130 void si_meminfo(struct sysinfo
*val
)
4132 val
->totalram
= totalram_pages
;
4133 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4134 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4135 val
->bufferram
= nr_blockdev_pages();
4136 val
->totalhigh
= totalhigh_pages
;
4137 val
->freehigh
= nr_free_highpages();
4138 val
->mem_unit
= PAGE_SIZE
;
4141 EXPORT_SYMBOL(si_meminfo
);
4144 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4146 int zone_type
; /* needs to be signed */
4147 unsigned long managed_pages
= 0;
4148 unsigned long managed_highpages
= 0;
4149 unsigned long free_highpages
= 0;
4150 pg_data_t
*pgdat
= NODE_DATA(nid
);
4152 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4153 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4154 val
->totalram
= managed_pages
;
4155 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4156 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4157 #ifdef CONFIG_HIGHMEM
4158 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4159 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4161 if (is_highmem(zone
)) {
4162 managed_highpages
+= zone
->managed_pages
;
4163 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4166 val
->totalhigh
= managed_highpages
;
4167 val
->freehigh
= free_highpages
;
4169 val
->totalhigh
= managed_highpages
;
4170 val
->freehigh
= free_highpages
;
4172 val
->mem_unit
= PAGE_SIZE
;
4177 * Determine whether the node should be displayed or not, depending on whether
4178 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4180 bool skip_free_areas_node(unsigned int flags
, int nid
)
4183 unsigned int cpuset_mems_cookie
;
4185 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4189 cpuset_mems_cookie
= read_mems_allowed_begin();
4190 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
4191 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
4196 #define K(x) ((x) << (PAGE_SHIFT-10))
4198 static void show_migration_types(unsigned char type
)
4200 static const char types
[MIGRATE_TYPES
] = {
4201 [MIGRATE_UNMOVABLE
] = 'U',
4202 [MIGRATE_MOVABLE
] = 'M',
4203 [MIGRATE_RECLAIMABLE
] = 'E',
4204 [MIGRATE_HIGHATOMIC
] = 'H',
4206 [MIGRATE_CMA
] = 'C',
4208 #ifdef CONFIG_MEMORY_ISOLATION
4209 [MIGRATE_ISOLATE
] = 'I',
4212 char tmp
[MIGRATE_TYPES
+ 1];
4216 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4217 if (type
& (1 << i
))
4222 printk("(%s) ", tmp
);
4226 * Show free area list (used inside shift_scroll-lock stuff)
4227 * We also calculate the percentage fragmentation. We do this by counting the
4228 * memory on each free list with the exception of the first item on the list.
4231 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4234 void show_free_areas(unsigned int filter
)
4236 unsigned long free_pcp
= 0;
4241 for_each_populated_zone(zone
) {
4242 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4245 for_each_online_cpu(cpu
)
4246 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4249 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4250 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4251 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4252 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4253 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4254 " free:%lu free_pcp:%lu free_cma:%lu\n",
4255 global_node_page_state(NR_ACTIVE_ANON
),
4256 global_node_page_state(NR_INACTIVE_ANON
),
4257 global_node_page_state(NR_ISOLATED_ANON
),
4258 global_node_page_state(NR_ACTIVE_FILE
),
4259 global_node_page_state(NR_INACTIVE_FILE
),
4260 global_node_page_state(NR_ISOLATED_FILE
),
4261 global_node_page_state(NR_UNEVICTABLE
),
4262 global_node_page_state(NR_FILE_DIRTY
),
4263 global_node_page_state(NR_WRITEBACK
),
4264 global_node_page_state(NR_UNSTABLE_NFS
),
4265 global_page_state(NR_SLAB_RECLAIMABLE
),
4266 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4267 global_node_page_state(NR_FILE_MAPPED
),
4268 global_node_page_state(NR_SHMEM
),
4269 global_page_state(NR_PAGETABLE
),
4270 global_page_state(NR_BOUNCE
),
4271 global_page_state(NR_FREE_PAGES
),
4273 global_page_state(NR_FREE_CMA_PAGES
));
4275 for_each_online_pgdat(pgdat
) {
4277 " active_anon:%lukB"
4278 " inactive_anon:%lukB"
4279 " active_file:%lukB"
4280 " inactive_file:%lukB"
4281 " unevictable:%lukB"
4282 " isolated(anon):%lukB"
4283 " isolated(file):%lukB"
4288 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4290 " shmem_pmdmapped: %lukB"
4293 " writeback_tmp:%lukB"
4295 " pages_scanned:%lu"
4296 " all_unreclaimable? %s"
4299 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4300 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4301 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4302 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4303 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4304 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4305 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4306 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4307 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4308 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4309 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4310 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4311 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4313 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4315 K(node_page_state(pgdat
, NR_SHMEM
)),
4316 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4317 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4318 node_page_state(pgdat
, NR_PAGES_SCANNED
),
4319 !pgdat_reclaimable(pgdat
) ? "yes" : "no");
4322 for_each_populated_zone(zone
) {
4325 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4329 for_each_online_cpu(cpu
)
4330 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4338 " active_anon:%lukB"
4339 " inactive_anon:%lukB"
4340 " active_file:%lukB"
4341 " inactive_file:%lukB"
4342 " unevictable:%lukB"
4343 " writepending:%lukB"
4347 " slab_reclaimable:%lukB"
4348 " slab_unreclaimable:%lukB"
4349 " kernel_stack:%lukB"
4357 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4358 K(min_wmark_pages(zone
)),
4359 K(low_wmark_pages(zone
)),
4360 K(high_wmark_pages(zone
)),
4361 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4362 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4363 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4364 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4365 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4366 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4367 K(zone
->present_pages
),
4368 K(zone
->managed_pages
),
4369 K(zone_page_state(zone
, NR_MLOCK
)),
4370 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4371 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4372 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4373 K(zone_page_state(zone
, NR_PAGETABLE
)),
4374 K(zone_page_state(zone
, NR_BOUNCE
)),
4376 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4377 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4378 printk("lowmem_reserve[]:");
4379 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4380 printk(" %ld", zone
->lowmem_reserve
[i
]);
4384 for_each_populated_zone(zone
) {
4386 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4387 unsigned char types
[MAX_ORDER
];
4389 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4392 printk("%s: ", zone
->name
);
4394 spin_lock_irqsave(&zone
->lock
, flags
);
4395 for (order
= 0; order
< MAX_ORDER
; order
++) {
4396 struct free_area
*area
= &zone
->free_area
[order
];
4399 nr
[order
] = area
->nr_free
;
4400 total
+= nr
[order
] << order
;
4403 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4404 if (!list_empty(&area
->free_list
[type
]))
4405 types
[order
] |= 1 << type
;
4408 spin_unlock_irqrestore(&zone
->lock
, flags
);
4409 for (order
= 0; order
< MAX_ORDER
; order
++) {
4410 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
4412 show_migration_types(types
[order
]);
4414 printk("= %lukB\n", K(total
));
4417 hugetlb_show_meminfo();
4419 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4421 show_swap_cache_info();
4424 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4426 zoneref
->zone
= zone
;
4427 zoneref
->zone_idx
= zone_idx(zone
);
4431 * Builds allocation fallback zone lists.
4433 * Add all populated zones of a node to the zonelist.
4435 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4439 enum zone_type zone_type
= MAX_NR_ZONES
;
4443 zone
= pgdat
->node_zones
+ zone_type
;
4444 if (managed_zone(zone
)) {
4445 zoneref_set_zone(zone
,
4446 &zonelist
->_zonerefs
[nr_zones
++]);
4447 check_highest_zone(zone_type
);
4449 } while (zone_type
);
4457 * 0 = automatic detection of better ordering.
4458 * 1 = order by ([node] distance, -zonetype)
4459 * 2 = order by (-zonetype, [node] distance)
4461 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4462 * the same zonelist. So only NUMA can configure this param.
4464 #define ZONELIST_ORDER_DEFAULT 0
4465 #define ZONELIST_ORDER_NODE 1
4466 #define ZONELIST_ORDER_ZONE 2
4468 /* zonelist order in the kernel.
4469 * set_zonelist_order() will set this to NODE or ZONE.
4471 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4472 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4476 /* The value user specified ....changed by config */
4477 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4478 /* string for sysctl */
4479 #define NUMA_ZONELIST_ORDER_LEN 16
4480 char numa_zonelist_order
[16] = "default";
4483 * interface for configure zonelist ordering.
4484 * command line option "numa_zonelist_order"
4485 * = "[dD]efault - default, automatic configuration.
4486 * = "[nN]ode - order by node locality, then by zone within node
4487 * = "[zZ]one - order by zone, then by locality within zone
4490 static int __parse_numa_zonelist_order(char *s
)
4492 if (*s
== 'd' || *s
== 'D') {
4493 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4494 } else if (*s
== 'n' || *s
== 'N') {
4495 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4496 } else if (*s
== 'z' || *s
== 'Z') {
4497 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4499 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4505 static __init
int setup_numa_zonelist_order(char *s
)
4512 ret
= __parse_numa_zonelist_order(s
);
4514 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4518 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4521 * sysctl handler for numa_zonelist_order
4523 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4524 void __user
*buffer
, size_t *length
,
4527 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4529 static DEFINE_MUTEX(zl_order_mutex
);
4531 mutex_lock(&zl_order_mutex
);
4533 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4537 strcpy(saved_string
, (char *)table
->data
);
4539 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4543 int oldval
= user_zonelist_order
;
4545 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4548 * bogus value. restore saved string
4550 strncpy((char *)table
->data
, saved_string
,
4551 NUMA_ZONELIST_ORDER_LEN
);
4552 user_zonelist_order
= oldval
;
4553 } else if (oldval
!= user_zonelist_order
) {
4554 mutex_lock(&zonelists_mutex
);
4555 build_all_zonelists(NULL
, NULL
);
4556 mutex_unlock(&zonelists_mutex
);
4560 mutex_unlock(&zl_order_mutex
);
4565 #define MAX_NODE_LOAD (nr_online_nodes)
4566 static int node_load
[MAX_NUMNODES
];
4569 * find_next_best_node - find the next node that should appear in a given node's fallback list
4570 * @node: node whose fallback list we're appending
4571 * @used_node_mask: nodemask_t of already used nodes
4573 * We use a number of factors to determine which is the next node that should
4574 * appear on a given node's fallback list. The node should not have appeared
4575 * already in @node's fallback list, and it should be the next closest node
4576 * according to the distance array (which contains arbitrary distance values
4577 * from each node to each node in the system), and should also prefer nodes
4578 * with no CPUs, since presumably they'll have very little allocation pressure
4579 * on them otherwise.
4580 * It returns -1 if no node is found.
4582 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4585 int min_val
= INT_MAX
;
4586 int best_node
= NUMA_NO_NODE
;
4587 const struct cpumask
*tmp
= cpumask_of_node(0);
4589 /* Use the local node if we haven't already */
4590 if (!node_isset(node
, *used_node_mask
)) {
4591 node_set(node
, *used_node_mask
);
4595 for_each_node_state(n
, N_MEMORY
) {
4597 /* Don't want a node to appear more than once */
4598 if (node_isset(n
, *used_node_mask
))
4601 /* Use the distance array to find the distance */
4602 val
= node_distance(node
, n
);
4604 /* Penalize nodes under us ("prefer the next node") */
4607 /* Give preference to headless and unused nodes */
4608 tmp
= cpumask_of_node(n
);
4609 if (!cpumask_empty(tmp
))
4610 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4612 /* Slight preference for less loaded node */
4613 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4614 val
+= node_load
[n
];
4616 if (val
< min_val
) {
4623 node_set(best_node
, *used_node_mask
);
4630 * Build zonelists ordered by node and zones within node.
4631 * This results in maximum locality--normal zone overflows into local
4632 * DMA zone, if any--but risks exhausting DMA zone.
4634 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4637 struct zonelist
*zonelist
;
4639 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4640 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4642 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4643 zonelist
->_zonerefs
[j
].zone
= NULL
;
4644 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4648 * Build gfp_thisnode zonelists
4650 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4653 struct zonelist
*zonelist
;
4655 zonelist
= &pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
];
4656 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4657 zonelist
->_zonerefs
[j
].zone
= NULL
;
4658 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4662 * Build zonelists ordered by zone and nodes within zones.
4663 * This results in conserving DMA zone[s] until all Normal memory is
4664 * exhausted, but results in overflowing to remote node while memory
4665 * may still exist in local DMA zone.
4667 static int node_order
[MAX_NUMNODES
];
4669 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4672 int zone_type
; /* needs to be signed */
4674 struct zonelist
*zonelist
;
4676 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4678 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4679 for (j
= 0; j
< nr_nodes
; j
++) {
4680 node
= node_order
[j
];
4681 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4682 if (managed_zone(z
)) {
4684 &zonelist
->_zonerefs
[pos
++]);
4685 check_highest_zone(zone_type
);
4689 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4690 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4693 #if defined(CONFIG_64BIT)
4695 * Devices that require DMA32/DMA are relatively rare and do not justify a
4696 * penalty to every machine in case the specialised case applies. Default
4697 * to Node-ordering on 64-bit NUMA machines
4699 static int default_zonelist_order(void)
4701 return ZONELIST_ORDER_NODE
;
4705 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4706 * by the kernel. If processes running on node 0 deplete the low memory zone
4707 * then reclaim will occur more frequency increasing stalls and potentially
4708 * be easier to OOM if a large percentage of the zone is under writeback or
4709 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4710 * Hence, default to zone ordering on 32-bit.
4712 static int default_zonelist_order(void)
4714 return ZONELIST_ORDER_ZONE
;
4716 #endif /* CONFIG_64BIT */
4718 static void set_zonelist_order(void)
4720 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4721 current_zonelist_order
= default_zonelist_order();
4723 current_zonelist_order
= user_zonelist_order
;
4726 static void build_zonelists(pg_data_t
*pgdat
)
4729 nodemask_t used_mask
;
4730 int local_node
, prev_node
;
4731 struct zonelist
*zonelist
;
4732 unsigned int order
= current_zonelist_order
;
4734 /* initialize zonelists */
4735 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4736 zonelist
= pgdat
->node_zonelists
+ i
;
4737 zonelist
->_zonerefs
[0].zone
= NULL
;
4738 zonelist
->_zonerefs
[0].zone_idx
= 0;
4741 /* NUMA-aware ordering of nodes */
4742 local_node
= pgdat
->node_id
;
4743 load
= nr_online_nodes
;
4744 prev_node
= local_node
;
4745 nodes_clear(used_mask
);
4747 memset(node_order
, 0, sizeof(node_order
));
4750 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4752 * We don't want to pressure a particular node.
4753 * So adding penalty to the first node in same
4754 * distance group to make it round-robin.
4756 if (node_distance(local_node
, node
) !=
4757 node_distance(local_node
, prev_node
))
4758 node_load
[node
] = load
;
4762 if (order
== ZONELIST_ORDER_NODE
)
4763 build_zonelists_in_node_order(pgdat
, node
);
4765 node_order
[i
++] = node
; /* remember order */
4768 if (order
== ZONELIST_ORDER_ZONE
) {
4769 /* calculate node order -- i.e., DMA last! */
4770 build_zonelists_in_zone_order(pgdat
, i
);
4773 build_thisnode_zonelists(pgdat
);
4776 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4778 * Return node id of node used for "local" allocations.
4779 * I.e., first node id of first zone in arg node's generic zonelist.
4780 * Used for initializing percpu 'numa_mem', which is used primarily
4781 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4783 int local_memory_node(int node
)
4787 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4788 gfp_zone(GFP_KERNEL
),
4790 return z
->zone
->node
;
4794 static void setup_min_unmapped_ratio(void);
4795 static void setup_min_slab_ratio(void);
4796 #else /* CONFIG_NUMA */
4798 static void set_zonelist_order(void)
4800 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4803 static void build_zonelists(pg_data_t
*pgdat
)
4805 int node
, local_node
;
4807 struct zonelist
*zonelist
;
4809 local_node
= pgdat
->node_id
;
4811 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4812 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4815 * Now we build the zonelist so that it contains the zones
4816 * of all the other nodes.
4817 * We don't want to pressure a particular node, so when
4818 * building the zones for node N, we make sure that the
4819 * zones coming right after the local ones are those from
4820 * node N+1 (modulo N)
4822 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4823 if (!node_online(node
))
4825 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4827 for (node
= 0; node
< local_node
; node
++) {
4828 if (!node_online(node
))
4830 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4833 zonelist
->_zonerefs
[j
].zone
= NULL
;
4834 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4837 #endif /* CONFIG_NUMA */
4840 * Boot pageset table. One per cpu which is going to be used for all
4841 * zones and all nodes. The parameters will be set in such a way
4842 * that an item put on a list will immediately be handed over to
4843 * the buddy list. This is safe since pageset manipulation is done
4844 * with interrupts disabled.
4846 * The boot_pagesets must be kept even after bootup is complete for
4847 * unused processors and/or zones. They do play a role for bootstrapping
4848 * hotplugged processors.
4850 * zoneinfo_show() and maybe other functions do
4851 * not check if the processor is online before following the pageset pointer.
4852 * Other parts of the kernel may not check if the zone is available.
4854 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4855 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4856 static void setup_zone_pageset(struct zone
*zone
);
4859 * Global mutex to protect against size modification of zonelists
4860 * as well as to serialize pageset setup for the new populated zone.
4862 DEFINE_MUTEX(zonelists_mutex
);
4864 /* return values int ....just for stop_machine() */
4865 static int __build_all_zonelists(void *data
)
4869 pg_data_t
*self
= data
;
4872 memset(node_load
, 0, sizeof(node_load
));
4875 if (self
&& !node_online(self
->node_id
)) {
4876 build_zonelists(self
);
4879 for_each_online_node(nid
) {
4880 pg_data_t
*pgdat
= NODE_DATA(nid
);
4882 build_zonelists(pgdat
);
4886 * Initialize the boot_pagesets that are going to be used
4887 * for bootstrapping processors. The real pagesets for
4888 * each zone will be allocated later when the per cpu
4889 * allocator is available.
4891 * boot_pagesets are used also for bootstrapping offline
4892 * cpus if the system is already booted because the pagesets
4893 * are needed to initialize allocators on a specific cpu too.
4894 * F.e. the percpu allocator needs the page allocator which
4895 * needs the percpu allocator in order to allocate its pagesets
4896 * (a chicken-egg dilemma).
4898 for_each_possible_cpu(cpu
) {
4899 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4901 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4903 * We now know the "local memory node" for each node--
4904 * i.e., the node of the first zone in the generic zonelist.
4905 * Set up numa_mem percpu variable for on-line cpus. During
4906 * boot, only the boot cpu should be on-line; we'll init the
4907 * secondary cpus' numa_mem as they come on-line. During
4908 * node/memory hotplug, we'll fixup all on-line cpus.
4910 if (cpu_online(cpu
))
4911 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4918 static noinline
void __init
4919 build_all_zonelists_init(void)
4921 __build_all_zonelists(NULL
);
4922 mminit_verify_zonelist();
4923 cpuset_init_current_mems_allowed();
4927 * Called with zonelists_mutex held always
4928 * unless system_state == SYSTEM_BOOTING.
4930 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4931 * [we're only called with non-NULL zone through __meminit paths] and
4932 * (2) call of __init annotated helper build_all_zonelists_init
4933 * [protected by SYSTEM_BOOTING].
4935 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4937 set_zonelist_order();
4939 if (system_state
== SYSTEM_BOOTING
) {
4940 build_all_zonelists_init();
4942 #ifdef CONFIG_MEMORY_HOTPLUG
4944 setup_zone_pageset(zone
);
4946 /* we have to stop all cpus to guarantee there is no user
4948 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4949 /* cpuset refresh routine should be here */
4951 vm_total_pages
= nr_free_pagecache_pages();
4953 * Disable grouping by mobility if the number of pages in the
4954 * system is too low to allow the mechanism to work. It would be
4955 * more accurate, but expensive to check per-zone. This check is
4956 * made on memory-hotadd so a system can start with mobility
4957 * disabled and enable it later
4959 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4960 page_group_by_mobility_disabled
= 1;
4962 page_group_by_mobility_disabled
= 0;
4964 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4966 zonelist_order_name
[current_zonelist_order
],
4967 page_group_by_mobility_disabled
? "off" : "on",
4970 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4975 * Helper functions to size the waitqueue hash table.
4976 * Essentially these want to choose hash table sizes sufficiently
4977 * large so that collisions trying to wait on pages are rare.
4978 * But in fact, the number of active page waitqueues on typical
4979 * systems is ridiculously low, less than 200. So this is even
4980 * conservative, even though it seems large.
4982 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4983 * waitqueues, i.e. the size of the waitq table given the number of pages.
4985 #define PAGES_PER_WAITQUEUE 256
4987 #ifndef CONFIG_MEMORY_HOTPLUG
4988 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4990 unsigned long size
= 1;
4992 pages
/= PAGES_PER_WAITQUEUE
;
4994 while (size
< pages
)
4998 * Once we have dozens or even hundreds of threads sleeping
4999 * on IO we've got bigger problems than wait queue collision.
5000 * Limit the size of the wait table to a reasonable size.
5002 size
= min(size
, 4096UL);
5004 return max(size
, 4UL);
5008 * A zone's size might be changed by hot-add, so it is not possible to determine
5009 * a suitable size for its wait_table. So we use the maximum size now.
5011 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5013 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5014 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5015 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5017 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5018 * or more by the traditional way. (See above). It equals:
5020 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5021 * ia64(16K page size) : = ( 8G + 4M)byte.
5022 * powerpc (64K page size) : = (32G +16M)byte.
5024 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
5031 * This is an integer logarithm so that shifts can be used later
5032 * to extract the more random high bits from the multiplicative
5033 * hash function before the remainder is taken.
5035 static inline unsigned long wait_table_bits(unsigned long size
)
5041 * Initially all pages are reserved - free ones are freed
5042 * up by free_all_bootmem() once the early boot process is
5043 * done. Non-atomic initialization, single-pass.
5045 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5046 unsigned long start_pfn
, enum memmap_context context
)
5048 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5049 unsigned long end_pfn
= start_pfn
+ size
;
5050 pg_data_t
*pgdat
= NODE_DATA(nid
);
5052 unsigned long nr_initialised
= 0;
5053 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5054 struct memblock_region
*r
= NULL
, *tmp
;
5057 if (highest_memmap_pfn
< end_pfn
- 1)
5058 highest_memmap_pfn
= end_pfn
- 1;
5061 * Honor reservation requested by the driver for this ZONE_DEVICE
5064 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5065 start_pfn
+= altmap
->reserve
;
5067 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5069 * There can be holes in boot-time mem_map[]s handed to this
5070 * function. They do not exist on hotplugged memory.
5072 if (context
!= MEMMAP_EARLY
)
5075 if (!early_pfn_valid(pfn
))
5077 if (!early_pfn_in_nid(pfn
, nid
))
5079 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5082 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5084 * Check given memblock attribute by firmware which can affect
5085 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5086 * mirrored, it's an overlapped memmap init. skip it.
5088 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5089 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5090 for_each_memblock(memory
, tmp
)
5091 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5095 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5096 memblock_is_mirror(r
)) {
5097 /* already initialized as NORMAL */
5098 pfn
= memblock_region_memory_end_pfn(r
);
5106 * Mark the block movable so that blocks are reserved for
5107 * movable at startup. This will force kernel allocations
5108 * to reserve their blocks rather than leaking throughout
5109 * the address space during boot when many long-lived
5110 * kernel allocations are made.
5112 * bitmap is created for zone's valid pfn range. but memmap
5113 * can be created for invalid pages (for alignment)
5114 * check here not to call set_pageblock_migratetype() against
5117 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5118 struct page
*page
= pfn_to_page(pfn
);
5120 __init_single_page(page
, pfn
, zone
, nid
);
5121 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5123 __init_single_pfn(pfn
, zone
, nid
);
5128 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5130 unsigned int order
, t
;
5131 for_each_migratetype_order(order
, t
) {
5132 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5133 zone
->free_area
[order
].nr_free
= 0;
5137 #ifndef __HAVE_ARCH_MEMMAP_INIT
5138 #define memmap_init(size, nid, zone, start_pfn) \
5139 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5142 static int zone_batchsize(struct zone
*zone
)
5148 * The per-cpu-pages pools are set to around 1000th of the
5149 * size of the zone. But no more than 1/2 of a meg.
5151 * OK, so we don't know how big the cache is. So guess.
5153 batch
= zone
->managed_pages
/ 1024;
5154 if (batch
* PAGE_SIZE
> 512 * 1024)
5155 batch
= (512 * 1024) / PAGE_SIZE
;
5156 batch
/= 4; /* We effectively *= 4 below */
5161 * Clamp the batch to a 2^n - 1 value. Having a power
5162 * of 2 value was found to be more likely to have
5163 * suboptimal cache aliasing properties in some cases.
5165 * For example if 2 tasks are alternately allocating
5166 * batches of pages, one task can end up with a lot
5167 * of pages of one half of the possible page colors
5168 * and the other with pages of the other colors.
5170 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5175 /* The deferral and batching of frees should be suppressed under NOMMU
5178 * The problem is that NOMMU needs to be able to allocate large chunks
5179 * of contiguous memory as there's no hardware page translation to
5180 * assemble apparent contiguous memory from discontiguous pages.
5182 * Queueing large contiguous runs of pages for batching, however,
5183 * causes the pages to actually be freed in smaller chunks. As there
5184 * can be a significant delay between the individual batches being
5185 * recycled, this leads to the once large chunks of space being
5186 * fragmented and becoming unavailable for high-order allocations.
5193 * pcp->high and pcp->batch values are related and dependent on one another:
5194 * ->batch must never be higher then ->high.
5195 * The following function updates them in a safe manner without read side
5198 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5199 * those fields changing asynchronously (acording the the above rule).
5201 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5202 * outside of boot time (or some other assurance that no concurrent updaters
5205 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5206 unsigned long batch
)
5208 /* start with a fail safe value for batch */
5212 /* Update high, then batch, in order */
5219 /* a companion to pageset_set_high() */
5220 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5222 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5225 static void pageset_init(struct per_cpu_pageset
*p
)
5227 struct per_cpu_pages
*pcp
;
5230 memset(p
, 0, sizeof(*p
));
5234 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5235 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5238 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5241 pageset_set_batch(p
, batch
);
5245 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5246 * to the value high for the pageset p.
5248 static void pageset_set_high(struct per_cpu_pageset
*p
,
5251 unsigned long batch
= max(1UL, high
/ 4);
5252 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5253 batch
= PAGE_SHIFT
* 8;
5255 pageset_update(&p
->pcp
, high
, batch
);
5258 static void pageset_set_high_and_batch(struct zone
*zone
,
5259 struct per_cpu_pageset
*pcp
)
5261 if (percpu_pagelist_fraction
)
5262 pageset_set_high(pcp
,
5263 (zone
->managed_pages
/
5264 percpu_pagelist_fraction
));
5266 pageset_set_batch(pcp
, zone_batchsize(zone
));
5269 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5271 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5274 pageset_set_high_and_batch(zone
, pcp
);
5277 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5280 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5281 for_each_possible_cpu(cpu
)
5282 zone_pageset_init(zone
, cpu
);
5286 * Allocate per cpu pagesets and initialize them.
5287 * Before this call only boot pagesets were available.
5289 void __init
setup_per_cpu_pageset(void)
5291 struct pglist_data
*pgdat
;
5294 for_each_populated_zone(zone
)
5295 setup_zone_pageset(zone
);
5297 for_each_online_pgdat(pgdat
)
5298 pgdat
->per_cpu_nodestats
=
5299 alloc_percpu(struct per_cpu_nodestat
);
5302 static noinline __ref
5303 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
5309 * The per-page waitqueue mechanism uses hashed waitqueues
5312 zone
->wait_table_hash_nr_entries
=
5313 wait_table_hash_nr_entries(zone_size_pages
);
5314 zone
->wait_table_bits
=
5315 wait_table_bits(zone
->wait_table_hash_nr_entries
);
5316 alloc_size
= zone
->wait_table_hash_nr_entries
5317 * sizeof(wait_queue_head_t
);
5319 if (!slab_is_available()) {
5320 zone
->wait_table
= (wait_queue_head_t
*)
5321 memblock_virt_alloc_node_nopanic(
5322 alloc_size
, zone
->zone_pgdat
->node_id
);
5325 * This case means that a zone whose size was 0 gets new memory
5326 * via memory hot-add.
5327 * But it may be the case that a new node was hot-added. In
5328 * this case vmalloc() will not be able to use this new node's
5329 * memory - this wait_table must be initialized to use this new
5330 * node itself as well.
5331 * To use this new node's memory, further consideration will be
5334 zone
->wait_table
= vmalloc(alloc_size
);
5336 if (!zone
->wait_table
)
5339 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
5340 init_waitqueue_head(zone
->wait_table
+ i
);
5345 static __meminit
void zone_pcp_init(struct zone
*zone
)
5348 * per cpu subsystem is not up at this point. The following code
5349 * relies on the ability of the linker to provide the
5350 * offset of a (static) per cpu variable into the per cpu area.
5352 zone
->pageset
= &boot_pageset
;
5354 if (populated_zone(zone
))
5355 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5356 zone
->name
, zone
->present_pages
,
5357 zone_batchsize(zone
));
5360 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5361 unsigned long zone_start_pfn
,
5364 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5366 ret
= zone_wait_table_init(zone
, size
);
5369 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5371 zone
->zone_start_pfn
= zone_start_pfn
;
5373 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5374 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5376 (unsigned long)zone_idx(zone
),
5377 zone_start_pfn
, (zone_start_pfn
+ size
));
5379 zone_init_free_lists(zone
);
5384 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5385 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5388 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5390 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5391 struct mminit_pfnnid_cache
*state
)
5393 unsigned long start_pfn
, end_pfn
;
5396 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5397 return state
->last_nid
;
5399 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5401 state
->last_start
= start_pfn
;
5402 state
->last_end
= end_pfn
;
5403 state
->last_nid
= nid
;
5408 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5411 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5412 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5413 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5415 * If an architecture guarantees that all ranges registered contain no holes
5416 * and may be freed, this this function may be used instead of calling
5417 * memblock_free_early_nid() manually.
5419 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5421 unsigned long start_pfn
, end_pfn
;
5424 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5425 start_pfn
= min(start_pfn
, max_low_pfn
);
5426 end_pfn
= min(end_pfn
, max_low_pfn
);
5428 if (start_pfn
< end_pfn
)
5429 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5430 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5436 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5437 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5439 * If an architecture guarantees that all ranges registered contain no holes and may
5440 * be freed, this function may be used instead of calling memory_present() manually.
5442 void __init
sparse_memory_present_with_active_regions(int nid
)
5444 unsigned long start_pfn
, end_pfn
;
5447 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5448 memory_present(this_nid
, start_pfn
, end_pfn
);
5452 * get_pfn_range_for_nid - Return the start and end page frames for a node
5453 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5454 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5455 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5457 * It returns the start and end page frame of a node based on information
5458 * provided by memblock_set_node(). If called for a node
5459 * with no available memory, a warning is printed and the start and end
5462 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5463 unsigned long *start_pfn
, unsigned long *end_pfn
)
5465 unsigned long this_start_pfn
, this_end_pfn
;
5471 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5472 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5473 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5476 if (*start_pfn
== -1UL)
5481 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5482 * assumption is made that zones within a node are ordered in monotonic
5483 * increasing memory addresses so that the "highest" populated zone is used
5485 static void __init
find_usable_zone_for_movable(void)
5488 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5489 if (zone_index
== ZONE_MOVABLE
)
5492 if (arch_zone_highest_possible_pfn
[zone_index
] >
5493 arch_zone_lowest_possible_pfn
[zone_index
])
5497 VM_BUG_ON(zone_index
== -1);
5498 movable_zone
= zone_index
;
5502 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5503 * because it is sized independent of architecture. Unlike the other zones,
5504 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5505 * in each node depending on the size of each node and how evenly kernelcore
5506 * is distributed. This helper function adjusts the zone ranges
5507 * provided by the architecture for a given node by using the end of the
5508 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5509 * zones within a node are in order of monotonic increases memory addresses
5511 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5512 unsigned long zone_type
,
5513 unsigned long node_start_pfn
,
5514 unsigned long node_end_pfn
,
5515 unsigned long *zone_start_pfn
,
5516 unsigned long *zone_end_pfn
)
5518 /* Only adjust if ZONE_MOVABLE is on this node */
5519 if (zone_movable_pfn
[nid
]) {
5520 /* Size ZONE_MOVABLE */
5521 if (zone_type
== ZONE_MOVABLE
) {
5522 *zone_start_pfn
= zone_movable_pfn
[nid
];
5523 *zone_end_pfn
= min(node_end_pfn
,
5524 arch_zone_highest_possible_pfn
[movable_zone
]);
5526 /* Adjust for ZONE_MOVABLE starting within this range */
5527 } else if (!mirrored_kernelcore
&&
5528 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5529 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5530 *zone_end_pfn
= zone_movable_pfn
[nid
];
5532 /* Check if this whole range is within ZONE_MOVABLE */
5533 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5534 *zone_start_pfn
= *zone_end_pfn
;
5539 * Return the number of pages a zone spans in a node, including holes
5540 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5542 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5543 unsigned long zone_type
,
5544 unsigned long node_start_pfn
,
5545 unsigned long node_end_pfn
,
5546 unsigned long *zone_start_pfn
,
5547 unsigned long *zone_end_pfn
,
5548 unsigned long *ignored
)
5550 /* When hotadd a new node from cpu_up(), the node should be empty */
5551 if (!node_start_pfn
&& !node_end_pfn
)
5554 /* Get the start and end of the zone */
5555 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5556 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5557 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5558 node_start_pfn
, node_end_pfn
,
5559 zone_start_pfn
, zone_end_pfn
);
5561 /* Check that this node has pages within the zone's required range */
5562 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5565 /* Move the zone boundaries inside the node if necessary */
5566 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5567 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5569 /* Return the spanned pages */
5570 return *zone_end_pfn
- *zone_start_pfn
;
5574 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5575 * then all holes in the requested range will be accounted for.
5577 unsigned long __meminit
__absent_pages_in_range(int nid
,
5578 unsigned long range_start_pfn
,
5579 unsigned long range_end_pfn
)
5581 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5582 unsigned long start_pfn
, end_pfn
;
5585 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5586 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5587 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5588 nr_absent
-= end_pfn
- start_pfn
;
5594 * absent_pages_in_range - Return number of page frames in holes within a range
5595 * @start_pfn: The start PFN to start searching for holes
5596 * @end_pfn: The end PFN to stop searching for holes
5598 * It returns the number of pages frames in memory holes within a range.
5600 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5601 unsigned long end_pfn
)
5603 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5606 /* Return the number of page frames in holes in a zone on a node */
5607 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5608 unsigned long zone_type
,
5609 unsigned long node_start_pfn
,
5610 unsigned long node_end_pfn
,
5611 unsigned long *ignored
)
5613 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5614 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5615 unsigned long zone_start_pfn
, zone_end_pfn
;
5616 unsigned long nr_absent
;
5618 /* When hotadd a new node from cpu_up(), the node should be empty */
5619 if (!node_start_pfn
&& !node_end_pfn
)
5622 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5623 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5625 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5626 node_start_pfn
, node_end_pfn
,
5627 &zone_start_pfn
, &zone_end_pfn
);
5628 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5631 * ZONE_MOVABLE handling.
5632 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5635 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5636 unsigned long start_pfn
, end_pfn
;
5637 struct memblock_region
*r
;
5639 for_each_memblock(memory
, r
) {
5640 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5641 zone_start_pfn
, zone_end_pfn
);
5642 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5643 zone_start_pfn
, zone_end_pfn
);
5645 if (zone_type
== ZONE_MOVABLE
&&
5646 memblock_is_mirror(r
))
5647 nr_absent
+= end_pfn
- start_pfn
;
5649 if (zone_type
== ZONE_NORMAL
&&
5650 !memblock_is_mirror(r
))
5651 nr_absent
+= end_pfn
- start_pfn
;
5658 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5659 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5660 unsigned long zone_type
,
5661 unsigned long node_start_pfn
,
5662 unsigned long node_end_pfn
,
5663 unsigned long *zone_start_pfn
,
5664 unsigned long *zone_end_pfn
,
5665 unsigned long *zones_size
)
5669 *zone_start_pfn
= node_start_pfn
;
5670 for (zone
= 0; zone
< zone_type
; zone
++)
5671 *zone_start_pfn
+= zones_size
[zone
];
5673 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5675 return zones_size
[zone_type
];
5678 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5679 unsigned long zone_type
,
5680 unsigned long node_start_pfn
,
5681 unsigned long node_end_pfn
,
5682 unsigned long *zholes_size
)
5687 return zholes_size
[zone_type
];
5690 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5692 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5693 unsigned long node_start_pfn
,
5694 unsigned long node_end_pfn
,
5695 unsigned long *zones_size
,
5696 unsigned long *zholes_size
)
5698 unsigned long realtotalpages
= 0, totalpages
= 0;
5701 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5702 struct zone
*zone
= pgdat
->node_zones
+ i
;
5703 unsigned long zone_start_pfn
, zone_end_pfn
;
5704 unsigned long size
, real_size
;
5706 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5712 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5713 node_start_pfn
, node_end_pfn
,
5716 zone
->zone_start_pfn
= zone_start_pfn
;
5718 zone
->zone_start_pfn
= 0;
5719 zone
->spanned_pages
= size
;
5720 zone
->present_pages
= real_size
;
5723 realtotalpages
+= real_size
;
5726 pgdat
->node_spanned_pages
= totalpages
;
5727 pgdat
->node_present_pages
= realtotalpages
;
5728 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5732 #ifndef CONFIG_SPARSEMEM
5734 * Calculate the size of the zone->blockflags rounded to an unsigned long
5735 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5736 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5737 * round what is now in bits to nearest long in bits, then return it in
5740 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5742 unsigned long usemapsize
;
5744 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5745 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5746 usemapsize
= usemapsize
>> pageblock_order
;
5747 usemapsize
*= NR_PAGEBLOCK_BITS
;
5748 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5750 return usemapsize
/ 8;
5753 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5755 unsigned long zone_start_pfn
,
5756 unsigned long zonesize
)
5758 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5759 zone
->pageblock_flags
= NULL
;
5761 zone
->pageblock_flags
=
5762 memblock_virt_alloc_node_nopanic(usemapsize
,
5766 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5767 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5768 #endif /* CONFIG_SPARSEMEM */
5770 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5772 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5773 void __paginginit
set_pageblock_order(void)
5777 /* Check that pageblock_nr_pages has not already been setup */
5778 if (pageblock_order
)
5781 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5782 order
= HUGETLB_PAGE_ORDER
;
5784 order
= MAX_ORDER
- 1;
5787 * Assume the largest contiguous order of interest is a huge page.
5788 * This value may be variable depending on boot parameters on IA64 and
5791 pageblock_order
= order
;
5793 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5796 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5797 * is unused as pageblock_order is set at compile-time. See
5798 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5801 void __paginginit
set_pageblock_order(void)
5805 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5807 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5808 unsigned long present_pages
)
5810 unsigned long pages
= spanned_pages
;
5813 * Provide a more accurate estimation if there are holes within
5814 * the zone and SPARSEMEM is in use. If there are holes within the
5815 * zone, each populated memory region may cost us one or two extra
5816 * memmap pages due to alignment because memmap pages for each
5817 * populated regions may not naturally algined on page boundary.
5818 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5820 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5821 IS_ENABLED(CONFIG_SPARSEMEM
))
5822 pages
= present_pages
;
5824 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5828 * Set up the zone data structures:
5829 * - mark all pages reserved
5830 * - mark all memory queues empty
5831 * - clear the memory bitmaps
5833 * NOTE: pgdat should get zeroed by caller.
5835 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5838 int nid
= pgdat
->node_id
;
5841 pgdat_resize_init(pgdat
);
5842 #ifdef CONFIG_NUMA_BALANCING
5843 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5844 pgdat
->numabalancing_migrate_nr_pages
= 0;
5845 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5847 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5848 spin_lock_init(&pgdat
->split_queue_lock
);
5849 INIT_LIST_HEAD(&pgdat
->split_queue
);
5850 pgdat
->split_queue_len
= 0;
5852 init_waitqueue_head(&pgdat
->kswapd_wait
);
5853 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5854 #ifdef CONFIG_COMPACTION
5855 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5857 pgdat_page_ext_init(pgdat
);
5858 spin_lock_init(&pgdat
->lru_lock
);
5859 lruvec_init(node_lruvec(pgdat
));
5861 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5862 struct zone
*zone
= pgdat
->node_zones
+ j
;
5863 unsigned long size
, realsize
, freesize
, memmap_pages
;
5864 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5866 size
= zone
->spanned_pages
;
5867 realsize
= freesize
= zone
->present_pages
;
5870 * Adjust freesize so that it accounts for how much memory
5871 * is used by this zone for memmap. This affects the watermark
5872 * and per-cpu initialisations
5874 memmap_pages
= calc_memmap_size(size
, realsize
);
5875 if (!is_highmem_idx(j
)) {
5876 if (freesize
>= memmap_pages
) {
5877 freesize
-= memmap_pages
;
5880 " %s zone: %lu pages used for memmap\n",
5881 zone_names
[j
], memmap_pages
);
5883 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5884 zone_names
[j
], memmap_pages
, freesize
);
5887 /* Account for reserved pages */
5888 if (j
== 0 && freesize
> dma_reserve
) {
5889 freesize
-= dma_reserve
;
5890 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5891 zone_names
[0], dma_reserve
);
5894 if (!is_highmem_idx(j
))
5895 nr_kernel_pages
+= freesize
;
5896 /* Charge for highmem memmap if there are enough kernel pages */
5897 else if (nr_kernel_pages
> memmap_pages
* 2)
5898 nr_kernel_pages
-= memmap_pages
;
5899 nr_all_pages
+= freesize
;
5902 * Set an approximate value for lowmem here, it will be adjusted
5903 * when the bootmem allocator frees pages into the buddy system.
5904 * And all highmem pages will be managed by the buddy system.
5906 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5910 zone
->name
= zone_names
[j
];
5911 zone
->zone_pgdat
= pgdat
;
5912 spin_lock_init(&zone
->lock
);
5913 zone_seqlock_init(zone
);
5914 zone_pcp_init(zone
);
5919 set_pageblock_order();
5920 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5921 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5923 memmap_init(size
, nid
, j
, zone_start_pfn
);
5927 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
5929 unsigned long __maybe_unused start
= 0;
5930 unsigned long __maybe_unused offset
= 0;
5932 /* Skip empty nodes */
5933 if (!pgdat
->node_spanned_pages
)
5936 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5937 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5938 offset
= pgdat
->node_start_pfn
- start
;
5939 /* ia64 gets its own node_mem_map, before this, without bootmem */
5940 if (!pgdat
->node_mem_map
) {
5941 unsigned long size
, end
;
5945 * The zone's endpoints aren't required to be MAX_ORDER
5946 * aligned but the node_mem_map endpoints must be in order
5947 * for the buddy allocator to function correctly.
5949 end
= pgdat_end_pfn(pgdat
);
5950 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5951 size
= (end
- start
) * sizeof(struct page
);
5952 map
= alloc_remap(pgdat
->node_id
, size
);
5954 map
= memblock_virt_alloc_node_nopanic(size
,
5956 pgdat
->node_mem_map
= map
+ offset
;
5958 #ifndef CONFIG_NEED_MULTIPLE_NODES
5960 * With no DISCONTIG, the global mem_map is just set as node 0's
5962 if (pgdat
== NODE_DATA(0)) {
5963 mem_map
= NODE_DATA(0)->node_mem_map
;
5964 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5965 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5967 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5970 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5973 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5974 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5976 pg_data_t
*pgdat
= NODE_DATA(nid
);
5977 unsigned long start_pfn
= 0;
5978 unsigned long end_pfn
= 0;
5980 /* pg_data_t should be reset to zero when it's allocated */
5981 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
5983 reset_deferred_meminit(pgdat
);
5984 pgdat
->node_id
= nid
;
5985 pgdat
->node_start_pfn
= node_start_pfn
;
5986 pgdat
->per_cpu_nodestats
= NULL
;
5987 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5988 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5989 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5990 (u64
)start_pfn
<< PAGE_SHIFT
,
5991 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5993 start_pfn
= node_start_pfn
;
5995 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5996 zones_size
, zholes_size
);
5998 alloc_node_mem_map(pgdat
);
5999 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6000 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6001 nid
, (unsigned long)pgdat
,
6002 (unsigned long)pgdat
->node_mem_map
);
6005 free_area_init_core(pgdat
);
6008 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6010 #if MAX_NUMNODES > 1
6012 * Figure out the number of possible node ids.
6014 void __init
setup_nr_node_ids(void)
6016 unsigned int highest
;
6018 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6019 nr_node_ids
= highest
+ 1;
6024 * node_map_pfn_alignment - determine the maximum internode alignment
6026 * This function should be called after node map is populated and sorted.
6027 * It calculates the maximum power of two alignment which can distinguish
6030 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6031 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6032 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6033 * shifted, 1GiB is enough and this function will indicate so.
6035 * This is used to test whether pfn -> nid mapping of the chosen memory
6036 * model has fine enough granularity to avoid incorrect mapping for the
6037 * populated node map.
6039 * Returns the determined alignment in pfn's. 0 if there is no alignment
6040 * requirement (single node).
6042 unsigned long __init
node_map_pfn_alignment(void)
6044 unsigned long accl_mask
= 0, last_end
= 0;
6045 unsigned long start
, end
, mask
;
6049 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6050 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6057 * Start with a mask granular enough to pin-point to the
6058 * start pfn and tick off bits one-by-one until it becomes
6059 * too coarse to separate the current node from the last.
6061 mask
= ~((1 << __ffs(start
)) - 1);
6062 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6065 /* accumulate all internode masks */
6069 /* convert mask to number of pages */
6070 return ~accl_mask
+ 1;
6073 /* Find the lowest pfn for a node */
6074 static unsigned long __init
find_min_pfn_for_node(int nid
)
6076 unsigned long min_pfn
= ULONG_MAX
;
6077 unsigned long start_pfn
;
6080 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6081 min_pfn
= min(min_pfn
, start_pfn
);
6083 if (min_pfn
== ULONG_MAX
) {
6084 pr_warn("Could not find start_pfn for node %d\n", nid
);
6092 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6094 * It returns the minimum PFN based on information provided via
6095 * memblock_set_node().
6097 unsigned long __init
find_min_pfn_with_active_regions(void)
6099 return find_min_pfn_for_node(MAX_NUMNODES
);
6103 * early_calculate_totalpages()
6104 * Sum pages in active regions for movable zone.
6105 * Populate N_MEMORY for calculating usable_nodes.
6107 static unsigned long __init
early_calculate_totalpages(void)
6109 unsigned long totalpages
= 0;
6110 unsigned long start_pfn
, end_pfn
;
6113 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6114 unsigned long pages
= end_pfn
- start_pfn
;
6116 totalpages
+= pages
;
6118 node_set_state(nid
, N_MEMORY
);
6124 * Find the PFN the Movable zone begins in each node. Kernel memory
6125 * is spread evenly between nodes as long as the nodes have enough
6126 * memory. When they don't, some nodes will have more kernelcore than
6129 static void __init
find_zone_movable_pfns_for_nodes(void)
6132 unsigned long usable_startpfn
;
6133 unsigned long kernelcore_node
, kernelcore_remaining
;
6134 /* save the state before borrow the nodemask */
6135 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6136 unsigned long totalpages
= early_calculate_totalpages();
6137 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6138 struct memblock_region
*r
;
6140 /* Need to find movable_zone earlier when movable_node is specified. */
6141 find_usable_zone_for_movable();
6144 * If movable_node is specified, ignore kernelcore and movablecore
6147 if (movable_node_is_enabled()) {
6148 for_each_memblock(memory
, r
) {
6149 if (!memblock_is_hotpluggable(r
))
6154 usable_startpfn
= PFN_DOWN(r
->base
);
6155 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6156 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6164 * If kernelcore=mirror is specified, ignore movablecore option
6166 if (mirrored_kernelcore
) {
6167 bool mem_below_4gb_not_mirrored
= false;
6169 for_each_memblock(memory
, r
) {
6170 if (memblock_is_mirror(r
))
6175 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6177 if (usable_startpfn
< 0x100000) {
6178 mem_below_4gb_not_mirrored
= true;
6182 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6183 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6187 if (mem_below_4gb_not_mirrored
)
6188 pr_warn("This configuration results in unmirrored kernel memory.");
6194 * If movablecore=nn[KMG] was specified, calculate what size of
6195 * kernelcore that corresponds so that memory usable for
6196 * any allocation type is evenly spread. If both kernelcore
6197 * and movablecore are specified, then the value of kernelcore
6198 * will be used for required_kernelcore if it's greater than
6199 * what movablecore would have allowed.
6201 if (required_movablecore
) {
6202 unsigned long corepages
;
6205 * Round-up so that ZONE_MOVABLE is at least as large as what
6206 * was requested by the user
6208 required_movablecore
=
6209 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6210 required_movablecore
= min(totalpages
, required_movablecore
);
6211 corepages
= totalpages
- required_movablecore
;
6213 required_kernelcore
= max(required_kernelcore
, corepages
);
6217 * If kernelcore was not specified or kernelcore size is larger
6218 * than totalpages, there is no ZONE_MOVABLE.
6220 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6223 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6224 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6227 /* Spread kernelcore memory as evenly as possible throughout nodes */
6228 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6229 for_each_node_state(nid
, N_MEMORY
) {
6230 unsigned long start_pfn
, end_pfn
;
6233 * Recalculate kernelcore_node if the division per node
6234 * now exceeds what is necessary to satisfy the requested
6235 * amount of memory for the kernel
6237 if (required_kernelcore
< kernelcore_node
)
6238 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6241 * As the map is walked, we track how much memory is usable
6242 * by the kernel using kernelcore_remaining. When it is
6243 * 0, the rest of the node is usable by ZONE_MOVABLE
6245 kernelcore_remaining
= kernelcore_node
;
6247 /* Go through each range of PFNs within this node */
6248 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6249 unsigned long size_pages
;
6251 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6252 if (start_pfn
>= end_pfn
)
6255 /* Account for what is only usable for kernelcore */
6256 if (start_pfn
< usable_startpfn
) {
6257 unsigned long kernel_pages
;
6258 kernel_pages
= min(end_pfn
, usable_startpfn
)
6261 kernelcore_remaining
-= min(kernel_pages
,
6262 kernelcore_remaining
);
6263 required_kernelcore
-= min(kernel_pages
,
6264 required_kernelcore
);
6266 /* Continue if range is now fully accounted */
6267 if (end_pfn
<= usable_startpfn
) {
6270 * Push zone_movable_pfn to the end so
6271 * that if we have to rebalance
6272 * kernelcore across nodes, we will
6273 * not double account here
6275 zone_movable_pfn
[nid
] = end_pfn
;
6278 start_pfn
= usable_startpfn
;
6282 * The usable PFN range for ZONE_MOVABLE is from
6283 * start_pfn->end_pfn. Calculate size_pages as the
6284 * number of pages used as kernelcore
6286 size_pages
= end_pfn
- start_pfn
;
6287 if (size_pages
> kernelcore_remaining
)
6288 size_pages
= kernelcore_remaining
;
6289 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6292 * Some kernelcore has been met, update counts and
6293 * break if the kernelcore for this node has been
6296 required_kernelcore
-= min(required_kernelcore
,
6298 kernelcore_remaining
-= size_pages
;
6299 if (!kernelcore_remaining
)
6305 * If there is still required_kernelcore, we do another pass with one
6306 * less node in the count. This will push zone_movable_pfn[nid] further
6307 * along on the nodes that still have memory until kernelcore is
6311 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6315 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6316 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6317 zone_movable_pfn
[nid
] =
6318 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6321 /* restore the node_state */
6322 node_states
[N_MEMORY
] = saved_node_state
;
6325 /* Any regular or high memory on that node ? */
6326 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6328 enum zone_type zone_type
;
6330 if (N_MEMORY
== N_NORMAL_MEMORY
)
6333 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6334 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6335 if (populated_zone(zone
)) {
6336 node_set_state(nid
, N_HIGH_MEMORY
);
6337 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6338 zone_type
<= ZONE_NORMAL
)
6339 node_set_state(nid
, N_NORMAL_MEMORY
);
6346 * free_area_init_nodes - Initialise all pg_data_t and zone data
6347 * @max_zone_pfn: an array of max PFNs for each zone
6349 * This will call free_area_init_node() for each active node in the system.
6350 * Using the page ranges provided by memblock_set_node(), the size of each
6351 * zone in each node and their holes is calculated. If the maximum PFN
6352 * between two adjacent zones match, it is assumed that the zone is empty.
6353 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6354 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6355 * starts where the previous one ended. For example, ZONE_DMA32 starts
6356 * at arch_max_dma_pfn.
6358 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6360 unsigned long start_pfn
, end_pfn
;
6363 /* Record where the zone boundaries are */
6364 memset(arch_zone_lowest_possible_pfn
, 0,
6365 sizeof(arch_zone_lowest_possible_pfn
));
6366 memset(arch_zone_highest_possible_pfn
, 0,
6367 sizeof(arch_zone_highest_possible_pfn
));
6369 start_pfn
= find_min_pfn_with_active_regions();
6371 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6372 if (i
== ZONE_MOVABLE
)
6375 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6376 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6377 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6379 start_pfn
= end_pfn
;
6381 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6382 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6384 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6385 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6386 find_zone_movable_pfns_for_nodes();
6388 /* Print out the zone ranges */
6389 pr_info("Zone ranges:\n");
6390 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6391 if (i
== ZONE_MOVABLE
)
6393 pr_info(" %-8s ", zone_names
[i
]);
6394 if (arch_zone_lowest_possible_pfn
[i
] ==
6395 arch_zone_highest_possible_pfn
[i
])
6398 pr_cont("[mem %#018Lx-%#018Lx]\n",
6399 (u64
)arch_zone_lowest_possible_pfn
[i
]
6401 ((u64
)arch_zone_highest_possible_pfn
[i
]
6402 << PAGE_SHIFT
) - 1);
6405 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6406 pr_info("Movable zone start for each node\n");
6407 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6408 if (zone_movable_pfn
[i
])
6409 pr_info(" Node %d: %#018Lx\n", i
,
6410 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6413 /* Print out the early node map */
6414 pr_info("Early memory node ranges\n");
6415 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6416 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6417 (u64
)start_pfn
<< PAGE_SHIFT
,
6418 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6420 /* Initialise every node */
6421 mminit_verify_pageflags_layout();
6422 setup_nr_node_ids();
6423 for_each_online_node(nid
) {
6424 pg_data_t
*pgdat
= NODE_DATA(nid
);
6425 free_area_init_node(nid
, NULL
,
6426 find_min_pfn_for_node(nid
), NULL
);
6428 /* Any memory on that node */
6429 if (pgdat
->node_present_pages
)
6430 node_set_state(nid
, N_MEMORY
);
6431 check_for_memory(pgdat
, nid
);
6435 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6437 unsigned long long coremem
;
6441 coremem
= memparse(p
, &p
);
6442 *core
= coremem
>> PAGE_SHIFT
;
6444 /* Paranoid check that UL is enough for the coremem value */
6445 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6451 * kernelcore=size sets the amount of memory for use for allocations that
6452 * cannot be reclaimed or migrated.
6454 static int __init
cmdline_parse_kernelcore(char *p
)
6456 /* parse kernelcore=mirror */
6457 if (parse_option_str(p
, "mirror")) {
6458 mirrored_kernelcore
= true;
6462 return cmdline_parse_core(p
, &required_kernelcore
);
6466 * movablecore=size sets the amount of memory for use for allocations that
6467 * can be reclaimed or migrated.
6469 static int __init
cmdline_parse_movablecore(char *p
)
6471 return cmdline_parse_core(p
, &required_movablecore
);
6474 early_param("kernelcore", cmdline_parse_kernelcore
);
6475 early_param("movablecore", cmdline_parse_movablecore
);
6477 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6479 void adjust_managed_page_count(struct page
*page
, long count
)
6481 spin_lock(&managed_page_count_lock
);
6482 page_zone(page
)->managed_pages
+= count
;
6483 totalram_pages
+= count
;
6484 #ifdef CONFIG_HIGHMEM
6485 if (PageHighMem(page
))
6486 totalhigh_pages
+= count
;
6488 spin_unlock(&managed_page_count_lock
);
6490 EXPORT_SYMBOL(adjust_managed_page_count
);
6492 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6495 unsigned long pages
= 0;
6497 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6498 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6499 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6500 if ((unsigned int)poison
<= 0xFF)
6501 memset(pos
, poison
, PAGE_SIZE
);
6502 free_reserved_page(virt_to_page(pos
));
6506 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6507 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6511 EXPORT_SYMBOL(free_reserved_area
);
6513 #ifdef CONFIG_HIGHMEM
6514 void free_highmem_page(struct page
*page
)
6516 __free_reserved_page(page
);
6518 page_zone(page
)->managed_pages
++;
6524 void __init
mem_init_print_info(const char *str
)
6526 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6527 unsigned long init_code_size
, init_data_size
;
6529 physpages
= get_num_physpages();
6530 codesize
= _etext
- _stext
;
6531 datasize
= _edata
- _sdata
;
6532 rosize
= __end_rodata
- __start_rodata
;
6533 bss_size
= __bss_stop
- __bss_start
;
6534 init_data_size
= __init_end
- __init_begin
;
6535 init_code_size
= _einittext
- _sinittext
;
6538 * Detect special cases and adjust section sizes accordingly:
6539 * 1) .init.* may be embedded into .data sections
6540 * 2) .init.text.* may be out of [__init_begin, __init_end],
6541 * please refer to arch/tile/kernel/vmlinux.lds.S.
6542 * 3) .rodata.* may be embedded into .text or .data sections.
6544 #define adj_init_size(start, end, size, pos, adj) \
6546 if (start <= pos && pos < end && size > adj) \
6550 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6551 _sinittext
, init_code_size
);
6552 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6553 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6554 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6555 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6557 #undef adj_init_size
6559 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6560 #ifdef CONFIG_HIGHMEM
6564 nr_free_pages() << (PAGE_SHIFT
- 10),
6565 physpages
<< (PAGE_SHIFT
- 10),
6566 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6567 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6568 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6569 totalcma_pages
<< (PAGE_SHIFT
- 10),
6570 #ifdef CONFIG_HIGHMEM
6571 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6573 str
? ", " : "", str
? str
: "");
6577 * set_dma_reserve - set the specified number of pages reserved in the first zone
6578 * @new_dma_reserve: The number of pages to mark reserved
6580 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6581 * In the DMA zone, a significant percentage may be consumed by kernel image
6582 * and other unfreeable allocations which can skew the watermarks badly. This
6583 * function may optionally be used to account for unfreeable pages in the
6584 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6585 * smaller per-cpu batchsize.
6587 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6589 dma_reserve
= new_dma_reserve
;
6592 void __init
free_area_init(unsigned long *zones_size
)
6594 free_area_init_node(0, zones_size
,
6595 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6598 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6599 unsigned long action
, void *hcpu
)
6601 int cpu
= (unsigned long)hcpu
;
6603 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6604 lru_add_drain_cpu(cpu
);
6608 * Spill the event counters of the dead processor
6609 * into the current processors event counters.
6610 * This artificially elevates the count of the current
6613 vm_events_fold_cpu(cpu
);
6616 * Zero the differential counters of the dead processor
6617 * so that the vm statistics are consistent.
6619 * This is only okay since the processor is dead and cannot
6620 * race with what we are doing.
6622 cpu_vm_stats_fold(cpu
);
6627 void __init
page_alloc_init(void)
6629 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6633 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6634 * or min_free_kbytes changes.
6636 static void calculate_totalreserve_pages(void)
6638 struct pglist_data
*pgdat
;
6639 unsigned long reserve_pages
= 0;
6640 enum zone_type i
, j
;
6642 for_each_online_pgdat(pgdat
) {
6644 pgdat
->totalreserve_pages
= 0;
6646 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6647 struct zone
*zone
= pgdat
->node_zones
+ i
;
6650 /* Find valid and maximum lowmem_reserve in the zone */
6651 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6652 if (zone
->lowmem_reserve
[j
] > max
)
6653 max
= zone
->lowmem_reserve
[j
];
6656 /* we treat the high watermark as reserved pages. */
6657 max
+= high_wmark_pages(zone
);
6659 if (max
> zone
->managed_pages
)
6660 max
= zone
->managed_pages
;
6662 pgdat
->totalreserve_pages
+= max
;
6664 reserve_pages
+= max
;
6667 totalreserve_pages
= reserve_pages
;
6671 * setup_per_zone_lowmem_reserve - called whenever
6672 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6673 * has a correct pages reserved value, so an adequate number of
6674 * pages are left in the zone after a successful __alloc_pages().
6676 static void setup_per_zone_lowmem_reserve(void)
6678 struct pglist_data
*pgdat
;
6679 enum zone_type j
, idx
;
6681 for_each_online_pgdat(pgdat
) {
6682 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6683 struct zone
*zone
= pgdat
->node_zones
+ j
;
6684 unsigned long managed_pages
= zone
->managed_pages
;
6686 zone
->lowmem_reserve
[j
] = 0;
6690 struct zone
*lower_zone
;
6694 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6695 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6697 lower_zone
= pgdat
->node_zones
+ idx
;
6698 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6699 sysctl_lowmem_reserve_ratio
[idx
];
6700 managed_pages
+= lower_zone
->managed_pages
;
6705 /* update totalreserve_pages */
6706 calculate_totalreserve_pages();
6709 static void __setup_per_zone_wmarks(void)
6711 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6712 unsigned long lowmem_pages
= 0;
6714 unsigned long flags
;
6716 /* Calculate total number of !ZONE_HIGHMEM pages */
6717 for_each_zone(zone
) {
6718 if (!is_highmem(zone
))
6719 lowmem_pages
+= zone
->managed_pages
;
6722 for_each_zone(zone
) {
6725 spin_lock_irqsave(&zone
->lock
, flags
);
6726 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6727 do_div(tmp
, lowmem_pages
);
6728 if (is_highmem(zone
)) {
6730 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6731 * need highmem pages, so cap pages_min to a small
6734 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6735 * deltas control asynch page reclaim, and so should
6736 * not be capped for highmem.
6738 unsigned long min_pages
;
6740 min_pages
= zone
->managed_pages
/ 1024;
6741 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6742 zone
->watermark
[WMARK_MIN
] = min_pages
;
6745 * If it's a lowmem zone, reserve a number of pages
6746 * proportionate to the zone's size.
6748 zone
->watermark
[WMARK_MIN
] = tmp
;
6752 * Set the kswapd watermarks distance according to the
6753 * scale factor in proportion to available memory, but
6754 * ensure a minimum size on small systems.
6756 tmp
= max_t(u64
, tmp
>> 2,
6757 mult_frac(zone
->managed_pages
,
6758 watermark_scale_factor
, 10000));
6760 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6761 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6763 spin_unlock_irqrestore(&zone
->lock
, flags
);
6766 /* update totalreserve_pages */
6767 calculate_totalreserve_pages();
6771 * setup_per_zone_wmarks - called when min_free_kbytes changes
6772 * or when memory is hot-{added|removed}
6774 * Ensures that the watermark[min,low,high] values for each zone are set
6775 * correctly with respect to min_free_kbytes.
6777 void setup_per_zone_wmarks(void)
6779 mutex_lock(&zonelists_mutex
);
6780 __setup_per_zone_wmarks();
6781 mutex_unlock(&zonelists_mutex
);
6785 * Initialise min_free_kbytes.
6787 * For small machines we want it small (128k min). For large machines
6788 * we want it large (64MB max). But it is not linear, because network
6789 * bandwidth does not increase linearly with machine size. We use
6791 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6792 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6808 int __meminit
init_per_zone_wmark_min(void)
6810 unsigned long lowmem_kbytes
;
6811 int new_min_free_kbytes
;
6813 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6814 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6816 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6817 min_free_kbytes
= new_min_free_kbytes
;
6818 if (min_free_kbytes
< 128)
6819 min_free_kbytes
= 128;
6820 if (min_free_kbytes
> 65536)
6821 min_free_kbytes
= 65536;
6823 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6824 new_min_free_kbytes
, user_min_free_kbytes
);
6826 setup_per_zone_wmarks();
6827 refresh_zone_stat_thresholds();
6828 setup_per_zone_lowmem_reserve();
6831 setup_min_unmapped_ratio();
6832 setup_min_slab_ratio();
6837 core_initcall(init_per_zone_wmark_min
)
6840 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6841 * that we can call two helper functions whenever min_free_kbytes
6844 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6845 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6849 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6854 user_min_free_kbytes
= min_free_kbytes
;
6855 setup_per_zone_wmarks();
6860 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6861 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6865 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6870 setup_per_zone_wmarks();
6876 static void setup_min_unmapped_ratio(void)
6881 for_each_online_pgdat(pgdat
)
6882 pgdat
->min_unmapped_pages
= 0;
6885 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
6886 sysctl_min_unmapped_ratio
) / 100;
6890 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6891 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6895 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6899 setup_min_unmapped_ratio();
6904 static void setup_min_slab_ratio(void)
6909 for_each_online_pgdat(pgdat
)
6910 pgdat
->min_slab_pages
= 0;
6913 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
6914 sysctl_min_slab_ratio
) / 100;
6917 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6918 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6922 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6926 setup_min_slab_ratio();
6933 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6934 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6935 * whenever sysctl_lowmem_reserve_ratio changes.
6937 * The reserve ratio obviously has absolutely no relation with the
6938 * minimum watermarks. The lowmem reserve ratio can only make sense
6939 * if in function of the boot time zone sizes.
6941 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6942 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6944 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6945 setup_per_zone_lowmem_reserve();
6950 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6951 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6952 * pagelist can have before it gets flushed back to buddy allocator.
6954 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6955 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6958 int old_percpu_pagelist_fraction
;
6961 mutex_lock(&pcp_batch_high_lock
);
6962 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6964 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6965 if (!write
|| ret
< 0)
6968 /* Sanity checking to avoid pcp imbalance */
6969 if (percpu_pagelist_fraction
&&
6970 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6971 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6977 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6980 for_each_populated_zone(zone
) {
6983 for_each_possible_cpu(cpu
)
6984 pageset_set_high_and_batch(zone
,
6985 per_cpu_ptr(zone
->pageset
, cpu
));
6988 mutex_unlock(&pcp_batch_high_lock
);
6993 int hashdist
= HASHDIST_DEFAULT
;
6995 static int __init
set_hashdist(char *str
)
6999 hashdist
= simple_strtoul(str
, &str
, 0);
7002 __setup("hashdist=", set_hashdist
);
7005 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7007 * Returns the number of pages that arch has reserved but
7008 * is not known to alloc_large_system_hash().
7010 static unsigned long __init
arch_reserved_kernel_pages(void)
7017 * allocate a large system hash table from bootmem
7018 * - it is assumed that the hash table must contain an exact power-of-2
7019 * quantity of entries
7020 * - limit is the number of hash buckets, not the total allocation size
7022 void *__init
alloc_large_system_hash(const char *tablename
,
7023 unsigned long bucketsize
,
7024 unsigned long numentries
,
7027 unsigned int *_hash_shift
,
7028 unsigned int *_hash_mask
,
7029 unsigned long low_limit
,
7030 unsigned long high_limit
)
7032 unsigned long long max
= high_limit
;
7033 unsigned long log2qty
, size
;
7036 /* allow the kernel cmdline to have a say */
7038 /* round applicable memory size up to nearest megabyte */
7039 numentries
= nr_kernel_pages
;
7040 numentries
-= arch_reserved_kernel_pages();
7042 /* It isn't necessary when PAGE_SIZE >= 1MB */
7043 if (PAGE_SHIFT
< 20)
7044 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7046 /* limit to 1 bucket per 2^scale bytes of low memory */
7047 if (scale
> PAGE_SHIFT
)
7048 numentries
>>= (scale
- PAGE_SHIFT
);
7050 numentries
<<= (PAGE_SHIFT
- scale
);
7052 /* Make sure we've got at least a 0-order allocation.. */
7053 if (unlikely(flags
& HASH_SMALL
)) {
7054 /* Makes no sense without HASH_EARLY */
7055 WARN_ON(!(flags
& HASH_EARLY
));
7056 if (!(numentries
>> *_hash_shift
)) {
7057 numentries
= 1UL << *_hash_shift
;
7058 BUG_ON(!numentries
);
7060 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7061 numentries
= PAGE_SIZE
/ bucketsize
;
7063 numentries
= roundup_pow_of_two(numentries
);
7065 /* limit allocation size to 1/16 total memory by default */
7067 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7068 do_div(max
, bucketsize
);
7070 max
= min(max
, 0x80000000ULL
);
7072 if (numentries
< low_limit
)
7073 numentries
= low_limit
;
7074 if (numentries
> max
)
7077 log2qty
= ilog2(numentries
);
7080 size
= bucketsize
<< log2qty
;
7081 if (flags
& HASH_EARLY
)
7082 table
= memblock_virt_alloc_nopanic(size
, 0);
7084 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
7087 * If bucketsize is not a power-of-two, we may free
7088 * some pages at the end of hash table which
7089 * alloc_pages_exact() automatically does
7091 if (get_order(size
) < MAX_ORDER
) {
7092 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
7093 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
7096 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7099 panic("Failed to allocate %s hash table\n", tablename
);
7101 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7102 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7105 *_hash_shift
= log2qty
;
7107 *_hash_mask
= (1 << log2qty
) - 1;
7113 * This function checks whether pageblock includes unmovable pages or not.
7114 * If @count is not zero, it is okay to include less @count unmovable pages
7116 * PageLRU check without isolation or lru_lock could race so that
7117 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7118 * expect this function should be exact.
7120 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7121 bool skip_hwpoisoned_pages
)
7123 unsigned long pfn
, iter
, found
;
7127 * For avoiding noise data, lru_add_drain_all() should be called
7128 * If ZONE_MOVABLE, the zone never contains unmovable pages
7130 if (zone_idx(zone
) == ZONE_MOVABLE
)
7132 mt
= get_pageblock_migratetype(page
);
7133 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7136 pfn
= page_to_pfn(page
);
7137 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7138 unsigned long check
= pfn
+ iter
;
7140 if (!pfn_valid_within(check
))
7143 page
= pfn_to_page(check
);
7146 * Hugepages are not in LRU lists, but they're movable.
7147 * We need not scan over tail pages bacause we don't
7148 * handle each tail page individually in migration.
7150 if (PageHuge(page
)) {
7151 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7156 * We can't use page_count without pin a page
7157 * because another CPU can free compound page.
7158 * This check already skips compound tails of THP
7159 * because their page->_refcount is zero at all time.
7161 if (!page_ref_count(page
)) {
7162 if (PageBuddy(page
))
7163 iter
+= (1 << page_order(page
)) - 1;
7168 * The HWPoisoned page may be not in buddy system, and
7169 * page_count() is not 0.
7171 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7177 * If there are RECLAIMABLE pages, we need to check
7178 * it. But now, memory offline itself doesn't call
7179 * shrink_node_slabs() and it still to be fixed.
7182 * If the page is not RAM, page_count()should be 0.
7183 * we don't need more check. This is an _used_ not-movable page.
7185 * The problematic thing here is PG_reserved pages. PG_reserved
7186 * is set to both of a memory hole page and a _used_ kernel
7195 bool is_pageblock_removable_nolock(struct page
*page
)
7201 * We have to be careful here because we are iterating over memory
7202 * sections which are not zone aware so we might end up outside of
7203 * the zone but still within the section.
7204 * We have to take care about the node as well. If the node is offline
7205 * its NODE_DATA will be NULL - see page_zone.
7207 if (!node_online(page_to_nid(page
)))
7210 zone
= page_zone(page
);
7211 pfn
= page_to_pfn(page
);
7212 if (!zone_spans_pfn(zone
, pfn
))
7215 return !has_unmovable_pages(zone
, page
, 0, true);
7218 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7220 static unsigned long pfn_max_align_down(unsigned long pfn
)
7222 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7223 pageblock_nr_pages
) - 1);
7226 static unsigned long pfn_max_align_up(unsigned long pfn
)
7228 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7229 pageblock_nr_pages
));
7232 /* [start, end) must belong to a single zone. */
7233 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7234 unsigned long start
, unsigned long end
)
7236 /* This function is based on compact_zone() from compaction.c. */
7237 unsigned long nr_reclaimed
;
7238 unsigned long pfn
= start
;
7239 unsigned int tries
= 0;
7244 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7245 if (fatal_signal_pending(current
)) {
7250 if (list_empty(&cc
->migratepages
)) {
7251 cc
->nr_migratepages
= 0;
7252 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7258 } else if (++tries
== 5) {
7259 ret
= ret
< 0 ? ret
: -EBUSY
;
7263 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7265 cc
->nr_migratepages
-= nr_reclaimed
;
7267 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7268 NULL
, 0, cc
->mode
, MR_CMA
);
7271 putback_movable_pages(&cc
->migratepages
);
7278 * alloc_contig_range() -- tries to allocate given range of pages
7279 * @start: start PFN to allocate
7280 * @end: one-past-the-last PFN to allocate
7281 * @migratetype: migratetype of the underlaying pageblocks (either
7282 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7283 * in range must have the same migratetype and it must
7284 * be either of the two.
7286 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7287 * aligned, however it's the caller's responsibility to guarantee that
7288 * we are the only thread that changes migrate type of pageblocks the
7291 * The PFN range must belong to a single zone.
7293 * Returns zero on success or negative error code. On success all
7294 * pages which PFN is in [start, end) are allocated for the caller and
7295 * need to be freed with free_contig_range().
7297 int alloc_contig_range(unsigned long start
, unsigned long end
,
7298 unsigned migratetype
)
7300 unsigned long outer_start
, outer_end
;
7304 struct compact_control cc
= {
7305 .nr_migratepages
= 0,
7307 .zone
= page_zone(pfn_to_page(start
)),
7308 .mode
= MIGRATE_SYNC
,
7309 .ignore_skip_hint
= true,
7311 INIT_LIST_HEAD(&cc
.migratepages
);
7314 * What we do here is we mark all pageblocks in range as
7315 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7316 * have different sizes, and due to the way page allocator
7317 * work, we align the range to biggest of the two pages so
7318 * that page allocator won't try to merge buddies from
7319 * different pageblocks and change MIGRATE_ISOLATE to some
7320 * other migration type.
7322 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7323 * migrate the pages from an unaligned range (ie. pages that
7324 * we are interested in). This will put all the pages in
7325 * range back to page allocator as MIGRATE_ISOLATE.
7327 * When this is done, we take the pages in range from page
7328 * allocator removing them from the buddy system. This way
7329 * page allocator will never consider using them.
7331 * This lets us mark the pageblocks back as
7332 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7333 * aligned range but not in the unaligned, original range are
7334 * put back to page allocator so that buddy can use them.
7337 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7338 pfn_max_align_up(end
), migratetype
,
7344 * In case of -EBUSY, we'd like to know which page causes problem.
7345 * So, just fall through. We will check it in test_pages_isolated().
7347 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7348 if (ret
&& ret
!= -EBUSY
)
7352 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7353 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7354 * more, all pages in [start, end) are free in page allocator.
7355 * What we are going to do is to allocate all pages from
7356 * [start, end) (that is remove them from page allocator).
7358 * The only problem is that pages at the beginning and at the
7359 * end of interesting range may be not aligned with pages that
7360 * page allocator holds, ie. they can be part of higher order
7361 * pages. Because of this, we reserve the bigger range and
7362 * once this is done free the pages we are not interested in.
7364 * We don't have to hold zone->lock here because the pages are
7365 * isolated thus they won't get removed from buddy.
7368 lru_add_drain_all();
7369 drain_all_pages(cc
.zone
);
7372 outer_start
= start
;
7373 while (!PageBuddy(pfn_to_page(outer_start
))) {
7374 if (++order
>= MAX_ORDER
) {
7375 outer_start
= start
;
7378 outer_start
&= ~0UL << order
;
7381 if (outer_start
!= start
) {
7382 order
= page_order(pfn_to_page(outer_start
));
7385 * outer_start page could be small order buddy page and
7386 * it doesn't include start page. Adjust outer_start
7387 * in this case to report failed page properly
7388 * on tracepoint in test_pages_isolated()
7390 if (outer_start
+ (1UL << order
) <= start
)
7391 outer_start
= start
;
7394 /* Make sure the range is really isolated. */
7395 if (test_pages_isolated(outer_start
, end
, false)) {
7396 pr_info("%s: [%lx, %lx) PFNs busy\n",
7397 __func__
, outer_start
, end
);
7402 /* Grab isolated pages from freelists. */
7403 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7409 /* Free head and tail (if any) */
7410 if (start
!= outer_start
)
7411 free_contig_range(outer_start
, start
- outer_start
);
7412 if (end
!= outer_end
)
7413 free_contig_range(end
, outer_end
- end
);
7416 undo_isolate_page_range(pfn_max_align_down(start
),
7417 pfn_max_align_up(end
), migratetype
);
7421 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7423 unsigned int count
= 0;
7425 for (; nr_pages
--; pfn
++) {
7426 struct page
*page
= pfn_to_page(pfn
);
7428 count
+= page_count(page
) != 1;
7431 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7435 #ifdef CONFIG_MEMORY_HOTPLUG
7437 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7438 * page high values need to be recalulated.
7440 void __meminit
zone_pcp_update(struct zone
*zone
)
7443 mutex_lock(&pcp_batch_high_lock
);
7444 for_each_possible_cpu(cpu
)
7445 pageset_set_high_and_batch(zone
,
7446 per_cpu_ptr(zone
->pageset
, cpu
));
7447 mutex_unlock(&pcp_batch_high_lock
);
7451 void zone_pcp_reset(struct zone
*zone
)
7453 unsigned long flags
;
7455 struct per_cpu_pageset
*pset
;
7457 /* avoid races with drain_pages() */
7458 local_irq_save(flags
);
7459 if (zone
->pageset
!= &boot_pageset
) {
7460 for_each_online_cpu(cpu
) {
7461 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7462 drain_zonestat(zone
, pset
);
7464 free_percpu(zone
->pageset
);
7465 zone
->pageset
= &boot_pageset
;
7467 local_irq_restore(flags
);
7470 #ifdef CONFIG_MEMORY_HOTREMOVE
7472 * All pages in the range must be in a single zone and isolated
7473 * before calling this.
7476 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7480 unsigned int order
, i
;
7482 unsigned long flags
;
7483 /* find the first valid pfn */
7484 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7489 zone
= page_zone(pfn_to_page(pfn
));
7490 spin_lock_irqsave(&zone
->lock
, flags
);
7492 while (pfn
< end_pfn
) {
7493 if (!pfn_valid(pfn
)) {
7497 page
= pfn_to_page(pfn
);
7499 * The HWPoisoned page may be not in buddy system, and
7500 * page_count() is not 0.
7502 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7504 SetPageReserved(page
);
7508 BUG_ON(page_count(page
));
7509 BUG_ON(!PageBuddy(page
));
7510 order
= page_order(page
);
7511 #ifdef CONFIG_DEBUG_VM
7512 pr_info("remove from free list %lx %d %lx\n",
7513 pfn
, 1 << order
, end_pfn
);
7515 list_del(&page
->lru
);
7516 rmv_page_order(page
);
7517 zone
->free_area
[order
].nr_free
--;
7518 for (i
= 0; i
< (1 << order
); i
++)
7519 SetPageReserved((page
+i
));
7520 pfn
+= (1 << order
);
7522 spin_unlock_irqrestore(&zone
->lock
, flags
);
7526 bool is_free_buddy_page(struct page
*page
)
7528 struct zone
*zone
= page_zone(page
);
7529 unsigned long pfn
= page_to_pfn(page
);
7530 unsigned long flags
;
7533 spin_lock_irqsave(&zone
->lock
, flags
);
7534 for (order
= 0; order
< MAX_ORDER
; order
++) {
7535 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7537 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7540 spin_unlock_irqrestore(&zone
->lock
, flags
);
7542 return order
< MAX_ORDER
;