2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
68 #include <asm/sections.h>
69 #include <asm/tlbflush.h>
70 #include <asm/div64.h>
73 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74 static DEFINE_MUTEX(pcp_batch_high_lock
);
75 #define MIN_PERCPU_PAGELIST_FRACTION (8)
77 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78 DEFINE_PER_CPU(int, numa_node
);
79 EXPORT_PER_CPU_SYMBOL(numa_node
);
82 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
89 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
90 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
91 int _node_numa_mem_
[MAX_NUMNODES
];
94 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
95 volatile unsigned long latent_entropy __latent_entropy
;
96 EXPORT_SYMBOL(latent_entropy
);
100 * Array of node states.
102 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
103 [N_POSSIBLE
] = NODE_MASK_ALL
,
104 [N_ONLINE
] = { { [0] = 1UL } },
106 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
107 #ifdef CONFIG_HIGHMEM
108 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
110 #ifdef CONFIG_MOVABLE_NODE
111 [N_MEMORY
] = { { [0] = 1UL } },
113 [N_CPU
] = { { [0] = 1UL } },
116 EXPORT_SYMBOL(node_states
);
118 /* Protect totalram_pages and zone->managed_pages */
119 static DEFINE_SPINLOCK(managed_page_count_lock
);
121 unsigned long totalram_pages __read_mostly
;
122 unsigned long totalreserve_pages __read_mostly
;
123 unsigned long totalcma_pages __read_mostly
;
125 int percpu_pagelist_fraction
;
126 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page
*page
)
141 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
143 page
->index
= migratetype
;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask
;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex
));
161 if (saved_gfp_mask
) {
162 gfp_allowed_mask
= saved_gfp_mask
;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex
));
170 WARN_ON(saved_gfp_mask
);
171 saved_gfp_mask
= gfp_allowed_mask
;
172 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly
;
187 static void __free_pages_ok(struct page
*page
, unsigned int order
);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages
);
215 static char * const zone_names
[MAX_NR_ZONES
] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 char * const migratetype_names
[MIGRATE_TYPES
] = {
240 #ifdef CONFIG_MEMORY_ISOLATION
245 compound_page_dtor
* const compound_page_dtors
[] = {
248 #ifdef CONFIG_HUGETLB_PAGE
251 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
256 int min_free_kbytes
= 1024;
257 int user_min_free_kbytes
= -1;
258 int watermark_scale_factor
= 10;
260 static unsigned long __meminitdata nr_kernel_pages
;
261 static unsigned long __meminitdata nr_all_pages
;
262 static unsigned long __meminitdata dma_reserve
;
264 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
265 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
266 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
267 static unsigned long __initdata required_kernelcore
;
268 static unsigned long __initdata required_movablecore
;
269 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
270 static bool mirrored_kernelcore
;
272 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
274 EXPORT_SYMBOL(movable_zone
);
275 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
278 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
279 int nr_online_nodes __read_mostly
= 1;
280 EXPORT_SYMBOL(nr_node_ids
);
281 EXPORT_SYMBOL(nr_online_nodes
);
284 int page_group_by_mobility_disabled __read_mostly
;
286 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
287 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
289 pgdat
->first_deferred_pfn
= ULONG_MAX
;
292 /* Returns true if the struct page for the pfn is uninitialised */
293 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
295 int nid
= early_pfn_to_nid(pfn
);
297 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
307 static inline bool update_defer_init(pg_data_t
*pgdat
,
308 unsigned long pfn
, unsigned long zone_end
,
309 unsigned long *nr_initialised
)
311 unsigned long max_initialise
;
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end
< pgdat_end_pfn(pgdat
))
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
320 max_initialise
= max(2UL << (30 - PAGE_SHIFT
),
321 (pgdat
->node_spanned_pages
>> 8));
324 if ((*nr_initialised
> max_initialise
) &&
325 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
326 pgdat
->first_deferred_pfn
= pfn
;
333 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
337 static inline bool early_page_uninitialised(unsigned long pfn
)
342 static inline bool update_defer_init(pg_data_t
*pgdat
,
343 unsigned long pfn
, unsigned long zone_end
,
344 unsigned long *nr_initialised
)
350 /* Return a pointer to the bitmap storing bits affecting a block of pages */
351 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
354 #ifdef CONFIG_SPARSEMEM
355 return __pfn_to_section(pfn
)->pageblock_flags
;
357 return page_zone(page
)->pageblock_flags
;
358 #endif /* CONFIG_SPARSEMEM */
361 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
363 #ifdef CONFIG_SPARSEMEM
364 pfn
&= (PAGES_PER_SECTION
-1);
365 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
367 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
368 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
369 #endif /* CONFIG_SPARSEMEM */
373 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
374 * @page: The page within the block of interest
375 * @pfn: The target page frame number
376 * @end_bitidx: The last bit of interest to retrieve
377 * @mask: mask of bits that the caller is interested in
379 * Return: pageblock_bits flags
381 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
383 unsigned long end_bitidx
,
386 unsigned long *bitmap
;
387 unsigned long bitidx
, word_bitidx
;
390 bitmap
= get_pageblock_bitmap(page
, pfn
);
391 bitidx
= pfn_to_bitidx(page
, pfn
);
392 word_bitidx
= bitidx
/ BITS_PER_LONG
;
393 bitidx
&= (BITS_PER_LONG
-1);
395 word
= bitmap
[word_bitidx
];
396 bitidx
+= end_bitidx
;
397 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
400 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
401 unsigned long end_bitidx
,
404 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
407 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
409 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
413 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
414 * @page: The page within the block of interest
415 * @flags: The flags to set
416 * @pfn: The target page frame number
417 * @end_bitidx: The last bit of interest
418 * @mask: mask of bits that the caller is interested in
420 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
422 unsigned long end_bitidx
,
425 unsigned long *bitmap
;
426 unsigned long bitidx
, word_bitidx
;
427 unsigned long old_word
, word
;
429 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
431 bitmap
= get_pageblock_bitmap(page
, pfn
);
432 bitidx
= pfn_to_bitidx(page
, pfn
);
433 word_bitidx
= bitidx
/ BITS_PER_LONG
;
434 bitidx
&= (BITS_PER_LONG
-1);
436 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
438 bitidx
+= end_bitidx
;
439 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
440 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
442 word
= READ_ONCE(bitmap
[word_bitidx
]);
444 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
445 if (word
== old_word
)
451 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
453 if (unlikely(page_group_by_mobility_disabled
&&
454 migratetype
< MIGRATE_PCPTYPES
))
455 migratetype
= MIGRATE_UNMOVABLE
;
457 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
458 PB_migrate
, PB_migrate_end
);
461 #ifdef CONFIG_DEBUG_VM
462 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
466 unsigned long pfn
= page_to_pfn(page
);
467 unsigned long sp
, start_pfn
;
470 seq
= zone_span_seqbegin(zone
);
471 start_pfn
= zone
->zone_start_pfn
;
472 sp
= zone
->spanned_pages
;
473 if (!zone_spans_pfn(zone
, pfn
))
475 } while (zone_span_seqretry(zone
, seq
));
478 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
479 pfn
, zone_to_nid(zone
), zone
->name
,
480 start_pfn
, start_pfn
+ sp
);
485 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
487 if (!pfn_valid_within(page_to_pfn(page
)))
489 if (zone
!= page_zone(page
))
495 * Temporary debugging check for pages not lying within a given zone.
497 static int bad_range(struct zone
*zone
, struct page
*page
)
499 if (page_outside_zone_boundaries(zone
, page
))
501 if (!page_is_consistent(zone
, page
))
507 static inline int bad_range(struct zone
*zone
, struct page
*page
)
513 static void bad_page(struct page
*page
, const char *reason
,
514 unsigned long bad_flags
)
516 static unsigned long resume
;
517 static unsigned long nr_shown
;
518 static unsigned long nr_unshown
;
521 * Allow a burst of 60 reports, then keep quiet for that minute;
522 * or allow a steady drip of one report per second.
524 if (nr_shown
== 60) {
525 if (time_before(jiffies
, resume
)) {
531 "BUG: Bad page state: %lu messages suppressed\n",
538 resume
= jiffies
+ 60 * HZ
;
540 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
541 current
->comm
, page_to_pfn(page
));
542 __dump_page(page
, reason
);
543 bad_flags
&= page
->flags
;
545 pr_alert("bad because of flags: %#lx(%pGp)\n",
546 bad_flags
, &bad_flags
);
547 dump_page_owner(page
);
552 /* Leave bad fields for debug, except PageBuddy could make trouble */
553 page_mapcount_reset(page
); /* remove PageBuddy */
554 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
558 * Higher-order pages are called "compound pages". They are structured thusly:
560 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
562 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
563 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
565 * The first tail page's ->compound_dtor holds the offset in array of compound
566 * page destructors. See compound_page_dtors.
568 * The first tail page's ->compound_order holds the order of allocation.
569 * This usage means that zero-order pages may not be compound.
572 void free_compound_page(struct page
*page
)
574 __free_pages_ok(page
, compound_order(page
));
577 void prep_compound_page(struct page
*page
, unsigned int order
)
580 int nr_pages
= 1 << order
;
582 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
583 set_compound_order(page
, order
);
585 for (i
= 1; i
< nr_pages
; i
++) {
586 struct page
*p
= page
+ i
;
587 set_page_count(p
, 0);
588 p
->mapping
= TAIL_MAPPING
;
589 set_compound_head(p
, page
);
591 atomic_set(compound_mapcount_ptr(page
), -1);
594 #ifdef CONFIG_DEBUG_PAGEALLOC
595 unsigned int _debug_guardpage_minorder
;
596 bool _debug_pagealloc_enabled __read_mostly
597 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
598 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
599 bool _debug_guardpage_enabled __read_mostly
;
601 static int __init
early_debug_pagealloc(char *buf
)
605 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
607 early_param("debug_pagealloc", early_debug_pagealloc
);
609 static bool need_debug_guardpage(void)
611 /* If we don't use debug_pagealloc, we don't need guard page */
612 if (!debug_pagealloc_enabled())
615 if (!debug_guardpage_minorder())
621 static void init_debug_guardpage(void)
623 if (!debug_pagealloc_enabled())
626 if (!debug_guardpage_minorder())
629 _debug_guardpage_enabled
= true;
632 struct page_ext_operations debug_guardpage_ops
= {
633 .need
= need_debug_guardpage
,
634 .init
= init_debug_guardpage
,
637 static int __init
debug_guardpage_minorder_setup(char *buf
)
641 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
642 pr_err("Bad debug_guardpage_minorder value\n");
645 _debug_guardpage_minorder
= res
;
646 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
649 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
651 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
652 unsigned int order
, int migratetype
)
654 struct page_ext
*page_ext
;
656 if (!debug_guardpage_enabled())
659 if (order
>= debug_guardpage_minorder())
662 page_ext
= lookup_page_ext(page
);
663 if (unlikely(!page_ext
))
666 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
668 INIT_LIST_HEAD(&page
->lru
);
669 set_page_private(page
, order
);
670 /* Guard pages are not available for any usage */
671 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
676 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
677 unsigned int order
, int migratetype
)
679 struct page_ext
*page_ext
;
681 if (!debug_guardpage_enabled())
684 page_ext
= lookup_page_ext(page
);
685 if (unlikely(!page_ext
))
688 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
690 set_page_private(page
, 0);
691 if (!is_migrate_isolate(migratetype
))
692 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
695 struct page_ext_operations debug_guardpage_ops
;
696 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
697 unsigned int order
, int migratetype
) { return false; }
698 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
699 unsigned int order
, int migratetype
) {}
702 static inline void set_page_order(struct page
*page
, unsigned int order
)
704 set_page_private(page
, order
);
705 __SetPageBuddy(page
);
708 static inline void rmv_page_order(struct page
*page
)
710 __ClearPageBuddy(page
);
711 set_page_private(page
, 0);
715 * This function checks whether a page is free && is the buddy
716 * we can do coalesce a page and its buddy if
717 * (a) the buddy is not in a hole &&
718 * (b) the buddy is in the buddy system &&
719 * (c) a page and its buddy have the same order &&
720 * (d) a page and its buddy are in the same zone.
722 * For recording whether a page is in the buddy system, we set ->_mapcount
723 * PAGE_BUDDY_MAPCOUNT_VALUE.
724 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
725 * serialized by zone->lock.
727 * For recording page's order, we use page_private(page).
729 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
732 if (!pfn_valid_within(page_to_pfn(buddy
)))
735 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
736 if (page_zone_id(page
) != page_zone_id(buddy
))
739 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
744 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
746 * zone check is done late to avoid uselessly
747 * calculating zone/node ids for pages that could
750 if (page_zone_id(page
) != page_zone_id(buddy
))
753 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
761 * Freeing function for a buddy system allocator.
763 * The concept of a buddy system is to maintain direct-mapped table
764 * (containing bit values) for memory blocks of various "orders".
765 * The bottom level table contains the map for the smallest allocatable
766 * units of memory (here, pages), and each level above it describes
767 * pairs of units from the levels below, hence, "buddies".
768 * At a high level, all that happens here is marking the table entry
769 * at the bottom level available, and propagating the changes upward
770 * as necessary, plus some accounting needed to play nicely with other
771 * parts of the VM system.
772 * At each level, we keep a list of pages, which are heads of continuous
773 * free pages of length of (1 << order) and marked with _mapcount
774 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
776 * So when we are allocating or freeing one, we can derive the state of the
777 * other. That is, if we allocate a small block, and both were
778 * free, the remainder of the region must be split into blocks.
779 * If a block is freed, and its buddy is also free, then this
780 * triggers coalescing into a block of larger size.
785 static inline void __free_one_page(struct page
*page
,
787 struct zone
*zone
, unsigned int order
,
790 unsigned long page_idx
;
791 unsigned long combined_idx
;
792 unsigned long uninitialized_var(buddy_idx
);
794 unsigned int max_order
;
796 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
798 VM_BUG_ON(!zone_is_initialized(zone
));
799 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
801 VM_BUG_ON(migratetype
== -1);
802 if (likely(!is_migrate_isolate(migratetype
)))
803 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
805 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
807 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
808 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
811 while (order
< max_order
- 1) {
812 buddy_idx
= __find_buddy_index(page_idx
, order
);
813 buddy
= page
+ (buddy_idx
- page_idx
);
814 if (!page_is_buddy(page
, buddy
, order
))
817 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
818 * merge with it and move up one order.
820 if (page_is_guard(buddy
)) {
821 clear_page_guard(zone
, buddy
, order
, migratetype
);
823 list_del(&buddy
->lru
);
824 zone
->free_area
[order
].nr_free
--;
825 rmv_page_order(buddy
);
827 combined_idx
= buddy_idx
& page_idx
;
828 page
= page
+ (combined_idx
- page_idx
);
829 page_idx
= combined_idx
;
832 if (max_order
< MAX_ORDER
) {
833 /* If we are here, it means order is >= pageblock_order.
834 * We want to prevent merge between freepages on isolate
835 * pageblock and normal pageblock. Without this, pageblock
836 * isolation could cause incorrect freepage or CMA accounting.
838 * We don't want to hit this code for the more frequent
841 if (unlikely(has_isolate_pageblock(zone
))) {
844 buddy_idx
= __find_buddy_index(page_idx
, order
);
845 buddy
= page
+ (buddy_idx
- page_idx
);
846 buddy_mt
= get_pageblock_migratetype(buddy
);
848 if (migratetype
!= buddy_mt
849 && (is_migrate_isolate(migratetype
) ||
850 is_migrate_isolate(buddy_mt
)))
854 goto continue_merging
;
858 set_page_order(page
, order
);
861 * If this is not the largest possible page, check if the buddy
862 * of the next-highest order is free. If it is, it's possible
863 * that pages are being freed that will coalesce soon. In case,
864 * that is happening, add the free page to the tail of the list
865 * so it's less likely to be used soon and more likely to be merged
866 * as a higher order page
868 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
869 struct page
*higher_page
, *higher_buddy
;
870 combined_idx
= buddy_idx
& page_idx
;
871 higher_page
= page
+ (combined_idx
- page_idx
);
872 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
873 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
874 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
875 list_add_tail(&page
->lru
,
876 &zone
->free_area
[order
].free_list
[migratetype
]);
881 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
883 zone
->free_area
[order
].nr_free
++;
887 * A bad page could be due to a number of fields. Instead of multiple branches,
888 * try and check multiple fields with one check. The caller must do a detailed
889 * check if necessary.
891 static inline bool page_expected_state(struct page
*page
,
892 unsigned long check_flags
)
894 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
897 if (unlikely((unsigned long)page
->mapping
|
898 page_ref_count(page
) |
900 (unsigned long)page
->mem_cgroup
|
902 (page
->flags
& check_flags
)))
908 static void free_pages_check_bad(struct page
*page
)
910 const char *bad_reason
;
911 unsigned long bad_flags
;
916 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
917 bad_reason
= "nonzero mapcount";
918 if (unlikely(page
->mapping
!= NULL
))
919 bad_reason
= "non-NULL mapping";
920 if (unlikely(page_ref_count(page
) != 0))
921 bad_reason
= "nonzero _refcount";
922 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
923 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
924 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
927 if (unlikely(page
->mem_cgroup
))
928 bad_reason
= "page still charged to cgroup";
930 bad_page(page
, bad_reason
, bad_flags
);
933 static inline int free_pages_check(struct page
*page
)
935 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
938 /* Something has gone sideways, find it */
939 free_pages_check_bad(page
);
943 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
948 * We rely page->lru.next never has bit 0 set, unless the page
949 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
951 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
953 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
957 switch (page
- head_page
) {
959 /* the first tail page: ->mapping is compound_mapcount() */
960 if (unlikely(compound_mapcount(page
))) {
961 bad_page(page
, "nonzero compound_mapcount", 0);
967 * the second tail page: ->mapping is
968 * page_deferred_list().next -- ignore value.
972 if (page
->mapping
!= TAIL_MAPPING
) {
973 bad_page(page
, "corrupted mapping in tail page", 0);
978 if (unlikely(!PageTail(page
))) {
979 bad_page(page
, "PageTail not set", 0);
982 if (unlikely(compound_head(page
) != head_page
)) {
983 bad_page(page
, "compound_head not consistent", 0);
988 page
->mapping
= NULL
;
989 clear_compound_head(page
);
993 static __always_inline
bool free_pages_prepare(struct page
*page
,
994 unsigned int order
, bool check_free
)
998 VM_BUG_ON_PAGE(PageTail(page
), page
);
1000 trace_mm_page_free(page
, order
);
1001 kmemcheck_free_shadow(page
, order
);
1004 * Check tail pages before head page information is cleared to
1005 * avoid checking PageCompound for order-0 pages.
1007 if (unlikely(order
)) {
1008 bool compound
= PageCompound(page
);
1011 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1014 ClearPageDoubleMap(page
);
1015 for (i
= 1; i
< (1 << order
); i
++) {
1017 bad
+= free_tail_pages_check(page
, page
+ i
);
1018 if (unlikely(free_pages_check(page
+ i
))) {
1022 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1025 if (PageMappingFlags(page
))
1026 page
->mapping
= NULL
;
1027 if (memcg_kmem_enabled() && PageKmemcg(page
))
1028 memcg_kmem_uncharge(page
, order
);
1030 bad
+= free_pages_check(page
);
1034 page_cpupid_reset_last(page
);
1035 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1036 reset_page_owner(page
, order
);
1038 if (!PageHighMem(page
)) {
1039 debug_check_no_locks_freed(page_address(page
),
1040 PAGE_SIZE
<< order
);
1041 debug_check_no_obj_freed(page_address(page
),
1042 PAGE_SIZE
<< order
);
1044 arch_free_page(page
, order
);
1045 kernel_poison_pages(page
, 1 << order
, 0);
1046 kernel_map_pages(page
, 1 << order
, 0);
1047 kasan_free_pages(page
, order
);
1052 #ifdef CONFIG_DEBUG_VM
1053 static inline bool free_pcp_prepare(struct page
*page
)
1055 return free_pages_prepare(page
, 0, true);
1058 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1063 static bool free_pcp_prepare(struct page
*page
)
1065 return free_pages_prepare(page
, 0, false);
1068 static bool bulkfree_pcp_prepare(struct page
*page
)
1070 return free_pages_check(page
);
1072 #endif /* CONFIG_DEBUG_VM */
1075 * Frees a number of pages from the PCP lists
1076 * Assumes all pages on list are in same zone, and of same order.
1077 * count is the number of pages to free.
1079 * If the zone was previously in an "all pages pinned" state then look to
1080 * see if this freeing clears that state.
1082 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1083 * pinned" detection logic.
1085 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1086 struct per_cpu_pages
*pcp
)
1088 int migratetype
= 0;
1090 unsigned long nr_scanned
;
1091 bool isolated_pageblocks
;
1093 spin_lock(&zone
->lock
);
1094 isolated_pageblocks
= has_isolate_pageblock(zone
);
1095 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1097 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1101 struct list_head
*list
;
1104 * Remove pages from lists in a round-robin fashion. A
1105 * batch_free count is maintained that is incremented when an
1106 * empty list is encountered. This is so more pages are freed
1107 * off fuller lists instead of spinning excessively around empty
1112 if (++migratetype
== MIGRATE_PCPTYPES
)
1114 list
= &pcp
->lists
[migratetype
];
1115 } while (list_empty(list
));
1117 /* This is the only non-empty list. Free them all. */
1118 if (batch_free
== MIGRATE_PCPTYPES
)
1122 int mt
; /* migratetype of the to-be-freed page */
1124 page
= list_last_entry(list
, struct page
, lru
);
1125 /* must delete as __free_one_page list manipulates */
1126 list_del(&page
->lru
);
1128 mt
= get_pcppage_migratetype(page
);
1129 /* MIGRATE_ISOLATE page should not go to pcplists */
1130 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1131 /* Pageblock could have been isolated meanwhile */
1132 if (unlikely(isolated_pageblocks
))
1133 mt
= get_pageblock_migratetype(page
);
1135 if (bulkfree_pcp_prepare(page
))
1138 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1139 trace_mm_page_pcpu_drain(page
, 0, mt
);
1140 } while (--count
&& --batch_free
&& !list_empty(list
));
1142 spin_unlock(&zone
->lock
);
1145 static void free_one_page(struct zone
*zone
,
1146 struct page
*page
, unsigned long pfn
,
1150 unsigned long nr_scanned
;
1151 spin_lock(&zone
->lock
);
1152 nr_scanned
= node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
);
1154 __mod_node_page_state(zone
->zone_pgdat
, NR_PAGES_SCANNED
, -nr_scanned
);
1156 if (unlikely(has_isolate_pageblock(zone
) ||
1157 is_migrate_isolate(migratetype
))) {
1158 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1160 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1161 spin_unlock(&zone
->lock
);
1164 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1165 unsigned long zone
, int nid
)
1167 set_page_links(page
, zone
, nid
, pfn
);
1168 init_page_count(page
);
1169 page_mapcount_reset(page
);
1170 page_cpupid_reset_last(page
);
1172 INIT_LIST_HEAD(&page
->lru
);
1173 #ifdef WANT_PAGE_VIRTUAL
1174 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1175 if (!is_highmem_idx(zone
))
1176 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1180 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1183 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1186 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1187 static void init_reserved_page(unsigned long pfn
)
1192 if (!early_page_uninitialised(pfn
))
1195 nid
= early_pfn_to_nid(pfn
);
1196 pgdat
= NODE_DATA(nid
);
1198 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1199 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1201 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1204 __init_single_pfn(pfn
, zid
, nid
);
1207 static inline void init_reserved_page(unsigned long pfn
)
1210 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1213 * Initialised pages do not have PageReserved set. This function is
1214 * called for each range allocated by the bootmem allocator and
1215 * marks the pages PageReserved. The remaining valid pages are later
1216 * sent to the buddy page allocator.
1218 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1220 unsigned long start_pfn
= PFN_DOWN(start
);
1221 unsigned long end_pfn
= PFN_UP(end
);
1223 for (; start_pfn
< end_pfn
; start_pfn
++) {
1224 if (pfn_valid(start_pfn
)) {
1225 struct page
*page
= pfn_to_page(start_pfn
);
1227 init_reserved_page(start_pfn
);
1229 /* Avoid false-positive PageTail() */
1230 INIT_LIST_HEAD(&page
->lru
);
1232 SetPageReserved(page
);
1237 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1239 unsigned long flags
;
1241 unsigned long pfn
= page_to_pfn(page
);
1243 if (!free_pages_prepare(page
, order
, true))
1246 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1247 local_irq_save(flags
);
1248 __count_vm_events(PGFREE
, 1 << order
);
1249 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1250 local_irq_restore(flags
);
1253 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1255 unsigned int nr_pages
= 1 << order
;
1256 struct page
*p
= page
;
1260 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1262 __ClearPageReserved(p
);
1263 set_page_count(p
, 0);
1265 __ClearPageReserved(p
);
1266 set_page_count(p
, 0);
1268 page_zone(page
)->managed_pages
+= nr_pages
;
1269 set_page_refcounted(page
);
1270 __free_pages(page
, order
);
1273 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1274 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1276 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1278 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1280 static DEFINE_SPINLOCK(early_pfn_lock
);
1283 spin_lock(&early_pfn_lock
);
1284 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1286 nid
= first_online_node
;
1287 spin_unlock(&early_pfn_lock
);
1293 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1294 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1295 struct mminit_pfnnid_cache
*state
)
1299 nid
= __early_pfn_to_nid(pfn
, state
);
1300 if (nid
>= 0 && nid
!= node
)
1305 /* Only safe to use early in boot when initialisation is single-threaded */
1306 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1308 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1313 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1317 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1318 struct mminit_pfnnid_cache
*state
)
1325 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1328 if (early_page_uninitialised(pfn
))
1330 return __free_pages_boot_core(page
, order
);
1334 * Check that the whole (or subset of) a pageblock given by the interval of
1335 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1336 * with the migration of free compaction scanner. The scanners then need to
1337 * use only pfn_valid_within() check for arches that allow holes within
1340 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1342 * It's possible on some configurations to have a setup like node0 node1 node0
1343 * i.e. it's possible that all pages within a zones range of pages do not
1344 * belong to a single zone. We assume that a border between node0 and node1
1345 * can occur within a single pageblock, but not a node0 node1 node0
1346 * interleaving within a single pageblock. It is therefore sufficient to check
1347 * the first and last page of a pageblock and avoid checking each individual
1348 * page in a pageblock.
1350 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1351 unsigned long end_pfn
, struct zone
*zone
)
1353 struct page
*start_page
;
1354 struct page
*end_page
;
1356 /* end_pfn is one past the range we are checking */
1359 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1362 start_page
= pfn_to_page(start_pfn
);
1364 if (page_zone(start_page
) != zone
)
1367 end_page
= pfn_to_page(end_pfn
);
1369 /* This gives a shorter code than deriving page_zone(end_page) */
1370 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1376 void set_zone_contiguous(struct zone
*zone
)
1378 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1379 unsigned long block_end_pfn
;
1381 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1382 for (; block_start_pfn
< zone_end_pfn(zone
);
1383 block_start_pfn
= block_end_pfn
,
1384 block_end_pfn
+= pageblock_nr_pages
) {
1386 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1388 if (!__pageblock_pfn_to_page(block_start_pfn
,
1389 block_end_pfn
, zone
))
1393 /* We confirm that there is no hole */
1394 zone
->contiguous
= true;
1397 void clear_zone_contiguous(struct zone
*zone
)
1399 zone
->contiguous
= false;
1402 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1403 static void __init
deferred_free_range(struct page
*page
,
1404 unsigned long pfn
, int nr_pages
)
1411 /* Free a large naturally-aligned chunk if possible */
1412 if (nr_pages
== pageblock_nr_pages
&&
1413 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1414 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1415 __free_pages_boot_core(page
, pageblock_order
);
1419 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1420 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1421 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1422 __free_pages_boot_core(page
, 0);
1426 /* Completion tracking for deferred_init_memmap() threads */
1427 static atomic_t pgdat_init_n_undone __initdata
;
1428 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1430 static inline void __init
pgdat_init_report_one_done(void)
1432 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1433 complete(&pgdat_init_all_done_comp
);
1436 /* Initialise remaining memory on a node */
1437 static int __init
deferred_init_memmap(void *data
)
1439 pg_data_t
*pgdat
= data
;
1440 int nid
= pgdat
->node_id
;
1441 struct mminit_pfnnid_cache nid_init_state
= { };
1442 unsigned long start
= jiffies
;
1443 unsigned long nr_pages
= 0;
1444 unsigned long walk_start
, walk_end
;
1447 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1448 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1450 if (first_init_pfn
== ULONG_MAX
) {
1451 pgdat_init_report_one_done();
1455 /* Bind memory initialisation thread to a local node if possible */
1456 if (!cpumask_empty(cpumask
))
1457 set_cpus_allowed_ptr(current
, cpumask
);
1459 /* Sanity check boundaries */
1460 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1461 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1462 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1464 /* Only the highest zone is deferred so find it */
1465 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1466 zone
= pgdat
->node_zones
+ zid
;
1467 if (first_init_pfn
< zone_end_pfn(zone
))
1471 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1472 unsigned long pfn
, end_pfn
;
1473 struct page
*page
= NULL
;
1474 struct page
*free_base_page
= NULL
;
1475 unsigned long free_base_pfn
= 0;
1478 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1479 pfn
= first_init_pfn
;
1480 if (pfn
< walk_start
)
1482 if (pfn
< zone
->zone_start_pfn
)
1483 pfn
= zone
->zone_start_pfn
;
1485 for (; pfn
< end_pfn
; pfn
++) {
1486 if (!pfn_valid_within(pfn
))
1490 * Ensure pfn_valid is checked every
1491 * pageblock_nr_pages for memory holes
1493 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1494 if (!pfn_valid(pfn
)) {
1500 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1505 /* Minimise pfn page lookups and scheduler checks */
1506 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1509 nr_pages
+= nr_to_free
;
1510 deferred_free_range(free_base_page
,
1511 free_base_pfn
, nr_to_free
);
1512 free_base_page
= NULL
;
1513 free_base_pfn
= nr_to_free
= 0;
1515 page
= pfn_to_page(pfn
);
1520 VM_BUG_ON(page_zone(page
) != zone
);
1524 __init_single_page(page
, pfn
, zid
, nid
);
1525 if (!free_base_page
) {
1526 free_base_page
= page
;
1527 free_base_pfn
= pfn
;
1532 /* Where possible, batch up pages for a single free */
1535 /* Free the current block of pages to allocator */
1536 nr_pages
+= nr_to_free
;
1537 deferred_free_range(free_base_page
, free_base_pfn
,
1539 free_base_page
= NULL
;
1540 free_base_pfn
= nr_to_free
= 0;
1542 /* Free the last block of pages to allocator */
1543 nr_pages
+= nr_to_free
;
1544 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1546 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1549 /* Sanity check that the next zone really is unpopulated */
1550 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1552 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1553 jiffies_to_msecs(jiffies
- start
));
1555 pgdat_init_report_one_done();
1558 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1560 void __init
page_alloc_init_late(void)
1564 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1567 /* There will be num_node_state(N_MEMORY) threads */
1568 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1569 for_each_node_state(nid
, N_MEMORY
) {
1570 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1573 /* Block until all are initialised */
1574 wait_for_completion(&pgdat_init_all_done_comp
);
1576 /* Reinit limits that are based on free pages after the kernel is up */
1577 files_maxfiles_init();
1580 for_each_populated_zone(zone
)
1581 set_zone_contiguous(zone
);
1585 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1586 void __init
init_cma_reserved_pageblock(struct page
*page
)
1588 unsigned i
= pageblock_nr_pages
;
1589 struct page
*p
= page
;
1592 __ClearPageReserved(p
);
1593 set_page_count(p
, 0);
1596 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1598 if (pageblock_order
>= MAX_ORDER
) {
1599 i
= pageblock_nr_pages
;
1602 set_page_refcounted(p
);
1603 __free_pages(p
, MAX_ORDER
- 1);
1604 p
+= MAX_ORDER_NR_PAGES
;
1605 } while (i
-= MAX_ORDER_NR_PAGES
);
1607 set_page_refcounted(page
);
1608 __free_pages(page
, pageblock_order
);
1611 adjust_managed_page_count(page
, pageblock_nr_pages
);
1616 * The order of subdivision here is critical for the IO subsystem.
1617 * Please do not alter this order without good reasons and regression
1618 * testing. Specifically, as large blocks of memory are subdivided,
1619 * the order in which smaller blocks are delivered depends on the order
1620 * they're subdivided in this function. This is the primary factor
1621 * influencing the order in which pages are delivered to the IO
1622 * subsystem according to empirical testing, and this is also justified
1623 * by considering the behavior of a buddy system containing a single
1624 * large block of memory acted on by a series of small allocations.
1625 * This behavior is a critical factor in sglist merging's success.
1629 static inline void expand(struct zone
*zone
, struct page
*page
,
1630 int low
, int high
, struct free_area
*area
,
1633 unsigned long size
= 1 << high
;
1635 while (high
> low
) {
1639 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1642 * Mark as guard pages (or page), that will allow to
1643 * merge back to allocator when buddy will be freed.
1644 * Corresponding page table entries will not be touched,
1645 * pages will stay not present in virtual address space
1647 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1650 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1652 set_page_order(&page
[size
], high
);
1656 static void check_new_page_bad(struct page
*page
)
1658 const char *bad_reason
= NULL
;
1659 unsigned long bad_flags
= 0;
1661 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1662 bad_reason
= "nonzero mapcount";
1663 if (unlikely(page
->mapping
!= NULL
))
1664 bad_reason
= "non-NULL mapping";
1665 if (unlikely(page_ref_count(page
) != 0))
1666 bad_reason
= "nonzero _count";
1667 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1668 bad_reason
= "HWPoisoned (hardware-corrupted)";
1669 bad_flags
= __PG_HWPOISON
;
1670 /* Don't complain about hwpoisoned pages */
1671 page_mapcount_reset(page
); /* remove PageBuddy */
1674 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1675 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1676 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1679 if (unlikely(page
->mem_cgroup
))
1680 bad_reason
= "page still charged to cgroup";
1682 bad_page(page
, bad_reason
, bad_flags
);
1686 * This page is about to be returned from the page allocator
1688 static inline int check_new_page(struct page
*page
)
1690 if (likely(page_expected_state(page
,
1691 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1694 check_new_page_bad(page
);
1698 static inline bool free_pages_prezeroed(bool poisoned
)
1700 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1701 page_poisoning_enabled() && poisoned
;
1704 #ifdef CONFIG_DEBUG_VM
1705 static bool check_pcp_refill(struct page
*page
)
1710 static bool check_new_pcp(struct page
*page
)
1712 return check_new_page(page
);
1715 static bool check_pcp_refill(struct page
*page
)
1717 return check_new_page(page
);
1719 static bool check_new_pcp(struct page
*page
)
1723 #endif /* CONFIG_DEBUG_VM */
1725 static bool check_new_pages(struct page
*page
, unsigned int order
)
1728 for (i
= 0; i
< (1 << order
); i
++) {
1729 struct page
*p
= page
+ i
;
1731 if (unlikely(check_new_page(p
)))
1738 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1741 set_page_private(page
, 0);
1742 set_page_refcounted(page
);
1744 arch_alloc_page(page
, order
);
1745 kernel_map_pages(page
, 1 << order
, 1);
1746 kernel_poison_pages(page
, 1 << order
, 1);
1747 kasan_alloc_pages(page
, order
);
1748 set_page_owner(page
, order
, gfp_flags
);
1751 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1752 unsigned int alloc_flags
)
1755 bool poisoned
= true;
1757 for (i
= 0; i
< (1 << order
); i
++) {
1758 struct page
*p
= page
+ i
;
1760 poisoned
&= page_is_poisoned(p
);
1763 post_alloc_hook(page
, order
, gfp_flags
);
1765 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1766 for (i
= 0; i
< (1 << order
); i
++)
1767 clear_highpage(page
+ i
);
1769 if (order
&& (gfp_flags
& __GFP_COMP
))
1770 prep_compound_page(page
, order
);
1773 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1774 * allocate the page. The expectation is that the caller is taking
1775 * steps that will free more memory. The caller should avoid the page
1776 * being used for !PFMEMALLOC purposes.
1778 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1779 set_page_pfmemalloc(page
);
1781 clear_page_pfmemalloc(page
);
1785 * Go through the free lists for the given migratetype and remove
1786 * the smallest available page from the freelists
1789 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1792 unsigned int current_order
;
1793 struct free_area
*area
;
1796 /* Find a page of the appropriate size in the preferred list */
1797 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1798 area
= &(zone
->free_area
[current_order
]);
1799 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1803 list_del(&page
->lru
);
1804 rmv_page_order(page
);
1806 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1807 set_pcppage_migratetype(page
, migratetype
);
1816 * This array describes the order lists are fallen back to when
1817 * the free lists for the desirable migrate type are depleted
1819 static int fallbacks
[MIGRATE_TYPES
][4] = {
1820 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1821 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1822 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1824 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1826 #ifdef CONFIG_MEMORY_ISOLATION
1827 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1832 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1835 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1838 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1839 unsigned int order
) { return NULL
; }
1843 * Move the free pages in a range to the free lists of the requested type.
1844 * Note that start_page and end_pages are not aligned on a pageblock
1845 * boundary. If alignment is required, use move_freepages_block()
1847 int move_freepages(struct zone
*zone
,
1848 struct page
*start_page
, struct page
*end_page
,
1853 int pages_moved
= 0;
1855 #ifndef CONFIG_HOLES_IN_ZONE
1857 * page_zone is not safe to call in this context when
1858 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1859 * anyway as we check zone boundaries in move_freepages_block().
1860 * Remove at a later date when no bug reports exist related to
1861 * grouping pages by mobility
1863 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1866 for (page
= start_page
; page
<= end_page
;) {
1867 /* Make sure we are not inadvertently changing nodes */
1868 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1870 if (!pfn_valid_within(page_to_pfn(page
))) {
1875 if (!PageBuddy(page
)) {
1880 order
= page_order(page
);
1881 list_move(&page
->lru
,
1882 &zone
->free_area
[order
].free_list
[migratetype
]);
1884 pages_moved
+= 1 << order
;
1890 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1893 unsigned long start_pfn
, end_pfn
;
1894 struct page
*start_page
, *end_page
;
1896 start_pfn
= page_to_pfn(page
);
1897 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1898 start_page
= pfn_to_page(start_pfn
);
1899 end_page
= start_page
+ pageblock_nr_pages
- 1;
1900 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1902 /* Do not cross zone boundaries */
1903 if (!zone_spans_pfn(zone
, start_pfn
))
1905 if (!zone_spans_pfn(zone
, end_pfn
))
1908 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1911 static void change_pageblock_range(struct page
*pageblock_page
,
1912 int start_order
, int migratetype
)
1914 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1916 while (nr_pageblocks
--) {
1917 set_pageblock_migratetype(pageblock_page
, migratetype
);
1918 pageblock_page
+= pageblock_nr_pages
;
1923 * When we are falling back to another migratetype during allocation, try to
1924 * steal extra free pages from the same pageblocks to satisfy further
1925 * allocations, instead of polluting multiple pageblocks.
1927 * If we are stealing a relatively large buddy page, it is likely there will
1928 * be more free pages in the pageblock, so try to steal them all. For
1929 * reclaimable and unmovable allocations, we steal regardless of page size,
1930 * as fragmentation caused by those allocations polluting movable pageblocks
1931 * is worse than movable allocations stealing from unmovable and reclaimable
1934 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1937 * Leaving this order check is intended, although there is
1938 * relaxed order check in next check. The reason is that
1939 * we can actually steal whole pageblock if this condition met,
1940 * but, below check doesn't guarantee it and that is just heuristic
1941 * so could be changed anytime.
1943 if (order
>= pageblock_order
)
1946 if (order
>= pageblock_order
/ 2 ||
1947 start_mt
== MIGRATE_RECLAIMABLE
||
1948 start_mt
== MIGRATE_UNMOVABLE
||
1949 page_group_by_mobility_disabled
)
1956 * This function implements actual steal behaviour. If order is large enough,
1957 * we can steal whole pageblock. If not, we first move freepages in this
1958 * pageblock and check whether half of pages are moved or not. If half of
1959 * pages are moved, we can change migratetype of pageblock and permanently
1960 * use it's pages as requested migratetype in the future.
1962 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1965 unsigned int current_order
= page_order(page
);
1968 /* Take ownership for orders >= pageblock_order */
1969 if (current_order
>= pageblock_order
) {
1970 change_pageblock_range(page
, current_order
, start_type
);
1974 pages
= move_freepages_block(zone
, page
, start_type
);
1976 /* Claim the whole block if over half of it is free */
1977 if (pages
>= (1 << (pageblock_order
-1)) ||
1978 page_group_by_mobility_disabled
)
1979 set_pageblock_migratetype(page
, start_type
);
1983 * Check whether there is a suitable fallback freepage with requested order.
1984 * If only_stealable is true, this function returns fallback_mt only if
1985 * we can steal other freepages all together. This would help to reduce
1986 * fragmentation due to mixed migratetype pages in one pageblock.
1988 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1989 int migratetype
, bool only_stealable
, bool *can_steal
)
1994 if (area
->nr_free
== 0)
1999 fallback_mt
= fallbacks
[migratetype
][i
];
2000 if (fallback_mt
== MIGRATE_TYPES
)
2003 if (list_empty(&area
->free_list
[fallback_mt
]))
2006 if (can_steal_fallback(order
, migratetype
))
2009 if (!only_stealable
)
2020 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2021 * there are no empty page blocks that contain a page with a suitable order
2023 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2024 unsigned int alloc_order
)
2027 unsigned long max_managed
, flags
;
2030 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2031 * Check is race-prone but harmless.
2033 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2034 if (zone
->nr_reserved_highatomic
>= max_managed
)
2037 spin_lock_irqsave(&zone
->lock
, flags
);
2039 /* Recheck the nr_reserved_highatomic limit under the lock */
2040 if (zone
->nr_reserved_highatomic
>= max_managed
)
2044 mt
= get_pageblock_migratetype(page
);
2045 if (mt
!= MIGRATE_HIGHATOMIC
&&
2046 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
2047 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2048 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2049 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
2053 spin_unlock_irqrestore(&zone
->lock
, flags
);
2057 * Used when an allocation is about to fail under memory pressure. This
2058 * potentially hurts the reliability of high-order allocations when under
2059 * intense memory pressure but failed atomic allocations should be easier
2060 * to recover from than an OOM.
2062 * If @force is true, try to unreserve a pageblock even though highatomic
2063 * pageblock is exhausted.
2065 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2068 struct zonelist
*zonelist
= ac
->zonelist
;
2069 unsigned long flags
;
2076 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2079 * Preserve at least one pageblock unless memory pressure
2082 if (!force
&& zone
->nr_reserved_highatomic
<=
2086 spin_lock_irqsave(&zone
->lock
, flags
);
2087 for (order
= 0; order
< MAX_ORDER
; order
++) {
2088 struct free_area
*area
= &(zone
->free_area
[order
]);
2090 page
= list_first_entry_or_null(
2091 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2097 * In page freeing path, migratetype change is racy so
2098 * we can counter several free pages in a pageblock
2099 * in this loop althoug we changed the pageblock type
2100 * from highatomic to ac->migratetype. So we should
2101 * adjust the count once.
2103 if (get_pageblock_migratetype(page
) ==
2104 MIGRATE_HIGHATOMIC
) {
2106 * It should never happen but changes to
2107 * locking could inadvertently allow a per-cpu
2108 * drain to add pages to MIGRATE_HIGHATOMIC
2109 * while unreserving so be safe and watch for
2112 zone
->nr_reserved_highatomic
-= min(
2114 zone
->nr_reserved_highatomic
);
2118 * Convert to ac->migratetype and avoid the normal
2119 * pageblock stealing heuristics. Minimally, the caller
2120 * is doing the work and needs the pages. More
2121 * importantly, if the block was always converted to
2122 * MIGRATE_UNMOVABLE or another type then the number
2123 * of pageblocks that cannot be completely freed
2126 set_pageblock_migratetype(page
, ac
->migratetype
);
2127 ret
= move_freepages_block(zone
, page
, ac
->migratetype
);
2129 spin_unlock_irqrestore(&zone
->lock
, flags
);
2133 spin_unlock_irqrestore(&zone
->lock
, flags
);
2139 /* Remove an element from the buddy allocator from the fallback list */
2140 static inline struct page
*
2141 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
2143 struct free_area
*area
;
2144 unsigned int current_order
;
2149 /* Find the largest possible block of pages in the other list */
2150 for (current_order
= MAX_ORDER
-1;
2151 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
2153 area
= &(zone
->free_area
[current_order
]);
2154 fallback_mt
= find_suitable_fallback(area
, current_order
,
2155 start_migratetype
, false, &can_steal
);
2156 if (fallback_mt
== -1)
2159 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2162 get_pageblock_migratetype(page
) != MIGRATE_HIGHATOMIC
)
2163 steal_suitable_fallback(zone
, page
, start_migratetype
);
2165 /* Remove the page from the freelists */
2167 list_del(&page
->lru
);
2168 rmv_page_order(page
);
2170 expand(zone
, page
, order
, current_order
, area
,
2173 * The pcppage_migratetype may differ from pageblock's
2174 * migratetype depending on the decisions in
2175 * find_suitable_fallback(). This is OK as long as it does not
2176 * differ for MIGRATE_CMA pageblocks. Those can be used as
2177 * fallback only via special __rmqueue_cma_fallback() function
2179 set_pcppage_migratetype(page
, start_migratetype
);
2181 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2182 start_migratetype
, fallback_mt
);
2191 * Do the hard work of removing an element from the buddy allocator.
2192 * Call me with the zone->lock already held.
2194 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2199 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2200 if (unlikely(!page
)) {
2201 if (migratetype
== MIGRATE_MOVABLE
)
2202 page
= __rmqueue_cma_fallback(zone
, order
);
2205 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2208 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2213 * Obtain a specified number of elements from the buddy allocator, all under
2214 * a single hold of the lock, for efficiency. Add them to the supplied list.
2215 * Returns the number of new pages which were placed at *list.
2217 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2218 unsigned long count
, struct list_head
*list
,
2219 int migratetype
, bool cold
)
2223 spin_lock(&zone
->lock
);
2224 for (i
= 0; i
< count
; ++i
) {
2225 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2226 if (unlikely(page
== NULL
))
2229 if (unlikely(check_pcp_refill(page
)))
2233 * Split buddy pages returned by expand() are received here
2234 * in physical page order. The page is added to the callers and
2235 * list and the list head then moves forward. From the callers
2236 * perspective, the linked list is ordered by page number in
2237 * some conditions. This is useful for IO devices that can
2238 * merge IO requests if the physical pages are ordered
2242 list_add(&page
->lru
, list
);
2244 list_add_tail(&page
->lru
, list
);
2246 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2247 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2250 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2251 spin_unlock(&zone
->lock
);
2257 * Called from the vmstat counter updater to drain pagesets of this
2258 * currently executing processor on remote nodes after they have
2261 * Note that this function must be called with the thread pinned to
2262 * a single processor.
2264 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2266 unsigned long flags
;
2267 int to_drain
, batch
;
2269 local_irq_save(flags
);
2270 batch
= READ_ONCE(pcp
->batch
);
2271 to_drain
= min(pcp
->count
, batch
);
2273 free_pcppages_bulk(zone
, to_drain
, pcp
);
2274 pcp
->count
-= to_drain
;
2276 local_irq_restore(flags
);
2281 * Drain pcplists of the indicated processor and zone.
2283 * The processor must either be the current processor and the
2284 * thread pinned to the current processor or a processor that
2287 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2289 unsigned long flags
;
2290 struct per_cpu_pageset
*pset
;
2291 struct per_cpu_pages
*pcp
;
2293 local_irq_save(flags
);
2294 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2298 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2301 local_irq_restore(flags
);
2305 * Drain pcplists of all zones on the indicated processor.
2307 * The processor must either be the current processor and the
2308 * thread pinned to the current processor or a processor that
2311 static void drain_pages(unsigned int cpu
)
2315 for_each_populated_zone(zone
) {
2316 drain_pages_zone(cpu
, zone
);
2321 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2323 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2324 * the single zone's pages.
2326 void drain_local_pages(struct zone
*zone
)
2328 int cpu
= smp_processor_id();
2331 drain_pages_zone(cpu
, zone
);
2337 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2339 * When zone parameter is non-NULL, spill just the single zone's pages.
2341 * Note that this code is protected against sending an IPI to an offline
2342 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2343 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2344 * nothing keeps CPUs from showing up after we populated the cpumask and
2345 * before the call to on_each_cpu_mask().
2347 void drain_all_pages(struct zone
*zone
)
2352 * Allocate in the BSS so we wont require allocation in
2353 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2355 static cpumask_t cpus_with_pcps
;
2358 * We don't care about racing with CPU hotplug event
2359 * as offline notification will cause the notified
2360 * cpu to drain that CPU pcps and on_each_cpu_mask
2361 * disables preemption as part of its processing
2363 for_each_online_cpu(cpu
) {
2364 struct per_cpu_pageset
*pcp
;
2366 bool has_pcps
= false;
2369 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2373 for_each_populated_zone(z
) {
2374 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2375 if (pcp
->pcp
.count
) {
2383 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2385 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2387 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
2391 #ifdef CONFIG_HIBERNATION
2393 void mark_free_pages(struct zone
*zone
)
2395 unsigned long pfn
, max_zone_pfn
;
2396 unsigned long flags
;
2397 unsigned int order
, t
;
2400 if (zone_is_empty(zone
))
2403 spin_lock_irqsave(&zone
->lock
, flags
);
2405 max_zone_pfn
= zone_end_pfn(zone
);
2406 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2407 if (pfn_valid(pfn
)) {
2408 page
= pfn_to_page(pfn
);
2410 if (page_zone(page
) != zone
)
2413 if (!swsusp_page_is_forbidden(page
))
2414 swsusp_unset_page_free(page
);
2417 for_each_migratetype_order(order
, t
) {
2418 list_for_each_entry(page
,
2419 &zone
->free_area
[order
].free_list
[t
], lru
) {
2422 pfn
= page_to_pfn(page
);
2423 for (i
= 0; i
< (1UL << order
); i
++)
2424 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2427 spin_unlock_irqrestore(&zone
->lock
, flags
);
2429 #endif /* CONFIG_PM */
2432 * Free a 0-order page
2433 * cold == true ? free a cold page : free a hot page
2435 void free_hot_cold_page(struct page
*page
, bool cold
)
2437 struct zone
*zone
= page_zone(page
);
2438 struct per_cpu_pages
*pcp
;
2439 unsigned long flags
;
2440 unsigned long pfn
= page_to_pfn(page
);
2443 if (!free_pcp_prepare(page
))
2446 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2447 set_pcppage_migratetype(page
, migratetype
);
2448 local_irq_save(flags
);
2449 __count_vm_event(PGFREE
);
2452 * We only track unmovable, reclaimable and movable on pcp lists.
2453 * Free ISOLATE pages back to the allocator because they are being
2454 * offlined but treat RESERVE as movable pages so we can get those
2455 * areas back if necessary. Otherwise, we may have to free
2456 * excessively into the page allocator
2458 if (migratetype
>= MIGRATE_PCPTYPES
) {
2459 if (unlikely(is_migrate_isolate(migratetype
))) {
2460 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2463 migratetype
= MIGRATE_MOVABLE
;
2466 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2468 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2470 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2472 if (pcp
->count
>= pcp
->high
) {
2473 unsigned long batch
= READ_ONCE(pcp
->batch
);
2474 free_pcppages_bulk(zone
, batch
, pcp
);
2475 pcp
->count
-= batch
;
2479 local_irq_restore(flags
);
2483 * Free a list of 0-order pages
2485 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2487 struct page
*page
, *next
;
2489 list_for_each_entry_safe(page
, next
, list
, lru
) {
2490 trace_mm_page_free_batched(page
, cold
);
2491 free_hot_cold_page(page
, cold
);
2496 * split_page takes a non-compound higher-order page, and splits it into
2497 * n (1<<order) sub-pages: page[0..n]
2498 * Each sub-page must be freed individually.
2500 * Note: this is probably too low level an operation for use in drivers.
2501 * Please consult with lkml before using this in your driver.
2503 void split_page(struct page
*page
, unsigned int order
)
2507 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2508 VM_BUG_ON_PAGE(!page_count(page
), page
);
2510 #ifdef CONFIG_KMEMCHECK
2512 * Split shadow pages too, because free(page[0]) would
2513 * otherwise free the whole shadow.
2515 if (kmemcheck_page_is_tracked(page
))
2516 split_page(virt_to_page(page
[0].shadow
), order
);
2519 for (i
= 1; i
< (1 << order
); i
++)
2520 set_page_refcounted(page
+ i
);
2521 split_page_owner(page
, order
);
2523 EXPORT_SYMBOL_GPL(split_page
);
2525 int __isolate_free_page(struct page
*page
, unsigned int order
)
2527 unsigned long watermark
;
2531 BUG_ON(!PageBuddy(page
));
2533 zone
= page_zone(page
);
2534 mt
= get_pageblock_migratetype(page
);
2536 if (!is_migrate_isolate(mt
)) {
2538 * Obey watermarks as if the page was being allocated. We can
2539 * emulate a high-order watermark check with a raised order-0
2540 * watermark, because we already know our high-order page
2543 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2544 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2547 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2550 /* Remove page from free list */
2551 list_del(&page
->lru
);
2552 zone
->free_area
[order
].nr_free
--;
2553 rmv_page_order(page
);
2556 * Set the pageblock if the isolated page is at least half of a
2559 if (order
>= pageblock_order
- 1) {
2560 struct page
*endpage
= page
+ (1 << order
) - 1;
2561 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2562 int mt
= get_pageblock_migratetype(page
);
2563 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2564 && mt
!= MIGRATE_HIGHATOMIC
)
2565 set_pageblock_migratetype(page
,
2571 return 1UL << order
;
2575 * Update NUMA hit/miss statistics
2577 * Must be called with interrupts disabled.
2579 * When __GFP_OTHER_NODE is set assume the node of the preferred
2580 * zone is the local node. This is useful for daemons who allocate
2581 * memory on behalf of other processes.
2583 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2587 int local_nid
= numa_node_id();
2588 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2590 if (unlikely(flags
& __GFP_OTHER_NODE
)) {
2591 local_stat
= NUMA_OTHER
;
2592 local_nid
= preferred_zone
->node
;
2595 if (z
->node
== local_nid
) {
2596 __inc_zone_state(z
, NUMA_HIT
);
2597 __inc_zone_state(z
, local_stat
);
2599 __inc_zone_state(z
, NUMA_MISS
);
2600 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2606 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2609 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2610 struct zone
*zone
, unsigned int order
,
2611 gfp_t gfp_flags
, unsigned int alloc_flags
,
2614 unsigned long flags
;
2616 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2618 if (likely(order
== 0)) {
2619 struct per_cpu_pages
*pcp
;
2620 struct list_head
*list
;
2622 local_irq_save(flags
);
2624 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2625 list
= &pcp
->lists
[migratetype
];
2626 if (list_empty(list
)) {
2627 pcp
->count
+= rmqueue_bulk(zone
, 0,
2630 if (unlikely(list_empty(list
)))
2635 page
= list_last_entry(list
, struct page
, lru
);
2637 page
= list_first_entry(list
, struct page
, lru
);
2639 list_del(&page
->lru
);
2642 } while (check_new_pcp(page
));
2645 * We most definitely don't want callers attempting to
2646 * allocate greater than order-1 page units with __GFP_NOFAIL.
2648 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2649 spin_lock_irqsave(&zone
->lock
, flags
);
2653 if (alloc_flags
& ALLOC_HARDER
) {
2654 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2656 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2659 page
= __rmqueue(zone
, order
, migratetype
);
2660 } while (page
&& check_new_pages(page
, order
));
2661 spin_unlock(&zone
->lock
);
2664 __mod_zone_freepage_state(zone
, -(1 << order
),
2665 get_pcppage_migratetype(page
));
2668 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2669 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2670 local_irq_restore(flags
);
2672 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2676 local_irq_restore(flags
);
2680 #ifdef CONFIG_FAIL_PAGE_ALLOC
2683 struct fault_attr attr
;
2685 bool ignore_gfp_highmem
;
2686 bool ignore_gfp_reclaim
;
2688 } fail_page_alloc
= {
2689 .attr
= FAULT_ATTR_INITIALIZER
,
2690 .ignore_gfp_reclaim
= true,
2691 .ignore_gfp_highmem
= true,
2695 static int __init
setup_fail_page_alloc(char *str
)
2697 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2699 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2701 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2703 if (order
< fail_page_alloc
.min_order
)
2705 if (gfp_mask
& __GFP_NOFAIL
)
2707 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2709 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2710 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2713 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2716 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2718 static int __init
fail_page_alloc_debugfs(void)
2720 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2723 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2724 &fail_page_alloc
.attr
);
2726 return PTR_ERR(dir
);
2728 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2729 &fail_page_alloc
.ignore_gfp_reclaim
))
2731 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2732 &fail_page_alloc
.ignore_gfp_highmem
))
2734 if (!debugfs_create_u32("min-order", mode
, dir
,
2735 &fail_page_alloc
.min_order
))
2740 debugfs_remove_recursive(dir
);
2745 late_initcall(fail_page_alloc_debugfs
);
2747 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2749 #else /* CONFIG_FAIL_PAGE_ALLOC */
2751 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2756 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2759 * Return true if free base pages are above 'mark'. For high-order checks it
2760 * will return true of the order-0 watermark is reached and there is at least
2761 * one free page of a suitable size. Checking now avoids taking the zone lock
2762 * to check in the allocation paths if no pages are free.
2764 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2765 int classzone_idx
, unsigned int alloc_flags
,
2770 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2772 /* free_pages may go negative - that's OK */
2773 free_pages
-= (1 << order
) - 1;
2775 if (alloc_flags
& ALLOC_HIGH
)
2779 * If the caller does not have rights to ALLOC_HARDER then subtract
2780 * the high-atomic reserves. This will over-estimate the size of the
2781 * atomic reserve but it avoids a search.
2783 if (likely(!alloc_harder
))
2784 free_pages
-= z
->nr_reserved_highatomic
;
2789 /* If allocation can't use CMA areas don't use free CMA pages */
2790 if (!(alloc_flags
& ALLOC_CMA
))
2791 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2795 * Check watermarks for an order-0 allocation request. If these
2796 * are not met, then a high-order request also cannot go ahead
2797 * even if a suitable page happened to be free.
2799 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2802 /* If this is an order-0 request then the watermark is fine */
2806 /* For a high-order request, check at least one suitable page is free */
2807 for (o
= order
; o
< MAX_ORDER
; o
++) {
2808 struct free_area
*area
= &z
->free_area
[o
];
2817 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2818 if (!list_empty(&area
->free_list
[mt
]))
2823 if ((alloc_flags
& ALLOC_CMA
) &&
2824 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2832 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2833 int classzone_idx
, unsigned int alloc_flags
)
2835 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2836 zone_page_state(z
, NR_FREE_PAGES
));
2839 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2840 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2842 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2846 /* If allocation can't use CMA areas don't use free CMA pages */
2847 if (!(alloc_flags
& ALLOC_CMA
))
2848 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2852 * Fast check for order-0 only. If this fails then the reserves
2853 * need to be calculated. There is a corner case where the check
2854 * passes but only the high-order atomic reserve are free. If
2855 * the caller is !atomic then it'll uselessly search the free
2856 * list. That corner case is then slower but it is harmless.
2858 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2861 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2865 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2866 unsigned long mark
, int classzone_idx
)
2868 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2870 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2871 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2873 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2878 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2880 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2883 #else /* CONFIG_NUMA */
2884 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2888 #endif /* CONFIG_NUMA */
2891 * get_page_from_freelist goes through the zonelist trying to allocate
2894 static struct page
*
2895 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2896 const struct alloc_context
*ac
)
2898 struct zoneref
*z
= ac
->preferred_zoneref
;
2900 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
2903 * Scan zonelist, looking for a zone with enough free.
2904 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2906 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2911 if (cpusets_enabled() &&
2912 (alloc_flags
& ALLOC_CPUSET
) &&
2913 !__cpuset_zone_allowed(zone
, gfp_mask
))
2916 * When allocating a page cache page for writing, we
2917 * want to get it from a node that is within its dirty
2918 * limit, such that no single node holds more than its
2919 * proportional share of globally allowed dirty pages.
2920 * The dirty limits take into account the node's
2921 * lowmem reserves and high watermark so that kswapd
2922 * should be able to balance it without having to
2923 * write pages from its LRU list.
2925 * XXX: For now, allow allocations to potentially
2926 * exceed the per-node dirty limit in the slowpath
2927 * (spread_dirty_pages unset) before going into reclaim,
2928 * which is important when on a NUMA setup the allowed
2929 * nodes are together not big enough to reach the
2930 * global limit. The proper fix for these situations
2931 * will require awareness of nodes in the
2932 * dirty-throttling and the flusher threads.
2934 if (ac
->spread_dirty_pages
) {
2935 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
2938 if (!node_dirty_ok(zone
->zone_pgdat
)) {
2939 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
2944 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2945 if (!zone_watermark_fast(zone
, order
, mark
,
2946 ac_classzone_idx(ac
), alloc_flags
)) {
2949 /* Checked here to keep the fast path fast */
2950 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2951 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2954 if (node_reclaim_mode
== 0 ||
2955 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
2958 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
2960 case NODE_RECLAIM_NOSCAN
:
2963 case NODE_RECLAIM_FULL
:
2964 /* scanned but unreclaimable */
2967 /* did we reclaim enough */
2968 if (zone_watermark_ok(zone
, order
, mark
,
2969 ac_classzone_idx(ac
), alloc_flags
))
2977 page
= buffered_rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
2978 gfp_mask
, alloc_flags
, ac
->migratetype
);
2980 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
2983 * If this is a high-order atomic allocation then check
2984 * if the pageblock should be reserved for the future
2986 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
2987 reserve_highatomic_pageblock(page
, zone
, order
);
2997 * Large machines with many possible nodes should not always dump per-node
2998 * meminfo in irq context.
3000 static inline bool should_suppress_show_mem(void)
3005 ret
= in_interrupt();
3010 static DEFINE_RATELIMIT_STATE(nopage_rs
,
3011 DEFAULT_RATELIMIT_INTERVAL
,
3012 DEFAULT_RATELIMIT_BURST
);
3014 void warn_alloc(gfp_t gfp_mask
, const char *fmt
, ...)
3016 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3017 struct va_format vaf
;
3020 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
3021 debug_guardpage_minorder() > 0)
3025 * This documents exceptions given to allocations in certain
3026 * contexts that are allowed to allocate outside current's set
3029 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3030 if (test_thread_flag(TIF_MEMDIE
) ||
3031 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3032 filter
&= ~SHOW_MEM_FILTER_NODES
;
3033 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3034 filter
&= ~SHOW_MEM_FILTER_NODES
;
3036 pr_warn("%s: ", current
->comm
);
3038 va_start(args
, fmt
);
3041 pr_cont("%pV", &vaf
);
3044 pr_cont(", mode:%#x(%pGg)\n", gfp_mask
, &gfp_mask
);
3047 if (!should_suppress_show_mem())
3051 static inline struct page
*
3052 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3053 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3055 struct oom_control oc
= {
3056 .zonelist
= ac
->zonelist
,
3057 .nodemask
= ac
->nodemask
,
3059 .gfp_mask
= gfp_mask
,
3064 *did_some_progress
= 0;
3067 * Acquire the oom lock. If that fails, somebody else is
3068 * making progress for us.
3070 if (!mutex_trylock(&oom_lock
)) {
3071 *did_some_progress
= 1;
3072 schedule_timeout_uninterruptible(1);
3077 * Go through the zonelist yet one more time, keep very high watermark
3078 * here, this is only to catch a parallel oom killing, we must fail if
3079 * we're still under heavy pressure.
3081 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
3082 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3086 if (!(gfp_mask
& __GFP_NOFAIL
)) {
3087 /* Coredumps can quickly deplete all memory reserves */
3088 if (current
->flags
& PF_DUMPCORE
)
3090 /* The OOM killer will not help higher order allocs */
3091 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3093 /* The OOM killer does not needlessly kill tasks for lowmem */
3094 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3096 if (pm_suspended_storage())
3099 * XXX: GFP_NOFS allocations should rather fail than rely on
3100 * other request to make a forward progress.
3101 * We are in an unfortunate situation where out_of_memory cannot
3102 * do much for this context but let's try it to at least get
3103 * access to memory reserved if the current task is killed (see
3104 * out_of_memory). Once filesystems are ready to handle allocation
3105 * failures more gracefully we should just bail out here.
3108 /* The OOM killer may not free memory on a specific node */
3109 if (gfp_mask
& __GFP_THISNODE
)
3112 /* Exhausted what can be done so it's blamo time */
3113 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3114 *did_some_progress
= 1;
3116 if (gfp_mask
& __GFP_NOFAIL
) {
3117 page
= get_page_from_freelist(gfp_mask
, order
,
3118 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
3120 * fallback to ignore cpuset restriction if our nodes
3124 page
= get_page_from_freelist(gfp_mask
, order
,
3125 ALLOC_NO_WATERMARKS
, ac
);
3129 mutex_unlock(&oom_lock
);
3134 * Maximum number of compaction retries wit a progress before OOM
3135 * killer is consider as the only way to move forward.
3137 #define MAX_COMPACT_RETRIES 16
3139 #ifdef CONFIG_COMPACTION
3140 /* Try memory compaction for high-order allocations before reclaim */
3141 static struct page
*
3142 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3143 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3144 enum compact_priority prio
, enum compact_result
*compact_result
)
3151 current
->flags
|= PF_MEMALLOC
;
3152 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3154 current
->flags
&= ~PF_MEMALLOC
;
3156 if (*compact_result
<= COMPACT_INACTIVE
)
3160 * At least in one zone compaction wasn't deferred or skipped, so let's
3161 * count a compaction stall
3163 count_vm_event(COMPACTSTALL
);
3165 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3168 struct zone
*zone
= page_zone(page
);
3170 zone
->compact_blockskip_flush
= false;
3171 compaction_defer_reset(zone
, order
, true);
3172 count_vm_event(COMPACTSUCCESS
);
3177 * It's bad if compaction run occurs and fails. The most likely reason
3178 * is that pages exist, but not enough to satisfy watermarks.
3180 count_vm_event(COMPACTFAIL
);
3188 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3189 enum compact_result compact_result
,
3190 enum compact_priority
*compact_priority
,
3191 int *compaction_retries
)
3193 int max_retries
= MAX_COMPACT_RETRIES
;
3199 if (compaction_made_progress(compact_result
))
3200 (*compaction_retries
)++;
3203 * compaction considers all the zone as desperately out of memory
3204 * so it doesn't really make much sense to retry except when the
3205 * failure could be caused by insufficient priority
3207 if (compaction_failed(compact_result
))
3208 goto check_priority
;
3211 * make sure the compaction wasn't deferred or didn't bail out early
3212 * due to locks contention before we declare that we should give up.
3213 * But do not retry if the given zonelist is not suitable for
3216 if (compaction_withdrawn(compact_result
))
3217 return compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3220 * !costly requests are much more important than __GFP_REPEAT
3221 * costly ones because they are de facto nofail and invoke OOM
3222 * killer to move on while costly can fail and users are ready
3223 * to cope with that. 1/4 retries is rather arbitrary but we
3224 * would need much more detailed feedback from compaction to
3225 * make a better decision.
3227 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3229 if (*compaction_retries
<= max_retries
)
3233 * Make sure there are attempts at the highest priority if we exhausted
3234 * all retries or failed at the lower priorities.
3237 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3238 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3239 if (*compact_priority
> min_priority
) {
3240 (*compact_priority
)--;
3241 *compaction_retries
= 0;
3247 static inline struct page
*
3248 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3249 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3250 enum compact_priority prio
, enum compact_result
*compact_result
)
3252 *compact_result
= COMPACT_SKIPPED
;
3257 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3258 enum compact_result compact_result
,
3259 enum compact_priority
*compact_priority
,
3260 int *compaction_retries
)
3265 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3269 * There are setups with compaction disabled which would prefer to loop
3270 * inside the allocator rather than hit the oom killer prematurely.
3271 * Let's give them a good hope and keep retrying while the order-0
3272 * watermarks are OK.
3274 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3276 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3277 ac_classzone_idx(ac
), alloc_flags
))
3282 #endif /* CONFIG_COMPACTION */
3284 /* Perform direct synchronous page reclaim */
3286 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3287 const struct alloc_context
*ac
)
3289 struct reclaim_state reclaim_state
;
3294 /* We now go into synchronous reclaim */
3295 cpuset_memory_pressure_bump();
3296 current
->flags
|= PF_MEMALLOC
;
3297 lockdep_set_current_reclaim_state(gfp_mask
);
3298 reclaim_state
.reclaimed_slab
= 0;
3299 current
->reclaim_state
= &reclaim_state
;
3301 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3304 current
->reclaim_state
= NULL
;
3305 lockdep_clear_current_reclaim_state();
3306 current
->flags
&= ~PF_MEMALLOC
;
3313 /* The really slow allocator path where we enter direct reclaim */
3314 static inline struct page
*
3315 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3316 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3317 unsigned long *did_some_progress
)
3319 struct page
*page
= NULL
;
3320 bool drained
= false;
3322 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3323 if (unlikely(!(*did_some_progress
)))
3327 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3330 * If an allocation failed after direct reclaim, it could be because
3331 * pages are pinned on the per-cpu lists or in high alloc reserves.
3332 * Shrink them them and try again
3334 if (!page
&& !drained
) {
3335 unreserve_highatomic_pageblock(ac
, false);
3336 drain_all_pages(NULL
);
3344 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3348 pg_data_t
*last_pgdat
= NULL
;
3350 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3351 ac
->high_zoneidx
, ac
->nodemask
) {
3352 if (last_pgdat
!= zone
->zone_pgdat
)
3353 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3354 last_pgdat
= zone
->zone_pgdat
;
3358 static inline unsigned int
3359 gfp_to_alloc_flags(gfp_t gfp_mask
)
3361 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3363 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3364 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3367 * The caller may dip into page reserves a bit more if the caller
3368 * cannot run direct reclaim, or if the caller has realtime scheduling
3369 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3370 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3372 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3374 if (gfp_mask
& __GFP_ATOMIC
) {
3376 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3377 * if it can't schedule.
3379 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3380 alloc_flags
|= ALLOC_HARDER
;
3382 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3383 * comment for __cpuset_node_allowed().
3385 alloc_flags
&= ~ALLOC_CPUSET
;
3386 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3387 alloc_flags
|= ALLOC_HARDER
;
3390 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3391 alloc_flags
|= ALLOC_CMA
;
3396 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3398 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3401 if (gfp_mask
& __GFP_MEMALLOC
)
3403 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3405 if (!in_interrupt() &&
3406 ((current
->flags
& PF_MEMALLOC
) ||
3407 unlikely(test_thread_flag(TIF_MEMDIE
))))
3414 * Maximum number of reclaim retries without any progress before OOM killer
3415 * is consider as the only way to move forward.
3417 #define MAX_RECLAIM_RETRIES 16
3420 * Checks whether it makes sense to retry the reclaim to make a forward progress
3421 * for the given allocation request.
3422 * The reclaim feedback represented by did_some_progress (any progress during
3423 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3424 * any progress in a row) is considered as well as the reclaimable pages on the
3425 * applicable zone list (with a backoff mechanism which is a function of
3426 * no_progress_loops).
3428 * Returns true if a retry is viable or false to enter the oom path.
3431 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3432 struct alloc_context
*ac
, int alloc_flags
,
3433 bool did_some_progress
, int *no_progress_loops
)
3439 * Costly allocations might have made a progress but this doesn't mean
3440 * their order will become available due to high fragmentation so
3441 * always increment the no progress counter for them
3443 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3444 *no_progress_loops
= 0;
3446 (*no_progress_loops
)++;
3449 * Make sure we converge to OOM if we cannot make any progress
3450 * several times in the row.
3452 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3453 /* Before OOM, exhaust highatomic_reserve */
3454 return unreserve_highatomic_pageblock(ac
, true);
3458 * Keep reclaiming pages while there is a chance this will lead
3459 * somewhere. If none of the target zones can satisfy our allocation
3460 * request even if all reclaimable pages are considered then we are
3461 * screwed and have to go OOM.
3463 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3465 unsigned long available
;
3466 unsigned long reclaimable
;
3468 available
= reclaimable
= zone_reclaimable_pages(zone
);
3469 available
-= DIV_ROUND_UP((*no_progress_loops
) * available
,
3470 MAX_RECLAIM_RETRIES
);
3471 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3474 * Would the allocation succeed if we reclaimed the whole
3477 if (__zone_watermark_ok(zone
, order
, min_wmark_pages(zone
),
3478 ac_classzone_idx(ac
), alloc_flags
, available
)) {
3480 * If we didn't make any progress and have a lot of
3481 * dirty + writeback pages then we should wait for
3482 * an IO to complete to slow down the reclaim and
3483 * prevent from pre mature OOM
3485 if (!did_some_progress
) {
3486 unsigned long write_pending
;
3488 write_pending
= zone_page_state_snapshot(zone
,
3489 NR_ZONE_WRITE_PENDING
);
3491 if (2 * write_pending
> reclaimable
) {
3492 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3498 * Memory allocation/reclaim might be called from a WQ
3499 * context and the current implementation of the WQ
3500 * concurrency control doesn't recognize that
3501 * a particular WQ is congested if the worker thread is
3502 * looping without ever sleeping. Therefore we have to
3503 * do a short sleep here rather than calling
3506 if (current
->flags
& PF_WQ_WORKER
)
3507 schedule_timeout_uninterruptible(1);
3518 static inline struct page
*
3519 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3520 struct alloc_context
*ac
)
3522 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3523 struct page
*page
= NULL
;
3524 unsigned int alloc_flags
;
3525 unsigned long did_some_progress
;
3526 enum compact_priority compact_priority
= DEF_COMPACT_PRIORITY
;
3527 enum compact_result compact_result
;
3528 int compaction_retries
= 0;
3529 int no_progress_loops
= 0;
3530 unsigned long alloc_start
= jiffies
;
3531 unsigned int stall_timeout
= 10 * HZ
;
3534 * In the slowpath, we sanity check order to avoid ever trying to
3535 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3536 * be using allocators in order of preference for an area that is
3539 if (order
>= MAX_ORDER
) {
3540 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3545 * We also sanity check to catch abuse of atomic reserves being used by
3546 * callers that are not in atomic context.
3548 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3549 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3550 gfp_mask
&= ~__GFP_ATOMIC
;
3553 * The fast path uses conservative alloc_flags to succeed only until
3554 * kswapd needs to be woken up, and to avoid the cost of setting up
3555 * alloc_flags precisely. So we do that now.
3557 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3559 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3560 wake_all_kswapds(order
, ac
);
3563 * The adjusted alloc_flags might result in immediate success, so try
3566 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3571 * For costly allocations, try direct compaction first, as it's likely
3572 * that we have enough base pages and don't need to reclaim. Don't try
3573 * that for allocations that are allowed to ignore watermarks, as the
3574 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3576 if (can_direct_reclaim
&& order
> PAGE_ALLOC_COSTLY_ORDER
&&
3577 !gfp_pfmemalloc_allowed(gfp_mask
)) {
3578 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3580 INIT_COMPACT_PRIORITY
,
3586 * Checks for costly allocations with __GFP_NORETRY, which
3587 * includes THP page fault allocations
3589 if (gfp_mask
& __GFP_NORETRY
) {
3591 * If compaction is deferred for high-order allocations,
3592 * it is because sync compaction recently failed. If
3593 * this is the case and the caller requested a THP
3594 * allocation, we do not want to heavily disrupt the
3595 * system, so we fail the allocation instead of entering
3598 if (compact_result
== COMPACT_DEFERRED
)
3602 * Looks like reclaim/compaction is worth trying, but
3603 * sync compaction could be very expensive, so keep
3604 * using async compaction.
3606 compact_priority
= INIT_COMPACT_PRIORITY
;
3611 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3612 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3613 wake_all_kswapds(order
, ac
);
3615 if (gfp_pfmemalloc_allowed(gfp_mask
))
3616 alloc_flags
= ALLOC_NO_WATERMARKS
;
3619 * Reset the zonelist iterators if memory policies can be ignored.
3620 * These allocations are high priority and system rather than user
3623 if (!(alloc_flags
& ALLOC_CPUSET
) || (alloc_flags
& ALLOC_NO_WATERMARKS
)) {
3624 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3625 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3626 ac
->high_zoneidx
, ac
->nodemask
);
3629 /* Attempt with potentially adjusted zonelist and alloc_flags */
3630 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3634 /* Caller is not willing to reclaim, we can't balance anything */
3635 if (!can_direct_reclaim
) {
3637 * All existing users of the __GFP_NOFAIL are blockable, so warn
3638 * of any new users that actually allow this type of allocation
3641 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3645 /* Avoid recursion of direct reclaim */
3646 if (current
->flags
& PF_MEMALLOC
) {
3648 * __GFP_NOFAIL request from this context is rather bizarre
3649 * because we cannot reclaim anything and only can loop waiting
3650 * for somebody to do a work for us.
3652 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3659 /* Avoid allocations with no watermarks from looping endlessly */
3660 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3664 /* Try direct reclaim and then allocating */
3665 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3666 &did_some_progress
);
3670 /* Try direct compaction and then allocating */
3671 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3672 compact_priority
, &compact_result
);
3676 /* Do not loop if specifically requested */
3677 if (gfp_mask
& __GFP_NORETRY
)
3681 * Do not retry costly high order allocations unless they are
3684 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_REPEAT
))
3687 /* Make sure we know about allocations which stall for too long */
3688 if (time_after(jiffies
, alloc_start
+ stall_timeout
)) {
3689 warn_alloc(gfp_mask
,
3690 "page allocation stalls for %ums, order:%u",
3691 jiffies_to_msecs(jiffies
-alloc_start
), order
);
3692 stall_timeout
+= 10 * HZ
;
3695 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
3696 did_some_progress
> 0, &no_progress_loops
))
3700 * It doesn't make any sense to retry for the compaction if the order-0
3701 * reclaim is not able to make any progress because the current
3702 * implementation of the compaction depends on the sufficient amount
3703 * of free memory (see __compaction_suitable)
3705 if (did_some_progress
> 0 &&
3706 should_compact_retry(ac
, order
, alloc_flags
,
3707 compact_result
, &compact_priority
,
3708 &compaction_retries
))
3711 /* Reclaim has failed us, start killing things */
3712 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3716 /* Retry as long as the OOM killer is making progress */
3717 if (did_some_progress
) {
3718 no_progress_loops
= 0;
3723 warn_alloc(gfp_mask
,
3724 "page allocation failure: order:%u", order
);
3730 * This is the 'heart' of the zoned buddy allocator.
3733 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3734 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3737 unsigned int cpuset_mems_cookie
;
3738 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
3739 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3740 struct alloc_context ac
= {
3741 .high_zoneidx
= gfp_zone(gfp_mask
),
3742 .zonelist
= zonelist
,
3743 .nodemask
= nodemask
,
3744 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3747 if (cpusets_enabled()) {
3748 alloc_mask
|= __GFP_HARDWALL
;
3749 alloc_flags
|= ALLOC_CPUSET
;
3751 ac
.nodemask
= &cpuset_current_mems_allowed
;
3754 gfp_mask
&= gfp_allowed_mask
;
3756 lockdep_trace_alloc(gfp_mask
);
3758 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3760 if (should_fail_alloc_page(gfp_mask
, order
))
3764 * Check the zones suitable for the gfp_mask contain at least one
3765 * valid zone. It's possible to have an empty zonelist as a result
3766 * of __GFP_THISNODE and a memoryless node
3768 if (unlikely(!zonelist
->_zonerefs
->zone
))
3771 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3772 alloc_flags
|= ALLOC_CMA
;
3775 cpuset_mems_cookie
= read_mems_allowed_begin();
3777 /* Dirty zone balancing only done in the fast path */
3778 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3781 * The preferred zone is used for statistics but crucially it is
3782 * also used as the starting point for the zonelist iterator. It
3783 * may get reset for allocations that ignore memory policies.
3785 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3786 ac
.high_zoneidx
, ac
.nodemask
);
3787 if (!ac
.preferred_zoneref
) {
3792 /* First allocation attempt */
3793 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3798 * Runtime PM, block IO and its error handling path can deadlock
3799 * because I/O on the device might not complete.
3801 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3802 ac
.spread_dirty_pages
= false;
3805 * Restore the original nodemask if it was potentially replaced with
3806 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3808 if (cpusets_enabled())
3809 ac
.nodemask
= nodemask
;
3810 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3814 * When updating a task's mems_allowed, it is possible to race with
3815 * parallel threads in such a way that an allocation can fail while
3816 * the mask is being updated. If a page allocation is about to fail,
3817 * check if the cpuset changed during allocation and if so, retry.
3819 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3820 alloc_mask
= gfp_mask
;
3825 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
3826 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
3827 __free_pages(page
, order
);
3831 if (kmemcheck_enabled
&& page
)
3832 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3834 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3838 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3841 * Common helper functions.
3843 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3848 * __get_free_pages() returns a 32-bit address, which cannot represent
3851 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3853 page
= alloc_pages(gfp_mask
, order
);
3856 return (unsigned long) page_address(page
);
3858 EXPORT_SYMBOL(__get_free_pages
);
3860 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3862 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3864 EXPORT_SYMBOL(get_zeroed_page
);
3866 void __free_pages(struct page
*page
, unsigned int order
)
3868 if (put_page_testzero(page
)) {
3870 free_hot_cold_page(page
, false);
3872 __free_pages_ok(page
, order
);
3876 EXPORT_SYMBOL(__free_pages
);
3878 void free_pages(unsigned long addr
, unsigned int order
)
3881 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3882 __free_pages(virt_to_page((void *)addr
), order
);
3886 EXPORT_SYMBOL(free_pages
);
3890 * An arbitrary-length arbitrary-offset area of memory which resides
3891 * within a 0 or higher order page. Multiple fragments within that page
3892 * are individually refcounted, in the page's reference counter.
3894 * The page_frag functions below provide a simple allocation framework for
3895 * page fragments. This is used by the network stack and network device
3896 * drivers to provide a backing region of memory for use as either an
3897 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3899 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3902 struct page
*page
= NULL
;
3903 gfp_t gfp
= gfp_mask
;
3905 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3906 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3908 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3909 PAGE_FRAG_CACHE_MAX_ORDER
);
3910 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3912 if (unlikely(!page
))
3913 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3915 nc
->va
= page
? page_address(page
) : NULL
;
3920 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3921 unsigned int fragsz
, gfp_t gfp_mask
)
3923 unsigned int size
= PAGE_SIZE
;
3927 if (unlikely(!nc
->va
)) {
3929 page
= __page_frag_refill(nc
, gfp_mask
);
3933 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3934 /* if size can vary use size else just use PAGE_SIZE */
3937 /* Even if we own the page, we do not use atomic_set().
3938 * This would break get_page_unless_zero() users.
3940 page_ref_add(page
, size
- 1);
3942 /* reset page count bias and offset to start of new frag */
3943 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3944 nc
->pagecnt_bias
= size
;
3948 offset
= nc
->offset
- fragsz
;
3949 if (unlikely(offset
< 0)) {
3950 page
= virt_to_page(nc
->va
);
3952 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3955 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3956 /* if size can vary use size else just use PAGE_SIZE */
3959 /* OK, page count is 0, we can safely set it */
3960 set_page_count(page
, size
);
3962 /* reset page count bias and offset to start of new frag */
3963 nc
->pagecnt_bias
= size
;
3964 offset
= size
- fragsz
;
3968 nc
->offset
= offset
;
3970 return nc
->va
+ offset
;
3972 EXPORT_SYMBOL(__alloc_page_frag
);
3975 * Frees a page fragment allocated out of either a compound or order 0 page.
3977 void __free_page_frag(void *addr
)
3979 struct page
*page
= virt_to_head_page(addr
);
3981 if (unlikely(put_page_testzero(page
)))
3982 __free_pages_ok(page
, compound_order(page
));
3984 EXPORT_SYMBOL(__free_page_frag
);
3986 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3990 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3991 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3993 split_page(virt_to_page((void *)addr
), order
);
3994 while (used
< alloc_end
) {
3999 return (void *)addr
;
4003 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4004 * @size: the number of bytes to allocate
4005 * @gfp_mask: GFP flags for the allocation
4007 * This function is similar to alloc_pages(), except that it allocates the
4008 * minimum number of pages to satisfy the request. alloc_pages() can only
4009 * allocate memory in power-of-two pages.
4011 * This function is also limited by MAX_ORDER.
4013 * Memory allocated by this function must be released by free_pages_exact().
4015 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4017 unsigned int order
= get_order(size
);
4020 addr
= __get_free_pages(gfp_mask
, order
);
4021 return make_alloc_exact(addr
, order
, size
);
4023 EXPORT_SYMBOL(alloc_pages_exact
);
4026 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4028 * @nid: the preferred node ID where memory should be allocated
4029 * @size: the number of bytes to allocate
4030 * @gfp_mask: GFP flags for the allocation
4032 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4035 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4037 unsigned int order
= get_order(size
);
4038 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4041 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4045 * free_pages_exact - release memory allocated via alloc_pages_exact()
4046 * @virt: the value returned by alloc_pages_exact.
4047 * @size: size of allocation, same value as passed to alloc_pages_exact().
4049 * Release the memory allocated by a previous call to alloc_pages_exact.
4051 void free_pages_exact(void *virt
, size_t size
)
4053 unsigned long addr
= (unsigned long)virt
;
4054 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4056 while (addr
< end
) {
4061 EXPORT_SYMBOL(free_pages_exact
);
4064 * nr_free_zone_pages - count number of pages beyond high watermark
4065 * @offset: The zone index of the highest zone
4067 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4068 * high watermark within all zones at or below a given zone index. For each
4069 * zone, the number of pages is calculated as:
4070 * managed_pages - high_pages
4072 static unsigned long nr_free_zone_pages(int offset
)
4077 /* Just pick one node, since fallback list is circular */
4078 unsigned long sum
= 0;
4080 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4082 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4083 unsigned long size
= zone
->managed_pages
;
4084 unsigned long high
= high_wmark_pages(zone
);
4093 * nr_free_buffer_pages - count number of pages beyond high watermark
4095 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4096 * watermark within ZONE_DMA and ZONE_NORMAL.
4098 unsigned long nr_free_buffer_pages(void)
4100 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4102 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4105 * nr_free_pagecache_pages - count number of pages beyond high watermark
4107 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4108 * high watermark within all zones.
4110 unsigned long nr_free_pagecache_pages(void)
4112 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4115 static inline void show_node(struct zone
*zone
)
4117 if (IS_ENABLED(CONFIG_NUMA
))
4118 printk("Node %d ", zone_to_nid(zone
));
4121 long si_mem_available(void)
4124 unsigned long pagecache
;
4125 unsigned long wmark_low
= 0;
4126 unsigned long pages
[NR_LRU_LISTS
];
4130 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4131 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4134 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4137 * Estimate the amount of memory available for userspace allocations,
4138 * without causing swapping.
4140 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4143 * Not all the page cache can be freed, otherwise the system will
4144 * start swapping. Assume at least half of the page cache, or the
4145 * low watermark worth of cache, needs to stay.
4147 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4148 pagecache
-= min(pagecache
/ 2, wmark_low
);
4149 available
+= pagecache
;
4152 * Part of the reclaimable slab consists of items that are in use,
4153 * and cannot be freed. Cap this estimate at the low watermark.
4155 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
4156 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
4162 EXPORT_SYMBOL_GPL(si_mem_available
);
4164 void si_meminfo(struct sysinfo
*val
)
4166 val
->totalram
= totalram_pages
;
4167 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4168 val
->freeram
= global_page_state(NR_FREE_PAGES
);
4169 val
->bufferram
= nr_blockdev_pages();
4170 val
->totalhigh
= totalhigh_pages
;
4171 val
->freehigh
= nr_free_highpages();
4172 val
->mem_unit
= PAGE_SIZE
;
4175 EXPORT_SYMBOL(si_meminfo
);
4178 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4180 int zone_type
; /* needs to be signed */
4181 unsigned long managed_pages
= 0;
4182 unsigned long managed_highpages
= 0;
4183 unsigned long free_highpages
= 0;
4184 pg_data_t
*pgdat
= NODE_DATA(nid
);
4186 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4187 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4188 val
->totalram
= managed_pages
;
4189 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4190 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4191 #ifdef CONFIG_HIGHMEM
4192 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4193 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4195 if (is_highmem(zone
)) {
4196 managed_highpages
+= zone
->managed_pages
;
4197 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4200 val
->totalhigh
= managed_highpages
;
4201 val
->freehigh
= free_highpages
;
4203 val
->totalhigh
= managed_highpages
;
4204 val
->freehigh
= free_highpages
;
4206 val
->mem_unit
= PAGE_SIZE
;
4211 * Determine whether the node should be displayed or not, depending on whether
4212 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4214 bool skip_free_areas_node(unsigned int flags
, int nid
)
4217 unsigned int cpuset_mems_cookie
;
4219 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4223 cpuset_mems_cookie
= read_mems_allowed_begin();
4224 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
4225 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
4230 #define K(x) ((x) << (PAGE_SHIFT-10))
4232 static void show_migration_types(unsigned char type
)
4234 static const char types
[MIGRATE_TYPES
] = {
4235 [MIGRATE_UNMOVABLE
] = 'U',
4236 [MIGRATE_MOVABLE
] = 'M',
4237 [MIGRATE_RECLAIMABLE
] = 'E',
4238 [MIGRATE_HIGHATOMIC
] = 'H',
4240 [MIGRATE_CMA
] = 'C',
4242 #ifdef CONFIG_MEMORY_ISOLATION
4243 [MIGRATE_ISOLATE
] = 'I',
4246 char tmp
[MIGRATE_TYPES
+ 1];
4250 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4251 if (type
& (1 << i
))
4256 printk(KERN_CONT
"(%s) ", tmp
);
4260 * Show free area list (used inside shift_scroll-lock stuff)
4261 * We also calculate the percentage fragmentation. We do this by counting the
4262 * memory on each free list with the exception of the first item on the list.
4265 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4268 void show_free_areas(unsigned int filter
)
4270 unsigned long free_pcp
= 0;
4275 for_each_populated_zone(zone
) {
4276 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4279 for_each_online_cpu(cpu
)
4280 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4283 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4284 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4285 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4286 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4287 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4288 " free:%lu free_pcp:%lu free_cma:%lu\n",
4289 global_node_page_state(NR_ACTIVE_ANON
),
4290 global_node_page_state(NR_INACTIVE_ANON
),
4291 global_node_page_state(NR_ISOLATED_ANON
),
4292 global_node_page_state(NR_ACTIVE_FILE
),
4293 global_node_page_state(NR_INACTIVE_FILE
),
4294 global_node_page_state(NR_ISOLATED_FILE
),
4295 global_node_page_state(NR_UNEVICTABLE
),
4296 global_node_page_state(NR_FILE_DIRTY
),
4297 global_node_page_state(NR_WRITEBACK
),
4298 global_node_page_state(NR_UNSTABLE_NFS
),
4299 global_page_state(NR_SLAB_RECLAIMABLE
),
4300 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4301 global_node_page_state(NR_FILE_MAPPED
),
4302 global_node_page_state(NR_SHMEM
),
4303 global_page_state(NR_PAGETABLE
),
4304 global_page_state(NR_BOUNCE
),
4305 global_page_state(NR_FREE_PAGES
),
4307 global_page_state(NR_FREE_CMA_PAGES
));
4309 for_each_online_pgdat(pgdat
) {
4311 " active_anon:%lukB"
4312 " inactive_anon:%lukB"
4313 " active_file:%lukB"
4314 " inactive_file:%lukB"
4315 " unevictable:%lukB"
4316 " isolated(anon):%lukB"
4317 " isolated(file):%lukB"
4322 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4324 " shmem_pmdmapped: %lukB"
4327 " writeback_tmp:%lukB"
4329 " pages_scanned:%lu"
4330 " all_unreclaimable? %s"
4333 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4334 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4335 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4336 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4337 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4338 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4339 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4340 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4341 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4342 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4343 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4344 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4345 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4347 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4349 K(node_page_state(pgdat
, NR_SHMEM
)),
4350 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4351 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4352 node_page_state(pgdat
, NR_PAGES_SCANNED
),
4353 !pgdat_reclaimable(pgdat
) ? "yes" : "no");
4356 for_each_populated_zone(zone
) {
4359 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4363 for_each_online_cpu(cpu
)
4364 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4373 " active_anon:%lukB"
4374 " inactive_anon:%lukB"
4375 " active_file:%lukB"
4376 " inactive_file:%lukB"
4377 " unevictable:%lukB"
4378 " writepending:%lukB"
4382 " slab_reclaimable:%lukB"
4383 " slab_unreclaimable:%lukB"
4384 " kernel_stack:%lukB"
4392 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4393 K(min_wmark_pages(zone
)),
4394 K(low_wmark_pages(zone
)),
4395 K(high_wmark_pages(zone
)),
4396 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4397 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4398 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4399 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4400 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4401 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4402 K(zone
->present_pages
),
4403 K(zone
->managed_pages
),
4404 K(zone_page_state(zone
, NR_MLOCK
)),
4405 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4406 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4407 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4408 K(zone_page_state(zone
, NR_PAGETABLE
)),
4409 K(zone_page_state(zone
, NR_BOUNCE
)),
4411 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4412 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4413 printk("lowmem_reserve[]:");
4414 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4415 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
4416 printk(KERN_CONT
"\n");
4419 for_each_populated_zone(zone
) {
4421 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4422 unsigned char types
[MAX_ORDER
];
4424 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4427 printk(KERN_CONT
"%s: ", zone
->name
);
4429 spin_lock_irqsave(&zone
->lock
, flags
);
4430 for (order
= 0; order
< MAX_ORDER
; order
++) {
4431 struct free_area
*area
= &zone
->free_area
[order
];
4434 nr
[order
] = area
->nr_free
;
4435 total
+= nr
[order
] << order
;
4438 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4439 if (!list_empty(&area
->free_list
[type
]))
4440 types
[order
] |= 1 << type
;
4443 spin_unlock_irqrestore(&zone
->lock
, flags
);
4444 for (order
= 0; order
< MAX_ORDER
; order
++) {
4445 printk(KERN_CONT
"%lu*%lukB ",
4446 nr
[order
], K(1UL) << order
);
4448 show_migration_types(types
[order
]);
4450 printk(KERN_CONT
"= %lukB\n", K(total
));
4453 hugetlb_show_meminfo();
4455 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4457 show_swap_cache_info();
4460 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4462 zoneref
->zone
= zone
;
4463 zoneref
->zone_idx
= zone_idx(zone
);
4467 * Builds allocation fallback zone lists.
4469 * Add all populated zones of a node to the zonelist.
4471 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4475 enum zone_type zone_type
= MAX_NR_ZONES
;
4479 zone
= pgdat
->node_zones
+ zone_type
;
4480 if (managed_zone(zone
)) {
4481 zoneref_set_zone(zone
,
4482 &zonelist
->_zonerefs
[nr_zones
++]);
4483 check_highest_zone(zone_type
);
4485 } while (zone_type
);
4493 * 0 = automatic detection of better ordering.
4494 * 1 = order by ([node] distance, -zonetype)
4495 * 2 = order by (-zonetype, [node] distance)
4497 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4498 * the same zonelist. So only NUMA can configure this param.
4500 #define ZONELIST_ORDER_DEFAULT 0
4501 #define ZONELIST_ORDER_NODE 1
4502 #define ZONELIST_ORDER_ZONE 2
4504 /* zonelist order in the kernel.
4505 * set_zonelist_order() will set this to NODE or ZONE.
4507 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4508 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4512 /* The value user specified ....changed by config */
4513 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4514 /* string for sysctl */
4515 #define NUMA_ZONELIST_ORDER_LEN 16
4516 char numa_zonelist_order
[16] = "default";
4519 * interface for configure zonelist ordering.
4520 * command line option "numa_zonelist_order"
4521 * = "[dD]efault - default, automatic configuration.
4522 * = "[nN]ode - order by node locality, then by zone within node
4523 * = "[zZ]one - order by zone, then by locality within zone
4526 static int __parse_numa_zonelist_order(char *s
)
4528 if (*s
== 'd' || *s
== 'D') {
4529 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4530 } else if (*s
== 'n' || *s
== 'N') {
4531 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4532 } else if (*s
== 'z' || *s
== 'Z') {
4533 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4535 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4541 static __init
int setup_numa_zonelist_order(char *s
)
4548 ret
= __parse_numa_zonelist_order(s
);
4550 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4554 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4557 * sysctl handler for numa_zonelist_order
4559 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4560 void __user
*buffer
, size_t *length
,
4563 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4565 static DEFINE_MUTEX(zl_order_mutex
);
4567 mutex_lock(&zl_order_mutex
);
4569 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4573 strcpy(saved_string
, (char *)table
->data
);
4575 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4579 int oldval
= user_zonelist_order
;
4581 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4584 * bogus value. restore saved string
4586 strncpy((char *)table
->data
, saved_string
,
4587 NUMA_ZONELIST_ORDER_LEN
);
4588 user_zonelist_order
= oldval
;
4589 } else if (oldval
!= user_zonelist_order
) {
4590 mutex_lock(&zonelists_mutex
);
4591 build_all_zonelists(NULL
, NULL
);
4592 mutex_unlock(&zonelists_mutex
);
4596 mutex_unlock(&zl_order_mutex
);
4601 #define MAX_NODE_LOAD (nr_online_nodes)
4602 static int node_load
[MAX_NUMNODES
];
4605 * find_next_best_node - find the next node that should appear in a given node's fallback list
4606 * @node: node whose fallback list we're appending
4607 * @used_node_mask: nodemask_t of already used nodes
4609 * We use a number of factors to determine which is the next node that should
4610 * appear on a given node's fallback list. The node should not have appeared
4611 * already in @node's fallback list, and it should be the next closest node
4612 * according to the distance array (which contains arbitrary distance values
4613 * from each node to each node in the system), and should also prefer nodes
4614 * with no CPUs, since presumably they'll have very little allocation pressure
4615 * on them otherwise.
4616 * It returns -1 if no node is found.
4618 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4621 int min_val
= INT_MAX
;
4622 int best_node
= NUMA_NO_NODE
;
4623 const struct cpumask
*tmp
= cpumask_of_node(0);
4625 /* Use the local node if we haven't already */
4626 if (!node_isset(node
, *used_node_mask
)) {
4627 node_set(node
, *used_node_mask
);
4631 for_each_node_state(n
, N_MEMORY
) {
4633 /* Don't want a node to appear more than once */
4634 if (node_isset(n
, *used_node_mask
))
4637 /* Use the distance array to find the distance */
4638 val
= node_distance(node
, n
);
4640 /* Penalize nodes under us ("prefer the next node") */
4643 /* Give preference to headless and unused nodes */
4644 tmp
= cpumask_of_node(n
);
4645 if (!cpumask_empty(tmp
))
4646 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4648 /* Slight preference for less loaded node */
4649 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4650 val
+= node_load
[n
];
4652 if (val
< min_val
) {
4659 node_set(best_node
, *used_node_mask
);
4666 * Build zonelists ordered by node and zones within node.
4667 * This results in maximum locality--normal zone overflows into local
4668 * DMA zone, if any--but risks exhausting DMA zone.
4670 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4673 struct zonelist
*zonelist
;
4675 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4676 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4678 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4679 zonelist
->_zonerefs
[j
].zone
= NULL
;
4680 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4684 * Build gfp_thisnode zonelists
4686 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4689 struct zonelist
*zonelist
;
4691 zonelist
= &pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
];
4692 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4693 zonelist
->_zonerefs
[j
].zone
= NULL
;
4694 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4698 * Build zonelists ordered by zone and nodes within zones.
4699 * This results in conserving DMA zone[s] until all Normal memory is
4700 * exhausted, but results in overflowing to remote node while memory
4701 * may still exist in local DMA zone.
4703 static int node_order
[MAX_NUMNODES
];
4705 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4708 int zone_type
; /* needs to be signed */
4710 struct zonelist
*zonelist
;
4712 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4714 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4715 for (j
= 0; j
< nr_nodes
; j
++) {
4716 node
= node_order
[j
];
4717 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4718 if (managed_zone(z
)) {
4720 &zonelist
->_zonerefs
[pos
++]);
4721 check_highest_zone(zone_type
);
4725 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4726 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4729 #if defined(CONFIG_64BIT)
4731 * Devices that require DMA32/DMA are relatively rare and do not justify a
4732 * penalty to every machine in case the specialised case applies. Default
4733 * to Node-ordering on 64-bit NUMA machines
4735 static int default_zonelist_order(void)
4737 return ZONELIST_ORDER_NODE
;
4741 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4742 * by the kernel. If processes running on node 0 deplete the low memory zone
4743 * then reclaim will occur more frequency increasing stalls and potentially
4744 * be easier to OOM if a large percentage of the zone is under writeback or
4745 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4746 * Hence, default to zone ordering on 32-bit.
4748 static int default_zonelist_order(void)
4750 return ZONELIST_ORDER_ZONE
;
4752 #endif /* CONFIG_64BIT */
4754 static void set_zonelist_order(void)
4756 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4757 current_zonelist_order
= default_zonelist_order();
4759 current_zonelist_order
= user_zonelist_order
;
4762 static void build_zonelists(pg_data_t
*pgdat
)
4765 nodemask_t used_mask
;
4766 int local_node
, prev_node
;
4767 struct zonelist
*zonelist
;
4768 unsigned int order
= current_zonelist_order
;
4770 /* initialize zonelists */
4771 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4772 zonelist
= pgdat
->node_zonelists
+ i
;
4773 zonelist
->_zonerefs
[0].zone
= NULL
;
4774 zonelist
->_zonerefs
[0].zone_idx
= 0;
4777 /* NUMA-aware ordering of nodes */
4778 local_node
= pgdat
->node_id
;
4779 load
= nr_online_nodes
;
4780 prev_node
= local_node
;
4781 nodes_clear(used_mask
);
4783 memset(node_order
, 0, sizeof(node_order
));
4786 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4788 * We don't want to pressure a particular node.
4789 * So adding penalty to the first node in same
4790 * distance group to make it round-robin.
4792 if (node_distance(local_node
, node
) !=
4793 node_distance(local_node
, prev_node
))
4794 node_load
[node
] = load
;
4798 if (order
== ZONELIST_ORDER_NODE
)
4799 build_zonelists_in_node_order(pgdat
, node
);
4801 node_order
[i
++] = node
; /* remember order */
4804 if (order
== ZONELIST_ORDER_ZONE
) {
4805 /* calculate node order -- i.e., DMA last! */
4806 build_zonelists_in_zone_order(pgdat
, i
);
4809 build_thisnode_zonelists(pgdat
);
4812 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4814 * Return node id of node used for "local" allocations.
4815 * I.e., first node id of first zone in arg node's generic zonelist.
4816 * Used for initializing percpu 'numa_mem', which is used primarily
4817 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4819 int local_memory_node(int node
)
4823 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4824 gfp_zone(GFP_KERNEL
),
4826 return z
->zone
->node
;
4830 static void setup_min_unmapped_ratio(void);
4831 static void setup_min_slab_ratio(void);
4832 #else /* CONFIG_NUMA */
4834 static void set_zonelist_order(void)
4836 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4839 static void build_zonelists(pg_data_t
*pgdat
)
4841 int node
, local_node
;
4843 struct zonelist
*zonelist
;
4845 local_node
= pgdat
->node_id
;
4847 zonelist
= &pgdat
->node_zonelists
[ZONELIST_FALLBACK
];
4848 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4851 * Now we build the zonelist so that it contains the zones
4852 * of all the other nodes.
4853 * We don't want to pressure a particular node, so when
4854 * building the zones for node N, we make sure that the
4855 * zones coming right after the local ones are those from
4856 * node N+1 (modulo N)
4858 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4859 if (!node_online(node
))
4861 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4863 for (node
= 0; node
< local_node
; node
++) {
4864 if (!node_online(node
))
4866 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4869 zonelist
->_zonerefs
[j
].zone
= NULL
;
4870 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4873 #endif /* CONFIG_NUMA */
4876 * Boot pageset table. One per cpu which is going to be used for all
4877 * zones and all nodes. The parameters will be set in such a way
4878 * that an item put on a list will immediately be handed over to
4879 * the buddy list. This is safe since pageset manipulation is done
4880 * with interrupts disabled.
4882 * The boot_pagesets must be kept even after bootup is complete for
4883 * unused processors and/or zones. They do play a role for bootstrapping
4884 * hotplugged processors.
4886 * zoneinfo_show() and maybe other functions do
4887 * not check if the processor is online before following the pageset pointer.
4888 * Other parts of the kernel may not check if the zone is available.
4890 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4891 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4892 static void setup_zone_pageset(struct zone
*zone
);
4895 * Global mutex to protect against size modification of zonelists
4896 * as well as to serialize pageset setup for the new populated zone.
4898 DEFINE_MUTEX(zonelists_mutex
);
4900 /* return values int ....just for stop_machine() */
4901 static int __build_all_zonelists(void *data
)
4905 pg_data_t
*self
= data
;
4908 memset(node_load
, 0, sizeof(node_load
));
4911 if (self
&& !node_online(self
->node_id
)) {
4912 build_zonelists(self
);
4915 for_each_online_node(nid
) {
4916 pg_data_t
*pgdat
= NODE_DATA(nid
);
4918 build_zonelists(pgdat
);
4922 * Initialize the boot_pagesets that are going to be used
4923 * for bootstrapping processors. The real pagesets for
4924 * each zone will be allocated later when the per cpu
4925 * allocator is available.
4927 * boot_pagesets are used also for bootstrapping offline
4928 * cpus if the system is already booted because the pagesets
4929 * are needed to initialize allocators on a specific cpu too.
4930 * F.e. the percpu allocator needs the page allocator which
4931 * needs the percpu allocator in order to allocate its pagesets
4932 * (a chicken-egg dilemma).
4934 for_each_possible_cpu(cpu
) {
4935 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4937 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4939 * We now know the "local memory node" for each node--
4940 * i.e., the node of the first zone in the generic zonelist.
4941 * Set up numa_mem percpu variable for on-line cpus. During
4942 * boot, only the boot cpu should be on-line; we'll init the
4943 * secondary cpus' numa_mem as they come on-line. During
4944 * node/memory hotplug, we'll fixup all on-line cpus.
4946 if (cpu_online(cpu
))
4947 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4954 static noinline
void __init
4955 build_all_zonelists_init(void)
4957 __build_all_zonelists(NULL
);
4958 mminit_verify_zonelist();
4959 cpuset_init_current_mems_allowed();
4963 * Called with zonelists_mutex held always
4964 * unless system_state == SYSTEM_BOOTING.
4966 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4967 * [we're only called with non-NULL zone through __meminit paths] and
4968 * (2) call of __init annotated helper build_all_zonelists_init
4969 * [protected by SYSTEM_BOOTING].
4971 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4973 set_zonelist_order();
4975 if (system_state
== SYSTEM_BOOTING
) {
4976 build_all_zonelists_init();
4978 #ifdef CONFIG_MEMORY_HOTPLUG
4980 setup_zone_pageset(zone
);
4982 /* we have to stop all cpus to guarantee there is no user
4984 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4985 /* cpuset refresh routine should be here */
4987 vm_total_pages
= nr_free_pagecache_pages();
4989 * Disable grouping by mobility if the number of pages in the
4990 * system is too low to allow the mechanism to work. It would be
4991 * more accurate, but expensive to check per-zone. This check is
4992 * made on memory-hotadd so a system can start with mobility
4993 * disabled and enable it later
4995 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4996 page_group_by_mobility_disabled
= 1;
4998 page_group_by_mobility_disabled
= 0;
5000 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5002 zonelist_order_name
[current_zonelist_order
],
5003 page_group_by_mobility_disabled
? "off" : "on",
5006 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5011 * Initially all pages are reserved - free ones are freed
5012 * up by free_all_bootmem() once the early boot process is
5013 * done. Non-atomic initialization, single-pass.
5015 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5016 unsigned long start_pfn
, enum memmap_context context
)
5018 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5019 unsigned long end_pfn
= start_pfn
+ size
;
5020 pg_data_t
*pgdat
= NODE_DATA(nid
);
5022 unsigned long nr_initialised
= 0;
5023 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5024 struct memblock_region
*r
= NULL
, *tmp
;
5027 if (highest_memmap_pfn
< end_pfn
- 1)
5028 highest_memmap_pfn
= end_pfn
- 1;
5031 * Honor reservation requested by the driver for this ZONE_DEVICE
5034 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5035 start_pfn
+= altmap
->reserve
;
5037 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5039 * There can be holes in boot-time mem_map[]s handed to this
5040 * function. They do not exist on hotplugged memory.
5042 if (context
!= MEMMAP_EARLY
)
5045 if (!early_pfn_valid(pfn
))
5047 if (!early_pfn_in_nid(pfn
, nid
))
5049 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5052 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5054 * Check given memblock attribute by firmware which can affect
5055 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5056 * mirrored, it's an overlapped memmap init. skip it.
5058 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5059 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5060 for_each_memblock(memory
, tmp
)
5061 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5065 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5066 memblock_is_mirror(r
)) {
5067 /* already initialized as NORMAL */
5068 pfn
= memblock_region_memory_end_pfn(r
);
5076 * Mark the block movable so that blocks are reserved for
5077 * movable at startup. This will force kernel allocations
5078 * to reserve their blocks rather than leaking throughout
5079 * the address space during boot when many long-lived
5080 * kernel allocations are made.
5082 * bitmap is created for zone's valid pfn range. but memmap
5083 * can be created for invalid pages (for alignment)
5084 * check here not to call set_pageblock_migratetype() against
5087 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5088 struct page
*page
= pfn_to_page(pfn
);
5090 __init_single_page(page
, pfn
, zone
, nid
);
5091 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5093 __init_single_pfn(pfn
, zone
, nid
);
5098 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5100 unsigned int order
, t
;
5101 for_each_migratetype_order(order
, t
) {
5102 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5103 zone
->free_area
[order
].nr_free
= 0;
5107 #ifndef __HAVE_ARCH_MEMMAP_INIT
5108 #define memmap_init(size, nid, zone, start_pfn) \
5109 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5112 static int zone_batchsize(struct zone
*zone
)
5118 * The per-cpu-pages pools are set to around 1000th of the
5119 * size of the zone. But no more than 1/2 of a meg.
5121 * OK, so we don't know how big the cache is. So guess.
5123 batch
= zone
->managed_pages
/ 1024;
5124 if (batch
* PAGE_SIZE
> 512 * 1024)
5125 batch
= (512 * 1024) / PAGE_SIZE
;
5126 batch
/= 4; /* We effectively *= 4 below */
5131 * Clamp the batch to a 2^n - 1 value. Having a power
5132 * of 2 value was found to be more likely to have
5133 * suboptimal cache aliasing properties in some cases.
5135 * For example if 2 tasks are alternately allocating
5136 * batches of pages, one task can end up with a lot
5137 * of pages of one half of the possible page colors
5138 * and the other with pages of the other colors.
5140 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5145 /* The deferral and batching of frees should be suppressed under NOMMU
5148 * The problem is that NOMMU needs to be able to allocate large chunks
5149 * of contiguous memory as there's no hardware page translation to
5150 * assemble apparent contiguous memory from discontiguous pages.
5152 * Queueing large contiguous runs of pages for batching, however,
5153 * causes the pages to actually be freed in smaller chunks. As there
5154 * can be a significant delay between the individual batches being
5155 * recycled, this leads to the once large chunks of space being
5156 * fragmented and becoming unavailable for high-order allocations.
5163 * pcp->high and pcp->batch values are related and dependent on one another:
5164 * ->batch must never be higher then ->high.
5165 * The following function updates them in a safe manner without read side
5168 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5169 * those fields changing asynchronously (acording the the above rule).
5171 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5172 * outside of boot time (or some other assurance that no concurrent updaters
5175 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5176 unsigned long batch
)
5178 /* start with a fail safe value for batch */
5182 /* Update high, then batch, in order */
5189 /* a companion to pageset_set_high() */
5190 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5192 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5195 static void pageset_init(struct per_cpu_pageset
*p
)
5197 struct per_cpu_pages
*pcp
;
5200 memset(p
, 0, sizeof(*p
));
5204 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5205 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5208 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5211 pageset_set_batch(p
, batch
);
5215 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5216 * to the value high for the pageset p.
5218 static void pageset_set_high(struct per_cpu_pageset
*p
,
5221 unsigned long batch
= max(1UL, high
/ 4);
5222 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5223 batch
= PAGE_SHIFT
* 8;
5225 pageset_update(&p
->pcp
, high
, batch
);
5228 static void pageset_set_high_and_batch(struct zone
*zone
,
5229 struct per_cpu_pageset
*pcp
)
5231 if (percpu_pagelist_fraction
)
5232 pageset_set_high(pcp
,
5233 (zone
->managed_pages
/
5234 percpu_pagelist_fraction
));
5236 pageset_set_batch(pcp
, zone_batchsize(zone
));
5239 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5241 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5244 pageset_set_high_and_batch(zone
, pcp
);
5247 static void __meminit
setup_zone_pageset(struct zone
*zone
)
5250 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5251 for_each_possible_cpu(cpu
)
5252 zone_pageset_init(zone
, cpu
);
5256 * Allocate per cpu pagesets and initialize them.
5257 * Before this call only boot pagesets were available.
5259 void __init
setup_per_cpu_pageset(void)
5261 struct pglist_data
*pgdat
;
5264 for_each_populated_zone(zone
)
5265 setup_zone_pageset(zone
);
5267 for_each_online_pgdat(pgdat
)
5268 pgdat
->per_cpu_nodestats
=
5269 alloc_percpu(struct per_cpu_nodestat
);
5272 static __meminit
void zone_pcp_init(struct zone
*zone
)
5275 * per cpu subsystem is not up at this point. The following code
5276 * relies on the ability of the linker to provide the
5277 * offset of a (static) per cpu variable into the per cpu area.
5279 zone
->pageset
= &boot_pageset
;
5281 if (populated_zone(zone
))
5282 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5283 zone
->name
, zone
->present_pages
,
5284 zone_batchsize(zone
));
5287 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5288 unsigned long zone_start_pfn
,
5291 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5293 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5295 zone
->zone_start_pfn
= zone_start_pfn
;
5297 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5298 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5300 (unsigned long)zone_idx(zone
),
5301 zone_start_pfn
, (zone_start_pfn
+ size
));
5303 zone_init_free_lists(zone
);
5304 zone
->initialized
= 1;
5309 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5310 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5313 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5315 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5316 struct mminit_pfnnid_cache
*state
)
5318 unsigned long start_pfn
, end_pfn
;
5321 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5322 return state
->last_nid
;
5324 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5326 state
->last_start
= start_pfn
;
5327 state
->last_end
= end_pfn
;
5328 state
->last_nid
= nid
;
5333 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5336 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5337 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5338 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5340 * If an architecture guarantees that all ranges registered contain no holes
5341 * and may be freed, this this function may be used instead of calling
5342 * memblock_free_early_nid() manually.
5344 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5346 unsigned long start_pfn
, end_pfn
;
5349 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5350 start_pfn
= min(start_pfn
, max_low_pfn
);
5351 end_pfn
= min(end_pfn
, max_low_pfn
);
5353 if (start_pfn
< end_pfn
)
5354 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5355 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5361 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5362 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5364 * If an architecture guarantees that all ranges registered contain no holes and may
5365 * be freed, this function may be used instead of calling memory_present() manually.
5367 void __init
sparse_memory_present_with_active_regions(int nid
)
5369 unsigned long start_pfn
, end_pfn
;
5372 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5373 memory_present(this_nid
, start_pfn
, end_pfn
);
5377 * get_pfn_range_for_nid - Return the start and end page frames for a node
5378 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5379 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5380 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5382 * It returns the start and end page frame of a node based on information
5383 * provided by memblock_set_node(). If called for a node
5384 * with no available memory, a warning is printed and the start and end
5387 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5388 unsigned long *start_pfn
, unsigned long *end_pfn
)
5390 unsigned long this_start_pfn
, this_end_pfn
;
5396 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5397 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5398 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5401 if (*start_pfn
== -1UL)
5406 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5407 * assumption is made that zones within a node are ordered in monotonic
5408 * increasing memory addresses so that the "highest" populated zone is used
5410 static void __init
find_usable_zone_for_movable(void)
5413 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5414 if (zone_index
== ZONE_MOVABLE
)
5417 if (arch_zone_highest_possible_pfn
[zone_index
] >
5418 arch_zone_lowest_possible_pfn
[zone_index
])
5422 VM_BUG_ON(zone_index
== -1);
5423 movable_zone
= zone_index
;
5427 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5428 * because it is sized independent of architecture. Unlike the other zones,
5429 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5430 * in each node depending on the size of each node and how evenly kernelcore
5431 * is distributed. This helper function adjusts the zone ranges
5432 * provided by the architecture for a given node by using the end of the
5433 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5434 * zones within a node are in order of monotonic increases memory addresses
5436 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5437 unsigned long zone_type
,
5438 unsigned long node_start_pfn
,
5439 unsigned long node_end_pfn
,
5440 unsigned long *zone_start_pfn
,
5441 unsigned long *zone_end_pfn
)
5443 /* Only adjust if ZONE_MOVABLE is on this node */
5444 if (zone_movable_pfn
[nid
]) {
5445 /* Size ZONE_MOVABLE */
5446 if (zone_type
== ZONE_MOVABLE
) {
5447 *zone_start_pfn
= zone_movable_pfn
[nid
];
5448 *zone_end_pfn
= min(node_end_pfn
,
5449 arch_zone_highest_possible_pfn
[movable_zone
]);
5451 /* Adjust for ZONE_MOVABLE starting within this range */
5452 } else if (!mirrored_kernelcore
&&
5453 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5454 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5455 *zone_end_pfn
= zone_movable_pfn
[nid
];
5457 /* Check if this whole range is within ZONE_MOVABLE */
5458 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5459 *zone_start_pfn
= *zone_end_pfn
;
5464 * Return the number of pages a zone spans in a node, including holes
5465 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5467 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5468 unsigned long zone_type
,
5469 unsigned long node_start_pfn
,
5470 unsigned long node_end_pfn
,
5471 unsigned long *zone_start_pfn
,
5472 unsigned long *zone_end_pfn
,
5473 unsigned long *ignored
)
5475 /* When hotadd a new node from cpu_up(), the node should be empty */
5476 if (!node_start_pfn
&& !node_end_pfn
)
5479 /* Get the start and end of the zone */
5480 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5481 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5482 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5483 node_start_pfn
, node_end_pfn
,
5484 zone_start_pfn
, zone_end_pfn
);
5486 /* Check that this node has pages within the zone's required range */
5487 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5490 /* Move the zone boundaries inside the node if necessary */
5491 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5492 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5494 /* Return the spanned pages */
5495 return *zone_end_pfn
- *zone_start_pfn
;
5499 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5500 * then all holes in the requested range will be accounted for.
5502 unsigned long __meminit
__absent_pages_in_range(int nid
,
5503 unsigned long range_start_pfn
,
5504 unsigned long range_end_pfn
)
5506 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5507 unsigned long start_pfn
, end_pfn
;
5510 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5511 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5512 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5513 nr_absent
-= end_pfn
- start_pfn
;
5519 * absent_pages_in_range - Return number of page frames in holes within a range
5520 * @start_pfn: The start PFN to start searching for holes
5521 * @end_pfn: The end PFN to stop searching for holes
5523 * It returns the number of pages frames in memory holes within a range.
5525 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5526 unsigned long end_pfn
)
5528 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5531 /* Return the number of page frames in holes in a zone on a node */
5532 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5533 unsigned long zone_type
,
5534 unsigned long node_start_pfn
,
5535 unsigned long node_end_pfn
,
5536 unsigned long *ignored
)
5538 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5539 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5540 unsigned long zone_start_pfn
, zone_end_pfn
;
5541 unsigned long nr_absent
;
5543 /* When hotadd a new node from cpu_up(), the node should be empty */
5544 if (!node_start_pfn
&& !node_end_pfn
)
5547 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5548 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5550 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5551 node_start_pfn
, node_end_pfn
,
5552 &zone_start_pfn
, &zone_end_pfn
);
5553 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5556 * ZONE_MOVABLE handling.
5557 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5560 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5561 unsigned long start_pfn
, end_pfn
;
5562 struct memblock_region
*r
;
5564 for_each_memblock(memory
, r
) {
5565 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5566 zone_start_pfn
, zone_end_pfn
);
5567 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5568 zone_start_pfn
, zone_end_pfn
);
5570 if (zone_type
== ZONE_MOVABLE
&&
5571 memblock_is_mirror(r
))
5572 nr_absent
+= end_pfn
- start_pfn
;
5574 if (zone_type
== ZONE_NORMAL
&&
5575 !memblock_is_mirror(r
))
5576 nr_absent
+= end_pfn
- start_pfn
;
5583 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5584 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5585 unsigned long zone_type
,
5586 unsigned long node_start_pfn
,
5587 unsigned long node_end_pfn
,
5588 unsigned long *zone_start_pfn
,
5589 unsigned long *zone_end_pfn
,
5590 unsigned long *zones_size
)
5594 *zone_start_pfn
= node_start_pfn
;
5595 for (zone
= 0; zone
< zone_type
; zone
++)
5596 *zone_start_pfn
+= zones_size
[zone
];
5598 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5600 return zones_size
[zone_type
];
5603 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5604 unsigned long zone_type
,
5605 unsigned long node_start_pfn
,
5606 unsigned long node_end_pfn
,
5607 unsigned long *zholes_size
)
5612 return zholes_size
[zone_type
];
5615 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5617 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5618 unsigned long node_start_pfn
,
5619 unsigned long node_end_pfn
,
5620 unsigned long *zones_size
,
5621 unsigned long *zholes_size
)
5623 unsigned long realtotalpages
= 0, totalpages
= 0;
5626 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5627 struct zone
*zone
= pgdat
->node_zones
+ i
;
5628 unsigned long zone_start_pfn
, zone_end_pfn
;
5629 unsigned long size
, real_size
;
5631 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5637 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5638 node_start_pfn
, node_end_pfn
,
5641 zone
->zone_start_pfn
= zone_start_pfn
;
5643 zone
->zone_start_pfn
= 0;
5644 zone
->spanned_pages
= size
;
5645 zone
->present_pages
= real_size
;
5648 realtotalpages
+= real_size
;
5651 pgdat
->node_spanned_pages
= totalpages
;
5652 pgdat
->node_present_pages
= realtotalpages
;
5653 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5657 #ifndef CONFIG_SPARSEMEM
5659 * Calculate the size of the zone->blockflags rounded to an unsigned long
5660 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5661 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5662 * round what is now in bits to nearest long in bits, then return it in
5665 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5667 unsigned long usemapsize
;
5669 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5670 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5671 usemapsize
= usemapsize
>> pageblock_order
;
5672 usemapsize
*= NR_PAGEBLOCK_BITS
;
5673 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5675 return usemapsize
/ 8;
5678 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5680 unsigned long zone_start_pfn
,
5681 unsigned long zonesize
)
5683 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5684 zone
->pageblock_flags
= NULL
;
5686 zone
->pageblock_flags
=
5687 memblock_virt_alloc_node_nopanic(usemapsize
,
5691 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5692 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5693 #endif /* CONFIG_SPARSEMEM */
5695 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5697 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5698 void __paginginit
set_pageblock_order(void)
5702 /* Check that pageblock_nr_pages has not already been setup */
5703 if (pageblock_order
)
5706 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5707 order
= HUGETLB_PAGE_ORDER
;
5709 order
= MAX_ORDER
- 1;
5712 * Assume the largest contiguous order of interest is a huge page.
5713 * This value may be variable depending on boot parameters on IA64 and
5716 pageblock_order
= order
;
5718 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5721 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5722 * is unused as pageblock_order is set at compile-time. See
5723 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5726 void __paginginit
set_pageblock_order(void)
5730 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5732 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5733 unsigned long present_pages
)
5735 unsigned long pages
= spanned_pages
;
5738 * Provide a more accurate estimation if there are holes within
5739 * the zone and SPARSEMEM is in use. If there are holes within the
5740 * zone, each populated memory region may cost us one or two extra
5741 * memmap pages due to alignment because memmap pages for each
5742 * populated regions may not naturally algined on page boundary.
5743 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5745 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5746 IS_ENABLED(CONFIG_SPARSEMEM
))
5747 pages
= present_pages
;
5749 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5753 * Set up the zone data structures:
5754 * - mark all pages reserved
5755 * - mark all memory queues empty
5756 * - clear the memory bitmaps
5758 * NOTE: pgdat should get zeroed by caller.
5760 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5763 int nid
= pgdat
->node_id
;
5766 pgdat_resize_init(pgdat
);
5767 #ifdef CONFIG_NUMA_BALANCING
5768 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5769 pgdat
->numabalancing_migrate_nr_pages
= 0;
5770 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5772 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5773 spin_lock_init(&pgdat
->split_queue_lock
);
5774 INIT_LIST_HEAD(&pgdat
->split_queue
);
5775 pgdat
->split_queue_len
= 0;
5777 init_waitqueue_head(&pgdat
->kswapd_wait
);
5778 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5779 #ifdef CONFIG_COMPACTION
5780 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5782 pgdat_page_ext_init(pgdat
);
5783 spin_lock_init(&pgdat
->lru_lock
);
5784 lruvec_init(node_lruvec(pgdat
));
5786 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5787 struct zone
*zone
= pgdat
->node_zones
+ j
;
5788 unsigned long size
, realsize
, freesize
, memmap_pages
;
5789 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5791 size
= zone
->spanned_pages
;
5792 realsize
= freesize
= zone
->present_pages
;
5795 * Adjust freesize so that it accounts for how much memory
5796 * is used by this zone for memmap. This affects the watermark
5797 * and per-cpu initialisations
5799 memmap_pages
= calc_memmap_size(size
, realsize
);
5800 if (!is_highmem_idx(j
)) {
5801 if (freesize
>= memmap_pages
) {
5802 freesize
-= memmap_pages
;
5805 " %s zone: %lu pages used for memmap\n",
5806 zone_names
[j
], memmap_pages
);
5808 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5809 zone_names
[j
], memmap_pages
, freesize
);
5812 /* Account for reserved pages */
5813 if (j
== 0 && freesize
> dma_reserve
) {
5814 freesize
-= dma_reserve
;
5815 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5816 zone_names
[0], dma_reserve
);
5819 if (!is_highmem_idx(j
))
5820 nr_kernel_pages
+= freesize
;
5821 /* Charge for highmem memmap if there are enough kernel pages */
5822 else if (nr_kernel_pages
> memmap_pages
* 2)
5823 nr_kernel_pages
-= memmap_pages
;
5824 nr_all_pages
+= freesize
;
5827 * Set an approximate value for lowmem here, it will be adjusted
5828 * when the bootmem allocator frees pages into the buddy system.
5829 * And all highmem pages will be managed by the buddy system.
5831 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5835 zone
->name
= zone_names
[j
];
5836 zone
->zone_pgdat
= pgdat
;
5837 spin_lock_init(&zone
->lock
);
5838 zone_seqlock_init(zone
);
5839 zone_pcp_init(zone
);
5844 set_pageblock_order();
5845 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5846 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5848 memmap_init(size
, nid
, j
, zone_start_pfn
);
5852 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
5854 unsigned long __maybe_unused start
= 0;
5855 unsigned long __maybe_unused offset
= 0;
5857 /* Skip empty nodes */
5858 if (!pgdat
->node_spanned_pages
)
5861 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5862 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5863 offset
= pgdat
->node_start_pfn
- start
;
5864 /* ia64 gets its own node_mem_map, before this, without bootmem */
5865 if (!pgdat
->node_mem_map
) {
5866 unsigned long size
, end
;
5870 * The zone's endpoints aren't required to be MAX_ORDER
5871 * aligned but the node_mem_map endpoints must be in order
5872 * for the buddy allocator to function correctly.
5874 end
= pgdat_end_pfn(pgdat
);
5875 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5876 size
= (end
- start
) * sizeof(struct page
);
5877 map
= alloc_remap(pgdat
->node_id
, size
);
5879 map
= memblock_virt_alloc_node_nopanic(size
,
5881 pgdat
->node_mem_map
= map
+ offset
;
5883 #ifndef CONFIG_NEED_MULTIPLE_NODES
5885 * With no DISCONTIG, the global mem_map is just set as node 0's
5887 if (pgdat
== NODE_DATA(0)) {
5888 mem_map
= NODE_DATA(0)->node_mem_map
;
5889 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5890 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5892 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5895 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5898 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5899 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5901 pg_data_t
*pgdat
= NODE_DATA(nid
);
5902 unsigned long start_pfn
= 0;
5903 unsigned long end_pfn
= 0;
5905 /* pg_data_t should be reset to zero when it's allocated */
5906 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
5908 reset_deferred_meminit(pgdat
);
5909 pgdat
->node_id
= nid
;
5910 pgdat
->node_start_pfn
= node_start_pfn
;
5911 pgdat
->per_cpu_nodestats
= NULL
;
5912 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5913 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5914 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5915 (u64
)start_pfn
<< PAGE_SHIFT
,
5916 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5918 start_pfn
= node_start_pfn
;
5920 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5921 zones_size
, zholes_size
);
5923 alloc_node_mem_map(pgdat
);
5924 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5925 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5926 nid
, (unsigned long)pgdat
,
5927 (unsigned long)pgdat
->node_mem_map
);
5930 free_area_init_core(pgdat
);
5933 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5935 #if MAX_NUMNODES > 1
5937 * Figure out the number of possible node ids.
5939 void __init
setup_nr_node_ids(void)
5941 unsigned int highest
;
5943 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5944 nr_node_ids
= highest
+ 1;
5949 * node_map_pfn_alignment - determine the maximum internode alignment
5951 * This function should be called after node map is populated and sorted.
5952 * It calculates the maximum power of two alignment which can distinguish
5955 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5956 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5957 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5958 * shifted, 1GiB is enough and this function will indicate so.
5960 * This is used to test whether pfn -> nid mapping of the chosen memory
5961 * model has fine enough granularity to avoid incorrect mapping for the
5962 * populated node map.
5964 * Returns the determined alignment in pfn's. 0 if there is no alignment
5965 * requirement (single node).
5967 unsigned long __init
node_map_pfn_alignment(void)
5969 unsigned long accl_mask
= 0, last_end
= 0;
5970 unsigned long start
, end
, mask
;
5974 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5975 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5982 * Start with a mask granular enough to pin-point to the
5983 * start pfn and tick off bits one-by-one until it becomes
5984 * too coarse to separate the current node from the last.
5986 mask
= ~((1 << __ffs(start
)) - 1);
5987 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5990 /* accumulate all internode masks */
5994 /* convert mask to number of pages */
5995 return ~accl_mask
+ 1;
5998 /* Find the lowest pfn for a node */
5999 static unsigned long __init
find_min_pfn_for_node(int nid
)
6001 unsigned long min_pfn
= ULONG_MAX
;
6002 unsigned long start_pfn
;
6005 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6006 min_pfn
= min(min_pfn
, start_pfn
);
6008 if (min_pfn
== ULONG_MAX
) {
6009 pr_warn("Could not find start_pfn for node %d\n", nid
);
6017 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6019 * It returns the minimum PFN based on information provided via
6020 * memblock_set_node().
6022 unsigned long __init
find_min_pfn_with_active_regions(void)
6024 return find_min_pfn_for_node(MAX_NUMNODES
);
6028 * early_calculate_totalpages()
6029 * Sum pages in active regions for movable zone.
6030 * Populate N_MEMORY for calculating usable_nodes.
6032 static unsigned long __init
early_calculate_totalpages(void)
6034 unsigned long totalpages
= 0;
6035 unsigned long start_pfn
, end_pfn
;
6038 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6039 unsigned long pages
= end_pfn
- start_pfn
;
6041 totalpages
+= pages
;
6043 node_set_state(nid
, N_MEMORY
);
6049 * Find the PFN the Movable zone begins in each node. Kernel memory
6050 * is spread evenly between nodes as long as the nodes have enough
6051 * memory. When they don't, some nodes will have more kernelcore than
6054 static void __init
find_zone_movable_pfns_for_nodes(void)
6057 unsigned long usable_startpfn
;
6058 unsigned long kernelcore_node
, kernelcore_remaining
;
6059 /* save the state before borrow the nodemask */
6060 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6061 unsigned long totalpages
= early_calculate_totalpages();
6062 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6063 struct memblock_region
*r
;
6065 /* Need to find movable_zone earlier when movable_node is specified. */
6066 find_usable_zone_for_movable();
6069 * If movable_node is specified, ignore kernelcore and movablecore
6072 if (movable_node_is_enabled()) {
6073 for_each_memblock(memory
, r
) {
6074 if (!memblock_is_hotpluggable(r
))
6079 usable_startpfn
= PFN_DOWN(r
->base
);
6080 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6081 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6089 * If kernelcore=mirror is specified, ignore movablecore option
6091 if (mirrored_kernelcore
) {
6092 bool mem_below_4gb_not_mirrored
= false;
6094 for_each_memblock(memory
, r
) {
6095 if (memblock_is_mirror(r
))
6100 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6102 if (usable_startpfn
< 0x100000) {
6103 mem_below_4gb_not_mirrored
= true;
6107 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6108 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6112 if (mem_below_4gb_not_mirrored
)
6113 pr_warn("This configuration results in unmirrored kernel memory.");
6119 * If movablecore=nn[KMG] was specified, calculate what size of
6120 * kernelcore that corresponds so that memory usable for
6121 * any allocation type is evenly spread. If both kernelcore
6122 * and movablecore are specified, then the value of kernelcore
6123 * will be used for required_kernelcore if it's greater than
6124 * what movablecore would have allowed.
6126 if (required_movablecore
) {
6127 unsigned long corepages
;
6130 * Round-up so that ZONE_MOVABLE is at least as large as what
6131 * was requested by the user
6133 required_movablecore
=
6134 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6135 required_movablecore
= min(totalpages
, required_movablecore
);
6136 corepages
= totalpages
- required_movablecore
;
6138 required_kernelcore
= max(required_kernelcore
, corepages
);
6142 * If kernelcore was not specified or kernelcore size is larger
6143 * than totalpages, there is no ZONE_MOVABLE.
6145 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6148 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6149 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6152 /* Spread kernelcore memory as evenly as possible throughout nodes */
6153 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6154 for_each_node_state(nid
, N_MEMORY
) {
6155 unsigned long start_pfn
, end_pfn
;
6158 * Recalculate kernelcore_node if the division per node
6159 * now exceeds what is necessary to satisfy the requested
6160 * amount of memory for the kernel
6162 if (required_kernelcore
< kernelcore_node
)
6163 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6166 * As the map is walked, we track how much memory is usable
6167 * by the kernel using kernelcore_remaining. When it is
6168 * 0, the rest of the node is usable by ZONE_MOVABLE
6170 kernelcore_remaining
= kernelcore_node
;
6172 /* Go through each range of PFNs within this node */
6173 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6174 unsigned long size_pages
;
6176 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6177 if (start_pfn
>= end_pfn
)
6180 /* Account for what is only usable for kernelcore */
6181 if (start_pfn
< usable_startpfn
) {
6182 unsigned long kernel_pages
;
6183 kernel_pages
= min(end_pfn
, usable_startpfn
)
6186 kernelcore_remaining
-= min(kernel_pages
,
6187 kernelcore_remaining
);
6188 required_kernelcore
-= min(kernel_pages
,
6189 required_kernelcore
);
6191 /* Continue if range is now fully accounted */
6192 if (end_pfn
<= usable_startpfn
) {
6195 * Push zone_movable_pfn to the end so
6196 * that if we have to rebalance
6197 * kernelcore across nodes, we will
6198 * not double account here
6200 zone_movable_pfn
[nid
] = end_pfn
;
6203 start_pfn
= usable_startpfn
;
6207 * The usable PFN range for ZONE_MOVABLE is from
6208 * start_pfn->end_pfn. Calculate size_pages as the
6209 * number of pages used as kernelcore
6211 size_pages
= end_pfn
- start_pfn
;
6212 if (size_pages
> kernelcore_remaining
)
6213 size_pages
= kernelcore_remaining
;
6214 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6217 * Some kernelcore has been met, update counts and
6218 * break if the kernelcore for this node has been
6221 required_kernelcore
-= min(required_kernelcore
,
6223 kernelcore_remaining
-= size_pages
;
6224 if (!kernelcore_remaining
)
6230 * If there is still required_kernelcore, we do another pass with one
6231 * less node in the count. This will push zone_movable_pfn[nid] further
6232 * along on the nodes that still have memory until kernelcore is
6236 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6240 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6241 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6242 zone_movable_pfn
[nid
] =
6243 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6246 /* restore the node_state */
6247 node_states
[N_MEMORY
] = saved_node_state
;
6250 /* Any regular or high memory on that node ? */
6251 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6253 enum zone_type zone_type
;
6255 if (N_MEMORY
== N_NORMAL_MEMORY
)
6258 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6259 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6260 if (populated_zone(zone
)) {
6261 node_set_state(nid
, N_HIGH_MEMORY
);
6262 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6263 zone_type
<= ZONE_NORMAL
)
6264 node_set_state(nid
, N_NORMAL_MEMORY
);
6271 * free_area_init_nodes - Initialise all pg_data_t and zone data
6272 * @max_zone_pfn: an array of max PFNs for each zone
6274 * This will call free_area_init_node() for each active node in the system.
6275 * Using the page ranges provided by memblock_set_node(), the size of each
6276 * zone in each node and their holes is calculated. If the maximum PFN
6277 * between two adjacent zones match, it is assumed that the zone is empty.
6278 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6279 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6280 * starts where the previous one ended. For example, ZONE_DMA32 starts
6281 * at arch_max_dma_pfn.
6283 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6285 unsigned long start_pfn
, end_pfn
;
6288 /* Record where the zone boundaries are */
6289 memset(arch_zone_lowest_possible_pfn
, 0,
6290 sizeof(arch_zone_lowest_possible_pfn
));
6291 memset(arch_zone_highest_possible_pfn
, 0,
6292 sizeof(arch_zone_highest_possible_pfn
));
6294 start_pfn
= find_min_pfn_with_active_regions();
6296 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6297 if (i
== ZONE_MOVABLE
)
6300 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6301 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6302 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6304 start_pfn
= end_pfn
;
6306 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6307 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6309 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6310 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6311 find_zone_movable_pfns_for_nodes();
6313 /* Print out the zone ranges */
6314 pr_info("Zone ranges:\n");
6315 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6316 if (i
== ZONE_MOVABLE
)
6318 pr_info(" %-8s ", zone_names
[i
]);
6319 if (arch_zone_lowest_possible_pfn
[i
] ==
6320 arch_zone_highest_possible_pfn
[i
])
6323 pr_cont("[mem %#018Lx-%#018Lx]\n",
6324 (u64
)arch_zone_lowest_possible_pfn
[i
]
6326 ((u64
)arch_zone_highest_possible_pfn
[i
]
6327 << PAGE_SHIFT
) - 1);
6330 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6331 pr_info("Movable zone start for each node\n");
6332 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6333 if (zone_movable_pfn
[i
])
6334 pr_info(" Node %d: %#018Lx\n", i
,
6335 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6338 /* Print out the early node map */
6339 pr_info("Early memory node ranges\n");
6340 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6341 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6342 (u64
)start_pfn
<< PAGE_SHIFT
,
6343 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6345 /* Initialise every node */
6346 mminit_verify_pageflags_layout();
6347 setup_nr_node_ids();
6348 for_each_online_node(nid
) {
6349 pg_data_t
*pgdat
= NODE_DATA(nid
);
6350 free_area_init_node(nid
, NULL
,
6351 find_min_pfn_for_node(nid
), NULL
);
6353 /* Any memory on that node */
6354 if (pgdat
->node_present_pages
)
6355 node_set_state(nid
, N_MEMORY
);
6356 check_for_memory(pgdat
, nid
);
6360 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6362 unsigned long long coremem
;
6366 coremem
= memparse(p
, &p
);
6367 *core
= coremem
>> PAGE_SHIFT
;
6369 /* Paranoid check that UL is enough for the coremem value */
6370 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6376 * kernelcore=size sets the amount of memory for use for allocations that
6377 * cannot be reclaimed or migrated.
6379 static int __init
cmdline_parse_kernelcore(char *p
)
6381 /* parse kernelcore=mirror */
6382 if (parse_option_str(p
, "mirror")) {
6383 mirrored_kernelcore
= true;
6387 return cmdline_parse_core(p
, &required_kernelcore
);
6391 * movablecore=size sets the amount of memory for use for allocations that
6392 * can be reclaimed or migrated.
6394 static int __init
cmdline_parse_movablecore(char *p
)
6396 return cmdline_parse_core(p
, &required_movablecore
);
6399 early_param("kernelcore", cmdline_parse_kernelcore
);
6400 early_param("movablecore", cmdline_parse_movablecore
);
6402 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6404 void adjust_managed_page_count(struct page
*page
, long count
)
6406 spin_lock(&managed_page_count_lock
);
6407 page_zone(page
)->managed_pages
+= count
;
6408 totalram_pages
+= count
;
6409 #ifdef CONFIG_HIGHMEM
6410 if (PageHighMem(page
))
6411 totalhigh_pages
+= count
;
6413 spin_unlock(&managed_page_count_lock
);
6415 EXPORT_SYMBOL(adjust_managed_page_count
);
6417 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6420 unsigned long pages
= 0;
6422 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6423 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6424 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6425 if ((unsigned int)poison
<= 0xFF)
6426 memset(pos
, poison
, PAGE_SIZE
);
6427 free_reserved_page(virt_to_page(pos
));
6431 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6432 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6436 EXPORT_SYMBOL(free_reserved_area
);
6438 #ifdef CONFIG_HIGHMEM
6439 void free_highmem_page(struct page
*page
)
6441 __free_reserved_page(page
);
6443 page_zone(page
)->managed_pages
++;
6449 void __init
mem_init_print_info(const char *str
)
6451 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6452 unsigned long init_code_size
, init_data_size
;
6454 physpages
= get_num_physpages();
6455 codesize
= _etext
- _stext
;
6456 datasize
= _edata
- _sdata
;
6457 rosize
= __end_rodata
- __start_rodata
;
6458 bss_size
= __bss_stop
- __bss_start
;
6459 init_data_size
= __init_end
- __init_begin
;
6460 init_code_size
= _einittext
- _sinittext
;
6463 * Detect special cases and adjust section sizes accordingly:
6464 * 1) .init.* may be embedded into .data sections
6465 * 2) .init.text.* may be out of [__init_begin, __init_end],
6466 * please refer to arch/tile/kernel/vmlinux.lds.S.
6467 * 3) .rodata.* may be embedded into .text or .data sections.
6469 #define adj_init_size(start, end, size, pos, adj) \
6471 if (start <= pos && pos < end && size > adj) \
6475 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6476 _sinittext
, init_code_size
);
6477 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6478 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6479 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6480 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6482 #undef adj_init_size
6484 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6485 #ifdef CONFIG_HIGHMEM
6489 nr_free_pages() << (PAGE_SHIFT
- 10),
6490 physpages
<< (PAGE_SHIFT
- 10),
6491 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6492 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6493 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6494 totalcma_pages
<< (PAGE_SHIFT
- 10),
6495 #ifdef CONFIG_HIGHMEM
6496 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6498 str
? ", " : "", str
? str
: "");
6502 * set_dma_reserve - set the specified number of pages reserved in the first zone
6503 * @new_dma_reserve: The number of pages to mark reserved
6505 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6506 * In the DMA zone, a significant percentage may be consumed by kernel image
6507 * and other unfreeable allocations which can skew the watermarks badly. This
6508 * function may optionally be used to account for unfreeable pages in the
6509 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6510 * smaller per-cpu batchsize.
6512 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6514 dma_reserve
= new_dma_reserve
;
6517 void __init
free_area_init(unsigned long *zones_size
)
6519 free_area_init_node(0, zones_size
,
6520 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6523 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6524 unsigned long action
, void *hcpu
)
6526 int cpu
= (unsigned long)hcpu
;
6528 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6529 lru_add_drain_cpu(cpu
);
6533 * Spill the event counters of the dead processor
6534 * into the current processors event counters.
6535 * This artificially elevates the count of the current
6538 vm_events_fold_cpu(cpu
);
6541 * Zero the differential counters of the dead processor
6542 * so that the vm statistics are consistent.
6544 * This is only okay since the processor is dead and cannot
6545 * race with what we are doing.
6547 cpu_vm_stats_fold(cpu
);
6552 void __init
page_alloc_init(void)
6554 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6558 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6559 * or min_free_kbytes changes.
6561 static void calculate_totalreserve_pages(void)
6563 struct pglist_data
*pgdat
;
6564 unsigned long reserve_pages
= 0;
6565 enum zone_type i
, j
;
6567 for_each_online_pgdat(pgdat
) {
6569 pgdat
->totalreserve_pages
= 0;
6571 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6572 struct zone
*zone
= pgdat
->node_zones
+ i
;
6575 /* Find valid and maximum lowmem_reserve in the zone */
6576 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6577 if (zone
->lowmem_reserve
[j
] > max
)
6578 max
= zone
->lowmem_reserve
[j
];
6581 /* we treat the high watermark as reserved pages. */
6582 max
+= high_wmark_pages(zone
);
6584 if (max
> zone
->managed_pages
)
6585 max
= zone
->managed_pages
;
6587 pgdat
->totalreserve_pages
+= max
;
6589 reserve_pages
+= max
;
6592 totalreserve_pages
= reserve_pages
;
6596 * setup_per_zone_lowmem_reserve - called whenever
6597 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6598 * has a correct pages reserved value, so an adequate number of
6599 * pages are left in the zone after a successful __alloc_pages().
6601 static void setup_per_zone_lowmem_reserve(void)
6603 struct pglist_data
*pgdat
;
6604 enum zone_type j
, idx
;
6606 for_each_online_pgdat(pgdat
) {
6607 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6608 struct zone
*zone
= pgdat
->node_zones
+ j
;
6609 unsigned long managed_pages
= zone
->managed_pages
;
6611 zone
->lowmem_reserve
[j
] = 0;
6615 struct zone
*lower_zone
;
6619 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6620 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6622 lower_zone
= pgdat
->node_zones
+ idx
;
6623 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6624 sysctl_lowmem_reserve_ratio
[idx
];
6625 managed_pages
+= lower_zone
->managed_pages
;
6630 /* update totalreserve_pages */
6631 calculate_totalreserve_pages();
6634 static void __setup_per_zone_wmarks(void)
6636 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6637 unsigned long lowmem_pages
= 0;
6639 unsigned long flags
;
6641 /* Calculate total number of !ZONE_HIGHMEM pages */
6642 for_each_zone(zone
) {
6643 if (!is_highmem(zone
))
6644 lowmem_pages
+= zone
->managed_pages
;
6647 for_each_zone(zone
) {
6650 spin_lock_irqsave(&zone
->lock
, flags
);
6651 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6652 do_div(tmp
, lowmem_pages
);
6653 if (is_highmem(zone
)) {
6655 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6656 * need highmem pages, so cap pages_min to a small
6659 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6660 * deltas control asynch page reclaim, and so should
6661 * not be capped for highmem.
6663 unsigned long min_pages
;
6665 min_pages
= zone
->managed_pages
/ 1024;
6666 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6667 zone
->watermark
[WMARK_MIN
] = min_pages
;
6670 * If it's a lowmem zone, reserve a number of pages
6671 * proportionate to the zone's size.
6673 zone
->watermark
[WMARK_MIN
] = tmp
;
6677 * Set the kswapd watermarks distance according to the
6678 * scale factor in proportion to available memory, but
6679 * ensure a minimum size on small systems.
6681 tmp
= max_t(u64
, tmp
>> 2,
6682 mult_frac(zone
->managed_pages
,
6683 watermark_scale_factor
, 10000));
6685 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6686 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6688 spin_unlock_irqrestore(&zone
->lock
, flags
);
6691 /* update totalreserve_pages */
6692 calculate_totalreserve_pages();
6696 * setup_per_zone_wmarks - called when min_free_kbytes changes
6697 * or when memory is hot-{added|removed}
6699 * Ensures that the watermark[min,low,high] values for each zone are set
6700 * correctly with respect to min_free_kbytes.
6702 void setup_per_zone_wmarks(void)
6704 mutex_lock(&zonelists_mutex
);
6705 __setup_per_zone_wmarks();
6706 mutex_unlock(&zonelists_mutex
);
6710 * Initialise min_free_kbytes.
6712 * For small machines we want it small (128k min). For large machines
6713 * we want it large (64MB max). But it is not linear, because network
6714 * bandwidth does not increase linearly with machine size. We use
6716 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6717 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6733 int __meminit
init_per_zone_wmark_min(void)
6735 unsigned long lowmem_kbytes
;
6736 int new_min_free_kbytes
;
6738 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6739 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6741 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6742 min_free_kbytes
= new_min_free_kbytes
;
6743 if (min_free_kbytes
< 128)
6744 min_free_kbytes
= 128;
6745 if (min_free_kbytes
> 65536)
6746 min_free_kbytes
= 65536;
6748 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6749 new_min_free_kbytes
, user_min_free_kbytes
);
6751 setup_per_zone_wmarks();
6752 refresh_zone_stat_thresholds();
6753 setup_per_zone_lowmem_reserve();
6756 setup_min_unmapped_ratio();
6757 setup_min_slab_ratio();
6762 core_initcall(init_per_zone_wmark_min
)
6765 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6766 * that we can call two helper functions whenever min_free_kbytes
6769 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6770 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6774 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6779 user_min_free_kbytes
= min_free_kbytes
;
6780 setup_per_zone_wmarks();
6785 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6786 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6790 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6795 setup_per_zone_wmarks();
6801 static void setup_min_unmapped_ratio(void)
6806 for_each_online_pgdat(pgdat
)
6807 pgdat
->min_unmapped_pages
= 0;
6810 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
6811 sysctl_min_unmapped_ratio
) / 100;
6815 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6816 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6820 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6824 setup_min_unmapped_ratio();
6829 static void setup_min_slab_ratio(void)
6834 for_each_online_pgdat(pgdat
)
6835 pgdat
->min_slab_pages
= 0;
6838 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
6839 sysctl_min_slab_ratio
) / 100;
6842 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6843 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6847 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6851 setup_min_slab_ratio();
6858 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6859 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6860 * whenever sysctl_lowmem_reserve_ratio changes.
6862 * The reserve ratio obviously has absolutely no relation with the
6863 * minimum watermarks. The lowmem reserve ratio can only make sense
6864 * if in function of the boot time zone sizes.
6866 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6867 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6869 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6870 setup_per_zone_lowmem_reserve();
6875 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6876 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6877 * pagelist can have before it gets flushed back to buddy allocator.
6879 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6880 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6883 int old_percpu_pagelist_fraction
;
6886 mutex_lock(&pcp_batch_high_lock
);
6887 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6889 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6890 if (!write
|| ret
< 0)
6893 /* Sanity checking to avoid pcp imbalance */
6894 if (percpu_pagelist_fraction
&&
6895 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6896 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6902 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6905 for_each_populated_zone(zone
) {
6908 for_each_possible_cpu(cpu
)
6909 pageset_set_high_and_batch(zone
,
6910 per_cpu_ptr(zone
->pageset
, cpu
));
6913 mutex_unlock(&pcp_batch_high_lock
);
6918 int hashdist
= HASHDIST_DEFAULT
;
6920 static int __init
set_hashdist(char *str
)
6924 hashdist
= simple_strtoul(str
, &str
, 0);
6927 __setup("hashdist=", set_hashdist
);
6930 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6932 * Returns the number of pages that arch has reserved but
6933 * is not known to alloc_large_system_hash().
6935 static unsigned long __init
arch_reserved_kernel_pages(void)
6942 * allocate a large system hash table from bootmem
6943 * - it is assumed that the hash table must contain an exact power-of-2
6944 * quantity of entries
6945 * - limit is the number of hash buckets, not the total allocation size
6947 void *__init
alloc_large_system_hash(const char *tablename
,
6948 unsigned long bucketsize
,
6949 unsigned long numentries
,
6952 unsigned int *_hash_shift
,
6953 unsigned int *_hash_mask
,
6954 unsigned long low_limit
,
6955 unsigned long high_limit
)
6957 unsigned long long max
= high_limit
;
6958 unsigned long log2qty
, size
;
6961 /* allow the kernel cmdline to have a say */
6963 /* round applicable memory size up to nearest megabyte */
6964 numentries
= nr_kernel_pages
;
6965 numentries
-= arch_reserved_kernel_pages();
6967 /* It isn't necessary when PAGE_SIZE >= 1MB */
6968 if (PAGE_SHIFT
< 20)
6969 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6971 /* limit to 1 bucket per 2^scale bytes of low memory */
6972 if (scale
> PAGE_SHIFT
)
6973 numentries
>>= (scale
- PAGE_SHIFT
);
6975 numentries
<<= (PAGE_SHIFT
- scale
);
6977 /* Make sure we've got at least a 0-order allocation.. */
6978 if (unlikely(flags
& HASH_SMALL
)) {
6979 /* Makes no sense without HASH_EARLY */
6980 WARN_ON(!(flags
& HASH_EARLY
));
6981 if (!(numentries
>> *_hash_shift
)) {
6982 numentries
= 1UL << *_hash_shift
;
6983 BUG_ON(!numentries
);
6985 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6986 numentries
= PAGE_SIZE
/ bucketsize
;
6988 numentries
= roundup_pow_of_two(numentries
);
6990 /* limit allocation size to 1/16 total memory by default */
6992 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6993 do_div(max
, bucketsize
);
6995 max
= min(max
, 0x80000000ULL
);
6997 if (numentries
< low_limit
)
6998 numentries
= low_limit
;
6999 if (numentries
> max
)
7002 log2qty
= ilog2(numentries
);
7005 size
= bucketsize
<< log2qty
;
7006 if (flags
& HASH_EARLY
)
7007 table
= memblock_virt_alloc_nopanic(size
, 0);
7009 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
7012 * If bucketsize is not a power-of-two, we may free
7013 * some pages at the end of hash table which
7014 * alloc_pages_exact() automatically does
7016 if (get_order(size
) < MAX_ORDER
) {
7017 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
7018 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
7021 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7024 panic("Failed to allocate %s hash table\n", tablename
);
7026 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7027 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7030 *_hash_shift
= log2qty
;
7032 *_hash_mask
= (1 << log2qty
) - 1;
7038 * This function checks whether pageblock includes unmovable pages or not.
7039 * If @count is not zero, it is okay to include less @count unmovable pages
7041 * PageLRU check without isolation or lru_lock could race so that
7042 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7043 * expect this function should be exact.
7045 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7046 bool skip_hwpoisoned_pages
)
7048 unsigned long pfn
, iter
, found
;
7052 * For avoiding noise data, lru_add_drain_all() should be called
7053 * If ZONE_MOVABLE, the zone never contains unmovable pages
7055 if (zone_idx(zone
) == ZONE_MOVABLE
)
7057 mt
= get_pageblock_migratetype(page
);
7058 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7061 pfn
= page_to_pfn(page
);
7062 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7063 unsigned long check
= pfn
+ iter
;
7065 if (!pfn_valid_within(check
))
7068 page
= pfn_to_page(check
);
7071 * Hugepages are not in LRU lists, but they're movable.
7072 * We need not scan over tail pages bacause we don't
7073 * handle each tail page individually in migration.
7075 if (PageHuge(page
)) {
7076 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7081 * We can't use page_count without pin a page
7082 * because another CPU can free compound page.
7083 * This check already skips compound tails of THP
7084 * because their page->_refcount is zero at all time.
7086 if (!page_ref_count(page
)) {
7087 if (PageBuddy(page
))
7088 iter
+= (1 << page_order(page
)) - 1;
7093 * The HWPoisoned page may be not in buddy system, and
7094 * page_count() is not 0.
7096 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7102 * If there are RECLAIMABLE pages, we need to check
7103 * it. But now, memory offline itself doesn't call
7104 * shrink_node_slabs() and it still to be fixed.
7107 * If the page is not RAM, page_count()should be 0.
7108 * we don't need more check. This is an _used_ not-movable page.
7110 * The problematic thing here is PG_reserved pages. PG_reserved
7111 * is set to both of a memory hole page and a _used_ kernel
7120 bool is_pageblock_removable_nolock(struct page
*page
)
7126 * We have to be careful here because we are iterating over memory
7127 * sections which are not zone aware so we might end up outside of
7128 * the zone but still within the section.
7129 * We have to take care about the node as well. If the node is offline
7130 * its NODE_DATA will be NULL - see page_zone.
7132 if (!node_online(page_to_nid(page
)))
7135 zone
= page_zone(page
);
7136 pfn
= page_to_pfn(page
);
7137 if (!zone_spans_pfn(zone
, pfn
))
7140 return !has_unmovable_pages(zone
, page
, 0, true);
7143 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7145 static unsigned long pfn_max_align_down(unsigned long pfn
)
7147 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7148 pageblock_nr_pages
) - 1);
7151 static unsigned long pfn_max_align_up(unsigned long pfn
)
7153 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7154 pageblock_nr_pages
));
7157 /* [start, end) must belong to a single zone. */
7158 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7159 unsigned long start
, unsigned long end
)
7161 /* This function is based on compact_zone() from compaction.c. */
7162 unsigned long nr_reclaimed
;
7163 unsigned long pfn
= start
;
7164 unsigned int tries
= 0;
7169 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7170 if (fatal_signal_pending(current
)) {
7175 if (list_empty(&cc
->migratepages
)) {
7176 cc
->nr_migratepages
= 0;
7177 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7183 } else if (++tries
== 5) {
7184 ret
= ret
< 0 ? ret
: -EBUSY
;
7188 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7190 cc
->nr_migratepages
-= nr_reclaimed
;
7192 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7193 NULL
, 0, cc
->mode
, MR_CMA
);
7196 putback_movable_pages(&cc
->migratepages
);
7203 * alloc_contig_range() -- tries to allocate given range of pages
7204 * @start: start PFN to allocate
7205 * @end: one-past-the-last PFN to allocate
7206 * @migratetype: migratetype of the underlaying pageblocks (either
7207 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7208 * in range must have the same migratetype and it must
7209 * be either of the two.
7211 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7212 * aligned, however it's the caller's responsibility to guarantee that
7213 * we are the only thread that changes migrate type of pageblocks the
7216 * The PFN range must belong to a single zone.
7218 * Returns zero on success or negative error code. On success all
7219 * pages which PFN is in [start, end) are allocated for the caller and
7220 * need to be freed with free_contig_range().
7222 int alloc_contig_range(unsigned long start
, unsigned long end
,
7223 unsigned migratetype
)
7225 unsigned long outer_start
, outer_end
;
7229 struct compact_control cc
= {
7230 .nr_migratepages
= 0,
7232 .zone
= page_zone(pfn_to_page(start
)),
7233 .mode
= MIGRATE_SYNC
,
7234 .ignore_skip_hint
= true,
7236 INIT_LIST_HEAD(&cc
.migratepages
);
7239 * What we do here is we mark all pageblocks in range as
7240 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7241 * have different sizes, and due to the way page allocator
7242 * work, we align the range to biggest of the two pages so
7243 * that page allocator won't try to merge buddies from
7244 * different pageblocks and change MIGRATE_ISOLATE to some
7245 * other migration type.
7247 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7248 * migrate the pages from an unaligned range (ie. pages that
7249 * we are interested in). This will put all the pages in
7250 * range back to page allocator as MIGRATE_ISOLATE.
7252 * When this is done, we take the pages in range from page
7253 * allocator removing them from the buddy system. This way
7254 * page allocator will never consider using them.
7256 * This lets us mark the pageblocks back as
7257 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7258 * aligned range but not in the unaligned, original range are
7259 * put back to page allocator so that buddy can use them.
7262 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7263 pfn_max_align_up(end
), migratetype
,
7269 * In case of -EBUSY, we'd like to know which page causes problem.
7270 * So, just fall through. We will check it in test_pages_isolated().
7272 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7273 if (ret
&& ret
!= -EBUSY
)
7277 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7278 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7279 * more, all pages in [start, end) are free in page allocator.
7280 * What we are going to do is to allocate all pages from
7281 * [start, end) (that is remove them from page allocator).
7283 * The only problem is that pages at the beginning and at the
7284 * end of interesting range may be not aligned with pages that
7285 * page allocator holds, ie. they can be part of higher order
7286 * pages. Because of this, we reserve the bigger range and
7287 * once this is done free the pages we are not interested in.
7289 * We don't have to hold zone->lock here because the pages are
7290 * isolated thus they won't get removed from buddy.
7293 lru_add_drain_all();
7294 drain_all_pages(cc
.zone
);
7297 outer_start
= start
;
7298 while (!PageBuddy(pfn_to_page(outer_start
))) {
7299 if (++order
>= MAX_ORDER
) {
7300 outer_start
= start
;
7303 outer_start
&= ~0UL << order
;
7306 if (outer_start
!= start
) {
7307 order
= page_order(pfn_to_page(outer_start
));
7310 * outer_start page could be small order buddy page and
7311 * it doesn't include start page. Adjust outer_start
7312 * in this case to report failed page properly
7313 * on tracepoint in test_pages_isolated()
7315 if (outer_start
+ (1UL << order
) <= start
)
7316 outer_start
= start
;
7319 /* Make sure the range is really isolated. */
7320 if (test_pages_isolated(outer_start
, end
, false)) {
7321 pr_info("%s: [%lx, %lx) PFNs busy\n",
7322 __func__
, outer_start
, end
);
7327 /* Grab isolated pages from freelists. */
7328 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7334 /* Free head and tail (if any) */
7335 if (start
!= outer_start
)
7336 free_contig_range(outer_start
, start
- outer_start
);
7337 if (end
!= outer_end
)
7338 free_contig_range(end
, outer_end
- end
);
7341 undo_isolate_page_range(pfn_max_align_down(start
),
7342 pfn_max_align_up(end
), migratetype
);
7346 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7348 unsigned int count
= 0;
7350 for (; nr_pages
--; pfn
++) {
7351 struct page
*page
= pfn_to_page(pfn
);
7353 count
+= page_count(page
) != 1;
7356 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7360 #ifdef CONFIG_MEMORY_HOTPLUG
7362 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7363 * page high values need to be recalulated.
7365 void __meminit
zone_pcp_update(struct zone
*zone
)
7368 mutex_lock(&pcp_batch_high_lock
);
7369 for_each_possible_cpu(cpu
)
7370 pageset_set_high_and_batch(zone
,
7371 per_cpu_ptr(zone
->pageset
, cpu
));
7372 mutex_unlock(&pcp_batch_high_lock
);
7376 void zone_pcp_reset(struct zone
*zone
)
7378 unsigned long flags
;
7380 struct per_cpu_pageset
*pset
;
7382 /* avoid races with drain_pages() */
7383 local_irq_save(flags
);
7384 if (zone
->pageset
!= &boot_pageset
) {
7385 for_each_online_cpu(cpu
) {
7386 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7387 drain_zonestat(zone
, pset
);
7389 free_percpu(zone
->pageset
);
7390 zone
->pageset
= &boot_pageset
;
7392 local_irq_restore(flags
);
7395 #ifdef CONFIG_MEMORY_HOTREMOVE
7397 * All pages in the range must be in a single zone and isolated
7398 * before calling this.
7401 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7405 unsigned int order
, i
;
7407 unsigned long flags
;
7408 /* find the first valid pfn */
7409 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7414 zone
= page_zone(pfn_to_page(pfn
));
7415 spin_lock_irqsave(&zone
->lock
, flags
);
7417 while (pfn
< end_pfn
) {
7418 if (!pfn_valid(pfn
)) {
7422 page
= pfn_to_page(pfn
);
7424 * The HWPoisoned page may be not in buddy system, and
7425 * page_count() is not 0.
7427 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7429 SetPageReserved(page
);
7433 BUG_ON(page_count(page
));
7434 BUG_ON(!PageBuddy(page
));
7435 order
= page_order(page
);
7436 #ifdef CONFIG_DEBUG_VM
7437 pr_info("remove from free list %lx %d %lx\n",
7438 pfn
, 1 << order
, end_pfn
);
7440 list_del(&page
->lru
);
7441 rmv_page_order(page
);
7442 zone
->free_area
[order
].nr_free
--;
7443 for (i
= 0; i
< (1 << order
); i
++)
7444 SetPageReserved((page
+i
));
7445 pfn
+= (1 << order
);
7447 spin_unlock_irqrestore(&zone
->lock
, flags
);
7451 bool is_free_buddy_page(struct page
*page
)
7453 struct zone
*zone
= page_zone(page
);
7454 unsigned long pfn
= page_to_pfn(page
);
7455 unsigned long flags
;
7458 spin_lock_irqsave(&zone
->lock
, flags
);
7459 for (order
= 0; order
< MAX_ORDER
; order
++) {
7460 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7462 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7465 spin_unlock_irqrestore(&zone
->lock
, flags
);
7467 return order
< MAX_ORDER
;