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/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
78 static DEFINE_MUTEX(pcp_batch_high_lock
);
79 #define MIN_PERCPU_PAGELIST_FRACTION (8)
81 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
82 DEFINE_PER_CPU(int, numa_node
);
83 EXPORT_PER_CPU_SYMBOL(numa_node
);
86 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
88 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
89 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
90 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
91 * defined in <linux/topology.h>.
93 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
94 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
95 int _node_numa_mem_
[MAX_NUMNODES
];
98 /* work_structs for global per-cpu drains */
99 DEFINE_MUTEX(pcpu_drain_mutex
);
100 DEFINE_PER_CPU(struct work_struct
, pcpu_drain
);
102 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
103 volatile unsigned long latent_entropy __latent_entropy
;
104 EXPORT_SYMBOL(latent_entropy
);
108 * Array of node states.
110 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
111 [N_POSSIBLE
] = NODE_MASK_ALL
,
112 [N_ONLINE
] = { { [0] = 1UL } },
114 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
115 #ifdef CONFIG_HIGHMEM
116 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
118 [N_MEMORY
] = { { [0] = 1UL } },
119 [N_CPU
] = { { [0] = 1UL } },
122 EXPORT_SYMBOL(node_states
);
124 /* Protect totalram_pages and zone->managed_pages */
125 static DEFINE_SPINLOCK(managed_page_count_lock
);
127 unsigned long totalram_pages __read_mostly
;
128 unsigned long totalreserve_pages __read_mostly
;
129 unsigned long totalcma_pages __read_mostly
;
131 int percpu_pagelist_fraction
;
132 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
134 #ifdef CONFIG_PM_SLEEP
136 * The following functions are used by the suspend/hibernate code to temporarily
137 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
138 * while devices are suspended. To avoid races with the suspend/hibernate code,
139 * they should always be called with pm_mutex held (gfp_allowed_mask also should
140 * only be modified with pm_mutex held, unless the suspend/hibernate code is
141 * guaranteed not to run in parallel with that modification).
144 static gfp_t saved_gfp_mask
;
146 void pm_restore_gfp_mask(void)
148 WARN_ON(!mutex_is_locked(&pm_mutex
));
149 if (saved_gfp_mask
) {
150 gfp_allowed_mask
= saved_gfp_mask
;
155 void pm_restrict_gfp_mask(void)
157 WARN_ON(!mutex_is_locked(&pm_mutex
));
158 WARN_ON(saved_gfp_mask
);
159 saved_gfp_mask
= gfp_allowed_mask
;
160 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
163 bool pm_suspended_storage(void)
165 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
169 #endif /* CONFIG_PM_SLEEP */
171 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
172 unsigned int pageblock_order __read_mostly
;
175 static void __free_pages_ok(struct page
*page
, unsigned int order
);
178 * results with 256, 32 in the lowmem_reserve sysctl:
179 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
180 * 1G machine -> (16M dma, 784M normal, 224M high)
181 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
182 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
183 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
185 * TBD: should special case ZONE_DMA32 machines here - in those we normally
186 * don't need any ZONE_NORMAL reservation
188 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
189 #ifdef CONFIG_ZONE_DMA
192 #ifdef CONFIG_ZONE_DMA32
195 #ifdef CONFIG_HIGHMEM
201 EXPORT_SYMBOL(totalram_pages
);
203 static char * const zone_names
[MAX_NR_ZONES
] = {
204 #ifdef CONFIG_ZONE_DMA
207 #ifdef CONFIG_ZONE_DMA32
211 #ifdef CONFIG_HIGHMEM
215 #ifdef CONFIG_ZONE_DEVICE
220 char * const migratetype_names
[MIGRATE_TYPES
] = {
228 #ifdef CONFIG_MEMORY_ISOLATION
233 compound_page_dtor
* const compound_page_dtors
[] = {
236 #ifdef CONFIG_HUGETLB_PAGE
239 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
245 * Try to keep at least this much lowmem free. Do not allow normal
246 * allocations below this point, only high priority ones. Automatically
247 * tuned according to the amount of memory in the system.
249 int min_free_kbytes
= 1024;
250 int user_min_free_kbytes
= -1;
251 int watermark_scale_factor
= 10;
254 * Extra memory for the system to try freeing. Used to temporarily
255 * free memory, to make space for new workloads. Anyone can allocate
256 * down to the min watermarks controlled by min_free_kbytes above.
258 int extra_free_kbytes
= 0;
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
289 * Determine how many pages need to be initialized durig early boot
290 * (non-deferred initialization).
291 * The value of first_deferred_pfn will be set later, once non-deferred pages
292 * are initialized, but for now set it ULONG_MAX.
294 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
296 phys_addr_t start_addr
, end_addr
;
297 unsigned long max_pgcnt
;
298 unsigned long reserved
;
301 * Initialise at least 2G of a node but also take into account that
302 * two large system hashes that can take up 1GB for 0.25TB/node.
304 max_pgcnt
= max(2UL << (30 - PAGE_SHIFT
),
305 (pgdat
->node_spanned_pages
>> 8));
308 * Compensate the all the memblock reservations (e.g. crash kernel)
309 * from the initial estimation to make sure we will initialize enough
312 start_addr
= PFN_PHYS(pgdat
->node_start_pfn
);
313 end_addr
= PFN_PHYS(pgdat
->node_start_pfn
+ max_pgcnt
);
314 reserved
= memblock_reserved_memory_within(start_addr
, end_addr
);
315 max_pgcnt
+= PHYS_PFN(reserved
);
317 pgdat
->static_init_pgcnt
= min(max_pgcnt
, pgdat
->node_spanned_pages
);
318 pgdat
->first_deferred_pfn
= ULONG_MAX
;
321 /* Returns true if the struct page for the pfn is uninitialised */
322 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
324 int nid
= early_pfn_to_nid(pfn
);
326 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
333 * Returns false when the remaining initialisation should be deferred until
334 * later in the boot cycle when it can be parallelised.
336 static inline bool update_defer_init(pg_data_t
*pgdat
,
337 unsigned long pfn
, unsigned long zone_end
,
338 unsigned long *nr_initialised
)
340 /* Always populate low zones for address-contrained allocations */
341 if (zone_end
< pgdat_end_pfn(pgdat
))
344 if ((*nr_initialised
> pgdat
->static_init_pgcnt
) &&
345 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
346 pgdat
->first_deferred_pfn
= pfn
;
353 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
357 static inline bool early_page_uninitialised(unsigned long pfn
)
362 static inline bool update_defer_init(pg_data_t
*pgdat
,
363 unsigned long pfn
, unsigned long zone_end
,
364 unsigned long *nr_initialised
)
370 /* Return a pointer to the bitmap storing bits affecting a block of pages */
371 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
374 #ifdef CONFIG_SPARSEMEM
375 return __pfn_to_section(pfn
)->pageblock_flags
;
377 return page_zone(page
)->pageblock_flags
;
378 #endif /* CONFIG_SPARSEMEM */
381 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
383 #ifdef CONFIG_SPARSEMEM
384 pfn
&= (PAGES_PER_SECTION
-1);
385 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
387 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
388 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
389 #endif /* CONFIG_SPARSEMEM */
393 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
394 * @page: The page within the block of interest
395 * @pfn: The target page frame number
396 * @end_bitidx: The last bit of interest to retrieve
397 * @mask: mask of bits that the caller is interested in
399 * Return: pageblock_bits flags
401 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
403 unsigned long end_bitidx
,
406 unsigned long *bitmap
;
407 unsigned long bitidx
, word_bitidx
;
410 bitmap
= get_pageblock_bitmap(page
, pfn
);
411 bitidx
= pfn_to_bitidx(page
, pfn
);
412 word_bitidx
= bitidx
/ BITS_PER_LONG
;
413 bitidx
&= (BITS_PER_LONG
-1);
415 word
= bitmap
[word_bitidx
];
416 bitidx
+= end_bitidx
;
417 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
420 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
421 unsigned long end_bitidx
,
424 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
427 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
429 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
433 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
434 * @page: The page within the block of interest
435 * @flags: The flags to set
436 * @pfn: The target page frame number
437 * @end_bitidx: The last bit of interest
438 * @mask: mask of bits that the caller is interested in
440 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
442 unsigned long end_bitidx
,
445 unsigned long *bitmap
;
446 unsigned long bitidx
, word_bitidx
;
447 unsigned long old_word
, word
;
449 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
451 bitmap
= get_pageblock_bitmap(page
, pfn
);
452 bitidx
= pfn_to_bitidx(page
, pfn
);
453 word_bitidx
= bitidx
/ BITS_PER_LONG
;
454 bitidx
&= (BITS_PER_LONG
-1);
456 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
458 bitidx
+= end_bitidx
;
459 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
460 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
462 word
= READ_ONCE(bitmap
[word_bitidx
]);
464 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
465 if (word
== old_word
)
471 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
473 if (unlikely(page_group_by_mobility_disabled
&&
474 migratetype
< MIGRATE_PCPTYPES
))
475 migratetype
= MIGRATE_UNMOVABLE
;
477 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
478 PB_migrate
, PB_migrate_end
);
481 #ifdef CONFIG_DEBUG_VM
482 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
486 unsigned long pfn
= page_to_pfn(page
);
487 unsigned long sp
, start_pfn
;
490 seq
= zone_span_seqbegin(zone
);
491 start_pfn
= zone
->zone_start_pfn
;
492 sp
= zone
->spanned_pages
;
493 if (!zone_spans_pfn(zone
, pfn
))
495 } while (zone_span_seqretry(zone
, seq
));
498 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
499 pfn
, zone_to_nid(zone
), zone
->name
,
500 start_pfn
, start_pfn
+ sp
);
505 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
507 if (!pfn_valid_within(page_to_pfn(page
)))
509 if (zone
!= page_zone(page
))
515 * Temporary debugging check for pages not lying within a given zone.
517 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
519 if (page_outside_zone_boundaries(zone
, page
))
521 if (!page_is_consistent(zone
, page
))
527 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
533 static void bad_page(struct page
*page
, const char *reason
,
534 unsigned long bad_flags
)
536 static unsigned long resume
;
537 static unsigned long nr_shown
;
538 static unsigned long nr_unshown
;
541 * Allow a burst of 60 reports, then keep quiet for that minute;
542 * or allow a steady drip of one report per second.
544 if (nr_shown
== 60) {
545 if (time_before(jiffies
, resume
)) {
551 "BUG: Bad page state: %lu messages suppressed\n",
558 resume
= jiffies
+ 60 * HZ
;
560 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
561 current
->comm
, page_to_pfn(page
));
562 __dump_page(page
, reason
);
563 bad_flags
&= page
->flags
;
565 pr_alert("bad because of flags: %#lx(%pGp)\n",
566 bad_flags
, &bad_flags
);
567 dump_page_owner(page
);
572 /* Leave bad fields for debug, except PageBuddy could make trouble */
573 page_mapcount_reset(page
); /* remove PageBuddy */
574 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
578 * Higher-order pages are called "compound pages". They are structured thusly:
580 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
582 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
583 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
585 * The first tail page's ->compound_dtor holds the offset in array of compound
586 * page destructors. See compound_page_dtors.
588 * The first tail page's ->compound_order holds the order of allocation.
589 * This usage means that zero-order pages may not be compound.
592 void free_compound_page(struct page
*page
)
594 __free_pages_ok(page
, compound_order(page
));
597 void prep_compound_page(struct page
*page
, unsigned int order
)
600 int nr_pages
= 1 << order
;
602 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
603 set_compound_order(page
, order
);
605 for (i
= 1; i
< nr_pages
; i
++) {
606 struct page
*p
= page
+ i
;
607 set_page_count(p
, 0);
608 p
->mapping
= TAIL_MAPPING
;
609 set_compound_head(p
, page
);
611 atomic_set(compound_mapcount_ptr(page
), -1);
614 #ifdef CONFIG_DEBUG_PAGEALLOC
615 unsigned int _debug_guardpage_minorder
;
616 bool _debug_pagealloc_enabled __read_mostly
617 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
618 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
619 bool _debug_guardpage_enabled __read_mostly
;
621 static int __init
early_debug_pagealloc(char *buf
)
625 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
627 early_param("debug_pagealloc", early_debug_pagealloc
);
629 static bool need_debug_guardpage(void)
631 /* If we don't use debug_pagealloc, we don't need guard page */
632 if (!debug_pagealloc_enabled())
635 if (!debug_guardpage_minorder())
641 static void init_debug_guardpage(void)
643 if (!debug_pagealloc_enabled())
646 if (!debug_guardpage_minorder())
649 _debug_guardpage_enabled
= true;
652 struct page_ext_operations debug_guardpage_ops
= {
653 .need
= need_debug_guardpage
,
654 .init
= init_debug_guardpage
,
657 static int __init
debug_guardpage_minorder_setup(char *buf
)
661 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
662 pr_err("Bad debug_guardpage_minorder value\n");
665 _debug_guardpage_minorder
= res
;
666 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
669 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
671 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
672 unsigned int order
, int migratetype
)
674 struct page_ext
*page_ext
;
676 if (!debug_guardpage_enabled())
679 if (order
>= debug_guardpage_minorder())
682 page_ext
= lookup_page_ext(page
);
683 if (unlikely(!page_ext
))
686 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
688 INIT_LIST_HEAD(&page
->lru
);
689 set_page_private(page
, order
);
690 /* Guard pages are not available for any usage */
691 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
696 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
697 unsigned int order
, int migratetype
)
699 struct page_ext
*page_ext
;
701 if (!debug_guardpage_enabled())
704 page_ext
= lookup_page_ext(page
);
705 if (unlikely(!page_ext
))
708 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
710 set_page_private(page
, 0);
711 if (!is_migrate_isolate(migratetype
))
712 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
715 struct page_ext_operations debug_guardpage_ops
;
716 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
717 unsigned int order
, int migratetype
) { return false; }
718 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
719 unsigned int order
, int migratetype
) {}
722 static inline void set_page_order(struct page
*page
, unsigned int order
)
724 set_page_private(page
, order
);
725 __SetPageBuddy(page
);
728 static inline void rmv_page_order(struct page
*page
)
730 __ClearPageBuddy(page
);
731 set_page_private(page
, 0);
735 * This function checks whether a page is free && is the buddy
736 * we can do coalesce a page and its buddy if
737 * (a) the buddy is not in a hole (check before calling!) &&
738 * (b) the buddy is in the buddy system &&
739 * (c) a page and its buddy have the same order &&
740 * (d) a page and its buddy are in the same zone.
742 * For recording whether a page is in the buddy system, we set ->_mapcount
743 * PAGE_BUDDY_MAPCOUNT_VALUE.
744 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
745 * serialized by zone->lock.
747 * For recording page's order, we use page_private(page).
749 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
752 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
753 if (page_zone_id(page
) != page_zone_id(buddy
))
756 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
761 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
763 * zone check is done late to avoid uselessly
764 * calculating zone/node ids for pages that could
767 if (page_zone_id(page
) != page_zone_id(buddy
))
770 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
778 * Freeing function for a buddy system allocator.
780 * The concept of a buddy system is to maintain direct-mapped table
781 * (containing bit values) for memory blocks of various "orders".
782 * The bottom level table contains the map for the smallest allocatable
783 * units of memory (here, pages), and each level above it describes
784 * pairs of units from the levels below, hence, "buddies".
785 * At a high level, all that happens here is marking the table entry
786 * at the bottom level available, and propagating the changes upward
787 * as necessary, plus some accounting needed to play nicely with other
788 * parts of the VM system.
789 * At each level, we keep a list of pages, which are heads of continuous
790 * free pages of length of (1 << order) and marked with _mapcount
791 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
793 * So when we are allocating or freeing one, we can derive the state of the
794 * other. That is, if we allocate a small block, and both were
795 * free, the remainder of the region must be split into blocks.
796 * If a block is freed, and its buddy is also free, then this
797 * triggers coalescing into a block of larger size.
802 static inline void __free_one_page(struct page
*page
,
804 struct zone
*zone
, unsigned int order
,
807 unsigned long combined_pfn
;
808 unsigned long uninitialized_var(buddy_pfn
);
810 unsigned int max_order
;
812 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
814 VM_BUG_ON(!zone_is_initialized(zone
));
815 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
817 VM_BUG_ON(migratetype
== -1);
818 if (likely(!is_migrate_isolate(migratetype
)))
819 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
821 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
822 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
825 while (order
< max_order
- 1) {
826 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
827 buddy
= page
+ (buddy_pfn
- pfn
);
829 if (!pfn_valid_within(buddy_pfn
))
831 if (!page_is_buddy(page
, buddy
, order
))
834 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
835 * merge with it and move up one order.
837 if (page_is_guard(buddy
)) {
838 clear_page_guard(zone
, buddy
, order
, migratetype
);
840 list_del(&buddy
->lru
);
841 zone
->free_area
[order
].nr_free
--;
842 rmv_page_order(buddy
);
844 combined_pfn
= buddy_pfn
& pfn
;
845 page
= page
+ (combined_pfn
- pfn
);
849 if (max_order
< MAX_ORDER
) {
850 /* If we are here, it means order is >= pageblock_order.
851 * We want to prevent merge between freepages on isolate
852 * pageblock and normal pageblock. Without this, pageblock
853 * isolation could cause incorrect freepage or CMA accounting.
855 * We don't want to hit this code for the more frequent
858 if (unlikely(has_isolate_pageblock(zone
))) {
861 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
862 buddy
= page
+ (buddy_pfn
- pfn
);
863 buddy_mt
= get_pageblock_migratetype(buddy
);
865 if (migratetype
!= buddy_mt
866 && (is_migrate_isolate(migratetype
) ||
867 is_migrate_isolate(buddy_mt
)))
871 goto continue_merging
;
875 set_page_order(page
, order
);
878 * If this is not the largest possible page, check if the buddy
879 * of the next-highest order is free. If it is, it's possible
880 * that pages are being freed that will coalesce soon. In case,
881 * that is happening, add the free page to the tail of the list
882 * so it's less likely to be used soon and more likely to be merged
883 * as a higher order page
885 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
886 struct page
*higher_page
, *higher_buddy
;
887 combined_pfn
= buddy_pfn
& pfn
;
888 higher_page
= page
+ (combined_pfn
- pfn
);
889 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
890 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
891 if (pfn_valid_within(buddy_pfn
) &&
892 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
893 list_add_tail(&page
->lru
,
894 &zone
->free_area
[order
].free_list
[migratetype
]);
899 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
901 zone
->free_area
[order
].nr_free
++;
905 * A bad page could be due to a number of fields. Instead of multiple branches,
906 * try and check multiple fields with one check. The caller must do a detailed
907 * check if necessary.
909 static inline bool page_expected_state(struct page
*page
,
910 unsigned long check_flags
)
912 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
915 if (unlikely((unsigned long)page
->mapping
|
916 page_ref_count(page
) |
918 (unsigned long)page
->mem_cgroup
|
920 (page
->flags
& check_flags
)))
926 static void free_pages_check_bad(struct page
*page
)
928 const char *bad_reason
;
929 unsigned long bad_flags
;
934 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
935 bad_reason
= "nonzero mapcount";
936 if (unlikely(page
->mapping
!= NULL
))
937 bad_reason
= "non-NULL mapping";
938 if (unlikely(page_ref_count(page
) != 0))
939 bad_reason
= "nonzero _refcount";
940 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
941 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
942 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
945 if (unlikely(page
->mem_cgroup
))
946 bad_reason
= "page still charged to cgroup";
948 bad_page(page
, bad_reason
, bad_flags
);
951 static inline int free_pages_check(struct page
*page
)
953 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
956 /* Something has gone sideways, find it */
957 free_pages_check_bad(page
);
961 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
966 * We rely page->lru.next never has bit 0 set, unless the page
967 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
969 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
971 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
975 switch (page
- head_page
) {
977 /* the first tail page: ->mapping is compound_mapcount() */
978 if (unlikely(compound_mapcount(page
))) {
979 bad_page(page
, "nonzero compound_mapcount", 0);
985 * the second tail page: ->mapping is
986 * page_deferred_list().next -- ignore value.
990 if (page
->mapping
!= TAIL_MAPPING
) {
991 bad_page(page
, "corrupted mapping in tail page", 0);
996 if (unlikely(!PageTail(page
))) {
997 bad_page(page
, "PageTail not set", 0);
1000 if (unlikely(compound_head(page
) != head_page
)) {
1001 bad_page(page
, "compound_head not consistent", 0);
1006 page
->mapping
= NULL
;
1007 clear_compound_head(page
);
1011 static __always_inline
bool free_pages_prepare(struct page
*page
,
1012 unsigned int order
, bool check_free
)
1016 VM_BUG_ON_PAGE(PageTail(page
), page
);
1018 trace_mm_page_free(page
, order
);
1021 * Check tail pages before head page information is cleared to
1022 * avoid checking PageCompound for order-0 pages.
1024 if (unlikely(order
)) {
1025 bool compound
= PageCompound(page
);
1028 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1031 ClearPageDoubleMap(page
);
1032 for (i
= 1; i
< (1 << order
); i
++) {
1034 bad
+= free_tail_pages_check(page
, page
+ i
);
1035 if (unlikely(free_pages_check(page
+ i
))) {
1039 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1042 if (PageMappingFlags(page
))
1043 page
->mapping
= NULL
;
1044 if (memcg_kmem_enabled() && PageKmemcg(page
))
1045 memcg_kmem_uncharge(page
, order
);
1047 bad
+= free_pages_check(page
);
1051 page_cpupid_reset_last(page
);
1052 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1053 reset_page_owner(page
, order
);
1055 if (!PageHighMem(page
)) {
1056 debug_check_no_locks_freed(page_address(page
),
1057 PAGE_SIZE
<< order
);
1058 debug_check_no_obj_freed(page_address(page
),
1059 PAGE_SIZE
<< order
);
1061 arch_free_page(page
, order
);
1062 kernel_poison_pages(page
, 1 << order
, 0);
1063 kernel_map_pages(page
, 1 << order
, 0);
1064 kasan_free_pages(page
, order
);
1069 #ifdef CONFIG_DEBUG_VM
1070 static inline bool free_pcp_prepare(struct page
*page
)
1072 return free_pages_prepare(page
, 0, true);
1075 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1080 static bool free_pcp_prepare(struct page
*page
)
1082 return free_pages_prepare(page
, 0, false);
1085 static bool bulkfree_pcp_prepare(struct page
*page
)
1087 return free_pages_check(page
);
1089 #endif /* CONFIG_DEBUG_VM */
1092 * Frees a number of pages from the PCP lists
1093 * Assumes all pages on list are in same zone, and of same order.
1094 * count is the number of pages to free.
1096 * If the zone was previously in an "all pages pinned" state then look to
1097 * see if this freeing clears that state.
1099 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1100 * pinned" detection logic.
1102 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1103 struct per_cpu_pages
*pcp
)
1105 int migratetype
= 0;
1107 bool isolated_pageblocks
;
1109 spin_lock(&zone
->lock
);
1110 isolated_pageblocks
= has_isolate_pageblock(zone
);
1114 struct list_head
*list
;
1117 * Remove pages from lists in a round-robin fashion. A
1118 * batch_free count is maintained that is incremented when an
1119 * empty list is encountered. This is so more pages are freed
1120 * off fuller lists instead of spinning excessively around empty
1125 if (++migratetype
== MIGRATE_PCPTYPES
)
1127 list
= &pcp
->lists
[migratetype
];
1128 } while (list_empty(list
));
1130 /* This is the only non-empty list. Free them all. */
1131 if (batch_free
== MIGRATE_PCPTYPES
)
1135 int mt
; /* migratetype of the to-be-freed page */
1137 page
= list_last_entry(list
, struct page
, lru
);
1138 /* must delete as __free_one_page list manipulates */
1139 list_del(&page
->lru
);
1141 mt
= get_pcppage_migratetype(page
);
1142 /* MIGRATE_ISOLATE page should not go to pcplists */
1143 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1144 /* Pageblock could have been isolated meanwhile */
1145 if (unlikely(isolated_pageblocks
))
1146 mt
= get_pageblock_migratetype(page
);
1148 if (bulkfree_pcp_prepare(page
))
1151 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1152 trace_mm_page_pcpu_drain(page
, 0, mt
);
1153 } while (--count
&& --batch_free
&& !list_empty(list
));
1155 spin_unlock(&zone
->lock
);
1158 static void free_one_page(struct zone
*zone
,
1159 struct page
*page
, unsigned long pfn
,
1163 spin_lock(&zone
->lock
);
1164 if (unlikely(has_isolate_pageblock(zone
) ||
1165 is_migrate_isolate(migratetype
))) {
1166 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1168 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1169 spin_unlock(&zone
->lock
);
1172 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1173 unsigned long zone
, int nid
)
1175 set_page_links(page
, zone
, nid
, pfn
);
1176 init_page_count(page
);
1177 page_mapcount_reset(page
);
1178 page_cpupid_reset_last(page
);
1180 INIT_LIST_HEAD(&page
->lru
);
1181 #ifdef WANT_PAGE_VIRTUAL
1182 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1183 if (!is_highmem_idx(zone
))
1184 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1188 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1191 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1194 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1195 static void __meminit
init_reserved_page(unsigned long pfn
)
1200 if (!early_page_uninitialised(pfn
))
1203 nid
= early_pfn_to_nid(pfn
);
1204 pgdat
= NODE_DATA(nid
);
1206 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1207 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1209 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1212 __init_single_pfn(pfn
, zid
, nid
);
1215 static inline void init_reserved_page(unsigned long pfn
)
1218 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1221 * Initialised pages do not have PageReserved set. This function is
1222 * called for each range allocated by the bootmem allocator and
1223 * marks the pages PageReserved. The remaining valid pages are later
1224 * sent to the buddy page allocator.
1226 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1228 unsigned long start_pfn
= PFN_DOWN(start
);
1229 unsigned long end_pfn
= PFN_UP(end
);
1231 for (; start_pfn
< end_pfn
; start_pfn
++) {
1232 if (pfn_valid(start_pfn
)) {
1233 struct page
*page
= pfn_to_page(start_pfn
);
1235 init_reserved_page(start_pfn
);
1237 /* Avoid false-positive PageTail() */
1238 INIT_LIST_HEAD(&page
->lru
);
1240 SetPageReserved(page
);
1245 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1247 unsigned long flags
;
1249 unsigned long pfn
= page_to_pfn(page
);
1251 if (!free_pages_prepare(page
, order
, true))
1254 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1255 local_irq_save(flags
);
1256 __count_vm_events(PGFREE
, 1 << order
);
1257 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1258 local_irq_restore(flags
);
1261 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1263 unsigned int nr_pages
= 1 << order
;
1264 struct page
*p
= page
;
1268 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1270 __ClearPageReserved(p
);
1271 set_page_count(p
, 0);
1273 __ClearPageReserved(p
);
1274 set_page_count(p
, 0);
1276 page_zone(page
)->managed_pages
+= nr_pages
;
1277 set_page_refcounted(page
);
1278 __free_pages(page
, order
);
1281 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1282 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1284 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1286 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1288 static DEFINE_SPINLOCK(early_pfn_lock
);
1291 spin_lock(&early_pfn_lock
);
1292 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1294 nid
= first_online_node
;
1295 spin_unlock(&early_pfn_lock
);
1301 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1302 static inline bool __meminit __maybe_unused
1303 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1304 struct mminit_pfnnid_cache
*state
)
1308 nid
= __early_pfn_to_nid(pfn
, state
);
1309 if (nid
>= 0 && nid
!= node
)
1314 /* Only safe to use early in boot when initialisation is single-threaded */
1315 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1317 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1322 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1326 static inline bool __meminit __maybe_unused
1327 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1328 struct mminit_pfnnid_cache
*state
)
1335 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1338 if (early_page_uninitialised(pfn
))
1340 return __free_pages_boot_core(page
, order
);
1344 * Check that the whole (or subset of) a pageblock given by the interval of
1345 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1346 * with the migration of free compaction scanner. The scanners then need to
1347 * use only pfn_valid_within() check for arches that allow holes within
1350 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1352 * It's possible on some configurations to have a setup like node0 node1 node0
1353 * i.e. it's possible that all pages within a zones range of pages do not
1354 * belong to a single zone. We assume that a border between node0 and node1
1355 * can occur within a single pageblock, but not a node0 node1 node0
1356 * interleaving within a single pageblock. It is therefore sufficient to check
1357 * the first and last page of a pageblock and avoid checking each individual
1358 * page in a pageblock.
1360 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1361 unsigned long end_pfn
, struct zone
*zone
)
1363 struct page
*start_page
;
1364 struct page
*end_page
;
1366 /* end_pfn is one past the range we are checking */
1369 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1372 start_page
= pfn_to_online_page(start_pfn
);
1376 if (page_zone(start_page
) != zone
)
1379 end_page
= pfn_to_page(end_pfn
);
1381 /* This gives a shorter code than deriving page_zone(end_page) */
1382 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1388 void set_zone_contiguous(struct zone
*zone
)
1390 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1391 unsigned long block_end_pfn
;
1393 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1394 for (; block_start_pfn
< zone_end_pfn(zone
);
1395 block_start_pfn
= block_end_pfn
,
1396 block_end_pfn
+= pageblock_nr_pages
) {
1398 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1400 if (!__pageblock_pfn_to_page(block_start_pfn
,
1401 block_end_pfn
, zone
))
1405 /* We confirm that there is no hole */
1406 zone
->contiguous
= true;
1409 void clear_zone_contiguous(struct zone
*zone
)
1411 zone
->contiguous
= false;
1414 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1415 static void __init
deferred_free_range(struct page
*page
,
1416 unsigned long pfn
, int nr_pages
)
1423 /* Free a large naturally-aligned chunk if possible */
1424 if (nr_pages
== pageblock_nr_pages
&&
1425 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1426 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1427 __free_pages_boot_core(page
, pageblock_order
);
1431 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1432 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1433 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1434 __free_pages_boot_core(page
, 0);
1438 /* Completion tracking for deferred_init_memmap() threads */
1439 static atomic_t pgdat_init_n_undone __initdata
;
1440 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1442 static inline void __init
pgdat_init_report_one_done(void)
1444 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1445 complete(&pgdat_init_all_done_comp
);
1448 /* Initialise remaining memory on a node */
1449 static int __init
deferred_init_memmap(void *data
)
1451 pg_data_t
*pgdat
= data
;
1452 int nid
= pgdat
->node_id
;
1453 struct mminit_pfnnid_cache nid_init_state
= { };
1454 unsigned long start
= jiffies
;
1455 unsigned long nr_pages
= 0;
1456 unsigned long walk_start
, walk_end
;
1459 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1460 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1462 if (first_init_pfn
== ULONG_MAX
) {
1463 pgdat_init_report_one_done();
1467 /* Bind memory initialisation thread to a local node if possible */
1468 if (!cpumask_empty(cpumask
))
1469 set_cpus_allowed_ptr(current
, cpumask
);
1471 /* Sanity check boundaries */
1472 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1473 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1474 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1476 /* Only the highest zone is deferred so find it */
1477 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1478 zone
= pgdat
->node_zones
+ zid
;
1479 if (first_init_pfn
< zone_end_pfn(zone
))
1483 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1484 unsigned long pfn
, end_pfn
;
1485 struct page
*page
= NULL
;
1486 struct page
*free_base_page
= NULL
;
1487 unsigned long free_base_pfn
= 0;
1490 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1491 pfn
= first_init_pfn
;
1492 if (pfn
< walk_start
)
1494 if (pfn
< zone
->zone_start_pfn
)
1495 pfn
= zone
->zone_start_pfn
;
1497 for (; pfn
< end_pfn
; pfn
++) {
1498 if (!pfn_valid_within(pfn
))
1502 * Ensure pfn_valid is checked every
1503 * pageblock_nr_pages for memory holes
1505 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1506 if (!pfn_valid(pfn
)) {
1512 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1517 /* Minimise pfn page lookups and scheduler checks */
1518 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1521 nr_pages
+= nr_to_free
;
1522 deferred_free_range(free_base_page
,
1523 free_base_pfn
, nr_to_free
);
1524 free_base_page
= NULL
;
1525 free_base_pfn
= nr_to_free
= 0;
1527 page
= pfn_to_page(pfn
);
1532 VM_BUG_ON(page_zone(page
) != zone
);
1536 __init_single_page(page
, pfn
, zid
, nid
);
1537 if (!free_base_page
) {
1538 free_base_page
= page
;
1539 free_base_pfn
= pfn
;
1544 /* Where possible, batch up pages for a single free */
1547 /* Free the current block of pages to allocator */
1548 nr_pages
+= nr_to_free
;
1549 deferred_free_range(free_base_page
, free_base_pfn
,
1551 free_base_page
= NULL
;
1552 free_base_pfn
= nr_to_free
= 0;
1554 /* Free the last block of pages to allocator */
1555 nr_pages
+= nr_to_free
;
1556 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1558 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1561 /* Sanity check that the next zone really is unpopulated */
1562 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1564 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1565 jiffies_to_msecs(jiffies
- start
));
1567 pgdat_init_report_one_done();
1570 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1572 void __init
page_alloc_init_late(void)
1576 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1579 /* There will be num_node_state(N_MEMORY) threads */
1580 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1581 for_each_node_state(nid
, N_MEMORY
) {
1582 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1585 /* Block until all are initialised */
1586 wait_for_completion(&pgdat_init_all_done_comp
);
1588 /* Reinit limits that are based on free pages after the kernel is up */
1589 files_maxfiles_init();
1591 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1592 /* Discard memblock private memory */
1596 for_each_populated_zone(zone
)
1597 set_zone_contiguous(zone
);
1601 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1602 void __init
init_cma_reserved_pageblock(struct page
*page
)
1604 unsigned i
= pageblock_nr_pages
;
1605 struct page
*p
= page
;
1608 __ClearPageReserved(p
);
1609 set_page_count(p
, 0);
1612 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1614 if (pageblock_order
>= MAX_ORDER
) {
1615 i
= pageblock_nr_pages
;
1618 set_page_refcounted(p
);
1619 __free_pages(p
, MAX_ORDER
- 1);
1620 p
+= MAX_ORDER_NR_PAGES
;
1621 } while (i
-= MAX_ORDER_NR_PAGES
);
1623 set_page_refcounted(page
);
1624 __free_pages(page
, pageblock_order
);
1627 adjust_managed_page_count(page
, pageblock_nr_pages
);
1632 * The order of subdivision here is critical for the IO subsystem.
1633 * Please do not alter this order without good reasons and regression
1634 * testing. Specifically, as large blocks of memory are subdivided,
1635 * the order in which smaller blocks are delivered depends on the order
1636 * they're subdivided in this function. This is the primary factor
1637 * influencing the order in which pages are delivered to the IO
1638 * subsystem according to empirical testing, and this is also justified
1639 * by considering the behavior of a buddy system containing a single
1640 * large block of memory acted on by a series of small allocations.
1641 * This behavior is a critical factor in sglist merging's success.
1645 static inline void expand(struct zone
*zone
, struct page
*page
,
1646 int low
, int high
, struct free_area
*area
,
1649 unsigned long size
= 1 << high
;
1651 while (high
> low
) {
1655 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1658 * Mark as guard pages (or page), that will allow to
1659 * merge back to allocator when buddy will be freed.
1660 * Corresponding page table entries will not be touched,
1661 * pages will stay not present in virtual address space
1663 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1666 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1668 set_page_order(&page
[size
], high
);
1672 static void check_new_page_bad(struct page
*page
)
1674 const char *bad_reason
= NULL
;
1675 unsigned long bad_flags
= 0;
1677 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1678 bad_reason
= "nonzero mapcount";
1679 if (unlikely(page
->mapping
!= NULL
))
1680 bad_reason
= "non-NULL mapping";
1681 if (unlikely(page_ref_count(page
) != 0))
1682 bad_reason
= "nonzero _count";
1683 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1684 bad_reason
= "HWPoisoned (hardware-corrupted)";
1685 bad_flags
= __PG_HWPOISON
;
1686 /* Don't complain about hwpoisoned pages */
1687 page_mapcount_reset(page
); /* remove PageBuddy */
1690 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1691 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1692 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1695 if (unlikely(page
->mem_cgroup
))
1696 bad_reason
= "page still charged to cgroup";
1698 bad_page(page
, bad_reason
, bad_flags
);
1702 * This page is about to be returned from the page allocator
1704 static inline int check_new_page(struct page
*page
)
1706 if (likely(page_expected_state(page
,
1707 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1710 check_new_page_bad(page
);
1714 static inline bool free_pages_prezeroed(void)
1716 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1717 page_poisoning_enabled();
1720 #ifdef CONFIG_DEBUG_VM
1721 static bool check_pcp_refill(struct page
*page
)
1726 static bool check_new_pcp(struct page
*page
)
1728 return check_new_page(page
);
1731 static bool check_pcp_refill(struct page
*page
)
1733 return check_new_page(page
);
1735 static bool check_new_pcp(struct page
*page
)
1739 #endif /* CONFIG_DEBUG_VM */
1741 static bool check_new_pages(struct page
*page
, unsigned int order
)
1744 for (i
= 0; i
< (1 << order
); i
++) {
1745 struct page
*p
= page
+ i
;
1747 if (unlikely(check_new_page(p
)))
1754 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1757 set_page_private(page
, 0);
1758 set_page_refcounted(page
);
1760 arch_alloc_page(page
, order
);
1761 kernel_map_pages(page
, 1 << order
, 1);
1762 kasan_alloc_pages(page
, order
);
1763 kernel_poison_pages(page
, 1 << order
, 1);
1764 set_page_owner(page
, order
, gfp_flags
);
1767 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1768 unsigned int alloc_flags
)
1772 post_alloc_hook(page
, order
, gfp_flags
);
1774 if (!free_pages_prezeroed() && (gfp_flags
& __GFP_ZERO
))
1775 for (i
= 0; i
< (1 << order
); i
++)
1776 clear_highpage(page
+ i
);
1778 if (order
&& (gfp_flags
& __GFP_COMP
))
1779 prep_compound_page(page
, order
);
1782 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1783 * allocate the page. The expectation is that the caller is taking
1784 * steps that will free more memory. The caller should avoid the page
1785 * being used for !PFMEMALLOC purposes.
1787 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1788 set_page_pfmemalloc(page
);
1790 clear_page_pfmemalloc(page
);
1794 * Go through the free lists for the given migratetype and remove
1795 * the smallest available page from the freelists
1798 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1801 unsigned int current_order
;
1802 struct free_area
*area
;
1805 /* Find a page of the appropriate size in the preferred list */
1806 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1807 area
= &(zone
->free_area
[current_order
]);
1808 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1812 list_del(&page
->lru
);
1813 rmv_page_order(page
);
1815 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1816 set_pcppage_migratetype(page
, migratetype
);
1825 * This array describes the order lists are fallen back to when
1826 * the free lists for the desirable migrate type are depleted
1828 static int fallbacks
[MIGRATE_TYPES
][4] = {
1829 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1830 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1831 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1833 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1835 #ifdef CONFIG_MEMORY_ISOLATION
1836 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1841 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1844 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1847 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1848 unsigned int order
) { return NULL
; }
1852 * Move the free pages in a range to the free lists of the requested type.
1853 * Note that start_page and end_pages are not aligned on a pageblock
1854 * boundary. If alignment is required, use move_freepages_block()
1856 static int move_freepages(struct zone
*zone
,
1857 struct page
*start_page
, struct page
*end_page
,
1858 int migratetype
, int *num_movable
)
1862 int pages_moved
= 0;
1864 #ifndef CONFIG_HOLES_IN_ZONE
1866 * page_zone is not safe to call in this context when
1867 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1868 * anyway as we check zone boundaries in move_freepages_block().
1869 * Remove at a later date when no bug reports exist related to
1870 * grouping pages by mobility
1872 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1878 for (page
= start_page
; page
<= end_page
;) {
1879 if (!pfn_valid_within(page_to_pfn(page
))) {
1884 /* Make sure we are not inadvertently changing nodes */
1885 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1887 if (!PageBuddy(page
)) {
1889 * We assume that pages that could be isolated for
1890 * migration are movable. But we don't actually try
1891 * isolating, as that would be expensive.
1894 (PageLRU(page
) || __PageMovable(page
)))
1901 order
= page_order(page
);
1902 list_move(&page
->lru
,
1903 &zone
->free_area
[order
].free_list
[migratetype
]);
1905 pages_moved
+= 1 << order
;
1911 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1912 int migratetype
, int *num_movable
)
1914 unsigned long start_pfn
, end_pfn
;
1915 struct page
*start_page
, *end_page
;
1917 start_pfn
= page_to_pfn(page
);
1918 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1919 start_page
= pfn_to_page(start_pfn
);
1920 end_page
= start_page
+ pageblock_nr_pages
- 1;
1921 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1923 /* Do not cross zone boundaries */
1924 if (!zone_spans_pfn(zone
, start_pfn
))
1926 if (!zone_spans_pfn(zone
, end_pfn
))
1929 return move_freepages(zone
, start_page
, end_page
, migratetype
,
1933 static void change_pageblock_range(struct page
*pageblock_page
,
1934 int start_order
, int migratetype
)
1936 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1938 while (nr_pageblocks
--) {
1939 set_pageblock_migratetype(pageblock_page
, migratetype
);
1940 pageblock_page
+= pageblock_nr_pages
;
1945 * When we are falling back to another migratetype during allocation, try to
1946 * steal extra free pages from the same pageblocks to satisfy further
1947 * allocations, instead of polluting multiple pageblocks.
1949 * If we are stealing a relatively large buddy page, it is likely there will
1950 * be more free pages in the pageblock, so try to steal them all. For
1951 * reclaimable and unmovable allocations, we steal regardless of page size,
1952 * as fragmentation caused by those allocations polluting movable pageblocks
1953 * is worse than movable allocations stealing from unmovable and reclaimable
1956 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1959 * Leaving this order check is intended, although there is
1960 * relaxed order check in next check. The reason is that
1961 * we can actually steal whole pageblock if this condition met,
1962 * but, below check doesn't guarantee it and that is just heuristic
1963 * so could be changed anytime.
1965 if (order
>= pageblock_order
)
1968 if (order
>= pageblock_order
/ 2 ||
1969 start_mt
== MIGRATE_RECLAIMABLE
||
1970 start_mt
== MIGRATE_UNMOVABLE
||
1971 page_group_by_mobility_disabled
)
1978 * This function implements actual steal behaviour. If order is large enough,
1979 * we can steal whole pageblock. If not, we first move freepages in this
1980 * pageblock to our migratetype and determine how many already-allocated pages
1981 * are there in the pageblock with a compatible migratetype. If at least half
1982 * of pages are free or compatible, we can change migratetype of the pageblock
1983 * itself, so pages freed in the future will be put on the correct free list.
1985 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1986 int start_type
, bool whole_block
)
1988 unsigned int current_order
= page_order(page
);
1989 struct free_area
*area
;
1990 int free_pages
, movable_pages
, alike_pages
;
1993 old_block_type
= get_pageblock_migratetype(page
);
1996 * This can happen due to races and we want to prevent broken
1997 * highatomic accounting.
1999 if (is_migrate_highatomic(old_block_type
))
2002 /* Take ownership for orders >= pageblock_order */
2003 if (current_order
>= pageblock_order
) {
2004 change_pageblock_range(page
, current_order
, start_type
);
2008 /* We are not allowed to try stealing from the whole block */
2012 free_pages
= move_freepages_block(zone
, page
, start_type
,
2015 * Determine how many pages are compatible with our allocation.
2016 * For movable allocation, it's the number of movable pages which
2017 * we just obtained. For other types it's a bit more tricky.
2019 if (start_type
== MIGRATE_MOVABLE
) {
2020 alike_pages
= movable_pages
;
2023 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2024 * to MOVABLE pageblock, consider all non-movable pages as
2025 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2026 * vice versa, be conservative since we can't distinguish the
2027 * exact migratetype of non-movable pages.
2029 if (old_block_type
== MIGRATE_MOVABLE
)
2030 alike_pages
= pageblock_nr_pages
2031 - (free_pages
+ movable_pages
);
2036 /* moving whole block can fail due to zone boundary conditions */
2041 * If a sufficient number of pages in the block are either free or of
2042 * comparable migratability as our allocation, claim the whole block.
2044 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2045 page_group_by_mobility_disabled
)
2046 set_pageblock_migratetype(page
, start_type
);
2051 area
= &zone
->free_area
[current_order
];
2052 list_move(&page
->lru
, &area
->free_list
[start_type
]);
2056 * Check whether there is a suitable fallback freepage with requested order.
2057 * If only_stealable is true, this function returns fallback_mt only if
2058 * we can steal other freepages all together. This would help to reduce
2059 * fragmentation due to mixed migratetype pages in one pageblock.
2061 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2062 int migratetype
, bool only_stealable
, bool *can_steal
)
2067 if (area
->nr_free
== 0)
2072 fallback_mt
= fallbacks
[migratetype
][i
];
2073 if (fallback_mt
== MIGRATE_TYPES
)
2076 if (list_empty(&area
->free_list
[fallback_mt
]))
2079 if (can_steal_fallback(order
, migratetype
))
2082 if (!only_stealable
)
2093 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2094 * there are no empty page blocks that contain a page with a suitable order
2096 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2097 unsigned int alloc_order
)
2100 unsigned long max_managed
, flags
;
2103 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2104 * Check is race-prone but harmless.
2106 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2107 if (zone
->nr_reserved_highatomic
>= max_managed
)
2110 spin_lock_irqsave(&zone
->lock
, flags
);
2112 /* Recheck the nr_reserved_highatomic limit under the lock */
2113 if (zone
->nr_reserved_highatomic
>= max_managed
)
2117 mt
= get_pageblock_migratetype(page
);
2118 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2119 && !is_migrate_cma(mt
)) {
2120 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2121 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2122 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2126 spin_unlock_irqrestore(&zone
->lock
, flags
);
2130 * Used when an allocation is about to fail under memory pressure. This
2131 * potentially hurts the reliability of high-order allocations when under
2132 * intense memory pressure but failed atomic allocations should be easier
2133 * to recover from than an OOM.
2135 * If @force is true, try to unreserve a pageblock even though highatomic
2136 * pageblock is exhausted.
2138 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2141 struct zonelist
*zonelist
= ac
->zonelist
;
2142 unsigned long flags
;
2149 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2152 * Preserve at least one pageblock unless memory pressure
2155 if (!force
&& zone
->nr_reserved_highatomic
<=
2159 spin_lock_irqsave(&zone
->lock
, flags
);
2160 for (order
= 0; order
< MAX_ORDER
; order
++) {
2161 struct free_area
*area
= &(zone
->free_area
[order
]);
2163 page
= list_first_entry_or_null(
2164 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2170 * In page freeing path, migratetype change is racy so
2171 * we can counter several free pages in a pageblock
2172 * in this loop althoug we changed the pageblock type
2173 * from highatomic to ac->migratetype. So we should
2174 * adjust the count once.
2176 if (is_migrate_highatomic_page(page
)) {
2178 * It should never happen but changes to
2179 * locking could inadvertently allow a per-cpu
2180 * drain to add pages to MIGRATE_HIGHATOMIC
2181 * while unreserving so be safe and watch for
2184 zone
->nr_reserved_highatomic
-= min(
2186 zone
->nr_reserved_highatomic
);
2190 * Convert to ac->migratetype and avoid the normal
2191 * pageblock stealing heuristics. Minimally, the caller
2192 * is doing the work and needs the pages. More
2193 * importantly, if the block was always converted to
2194 * MIGRATE_UNMOVABLE or another type then the number
2195 * of pageblocks that cannot be completely freed
2198 set_pageblock_migratetype(page
, ac
->migratetype
);
2199 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2202 spin_unlock_irqrestore(&zone
->lock
, flags
);
2206 spin_unlock_irqrestore(&zone
->lock
, flags
);
2213 * Try finding a free buddy page on the fallback list and put it on the free
2214 * list of requested migratetype, possibly along with other pages from the same
2215 * block, depending on fragmentation avoidance heuristics. Returns true if
2216 * fallback was found so that __rmqueue_smallest() can grab it.
2218 * The use of signed ints for order and current_order is a deliberate
2219 * deviation from the rest of this file, to make the for loop
2220 * condition simpler.
2223 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
2225 struct free_area
*area
;
2232 * Find the largest available free page in the other list. This roughly
2233 * approximates finding the pageblock with the most free pages, which
2234 * would be too costly to do exactly.
2236 for (current_order
= MAX_ORDER
- 1; current_order
>= order
;
2238 area
= &(zone
->free_area
[current_order
]);
2239 fallback_mt
= find_suitable_fallback(area
, current_order
,
2240 start_migratetype
, false, &can_steal
);
2241 if (fallback_mt
== -1)
2245 * We cannot steal all free pages from the pageblock and the
2246 * requested migratetype is movable. In that case it's better to
2247 * steal and split the smallest available page instead of the
2248 * largest available page, because even if the next movable
2249 * allocation falls back into a different pageblock than this
2250 * one, it won't cause permanent fragmentation.
2252 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2253 && current_order
> order
)
2262 for (current_order
= order
; current_order
< MAX_ORDER
;
2264 area
= &(zone
->free_area
[current_order
]);
2265 fallback_mt
= find_suitable_fallback(area
, current_order
,
2266 start_migratetype
, false, &can_steal
);
2267 if (fallback_mt
!= -1)
2272 * This should not happen - we already found a suitable fallback
2273 * when looking for the largest page.
2275 VM_BUG_ON(current_order
== MAX_ORDER
);
2278 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2281 steal_suitable_fallback(zone
, page
, start_migratetype
, can_steal
);
2283 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2284 start_migratetype
, fallback_mt
);
2291 * Do the hard work of removing an element from the buddy allocator.
2292 * Call me with the zone->lock already held.
2293 * If gfp mask of the page allocation has GFP_HIGHUSER_MOVABLE, @migratetype
2294 * is changed from MIGRATE_MOVABLE to MIGRATE_CMA in rmqueue() to select the
2295 * free list of MIGRATE_CMA. It helps depleting CMA free pages so that
2296 * evaluation of watermark for unmovable page allocations is not too different
2297 * from movable page allocations.
2298 * If @migratetype is MIGRATE_CMA, it should be corrected to MIGRATE_MOVABLE
2299 * after the free list of MIGRATE_CMA is searched to have a chance to search the
2300 * free list of MIGRATE_MOVABLE. It also records correct migrate type in the
2301 * trace as intended by the page allocation.
2303 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2306 struct page
*page
= NULL
;
2309 if (migratetype
== MIGRATE_CMA
) {
2311 if (migratetype
== MIGRATE_MOVABLE
) {
2313 page
= __rmqueue_cma_fallback(zone
, order
);
2314 migratetype
= MIGRATE_MOVABLE
;
2318 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2319 if (unlikely(!page
) &&
2320 !__rmqueue_fallback(zone
, order
, migratetype
))
2324 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2329 * Obtain a specified number of elements from the buddy allocator, all under
2330 * a single hold of the lock, for efficiency. Add them to the supplied list.
2331 * Returns the number of new pages which were placed at *list.
2333 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2334 unsigned long count
, struct list_head
*list
,
2335 int migratetype
, bool cold
)
2339 spin_lock(&zone
->lock
);
2340 for (i
= 0; i
< count
; ++i
) {
2341 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2342 if (unlikely(page
== NULL
))
2345 if (unlikely(check_pcp_refill(page
)))
2349 * Split buddy pages returned by expand() are received here
2350 * in physical page order. The page is added to the callers and
2351 * list and the list head then moves forward. From the callers
2352 * perspective, the linked list is ordered by page number in
2353 * some conditions. This is useful for IO devices that can
2354 * merge IO requests if the physical pages are ordered
2358 list_add(&page
->lru
, list
);
2360 list_add_tail(&page
->lru
, list
);
2363 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2364 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2369 * i pages were removed from the buddy list even if some leak due
2370 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2371 * on i. Do not confuse with 'alloced' which is the number of
2372 * pages added to the pcp list.
2374 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2375 spin_unlock(&zone
->lock
);
2381 * Called from the vmstat counter updater to drain pagesets of this
2382 * currently executing processor on remote nodes after they have
2385 * Note that this function must be called with the thread pinned to
2386 * a single processor.
2388 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2390 unsigned long flags
;
2391 int to_drain
, batch
;
2393 local_irq_save(flags
);
2394 batch
= READ_ONCE(pcp
->batch
);
2395 to_drain
= min(pcp
->count
, batch
);
2397 free_pcppages_bulk(zone
, to_drain
, pcp
);
2398 pcp
->count
-= to_drain
;
2400 local_irq_restore(flags
);
2405 * Drain pcplists of the indicated processor and zone.
2407 * The processor must either be the current processor and the
2408 * thread pinned to the current processor or a processor that
2411 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2413 unsigned long flags
;
2414 struct per_cpu_pageset
*pset
;
2415 struct per_cpu_pages
*pcp
;
2417 local_irq_save(flags
);
2418 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2422 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2425 local_irq_restore(flags
);
2429 * Drain pcplists of all zones on the indicated processor.
2431 * The processor must either be the current processor and the
2432 * thread pinned to the current processor or a processor that
2435 static void drain_pages(unsigned int cpu
)
2439 for_each_populated_zone(zone
) {
2440 drain_pages_zone(cpu
, zone
);
2445 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2447 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2448 * the single zone's pages.
2450 void drain_local_pages(struct zone
*zone
)
2452 int cpu
= smp_processor_id();
2455 drain_pages_zone(cpu
, zone
);
2460 static void drain_local_pages_wq(struct work_struct
*work
)
2463 * drain_all_pages doesn't use proper cpu hotplug protection so
2464 * we can race with cpu offline when the WQ can move this from
2465 * a cpu pinned worker to an unbound one. We can operate on a different
2466 * cpu which is allright but we also have to make sure to not move to
2470 drain_local_pages(NULL
);
2475 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2477 * When zone parameter is non-NULL, spill just the single zone's pages.
2479 * Note that this can be extremely slow as the draining happens in a workqueue.
2481 void drain_all_pages(struct zone
*zone
)
2486 * Allocate in the BSS so we wont require allocation in
2487 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2489 static cpumask_t cpus_with_pcps
;
2492 * Make sure nobody triggers this path before mm_percpu_wq is fully
2495 if (WARN_ON_ONCE(!mm_percpu_wq
))
2499 * Do not drain if one is already in progress unless it's specific to
2500 * a zone. Such callers are primarily CMA and memory hotplug and need
2501 * the drain to be complete when the call returns.
2503 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2506 mutex_lock(&pcpu_drain_mutex
);
2510 * We don't care about racing with CPU hotplug event
2511 * as offline notification will cause the notified
2512 * cpu to drain that CPU pcps and on_each_cpu_mask
2513 * disables preemption as part of its processing
2515 for_each_online_cpu(cpu
) {
2516 struct per_cpu_pageset
*pcp
;
2518 bool has_pcps
= false;
2521 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2525 for_each_populated_zone(z
) {
2526 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2527 if (pcp
->pcp
.count
) {
2535 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2537 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2540 for_each_cpu(cpu
, &cpus_with_pcps
) {
2541 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2542 INIT_WORK(work
, drain_local_pages_wq
);
2543 queue_work_on(cpu
, mm_percpu_wq
, work
);
2545 for_each_cpu(cpu
, &cpus_with_pcps
)
2546 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2548 mutex_unlock(&pcpu_drain_mutex
);
2551 #ifdef CONFIG_HIBERNATION
2554 * Touch the watchdog for every WD_PAGE_COUNT pages.
2556 #define WD_PAGE_COUNT (128*1024)
2558 void mark_free_pages(struct zone
*zone
)
2560 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2561 unsigned long flags
;
2562 unsigned int order
, t
;
2565 if (zone_is_empty(zone
))
2568 spin_lock_irqsave(&zone
->lock
, flags
);
2570 max_zone_pfn
= zone_end_pfn(zone
);
2571 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2572 if (pfn_valid(pfn
)) {
2573 page
= pfn_to_page(pfn
);
2575 if (!--page_count
) {
2576 touch_nmi_watchdog();
2577 page_count
= WD_PAGE_COUNT
;
2580 if (page_zone(page
) != zone
)
2583 if (!swsusp_page_is_forbidden(page
))
2584 swsusp_unset_page_free(page
);
2587 for_each_migratetype_order(order
, t
) {
2588 list_for_each_entry(page
,
2589 &zone
->free_area
[order
].free_list
[t
], lru
) {
2592 pfn
= page_to_pfn(page
);
2593 for (i
= 0; i
< (1UL << order
); i
++) {
2594 if (!--page_count
) {
2595 touch_nmi_watchdog();
2596 page_count
= WD_PAGE_COUNT
;
2598 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2602 spin_unlock_irqrestore(&zone
->lock
, flags
);
2604 #endif /* CONFIG_PM */
2607 * Free a 0-order page
2608 * cold == true ? free a cold page : free a hot page
2610 void free_hot_cold_page(struct page
*page
, bool cold
)
2612 struct zone
*zone
= page_zone(page
);
2613 struct per_cpu_pages
*pcp
;
2614 unsigned long flags
;
2615 unsigned long pfn
= page_to_pfn(page
);
2618 if (!free_pcp_prepare(page
))
2621 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2622 set_pcppage_migratetype(page
, migratetype
);
2623 local_irq_save(flags
);
2624 __count_vm_event(PGFREE
);
2627 * We only track unmovable, reclaimable and movable on pcp lists.
2628 * Free ISOLATE pages back to the allocator because they are being
2629 * offlined but treat HIGHATOMIC as movable pages so we can get those
2630 * areas back if necessary. Otherwise, we may have to free
2631 * excessively into the page allocator
2633 if (migratetype
>= MIGRATE_PCPTYPES
) {
2634 if (unlikely(is_migrate_isolate(migratetype
))) {
2635 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2638 migratetype
= MIGRATE_MOVABLE
;
2641 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2643 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2645 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2647 if (pcp
->count
>= pcp
->high
) {
2648 unsigned long batch
= READ_ONCE(pcp
->batch
);
2649 free_pcppages_bulk(zone
, batch
, pcp
);
2650 pcp
->count
-= batch
;
2654 local_irq_restore(flags
);
2658 * Free a list of 0-order pages
2660 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2662 struct page
*page
, *next
;
2664 list_for_each_entry_safe(page
, next
, list
, lru
) {
2665 trace_mm_page_free_batched(page
, cold
);
2666 free_hot_cold_page(page
, cold
);
2671 * split_page takes a non-compound higher-order page, and splits it into
2672 * n (1<<order) sub-pages: page[0..n]
2673 * Each sub-page must be freed individually.
2675 * Note: this is probably too low level an operation for use in drivers.
2676 * Please consult with lkml before using this in your driver.
2678 void split_page(struct page
*page
, unsigned int order
)
2682 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2683 VM_BUG_ON_PAGE(!page_count(page
), page
);
2685 for (i
= 1; i
< (1 << order
); i
++)
2686 set_page_refcounted(page
+ i
);
2687 split_page_owner(page
, order
);
2689 EXPORT_SYMBOL_GPL(split_page
);
2691 int __isolate_free_page(struct page
*page
, unsigned int order
)
2693 unsigned long watermark
;
2697 BUG_ON(!PageBuddy(page
));
2699 zone
= page_zone(page
);
2700 mt
= get_pageblock_migratetype(page
);
2702 if (!is_migrate_isolate(mt
)) {
2704 * Obey watermarks as if the page was being allocated. We can
2705 * emulate a high-order watermark check with a raised order-0
2706 * watermark, because we already know our high-order page
2709 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2710 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
2713 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2716 /* Remove page from free list */
2717 list_del(&page
->lru
);
2718 zone
->free_area
[order
].nr_free
--;
2719 rmv_page_order(page
);
2722 * Set the pageblock if the isolated page is at least half of a
2725 if (order
>= pageblock_order
- 1) {
2726 struct page
*endpage
= page
+ (1 << order
) - 1;
2727 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2728 int mt
= get_pageblock_migratetype(page
);
2729 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2730 && !is_migrate_highatomic(mt
))
2731 set_pageblock_migratetype(page
,
2737 return 1UL << order
;
2741 * Update NUMA hit/miss statistics
2743 * Must be called with interrupts disabled.
2745 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2748 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2750 if (z
->node
!= numa_node_id())
2751 local_stat
= NUMA_OTHER
;
2753 if (z
->node
== preferred_zone
->node
)
2754 __inc_numa_state(z
, NUMA_HIT
);
2756 __inc_numa_state(z
, NUMA_MISS
);
2757 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
2759 __inc_numa_state(z
, local_stat
);
2763 /* Remove page from the per-cpu list, caller must protect the list */
2764 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2765 bool cold
, struct per_cpu_pages
*pcp
,
2766 struct list_head
*list
)
2771 if (list_empty(list
)) {
2772 pcp
->count
+= rmqueue_bulk(zone
, 0,
2775 if (unlikely(list_empty(list
)))
2780 page
= list_last_entry(list
, struct page
, lru
);
2782 page
= list_first_entry(list
, struct page
, lru
);
2785 * If the head or the tail page in the pcp list is CMA page and
2786 * the gfp flags is not GFP_HIGHUSER_MOVABLE, do not allocate a
2787 * page from the pcp list. The free list of MIGRATE_CMA is a
2788 * special case of the free list of MIGRATE_MOVABLE and the
2789 * pages from the free list of MIGRATE_CMA are pushed to the pcp
2790 * list of MIGRATE_MOVABLE. Since the pcp list of
2791 * MIGRATE_MOVABLE is selected if the gfp flags has GFP_MOVABLE,
2792 * we should avoid the case that a cma page in the pcp list of
2793 * MIGRATE_MOVABLE is allocated to a movable allocation without
2794 * GFP_HIGHUSER_MOVABLE.
2795 * If this is the case, allocate a movable page from the free
2796 * list of MIGRATE_MOVABLE instead of pcp list of
2800 if (is_migrate_cma_page(page
) && (migratetype
!= MIGRATE_CMA
))
2803 list_del(&page
->lru
);
2805 } while (check_new_pcp(page
));
2810 /* Lock and remove page from the per-cpu list */
2811 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2812 struct zone
*zone
, unsigned int order
,
2813 gfp_t gfp_flags
, int migratetype
,
2814 int migratetype_rmqueue
)
2816 struct per_cpu_pages
*pcp
;
2817 struct list_head
*list
;
2818 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2820 unsigned long flags
;
2822 local_irq_save(flags
);
2823 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2824 list
= &pcp
->lists
[migratetype
];
2825 page
= __rmqueue_pcplist(zone
, migratetype_rmqueue
, cold
, pcp
, list
);
2827 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2828 zone_statistics(preferred_zone
, zone
);
2830 local_irq_restore(flags
);
2835 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2838 struct page
*rmqueue(struct zone
*preferred_zone
,
2839 struct zone
*zone
, unsigned int order
,
2840 gfp_t gfp_flags
, unsigned int alloc_flags
,
2843 unsigned long flags
;
2845 int migratetype_rmqueue
= migratetype
;
2848 if ((migratetype_rmqueue
== MIGRATE_MOVABLE
) &&
2849 ((gfp_flags
& GFP_HIGHUSER_MOVABLE
) == GFP_HIGHUSER_MOVABLE
))
2850 migratetype_rmqueue
= MIGRATE_CMA
;
2852 if (likely(order
== 0)) {
2853 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
2854 gfp_flags
, migratetype
, migratetype_rmqueue
);
2856 * Allocation with GFP_MOVABLE and !GFP_HIGHMEM will have
2857 * another chance of page allocation from the free list.
2858 * See the comment in __rmqueue_pcplist().
2861 if (likely(page
) || (migratetype_rmqueue
!= MIGRATE_MOVABLE
))
2867 * We most definitely don't want callers attempting to
2868 * allocate greater than order-1 page units with __GFP_NOFAIL.
2870 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2871 spin_lock_irqsave(&zone
->lock
, flags
);
2875 if (alloc_flags
& ALLOC_HARDER
) {
2876 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2878 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2881 page
= __rmqueue(zone
, order
, migratetype_rmqueue
);
2882 } while (page
&& check_new_pages(page
, order
));
2883 spin_unlock(&zone
->lock
);
2886 __mod_zone_freepage_state(zone
, -(1 << order
),
2887 get_pcppage_migratetype(page
));
2889 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2890 zone_statistics(preferred_zone
, zone
);
2891 local_irq_restore(flags
);
2894 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
2898 local_irq_restore(flags
);
2902 #ifdef CONFIG_FAIL_PAGE_ALLOC
2905 struct fault_attr attr
;
2907 bool ignore_gfp_highmem
;
2908 bool ignore_gfp_reclaim
;
2910 } fail_page_alloc
= {
2911 .attr
= FAULT_ATTR_INITIALIZER
,
2912 .ignore_gfp_reclaim
= true,
2913 .ignore_gfp_highmem
= true,
2917 static int __init
setup_fail_page_alloc(char *str
)
2919 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2921 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2923 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2925 if (order
< fail_page_alloc
.min_order
)
2927 if (gfp_mask
& __GFP_NOFAIL
)
2929 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2931 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2932 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2935 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2938 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2940 static int __init
fail_page_alloc_debugfs(void)
2942 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2945 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2946 &fail_page_alloc
.attr
);
2948 return PTR_ERR(dir
);
2950 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2951 &fail_page_alloc
.ignore_gfp_reclaim
))
2953 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2954 &fail_page_alloc
.ignore_gfp_highmem
))
2956 if (!debugfs_create_u32("min-order", mode
, dir
,
2957 &fail_page_alloc
.min_order
))
2962 debugfs_remove_recursive(dir
);
2967 late_initcall(fail_page_alloc_debugfs
);
2969 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2971 #else /* CONFIG_FAIL_PAGE_ALLOC */
2973 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2978 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2981 * Return true if free base pages are above 'mark'. For high-order checks it
2982 * will return true of the order-0 watermark is reached and there is at least
2983 * one free page of a suitable size. Checking now avoids taking the zone lock
2984 * to check in the allocation paths if no pages are free.
2986 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2987 int classzone_idx
, unsigned int alloc_flags
,
2992 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
2994 /* free_pages may go negative - that's OK */
2995 free_pages
-= (1 << order
) - 1;
2997 if (alloc_flags
& ALLOC_HIGH
)
3001 * If the caller does not have rights to ALLOC_HARDER then subtract
3002 * the high-atomic reserves. This will over-estimate the size of the
3003 * atomic reserve but it avoids a search.
3005 if (likely(!alloc_harder
)) {
3006 free_pages
-= z
->nr_reserved_highatomic
;
3009 * OOM victims can try even harder than normal ALLOC_HARDER
3010 * users on the grounds that it's definitely going to be in
3011 * the exit path shortly and free memory. Any allocation it
3012 * makes during the free path will be small and short-lived.
3014 if (alloc_flags
& ALLOC_OOM
)
3022 /* If allocation can't use CMA areas don't use free CMA pages */
3023 if (!(alloc_flags
& ALLOC_CMA
))
3024 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3028 * Check watermarks for an order-0 allocation request. If these
3029 * are not met, then a high-order request also cannot go ahead
3030 * even if a suitable page happened to be free.
3032 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3035 /* If this is an order-0 request then the watermark is fine */
3039 /* For a high-order request, check at least one suitable page is free */
3040 for (o
= order
; o
< MAX_ORDER
; o
++) {
3041 struct free_area
*area
= &z
->free_area
[o
];
3047 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3048 if (!list_empty(&area
->free_list
[mt
]))
3053 if ((alloc_flags
& ALLOC_CMA
) &&
3054 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
3059 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3065 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3066 int classzone_idx
, unsigned int alloc_flags
)
3068 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3069 zone_page_state(z
, NR_FREE_PAGES
));
3072 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3073 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3075 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3079 /* If allocation can't use CMA areas don't use free CMA pages */
3080 if (!(alloc_flags
& ALLOC_CMA
))
3081 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3085 * Fast check for order-0 only. If this fails then the reserves
3086 * need to be calculated. There is a corner case where the check
3087 * passes but only the high-order atomic reserve are free. If
3088 * the caller is !atomic then it'll uselessly search the free
3089 * list. That corner case is then slower but it is harmless.
3091 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3094 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3098 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3099 unsigned long mark
, int classzone_idx
)
3101 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3103 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3104 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3106 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3111 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3113 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3116 #else /* CONFIG_NUMA */
3117 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3121 #endif /* CONFIG_NUMA */
3124 * get_page_from_freelist goes through the zonelist trying to allocate
3127 static struct page
*
3128 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3129 const struct alloc_context
*ac
)
3131 struct zoneref
*z
= ac
->preferred_zoneref
;
3133 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3136 * Scan zonelist, looking for a zone with enough free.
3137 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3139 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3144 if (cpusets_enabled() &&
3145 (alloc_flags
& ALLOC_CPUSET
) &&
3146 !__cpuset_zone_allowed(zone
, gfp_mask
))
3149 * When allocating a page cache page for writing, we
3150 * want to get it from a node that is within its dirty
3151 * limit, such that no single node holds more than its
3152 * proportional share of globally allowed dirty pages.
3153 * The dirty limits take into account the node's
3154 * lowmem reserves and high watermark so that kswapd
3155 * should be able to balance it without having to
3156 * write pages from its LRU list.
3158 * XXX: For now, allow allocations to potentially
3159 * exceed the per-node dirty limit in the slowpath
3160 * (spread_dirty_pages unset) before going into reclaim,
3161 * which is important when on a NUMA setup the allowed
3162 * nodes are together not big enough to reach the
3163 * global limit. The proper fix for these situations
3164 * will require awareness of nodes in the
3165 * dirty-throttling and the flusher threads.
3167 if (ac
->spread_dirty_pages
) {
3168 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3171 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3172 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3177 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
3178 if (!zone_watermark_fast(zone
, order
, mark
,
3179 ac_classzone_idx(ac
), alloc_flags
)) {
3182 /* Checked here to keep the fast path fast */
3183 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3184 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3187 if (node_reclaim_mode
== 0 ||
3188 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3191 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3193 case NODE_RECLAIM_NOSCAN
:
3196 case NODE_RECLAIM_FULL
:
3197 /* scanned but unreclaimable */
3200 /* did we reclaim enough */
3201 if (zone_watermark_ok(zone
, order
, mark
,
3202 ac_classzone_idx(ac
), alloc_flags
))
3210 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3211 gfp_mask
, alloc_flags
, ac
->migratetype
);
3213 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3216 * If this is a high-order atomic allocation then check
3217 * if the pageblock should be reserved for the future
3219 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3220 reserve_highatomic_pageblock(page
, zone
, order
);
3230 * Large machines with many possible nodes should not always dump per-node
3231 * meminfo in irq context.
3233 static inline bool should_suppress_show_mem(void)
3238 ret
= in_interrupt();
3243 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3245 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3246 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3248 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs
))
3252 * This documents exceptions given to allocations in certain
3253 * contexts that are allowed to allocate outside current's set
3256 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3257 if (tsk_is_oom_victim(current
) ||
3258 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3259 filter
&= ~SHOW_MEM_FILTER_NODES
;
3260 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3261 filter
&= ~SHOW_MEM_FILTER_NODES
;
3263 show_mem(filter
, nodemask
);
3266 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3268 struct va_format vaf
;
3270 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3271 DEFAULT_RATELIMIT_BURST
);
3273 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3276 pr_warn("%s: ", current
->comm
);
3278 va_start(args
, fmt
);
3281 pr_cont("%pV", &vaf
);
3284 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask
, &gfp_mask
);
3286 pr_cont("%*pbl\n", nodemask_pr_args(nodemask
));
3288 pr_cont("(null)\n");
3290 cpuset_print_current_mems_allowed();
3293 warn_alloc_show_mem(gfp_mask
, nodemask
);
3296 static inline struct page
*
3297 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3298 unsigned int alloc_flags
,
3299 const struct alloc_context
*ac
)
3303 page
= get_page_from_freelist(gfp_mask
, order
,
3304 alloc_flags
|ALLOC_CPUSET
, ac
);
3306 * fallback to ignore cpuset restriction if our nodes
3310 page
= get_page_from_freelist(gfp_mask
, order
,
3316 static inline struct page
*
3317 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3318 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3320 struct oom_control oc
= {
3321 .zonelist
= ac
->zonelist
,
3322 .nodemask
= ac
->nodemask
,
3324 .gfp_mask
= gfp_mask
,
3329 *did_some_progress
= 0;
3332 * Acquire the oom lock. If that fails, somebody else is
3333 * making progress for us.
3335 if (!mutex_trylock(&oom_lock
)) {
3336 *did_some_progress
= 1;
3337 schedule_timeout_uninterruptible(1);
3342 * Go through the zonelist yet one more time, keep very high watermark
3343 * here, this is only to catch a parallel oom killing, we must fail if
3344 * we're still under heavy pressure. But make sure that this reclaim
3345 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3346 * allocation which will never fail due to oom_lock already held.
3348 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3349 ~__GFP_DIRECT_RECLAIM
, order
,
3350 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3354 /* Coredumps can quickly deplete all memory reserves */
3355 if (current
->flags
& PF_DUMPCORE
)
3357 /* The OOM killer will not help higher order allocs */
3358 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3361 * We have already exhausted all our reclaim opportunities without any
3362 * success so it is time to admit defeat. We will skip the OOM killer
3363 * because it is very likely that the caller has a more reasonable
3364 * fallback than shooting a random task.
3366 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3368 /* The OOM killer does not needlessly kill tasks for lowmem */
3369 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3371 if (pm_suspended_storage())
3374 * XXX: GFP_NOFS allocations should rather fail than rely on
3375 * other request to make a forward progress.
3376 * We are in an unfortunate situation where out_of_memory cannot
3377 * do much for this context but let's try it to at least get
3378 * access to memory reserved if the current task is killed (see
3379 * out_of_memory). Once filesystems are ready to handle allocation
3380 * failures more gracefully we should just bail out here.
3383 /* The OOM killer may not free memory on a specific node */
3384 if (gfp_mask
& __GFP_THISNODE
)
3387 /* Exhausted what can be done so it's blamo time */
3388 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3389 *did_some_progress
= 1;
3392 * Help non-failing allocations by giving them access to memory
3395 if (gfp_mask
& __GFP_NOFAIL
)
3396 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3397 ALLOC_NO_WATERMARKS
, ac
);
3400 mutex_unlock(&oom_lock
);
3405 * Maximum number of compaction retries wit a progress before OOM
3406 * killer is consider as the only way to move forward.
3408 #define MAX_COMPACT_RETRIES 16
3410 #ifdef CONFIG_COMPACTION
3411 /* Try memory compaction for high-order allocations before reclaim */
3412 static struct page
*
3413 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3414 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3415 enum compact_priority prio
, enum compact_result
*compact_result
)
3418 unsigned long pflags
;
3419 unsigned int noreclaim_flag
;
3424 psi_memstall_enter(&pflags
);
3425 noreclaim_flag
= memalloc_noreclaim_save();
3427 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3430 memalloc_noreclaim_restore(noreclaim_flag
);
3431 psi_memstall_leave(&pflags
);
3433 if (*compact_result
<= COMPACT_INACTIVE
)
3437 * At least in one zone compaction wasn't deferred or skipped, so let's
3438 * count a compaction stall
3440 count_vm_event(COMPACTSTALL
);
3442 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3445 struct zone
*zone
= page_zone(page
);
3447 zone
->compact_blockskip_flush
= false;
3448 compaction_defer_reset(zone
, order
, true);
3449 count_vm_event(COMPACTSUCCESS
);
3454 * It's bad if compaction run occurs and fails. The most likely reason
3455 * is that pages exist, but not enough to satisfy watermarks.
3457 count_vm_event(COMPACTFAIL
);
3465 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3466 enum compact_result compact_result
,
3467 enum compact_priority
*compact_priority
,
3468 int *compaction_retries
)
3470 int max_retries
= MAX_COMPACT_RETRIES
;
3473 int retries
= *compaction_retries
;
3474 enum compact_priority priority
= *compact_priority
;
3479 if (compaction_made_progress(compact_result
))
3480 (*compaction_retries
)++;
3483 * compaction considers all the zone as desperately out of memory
3484 * so it doesn't really make much sense to retry except when the
3485 * failure could be caused by insufficient priority
3487 if (compaction_failed(compact_result
))
3488 goto check_priority
;
3491 * make sure the compaction wasn't deferred or didn't bail out early
3492 * due to locks contention before we declare that we should give up.
3493 * But do not retry if the given zonelist is not suitable for
3496 if (compaction_withdrawn(compact_result
)) {
3497 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3502 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3503 * costly ones because they are de facto nofail and invoke OOM
3504 * killer to move on while costly can fail and users are ready
3505 * to cope with that. 1/4 retries is rather arbitrary but we
3506 * would need much more detailed feedback from compaction to
3507 * make a better decision.
3509 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3511 if (*compaction_retries
<= max_retries
) {
3517 * Make sure there are attempts at the highest priority if we exhausted
3518 * all retries or failed at the lower priorities.
3521 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3522 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3524 if (*compact_priority
> min_priority
) {
3525 (*compact_priority
)--;
3526 *compaction_retries
= 0;
3530 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3534 static inline struct page
*
3535 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3536 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3537 enum compact_priority prio
, enum compact_result
*compact_result
)
3539 *compact_result
= COMPACT_SKIPPED
;
3544 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3545 enum compact_result compact_result
,
3546 enum compact_priority
*compact_priority
,
3547 int *compaction_retries
)
3552 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3556 * There are setups with compaction disabled which would prefer to loop
3557 * inside the allocator rather than hit the oom killer prematurely.
3558 * Let's give them a good hope and keep retrying while the order-0
3559 * watermarks are OK.
3561 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3563 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3564 ac_classzone_idx(ac
), alloc_flags
))
3569 #endif /* CONFIG_COMPACTION */
3571 #ifdef CONFIG_LOCKDEP
3572 struct lockdep_map __fs_reclaim_map
=
3573 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3575 static bool __need_fs_reclaim(gfp_t gfp_mask
)
3577 gfp_mask
= current_gfp_context(gfp_mask
);
3579 /* no reclaim without waiting on it */
3580 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3583 /* this guy won't enter reclaim */
3584 if (current
->flags
& PF_MEMALLOC
)
3587 /* We're only interested __GFP_FS allocations for now */
3588 if (!(gfp_mask
& __GFP_FS
))
3591 if (gfp_mask
& __GFP_NOLOCKDEP
)
3597 void fs_reclaim_acquire(gfp_t gfp_mask
)
3599 if (__need_fs_reclaim(gfp_mask
))
3600 lock_map_acquire(&__fs_reclaim_map
);
3602 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3604 void fs_reclaim_release(gfp_t gfp_mask
)
3606 if (__need_fs_reclaim(gfp_mask
))
3607 lock_map_release(&__fs_reclaim_map
);
3609 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3612 /* Perform direct synchronous page reclaim */
3614 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3615 const struct alloc_context
*ac
)
3617 struct reclaim_state reclaim_state
;
3619 unsigned int noreclaim_flag
;
3620 unsigned long pflags
;
3624 /* We now go into synchronous reclaim */
3625 cpuset_memory_pressure_bump();
3626 psi_memstall_enter(&pflags
);
3627 noreclaim_flag
= memalloc_noreclaim_save();
3628 fs_reclaim_acquire(gfp_mask
);
3629 reclaim_state
.reclaimed_slab
= 0;
3630 current
->reclaim_state
= &reclaim_state
;
3632 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3635 current
->reclaim_state
= NULL
;
3636 fs_reclaim_release(gfp_mask
);
3637 memalloc_noreclaim_restore(noreclaim_flag
);
3638 psi_memstall_leave(&pflags
);
3645 /* The really slow allocator path where we enter direct reclaim */
3646 static inline struct page
*
3647 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3648 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3649 unsigned long *did_some_progress
)
3651 struct page
*page
= NULL
;
3652 bool drained
= false;
3654 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3655 if (unlikely(!(*did_some_progress
)))
3659 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3662 * If an allocation failed after direct reclaim, it could be because
3663 * pages are pinned on the per-cpu lists or in high alloc reserves.
3664 * Shrink them them and try again
3666 if (!page
&& !drained
) {
3667 unreserve_highatomic_pageblock(ac
, false);
3668 drain_all_pages(NULL
);
3676 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3680 pg_data_t
*last_pgdat
= NULL
;
3682 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3683 ac
->high_zoneidx
, ac
->nodemask
) {
3684 if (last_pgdat
!= zone
->zone_pgdat
)
3685 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3686 last_pgdat
= zone
->zone_pgdat
;
3690 static inline unsigned int
3691 gfp_to_alloc_flags(gfp_t gfp_mask
)
3693 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3695 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3696 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3699 * The caller may dip into page reserves a bit more if the caller
3700 * cannot run direct reclaim, or if the caller has realtime scheduling
3701 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3702 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3704 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3706 if (gfp_mask
& __GFP_ATOMIC
) {
3708 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3709 * if it can't schedule.
3711 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3712 alloc_flags
|= ALLOC_HARDER
;
3714 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3715 * comment for __cpuset_node_allowed().
3717 alloc_flags
&= ~ALLOC_CPUSET
;
3718 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3719 alloc_flags
|= ALLOC_HARDER
;
3722 if ((gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
) ||
3723 ((gfp_mask
& GFP_HIGHUSER_MOVABLE
) == GFP_HIGHUSER_MOVABLE
))
3724 alloc_flags
|= ALLOC_CMA
;
3729 static bool oom_reserves_allowed(struct task_struct
*tsk
)
3731 if (!tsk_is_oom_victim(tsk
))
3735 * !MMU doesn't have oom reaper so give access to memory reserves
3736 * only to the thread with TIF_MEMDIE set
3738 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
3745 * Distinguish requests which really need access to full memory
3746 * reserves from oom victims which can live with a portion of it
3748 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
3750 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3752 if (gfp_mask
& __GFP_MEMALLOC
)
3753 return ALLOC_NO_WATERMARKS
;
3754 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3755 return ALLOC_NO_WATERMARKS
;
3756 if (!in_interrupt()) {
3757 if (current
->flags
& PF_MEMALLOC
)
3758 return ALLOC_NO_WATERMARKS
;
3759 else if (oom_reserves_allowed(current
))
3766 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3768 return !!__gfp_pfmemalloc_flags(gfp_mask
);
3772 * Checks whether it makes sense to retry the reclaim to make a forward progress
3773 * for the given allocation request.
3775 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3776 * without success, or when we couldn't even meet the watermark if we
3777 * reclaimed all remaining pages on the LRU lists.
3779 * Returns true if a retry is viable or false to enter the oom path.
3782 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3783 struct alloc_context
*ac
, int alloc_flags
,
3784 bool did_some_progress
, int *no_progress_loops
)
3790 * Costly allocations might have made a progress but this doesn't mean
3791 * their order will become available due to high fragmentation so
3792 * always increment the no progress counter for them
3794 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3795 *no_progress_loops
= 0;
3797 (*no_progress_loops
)++;
3800 * Make sure we converge to OOM if we cannot make any progress
3801 * several times in the row.
3803 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3804 /* Before OOM, exhaust highatomic_reserve */
3805 return unreserve_highatomic_pageblock(ac
, true);
3809 * Keep reclaiming pages while there is a chance this will lead
3810 * somewhere. If none of the target zones can satisfy our allocation
3811 * request even if all reclaimable pages are considered then we are
3812 * screwed and have to go OOM.
3814 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3816 unsigned long available
;
3817 unsigned long reclaimable
;
3818 unsigned long min_wmark
= min_wmark_pages(zone
);
3821 available
= reclaimable
= zone_reclaimable_pages(zone
);
3822 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3825 * Would the allocation succeed if we reclaimed all
3826 * reclaimable pages?
3828 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3829 ac_classzone_idx(ac
), alloc_flags
, available
);
3830 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3831 available
, min_wmark
, *no_progress_loops
, wmark
);
3834 * If we didn't make any progress and have a lot of
3835 * dirty + writeback pages then we should wait for
3836 * an IO to complete to slow down the reclaim and
3837 * prevent from pre mature OOM
3839 if (!did_some_progress
) {
3840 unsigned long write_pending
;
3842 write_pending
= zone_page_state_snapshot(zone
,
3843 NR_ZONE_WRITE_PENDING
);
3845 if (2 * write_pending
> reclaimable
) {
3846 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3852 * Memory allocation/reclaim might be called from a WQ
3853 * context and the current implementation of the WQ
3854 * concurrency control doesn't recognize that
3855 * a particular WQ is congested if the worker thread is
3856 * looping without ever sleeping. Therefore we have to
3857 * do a short sleep here rather than calling
3860 if (current
->flags
& PF_WQ_WORKER
)
3861 schedule_timeout_uninterruptible(1);
3873 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
3876 * It's possible that cpuset's mems_allowed and the nodemask from
3877 * mempolicy don't intersect. This should be normally dealt with by
3878 * policy_nodemask(), but it's possible to race with cpuset update in
3879 * such a way the check therein was true, and then it became false
3880 * before we got our cpuset_mems_cookie here.
3881 * This assumes that for all allocations, ac->nodemask can come only
3882 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3883 * when it does not intersect with the cpuset restrictions) or the
3884 * caller can deal with a violated nodemask.
3886 if (cpusets_enabled() && ac
->nodemask
&&
3887 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
3888 ac
->nodemask
= NULL
;
3893 * When updating a task's mems_allowed or mempolicy nodemask, it is
3894 * possible to race with parallel threads in such a way that our
3895 * allocation can fail while the mask is being updated. If we are about
3896 * to fail, check if the cpuset changed during allocation and if so,
3899 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3905 static inline struct page
*
3906 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3907 struct alloc_context
*ac
)
3909 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3910 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
3911 struct page
*page
= NULL
;
3912 unsigned int alloc_flags
;
3913 unsigned long did_some_progress
;
3914 enum compact_priority compact_priority
;
3915 enum compact_result compact_result
;
3916 int compaction_retries
;
3917 int no_progress_loops
;
3918 unsigned int cpuset_mems_cookie
;
3922 * We also sanity check to catch abuse of atomic reserves being used by
3923 * callers that are not in atomic context.
3925 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3926 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3927 gfp_mask
&= ~__GFP_ATOMIC
;
3930 compaction_retries
= 0;
3931 no_progress_loops
= 0;
3932 compact_priority
= DEF_COMPACT_PRIORITY
;
3933 cpuset_mems_cookie
= read_mems_allowed_begin();
3936 * The fast path uses conservative alloc_flags to succeed only until
3937 * kswapd needs to be woken up, and to avoid the cost of setting up
3938 * alloc_flags precisely. So we do that now.
3940 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3943 * We need to recalculate the starting point for the zonelist iterator
3944 * because we might have used different nodemask in the fast path, or
3945 * there was a cpuset modification and we are retrying - otherwise we
3946 * could end up iterating over non-eligible zones endlessly.
3948 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3949 ac
->high_zoneidx
, ac
->nodemask
);
3950 if (!ac
->preferred_zoneref
->zone
)
3953 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3954 wake_all_kswapds(order
, ac
);
3957 * The adjusted alloc_flags might result in immediate success, so try
3960 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3965 * For costly allocations, try direct compaction first, as it's likely
3966 * that we have enough base pages and don't need to reclaim. For non-
3967 * movable high-order allocations, do that as well, as compaction will
3968 * try prevent permanent fragmentation by migrating from blocks of the
3970 * Don't try this for allocations that are allowed to ignore
3971 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3973 if (can_direct_reclaim
&&
3975 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
3976 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
3977 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3979 INIT_COMPACT_PRIORITY
,
3985 * Checks for costly allocations with __GFP_NORETRY, which
3986 * includes THP page fault allocations
3988 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
3990 * If compaction is deferred for high-order allocations,
3991 * it is because sync compaction recently failed. If
3992 * this is the case and the caller requested a THP
3993 * allocation, we do not want to heavily disrupt the
3994 * system, so we fail the allocation instead of entering
3997 if (compact_result
== COMPACT_DEFERRED
)
4001 * Looks like reclaim/compaction is worth trying, but
4002 * sync compaction could be very expensive, so keep
4003 * using async compaction.
4005 compact_priority
= INIT_COMPACT_PRIORITY
;
4010 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4011 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4012 wake_all_kswapds(order
, ac
);
4014 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4016 alloc_flags
= reserve_flags
;
4019 * Reset the zonelist iterators if memory policies can be ignored.
4020 * These allocations are high priority and system rather than user
4023 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4024 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4025 ac
->high_zoneidx
, ac
->nodemask
);
4028 /* Attempt with potentially adjusted zonelist and alloc_flags */
4029 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4033 /* Caller is not willing to reclaim, we can't balance anything */
4034 if (!can_direct_reclaim
)
4037 /* Avoid recursion of direct reclaim */
4038 if (current
->flags
& PF_MEMALLOC
)
4041 /* Try direct reclaim and then allocating */
4042 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4043 &did_some_progress
);
4047 /* Try direct compaction and then allocating */
4048 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4049 compact_priority
, &compact_result
);
4053 /* Do not loop if specifically requested */
4054 if (gfp_mask
& __GFP_NORETRY
)
4058 * Do not retry costly high order allocations unless they are
4059 * __GFP_RETRY_MAYFAIL
4061 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4064 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4065 did_some_progress
> 0, &no_progress_loops
))
4069 * It doesn't make any sense to retry for the compaction if the order-0
4070 * reclaim is not able to make any progress because the current
4071 * implementation of the compaction depends on the sufficient amount
4072 * of free memory (see __compaction_suitable)
4074 if (did_some_progress
> 0 &&
4075 should_compact_retry(ac
, order
, alloc_flags
,
4076 compact_result
, &compact_priority
,
4077 &compaction_retries
))
4081 /* Deal with possible cpuset update races before we start OOM killing */
4082 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4085 /* Reclaim has failed us, start killing things */
4086 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4090 /* Avoid allocations with no watermarks from looping endlessly */
4091 if (tsk_is_oom_victim(current
) &&
4092 (alloc_flags
== ALLOC_OOM
||
4093 (gfp_mask
& __GFP_NOMEMALLOC
)))
4096 /* Retry as long as the OOM killer is making progress */
4097 if (did_some_progress
) {
4098 no_progress_loops
= 0;
4103 /* Deal with possible cpuset update races before we fail */
4104 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4108 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4111 if (gfp_mask
& __GFP_NOFAIL
) {
4113 * All existing users of the __GFP_NOFAIL are blockable, so warn
4114 * of any new users that actually require GFP_NOWAIT
4116 if (WARN_ON_ONCE(!can_direct_reclaim
))
4120 * PF_MEMALLOC request from this context is rather bizarre
4121 * because we cannot reclaim anything and only can loop waiting
4122 * for somebody to do a work for us
4124 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4127 * non failing costly orders are a hard requirement which we
4128 * are not prepared for much so let's warn about these users
4129 * so that we can identify them and convert them to something
4132 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4135 * Help non-failing allocations by giving them access to memory
4136 * reserves but do not use ALLOC_NO_WATERMARKS because this
4137 * could deplete whole memory reserves which would just make
4138 * the situation worse
4140 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4148 warn_alloc(gfp_mask
, ac
->nodemask
,
4149 "page allocation failure: order:%u", order
);
4154 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4155 int preferred_nid
, nodemask_t
*nodemask
,
4156 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4157 unsigned int *alloc_flags
)
4159 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4160 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4161 ac
->nodemask
= nodemask
;
4162 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4164 if (cpusets_enabled()) {
4165 *alloc_mask
|= __GFP_HARDWALL
;
4167 ac
->nodemask
= &cpuset_current_mems_allowed
;
4169 *alloc_flags
|= ALLOC_CPUSET
;
4172 fs_reclaim_acquire(gfp_mask
);
4173 fs_reclaim_release(gfp_mask
);
4175 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4177 if (should_fail_alloc_page(gfp_mask
, order
))
4180 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4181 *alloc_flags
|= ALLOC_CMA
;
4186 /* Determine whether to spread dirty pages and what the first usable zone */
4187 static inline void finalise_ac(gfp_t gfp_mask
,
4188 unsigned int order
, struct alloc_context
*ac
)
4190 /* Dirty zone balancing only done in the fast path */
4191 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4194 * The preferred zone is used for statistics but crucially it is
4195 * also used as the starting point for the zonelist iterator. It
4196 * may get reset for allocations that ignore memory policies.
4198 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4199 ac
->high_zoneidx
, ac
->nodemask
);
4203 * This is the 'heart' of the zoned buddy allocator.
4206 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4207 nodemask_t
*nodemask
)
4210 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4211 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4212 struct alloc_context ac
= { };
4215 * There are several places where we assume that the order value is sane
4216 * so bail out early if the request is out of bound.
4218 if (unlikely(order
>= MAX_ORDER
)) {
4219 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4223 gfp_mask
&= gfp_allowed_mask
;
4224 alloc_mask
= gfp_mask
;
4225 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4228 finalise_ac(gfp_mask
, order
, &ac
);
4230 /* First allocation attempt */
4231 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4236 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4237 * resp. GFP_NOIO which has to be inherited for all allocation requests
4238 * from a particular context which has been marked by
4239 * memalloc_no{fs,io}_{save,restore}.
4241 alloc_mask
= current_gfp_context(gfp_mask
);
4242 ac
.spread_dirty_pages
= false;
4245 * Restore the original nodemask if it was potentially replaced with
4246 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4248 if (unlikely(ac
.nodemask
!= nodemask
))
4249 ac
.nodemask
= nodemask
;
4251 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4254 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4255 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4256 __free_pages(page
, order
);
4260 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4264 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4267 * Common helper functions.
4269 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4274 * __get_free_pages() returns a 32-bit address, which cannot represent
4277 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
4279 page
= alloc_pages(gfp_mask
, order
);
4282 return (unsigned long) page_address(page
);
4284 EXPORT_SYMBOL(__get_free_pages
);
4286 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4288 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4290 EXPORT_SYMBOL(get_zeroed_page
);
4292 void __free_pages(struct page
*page
, unsigned int order
)
4294 if (put_page_testzero(page
)) {
4296 free_hot_cold_page(page
, false);
4298 __free_pages_ok(page
, order
);
4302 EXPORT_SYMBOL(__free_pages
);
4304 void free_pages(unsigned long addr
, unsigned int order
)
4307 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4308 __free_pages(virt_to_page((void *)addr
), order
);
4312 EXPORT_SYMBOL(free_pages
);
4316 * An arbitrary-length arbitrary-offset area of memory which resides
4317 * within a 0 or higher order page. Multiple fragments within that page
4318 * are individually refcounted, in the page's reference counter.
4320 * The page_frag functions below provide a simple allocation framework for
4321 * page fragments. This is used by the network stack and network device
4322 * drivers to provide a backing region of memory for use as either an
4323 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4325 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4328 struct page
*page
= NULL
;
4329 gfp_t gfp
= gfp_mask
;
4331 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4332 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4334 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4335 PAGE_FRAG_CACHE_MAX_ORDER
);
4336 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4338 if (unlikely(!page
))
4339 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4341 nc
->va
= page
? page_address(page
) : NULL
;
4346 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4348 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4350 if (page_ref_sub_and_test(page
, count
)) {
4351 unsigned int order
= compound_order(page
);
4354 free_hot_cold_page(page
, false);
4356 __free_pages_ok(page
, order
);
4359 EXPORT_SYMBOL(__page_frag_cache_drain
);
4361 void *page_frag_alloc(struct page_frag_cache
*nc
,
4362 unsigned int fragsz
, gfp_t gfp_mask
)
4364 unsigned int size
= PAGE_SIZE
;
4368 if (unlikely(!nc
->va
)) {
4370 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4374 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4375 /* if size can vary use size else just use PAGE_SIZE */
4378 /* Even if we own the page, we do not use atomic_set().
4379 * This would break get_page_unless_zero() users.
4381 page_ref_add(page
, size
);
4383 /* reset page count bias and offset to start of new frag */
4384 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4385 nc
->pagecnt_bias
= size
+ 1;
4389 offset
= nc
->offset
- fragsz
;
4390 if (unlikely(offset
< 0)) {
4391 page
= virt_to_page(nc
->va
);
4393 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4396 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4397 /* if size can vary use size else just use PAGE_SIZE */
4400 /* OK, page count is 0, we can safely set it */
4401 set_page_count(page
, size
+ 1);
4403 /* reset page count bias and offset to start of new frag */
4404 nc
->pagecnt_bias
= size
+ 1;
4405 offset
= size
- fragsz
;
4409 nc
->offset
= offset
;
4411 return nc
->va
+ offset
;
4413 EXPORT_SYMBOL(page_frag_alloc
);
4416 * Frees a page fragment allocated out of either a compound or order 0 page.
4418 void page_frag_free(void *addr
)
4420 struct page
*page
= virt_to_head_page(addr
);
4422 if (unlikely(put_page_testzero(page
)))
4423 __free_pages_ok(page
, compound_order(page
));
4425 EXPORT_SYMBOL(page_frag_free
);
4427 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4431 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4432 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4434 split_page(virt_to_page((void *)addr
), order
);
4435 while (used
< alloc_end
) {
4440 return (void *)addr
;
4444 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4445 * @size: the number of bytes to allocate
4446 * @gfp_mask: GFP flags for the allocation
4448 * This function is similar to alloc_pages(), except that it allocates the
4449 * minimum number of pages to satisfy the request. alloc_pages() can only
4450 * allocate memory in power-of-two pages.
4452 * This function is also limited by MAX_ORDER.
4454 * Memory allocated by this function must be released by free_pages_exact().
4456 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4458 unsigned int order
= get_order(size
);
4461 addr
= __get_free_pages(gfp_mask
, order
);
4462 return make_alloc_exact(addr
, order
, size
);
4464 EXPORT_SYMBOL(alloc_pages_exact
);
4467 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4469 * @nid: the preferred node ID where memory should be allocated
4470 * @size: the number of bytes to allocate
4471 * @gfp_mask: GFP flags for the allocation
4473 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4476 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4478 unsigned int order
= get_order(size
);
4479 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4482 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4486 * free_pages_exact - release memory allocated via alloc_pages_exact()
4487 * @virt: the value returned by alloc_pages_exact.
4488 * @size: size of allocation, same value as passed to alloc_pages_exact().
4490 * Release the memory allocated by a previous call to alloc_pages_exact.
4492 void free_pages_exact(void *virt
, size_t size
)
4494 unsigned long addr
= (unsigned long)virt
;
4495 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4497 while (addr
< end
) {
4502 EXPORT_SYMBOL(free_pages_exact
);
4505 * nr_free_zone_pages - count number of pages beyond high watermark
4506 * @offset: The zone index of the highest zone
4508 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4509 * high watermark within all zones at or below a given zone index. For each
4510 * zone, the number of pages is calculated as:
4512 * nr_free_zone_pages = managed_pages - high_pages
4514 static unsigned long nr_free_zone_pages(int offset
)
4519 /* Just pick one node, since fallback list is circular */
4520 unsigned long sum
= 0;
4522 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4524 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4525 unsigned long size
= zone
->managed_pages
;
4526 unsigned long high
= high_wmark_pages(zone
);
4535 * nr_free_buffer_pages - count number of pages beyond high watermark
4537 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4538 * watermark within ZONE_DMA and ZONE_NORMAL.
4540 unsigned long nr_free_buffer_pages(void)
4542 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4544 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4547 * nr_free_pagecache_pages - count number of pages beyond high watermark
4549 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4550 * high watermark within all zones.
4552 unsigned long nr_free_pagecache_pages(void)
4554 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4557 static inline void show_node(struct zone
*zone
)
4559 if (IS_ENABLED(CONFIG_NUMA
))
4560 printk("Node %d ", zone_to_nid(zone
));
4563 long si_mem_available(void)
4566 unsigned long pagecache
;
4567 unsigned long wmark_low
= 0;
4568 unsigned long pages
[NR_LRU_LISTS
];
4572 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4573 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4576 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4579 * Estimate the amount of memory available for userspace allocations,
4580 * without causing swapping.
4582 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4585 * Not all the page cache can be freed, otherwise the system will
4586 * start swapping. Assume at least half of the page cache, or the
4587 * low watermark worth of cache, needs to stay.
4589 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4590 pagecache
-= min(pagecache
/ 2, wmark_low
);
4591 available
+= pagecache
;
4594 * Part of the reclaimable slab consists of items that are in use,
4595 * and cannot be freed. Cap this estimate at the low watermark.
4597 available
+= global_node_page_state(NR_SLAB_RECLAIMABLE
) -
4598 min(global_node_page_state(NR_SLAB_RECLAIMABLE
) / 2,
4602 * Part of the kernel memory, which can be released under memory
4605 available
+= global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES
) >>
4612 EXPORT_SYMBOL_GPL(si_mem_available
);
4614 void si_meminfo(struct sysinfo
*val
)
4616 val
->totalram
= totalram_pages
;
4617 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4618 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
4619 val
->bufferram
= nr_blockdev_pages();
4620 val
->totalhigh
= totalhigh_pages
;
4621 val
->freehigh
= nr_free_highpages();
4622 val
->mem_unit
= PAGE_SIZE
;
4625 EXPORT_SYMBOL(si_meminfo
);
4628 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4630 int zone_type
; /* needs to be signed */
4631 unsigned long managed_pages
= 0;
4632 unsigned long managed_highpages
= 0;
4633 unsigned long free_highpages
= 0;
4634 pg_data_t
*pgdat
= NODE_DATA(nid
);
4636 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4637 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4638 val
->totalram
= managed_pages
;
4639 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4640 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4641 #ifdef CONFIG_HIGHMEM
4642 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4643 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4645 if (is_highmem(zone
)) {
4646 managed_highpages
+= zone
->managed_pages
;
4647 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4650 val
->totalhigh
= managed_highpages
;
4651 val
->freehigh
= free_highpages
;
4653 val
->totalhigh
= managed_highpages
;
4654 val
->freehigh
= free_highpages
;
4656 val
->mem_unit
= PAGE_SIZE
;
4661 * Determine whether the node should be displayed or not, depending on whether
4662 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4664 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4666 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4670 * no node mask - aka implicit memory numa policy. Do not bother with
4671 * the synchronization - read_mems_allowed_begin - because we do not
4672 * have to be precise here.
4675 nodemask
= &cpuset_current_mems_allowed
;
4677 return !node_isset(nid
, *nodemask
);
4680 #define K(x) ((x) << (PAGE_SHIFT-10))
4682 static void show_migration_types(unsigned char type
)
4684 static const char types
[MIGRATE_TYPES
] = {
4685 [MIGRATE_UNMOVABLE
] = 'U',
4686 [MIGRATE_MOVABLE
] = 'M',
4687 [MIGRATE_RECLAIMABLE
] = 'E',
4688 [MIGRATE_HIGHATOMIC
] = 'H',
4690 [MIGRATE_CMA
] = 'C',
4692 #ifdef CONFIG_MEMORY_ISOLATION
4693 [MIGRATE_ISOLATE
] = 'I',
4696 char tmp
[MIGRATE_TYPES
+ 1];
4700 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4701 if (type
& (1 << i
))
4706 printk(KERN_CONT
"(%s) ", tmp
);
4710 * Show free area list (used inside shift_scroll-lock stuff)
4711 * We also calculate the percentage fragmentation. We do this by counting the
4712 * memory on each free list with the exception of the first item on the list.
4715 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4718 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4720 unsigned long free_pcp
= 0;
4725 for_each_populated_zone(zone
) {
4726 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4729 for_each_online_cpu(cpu
)
4730 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4733 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4734 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4735 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4736 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4737 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4738 " free:%lu free_pcp:%lu free_cma:%lu\n",
4739 global_node_page_state(NR_ACTIVE_ANON
),
4740 global_node_page_state(NR_INACTIVE_ANON
),
4741 global_node_page_state(NR_ISOLATED_ANON
),
4742 global_node_page_state(NR_ACTIVE_FILE
),
4743 global_node_page_state(NR_INACTIVE_FILE
),
4744 global_node_page_state(NR_ISOLATED_FILE
),
4745 global_node_page_state(NR_UNEVICTABLE
),
4746 global_node_page_state(NR_FILE_DIRTY
),
4747 global_node_page_state(NR_WRITEBACK
),
4748 global_node_page_state(NR_UNSTABLE_NFS
),
4749 global_node_page_state(NR_SLAB_RECLAIMABLE
),
4750 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
4751 global_node_page_state(NR_FILE_MAPPED
),
4752 global_node_page_state(NR_SHMEM
),
4753 global_zone_page_state(NR_PAGETABLE
),
4754 global_zone_page_state(NR_BOUNCE
),
4755 global_zone_page_state(NR_FREE_PAGES
),
4757 global_zone_page_state(NR_FREE_CMA_PAGES
));
4759 for_each_online_pgdat(pgdat
) {
4760 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
4764 " active_anon:%lukB"
4765 " inactive_anon:%lukB"
4766 " active_file:%lukB"
4767 " inactive_file:%lukB"
4768 " unevictable:%lukB"
4769 " isolated(anon):%lukB"
4770 " isolated(file):%lukB"
4775 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4777 " shmem_pmdmapped: %lukB"
4780 " writeback_tmp:%lukB"
4782 " all_unreclaimable? %s"
4785 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4786 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4787 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4788 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4789 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4790 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4791 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4792 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4793 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4794 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4795 K(node_page_state(pgdat
, NR_SHMEM
)),
4796 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4797 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4798 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4800 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4802 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4803 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4804 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
4808 for_each_populated_zone(zone
) {
4811 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4815 for_each_online_cpu(cpu
)
4816 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4825 " active_anon:%lukB"
4826 " inactive_anon:%lukB"
4827 " active_file:%lukB"
4828 " inactive_file:%lukB"
4829 " unevictable:%lukB"
4830 " writepending:%lukB"
4834 " kernel_stack:%lukB"
4842 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4843 K(min_wmark_pages(zone
)),
4844 K(low_wmark_pages(zone
)),
4845 K(high_wmark_pages(zone
)),
4846 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4847 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4848 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4849 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4850 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4851 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4852 K(zone
->present_pages
),
4853 K(zone
->managed_pages
),
4854 K(zone_page_state(zone
, NR_MLOCK
)),
4855 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4856 K(zone_page_state(zone
, NR_PAGETABLE
)),
4857 K(zone_page_state(zone
, NR_BOUNCE
)),
4859 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4860 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4861 printk("lowmem_reserve[]:");
4862 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4863 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
4864 printk(KERN_CONT
"\n");
4867 for_each_populated_zone(zone
) {
4869 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4870 unsigned char types
[MAX_ORDER
];
4872 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4875 printk(KERN_CONT
"%s: ", zone
->name
);
4877 spin_lock_irqsave(&zone
->lock
, flags
);
4878 for (order
= 0; order
< MAX_ORDER
; order
++) {
4879 struct free_area
*area
= &zone
->free_area
[order
];
4882 nr
[order
] = area
->nr_free
;
4883 total
+= nr
[order
] << order
;
4886 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4887 if (!list_empty(&area
->free_list
[type
]))
4888 types
[order
] |= 1 << type
;
4891 spin_unlock_irqrestore(&zone
->lock
, flags
);
4892 for (order
= 0; order
< MAX_ORDER
; order
++) {
4893 printk(KERN_CONT
"%lu*%lukB ",
4894 nr
[order
], K(1UL) << order
);
4896 show_migration_types(types
[order
]);
4898 printk(KERN_CONT
"= %lukB\n", K(total
));
4901 hugetlb_show_meminfo();
4903 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4905 show_swap_cache_info();
4908 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4910 zoneref
->zone
= zone
;
4911 zoneref
->zone_idx
= zone_idx(zone
);
4915 * Builds allocation fallback zone lists.
4917 * Add all populated zones of a node to the zonelist.
4919 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
4922 enum zone_type zone_type
= MAX_NR_ZONES
;
4927 zone
= pgdat
->node_zones
+ zone_type
;
4928 if (managed_zone(zone
)) {
4929 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
4930 check_highest_zone(zone_type
);
4932 } while (zone_type
);
4939 static int __parse_numa_zonelist_order(char *s
)
4942 * We used to support different zonlists modes but they turned
4943 * out to be just not useful. Let's keep the warning in place
4944 * if somebody still use the cmd line parameter so that we do
4945 * not fail it silently
4947 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
4948 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
4954 static __init
int setup_numa_zonelist_order(char *s
)
4959 return __parse_numa_zonelist_order(s
);
4961 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4963 char numa_zonelist_order
[] = "Node";
4966 * sysctl handler for numa_zonelist_order
4968 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4969 void __user
*buffer
, size_t *length
,
4976 return proc_dostring(table
, write
, buffer
, length
, ppos
);
4977 str
= memdup_user_nul(buffer
, 16);
4979 return PTR_ERR(str
);
4981 ret
= __parse_numa_zonelist_order(str
);
4987 #define MAX_NODE_LOAD (nr_online_nodes)
4988 static int node_load
[MAX_NUMNODES
];
4991 * find_next_best_node - find the next node that should appear in a given node's fallback list
4992 * @node: node whose fallback list we're appending
4993 * @used_node_mask: nodemask_t of already used nodes
4995 * We use a number of factors to determine which is the next node that should
4996 * appear on a given node's fallback list. The node should not have appeared
4997 * already in @node's fallback list, and it should be the next closest node
4998 * according to the distance array (which contains arbitrary distance values
4999 * from each node to each node in the system), and should also prefer nodes
5000 * with no CPUs, since presumably they'll have very little allocation pressure
5001 * on them otherwise.
5002 * It returns -1 if no node is found.
5004 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5007 int min_val
= INT_MAX
;
5008 int best_node
= NUMA_NO_NODE
;
5009 const struct cpumask
*tmp
= cpumask_of_node(0);
5011 /* Use the local node if we haven't already */
5012 if (!node_isset(node
, *used_node_mask
)) {
5013 node_set(node
, *used_node_mask
);
5017 for_each_node_state(n
, N_MEMORY
) {
5019 /* Don't want a node to appear more than once */
5020 if (node_isset(n
, *used_node_mask
))
5023 /* Use the distance array to find the distance */
5024 val
= node_distance(node
, n
);
5026 /* Penalize nodes under us ("prefer the next node") */
5029 /* Give preference to headless and unused nodes */
5030 tmp
= cpumask_of_node(n
);
5031 if (!cpumask_empty(tmp
))
5032 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5034 /* Slight preference for less loaded node */
5035 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5036 val
+= node_load
[n
];
5038 if (val
< min_val
) {
5045 node_set(best_node
, *used_node_mask
);
5052 * Build zonelists ordered by node and zones within node.
5053 * This results in maximum locality--normal zone overflows into local
5054 * DMA zone, if any--but risks exhausting DMA zone.
5056 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5059 struct zoneref
*zonerefs
;
5062 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5064 for (i
= 0; i
< nr_nodes
; i
++) {
5067 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5069 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5070 zonerefs
+= nr_zones
;
5072 zonerefs
->zone
= NULL
;
5073 zonerefs
->zone_idx
= 0;
5077 * Build gfp_thisnode zonelists
5079 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5081 struct zoneref
*zonerefs
;
5084 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5085 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5086 zonerefs
+= nr_zones
;
5087 zonerefs
->zone
= NULL
;
5088 zonerefs
->zone_idx
= 0;
5092 * Build zonelists ordered by zone and nodes within zones.
5093 * This results in conserving DMA zone[s] until all Normal memory is
5094 * exhausted, but results in overflowing to remote node while memory
5095 * may still exist in local DMA zone.
5098 static void build_zonelists(pg_data_t
*pgdat
)
5100 static int node_order
[MAX_NUMNODES
];
5101 int node
, load
, nr_nodes
= 0;
5102 nodemask_t used_mask
;
5103 int local_node
, prev_node
;
5105 /* NUMA-aware ordering of nodes */
5106 local_node
= pgdat
->node_id
;
5107 load
= nr_online_nodes
;
5108 prev_node
= local_node
;
5109 nodes_clear(used_mask
);
5111 memset(node_order
, 0, sizeof(node_order
));
5112 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5114 * We don't want to pressure a particular node.
5115 * So adding penalty to the first node in same
5116 * distance group to make it round-robin.
5118 if (node_distance(local_node
, node
) !=
5119 node_distance(local_node
, prev_node
))
5120 node_load
[node
] = load
;
5122 node_order
[nr_nodes
++] = node
;
5127 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5128 build_thisnode_zonelists(pgdat
);
5131 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5133 * Return node id of node used for "local" allocations.
5134 * I.e., first node id of first zone in arg node's generic zonelist.
5135 * Used for initializing percpu 'numa_mem', which is used primarily
5136 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5138 int local_memory_node(int node
)
5142 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5143 gfp_zone(GFP_KERNEL
),
5145 return z
->zone
->node
;
5149 static void setup_min_unmapped_ratio(void);
5150 static void setup_min_slab_ratio(void);
5151 #else /* CONFIG_NUMA */
5153 static void build_zonelists(pg_data_t
*pgdat
)
5155 int node
, local_node
;
5156 struct zoneref
*zonerefs
;
5159 local_node
= pgdat
->node_id
;
5161 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5162 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5163 zonerefs
+= nr_zones
;
5166 * Now we build the zonelist so that it contains the zones
5167 * of all the other nodes.
5168 * We don't want to pressure a particular node, so when
5169 * building the zones for node N, we make sure that the
5170 * zones coming right after the local ones are those from
5171 * node N+1 (modulo N)
5173 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5174 if (!node_online(node
))
5176 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5177 zonerefs
+= nr_zones
;
5179 for (node
= 0; node
< local_node
; node
++) {
5180 if (!node_online(node
))
5182 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5183 zonerefs
+= nr_zones
;
5186 zonerefs
->zone
= NULL
;
5187 zonerefs
->zone_idx
= 0;
5190 #endif /* CONFIG_NUMA */
5193 * Boot pageset table. One per cpu which is going to be used for all
5194 * zones and all nodes. The parameters will be set in such a way
5195 * that an item put on a list will immediately be handed over to
5196 * the buddy list. This is safe since pageset manipulation is done
5197 * with interrupts disabled.
5199 * The boot_pagesets must be kept even after bootup is complete for
5200 * unused processors and/or zones. They do play a role for bootstrapping
5201 * hotplugged processors.
5203 * zoneinfo_show() and maybe other functions do
5204 * not check if the processor is online before following the pageset pointer.
5205 * Other parts of the kernel may not check if the zone is available.
5207 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5208 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5209 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5211 static void __build_all_zonelists(void *data
)
5214 int __maybe_unused cpu
;
5215 pg_data_t
*self
= data
;
5216 static DEFINE_SPINLOCK(lock
);
5221 memset(node_load
, 0, sizeof(node_load
));
5225 * This node is hotadded and no memory is yet present. So just
5226 * building zonelists is fine - no need to touch other nodes.
5228 if (self
&& !node_online(self
->node_id
)) {
5229 build_zonelists(self
);
5231 for_each_online_node(nid
) {
5232 pg_data_t
*pgdat
= NODE_DATA(nid
);
5234 build_zonelists(pgdat
);
5237 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5239 * We now know the "local memory node" for each node--
5240 * i.e., the node of the first zone in the generic zonelist.
5241 * Set up numa_mem percpu variable for on-line cpus. During
5242 * boot, only the boot cpu should be on-line; we'll init the
5243 * secondary cpus' numa_mem as they come on-line. During
5244 * node/memory hotplug, we'll fixup all on-line cpus.
5246 for_each_online_cpu(cpu
)
5247 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5254 static noinline
void __init
5255 build_all_zonelists_init(void)
5259 __build_all_zonelists(NULL
);
5262 * Initialize the boot_pagesets that are going to be used
5263 * for bootstrapping processors. The real pagesets for
5264 * each zone will be allocated later when the per cpu
5265 * allocator is available.
5267 * boot_pagesets are used also for bootstrapping offline
5268 * cpus if the system is already booted because the pagesets
5269 * are needed to initialize allocators on a specific cpu too.
5270 * F.e. the percpu allocator needs the page allocator which
5271 * needs the percpu allocator in order to allocate its pagesets
5272 * (a chicken-egg dilemma).
5274 for_each_possible_cpu(cpu
)
5275 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5277 mminit_verify_zonelist();
5278 cpuset_init_current_mems_allowed();
5282 * unless system_state == SYSTEM_BOOTING.
5284 * __ref due to call of __init annotated helper build_all_zonelists_init
5285 * [protected by SYSTEM_BOOTING].
5287 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5289 if (system_state
== SYSTEM_BOOTING
) {
5290 build_all_zonelists_init();
5292 __build_all_zonelists(pgdat
);
5293 /* cpuset refresh routine should be here */
5295 vm_total_pages
= nr_free_pagecache_pages();
5297 * Disable grouping by mobility if the number of pages in the
5298 * system is too low to allow the mechanism to work. It would be
5299 * more accurate, but expensive to check per-zone. This check is
5300 * made on memory-hotadd so a system can start with mobility
5301 * disabled and enable it later
5303 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5304 page_group_by_mobility_disabled
= 1;
5306 page_group_by_mobility_disabled
= 0;
5308 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5310 page_group_by_mobility_disabled
? "off" : "on",
5313 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5318 * Initially all pages are reserved - free ones are freed
5319 * up by free_all_bootmem() once the early boot process is
5320 * done. Non-atomic initialization, single-pass.
5322 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5323 unsigned long start_pfn
, enum memmap_context context
)
5325 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5326 unsigned long end_pfn
= start_pfn
+ size
;
5327 pg_data_t
*pgdat
= NODE_DATA(nid
);
5329 unsigned long nr_initialised
= 0;
5330 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5331 struct memblock_region
*r
= NULL
, *tmp
;
5334 if (highest_memmap_pfn
< end_pfn
- 1)
5335 highest_memmap_pfn
= end_pfn
- 1;
5338 * Honor reservation requested by the driver for this ZONE_DEVICE
5341 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5342 start_pfn
+= altmap
->reserve
;
5344 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5346 * There can be holes in boot-time mem_map[]s handed to this
5347 * function. They do not exist on hotplugged memory.
5349 if (context
!= MEMMAP_EARLY
)
5352 if (!early_pfn_valid(pfn
))
5354 if (!early_pfn_in_nid(pfn
, nid
))
5356 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5359 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5361 * Check given memblock attribute by firmware which can affect
5362 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5363 * mirrored, it's an overlapped memmap init. skip it.
5365 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5366 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5367 for_each_memblock(memory
, tmp
)
5368 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5372 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5373 memblock_is_mirror(r
)) {
5374 /* already initialized as NORMAL */
5375 pfn
= memblock_region_memory_end_pfn(r
);
5383 * Mark the block movable so that blocks are reserved for
5384 * movable at startup. This will force kernel allocations
5385 * to reserve their blocks rather than leaking throughout
5386 * the address space during boot when many long-lived
5387 * kernel allocations are made.
5389 * bitmap is created for zone's valid pfn range. but memmap
5390 * can be created for invalid pages (for alignment)
5391 * check here not to call set_pageblock_migratetype() against
5394 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5395 struct page
*page
= pfn_to_page(pfn
);
5397 __init_single_page(page
, pfn
, zone
, nid
);
5398 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5401 __init_single_pfn(pfn
, zone
, nid
);
5406 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5408 unsigned int order
, t
;
5409 for_each_migratetype_order(order
, t
) {
5410 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5411 zone
->free_area
[order
].nr_free
= 0;
5415 #ifndef __HAVE_ARCH_MEMMAP_INIT
5416 #define memmap_init(size, nid, zone, start_pfn) \
5417 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5420 static int zone_batchsize(struct zone
*zone
)
5426 * The per-cpu-pages pools are set to around 1000th of the
5427 * size of the zone. But no more than 1/2 of a meg.
5429 * OK, so we don't know how big the cache is. So guess.
5431 batch
= zone
->managed_pages
/ 1024;
5432 if (batch
* PAGE_SIZE
> 512 * 1024)
5433 batch
= (512 * 1024) / PAGE_SIZE
;
5434 batch
/= 4; /* We effectively *= 4 below */
5439 * Clamp the batch to a 2^n - 1 value. Having a power
5440 * of 2 value was found to be more likely to have
5441 * suboptimal cache aliasing properties in some cases.
5443 * For example if 2 tasks are alternately allocating
5444 * batches of pages, one task can end up with a lot
5445 * of pages of one half of the possible page colors
5446 * and the other with pages of the other colors.
5448 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5453 /* The deferral and batching of frees should be suppressed under NOMMU
5456 * The problem is that NOMMU needs to be able to allocate large chunks
5457 * of contiguous memory as there's no hardware page translation to
5458 * assemble apparent contiguous memory from discontiguous pages.
5460 * Queueing large contiguous runs of pages for batching, however,
5461 * causes the pages to actually be freed in smaller chunks. As there
5462 * can be a significant delay between the individual batches being
5463 * recycled, this leads to the once large chunks of space being
5464 * fragmented and becoming unavailable for high-order allocations.
5471 * pcp->high and pcp->batch values are related and dependent on one another:
5472 * ->batch must never be higher then ->high.
5473 * The following function updates them in a safe manner without read side
5476 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5477 * those fields changing asynchronously (acording the the above rule).
5479 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5480 * outside of boot time (or some other assurance that no concurrent updaters
5483 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5484 unsigned long batch
)
5486 /* start with a fail safe value for batch */
5490 /* Update high, then batch, in order */
5497 /* a companion to pageset_set_high() */
5498 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5500 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5503 static void pageset_init(struct per_cpu_pageset
*p
)
5505 struct per_cpu_pages
*pcp
;
5508 memset(p
, 0, sizeof(*p
));
5512 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5513 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5516 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5519 pageset_set_batch(p
, batch
);
5523 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5524 * to the value high for the pageset p.
5526 static void pageset_set_high(struct per_cpu_pageset
*p
,
5529 unsigned long batch
= max(1UL, high
/ 4);
5530 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5531 batch
= PAGE_SHIFT
* 8;
5533 pageset_update(&p
->pcp
, high
, batch
);
5536 static void pageset_set_high_and_batch(struct zone
*zone
,
5537 struct per_cpu_pageset
*pcp
)
5539 if (percpu_pagelist_fraction
)
5540 pageset_set_high(pcp
,
5541 (zone
->managed_pages
/
5542 percpu_pagelist_fraction
));
5544 pageset_set_batch(pcp
, zone_batchsize(zone
));
5547 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5549 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5552 pageset_set_high_and_batch(zone
, pcp
);
5555 void __meminit
setup_zone_pageset(struct zone
*zone
)
5558 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5559 for_each_possible_cpu(cpu
)
5560 zone_pageset_init(zone
, cpu
);
5564 * Allocate per cpu pagesets and initialize them.
5565 * Before this call only boot pagesets were available.
5567 void __init
setup_per_cpu_pageset(void)
5569 struct pglist_data
*pgdat
;
5572 for_each_populated_zone(zone
)
5573 setup_zone_pageset(zone
);
5575 for_each_online_pgdat(pgdat
)
5576 pgdat
->per_cpu_nodestats
=
5577 alloc_percpu(struct per_cpu_nodestat
);
5580 static __meminit
void zone_pcp_init(struct zone
*zone
)
5583 * per cpu subsystem is not up at this point. The following code
5584 * relies on the ability of the linker to provide the
5585 * offset of a (static) per cpu variable into the per cpu area.
5587 zone
->pageset
= &boot_pageset
;
5589 if (populated_zone(zone
))
5590 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5591 zone
->name
, zone
->present_pages
,
5592 zone_batchsize(zone
));
5595 void __meminit
init_currently_empty_zone(struct zone
*zone
,
5596 unsigned long zone_start_pfn
,
5599 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5600 int zone_idx
= zone_idx(zone
) + 1;
5602 if (zone_idx
> pgdat
->nr_zones
)
5603 pgdat
->nr_zones
= zone_idx
;
5605 zone
->zone_start_pfn
= zone_start_pfn
;
5607 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5608 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5610 (unsigned long)zone_idx(zone
),
5611 zone_start_pfn
, (zone_start_pfn
+ size
));
5613 zone_init_free_lists(zone
);
5614 zone
->initialized
= 1;
5617 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5618 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5621 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5623 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5624 struct mminit_pfnnid_cache
*state
)
5626 unsigned long start_pfn
, end_pfn
;
5629 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5630 return state
->last_nid
;
5632 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5634 state
->last_start
= start_pfn
;
5635 state
->last_end
= end_pfn
;
5636 state
->last_nid
= nid
;
5641 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5644 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5645 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5646 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5648 * If an architecture guarantees that all ranges registered contain no holes
5649 * and may be freed, this this function may be used instead of calling
5650 * memblock_free_early_nid() manually.
5652 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5654 unsigned long start_pfn
, end_pfn
;
5657 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5658 start_pfn
= min(start_pfn
, max_low_pfn
);
5659 end_pfn
= min(end_pfn
, max_low_pfn
);
5661 if (start_pfn
< end_pfn
)
5662 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5663 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5669 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5670 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5672 * If an architecture guarantees that all ranges registered contain no holes and may
5673 * be freed, this function may be used instead of calling memory_present() manually.
5675 void __init
sparse_memory_present_with_active_regions(int nid
)
5677 unsigned long start_pfn
, end_pfn
;
5680 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5681 memory_present(this_nid
, start_pfn
, end_pfn
);
5685 * get_pfn_range_for_nid - Return the start and end page frames for a node
5686 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5687 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5688 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5690 * It returns the start and end page frame of a node based on information
5691 * provided by memblock_set_node(). If called for a node
5692 * with no available memory, a warning is printed and the start and end
5695 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5696 unsigned long *start_pfn
, unsigned long *end_pfn
)
5698 unsigned long this_start_pfn
, this_end_pfn
;
5704 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5705 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5706 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5709 if (*start_pfn
== -1UL)
5714 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5715 * assumption is made that zones within a node are ordered in monotonic
5716 * increasing memory addresses so that the "highest" populated zone is used
5718 static void __init
find_usable_zone_for_movable(void)
5721 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5722 if (zone_index
== ZONE_MOVABLE
)
5725 if (arch_zone_highest_possible_pfn
[zone_index
] >
5726 arch_zone_lowest_possible_pfn
[zone_index
])
5730 VM_BUG_ON(zone_index
== -1);
5731 movable_zone
= zone_index
;
5735 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5736 * because it is sized independent of architecture. Unlike the other zones,
5737 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5738 * in each node depending on the size of each node and how evenly kernelcore
5739 * is distributed. This helper function adjusts the zone ranges
5740 * provided by the architecture for a given node by using the end of the
5741 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5742 * zones within a node are in order of monotonic increases memory addresses
5744 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5745 unsigned long zone_type
,
5746 unsigned long node_start_pfn
,
5747 unsigned long node_end_pfn
,
5748 unsigned long *zone_start_pfn
,
5749 unsigned long *zone_end_pfn
)
5751 /* Only adjust if ZONE_MOVABLE is on this node */
5752 if (zone_movable_pfn
[nid
]) {
5753 /* Size ZONE_MOVABLE */
5754 if (zone_type
== ZONE_MOVABLE
) {
5755 *zone_start_pfn
= zone_movable_pfn
[nid
];
5756 *zone_end_pfn
= min(node_end_pfn
,
5757 arch_zone_highest_possible_pfn
[movable_zone
]);
5759 /* Adjust for ZONE_MOVABLE starting within this range */
5760 } else if (!mirrored_kernelcore
&&
5761 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5762 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5763 *zone_end_pfn
= zone_movable_pfn
[nid
];
5765 /* Check if this whole range is within ZONE_MOVABLE */
5766 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5767 *zone_start_pfn
= *zone_end_pfn
;
5772 * Return the number of pages a zone spans in a node, including holes
5773 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5775 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5776 unsigned long zone_type
,
5777 unsigned long node_start_pfn
,
5778 unsigned long node_end_pfn
,
5779 unsigned long *zone_start_pfn
,
5780 unsigned long *zone_end_pfn
,
5781 unsigned long *ignored
)
5783 /* When hotadd a new node from cpu_up(), the node should be empty */
5784 if (!node_start_pfn
&& !node_end_pfn
)
5787 /* Get the start and end of the zone */
5788 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5789 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5790 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5791 node_start_pfn
, node_end_pfn
,
5792 zone_start_pfn
, zone_end_pfn
);
5794 /* Check that this node has pages within the zone's required range */
5795 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5798 /* Move the zone boundaries inside the node if necessary */
5799 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5800 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5802 /* Return the spanned pages */
5803 return *zone_end_pfn
- *zone_start_pfn
;
5807 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5808 * then all holes in the requested range will be accounted for.
5810 unsigned long __meminit
__absent_pages_in_range(int nid
,
5811 unsigned long range_start_pfn
,
5812 unsigned long range_end_pfn
)
5814 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5815 unsigned long start_pfn
, end_pfn
;
5818 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5819 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5820 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5821 nr_absent
-= end_pfn
- start_pfn
;
5827 * absent_pages_in_range - Return number of page frames in holes within a range
5828 * @start_pfn: The start PFN to start searching for holes
5829 * @end_pfn: The end PFN to stop searching for holes
5831 * It returns the number of pages frames in memory holes within a range.
5833 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5834 unsigned long end_pfn
)
5836 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5839 /* Return the number of page frames in holes in a zone on a node */
5840 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5841 unsigned long zone_type
,
5842 unsigned long node_start_pfn
,
5843 unsigned long node_end_pfn
,
5844 unsigned long *ignored
)
5846 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5847 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5848 unsigned long zone_start_pfn
, zone_end_pfn
;
5849 unsigned long nr_absent
;
5851 /* When hotadd a new node from cpu_up(), the node should be empty */
5852 if (!node_start_pfn
&& !node_end_pfn
)
5855 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5856 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5858 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5859 node_start_pfn
, node_end_pfn
,
5860 &zone_start_pfn
, &zone_end_pfn
);
5861 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5864 * ZONE_MOVABLE handling.
5865 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5868 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5869 unsigned long start_pfn
, end_pfn
;
5870 struct memblock_region
*r
;
5872 for_each_memblock(memory
, r
) {
5873 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5874 zone_start_pfn
, zone_end_pfn
);
5875 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5876 zone_start_pfn
, zone_end_pfn
);
5878 if (zone_type
== ZONE_MOVABLE
&&
5879 memblock_is_mirror(r
))
5880 nr_absent
+= end_pfn
- start_pfn
;
5882 if (zone_type
== ZONE_NORMAL
&&
5883 !memblock_is_mirror(r
))
5884 nr_absent
+= end_pfn
- start_pfn
;
5891 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5892 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5893 unsigned long zone_type
,
5894 unsigned long node_start_pfn
,
5895 unsigned long node_end_pfn
,
5896 unsigned long *zone_start_pfn
,
5897 unsigned long *zone_end_pfn
,
5898 unsigned long *zones_size
)
5902 *zone_start_pfn
= node_start_pfn
;
5903 for (zone
= 0; zone
< zone_type
; zone
++)
5904 *zone_start_pfn
+= zones_size
[zone
];
5906 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5908 return zones_size
[zone_type
];
5911 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5912 unsigned long zone_type
,
5913 unsigned long node_start_pfn
,
5914 unsigned long node_end_pfn
,
5915 unsigned long *zholes_size
)
5920 return zholes_size
[zone_type
];
5923 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5925 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5926 unsigned long node_start_pfn
,
5927 unsigned long node_end_pfn
,
5928 unsigned long *zones_size
,
5929 unsigned long *zholes_size
)
5931 unsigned long realtotalpages
= 0, totalpages
= 0;
5934 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5935 struct zone
*zone
= pgdat
->node_zones
+ i
;
5936 unsigned long zone_start_pfn
, zone_end_pfn
;
5937 unsigned long size
, real_size
;
5939 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5945 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5946 node_start_pfn
, node_end_pfn
,
5949 zone
->zone_start_pfn
= zone_start_pfn
;
5951 zone
->zone_start_pfn
= 0;
5952 zone
->spanned_pages
= size
;
5953 zone
->present_pages
= real_size
;
5956 realtotalpages
+= real_size
;
5959 pgdat
->node_spanned_pages
= totalpages
;
5960 pgdat
->node_present_pages
= realtotalpages
;
5961 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5965 #ifndef CONFIG_SPARSEMEM
5967 * Calculate the size of the zone->blockflags rounded to an unsigned long
5968 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5969 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5970 * round what is now in bits to nearest long in bits, then return it in
5973 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5975 unsigned long usemapsize
;
5977 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5978 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5979 usemapsize
= usemapsize
>> pageblock_order
;
5980 usemapsize
*= NR_PAGEBLOCK_BITS
;
5981 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5983 return usemapsize
/ 8;
5986 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5988 unsigned long zone_start_pfn
,
5989 unsigned long zonesize
)
5991 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5992 zone
->pageblock_flags
= NULL
;
5994 zone
->pageblock_flags
=
5995 memblock_virt_alloc_node_nopanic(usemapsize
,
5999 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6000 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6001 #endif /* CONFIG_SPARSEMEM */
6003 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6005 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6006 void __paginginit
set_pageblock_order(void)
6010 /* Check that pageblock_nr_pages has not already been setup */
6011 if (pageblock_order
)
6014 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6015 order
= HUGETLB_PAGE_ORDER
;
6017 order
= MAX_ORDER
- 1;
6020 * Assume the largest contiguous order of interest is a huge page.
6021 * This value may be variable depending on boot parameters on IA64 and
6024 pageblock_order
= order
;
6026 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6029 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6030 * is unused as pageblock_order is set at compile-time. See
6031 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6034 void __paginginit
set_pageblock_order(void)
6038 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6040 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
6041 unsigned long present_pages
)
6043 unsigned long pages
= spanned_pages
;
6046 * Provide a more accurate estimation if there are holes within
6047 * the zone and SPARSEMEM is in use. If there are holes within the
6048 * zone, each populated memory region may cost us one or two extra
6049 * memmap pages due to alignment because memmap pages for each
6050 * populated regions may not be naturally aligned on page boundary.
6051 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6053 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6054 IS_ENABLED(CONFIG_SPARSEMEM
))
6055 pages
= present_pages
;
6057 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6061 * Set up the zone data structures:
6062 * - mark all pages reserved
6063 * - mark all memory queues empty
6064 * - clear the memory bitmaps
6066 * NOTE: pgdat should get zeroed by caller.
6068 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
6071 int nid
= pgdat
->node_id
;
6073 pgdat_resize_init(pgdat
);
6074 #ifdef CONFIG_NUMA_BALANCING
6075 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
6076 pgdat
->numabalancing_migrate_nr_pages
= 0;
6077 pgdat
->numabalancing_migrate_next_window
= jiffies
;
6079 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6080 spin_lock_init(&pgdat
->split_queue_lock
);
6081 INIT_LIST_HEAD(&pgdat
->split_queue
);
6082 pgdat
->split_queue_len
= 0;
6084 init_waitqueue_head(&pgdat
->kswapd_wait
);
6085 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6086 #ifdef CONFIG_COMPACTION
6087 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6089 pgdat_page_ext_init(pgdat
);
6090 spin_lock_init(&pgdat
->lru_lock
);
6091 lruvec_init(node_lruvec(pgdat
));
6093 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6095 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6096 struct zone
*zone
= pgdat
->node_zones
+ j
;
6097 unsigned long size
, realsize
, freesize
, memmap_pages
;
6098 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6100 size
= zone
->spanned_pages
;
6101 realsize
= freesize
= zone
->present_pages
;
6104 * Adjust freesize so that it accounts for how much memory
6105 * is used by this zone for memmap. This affects the watermark
6106 * and per-cpu initialisations
6108 memmap_pages
= calc_memmap_size(size
, realsize
);
6109 if (!is_highmem_idx(j
)) {
6110 if (freesize
>= memmap_pages
) {
6111 freesize
-= memmap_pages
;
6114 " %s zone: %lu pages used for memmap\n",
6115 zone_names
[j
], memmap_pages
);
6117 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6118 zone_names
[j
], memmap_pages
, freesize
);
6121 /* Account for reserved pages */
6122 if (j
== 0 && freesize
> dma_reserve
) {
6123 freesize
-= dma_reserve
;
6124 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6125 zone_names
[0], dma_reserve
);
6128 if (!is_highmem_idx(j
))
6129 nr_kernel_pages
+= freesize
;
6130 /* Charge for highmem memmap if there are enough kernel pages */
6131 else if (nr_kernel_pages
> memmap_pages
* 2)
6132 nr_kernel_pages
-= memmap_pages
;
6133 nr_all_pages
+= freesize
;
6136 * Set an approximate value for lowmem here, it will be adjusted
6137 * when the bootmem allocator frees pages into the buddy system.
6138 * And all highmem pages will be managed by the buddy system.
6140 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
6144 zone
->name
= zone_names
[j
];
6145 zone
->zone_pgdat
= pgdat
;
6146 spin_lock_init(&zone
->lock
);
6147 zone_seqlock_init(zone
);
6148 zone_pcp_init(zone
);
6153 set_pageblock_order();
6154 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6155 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6156 memmap_init(size
, nid
, j
, zone_start_pfn
);
6160 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6162 unsigned long __maybe_unused start
= 0;
6163 unsigned long __maybe_unused offset
= 0;
6165 /* Skip empty nodes */
6166 if (!pgdat
->node_spanned_pages
)
6169 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6170 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6171 offset
= pgdat
->node_start_pfn
- start
;
6172 /* ia64 gets its own node_mem_map, before this, without bootmem */
6173 if (!pgdat
->node_mem_map
) {
6174 unsigned long size
, end
;
6178 * The zone's endpoints aren't required to be MAX_ORDER
6179 * aligned but the node_mem_map endpoints must be in order
6180 * for the buddy allocator to function correctly.
6182 end
= pgdat_end_pfn(pgdat
);
6183 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6184 size
= (end
- start
) * sizeof(struct page
);
6185 map
= alloc_remap(pgdat
->node_id
, size
);
6187 map
= memblock_virt_alloc_node_nopanic(size
,
6189 pgdat
->node_mem_map
= map
+ offset
;
6191 #ifndef CONFIG_NEED_MULTIPLE_NODES
6193 * With no DISCONTIG, the global mem_map is just set as node 0's
6195 if (pgdat
== NODE_DATA(0)) {
6196 mem_map
= NODE_DATA(0)->node_mem_map
;
6197 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6198 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6200 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6203 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6206 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
6207 unsigned long node_start_pfn
, unsigned long *zholes_size
)
6209 pg_data_t
*pgdat
= NODE_DATA(nid
);
6210 unsigned long start_pfn
= 0;
6211 unsigned long end_pfn
= 0;
6213 /* pg_data_t should be reset to zero when it's allocated */
6214 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6216 pgdat
->node_id
= nid
;
6217 pgdat
->node_start_pfn
= node_start_pfn
;
6218 pgdat
->per_cpu_nodestats
= NULL
;
6219 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6220 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6221 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6222 (u64
)start_pfn
<< PAGE_SHIFT
,
6223 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6225 start_pfn
= node_start_pfn
;
6227 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6228 zones_size
, zholes_size
);
6230 alloc_node_mem_map(pgdat
);
6231 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6232 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6233 nid
, (unsigned long)pgdat
,
6234 (unsigned long)pgdat
->node_mem_map
);
6237 reset_deferred_meminit(pgdat
);
6238 free_area_init_core(pgdat
);
6241 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6243 #if MAX_NUMNODES > 1
6245 * Figure out the number of possible node ids.
6247 void __init
setup_nr_node_ids(void)
6249 unsigned int highest
;
6251 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6252 nr_node_ids
= highest
+ 1;
6257 * node_map_pfn_alignment - determine the maximum internode alignment
6259 * This function should be called after node map is populated and sorted.
6260 * It calculates the maximum power of two alignment which can distinguish
6263 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6264 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6265 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6266 * shifted, 1GiB is enough and this function will indicate so.
6268 * This is used to test whether pfn -> nid mapping of the chosen memory
6269 * model has fine enough granularity to avoid incorrect mapping for the
6270 * populated node map.
6272 * Returns the determined alignment in pfn's. 0 if there is no alignment
6273 * requirement (single node).
6275 unsigned long __init
node_map_pfn_alignment(void)
6277 unsigned long accl_mask
= 0, last_end
= 0;
6278 unsigned long start
, end
, mask
;
6282 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6283 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6290 * Start with a mask granular enough to pin-point to the
6291 * start pfn and tick off bits one-by-one until it becomes
6292 * too coarse to separate the current node from the last.
6294 mask
= ~((1 << __ffs(start
)) - 1);
6295 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6298 /* accumulate all internode masks */
6302 /* convert mask to number of pages */
6303 return ~accl_mask
+ 1;
6306 /* Find the lowest pfn for a node */
6307 static unsigned long __init
find_min_pfn_for_node(int nid
)
6309 unsigned long min_pfn
= ULONG_MAX
;
6310 unsigned long start_pfn
;
6313 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6314 min_pfn
= min(min_pfn
, start_pfn
);
6316 if (min_pfn
== ULONG_MAX
) {
6317 pr_warn("Could not find start_pfn for node %d\n", nid
);
6325 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6327 * It returns the minimum PFN based on information provided via
6328 * memblock_set_node().
6330 unsigned long __init
find_min_pfn_with_active_regions(void)
6332 return find_min_pfn_for_node(MAX_NUMNODES
);
6336 * early_calculate_totalpages()
6337 * Sum pages in active regions for movable zone.
6338 * Populate N_MEMORY for calculating usable_nodes.
6340 static unsigned long __init
early_calculate_totalpages(void)
6342 unsigned long totalpages
= 0;
6343 unsigned long start_pfn
, end_pfn
;
6346 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6347 unsigned long pages
= end_pfn
- start_pfn
;
6349 totalpages
+= pages
;
6351 node_set_state(nid
, N_MEMORY
);
6357 * Find the PFN the Movable zone begins in each node. Kernel memory
6358 * is spread evenly between nodes as long as the nodes have enough
6359 * memory. When they don't, some nodes will have more kernelcore than
6362 static void __init
find_zone_movable_pfns_for_nodes(void)
6365 unsigned long usable_startpfn
;
6366 unsigned long kernelcore_node
, kernelcore_remaining
;
6367 /* save the state before borrow the nodemask */
6368 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6369 unsigned long totalpages
= early_calculate_totalpages();
6370 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6371 struct memblock_region
*r
;
6373 /* Need to find movable_zone earlier when movable_node is specified. */
6374 find_usable_zone_for_movable();
6377 * If movable_node is specified, ignore kernelcore and movablecore
6380 if (movable_node_is_enabled()) {
6381 for_each_memblock(memory
, r
) {
6382 if (!memblock_is_hotpluggable(r
))
6387 usable_startpfn
= PFN_DOWN(r
->base
);
6388 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6389 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6397 * If kernelcore=mirror is specified, ignore movablecore option
6399 if (mirrored_kernelcore
) {
6400 bool mem_below_4gb_not_mirrored
= false;
6402 for_each_memblock(memory
, r
) {
6403 if (memblock_is_mirror(r
))
6408 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6410 if (usable_startpfn
< 0x100000) {
6411 mem_below_4gb_not_mirrored
= true;
6415 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6416 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6420 if (mem_below_4gb_not_mirrored
)
6421 pr_warn("This configuration results in unmirrored kernel memory.");
6427 * If movablecore=nn[KMG] was specified, calculate what size of
6428 * kernelcore that corresponds so that memory usable for
6429 * any allocation type is evenly spread. If both kernelcore
6430 * and movablecore are specified, then the value of kernelcore
6431 * will be used for required_kernelcore if it's greater than
6432 * what movablecore would have allowed.
6434 if (required_movablecore
) {
6435 unsigned long corepages
;
6438 * Round-up so that ZONE_MOVABLE is at least as large as what
6439 * was requested by the user
6441 required_movablecore
=
6442 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6443 required_movablecore
= min(totalpages
, required_movablecore
);
6444 corepages
= totalpages
- required_movablecore
;
6446 required_kernelcore
= max(required_kernelcore
, corepages
);
6450 * If kernelcore was not specified or kernelcore size is larger
6451 * than totalpages, there is no ZONE_MOVABLE.
6453 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6456 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6457 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6460 /* Spread kernelcore memory as evenly as possible throughout nodes */
6461 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6462 for_each_node_state(nid
, N_MEMORY
) {
6463 unsigned long start_pfn
, end_pfn
;
6466 * Recalculate kernelcore_node if the division per node
6467 * now exceeds what is necessary to satisfy the requested
6468 * amount of memory for the kernel
6470 if (required_kernelcore
< kernelcore_node
)
6471 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6474 * As the map is walked, we track how much memory is usable
6475 * by the kernel using kernelcore_remaining. When it is
6476 * 0, the rest of the node is usable by ZONE_MOVABLE
6478 kernelcore_remaining
= kernelcore_node
;
6480 /* Go through each range of PFNs within this node */
6481 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6482 unsigned long size_pages
;
6484 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6485 if (start_pfn
>= end_pfn
)
6488 /* Account for what is only usable for kernelcore */
6489 if (start_pfn
< usable_startpfn
) {
6490 unsigned long kernel_pages
;
6491 kernel_pages
= min(end_pfn
, usable_startpfn
)
6494 kernelcore_remaining
-= min(kernel_pages
,
6495 kernelcore_remaining
);
6496 required_kernelcore
-= min(kernel_pages
,
6497 required_kernelcore
);
6499 /* Continue if range is now fully accounted */
6500 if (end_pfn
<= usable_startpfn
) {
6503 * Push zone_movable_pfn to the end so
6504 * that if we have to rebalance
6505 * kernelcore across nodes, we will
6506 * not double account here
6508 zone_movable_pfn
[nid
] = end_pfn
;
6511 start_pfn
= usable_startpfn
;
6515 * The usable PFN range for ZONE_MOVABLE is from
6516 * start_pfn->end_pfn. Calculate size_pages as the
6517 * number of pages used as kernelcore
6519 size_pages
= end_pfn
- start_pfn
;
6520 if (size_pages
> kernelcore_remaining
)
6521 size_pages
= kernelcore_remaining
;
6522 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6525 * Some kernelcore has been met, update counts and
6526 * break if the kernelcore for this node has been
6529 required_kernelcore
-= min(required_kernelcore
,
6531 kernelcore_remaining
-= size_pages
;
6532 if (!kernelcore_remaining
)
6538 * If there is still required_kernelcore, we do another pass with one
6539 * less node in the count. This will push zone_movable_pfn[nid] further
6540 * along on the nodes that still have memory until kernelcore is
6544 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6548 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6549 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6550 zone_movable_pfn
[nid
] =
6551 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6554 /* restore the node_state */
6555 node_states
[N_MEMORY
] = saved_node_state
;
6558 /* Any regular or high memory on that node ? */
6559 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6561 enum zone_type zone_type
;
6563 if (N_MEMORY
== N_NORMAL_MEMORY
)
6566 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6567 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6568 if (populated_zone(zone
)) {
6569 node_set_state(nid
, N_HIGH_MEMORY
);
6570 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6571 zone_type
<= ZONE_NORMAL
)
6572 node_set_state(nid
, N_NORMAL_MEMORY
);
6579 * free_area_init_nodes - Initialise all pg_data_t and zone data
6580 * @max_zone_pfn: an array of max PFNs for each zone
6582 * This will call free_area_init_node() for each active node in the system.
6583 * Using the page ranges provided by memblock_set_node(), the size of each
6584 * zone in each node and their holes is calculated. If the maximum PFN
6585 * between two adjacent zones match, it is assumed that the zone is empty.
6586 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6587 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6588 * starts where the previous one ended. For example, ZONE_DMA32 starts
6589 * at arch_max_dma_pfn.
6591 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6593 unsigned long start_pfn
, end_pfn
;
6596 /* Record where the zone boundaries are */
6597 memset(arch_zone_lowest_possible_pfn
, 0,
6598 sizeof(arch_zone_lowest_possible_pfn
));
6599 memset(arch_zone_highest_possible_pfn
, 0,
6600 sizeof(arch_zone_highest_possible_pfn
));
6602 start_pfn
= find_min_pfn_with_active_regions();
6604 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6605 if (i
== ZONE_MOVABLE
)
6608 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6609 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6610 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6612 start_pfn
= end_pfn
;
6615 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6616 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6617 find_zone_movable_pfns_for_nodes();
6619 /* Print out the zone ranges */
6620 pr_info("Zone ranges:\n");
6621 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6622 if (i
== ZONE_MOVABLE
)
6624 pr_info(" %-8s ", zone_names
[i
]);
6625 if (arch_zone_lowest_possible_pfn
[i
] ==
6626 arch_zone_highest_possible_pfn
[i
])
6629 pr_cont("[mem %#018Lx-%#018Lx]\n",
6630 (u64
)arch_zone_lowest_possible_pfn
[i
]
6632 ((u64
)arch_zone_highest_possible_pfn
[i
]
6633 << PAGE_SHIFT
) - 1);
6636 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6637 pr_info("Movable zone start for each node\n");
6638 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6639 if (zone_movable_pfn
[i
])
6640 pr_info(" Node %d: %#018Lx\n", i
,
6641 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6644 /* Print out the early node map */
6645 pr_info("Early memory node ranges\n");
6646 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6647 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6648 (u64
)start_pfn
<< PAGE_SHIFT
,
6649 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6651 /* Initialise every node */
6652 mminit_verify_pageflags_layout();
6653 setup_nr_node_ids();
6654 for_each_online_node(nid
) {
6655 pg_data_t
*pgdat
= NODE_DATA(nid
);
6656 free_area_init_node(nid
, NULL
,
6657 find_min_pfn_for_node(nid
), NULL
);
6659 /* Any memory on that node */
6660 if (pgdat
->node_present_pages
)
6661 node_set_state(nid
, N_MEMORY
);
6662 check_for_memory(pgdat
, nid
);
6666 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6668 unsigned long long coremem
;
6672 coremem
= memparse(p
, &p
);
6673 *core
= coremem
>> PAGE_SHIFT
;
6675 /* Paranoid check that UL is enough for the coremem value */
6676 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6682 * kernelcore=size sets the amount of memory for use for allocations that
6683 * cannot be reclaimed or migrated.
6685 static int __init
cmdline_parse_kernelcore(char *p
)
6687 /* parse kernelcore=mirror */
6688 if (parse_option_str(p
, "mirror")) {
6689 mirrored_kernelcore
= true;
6693 return cmdline_parse_core(p
, &required_kernelcore
);
6697 * movablecore=size sets the amount of memory for use for allocations that
6698 * can be reclaimed or migrated.
6700 static int __init
cmdline_parse_movablecore(char *p
)
6702 return cmdline_parse_core(p
, &required_movablecore
);
6705 early_param("kernelcore", cmdline_parse_kernelcore
);
6706 early_param("movablecore", cmdline_parse_movablecore
);
6708 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6710 void adjust_managed_page_count(struct page
*page
, long count
)
6712 spin_lock(&managed_page_count_lock
);
6713 page_zone(page
)->managed_pages
+= count
;
6714 totalram_pages
+= count
;
6715 #ifdef CONFIG_HIGHMEM
6716 if (PageHighMem(page
))
6717 totalhigh_pages
+= count
;
6719 spin_unlock(&managed_page_count_lock
);
6721 EXPORT_SYMBOL(adjust_managed_page_count
);
6723 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6726 unsigned long pages
= 0;
6728 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6729 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6730 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6731 if ((unsigned int)poison
<= 0xFF)
6732 memset(pos
, poison
, PAGE_SIZE
);
6733 free_reserved_page(virt_to_page(pos
));
6737 pr_info("Freeing %s memory: %ldK\n",
6738 s
, pages
<< (PAGE_SHIFT
- 10));
6742 EXPORT_SYMBOL(free_reserved_area
);
6744 #ifdef CONFIG_HIGHMEM
6745 void free_highmem_page(struct page
*page
)
6747 __free_reserved_page(page
);
6749 page_zone(page
)->managed_pages
++;
6755 void __init
mem_init_print_info(const char *str
)
6757 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6758 unsigned long init_code_size
, init_data_size
;
6760 physpages
= get_num_physpages();
6761 codesize
= _etext
- _stext
;
6762 datasize
= _edata
- _sdata
;
6763 rosize
= __end_rodata
- __start_rodata
;
6764 bss_size
= __bss_stop
- __bss_start
;
6765 init_data_size
= __init_end
- __init_begin
;
6766 init_code_size
= _einittext
- _sinittext
;
6769 * Detect special cases and adjust section sizes accordingly:
6770 * 1) .init.* may be embedded into .data sections
6771 * 2) .init.text.* may be out of [__init_begin, __init_end],
6772 * please refer to arch/tile/kernel/vmlinux.lds.S.
6773 * 3) .rodata.* may be embedded into .text or .data sections.
6775 #define adj_init_size(start, end, size, pos, adj) \
6777 if (start <= pos && pos < end && size > adj) \
6781 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6782 _sinittext
, init_code_size
);
6783 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6784 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6785 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6786 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6788 #undef adj_init_size
6790 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6791 #ifdef CONFIG_HIGHMEM
6795 nr_free_pages() << (PAGE_SHIFT
- 10),
6796 physpages
<< (PAGE_SHIFT
- 10),
6797 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6798 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6799 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6800 totalcma_pages
<< (PAGE_SHIFT
- 10),
6801 #ifdef CONFIG_HIGHMEM
6802 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6804 str
? ", " : "", str
? str
: "");
6808 * set_dma_reserve - set the specified number of pages reserved in the first zone
6809 * @new_dma_reserve: The number of pages to mark reserved
6811 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6812 * In the DMA zone, a significant percentage may be consumed by kernel image
6813 * and other unfreeable allocations which can skew the watermarks badly. This
6814 * function may optionally be used to account for unfreeable pages in the
6815 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6816 * smaller per-cpu batchsize.
6818 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6820 dma_reserve
= new_dma_reserve
;
6823 void __init
free_area_init(unsigned long *zones_size
)
6825 free_area_init_node(0, zones_size
,
6826 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6829 static int page_alloc_cpu_dead(unsigned int cpu
)
6832 lru_add_drain_cpu(cpu
);
6836 * Spill the event counters of the dead processor
6837 * into the current processors event counters.
6838 * This artificially elevates the count of the current
6841 vm_events_fold_cpu(cpu
);
6844 * Zero the differential counters of the dead processor
6845 * so that the vm statistics are consistent.
6847 * This is only okay since the processor is dead and cannot
6848 * race with what we are doing.
6850 cpu_vm_stats_fold(cpu
);
6854 void __init
page_alloc_init(void)
6858 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
6859 "mm/page_alloc:dead", NULL
,
6860 page_alloc_cpu_dead
);
6865 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6866 * or min_free_kbytes changes.
6868 static void calculate_totalreserve_pages(void)
6870 struct pglist_data
*pgdat
;
6871 unsigned long reserve_pages
= 0;
6872 enum zone_type i
, j
;
6874 for_each_online_pgdat(pgdat
) {
6876 pgdat
->totalreserve_pages
= 0;
6878 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6879 struct zone
*zone
= pgdat
->node_zones
+ i
;
6882 /* Find valid and maximum lowmem_reserve in the zone */
6883 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6884 if (zone
->lowmem_reserve
[j
] > max
)
6885 max
= zone
->lowmem_reserve
[j
];
6888 /* we treat the high watermark as reserved pages. */
6889 max
+= high_wmark_pages(zone
);
6891 if (max
> zone
->managed_pages
)
6892 max
= zone
->managed_pages
;
6894 pgdat
->totalreserve_pages
+= max
;
6896 reserve_pages
+= max
;
6899 totalreserve_pages
= reserve_pages
;
6903 * setup_per_zone_lowmem_reserve - called whenever
6904 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6905 * has a correct pages reserved value, so an adequate number of
6906 * pages are left in the zone after a successful __alloc_pages().
6908 static void setup_per_zone_lowmem_reserve(void)
6910 struct pglist_data
*pgdat
;
6911 enum zone_type j
, idx
;
6913 for_each_online_pgdat(pgdat
) {
6914 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6915 struct zone
*zone
= pgdat
->node_zones
+ j
;
6916 unsigned long managed_pages
= zone
->managed_pages
;
6918 zone
->lowmem_reserve
[j
] = 0;
6922 struct zone
*lower_zone
;
6926 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6927 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6929 lower_zone
= pgdat
->node_zones
+ idx
;
6930 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6931 sysctl_lowmem_reserve_ratio
[idx
];
6932 managed_pages
+= lower_zone
->managed_pages
;
6937 /* update totalreserve_pages */
6938 calculate_totalreserve_pages();
6941 static void __setup_per_zone_wmarks(void)
6943 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6944 unsigned long pages_low
= extra_free_kbytes
>> (PAGE_SHIFT
- 10);
6945 unsigned long lowmem_pages
= 0;
6947 unsigned long flags
;
6949 /* Calculate total number of !ZONE_HIGHMEM pages */
6950 for_each_zone(zone
) {
6951 if (!is_highmem(zone
))
6952 lowmem_pages
+= zone
->managed_pages
;
6955 for_each_zone(zone
) {
6958 spin_lock_irqsave(&zone
->lock
, flags
);
6959 min
= (u64
)pages_min
* zone
->managed_pages
;
6960 do_div(min
, lowmem_pages
);
6961 low
= (u64
)pages_low
* zone
->managed_pages
;
6962 do_div(low
, vm_total_pages
);
6964 if (is_highmem(zone
)) {
6966 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6967 * need highmem pages, so cap pages_min to a small
6970 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6971 * deltas control asynch page reclaim, and so should
6972 * not be capped for highmem.
6974 unsigned long min_pages
;
6976 min_pages
= zone
->managed_pages
/ 1024;
6977 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6978 zone
->watermark
[WMARK_MIN
] = min_pages
;
6981 * If it's a lowmem zone, reserve a number of pages
6982 * proportionate to the zone's size.
6984 zone
->watermark
[WMARK_MIN
] = min
;
6988 * Set the kswapd watermarks distance according to the
6989 * scale factor in proportion to available memory, but
6990 * ensure a minimum size on small systems.
6992 min
= max_t(u64
, min
>> 2,
6993 mult_frac(zone
->managed_pages
,
6994 watermark_scale_factor
, 10000));
6996 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) +
6998 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) +
7001 spin_unlock_irqrestore(&zone
->lock
, flags
);
7004 /* update totalreserve_pages */
7005 calculate_totalreserve_pages();
7009 * setup_per_zone_wmarks - called when min_free_kbytes changes
7010 * or when memory is hot-{added|removed}
7012 * Ensures that the watermark[min,low,high] values for each zone are set
7013 * correctly with respect to min_free_kbytes.
7015 void setup_per_zone_wmarks(void)
7017 static DEFINE_SPINLOCK(lock
);
7020 __setup_per_zone_wmarks();
7025 * Initialise min_free_kbytes.
7027 * For small machines we want it small (128k min). For large machines
7028 * we want it large (64MB max). But it is not linear, because network
7029 * bandwidth does not increase linearly with machine size. We use
7031 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7032 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7048 int __meminit
init_per_zone_wmark_min(void)
7050 unsigned long lowmem_kbytes
;
7051 int new_min_free_kbytes
;
7053 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7054 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7056 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7057 min_free_kbytes
= new_min_free_kbytes
;
7058 if (min_free_kbytes
< 128)
7059 min_free_kbytes
= 128;
7060 if (min_free_kbytes
> 65536)
7061 min_free_kbytes
= 65536;
7063 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7064 new_min_free_kbytes
, user_min_free_kbytes
);
7066 setup_per_zone_wmarks();
7067 refresh_zone_stat_thresholds();
7068 setup_per_zone_lowmem_reserve();
7071 setup_min_unmapped_ratio();
7072 setup_min_slab_ratio();
7077 core_initcall(init_per_zone_wmark_min
)
7080 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7081 * that we can call two helper functions whenever min_free_kbytes
7082 * or extra_free_kbytes changes.
7084 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7085 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7089 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7094 user_min_free_kbytes
= min_free_kbytes
;
7095 setup_per_zone_wmarks();
7100 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7101 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7105 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7110 setup_per_zone_wmarks();
7116 static void setup_min_unmapped_ratio(void)
7121 for_each_online_pgdat(pgdat
)
7122 pgdat
->min_unmapped_pages
= 0;
7125 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
7126 sysctl_min_unmapped_ratio
) / 100;
7130 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7131 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7135 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7139 setup_min_unmapped_ratio();
7144 static void setup_min_slab_ratio(void)
7149 for_each_online_pgdat(pgdat
)
7150 pgdat
->min_slab_pages
= 0;
7153 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
7154 sysctl_min_slab_ratio
) / 100;
7157 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7158 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7162 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7166 setup_min_slab_ratio();
7173 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7174 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7175 * whenever sysctl_lowmem_reserve_ratio changes.
7177 * The reserve ratio obviously has absolutely no relation with the
7178 * minimum watermarks. The lowmem reserve ratio can only make sense
7179 * if in function of the boot time zone sizes.
7181 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7182 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7184 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7185 setup_per_zone_lowmem_reserve();
7190 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7191 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7192 * pagelist can have before it gets flushed back to buddy allocator.
7194 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7195 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7198 int old_percpu_pagelist_fraction
;
7201 mutex_lock(&pcp_batch_high_lock
);
7202 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7204 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7205 if (!write
|| ret
< 0)
7208 /* Sanity checking to avoid pcp imbalance */
7209 if (percpu_pagelist_fraction
&&
7210 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7211 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7217 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7220 for_each_populated_zone(zone
) {
7223 for_each_possible_cpu(cpu
)
7224 pageset_set_high_and_batch(zone
,
7225 per_cpu_ptr(zone
->pageset
, cpu
));
7228 mutex_unlock(&pcp_batch_high_lock
);
7233 int hashdist
= HASHDIST_DEFAULT
;
7235 static int __init
set_hashdist(char *str
)
7239 hashdist
= simple_strtoul(str
, &str
, 0);
7242 __setup("hashdist=", set_hashdist
);
7245 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7247 * Returns the number of pages that arch has reserved but
7248 * is not known to alloc_large_system_hash().
7250 static unsigned long __init
arch_reserved_kernel_pages(void)
7257 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7258 * machines. As memory size is increased the scale is also increased but at
7259 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7260 * quadruples the scale is increased by one, which means the size of hash table
7261 * only doubles, instead of quadrupling as well.
7262 * Because 32-bit systems cannot have large physical memory, where this scaling
7263 * makes sense, it is disabled on such platforms.
7265 #if __BITS_PER_LONG > 32
7266 #define ADAPT_SCALE_BASE (64ul << 30)
7267 #define ADAPT_SCALE_SHIFT 2
7268 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7272 * allocate a large system hash table from bootmem
7273 * - it is assumed that the hash table must contain an exact power-of-2
7274 * quantity of entries
7275 * - limit is the number of hash buckets, not the total allocation size
7277 void *__init
alloc_large_system_hash(const char *tablename
,
7278 unsigned long bucketsize
,
7279 unsigned long numentries
,
7282 unsigned int *_hash_shift
,
7283 unsigned int *_hash_mask
,
7284 unsigned long low_limit
,
7285 unsigned long high_limit
)
7287 unsigned long long max
= high_limit
;
7288 unsigned long log2qty
, size
;
7292 /* allow the kernel cmdline to have a say */
7294 /* round applicable memory size up to nearest megabyte */
7295 numentries
= nr_kernel_pages
;
7296 numentries
-= arch_reserved_kernel_pages();
7298 /* It isn't necessary when PAGE_SIZE >= 1MB */
7299 if (PAGE_SHIFT
< 20)
7300 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7302 #if __BITS_PER_LONG > 32
7304 unsigned long adapt
;
7306 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
7307 adapt
<<= ADAPT_SCALE_SHIFT
)
7312 /* limit to 1 bucket per 2^scale bytes of low memory */
7313 if (scale
> PAGE_SHIFT
)
7314 numentries
>>= (scale
- PAGE_SHIFT
);
7316 numentries
<<= (PAGE_SHIFT
- scale
);
7318 /* Make sure we've got at least a 0-order allocation.. */
7319 if (unlikely(flags
& HASH_SMALL
)) {
7320 /* Makes no sense without HASH_EARLY */
7321 WARN_ON(!(flags
& HASH_EARLY
));
7322 if (!(numentries
>> *_hash_shift
)) {
7323 numentries
= 1UL << *_hash_shift
;
7324 BUG_ON(!numentries
);
7326 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7327 numentries
= PAGE_SIZE
/ bucketsize
;
7329 numentries
= roundup_pow_of_two(numentries
);
7331 /* limit allocation size to 1/16 total memory by default */
7333 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7334 do_div(max
, bucketsize
);
7336 max
= min(max
, 0x80000000ULL
);
7338 if (numentries
< low_limit
)
7339 numentries
= low_limit
;
7340 if (numentries
> max
)
7343 log2qty
= ilog2(numentries
);
7346 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7347 * currently not used when HASH_EARLY is specified.
7349 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
7351 size
= bucketsize
<< log2qty
;
7352 if (flags
& HASH_EARLY
)
7353 table
= memblock_virt_alloc_nopanic(size
, 0);
7355 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
7358 * If bucketsize is not a power-of-two, we may free
7359 * some pages at the end of hash table which
7360 * alloc_pages_exact() automatically does
7362 if (get_order(size
) < MAX_ORDER
) {
7363 table
= alloc_pages_exact(size
, gfp_flags
);
7364 kmemleak_alloc(table
, size
, 1, gfp_flags
);
7367 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7370 panic("Failed to allocate %s hash table\n", tablename
);
7372 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7373 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7376 *_hash_shift
= log2qty
;
7378 *_hash_mask
= (1 << log2qty
) - 1;
7384 * This function checks whether pageblock includes unmovable pages or not.
7385 * If @count is not zero, it is okay to include less @count unmovable pages
7387 * PageLRU check without isolation or lru_lock could race so that
7388 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7389 * check without lock_page also may miss some movable non-lru pages at
7390 * race condition. So you can't expect this function should be exact.
7392 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7393 bool skip_hwpoisoned_pages
)
7395 unsigned long pfn
, iter
, found
;
7399 * For avoiding noise data, lru_add_drain_all() should be called
7400 * If ZONE_MOVABLE, the zone never contains unmovable pages
7402 if (zone_idx(zone
) == ZONE_MOVABLE
)
7404 mt
= get_pageblock_migratetype(page
);
7405 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7408 pfn
= page_to_pfn(page
);
7409 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7410 unsigned long check
= pfn
+ iter
;
7412 if (!pfn_valid_within(check
))
7415 page
= pfn_to_page(check
);
7418 * Hugepages are not in LRU lists, but they're movable.
7419 * We need not scan over tail pages bacause we don't
7420 * handle each tail page individually in migration.
7422 if (PageHuge(page
)) {
7423 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7428 * We can't use page_count without pin a page
7429 * because another CPU can free compound page.
7430 * This check already skips compound tails of THP
7431 * because their page->_refcount is zero at all time.
7433 if (!page_ref_count(page
)) {
7434 if (PageBuddy(page
))
7435 iter
+= (1 << page_order(page
)) - 1;
7440 * The HWPoisoned page may be not in buddy system, and
7441 * page_count() is not 0.
7443 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7446 if (__PageMovable(page
))
7452 * If there are RECLAIMABLE pages, we need to check
7453 * it. But now, memory offline itself doesn't call
7454 * shrink_node_slabs() and it still to be fixed.
7457 * If the page is not RAM, page_count()should be 0.
7458 * we don't need more check. This is an _used_ not-movable page.
7460 * The problematic thing here is PG_reserved pages. PG_reserved
7461 * is set to both of a memory hole page and a _used_ kernel
7470 bool is_pageblock_removable_nolock(struct page
*page
)
7476 * We have to be careful here because we are iterating over memory
7477 * sections which are not zone aware so we might end up outside of
7478 * the zone but still within the section.
7479 * We have to take care about the node as well. If the node is offline
7480 * its NODE_DATA will be NULL - see page_zone.
7482 if (!node_online(page_to_nid(page
)))
7485 zone
= page_zone(page
);
7486 pfn
= page_to_pfn(page
);
7487 if (!zone_spans_pfn(zone
, pfn
))
7490 return !has_unmovable_pages(zone
, page
, 0, true);
7493 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7495 static unsigned long pfn_max_align_down(unsigned long pfn
)
7497 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7498 pageblock_nr_pages
) - 1);
7501 static unsigned long pfn_max_align_up(unsigned long pfn
)
7503 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7504 pageblock_nr_pages
));
7507 /* [start, end) must belong to a single zone. */
7508 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7509 unsigned long start
, unsigned long end
,
7512 /* This function is based on compact_zone() from compaction.c. */
7513 unsigned long nr_reclaimed
;
7514 unsigned long pfn
= start
;
7515 unsigned int tries
= 0;
7521 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7522 if (fatal_signal_pending(current
)) {
7527 if (list_empty(&cc
->migratepages
)) {
7528 cc
->nr_migratepages
= 0;
7529 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7535 } else if (++tries
== 5) {
7536 ret
= ret
< 0 ? ret
: -EBUSY
;
7540 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7542 cc
->nr_migratepages
-= nr_reclaimed
;
7544 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7545 NULL
, 0, cc
->mode
, drain
? MR_CMA
: MR_HPA
);
7548 putback_movable_pages(&cc
->migratepages
);
7555 * alloc_contig_range() -- tries to allocate given range of pages
7556 * @start: start PFN to allocate
7557 * @end: one-past-the-last PFN to allocate
7558 * @migratetype: migratetype of the underlaying pageblocks (either
7559 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7560 * in range must have the same migratetype and it must
7561 * be either of the two.
7562 * @gfp_mask: GFP mask to use during compaction
7564 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7565 * aligned, however it's the caller's responsibility to guarantee that
7566 * we are the only thread that changes migrate type of pageblocks the
7569 * The PFN range must belong to a single zone.
7571 * Returns zero on success or negative error code. On success all
7572 * pages which PFN is in [start, end) are allocated for the caller and
7573 * need to be freed with free_contig_range().
7575 int __alloc_contig_range(unsigned long start
, unsigned long end
,
7576 unsigned migratetype
, gfp_t gfp_mask
, bool drain
)
7578 unsigned long outer_start
, outer_end
;
7582 struct compact_control cc
= {
7583 .nr_migratepages
= 0,
7585 .zone
= page_zone(pfn_to_page(start
)),
7586 .mode
= MIGRATE_SYNC
,
7587 .ignore_skip_hint
= true,
7588 .gfp_mask
= current_gfp_context(gfp_mask
),
7590 INIT_LIST_HEAD(&cc
.migratepages
);
7593 * What we do here is we mark all pageblocks in range as
7594 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7595 * have different sizes, and due to the way page allocator
7596 * work, we align the range to biggest of the two pages so
7597 * that page allocator won't try to merge buddies from
7598 * different pageblocks and change MIGRATE_ISOLATE to some
7599 * other migration type.
7601 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7602 * migrate the pages from an unaligned range (ie. pages that
7603 * we are interested in). This will put all the pages in
7604 * range back to page allocator as MIGRATE_ISOLATE.
7606 * When this is done, we take the pages in range from page
7607 * allocator removing them from the buddy system. This way
7608 * page allocator will never consider using them.
7610 * This lets us mark the pageblocks back as
7611 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7612 * aligned range but not in the unaligned, original range are
7613 * put back to page allocator so that buddy can use them.
7616 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7617 pfn_max_align_up(end
), migratetype
,
7623 * In case of -EBUSY, we'd like to know which page causes problem.
7624 * So, just fall through. test_pages_isolated() has a tracepoint
7625 * which will report the busy page.
7627 * It is possible that busy pages could become available before
7628 * the call to test_pages_isolated, and the range will actually be
7629 * allocated. So, if we fall through be sure to clear ret so that
7630 * -EBUSY is not accidentally used or returned to caller.
7632 ret
= __alloc_contig_migrate_range(&cc
, start
, end
, drain
);
7633 if (ret
&& ret
!= -EBUSY
)
7638 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7639 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7640 * more, all pages in [start, end) are free in page allocator.
7641 * What we are going to do is to allocate all pages from
7642 * [start, end) (that is remove them from page allocator).
7644 * The only problem is that pages at the beginning and at the
7645 * end of interesting range may be not aligned with pages that
7646 * page allocator holds, ie. they can be part of higher order
7647 * pages. Because of this, we reserve the bigger range and
7648 * once this is done free the pages we are not interested in.
7650 * We don't have to hold zone->lock here because the pages are
7651 * isolated thus they won't get removed from buddy.
7655 outer_start
= start
;
7658 lru_add_drain_all();
7659 drain_all_pages(cc
.zone
);
7661 while (!PageBuddy(pfn_to_page(outer_start
))) {
7662 if (++order
>= MAX_ORDER
) {
7663 outer_start
= start
;
7666 outer_start
&= ~0UL << order
;
7669 if (outer_start
!= start
) {
7670 order
= page_order(pfn_to_page(outer_start
));
7673 * outer_start page could be small order buddy page and
7674 * it doesn't include start page. Adjust outer_start
7675 * in this case to report failed page properly
7676 * on tracepoint in test_pages_isolated()
7678 if (outer_start
+ (1UL << order
) <= start
)
7679 outer_start
= start
;
7682 /* Make sure the range is really isolated. */
7683 if (test_pages_isolated(outer_start
, end
, false)) {
7684 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7685 __func__
, outer_start
, end
);
7691 /* Grab isolated pages from freelists. */
7692 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7698 /* Free head and tail (if any) */
7699 if (start
!= outer_start
)
7700 free_contig_range(outer_start
, start
- outer_start
);
7701 if (end
!= outer_end
)
7702 free_contig_range(end
, outer_end
- end
);
7705 undo_isolate_page_range(pfn_max_align_down(start
),
7706 pfn_max_align_up(end
), migratetype
);
7710 int alloc_contig_range(unsigned long start
, unsigned long end
,
7711 unsigned migratetype
, gfp_t gfp_mask
)
7713 return __alloc_contig_range(start
, end
, migratetype
, gfp_mask
, true);
7716 int alloc_contig_range_fast(unsigned long start
, unsigned long end
,
7717 unsigned migratetype
)
7719 return __alloc_contig_range(start
, end
, migratetype
, GFP_KERNEL
, false);
7722 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7724 unsigned int count
= 0;
7726 for (; nr_pages
--; pfn
++) {
7727 struct page
*page
= pfn_to_page(pfn
);
7729 count
+= page_count(page
) != 1;
7732 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7736 #ifdef CONFIG_MEMORY_HOTPLUG
7738 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7739 * page high values need to be recalulated.
7741 void __meminit
zone_pcp_update(struct zone
*zone
)
7744 mutex_lock(&pcp_batch_high_lock
);
7745 for_each_possible_cpu(cpu
)
7746 pageset_set_high_and_batch(zone
,
7747 per_cpu_ptr(zone
->pageset
, cpu
));
7748 mutex_unlock(&pcp_batch_high_lock
);
7752 void zone_pcp_reset(struct zone
*zone
)
7754 unsigned long flags
;
7756 struct per_cpu_pageset
*pset
;
7758 /* avoid races with drain_pages() */
7759 local_irq_save(flags
);
7760 if (zone
->pageset
!= &boot_pageset
) {
7761 for_each_online_cpu(cpu
) {
7762 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7763 drain_zonestat(zone
, pset
);
7765 free_percpu(zone
->pageset
);
7766 zone
->pageset
= &boot_pageset
;
7768 local_irq_restore(flags
);
7771 #ifdef CONFIG_MEMORY_HOTREMOVE
7773 * All pages in the range must be in a single zone and isolated
7774 * before calling this.
7777 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7781 unsigned int order
, i
;
7783 unsigned long flags
;
7784 /* find the first valid pfn */
7785 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7790 offline_mem_sections(pfn
, end_pfn
);
7791 zone
= page_zone(pfn_to_page(pfn
));
7792 spin_lock_irqsave(&zone
->lock
, flags
);
7794 while (pfn
< end_pfn
) {
7795 if (!pfn_valid(pfn
)) {
7799 page
= pfn_to_page(pfn
);
7801 * The HWPoisoned page may be not in buddy system, and
7802 * page_count() is not 0.
7804 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7806 SetPageReserved(page
);
7810 BUG_ON(page_count(page
));
7811 BUG_ON(!PageBuddy(page
));
7812 order
= page_order(page
);
7813 #ifdef CONFIG_DEBUG_VM
7814 pr_info("remove from free list %lx %d %lx\n",
7815 pfn
, 1 << order
, end_pfn
);
7817 list_del(&page
->lru
);
7818 rmv_page_order(page
);
7819 zone
->free_area
[order
].nr_free
--;
7820 for (i
= 0; i
< (1 << order
); i
++)
7821 SetPageReserved((page
+i
));
7822 pfn
+= (1 << order
);
7824 spin_unlock_irqrestore(&zone
->lock
, flags
);
7828 bool is_free_buddy_page(struct page
*page
)
7830 struct zone
*zone
= page_zone(page
);
7831 unsigned long pfn
= page_to_pfn(page
);
7832 unsigned long flags
;
7835 spin_lock_irqsave(&zone
->lock
, flags
);
7836 for (order
= 0; order
< MAX_ORDER
; order
++) {
7837 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7839 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7842 spin_unlock_irqrestore(&zone
->lock
, flags
);
7844 return order
< MAX_ORDER
;