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/highmem.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.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/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/psi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock
);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node
);
82 EXPORT_PER_CPU_SYMBOL(numa_node
);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
96 int _node_numa_mem_
[MAX_NUMNODES
];
99 /* work_structs for global per-cpu drains */
100 DEFINE_MUTEX(pcpu_drain_mutex
);
101 DEFINE_PER_CPU(struct work_struct
, pcpu_drain
);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy
;
105 EXPORT_SYMBOL(latent_entropy
);
109 * Array of node states.
111 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
112 [N_POSSIBLE
] = NODE_MASK_ALL
,
113 [N_ONLINE
] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
119 [N_MEMORY
] = { { [0] = 1UL } },
120 [N_CPU
] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states
);
125 atomic_long_t _totalram_pages __read_mostly
;
126 EXPORT_SYMBOL(_totalram_pages
);
127 unsigned long totalreserve_pages __read_mostly
;
128 unsigned long totalcma_pages __read_mostly
;
130 int percpu_pagelist_fraction
;
131 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
134 * A cached value of the page's pageblock's migratetype, used when the page is
135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
136 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
137 * Also the migratetype set in the page does not necessarily match the pcplist
138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
139 * other index - this ensures that it will be put on the correct CMA freelist.
141 static inline int get_pcppage_migratetype(struct page
*page
)
146 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
148 page
->index
= migratetype
;
151 #ifdef CONFIG_PM_SLEEP
153 * The following functions are used by the suspend/hibernate code to temporarily
154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
155 * while devices are suspended. To avoid races with the suspend/hibernate code,
156 * they should always be called with system_transition_mutex held
157 * (gfp_allowed_mask also should only be modified with system_transition_mutex
158 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
159 * with that modification).
162 static gfp_t saved_gfp_mask
;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
167 if (saved_gfp_mask
) {
168 gfp_allowed_mask
= saved_gfp_mask
;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
176 WARN_ON(saved_gfp_mask
);
177 saved_gfp_mask
= gfp_allowed_mask
;
178 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly
;
193 static void __free_pages_ok(struct page
*page
, unsigned int order
);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
214 #ifdef CONFIG_HIGHMEM
220 EXPORT_SYMBOL(totalram_pages
);
222 static char * const zone_names
[MAX_NR_ZONES
] = {
223 #ifdef CONFIG_ZONE_DMA
226 #ifdef CONFIG_ZONE_DMA32
230 #ifdef CONFIG_HIGHMEM
234 #ifdef CONFIG_ZONE_DEVICE
239 const char * const migratetype_names
[MIGRATE_TYPES
] = {
247 #ifdef CONFIG_MEMORY_ISOLATION
252 compound_page_dtor
* const compound_page_dtors
[] = {
255 #ifdef CONFIG_HUGETLB_PAGE
258 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
263 int min_free_kbytes
= 1024;
264 int user_min_free_kbytes
= -1;
265 int watermark_boost_factor __read_mostly
= 15000;
266 int watermark_scale_factor
= 10;
268 static unsigned long nr_kernel_pages __meminitdata
;
269 static unsigned long nr_all_pages __meminitdata
;
270 static unsigned long dma_reserve __meminitdata
;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __meminitdata
;
274 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __meminitdata
;
275 static unsigned long required_kernelcore __initdata
;
276 static unsigned long required_kernelcore_percent __initdata
;
277 static unsigned long required_movablecore __initdata
;
278 static unsigned long required_movablecore_percent __initdata
;
279 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __meminitdata
;
280 static bool mirrored_kernelcore __meminitdata
;
282 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
284 EXPORT_SYMBOL(movable_zone
);
285 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
288 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
289 int nr_online_nodes __read_mostly
= 1;
290 EXPORT_SYMBOL(nr_node_ids
);
291 EXPORT_SYMBOL(nr_online_nodes
);
294 int page_group_by_mobility_disabled __read_mostly
;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297 /* Returns true if the struct page for the pfn is uninitialised */
298 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
300 int nid
= early_pfn_to_nid(pfn
);
302 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
309 * Returns true when the remaining initialisation should be deferred until
310 * later in the boot cycle when it can be parallelised.
312 static bool __meminit
313 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
315 static unsigned long prev_end_pfn
, nr_initialised
;
318 * prev_end_pfn static that contains the end of previous zone
319 * No need to protect because called very early in boot before smp_init.
321 if (prev_end_pfn
!= end_pfn
) {
322 prev_end_pfn
= end_pfn
;
326 /* Always populate low zones for address-constrained allocations */
327 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
330 if ((nr_initialised
> NODE_DATA(nid
)->static_init_pgcnt
) &&
331 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
332 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
338 static inline bool early_page_uninitialised(unsigned long pfn
)
343 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
349 /* Return a pointer to the bitmap storing bits affecting a block of pages */
350 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
353 #ifdef CONFIG_SPARSEMEM
354 return __pfn_to_section(pfn
)->pageblock_flags
;
356 return page_zone(page
)->pageblock_flags
;
357 #endif /* CONFIG_SPARSEMEM */
360 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
362 #ifdef CONFIG_SPARSEMEM
363 pfn
&= (PAGES_PER_SECTION
-1);
364 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
366 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
367 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
368 #endif /* CONFIG_SPARSEMEM */
372 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
373 * @page: The page within the block of interest
374 * @pfn: The target page frame number
375 * @end_bitidx: The last bit of interest to retrieve
376 * @mask: mask of bits that the caller is interested in
378 * Return: pageblock_bits flags
380 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
382 unsigned long end_bitidx
,
385 unsigned long *bitmap
;
386 unsigned long bitidx
, word_bitidx
;
389 bitmap
= get_pageblock_bitmap(page
, pfn
);
390 bitidx
= pfn_to_bitidx(page
, pfn
);
391 word_bitidx
= bitidx
/ BITS_PER_LONG
;
392 bitidx
&= (BITS_PER_LONG
-1);
394 word
= bitmap
[word_bitidx
];
395 bitidx
+= end_bitidx
;
396 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
399 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
400 unsigned long end_bitidx
,
403 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
406 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
408 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
412 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
413 * @page: The page within the block of interest
414 * @flags: The flags to set
415 * @pfn: The target page frame number
416 * @end_bitidx: The last bit of interest
417 * @mask: mask of bits that the caller is interested in
419 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
421 unsigned long end_bitidx
,
424 unsigned long *bitmap
;
425 unsigned long bitidx
, word_bitidx
;
426 unsigned long old_word
, word
;
428 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
430 bitmap
= get_pageblock_bitmap(page
, pfn
);
431 bitidx
= pfn_to_bitidx(page
, pfn
);
432 word_bitidx
= bitidx
/ BITS_PER_LONG
;
433 bitidx
&= (BITS_PER_LONG
-1);
435 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
437 bitidx
+= end_bitidx
;
438 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
439 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
441 word
= READ_ONCE(bitmap
[word_bitidx
]);
443 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
444 if (word
== old_word
)
450 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
452 if (unlikely(page_group_by_mobility_disabled
&&
453 migratetype
< MIGRATE_PCPTYPES
))
454 migratetype
= MIGRATE_UNMOVABLE
;
456 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
457 PB_migrate
, PB_migrate_end
);
460 #ifdef CONFIG_DEBUG_VM
461 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
465 unsigned long pfn
= page_to_pfn(page
);
466 unsigned long sp
, start_pfn
;
469 seq
= zone_span_seqbegin(zone
);
470 start_pfn
= zone
->zone_start_pfn
;
471 sp
= zone
->spanned_pages
;
472 if (!zone_spans_pfn(zone
, pfn
))
474 } while (zone_span_seqretry(zone
, seq
));
477 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
478 pfn
, zone_to_nid(zone
), zone
->name
,
479 start_pfn
, start_pfn
+ sp
);
484 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
486 if (!pfn_valid_within(page_to_pfn(page
)))
488 if (zone
!= page_zone(page
))
494 * Temporary debugging check for pages not lying within a given zone.
496 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
498 if (page_outside_zone_boundaries(zone
, page
))
500 if (!page_is_consistent(zone
, page
))
506 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
512 static void bad_page(struct page
*page
, const char *reason
,
513 unsigned long bad_flags
)
515 static unsigned long resume
;
516 static unsigned long nr_shown
;
517 static unsigned long nr_unshown
;
520 * Allow a burst of 60 reports, then keep quiet for that minute;
521 * or allow a steady drip of one report per second.
523 if (nr_shown
== 60) {
524 if (time_before(jiffies
, resume
)) {
530 "BUG: Bad page state: %lu messages suppressed\n",
537 resume
= jiffies
+ 60 * HZ
;
539 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
540 current
->comm
, page_to_pfn(page
));
541 __dump_page(page
, reason
);
542 bad_flags
&= page
->flags
;
544 pr_alert("bad because of flags: %#lx(%pGp)\n",
545 bad_flags
, &bad_flags
);
546 dump_page_owner(page
);
551 /* Leave bad fields for debug, except PageBuddy could make trouble */
552 page_mapcount_reset(page
); /* remove PageBuddy */
553 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
557 * Higher-order pages are called "compound pages". They are structured thusly:
559 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
561 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
562 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
564 * The first tail page's ->compound_dtor holds the offset in array of compound
565 * page destructors. See compound_page_dtors.
567 * The first tail page's ->compound_order holds the order of allocation.
568 * This usage means that zero-order pages may not be compound.
571 void free_compound_page(struct page
*page
)
573 __free_pages_ok(page
, compound_order(page
));
576 void prep_compound_page(struct page
*page
, unsigned int order
)
579 int nr_pages
= 1 << order
;
581 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
582 set_compound_order(page
, order
);
584 for (i
= 1; i
< nr_pages
; i
++) {
585 struct page
*p
= page
+ i
;
586 set_page_count(p
, 0);
587 p
->mapping
= TAIL_MAPPING
;
588 set_compound_head(p
, page
);
590 atomic_set(compound_mapcount_ptr(page
), -1);
593 #ifdef CONFIG_DEBUG_PAGEALLOC
594 unsigned int _debug_guardpage_minorder
;
595 bool _debug_pagealloc_enabled __read_mostly
596 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
597 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
598 bool _debug_guardpage_enabled __read_mostly
;
600 static int __init
early_debug_pagealloc(char *buf
)
604 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
606 early_param("debug_pagealloc", early_debug_pagealloc
);
608 static bool need_debug_guardpage(void)
610 /* If we don't use debug_pagealloc, we don't need guard page */
611 if (!debug_pagealloc_enabled())
614 if (!debug_guardpage_minorder())
620 static void init_debug_guardpage(void)
622 if (!debug_pagealloc_enabled())
625 if (!debug_guardpage_minorder())
628 _debug_guardpage_enabled
= true;
631 struct page_ext_operations debug_guardpage_ops
= {
632 .need
= need_debug_guardpage
,
633 .init
= init_debug_guardpage
,
636 static int __init
debug_guardpage_minorder_setup(char *buf
)
640 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
641 pr_err("Bad debug_guardpage_minorder value\n");
644 _debug_guardpage_minorder
= res
;
645 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
648 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
650 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
651 unsigned int order
, int migratetype
)
653 struct page_ext
*page_ext
;
655 if (!debug_guardpage_enabled())
658 if (order
>= debug_guardpage_minorder())
661 page_ext
= lookup_page_ext(page
);
662 if (unlikely(!page_ext
))
665 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
667 INIT_LIST_HEAD(&page
->lru
);
668 set_page_private(page
, order
);
669 /* Guard pages are not available for any usage */
670 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
675 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
676 unsigned int order
, int migratetype
)
678 struct page_ext
*page_ext
;
680 if (!debug_guardpage_enabled())
683 page_ext
= lookup_page_ext(page
);
684 if (unlikely(!page_ext
))
687 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
689 set_page_private(page
, 0);
690 if (!is_migrate_isolate(migratetype
))
691 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
694 struct page_ext_operations debug_guardpage_ops
;
695 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
696 unsigned int order
, int migratetype
) { return false; }
697 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
698 unsigned int order
, int migratetype
) {}
701 static inline void set_page_order(struct page
*page
, unsigned int order
)
703 set_page_private(page
, order
);
704 __SetPageBuddy(page
);
707 static inline void rmv_page_order(struct page
*page
)
709 __ClearPageBuddy(page
);
710 set_page_private(page
, 0);
714 * This function checks whether a page is free && is the buddy
715 * we can coalesce a page and its buddy if
716 * (a) the buddy is not in a hole (check before calling!) &&
717 * (b) the buddy is in the buddy system &&
718 * (c) a page and its buddy have the same order &&
719 * (d) a page and its buddy are in the same zone.
721 * For recording whether a page is in the buddy system, we set PageBuddy.
722 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
724 * For recording page's order, we use page_private(page).
726 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
729 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
730 if (page_zone_id(page
) != page_zone_id(buddy
))
733 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
738 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
740 * zone check is done late to avoid uselessly
741 * calculating zone/node ids for pages that could
744 if (page_zone_id(page
) != page_zone_id(buddy
))
747 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
755 * Freeing function for a buddy system allocator.
757 * The concept of a buddy system is to maintain direct-mapped table
758 * (containing bit values) for memory blocks of various "orders".
759 * The bottom level table contains the map for the smallest allocatable
760 * units of memory (here, pages), and each level above it describes
761 * pairs of units from the levels below, hence, "buddies".
762 * At a high level, all that happens here is marking the table entry
763 * at the bottom level available, and propagating the changes upward
764 * as necessary, plus some accounting needed to play nicely with other
765 * parts of the VM system.
766 * At each level, we keep a list of pages, which are heads of continuous
767 * free pages of length of (1 << order) and marked with PageBuddy.
768 * Page's order is recorded in page_private(page) field.
769 * So when we are allocating or freeing one, we can derive the state of the
770 * other. That is, if we allocate a small block, and both were
771 * free, the remainder of the region must be split into blocks.
772 * If a block is freed, and its buddy is also free, then this
773 * triggers coalescing into a block of larger size.
778 static inline void __free_one_page(struct page
*page
,
780 struct zone
*zone
, unsigned int order
,
783 unsigned long combined_pfn
;
784 unsigned long uninitialized_var(buddy_pfn
);
786 unsigned int max_order
;
788 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
790 VM_BUG_ON(!zone_is_initialized(zone
));
791 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
793 VM_BUG_ON(migratetype
== -1);
794 if (likely(!is_migrate_isolate(migratetype
)))
795 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
797 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
798 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
801 while (order
< max_order
- 1) {
802 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
803 buddy
= page
+ (buddy_pfn
- pfn
);
805 if (!pfn_valid_within(buddy_pfn
))
807 if (!page_is_buddy(page
, buddy
, order
))
810 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
811 * merge with it and move up one order.
813 if (page_is_guard(buddy
)) {
814 clear_page_guard(zone
, buddy
, order
, migratetype
);
816 list_del(&buddy
->lru
);
817 zone
->free_area
[order
].nr_free
--;
818 rmv_page_order(buddy
);
820 combined_pfn
= buddy_pfn
& pfn
;
821 page
= page
+ (combined_pfn
- pfn
);
825 if (max_order
< MAX_ORDER
) {
826 /* If we are here, it means order is >= pageblock_order.
827 * We want to prevent merge between freepages on isolate
828 * pageblock and normal pageblock. Without this, pageblock
829 * isolation could cause incorrect freepage or CMA accounting.
831 * We don't want to hit this code for the more frequent
834 if (unlikely(has_isolate_pageblock(zone
))) {
837 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
838 buddy
= page
+ (buddy_pfn
- pfn
);
839 buddy_mt
= get_pageblock_migratetype(buddy
);
841 if (migratetype
!= buddy_mt
842 && (is_migrate_isolate(migratetype
) ||
843 is_migrate_isolate(buddy_mt
)))
847 goto continue_merging
;
851 set_page_order(page
, order
);
854 * If this is not the largest possible page, check if the buddy
855 * of the next-highest order is free. If it is, it's possible
856 * that pages are being freed that will coalesce soon. In case,
857 * that is happening, add the free page to the tail of the list
858 * so it's less likely to be used soon and more likely to be merged
859 * as a higher order page
861 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
862 struct page
*higher_page
, *higher_buddy
;
863 combined_pfn
= buddy_pfn
& pfn
;
864 higher_page
= page
+ (combined_pfn
- pfn
);
865 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
866 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
867 if (pfn_valid_within(buddy_pfn
) &&
868 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
869 list_add_tail(&page
->lru
,
870 &zone
->free_area
[order
].free_list
[migratetype
]);
875 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
877 zone
->free_area
[order
].nr_free
++;
881 * A bad page could be due to a number of fields. Instead of multiple branches,
882 * try and check multiple fields with one check. The caller must do a detailed
883 * check if necessary.
885 static inline bool page_expected_state(struct page
*page
,
886 unsigned long check_flags
)
888 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
891 if (unlikely((unsigned long)page
->mapping
|
892 page_ref_count(page
) |
894 (unsigned long)page
->mem_cgroup
|
896 (page
->flags
& check_flags
)))
902 static void free_pages_check_bad(struct page
*page
)
904 const char *bad_reason
;
905 unsigned long bad_flags
;
910 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
911 bad_reason
= "nonzero mapcount";
912 if (unlikely(page
->mapping
!= NULL
))
913 bad_reason
= "non-NULL mapping";
914 if (unlikely(page_ref_count(page
) != 0))
915 bad_reason
= "nonzero _refcount";
916 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
917 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
918 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
921 if (unlikely(page
->mem_cgroup
))
922 bad_reason
= "page still charged to cgroup";
924 bad_page(page
, bad_reason
, bad_flags
);
927 static inline int free_pages_check(struct page
*page
)
929 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
932 /* Something has gone sideways, find it */
933 free_pages_check_bad(page
);
937 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
942 * We rely page->lru.next never has bit 0 set, unless the page
943 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
945 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
947 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
951 switch (page
- head_page
) {
953 /* the first tail page: ->mapping may be compound_mapcount() */
954 if (unlikely(compound_mapcount(page
))) {
955 bad_page(page
, "nonzero compound_mapcount", 0);
961 * the second tail page: ->mapping is
962 * deferred_list.next -- ignore value.
966 if (page
->mapping
!= TAIL_MAPPING
) {
967 bad_page(page
, "corrupted mapping in tail page", 0);
972 if (unlikely(!PageTail(page
))) {
973 bad_page(page
, "PageTail not set", 0);
976 if (unlikely(compound_head(page
) != head_page
)) {
977 bad_page(page
, "compound_head not consistent", 0);
982 page
->mapping
= NULL
;
983 clear_compound_head(page
);
987 static __always_inline
bool free_pages_prepare(struct page
*page
,
988 unsigned int order
, bool check_free
)
992 VM_BUG_ON_PAGE(PageTail(page
), page
);
994 trace_mm_page_free(page
, order
);
997 * Check tail pages before head page information is cleared to
998 * avoid checking PageCompound for order-0 pages.
1000 if (unlikely(order
)) {
1001 bool compound
= PageCompound(page
);
1004 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1007 ClearPageDoubleMap(page
);
1008 for (i
= 1; i
< (1 << order
); i
++) {
1010 bad
+= free_tail_pages_check(page
, page
+ i
);
1011 if (unlikely(free_pages_check(page
+ i
))) {
1015 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1018 if (PageMappingFlags(page
))
1019 page
->mapping
= NULL
;
1020 if (memcg_kmem_enabled() && PageKmemcg(page
))
1021 memcg_kmem_uncharge(page
, order
);
1023 bad
+= free_pages_check(page
);
1027 page_cpupid_reset_last(page
);
1028 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1029 reset_page_owner(page
, order
);
1031 if (!PageHighMem(page
)) {
1032 debug_check_no_locks_freed(page_address(page
),
1033 PAGE_SIZE
<< order
);
1034 debug_check_no_obj_freed(page_address(page
),
1035 PAGE_SIZE
<< order
);
1037 arch_free_page(page
, order
);
1038 kernel_poison_pages(page
, 1 << order
, 0);
1039 kernel_map_pages(page
, 1 << order
, 0);
1040 kasan_free_pages(page
, order
);
1045 #ifdef CONFIG_DEBUG_VM
1046 static inline bool free_pcp_prepare(struct page
*page
)
1048 return free_pages_prepare(page
, 0, true);
1051 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1056 static bool free_pcp_prepare(struct page
*page
)
1058 return free_pages_prepare(page
, 0, false);
1061 static bool bulkfree_pcp_prepare(struct page
*page
)
1063 return free_pages_check(page
);
1065 #endif /* CONFIG_DEBUG_VM */
1067 static inline void prefetch_buddy(struct page
*page
)
1069 unsigned long pfn
= page_to_pfn(page
);
1070 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1071 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1077 * Frees a number of pages from the PCP lists
1078 * Assumes all pages on list are in same zone, and of same order.
1079 * count is the number of pages to free.
1081 * If the zone was previously in an "all pages pinned" state then look to
1082 * see if this freeing clears that state.
1084 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1085 * pinned" detection logic.
1087 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1088 struct per_cpu_pages
*pcp
)
1090 int migratetype
= 0;
1092 int prefetch_nr
= 0;
1093 bool isolated_pageblocks
;
1094 struct page
*page
, *tmp
;
1098 struct list_head
*list
;
1101 * Remove pages from lists in a round-robin fashion. A
1102 * batch_free count is maintained that is incremented when an
1103 * empty list is encountered. This is so more pages are freed
1104 * off fuller lists instead of spinning excessively around empty
1109 if (++migratetype
== MIGRATE_PCPTYPES
)
1111 list
= &pcp
->lists
[migratetype
];
1112 } while (list_empty(list
));
1114 /* This is the only non-empty list. Free them all. */
1115 if (batch_free
== MIGRATE_PCPTYPES
)
1119 page
= list_last_entry(list
, struct page
, lru
);
1120 /* must delete to avoid corrupting pcp list */
1121 list_del(&page
->lru
);
1124 if (bulkfree_pcp_prepare(page
))
1127 list_add_tail(&page
->lru
, &head
);
1130 * We are going to put the page back to the global
1131 * pool, prefetch its buddy to speed up later access
1132 * under zone->lock. It is believed the overhead of
1133 * an additional test and calculating buddy_pfn here
1134 * can be offset by reduced memory latency later. To
1135 * avoid excessive prefetching due to large count, only
1136 * prefetch buddy for the first pcp->batch nr of pages.
1138 if (prefetch_nr
++ < pcp
->batch
)
1139 prefetch_buddy(page
);
1140 } while (--count
&& --batch_free
&& !list_empty(list
));
1143 spin_lock(&zone
->lock
);
1144 isolated_pageblocks
= has_isolate_pageblock(zone
);
1147 * Use safe version since after __free_one_page(),
1148 * page->lru.next will not point to original list.
1150 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1151 int mt
= get_pcppage_migratetype(page
);
1152 /* MIGRATE_ISOLATE page should not go to pcplists */
1153 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1154 /* Pageblock could have been isolated meanwhile */
1155 if (unlikely(isolated_pageblocks
))
1156 mt
= get_pageblock_migratetype(page
);
1158 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1159 trace_mm_page_pcpu_drain(page
, 0, mt
);
1161 spin_unlock(&zone
->lock
);
1164 static void free_one_page(struct zone
*zone
,
1165 struct page
*page
, unsigned long pfn
,
1169 spin_lock(&zone
->lock
);
1170 if (unlikely(has_isolate_pageblock(zone
) ||
1171 is_migrate_isolate(migratetype
))) {
1172 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1174 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1175 spin_unlock(&zone
->lock
);
1178 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1179 unsigned long zone
, int nid
)
1181 mm_zero_struct_page(page
);
1182 set_page_links(page
, zone
, nid
, pfn
);
1183 init_page_count(page
);
1184 page_mapcount_reset(page
);
1185 page_cpupid_reset_last(page
);
1186 page_kasan_tag_reset(page
);
1188 INIT_LIST_HEAD(&page
->lru
);
1189 #ifdef WANT_PAGE_VIRTUAL
1190 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1191 if (!is_highmem_idx(zone
))
1192 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1196 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1197 static void __meminit
init_reserved_page(unsigned long pfn
)
1202 if (!early_page_uninitialised(pfn
))
1205 nid
= early_pfn_to_nid(pfn
);
1206 pgdat
= NODE_DATA(nid
);
1208 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1209 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1211 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1214 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1217 static inline void init_reserved_page(unsigned long pfn
)
1220 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1223 * Initialised pages do not have PageReserved set. This function is
1224 * called for each range allocated by the bootmem allocator and
1225 * marks the pages PageReserved. The remaining valid pages are later
1226 * sent to the buddy page allocator.
1228 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1230 unsigned long start_pfn
= PFN_DOWN(start
);
1231 unsigned long end_pfn
= PFN_UP(end
);
1233 for (; start_pfn
< end_pfn
; start_pfn
++) {
1234 if (pfn_valid(start_pfn
)) {
1235 struct page
*page
= pfn_to_page(start_pfn
);
1237 init_reserved_page(start_pfn
);
1239 /* Avoid false-positive PageTail() */
1240 INIT_LIST_HEAD(&page
->lru
);
1243 * no need for atomic set_bit because the struct
1244 * page is not visible yet so nobody should
1247 __SetPageReserved(page
);
1252 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1254 unsigned long flags
;
1256 unsigned long pfn
= page_to_pfn(page
);
1258 if (!free_pages_prepare(page
, order
, true))
1261 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1262 local_irq_save(flags
);
1263 __count_vm_events(PGFREE
, 1 << order
);
1264 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1265 local_irq_restore(flags
);
1268 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1270 unsigned int nr_pages
= 1 << order
;
1271 struct page
*p
= page
;
1275 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1277 __ClearPageReserved(p
);
1278 set_page_count(p
, 0);
1280 __ClearPageReserved(p
);
1281 set_page_count(p
, 0);
1283 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1284 set_page_refcounted(page
);
1285 __free_pages(page
, order
);
1288 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1289 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1291 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1293 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1295 static DEFINE_SPINLOCK(early_pfn_lock
);
1298 spin_lock(&early_pfn_lock
);
1299 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1301 nid
= first_online_node
;
1302 spin_unlock(&early_pfn_lock
);
1308 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1309 static inline bool __meminit __maybe_unused
1310 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1311 struct mminit_pfnnid_cache
*state
)
1315 nid
= __early_pfn_to_nid(pfn
, state
);
1316 if (nid
>= 0 && nid
!= node
)
1321 /* Only safe to use early in boot when initialisation is single-threaded */
1322 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1324 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1329 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1333 static inline bool __meminit __maybe_unused
1334 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1335 struct mminit_pfnnid_cache
*state
)
1342 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1345 if (early_page_uninitialised(pfn
))
1347 return __free_pages_boot_core(page
, order
);
1351 * Check that the whole (or subset of) a pageblock given by the interval of
1352 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1353 * with the migration of free compaction scanner. The scanners then need to
1354 * use only pfn_valid_within() check for arches that allow holes within
1357 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1359 * It's possible on some configurations to have a setup like node0 node1 node0
1360 * i.e. it's possible that all pages within a zones range of pages do not
1361 * belong to a single zone. We assume that a border between node0 and node1
1362 * can occur within a single pageblock, but not a node0 node1 node0
1363 * interleaving within a single pageblock. It is therefore sufficient to check
1364 * the first and last page of a pageblock and avoid checking each individual
1365 * page in a pageblock.
1367 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1368 unsigned long end_pfn
, struct zone
*zone
)
1370 struct page
*start_page
;
1371 struct page
*end_page
;
1373 /* end_pfn is one past the range we are checking */
1376 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1379 start_page
= pfn_to_online_page(start_pfn
);
1383 if (page_zone(start_page
) != zone
)
1386 end_page
= pfn_to_page(end_pfn
);
1388 /* This gives a shorter code than deriving page_zone(end_page) */
1389 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1395 void set_zone_contiguous(struct zone
*zone
)
1397 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1398 unsigned long block_end_pfn
;
1400 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1401 for (; block_start_pfn
< zone_end_pfn(zone
);
1402 block_start_pfn
= block_end_pfn
,
1403 block_end_pfn
+= pageblock_nr_pages
) {
1405 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1407 if (!__pageblock_pfn_to_page(block_start_pfn
,
1408 block_end_pfn
, zone
))
1412 /* We confirm that there is no hole */
1413 zone
->contiguous
= true;
1416 void clear_zone_contiguous(struct zone
*zone
)
1418 zone
->contiguous
= false;
1421 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1422 static void __init
deferred_free_range(unsigned long pfn
,
1423 unsigned long nr_pages
)
1431 page
= pfn_to_page(pfn
);
1433 /* Free a large naturally-aligned chunk if possible */
1434 if (nr_pages
== pageblock_nr_pages
&&
1435 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1436 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1437 __free_pages_boot_core(page
, pageblock_order
);
1441 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1442 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1443 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1444 __free_pages_boot_core(page
, 0);
1448 /* Completion tracking for deferred_init_memmap() threads */
1449 static atomic_t pgdat_init_n_undone __initdata
;
1450 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1452 static inline void __init
pgdat_init_report_one_done(void)
1454 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1455 complete(&pgdat_init_all_done_comp
);
1459 * Returns true if page needs to be initialized or freed to buddy allocator.
1461 * First we check if pfn is valid on architectures where it is possible to have
1462 * holes within pageblock_nr_pages. On systems where it is not possible, this
1463 * function is optimized out.
1465 * Then, we check if a current large page is valid by only checking the validity
1468 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1469 * within a node: a pfn is between start and end of a node, but does not belong
1470 * to this memory node.
1472 static inline bool __init
1473 deferred_pfn_valid(int nid
, unsigned long pfn
,
1474 struct mminit_pfnnid_cache
*nid_init_state
)
1476 if (!pfn_valid_within(pfn
))
1478 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1480 if (!meminit_pfn_in_nid(pfn
, nid
, nid_init_state
))
1486 * Free pages to buddy allocator. Try to free aligned pages in
1487 * pageblock_nr_pages sizes.
1489 static void __init
deferred_free_pages(int nid
, int zid
, unsigned long pfn
,
1490 unsigned long end_pfn
)
1492 struct mminit_pfnnid_cache nid_init_state
= { };
1493 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1494 unsigned long nr_free
= 0;
1496 for (; pfn
< end_pfn
; pfn
++) {
1497 if (!deferred_pfn_valid(nid
, pfn
, &nid_init_state
)) {
1498 deferred_free_range(pfn
- nr_free
, nr_free
);
1500 } else if (!(pfn
& nr_pgmask
)) {
1501 deferred_free_range(pfn
- nr_free
, nr_free
);
1503 touch_nmi_watchdog();
1508 /* Free the last block of pages to allocator */
1509 deferred_free_range(pfn
- nr_free
, nr_free
);
1513 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1514 * by performing it only once every pageblock_nr_pages.
1515 * Return number of pages initialized.
1517 static unsigned long __init
deferred_init_pages(int nid
, int zid
,
1519 unsigned long end_pfn
)
1521 struct mminit_pfnnid_cache nid_init_state
= { };
1522 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1523 unsigned long nr_pages
= 0;
1524 struct page
*page
= NULL
;
1526 for (; pfn
< end_pfn
; pfn
++) {
1527 if (!deferred_pfn_valid(nid
, pfn
, &nid_init_state
)) {
1530 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1531 page
= pfn_to_page(pfn
);
1532 touch_nmi_watchdog();
1536 __init_single_page(page
, pfn
, zid
, nid
);
1542 /* Initialise remaining memory on a node */
1543 static int __init
deferred_init_memmap(void *data
)
1545 pg_data_t
*pgdat
= data
;
1546 int nid
= pgdat
->node_id
;
1547 unsigned long start
= jiffies
;
1548 unsigned long nr_pages
= 0;
1549 unsigned long spfn
, epfn
, first_init_pfn
, flags
;
1550 phys_addr_t spa
, epa
;
1553 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1556 /* Bind memory initialisation thread to a local node if possible */
1557 if (!cpumask_empty(cpumask
))
1558 set_cpus_allowed_ptr(current
, cpumask
);
1560 pgdat_resize_lock(pgdat
, &flags
);
1561 first_init_pfn
= pgdat
->first_deferred_pfn
;
1562 if (first_init_pfn
== ULONG_MAX
) {
1563 pgdat_resize_unlock(pgdat
, &flags
);
1564 pgdat_init_report_one_done();
1568 /* Sanity check boundaries */
1569 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1570 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1571 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1573 /* Only the highest zone is deferred so find it */
1574 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1575 zone
= pgdat
->node_zones
+ zid
;
1576 if (first_init_pfn
< zone_end_pfn(zone
))
1579 first_init_pfn
= max(zone
->zone_start_pfn
, first_init_pfn
);
1582 * Initialize and free pages. We do it in two loops: first we initialize
1583 * struct page, than free to buddy allocator, because while we are
1584 * freeing pages we can access pages that are ahead (computing buddy
1585 * page in __free_one_page()).
1587 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1588 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1589 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1590 nr_pages
+= deferred_init_pages(nid
, zid
, spfn
, epfn
);
1592 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1593 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1594 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1595 deferred_free_pages(nid
, zid
, spfn
, epfn
);
1597 pgdat_resize_unlock(pgdat
, &flags
);
1599 /* Sanity check that the next zone really is unpopulated */
1600 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1602 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1603 jiffies_to_msecs(jiffies
- start
));
1605 pgdat_init_report_one_done();
1610 * During boot we initialize deferred pages on-demand, as needed, but once
1611 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1612 * and we can permanently disable that path.
1614 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
1617 * If this zone has deferred pages, try to grow it by initializing enough
1618 * deferred pages to satisfy the allocation specified by order, rounded up to
1619 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1620 * of SECTION_SIZE bytes by initializing struct pages in increments of
1621 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1623 * Return true when zone was grown, otherwise return false. We return true even
1624 * when we grow less than requested, to let the caller decide if there are
1625 * enough pages to satisfy the allocation.
1627 * Note: We use noinline because this function is needed only during boot, and
1628 * it is called from a __ref function _deferred_grow_zone. This way we are
1629 * making sure that it is not inlined into permanent text section.
1631 static noinline
bool __init
1632 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1634 int zid
= zone_idx(zone
);
1635 int nid
= zone_to_nid(zone
);
1636 pg_data_t
*pgdat
= NODE_DATA(nid
);
1637 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1638 unsigned long nr_pages
= 0;
1639 unsigned long first_init_pfn
, spfn
, epfn
, t
, flags
;
1640 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1641 phys_addr_t spa
, epa
;
1644 /* Only the last zone may have deferred pages */
1645 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1648 pgdat_resize_lock(pgdat
, &flags
);
1651 * If deferred pages have been initialized while we were waiting for
1652 * the lock, return true, as the zone was grown. The caller will retry
1653 * this zone. We won't return to this function since the caller also
1654 * has this static branch.
1656 if (!static_branch_unlikely(&deferred_pages
)) {
1657 pgdat_resize_unlock(pgdat
, &flags
);
1662 * If someone grew this zone while we were waiting for spinlock, return
1663 * true, as there might be enough pages already.
1665 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1666 pgdat_resize_unlock(pgdat
, &flags
);
1670 first_init_pfn
= max(zone
->zone_start_pfn
, first_deferred_pfn
);
1672 if (first_init_pfn
>= pgdat_end_pfn(pgdat
)) {
1673 pgdat_resize_unlock(pgdat
, &flags
);
1677 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1678 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1679 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1681 while (spfn
< epfn
&& nr_pages
< nr_pages_needed
) {
1682 t
= ALIGN(spfn
+ PAGES_PER_SECTION
, PAGES_PER_SECTION
);
1683 first_deferred_pfn
= min(t
, epfn
);
1684 nr_pages
+= deferred_init_pages(nid
, zid
, spfn
,
1685 first_deferred_pfn
);
1686 spfn
= first_deferred_pfn
;
1689 if (nr_pages
>= nr_pages_needed
)
1693 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1694 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1695 epfn
= min_t(unsigned long, first_deferred_pfn
, PFN_DOWN(epa
));
1696 deferred_free_pages(nid
, zid
, spfn
, epfn
);
1698 if (first_deferred_pfn
== epfn
)
1701 pgdat
->first_deferred_pfn
= first_deferred_pfn
;
1702 pgdat_resize_unlock(pgdat
, &flags
);
1704 return nr_pages
> 0;
1708 * deferred_grow_zone() is __init, but it is called from
1709 * get_page_from_freelist() during early boot until deferred_pages permanently
1710 * disables this call. This is why we have refdata wrapper to avoid warning,
1711 * and to ensure that the function body gets unloaded.
1714 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1716 return deferred_grow_zone(zone
, order
);
1719 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1721 void __init
page_alloc_init_late(void)
1725 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1728 /* There will be num_node_state(N_MEMORY) threads */
1729 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1730 for_each_node_state(nid
, N_MEMORY
) {
1731 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1734 /* Block until all are initialised */
1735 wait_for_completion(&pgdat_init_all_done_comp
);
1738 * We initialized the rest of the deferred pages. Permanently disable
1739 * on-demand struct page initialization.
1741 static_branch_disable(&deferred_pages
);
1743 /* Reinit limits that are based on free pages after the kernel is up */
1744 files_maxfiles_init();
1746 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1747 /* Discard memblock private memory */
1751 for_each_populated_zone(zone
)
1752 set_zone_contiguous(zone
);
1756 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1757 void __init
init_cma_reserved_pageblock(struct page
*page
)
1759 unsigned i
= pageblock_nr_pages
;
1760 struct page
*p
= page
;
1763 __ClearPageReserved(p
);
1764 set_page_count(p
, 0);
1767 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1769 if (pageblock_order
>= MAX_ORDER
) {
1770 i
= pageblock_nr_pages
;
1773 set_page_refcounted(p
);
1774 __free_pages(p
, MAX_ORDER
- 1);
1775 p
+= MAX_ORDER_NR_PAGES
;
1776 } while (i
-= MAX_ORDER_NR_PAGES
);
1778 set_page_refcounted(page
);
1779 __free_pages(page
, pageblock_order
);
1782 adjust_managed_page_count(page
, pageblock_nr_pages
);
1787 * The order of subdivision here is critical for the IO subsystem.
1788 * Please do not alter this order without good reasons and regression
1789 * testing. Specifically, as large blocks of memory are subdivided,
1790 * the order in which smaller blocks are delivered depends on the order
1791 * they're subdivided in this function. This is the primary factor
1792 * influencing the order in which pages are delivered to the IO
1793 * subsystem according to empirical testing, and this is also justified
1794 * by considering the behavior of a buddy system containing a single
1795 * large block of memory acted on by a series of small allocations.
1796 * This behavior is a critical factor in sglist merging's success.
1800 static inline void expand(struct zone
*zone
, struct page
*page
,
1801 int low
, int high
, struct free_area
*area
,
1804 unsigned long size
= 1 << high
;
1806 while (high
> low
) {
1810 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1813 * Mark as guard pages (or page), that will allow to
1814 * merge back to allocator when buddy will be freed.
1815 * Corresponding page table entries will not be touched,
1816 * pages will stay not present in virtual address space
1818 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1821 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1823 set_page_order(&page
[size
], high
);
1827 static void check_new_page_bad(struct page
*page
)
1829 const char *bad_reason
= NULL
;
1830 unsigned long bad_flags
= 0;
1832 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1833 bad_reason
= "nonzero mapcount";
1834 if (unlikely(page
->mapping
!= NULL
))
1835 bad_reason
= "non-NULL mapping";
1836 if (unlikely(page_ref_count(page
) != 0))
1837 bad_reason
= "nonzero _count";
1838 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1839 bad_reason
= "HWPoisoned (hardware-corrupted)";
1840 bad_flags
= __PG_HWPOISON
;
1841 /* Don't complain about hwpoisoned pages */
1842 page_mapcount_reset(page
); /* remove PageBuddy */
1845 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1846 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1847 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1850 if (unlikely(page
->mem_cgroup
))
1851 bad_reason
= "page still charged to cgroup";
1853 bad_page(page
, bad_reason
, bad_flags
);
1857 * This page is about to be returned from the page allocator
1859 static inline int check_new_page(struct page
*page
)
1861 if (likely(page_expected_state(page
,
1862 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1865 check_new_page_bad(page
);
1869 static inline bool free_pages_prezeroed(void)
1871 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1872 page_poisoning_enabled();
1875 #ifdef CONFIG_DEBUG_VM
1876 static bool check_pcp_refill(struct page
*page
)
1881 static bool check_new_pcp(struct page
*page
)
1883 return check_new_page(page
);
1886 static bool check_pcp_refill(struct page
*page
)
1888 return check_new_page(page
);
1890 static bool check_new_pcp(struct page
*page
)
1894 #endif /* CONFIG_DEBUG_VM */
1896 static bool check_new_pages(struct page
*page
, unsigned int order
)
1899 for (i
= 0; i
< (1 << order
); i
++) {
1900 struct page
*p
= page
+ i
;
1902 if (unlikely(check_new_page(p
)))
1909 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1912 set_page_private(page
, 0);
1913 set_page_refcounted(page
);
1915 arch_alloc_page(page
, order
);
1916 kernel_map_pages(page
, 1 << order
, 1);
1917 kernel_poison_pages(page
, 1 << order
, 1);
1918 kasan_alloc_pages(page
, order
);
1919 set_page_owner(page
, order
, gfp_flags
);
1922 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1923 unsigned int alloc_flags
)
1927 post_alloc_hook(page
, order
, gfp_flags
);
1929 if (!free_pages_prezeroed() && (gfp_flags
& __GFP_ZERO
))
1930 for (i
= 0; i
< (1 << order
); i
++)
1931 clear_highpage(page
+ i
);
1933 if (order
&& (gfp_flags
& __GFP_COMP
))
1934 prep_compound_page(page
, order
);
1937 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1938 * allocate the page. The expectation is that the caller is taking
1939 * steps that will free more memory. The caller should avoid the page
1940 * being used for !PFMEMALLOC purposes.
1942 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1943 set_page_pfmemalloc(page
);
1945 clear_page_pfmemalloc(page
);
1949 * Go through the free lists for the given migratetype and remove
1950 * the smallest available page from the freelists
1952 static __always_inline
1953 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1956 unsigned int current_order
;
1957 struct free_area
*area
;
1960 /* Find a page of the appropriate size in the preferred list */
1961 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1962 area
= &(zone
->free_area
[current_order
]);
1963 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1967 list_del(&page
->lru
);
1968 rmv_page_order(page
);
1970 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1971 set_pcppage_migratetype(page
, migratetype
);
1980 * This array describes the order lists are fallen back to when
1981 * the free lists for the desirable migrate type are depleted
1983 static int fallbacks
[MIGRATE_TYPES
][4] = {
1984 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1985 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1986 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1988 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1990 #ifdef CONFIG_MEMORY_ISOLATION
1991 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1996 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1999 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2002 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2003 unsigned int order
) { return NULL
; }
2007 * Move the free pages in a range to the free lists of the requested type.
2008 * Note that start_page and end_pages are not aligned on a pageblock
2009 * boundary. If alignment is required, use move_freepages_block()
2011 static int move_freepages(struct zone
*zone
,
2012 struct page
*start_page
, struct page
*end_page
,
2013 int migratetype
, int *num_movable
)
2017 int pages_moved
= 0;
2019 #ifndef CONFIG_HOLES_IN_ZONE
2021 * page_zone is not safe to call in this context when
2022 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2023 * anyway as we check zone boundaries in move_freepages_block().
2024 * Remove at a later date when no bug reports exist related to
2025 * grouping pages by mobility
2027 VM_BUG_ON(pfn_valid(page_to_pfn(start_page
)) &&
2028 pfn_valid(page_to_pfn(end_page
)) &&
2029 page_zone(start_page
) != page_zone(end_page
));
2031 for (page
= start_page
; page
<= end_page
;) {
2032 if (!pfn_valid_within(page_to_pfn(page
))) {
2037 /* Make sure we are not inadvertently changing nodes */
2038 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2040 if (!PageBuddy(page
)) {
2042 * We assume that pages that could be isolated for
2043 * migration are movable. But we don't actually try
2044 * isolating, as that would be expensive.
2047 (PageLRU(page
) || __PageMovable(page
)))
2054 order
= page_order(page
);
2055 list_move(&page
->lru
,
2056 &zone
->free_area
[order
].free_list
[migratetype
]);
2058 pages_moved
+= 1 << order
;
2064 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2065 int migratetype
, int *num_movable
)
2067 unsigned long start_pfn
, end_pfn
;
2068 struct page
*start_page
, *end_page
;
2073 start_pfn
= page_to_pfn(page
);
2074 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2075 start_page
= pfn_to_page(start_pfn
);
2076 end_page
= start_page
+ pageblock_nr_pages
- 1;
2077 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2079 /* Do not cross zone boundaries */
2080 if (!zone_spans_pfn(zone
, start_pfn
))
2082 if (!zone_spans_pfn(zone
, end_pfn
))
2085 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2089 static void change_pageblock_range(struct page
*pageblock_page
,
2090 int start_order
, int migratetype
)
2092 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2094 while (nr_pageblocks
--) {
2095 set_pageblock_migratetype(pageblock_page
, migratetype
);
2096 pageblock_page
+= pageblock_nr_pages
;
2101 * When we are falling back to another migratetype during allocation, try to
2102 * steal extra free pages from the same pageblocks to satisfy further
2103 * allocations, instead of polluting multiple pageblocks.
2105 * If we are stealing a relatively large buddy page, it is likely there will
2106 * be more free pages in the pageblock, so try to steal them all. For
2107 * reclaimable and unmovable allocations, we steal regardless of page size,
2108 * as fragmentation caused by those allocations polluting movable pageblocks
2109 * is worse than movable allocations stealing from unmovable and reclaimable
2112 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2115 * Leaving this order check is intended, although there is
2116 * relaxed order check in next check. The reason is that
2117 * we can actually steal whole pageblock if this condition met,
2118 * but, below check doesn't guarantee it and that is just heuristic
2119 * so could be changed anytime.
2121 if (order
>= pageblock_order
)
2124 if (order
>= pageblock_order
/ 2 ||
2125 start_mt
== MIGRATE_RECLAIMABLE
||
2126 start_mt
== MIGRATE_UNMOVABLE
||
2127 page_group_by_mobility_disabled
)
2133 static inline void boost_watermark(struct zone
*zone
)
2135 unsigned long max_boost
;
2137 if (!watermark_boost_factor
)
2140 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2141 watermark_boost_factor
, 10000);
2142 max_boost
= max(pageblock_nr_pages
, max_boost
);
2144 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2149 * This function implements actual steal behaviour. If order is large enough,
2150 * we can steal whole pageblock. If not, we first move freepages in this
2151 * pageblock to our migratetype and determine how many already-allocated pages
2152 * are there in the pageblock with a compatible migratetype. If at least half
2153 * of pages are free or compatible, we can change migratetype of the pageblock
2154 * itself, so pages freed in the future will be put on the correct free list.
2156 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2157 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2159 unsigned int current_order
= page_order(page
);
2160 struct free_area
*area
;
2161 int free_pages
, movable_pages
, alike_pages
;
2164 old_block_type
= get_pageblock_migratetype(page
);
2167 * This can happen due to races and we want to prevent broken
2168 * highatomic accounting.
2170 if (is_migrate_highatomic(old_block_type
))
2173 /* Take ownership for orders >= pageblock_order */
2174 if (current_order
>= pageblock_order
) {
2175 change_pageblock_range(page
, current_order
, start_type
);
2180 * Boost watermarks to increase reclaim pressure to reduce the
2181 * likelihood of future fallbacks. Wake kswapd now as the node
2182 * may be balanced overall and kswapd will not wake naturally.
2184 boost_watermark(zone
);
2185 if (alloc_flags
& ALLOC_KSWAPD
)
2186 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
2188 /* We are not allowed to try stealing from the whole block */
2192 free_pages
= move_freepages_block(zone
, page
, start_type
,
2195 * Determine how many pages are compatible with our allocation.
2196 * For movable allocation, it's the number of movable pages which
2197 * we just obtained. For other types it's a bit more tricky.
2199 if (start_type
== MIGRATE_MOVABLE
) {
2200 alike_pages
= movable_pages
;
2203 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2204 * to MOVABLE pageblock, consider all non-movable pages as
2205 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2206 * vice versa, be conservative since we can't distinguish the
2207 * exact migratetype of non-movable pages.
2209 if (old_block_type
== MIGRATE_MOVABLE
)
2210 alike_pages
= pageblock_nr_pages
2211 - (free_pages
+ movable_pages
);
2216 /* moving whole block can fail due to zone boundary conditions */
2221 * If a sufficient number of pages in the block are either free or of
2222 * comparable migratability as our allocation, claim the whole block.
2224 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2225 page_group_by_mobility_disabled
)
2226 set_pageblock_migratetype(page
, start_type
);
2231 area
= &zone
->free_area
[current_order
];
2232 list_move(&page
->lru
, &area
->free_list
[start_type
]);
2236 * Check whether there is a suitable fallback freepage with requested order.
2237 * If only_stealable is true, this function returns fallback_mt only if
2238 * we can steal other freepages all together. This would help to reduce
2239 * fragmentation due to mixed migratetype pages in one pageblock.
2241 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2242 int migratetype
, bool only_stealable
, bool *can_steal
)
2247 if (area
->nr_free
== 0)
2252 fallback_mt
= fallbacks
[migratetype
][i
];
2253 if (fallback_mt
== MIGRATE_TYPES
)
2256 if (list_empty(&area
->free_list
[fallback_mt
]))
2259 if (can_steal_fallback(order
, migratetype
))
2262 if (!only_stealable
)
2273 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2274 * there are no empty page blocks that contain a page with a suitable order
2276 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2277 unsigned int alloc_order
)
2280 unsigned long max_managed
, flags
;
2283 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2284 * Check is race-prone but harmless.
2286 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2287 if (zone
->nr_reserved_highatomic
>= max_managed
)
2290 spin_lock_irqsave(&zone
->lock
, flags
);
2292 /* Recheck the nr_reserved_highatomic limit under the lock */
2293 if (zone
->nr_reserved_highatomic
>= max_managed
)
2297 mt
= get_pageblock_migratetype(page
);
2298 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2299 && !is_migrate_cma(mt
)) {
2300 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2301 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2302 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2306 spin_unlock_irqrestore(&zone
->lock
, flags
);
2310 * Used when an allocation is about to fail under memory pressure. This
2311 * potentially hurts the reliability of high-order allocations when under
2312 * intense memory pressure but failed atomic allocations should be easier
2313 * to recover from than an OOM.
2315 * If @force is true, try to unreserve a pageblock even though highatomic
2316 * pageblock is exhausted.
2318 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2321 struct zonelist
*zonelist
= ac
->zonelist
;
2322 unsigned long flags
;
2329 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2332 * Preserve at least one pageblock unless memory pressure
2335 if (!force
&& zone
->nr_reserved_highatomic
<=
2339 spin_lock_irqsave(&zone
->lock
, flags
);
2340 for (order
= 0; order
< MAX_ORDER
; order
++) {
2341 struct free_area
*area
= &(zone
->free_area
[order
]);
2343 page
= list_first_entry_or_null(
2344 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2350 * In page freeing path, migratetype change is racy so
2351 * we can counter several free pages in a pageblock
2352 * in this loop althoug we changed the pageblock type
2353 * from highatomic to ac->migratetype. So we should
2354 * adjust the count once.
2356 if (is_migrate_highatomic_page(page
)) {
2358 * It should never happen but changes to
2359 * locking could inadvertently allow a per-cpu
2360 * drain to add pages to MIGRATE_HIGHATOMIC
2361 * while unreserving so be safe and watch for
2364 zone
->nr_reserved_highatomic
-= min(
2366 zone
->nr_reserved_highatomic
);
2370 * Convert to ac->migratetype and avoid the normal
2371 * pageblock stealing heuristics. Minimally, the caller
2372 * is doing the work and needs the pages. More
2373 * importantly, if the block was always converted to
2374 * MIGRATE_UNMOVABLE or another type then the number
2375 * of pageblocks that cannot be completely freed
2378 set_pageblock_migratetype(page
, ac
->migratetype
);
2379 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2382 spin_unlock_irqrestore(&zone
->lock
, flags
);
2386 spin_unlock_irqrestore(&zone
->lock
, flags
);
2393 * Try finding a free buddy page on the fallback list and put it on the free
2394 * list of requested migratetype, possibly along with other pages from the same
2395 * block, depending on fragmentation avoidance heuristics. Returns true if
2396 * fallback was found so that __rmqueue_smallest() can grab it.
2398 * The use of signed ints for order and current_order is a deliberate
2399 * deviation from the rest of this file, to make the for loop
2400 * condition simpler.
2402 static __always_inline
bool
2403 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2404 unsigned int alloc_flags
)
2406 struct free_area
*area
;
2408 int min_order
= order
;
2414 * Do not steal pages from freelists belonging to other pageblocks
2415 * i.e. orders < pageblock_order. If there are no local zones free,
2416 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2418 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2419 min_order
= pageblock_order
;
2422 * Find the largest available free page in the other list. This roughly
2423 * approximates finding the pageblock with the most free pages, which
2424 * would be too costly to do exactly.
2426 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2428 area
= &(zone
->free_area
[current_order
]);
2429 fallback_mt
= find_suitable_fallback(area
, current_order
,
2430 start_migratetype
, false, &can_steal
);
2431 if (fallback_mt
== -1)
2435 * We cannot steal all free pages from the pageblock and the
2436 * requested migratetype is movable. In that case it's better to
2437 * steal and split the smallest available page instead of the
2438 * largest available page, because even if the next movable
2439 * allocation falls back into a different pageblock than this
2440 * one, it won't cause permanent fragmentation.
2442 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2443 && current_order
> order
)
2452 for (current_order
= order
; current_order
< MAX_ORDER
;
2454 area
= &(zone
->free_area
[current_order
]);
2455 fallback_mt
= find_suitable_fallback(area
, current_order
,
2456 start_migratetype
, false, &can_steal
);
2457 if (fallback_mt
!= -1)
2462 * This should not happen - we already found a suitable fallback
2463 * when looking for the largest page.
2465 VM_BUG_ON(current_order
== MAX_ORDER
);
2468 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2471 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2474 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2475 start_migratetype
, fallback_mt
);
2482 * Do the hard work of removing an element from the buddy allocator.
2483 * Call me with the zone->lock already held.
2485 static __always_inline
struct page
*
2486 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2487 unsigned int alloc_flags
)
2492 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2493 if (unlikely(!page
)) {
2494 if (migratetype
== MIGRATE_MOVABLE
)
2495 page
= __rmqueue_cma_fallback(zone
, order
);
2497 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2502 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2507 * Obtain a specified number of elements from the buddy allocator, all under
2508 * a single hold of the lock, for efficiency. Add them to the supplied list.
2509 * Returns the number of new pages which were placed at *list.
2511 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2512 unsigned long count
, struct list_head
*list
,
2513 int migratetype
, unsigned int alloc_flags
)
2517 spin_lock(&zone
->lock
);
2518 for (i
= 0; i
< count
; ++i
) {
2519 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2521 if (unlikely(page
== NULL
))
2524 if (unlikely(check_pcp_refill(page
)))
2528 * Split buddy pages returned by expand() are received here in
2529 * physical page order. The page is added to the tail of
2530 * caller's list. From the callers perspective, the linked list
2531 * is ordered by page number under some conditions. This is
2532 * useful for IO devices that can forward direction from the
2533 * head, thus also in the physical page order. This is useful
2534 * for IO devices that can merge IO requests if the physical
2535 * pages are ordered properly.
2537 list_add_tail(&page
->lru
, list
);
2539 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2540 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2545 * i pages were removed from the buddy list even if some leak due
2546 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2547 * on i. Do not confuse with 'alloced' which is the number of
2548 * pages added to the pcp list.
2550 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2551 spin_unlock(&zone
->lock
);
2557 * Called from the vmstat counter updater to drain pagesets of this
2558 * currently executing processor on remote nodes after they have
2561 * Note that this function must be called with the thread pinned to
2562 * a single processor.
2564 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2566 unsigned long flags
;
2567 int to_drain
, batch
;
2569 local_irq_save(flags
);
2570 batch
= READ_ONCE(pcp
->batch
);
2571 to_drain
= min(pcp
->count
, batch
);
2573 free_pcppages_bulk(zone
, to_drain
, pcp
);
2574 local_irq_restore(flags
);
2579 * Drain pcplists of the indicated processor and zone.
2581 * The processor must either be the current processor and the
2582 * thread pinned to the current processor or a processor that
2585 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2587 unsigned long flags
;
2588 struct per_cpu_pageset
*pset
;
2589 struct per_cpu_pages
*pcp
;
2591 local_irq_save(flags
);
2592 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2596 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2597 local_irq_restore(flags
);
2601 * Drain pcplists of all zones on the indicated processor.
2603 * The processor must either be the current processor and the
2604 * thread pinned to the current processor or a processor that
2607 static void drain_pages(unsigned int cpu
)
2611 for_each_populated_zone(zone
) {
2612 drain_pages_zone(cpu
, zone
);
2617 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2619 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2620 * the single zone's pages.
2622 void drain_local_pages(struct zone
*zone
)
2624 int cpu
= smp_processor_id();
2627 drain_pages_zone(cpu
, zone
);
2632 static void drain_local_pages_wq(struct work_struct
*work
)
2635 * drain_all_pages doesn't use proper cpu hotplug protection so
2636 * we can race with cpu offline when the WQ can move this from
2637 * a cpu pinned worker to an unbound one. We can operate on a different
2638 * cpu which is allright but we also have to make sure to not move to
2642 drain_local_pages(NULL
);
2647 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2649 * When zone parameter is non-NULL, spill just the single zone's pages.
2651 * Note that this can be extremely slow as the draining happens in a workqueue.
2653 void drain_all_pages(struct zone
*zone
)
2658 * Allocate in the BSS so we wont require allocation in
2659 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2661 static cpumask_t cpus_with_pcps
;
2664 * Make sure nobody triggers this path before mm_percpu_wq is fully
2667 if (WARN_ON_ONCE(!mm_percpu_wq
))
2671 * Do not drain if one is already in progress unless it's specific to
2672 * a zone. Such callers are primarily CMA and memory hotplug and need
2673 * the drain to be complete when the call returns.
2675 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2678 mutex_lock(&pcpu_drain_mutex
);
2682 * We don't care about racing with CPU hotplug event
2683 * as offline notification will cause the notified
2684 * cpu to drain that CPU pcps and on_each_cpu_mask
2685 * disables preemption as part of its processing
2687 for_each_online_cpu(cpu
) {
2688 struct per_cpu_pageset
*pcp
;
2690 bool has_pcps
= false;
2693 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2697 for_each_populated_zone(z
) {
2698 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2699 if (pcp
->pcp
.count
) {
2707 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2709 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2712 for_each_cpu(cpu
, &cpus_with_pcps
) {
2713 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2714 INIT_WORK(work
, drain_local_pages_wq
);
2715 queue_work_on(cpu
, mm_percpu_wq
, work
);
2717 for_each_cpu(cpu
, &cpus_with_pcps
)
2718 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2720 mutex_unlock(&pcpu_drain_mutex
);
2723 #ifdef CONFIG_HIBERNATION
2726 * Touch the watchdog for every WD_PAGE_COUNT pages.
2728 #define WD_PAGE_COUNT (128*1024)
2730 void mark_free_pages(struct zone
*zone
)
2732 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2733 unsigned long flags
;
2734 unsigned int order
, t
;
2737 if (zone_is_empty(zone
))
2740 spin_lock_irqsave(&zone
->lock
, flags
);
2742 max_zone_pfn
= zone_end_pfn(zone
);
2743 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2744 if (pfn_valid(pfn
)) {
2745 page
= pfn_to_page(pfn
);
2747 if (!--page_count
) {
2748 touch_nmi_watchdog();
2749 page_count
= WD_PAGE_COUNT
;
2752 if (page_zone(page
) != zone
)
2755 if (!swsusp_page_is_forbidden(page
))
2756 swsusp_unset_page_free(page
);
2759 for_each_migratetype_order(order
, t
) {
2760 list_for_each_entry(page
,
2761 &zone
->free_area
[order
].free_list
[t
], lru
) {
2764 pfn
= page_to_pfn(page
);
2765 for (i
= 0; i
< (1UL << order
); i
++) {
2766 if (!--page_count
) {
2767 touch_nmi_watchdog();
2768 page_count
= WD_PAGE_COUNT
;
2770 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2774 spin_unlock_irqrestore(&zone
->lock
, flags
);
2776 #endif /* CONFIG_PM */
2778 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
2782 if (!free_pcp_prepare(page
))
2785 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2786 set_pcppage_migratetype(page
, migratetype
);
2790 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
2792 struct zone
*zone
= page_zone(page
);
2793 struct per_cpu_pages
*pcp
;
2796 migratetype
= get_pcppage_migratetype(page
);
2797 __count_vm_event(PGFREE
);
2800 * We only track unmovable, reclaimable and movable on pcp lists.
2801 * Free ISOLATE pages back to the allocator because they are being
2802 * offlined but treat HIGHATOMIC as movable pages so we can get those
2803 * areas back if necessary. Otherwise, we may have to free
2804 * excessively into the page allocator
2806 if (migratetype
>= MIGRATE_PCPTYPES
) {
2807 if (unlikely(is_migrate_isolate(migratetype
))) {
2808 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2811 migratetype
= MIGRATE_MOVABLE
;
2814 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2815 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2817 if (pcp
->count
>= pcp
->high
) {
2818 unsigned long batch
= READ_ONCE(pcp
->batch
);
2819 free_pcppages_bulk(zone
, batch
, pcp
);
2824 * Free a 0-order page
2826 void free_unref_page(struct page
*page
)
2828 unsigned long flags
;
2829 unsigned long pfn
= page_to_pfn(page
);
2831 if (!free_unref_page_prepare(page
, pfn
))
2834 local_irq_save(flags
);
2835 free_unref_page_commit(page
, pfn
);
2836 local_irq_restore(flags
);
2840 * Free a list of 0-order pages
2842 void free_unref_page_list(struct list_head
*list
)
2844 struct page
*page
, *next
;
2845 unsigned long flags
, pfn
;
2846 int batch_count
= 0;
2848 /* Prepare pages for freeing */
2849 list_for_each_entry_safe(page
, next
, list
, lru
) {
2850 pfn
= page_to_pfn(page
);
2851 if (!free_unref_page_prepare(page
, pfn
))
2852 list_del(&page
->lru
);
2853 set_page_private(page
, pfn
);
2856 local_irq_save(flags
);
2857 list_for_each_entry_safe(page
, next
, list
, lru
) {
2858 unsigned long pfn
= page_private(page
);
2860 set_page_private(page
, 0);
2861 trace_mm_page_free_batched(page
);
2862 free_unref_page_commit(page
, pfn
);
2865 * Guard against excessive IRQ disabled times when we get
2866 * a large list of pages to free.
2868 if (++batch_count
== SWAP_CLUSTER_MAX
) {
2869 local_irq_restore(flags
);
2871 local_irq_save(flags
);
2874 local_irq_restore(flags
);
2878 * split_page takes a non-compound higher-order page, and splits it into
2879 * n (1<<order) sub-pages: page[0..n]
2880 * Each sub-page must be freed individually.
2882 * Note: this is probably too low level an operation for use in drivers.
2883 * Please consult with lkml before using this in your driver.
2885 void split_page(struct page
*page
, unsigned int order
)
2889 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2890 VM_BUG_ON_PAGE(!page_count(page
), page
);
2892 for (i
= 1; i
< (1 << order
); i
++)
2893 set_page_refcounted(page
+ i
);
2894 split_page_owner(page
, order
);
2896 EXPORT_SYMBOL_GPL(split_page
);
2898 int __isolate_free_page(struct page
*page
, unsigned int order
)
2900 unsigned long watermark
;
2904 BUG_ON(!PageBuddy(page
));
2906 zone
= page_zone(page
);
2907 mt
= get_pageblock_migratetype(page
);
2909 if (!is_migrate_isolate(mt
)) {
2911 * Obey watermarks as if the page was being allocated. We can
2912 * emulate a high-order watermark check with a raised order-0
2913 * watermark, because we already know our high-order page
2916 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2917 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2920 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2923 /* Remove page from free list */
2924 list_del(&page
->lru
);
2925 zone
->free_area
[order
].nr_free
--;
2926 rmv_page_order(page
);
2929 * Set the pageblock if the isolated page is at least half of a
2932 if (order
>= pageblock_order
- 1) {
2933 struct page
*endpage
= page
+ (1 << order
) - 1;
2934 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2935 int mt
= get_pageblock_migratetype(page
);
2936 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2937 && !is_migrate_highatomic(mt
))
2938 set_pageblock_migratetype(page
,
2944 return 1UL << order
;
2948 * Update NUMA hit/miss statistics
2950 * Must be called with interrupts disabled.
2952 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2955 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2957 /* skip numa counters update if numa stats is disabled */
2958 if (!static_branch_likely(&vm_numa_stat_key
))
2961 if (zone_to_nid(z
) != numa_node_id())
2962 local_stat
= NUMA_OTHER
;
2964 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
2965 __inc_numa_state(z
, NUMA_HIT
);
2967 __inc_numa_state(z
, NUMA_MISS
);
2968 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
2970 __inc_numa_state(z
, local_stat
);
2974 /* Remove page from the per-cpu list, caller must protect the list */
2975 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2976 unsigned int alloc_flags
,
2977 struct per_cpu_pages
*pcp
,
2978 struct list_head
*list
)
2983 if (list_empty(list
)) {
2984 pcp
->count
+= rmqueue_bulk(zone
, 0,
2986 migratetype
, alloc_flags
);
2987 if (unlikely(list_empty(list
)))
2991 page
= list_first_entry(list
, struct page
, lru
);
2992 list_del(&page
->lru
);
2994 } while (check_new_pcp(page
));
2999 /* Lock and remove page from the per-cpu list */
3000 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3001 struct zone
*zone
, unsigned int order
,
3002 gfp_t gfp_flags
, int migratetype
,
3003 unsigned int alloc_flags
)
3005 struct per_cpu_pages
*pcp
;
3006 struct list_head
*list
;
3008 unsigned long flags
;
3010 local_irq_save(flags
);
3011 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3012 list
= &pcp
->lists
[migratetype
];
3013 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3015 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3016 zone_statistics(preferred_zone
, zone
);
3018 local_irq_restore(flags
);
3023 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3026 struct page
*rmqueue(struct zone
*preferred_zone
,
3027 struct zone
*zone
, unsigned int order
,
3028 gfp_t gfp_flags
, unsigned int alloc_flags
,
3031 unsigned long flags
;
3034 if (likely(order
== 0)) {
3035 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
3036 gfp_flags
, migratetype
, alloc_flags
);
3041 * We most definitely don't want callers attempting to
3042 * allocate greater than order-1 page units with __GFP_NOFAIL.
3044 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3045 spin_lock_irqsave(&zone
->lock
, flags
);
3049 if (alloc_flags
& ALLOC_HARDER
) {
3050 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3052 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3055 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3056 } while (page
&& check_new_pages(page
, order
));
3057 spin_unlock(&zone
->lock
);
3060 __mod_zone_freepage_state(zone
, -(1 << order
),
3061 get_pcppage_migratetype(page
));
3063 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3064 zone_statistics(preferred_zone
, zone
);
3065 local_irq_restore(flags
);
3068 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3072 local_irq_restore(flags
);
3076 #ifdef CONFIG_FAIL_PAGE_ALLOC
3079 struct fault_attr attr
;
3081 bool ignore_gfp_highmem
;
3082 bool ignore_gfp_reclaim
;
3084 } fail_page_alloc
= {
3085 .attr
= FAULT_ATTR_INITIALIZER
,
3086 .ignore_gfp_reclaim
= true,
3087 .ignore_gfp_highmem
= true,
3091 static int __init
setup_fail_page_alloc(char *str
)
3093 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3095 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3097 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3099 if (order
< fail_page_alloc
.min_order
)
3101 if (gfp_mask
& __GFP_NOFAIL
)
3103 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3105 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3106 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3109 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3112 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3114 static int __init
fail_page_alloc_debugfs(void)
3116 umode_t mode
= S_IFREG
| 0600;
3119 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3120 &fail_page_alloc
.attr
);
3122 return PTR_ERR(dir
);
3124 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3125 &fail_page_alloc
.ignore_gfp_reclaim
))
3127 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3128 &fail_page_alloc
.ignore_gfp_highmem
))
3130 if (!debugfs_create_u32("min-order", mode
, dir
,
3131 &fail_page_alloc
.min_order
))
3136 debugfs_remove_recursive(dir
);
3141 late_initcall(fail_page_alloc_debugfs
);
3143 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3145 #else /* CONFIG_FAIL_PAGE_ALLOC */
3147 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3152 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3155 * Return true if free base pages are above 'mark'. For high-order checks it
3156 * will return true of the order-0 watermark is reached and there is at least
3157 * one free page of a suitable size. Checking now avoids taking the zone lock
3158 * to check in the allocation paths if no pages are free.
3160 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3161 int classzone_idx
, unsigned int alloc_flags
,
3166 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3168 /* free_pages may go negative - that's OK */
3169 free_pages
-= (1 << order
) - 1;
3171 if (alloc_flags
& ALLOC_HIGH
)
3175 * If the caller does not have rights to ALLOC_HARDER then subtract
3176 * the high-atomic reserves. This will over-estimate the size of the
3177 * atomic reserve but it avoids a search.
3179 if (likely(!alloc_harder
)) {
3180 free_pages
-= z
->nr_reserved_highatomic
;
3183 * OOM victims can try even harder than normal ALLOC_HARDER
3184 * users on the grounds that it's definitely going to be in
3185 * the exit path shortly and free memory. Any allocation it
3186 * makes during the free path will be small and short-lived.
3188 if (alloc_flags
& ALLOC_OOM
)
3196 /* If allocation can't use CMA areas don't use free CMA pages */
3197 if (!(alloc_flags
& ALLOC_CMA
))
3198 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3202 * Check watermarks for an order-0 allocation request. If these
3203 * are not met, then a high-order request also cannot go ahead
3204 * even if a suitable page happened to be free.
3206 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3209 /* If this is an order-0 request then the watermark is fine */
3213 /* For a high-order request, check at least one suitable page is free */
3214 for (o
= order
; o
< MAX_ORDER
; o
++) {
3215 struct free_area
*area
= &z
->free_area
[o
];
3221 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3222 if (!list_empty(&area
->free_list
[mt
]))
3227 if ((alloc_flags
& ALLOC_CMA
) &&
3228 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
3233 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3239 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3240 int classzone_idx
, unsigned int alloc_flags
)
3242 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3243 zone_page_state(z
, NR_FREE_PAGES
));
3246 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3247 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3249 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3253 /* If allocation can't use CMA areas don't use free CMA pages */
3254 if (!(alloc_flags
& ALLOC_CMA
))
3255 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3259 * Fast check for order-0 only. If this fails then the reserves
3260 * need to be calculated. There is a corner case where the check
3261 * passes but only the high-order atomic reserve are free. If
3262 * the caller is !atomic then it'll uselessly search the free
3263 * list. That corner case is then slower but it is harmless.
3265 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3268 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3272 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3273 unsigned long mark
, int classzone_idx
)
3275 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3277 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3278 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3280 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3285 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3287 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3290 #else /* CONFIG_NUMA */
3291 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3295 #endif /* CONFIG_NUMA */
3298 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3299 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3300 * premature use of a lower zone may cause lowmem pressure problems that
3301 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3302 * probably too small. It only makes sense to spread allocations to avoid
3303 * fragmentation between the Normal and DMA32 zones.
3305 static inline unsigned int
3306 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3308 unsigned int alloc_flags
= 0;
3310 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3311 alloc_flags
|= ALLOC_KSWAPD
;
3313 #ifdef CONFIG_ZONE_DMA32
3314 if (zone_idx(zone
) != ZONE_NORMAL
)
3318 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3319 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3320 * on UMA that if Normal is populated then so is DMA32.
3322 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3323 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3327 #endif /* CONFIG_ZONE_DMA32 */
3332 * get_page_from_freelist goes through the zonelist trying to allocate
3335 static struct page
*
3336 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3337 const struct alloc_context
*ac
)
3341 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3346 * Scan zonelist, looking for a zone with enough free.
3347 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3349 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3350 z
= ac
->preferred_zoneref
;
3351 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3356 if (cpusets_enabled() &&
3357 (alloc_flags
& ALLOC_CPUSET
) &&
3358 !__cpuset_zone_allowed(zone
, gfp_mask
))
3361 * When allocating a page cache page for writing, we
3362 * want to get it from a node that is within its dirty
3363 * limit, such that no single node holds more than its
3364 * proportional share of globally allowed dirty pages.
3365 * The dirty limits take into account the node's
3366 * lowmem reserves and high watermark so that kswapd
3367 * should be able to balance it without having to
3368 * write pages from its LRU list.
3370 * XXX: For now, allow allocations to potentially
3371 * exceed the per-node dirty limit in the slowpath
3372 * (spread_dirty_pages unset) before going into reclaim,
3373 * which is important when on a NUMA setup the allowed
3374 * nodes are together not big enough to reach the
3375 * global limit. The proper fix for these situations
3376 * will require awareness of nodes in the
3377 * dirty-throttling and the flusher threads.
3379 if (ac
->spread_dirty_pages
) {
3380 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3383 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3384 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3389 if (no_fallback
&& nr_online_nodes
> 1 &&
3390 zone
!= ac
->preferred_zoneref
->zone
) {
3394 * If moving to a remote node, retry but allow
3395 * fragmenting fallbacks. Locality is more important
3396 * than fragmentation avoidance.
3398 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3399 if (zone_to_nid(zone
) != local_nid
) {
3400 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3405 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3406 if (!zone_watermark_fast(zone
, order
, mark
,
3407 ac_classzone_idx(ac
), alloc_flags
)) {
3410 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3412 * Watermark failed for this zone, but see if we can
3413 * grow this zone if it contains deferred pages.
3415 if (static_branch_unlikely(&deferred_pages
)) {
3416 if (_deferred_grow_zone(zone
, order
))
3420 /* Checked here to keep the fast path fast */
3421 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3422 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3425 if (node_reclaim_mode
== 0 ||
3426 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3429 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3431 case NODE_RECLAIM_NOSCAN
:
3434 case NODE_RECLAIM_FULL
:
3435 /* scanned but unreclaimable */
3438 /* did we reclaim enough */
3439 if (zone_watermark_ok(zone
, order
, mark
,
3440 ac_classzone_idx(ac
), alloc_flags
))
3448 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3449 gfp_mask
, alloc_flags
, ac
->migratetype
);
3451 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3454 * If this is a high-order atomic allocation then check
3455 * if the pageblock should be reserved for the future
3457 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3458 reserve_highatomic_pageblock(page
, zone
, order
);
3462 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3463 /* Try again if zone has deferred pages */
3464 if (static_branch_unlikely(&deferred_pages
)) {
3465 if (_deferred_grow_zone(zone
, order
))
3473 * It's possible on a UMA machine to get through all zones that are
3474 * fragmented. If avoiding fragmentation, reset and try again.
3477 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3484 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3486 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3487 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3489 if (!__ratelimit(&show_mem_rs
))
3493 * This documents exceptions given to allocations in certain
3494 * contexts that are allowed to allocate outside current's set
3497 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3498 if (tsk_is_oom_victim(current
) ||
3499 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3500 filter
&= ~SHOW_MEM_FILTER_NODES
;
3501 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3502 filter
&= ~SHOW_MEM_FILTER_NODES
;
3504 show_mem(filter
, nodemask
);
3507 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3509 struct va_format vaf
;
3511 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3512 DEFAULT_RATELIMIT_BURST
);
3514 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3517 va_start(args
, fmt
);
3520 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3521 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3522 nodemask_pr_args(nodemask
));
3525 cpuset_print_current_mems_allowed();
3528 warn_alloc_show_mem(gfp_mask
, nodemask
);
3531 static inline struct page
*
3532 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3533 unsigned int alloc_flags
,
3534 const struct alloc_context
*ac
)
3538 page
= get_page_from_freelist(gfp_mask
, order
,
3539 alloc_flags
|ALLOC_CPUSET
, ac
);
3541 * fallback to ignore cpuset restriction if our nodes
3545 page
= get_page_from_freelist(gfp_mask
, order
,
3551 static inline struct page
*
3552 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3553 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3555 struct oom_control oc
= {
3556 .zonelist
= ac
->zonelist
,
3557 .nodemask
= ac
->nodemask
,
3559 .gfp_mask
= gfp_mask
,
3564 *did_some_progress
= 0;
3567 * Acquire the oom lock. If that fails, somebody else is
3568 * making progress for us.
3570 if (!mutex_trylock(&oom_lock
)) {
3571 *did_some_progress
= 1;
3572 schedule_timeout_uninterruptible(1);
3577 * Go through the zonelist yet one more time, keep very high watermark
3578 * here, this is only to catch a parallel oom killing, we must fail if
3579 * we're still under heavy pressure. But make sure that this reclaim
3580 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3581 * allocation which will never fail due to oom_lock already held.
3583 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3584 ~__GFP_DIRECT_RECLAIM
, order
,
3585 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3589 /* Coredumps can quickly deplete all memory reserves */
3590 if (current
->flags
& PF_DUMPCORE
)
3592 /* The OOM killer will not help higher order allocs */
3593 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3596 * We have already exhausted all our reclaim opportunities without any
3597 * success so it is time to admit defeat. We will skip the OOM killer
3598 * because it is very likely that the caller has a more reasonable
3599 * fallback than shooting a random task.
3601 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3603 /* The OOM killer does not needlessly kill tasks for lowmem */
3604 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3606 if (pm_suspended_storage())
3609 * XXX: GFP_NOFS allocations should rather fail than rely on
3610 * other request to make a forward progress.
3611 * We are in an unfortunate situation where out_of_memory cannot
3612 * do much for this context but let's try it to at least get
3613 * access to memory reserved if the current task is killed (see
3614 * out_of_memory). Once filesystems are ready to handle allocation
3615 * failures more gracefully we should just bail out here.
3618 /* The OOM killer may not free memory on a specific node */
3619 if (gfp_mask
& __GFP_THISNODE
)
3622 /* Exhausted what can be done so it's blame time */
3623 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3624 *did_some_progress
= 1;
3627 * Help non-failing allocations by giving them access to memory
3630 if (gfp_mask
& __GFP_NOFAIL
)
3631 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3632 ALLOC_NO_WATERMARKS
, ac
);
3635 mutex_unlock(&oom_lock
);
3640 * Maximum number of compaction retries wit a progress before OOM
3641 * killer is consider as the only way to move forward.
3643 #define MAX_COMPACT_RETRIES 16
3645 #ifdef CONFIG_COMPACTION
3646 /* Try memory compaction for high-order allocations before reclaim */
3647 static struct page
*
3648 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3649 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3650 enum compact_priority prio
, enum compact_result
*compact_result
)
3653 unsigned long pflags
;
3654 unsigned int noreclaim_flag
;
3659 psi_memstall_enter(&pflags
);
3660 noreclaim_flag
= memalloc_noreclaim_save();
3662 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3665 memalloc_noreclaim_restore(noreclaim_flag
);
3666 psi_memstall_leave(&pflags
);
3668 if (*compact_result
<= COMPACT_INACTIVE
)
3672 * At least in one zone compaction wasn't deferred or skipped, so let's
3673 * count a compaction stall
3675 count_vm_event(COMPACTSTALL
);
3677 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3680 struct zone
*zone
= page_zone(page
);
3682 zone
->compact_blockskip_flush
= false;
3683 compaction_defer_reset(zone
, order
, true);
3684 count_vm_event(COMPACTSUCCESS
);
3689 * It's bad if compaction run occurs and fails. The most likely reason
3690 * is that pages exist, but not enough to satisfy watermarks.
3692 count_vm_event(COMPACTFAIL
);
3700 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3701 enum compact_result compact_result
,
3702 enum compact_priority
*compact_priority
,
3703 int *compaction_retries
)
3705 int max_retries
= MAX_COMPACT_RETRIES
;
3708 int retries
= *compaction_retries
;
3709 enum compact_priority priority
= *compact_priority
;
3714 if (compaction_made_progress(compact_result
))
3715 (*compaction_retries
)++;
3718 * compaction considers all the zone as desperately out of memory
3719 * so it doesn't really make much sense to retry except when the
3720 * failure could be caused by insufficient priority
3722 if (compaction_failed(compact_result
))
3723 goto check_priority
;
3726 * make sure the compaction wasn't deferred or didn't bail out early
3727 * due to locks contention before we declare that we should give up.
3728 * But do not retry if the given zonelist is not suitable for
3731 if (compaction_withdrawn(compact_result
)) {
3732 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3737 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3738 * costly ones because they are de facto nofail and invoke OOM
3739 * killer to move on while costly can fail and users are ready
3740 * to cope with that. 1/4 retries is rather arbitrary but we
3741 * would need much more detailed feedback from compaction to
3742 * make a better decision.
3744 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3746 if (*compaction_retries
<= max_retries
) {
3752 * Make sure there are attempts at the highest priority if we exhausted
3753 * all retries or failed at the lower priorities.
3756 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3757 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3759 if (*compact_priority
> min_priority
) {
3760 (*compact_priority
)--;
3761 *compaction_retries
= 0;
3765 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3769 static inline struct page
*
3770 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3771 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3772 enum compact_priority prio
, enum compact_result
*compact_result
)
3774 *compact_result
= COMPACT_SKIPPED
;
3779 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3780 enum compact_result compact_result
,
3781 enum compact_priority
*compact_priority
,
3782 int *compaction_retries
)
3787 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3791 * There are setups with compaction disabled which would prefer to loop
3792 * inside the allocator rather than hit the oom killer prematurely.
3793 * Let's give them a good hope and keep retrying while the order-0
3794 * watermarks are OK.
3796 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3798 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3799 ac_classzone_idx(ac
), alloc_flags
))
3804 #endif /* CONFIG_COMPACTION */
3806 #ifdef CONFIG_LOCKDEP
3807 static struct lockdep_map __fs_reclaim_map
=
3808 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3810 static bool __need_fs_reclaim(gfp_t gfp_mask
)
3812 gfp_mask
= current_gfp_context(gfp_mask
);
3814 /* no reclaim without waiting on it */
3815 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3818 /* this guy won't enter reclaim */
3819 if (current
->flags
& PF_MEMALLOC
)
3822 /* We're only interested __GFP_FS allocations for now */
3823 if (!(gfp_mask
& __GFP_FS
))
3826 if (gfp_mask
& __GFP_NOLOCKDEP
)
3832 void __fs_reclaim_acquire(void)
3834 lock_map_acquire(&__fs_reclaim_map
);
3837 void __fs_reclaim_release(void)
3839 lock_map_release(&__fs_reclaim_map
);
3842 void fs_reclaim_acquire(gfp_t gfp_mask
)
3844 if (__need_fs_reclaim(gfp_mask
))
3845 __fs_reclaim_acquire();
3847 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3849 void fs_reclaim_release(gfp_t gfp_mask
)
3851 if (__need_fs_reclaim(gfp_mask
))
3852 __fs_reclaim_release();
3854 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3857 /* Perform direct synchronous page reclaim */
3859 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3860 const struct alloc_context
*ac
)
3862 struct reclaim_state reclaim_state
;
3864 unsigned int noreclaim_flag
;
3865 unsigned long pflags
;
3869 /* We now go into synchronous reclaim */
3870 cpuset_memory_pressure_bump();
3871 psi_memstall_enter(&pflags
);
3872 fs_reclaim_acquire(gfp_mask
);
3873 noreclaim_flag
= memalloc_noreclaim_save();
3874 reclaim_state
.reclaimed_slab
= 0;
3875 current
->reclaim_state
= &reclaim_state
;
3877 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3880 current
->reclaim_state
= NULL
;
3881 memalloc_noreclaim_restore(noreclaim_flag
);
3882 fs_reclaim_release(gfp_mask
);
3883 psi_memstall_leave(&pflags
);
3890 /* The really slow allocator path where we enter direct reclaim */
3891 static inline struct page
*
3892 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3893 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3894 unsigned long *did_some_progress
)
3896 struct page
*page
= NULL
;
3897 bool drained
= false;
3899 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3900 if (unlikely(!(*did_some_progress
)))
3904 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3907 * If an allocation failed after direct reclaim, it could be because
3908 * pages are pinned on the per-cpu lists or in high alloc reserves.
3909 * Shrink them them and try again
3911 if (!page
&& !drained
) {
3912 unreserve_highatomic_pageblock(ac
, false);
3913 drain_all_pages(NULL
);
3921 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
3922 const struct alloc_context
*ac
)
3926 pg_data_t
*last_pgdat
= NULL
;
3927 enum zone_type high_zoneidx
= ac
->high_zoneidx
;
3929 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, high_zoneidx
,
3931 if (last_pgdat
!= zone
->zone_pgdat
)
3932 wakeup_kswapd(zone
, gfp_mask
, order
, high_zoneidx
);
3933 last_pgdat
= zone
->zone_pgdat
;
3937 static inline unsigned int
3938 gfp_to_alloc_flags(gfp_t gfp_mask
)
3940 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3942 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3943 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3946 * The caller may dip into page reserves a bit more if the caller
3947 * cannot run direct reclaim, or if the caller has realtime scheduling
3948 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3949 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3951 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3953 if (gfp_mask
& __GFP_ATOMIC
) {
3955 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3956 * if it can't schedule.
3958 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3959 alloc_flags
|= ALLOC_HARDER
;
3961 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3962 * comment for __cpuset_node_allowed().
3964 alloc_flags
&= ~ALLOC_CPUSET
;
3965 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3966 alloc_flags
|= ALLOC_HARDER
;
3968 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3969 alloc_flags
|= ALLOC_KSWAPD
;
3972 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3973 alloc_flags
|= ALLOC_CMA
;
3978 static bool oom_reserves_allowed(struct task_struct
*tsk
)
3980 if (!tsk_is_oom_victim(tsk
))
3984 * !MMU doesn't have oom reaper so give access to memory reserves
3985 * only to the thread with TIF_MEMDIE set
3987 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
3994 * Distinguish requests which really need access to full memory
3995 * reserves from oom victims which can live with a portion of it
3997 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
3999 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4001 if (gfp_mask
& __GFP_MEMALLOC
)
4002 return ALLOC_NO_WATERMARKS
;
4003 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4004 return ALLOC_NO_WATERMARKS
;
4005 if (!in_interrupt()) {
4006 if (current
->flags
& PF_MEMALLOC
)
4007 return ALLOC_NO_WATERMARKS
;
4008 else if (oom_reserves_allowed(current
))
4015 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4017 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4021 * Checks whether it makes sense to retry the reclaim to make a forward progress
4022 * for the given allocation request.
4024 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4025 * without success, or when we couldn't even meet the watermark if we
4026 * reclaimed all remaining pages on the LRU lists.
4028 * Returns true if a retry is viable or false to enter the oom path.
4031 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4032 struct alloc_context
*ac
, int alloc_flags
,
4033 bool did_some_progress
, int *no_progress_loops
)
4040 * Costly allocations might have made a progress but this doesn't mean
4041 * their order will become available due to high fragmentation so
4042 * always increment the no progress counter for them
4044 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4045 *no_progress_loops
= 0;
4047 (*no_progress_loops
)++;
4050 * Make sure we converge to OOM if we cannot make any progress
4051 * several times in the row.
4053 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4054 /* Before OOM, exhaust highatomic_reserve */
4055 return unreserve_highatomic_pageblock(ac
, true);
4059 * Keep reclaiming pages while there is a chance this will lead
4060 * somewhere. If none of the target zones can satisfy our allocation
4061 * request even if all reclaimable pages are considered then we are
4062 * screwed and have to go OOM.
4064 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4066 unsigned long available
;
4067 unsigned long reclaimable
;
4068 unsigned long min_wmark
= min_wmark_pages(zone
);
4071 available
= reclaimable
= zone_reclaimable_pages(zone
);
4072 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4075 * Would the allocation succeed if we reclaimed all
4076 * reclaimable pages?
4078 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4079 ac_classzone_idx(ac
), alloc_flags
, available
);
4080 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4081 available
, min_wmark
, *no_progress_loops
, wmark
);
4084 * If we didn't make any progress and have a lot of
4085 * dirty + writeback pages then we should wait for
4086 * an IO to complete to slow down the reclaim and
4087 * prevent from pre mature OOM
4089 if (!did_some_progress
) {
4090 unsigned long write_pending
;
4092 write_pending
= zone_page_state_snapshot(zone
,
4093 NR_ZONE_WRITE_PENDING
);
4095 if (2 * write_pending
> reclaimable
) {
4096 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4108 * Memory allocation/reclaim might be called from a WQ context and the
4109 * current implementation of the WQ concurrency control doesn't
4110 * recognize that a particular WQ is congested if the worker thread is
4111 * looping without ever sleeping. Therefore we have to do a short sleep
4112 * here rather than calling cond_resched().
4114 if (current
->flags
& PF_WQ_WORKER
)
4115 schedule_timeout_uninterruptible(1);
4122 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4125 * It's possible that cpuset's mems_allowed and the nodemask from
4126 * mempolicy don't intersect. This should be normally dealt with by
4127 * policy_nodemask(), but it's possible to race with cpuset update in
4128 * such a way the check therein was true, and then it became false
4129 * before we got our cpuset_mems_cookie here.
4130 * This assumes that for all allocations, ac->nodemask can come only
4131 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4132 * when it does not intersect with the cpuset restrictions) or the
4133 * caller can deal with a violated nodemask.
4135 if (cpusets_enabled() && ac
->nodemask
&&
4136 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4137 ac
->nodemask
= NULL
;
4142 * When updating a task's mems_allowed or mempolicy nodemask, it is
4143 * possible to race with parallel threads in such a way that our
4144 * allocation can fail while the mask is being updated. If we are about
4145 * to fail, check if the cpuset changed during allocation and if so,
4148 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4154 static inline struct page
*
4155 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4156 struct alloc_context
*ac
)
4158 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4159 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4160 struct page
*page
= NULL
;
4161 unsigned int alloc_flags
;
4162 unsigned long did_some_progress
;
4163 enum compact_priority compact_priority
;
4164 enum compact_result compact_result
;
4165 int compaction_retries
;
4166 int no_progress_loops
;
4167 unsigned int cpuset_mems_cookie
;
4171 * We also sanity check to catch abuse of atomic reserves being used by
4172 * callers that are not in atomic context.
4174 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4175 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4176 gfp_mask
&= ~__GFP_ATOMIC
;
4179 compaction_retries
= 0;
4180 no_progress_loops
= 0;
4181 compact_priority
= DEF_COMPACT_PRIORITY
;
4182 cpuset_mems_cookie
= read_mems_allowed_begin();
4185 * The fast path uses conservative alloc_flags to succeed only until
4186 * kswapd needs to be woken up, and to avoid the cost of setting up
4187 * alloc_flags precisely. So we do that now.
4189 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4192 * We need to recalculate the starting point for the zonelist iterator
4193 * because we might have used different nodemask in the fast path, or
4194 * there was a cpuset modification and we are retrying - otherwise we
4195 * could end up iterating over non-eligible zones endlessly.
4197 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4198 ac
->high_zoneidx
, ac
->nodemask
);
4199 if (!ac
->preferred_zoneref
->zone
)
4202 if (alloc_flags
& ALLOC_KSWAPD
)
4203 wake_all_kswapds(order
, gfp_mask
, ac
);
4206 * The adjusted alloc_flags might result in immediate success, so try
4209 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4214 * For costly allocations, try direct compaction first, as it's likely
4215 * that we have enough base pages and don't need to reclaim. For non-
4216 * movable high-order allocations, do that as well, as compaction will
4217 * try prevent permanent fragmentation by migrating from blocks of the
4219 * Don't try this for allocations that are allowed to ignore
4220 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4222 if (can_direct_reclaim
&&
4224 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4225 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4226 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4228 INIT_COMPACT_PRIORITY
,
4234 * Checks for costly allocations with __GFP_NORETRY, which
4235 * includes THP page fault allocations
4237 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4239 * If compaction is deferred for high-order allocations,
4240 * it is because sync compaction recently failed. If
4241 * this is the case and the caller requested a THP
4242 * allocation, we do not want to heavily disrupt the
4243 * system, so we fail the allocation instead of entering
4246 if (compact_result
== COMPACT_DEFERRED
)
4250 * Looks like reclaim/compaction is worth trying, but
4251 * sync compaction could be very expensive, so keep
4252 * using async compaction.
4254 compact_priority
= INIT_COMPACT_PRIORITY
;
4259 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4260 if (alloc_flags
& ALLOC_KSWAPD
)
4261 wake_all_kswapds(order
, gfp_mask
, ac
);
4263 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4265 alloc_flags
= reserve_flags
;
4268 * Reset the nodemask and zonelist iterators if memory policies can be
4269 * ignored. These allocations are high priority and system rather than
4272 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4273 ac
->nodemask
= NULL
;
4274 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4275 ac
->high_zoneidx
, ac
->nodemask
);
4278 /* Attempt with potentially adjusted zonelist and alloc_flags */
4279 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4283 /* Caller is not willing to reclaim, we can't balance anything */
4284 if (!can_direct_reclaim
)
4287 /* Avoid recursion of direct reclaim */
4288 if (current
->flags
& PF_MEMALLOC
)
4291 /* Try direct reclaim and then allocating */
4292 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4293 &did_some_progress
);
4297 /* Try direct compaction and then allocating */
4298 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4299 compact_priority
, &compact_result
);
4303 /* Do not loop if specifically requested */
4304 if (gfp_mask
& __GFP_NORETRY
)
4308 * Do not retry costly high order allocations unless they are
4309 * __GFP_RETRY_MAYFAIL
4311 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4314 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4315 did_some_progress
> 0, &no_progress_loops
))
4319 * It doesn't make any sense to retry for the compaction if the order-0
4320 * reclaim is not able to make any progress because the current
4321 * implementation of the compaction depends on the sufficient amount
4322 * of free memory (see __compaction_suitable)
4324 if (did_some_progress
> 0 &&
4325 should_compact_retry(ac
, order
, alloc_flags
,
4326 compact_result
, &compact_priority
,
4327 &compaction_retries
))
4331 /* Deal with possible cpuset update races before we start OOM killing */
4332 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4335 /* Reclaim has failed us, start killing things */
4336 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4340 /* Avoid allocations with no watermarks from looping endlessly */
4341 if (tsk_is_oom_victim(current
) &&
4342 (alloc_flags
== ALLOC_OOM
||
4343 (gfp_mask
& __GFP_NOMEMALLOC
)))
4346 /* Retry as long as the OOM killer is making progress */
4347 if (did_some_progress
) {
4348 no_progress_loops
= 0;
4353 /* Deal with possible cpuset update races before we fail */
4354 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4358 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4361 if (gfp_mask
& __GFP_NOFAIL
) {
4363 * All existing users of the __GFP_NOFAIL are blockable, so warn
4364 * of any new users that actually require GFP_NOWAIT
4366 if (WARN_ON_ONCE(!can_direct_reclaim
))
4370 * PF_MEMALLOC request from this context is rather bizarre
4371 * because we cannot reclaim anything and only can loop waiting
4372 * for somebody to do a work for us
4374 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4377 * non failing costly orders are a hard requirement which we
4378 * are not prepared for much so let's warn about these users
4379 * so that we can identify them and convert them to something
4382 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4385 * Help non-failing allocations by giving them access to memory
4386 * reserves but do not use ALLOC_NO_WATERMARKS because this
4387 * could deplete whole memory reserves which would just make
4388 * the situation worse
4390 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4398 warn_alloc(gfp_mask
, ac
->nodemask
,
4399 "page allocation failure: order:%u", order
);
4404 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4405 int preferred_nid
, nodemask_t
*nodemask
,
4406 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4407 unsigned int *alloc_flags
)
4409 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4410 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4411 ac
->nodemask
= nodemask
;
4412 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4414 if (cpusets_enabled()) {
4415 *alloc_mask
|= __GFP_HARDWALL
;
4417 ac
->nodemask
= &cpuset_current_mems_allowed
;
4419 *alloc_flags
|= ALLOC_CPUSET
;
4422 fs_reclaim_acquire(gfp_mask
);
4423 fs_reclaim_release(gfp_mask
);
4425 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4427 if (should_fail_alloc_page(gfp_mask
, order
))
4430 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4431 *alloc_flags
|= ALLOC_CMA
;
4436 /* Determine whether to spread dirty pages and what the first usable zone */
4437 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4439 /* Dirty zone balancing only done in the fast path */
4440 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4443 * The preferred zone is used for statistics but crucially it is
4444 * also used as the starting point for the zonelist iterator. It
4445 * may get reset for allocations that ignore memory policies.
4447 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4448 ac
->high_zoneidx
, ac
->nodemask
);
4452 * This is the 'heart' of the zoned buddy allocator.
4455 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4456 nodemask_t
*nodemask
)
4459 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4460 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4461 struct alloc_context ac
= { };
4464 * There are several places where we assume that the order value is sane
4465 * so bail out early if the request is out of bound.
4467 if (unlikely(order
>= MAX_ORDER
)) {
4468 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4472 gfp_mask
&= gfp_allowed_mask
;
4473 alloc_mask
= gfp_mask
;
4474 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4477 finalise_ac(gfp_mask
, &ac
);
4480 * Forbid the first pass from falling back to types that fragment
4481 * memory until all local zones are considered.
4483 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4485 /* First allocation attempt */
4486 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4491 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4492 * resp. GFP_NOIO which has to be inherited for all allocation requests
4493 * from a particular context which has been marked by
4494 * memalloc_no{fs,io}_{save,restore}.
4496 alloc_mask
= current_gfp_context(gfp_mask
);
4497 ac
.spread_dirty_pages
= false;
4500 * Restore the original nodemask if it was potentially replaced with
4501 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4503 if (unlikely(ac
.nodemask
!= nodemask
))
4504 ac
.nodemask
= nodemask
;
4506 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4509 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4510 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4511 __free_pages(page
, order
);
4515 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4519 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4522 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4523 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4524 * you need to access high mem.
4526 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4530 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4533 return (unsigned long) page_address(page
);
4535 EXPORT_SYMBOL(__get_free_pages
);
4537 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4539 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4541 EXPORT_SYMBOL(get_zeroed_page
);
4543 static inline void free_the_page(struct page
*page
, unsigned int order
)
4545 if (order
== 0) /* Via pcp? */
4546 free_unref_page(page
);
4548 __free_pages_ok(page
, order
);
4551 void __free_pages(struct page
*page
, unsigned int order
)
4553 if (put_page_testzero(page
))
4554 free_the_page(page
, order
);
4556 EXPORT_SYMBOL(__free_pages
);
4558 void free_pages(unsigned long addr
, unsigned int order
)
4561 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4562 __free_pages(virt_to_page((void *)addr
), order
);
4566 EXPORT_SYMBOL(free_pages
);
4570 * An arbitrary-length arbitrary-offset area of memory which resides
4571 * within a 0 or higher order page. Multiple fragments within that page
4572 * are individually refcounted, in the page's reference counter.
4574 * The page_frag functions below provide a simple allocation framework for
4575 * page fragments. This is used by the network stack and network device
4576 * drivers to provide a backing region of memory for use as either an
4577 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4579 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4582 struct page
*page
= NULL
;
4583 gfp_t gfp
= gfp_mask
;
4585 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4586 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4588 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4589 PAGE_FRAG_CACHE_MAX_ORDER
);
4590 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4592 if (unlikely(!page
))
4593 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4595 nc
->va
= page
? page_address(page
) : NULL
;
4600 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4602 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4604 if (page_ref_sub_and_test(page
, count
))
4605 free_the_page(page
, compound_order(page
));
4607 EXPORT_SYMBOL(__page_frag_cache_drain
);
4609 void *page_frag_alloc(struct page_frag_cache
*nc
,
4610 unsigned int fragsz
, gfp_t gfp_mask
)
4612 unsigned int size
= PAGE_SIZE
;
4616 if (unlikely(!nc
->va
)) {
4618 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4622 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4623 /* if size can vary use size else just use PAGE_SIZE */
4626 /* Even if we own the page, we do not use atomic_set().
4627 * This would break get_page_unless_zero() users.
4629 page_ref_add(page
, size
- 1);
4631 /* reset page count bias and offset to start of new frag */
4632 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4633 nc
->pagecnt_bias
= size
;
4637 offset
= nc
->offset
- fragsz
;
4638 if (unlikely(offset
< 0)) {
4639 page
= virt_to_page(nc
->va
);
4641 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4644 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4645 /* if size can vary use size else just use PAGE_SIZE */
4648 /* OK, page count is 0, we can safely set it */
4649 set_page_count(page
, size
);
4651 /* reset page count bias and offset to start of new frag */
4652 nc
->pagecnt_bias
= size
;
4653 offset
= size
- fragsz
;
4657 nc
->offset
= offset
;
4659 return nc
->va
+ offset
;
4661 EXPORT_SYMBOL(page_frag_alloc
);
4664 * Frees a page fragment allocated out of either a compound or order 0 page.
4666 void page_frag_free(void *addr
)
4668 struct page
*page
= virt_to_head_page(addr
);
4670 if (unlikely(put_page_testzero(page
)))
4671 free_the_page(page
, compound_order(page
));
4673 EXPORT_SYMBOL(page_frag_free
);
4675 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4679 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4680 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4682 split_page(virt_to_page((void *)addr
), order
);
4683 while (used
< alloc_end
) {
4688 return (void *)addr
;
4692 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4693 * @size: the number of bytes to allocate
4694 * @gfp_mask: GFP flags for the allocation
4696 * This function is similar to alloc_pages(), except that it allocates the
4697 * minimum number of pages to satisfy the request. alloc_pages() can only
4698 * allocate memory in power-of-two pages.
4700 * This function is also limited by MAX_ORDER.
4702 * Memory allocated by this function must be released by free_pages_exact().
4704 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4706 unsigned int order
= get_order(size
);
4709 addr
= __get_free_pages(gfp_mask
, order
);
4710 return make_alloc_exact(addr
, order
, size
);
4712 EXPORT_SYMBOL(alloc_pages_exact
);
4715 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4717 * @nid: the preferred node ID where memory should be allocated
4718 * @size: the number of bytes to allocate
4719 * @gfp_mask: GFP flags for the allocation
4721 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4724 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4726 unsigned int order
= get_order(size
);
4727 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4730 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4734 * free_pages_exact - release memory allocated via alloc_pages_exact()
4735 * @virt: the value returned by alloc_pages_exact.
4736 * @size: size of allocation, same value as passed to alloc_pages_exact().
4738 * Release the memory allocated by a previous call to alloc_pages_exact.
4740 void free_pages_exact(void *virt
, size_t size
)
4742 unsigned long addr
= (unsigned long)virt
;
4743 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4745 while (addr
< end
) {
4750 EXPORT_SYMBOL(free_pages_exact
);
4753 * nr_free_zone_pages - count number of pages beyond high watermark
4754 * @offset: The zone index of the highest zone
4756 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4757 * high watermark within all zones at or below a given zone index. For each
4758 * zone, the number of pages is calculated as:
4760 * nr_free_zone_pages = managed_pages - high_pages
4762 static unsigned long nr_free_zone_pages(int offset
)
4767 /* Just pick one node, since fallback list is circular */
4768 unsigned long sum
= 0;
4770 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4772 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4773 unsigned long size
= zone_managed_pages(zone
);
4774 unsigned long high
= high_wmark_pages(zone
);
4783 * nr_free_buffer_pages - count number of pages beyond high watermark
4785 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4786 * watermark within ZONE_DMA and ZONE_NORMAL.
4788 unsigned long nr_free_buffer_pages(void)
4790 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4792 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4795 * nr_free_pagecache_pages - count number of pages beyond high watermark
4797 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4798 * high watermark within all zones.
4800 unsigned long nr_free_pagecache_pages(void)
4802 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4805 static inline void show_node(struct zone
*zone
)
4807 if (IS_ENABLED(CONFIG_NUMA
))
4808 printk("Node %d ", zone_to_nid(zone
));
4811 long si_mem_available(void)
4814 unsigned long pagecache
;
4815 unsigned long wmark_low
= 0;
4816 unsigned long pages
[NR_LRU_LISTS
];
4817 unsigned long reclaimable
;
4821 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4822 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4825 wmark_low
+= low_wmark_pages(zone
);
4828 * Estimate the amount of memory available for userspace allocations,
4829 * without causing swapping.
4831 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4834 * Not all the page cache can be freed, otherwise the system will
4835 * start swapping. Assume at least half of the page cache, or the
4836 * low watermark worth of cache, needs to stay.
4838 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4839 pagecache
-= min(pagecache
/ 2, wmark_low
);
4840 available
+= pagecache
;
4843 * Part of the reclaimable slab and other kernel memory consists of
4844 * items that are in use, and cannot be freed. Cap this estimate at the
4847 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
4848 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
4849 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
4855 EXPORT_SYMBOL_GPL(si_mem_available
);
4857 void si_meminfo(struct sysinfo
*val
)
4859 val
->totalram
= totalram_pages();
4860 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4861 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
4862 val
->bufferram
= nr_blockdev_pages();
4863 val
->totalhigh
= totalhigh_pages();
4864 val
->freehigh
= nr_free_highpages();
4865 val
->mem_unit
= PAGE_SIZE
;
4868 EXPORT_SYMBOL(si_meminfo
);
4871 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4873 int zone_type
; /* needs to be signed */
4874 unsigned long managed_pages
= 0;
4875 unsigned long managed_highpages
= 0;
4876 unsigned long free_highpages
= 0;
4877 pg_data_t
*pgdat
= NODE_DATA(nid
);
4879 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4880 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
4881 val
->totalram
= managed_pages
;
4882 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4883 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4884 #ifdef CONFIG_HIGHMEM
4885 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4886 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4888 if (is_highmem(zone
)) {
4889 managed_highpages
+= zone_managed_pages(zone
);
4890 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4893 val
->totalhigh
= managed_highpages
;
4894 val
->freehigh
= free_highpages
;
4896 val
->totalhigh
= managed_highpages
;
4897 val
->freehigh
= free_highpages
;
4899 val
->mem_unit
= PAGE_SIZE
;
4904 * Determine whether the node should be displayed or not, depending on whether
4905 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4907 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4909 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4913 * no node mask - aka implicit memory numa policy. Do not bother with
4914 * the synchronization - read_mems_allowed_begin - because we do not
4915 * have to be precise here.
4918 nodemask
= &cpuset_current_mems_allowed
;
4920 return !node_isset(nid
, *nodemask
);
4923 #define K(x) ((x) << (PAGE_SHIFT-10))
4925 static void show_migration_types(unsigned char type
)
4927 static const char types
[MIGRATE_TYPES
] = {
4928 [MIGRATE_UNMOVABLE
] = 'U',
4929 [MIGRATE_MOVABLE
] = 'M',
4930 [MIGRATE_RECLAIMABLE
] = 'E',
4931 [MIGRATE_HIGHATOMIC
] = 'H',
4933 [MIGRATE_CMA
] = 'C',
4935 #ifdef CONFIG_MEMORY_ISOLATION
4936 [MIGRATE_ISOLATE
] = 'I',
4939 char tmp
[MIGRATE_TYPES
+ 1];
4943 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4944 if (type
& (1 << i
))
4949 printk(KERN_CONT
"(%s) ", tmp
);
4953 * Show free area list (used inside shift_scroll-lock stuff)
4954 * We also calculate the percentage fragmentation. We do this by counting the
4955 * memory on each free list with the exception of the first item on the list.
4958 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4961 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4963 unsigned long free_pcp
= 0;
4968 for_each_populated_zone(zone
) {
4969 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4972 for_each_online_cpu(cpu
)
4973 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4976 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4977 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4978 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4979 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4980 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4981 " free:%lu free_pcp:%lu free_cma:%lu\n",
4982 global_node_page_state(NR_ACTIVE_ANON
),
4983 global_node_page_state(NR_INACTIVE_ANON
),
4984 global_node_page_state(NR_ISOLATED_ANON
),
4985 global_node_page_state(NR_ACTIVE_FILE
),
4986 global_node_page_state(NR_INACTIVE_FILE
),
4987 global_node_page_state(NR_ISOLATED_FILE
),
4988 global_node_page_state(NR_UNEVICTABLE
),
4989 global_node_page_state(NR_FILE_DIRTY
),
4990 global_node_page_state(NR_WRITEBACK
),
4991 global_node_page_state(NR_UNSTABLE_NFS
),
4992 global_node_page_state(NR_SLAB_RECLAIMABLE
),
4993 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
4994 global_node_page_state(NR_FILE_MAPPED
),
4995 global_node_page_state(NR_SHMEM
),
4996 global_zone_page_state(NR_PAGETABLE
),
4997 global_zone_page_state(NR_BOUNCE
),
4998 global_zone_page_state(NR_FREE_PAGES
),
5000 global_zone_page_state(NR_FREE_CMA_PAGES
));
5002 for_each_online_pgdat(pgdat
) {
5003 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5007 " active_anon:%lukB"
5008 " inactive_anon:%lukB"
5009 " active_file:%lukB"
5010 " inactive_file:%lukB"
5011 " unevictable:%lukB"
5012 " isolated(anon):%lukB"
5013 " isolated(file):%lukB"
5018 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5020 " shmem_pmdmapped: %lukB"
5023 " writeback_tmp:%lukB"
5025 " all_unreclaimable? %s"
5028 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5029 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5030 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5031 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5032 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5033 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5034 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5035 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5036 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5037 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5038 K(node_page_state(pgdat
, NR_SHMEM
)),
5039 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5040 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5041 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5043 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5045 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5046 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
5047 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5051 for_each_populated_zone(zone
) {
5054 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5058 for_each_online_cpu(cpu
)
5059 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5068 " active_anon:%lukB"
5069 " inactive_anon:%lukB"
5070 " active_file:%lukB"
5071 " inactive_file:%lukB"
5072 " unevictable:%lukB"
5073 " writepending:%lukB"
5077 " kernel_stack:%lukB"
5085 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5086 K(min_wmark_pages(zone
)),
5087 K(low_wmark_pages(zone
)),
5088 K(high_wmark_pages(zone
)),
5089 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5090 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5091 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5092 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5093 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5094 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5095 K(zone
->present_pages
),
5096 K(zone_managed_pages(zone
)),
5097 K(zone_page_state(zone
, NR_MLOCK
)),
5098 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5099 K(zone_page_state(zone
, NR_PAGETABLE
)),
5100 K(zone_page_state(zone
, NR_BOUNCE
)),
5102 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5103 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5104 printk("lowmem_reserve[]:");
5105 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5106 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5107 printk(KERN_CONT
"\n");
5110 for_each_populated_zone(zone
) {
5112 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5113 unsigned char types
[MAX_ORDER
];
5115 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5118 printk(KERN_CONT
"%s: ", zone
->name
);
5120 spin_lock_irqsave(&zone
->lock
, flags
);
5121 for (order
= 0; order
< MAX_ORDER
; order
++) {
5122 struct free_area
*area
= &zone
->free_area
[order
];
5125 nr
[order
] = area
->nr_free
;
5126 total
+= nr
[order
] << order
;
5129 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5130 if (!list_empty(&area
->free_list
[type
]))
5131 types
[order
] |= 1 << type
;
5134 spin_unlock_irqrestore(&zone
->lock
, flags
);
5135 for (order
= 0; order
< MAX_ORDER
; order
++) {
5136 printk(KERN_CONT
"%lu*%lukB ",
5137 nr
[order
], K(1UL) << order
);
5139 show_migration_types(types
[order
]);
5141 printk(KERN_CONT
"= %lukB\n", K(total
));
5144 hugetlb_show_meminfo();
5146 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5148 show_swap_cache_info();
5151 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5153 zoneref
->zone
= zone
;
5154 zoneref
->zone_idx
= zone_idx(zone
);
5158 * Builds allocation fallback zone lists.
5160 * Add all populated zones of a node to the zonelist.
5162 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5165 enum zone_type zone_type
= MAX_NR_ZONES
;
5170 zone
= pgdat
->node_zones
+ zone_type
;
5171 if (managed_zone(zone
)) {
5172 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5173 check_highest_zone(zone_type
);
5175 } while (zone_type
);
5182 static int __parse_numa_zonelist_order(char *s
)
5185 * We used to support different zonlists modes but they turned
5186 * out to be just not useful. Let's keep the warning in place
5187 * if somebody still use the cmd line parameter so that we do
5188 * not fail it silently
5190 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5191 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5197 static __init
int setup_numa_zonelist_order(char *s
)
5202 return __parse_numa_zonelist_order(s
);
5204 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
5206 char numa_zonelist_order
[] = "Node";
5209 * sysctl handler for numa_zonelist_order
5211 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5212 void __user
*buffer
, size_t *length
,
5219 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5220 str
= memdup_user_nul(buffer
, 16);
5222 return PTR_ERR(str
);
5224 ret
= __parse_numa_zonelist_order(str
);
5230 #define MAX_NODE_LOAD (nr_online_nodes)
5231 static int node_load
[MAX_NUMNODES
];
5234 * find_next_best_node - find the next node that should appear in a given node's fallback list
5235 * @node: node whose fallback list we're appending
5236 * @used_node_mask: nodemask_t of already used nodes
5238 * We use a number of factors to determine which is the next node that should
5239 * appear on a given node's fallback list. The node should not have appeared
5240 * already in @node's fallback list, and it should be the next closest node
5241 * according to the distance array (which contains arbitrary distance values
5242 * from each node to each node in the system), and should also prefer nodes
5243 * with no CPUs, since presumably they'll have very little allocation pressure
5244 * on them otherwise.
5245 * It returns -1 if no node is found.
5247 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5250 int min_val
= INT_MAX
;
5251 int best_node
= NUMA_NO_NODE
;
5252 const struct cpumask
*tmp
= cpumask_of_node(0);
5254 /* Use the local node if we haven't already */
5255 if (!node_isset(node
, *used_node_mask
)) {
5256 node_set(node
, *used_node_mask
);
5260 for_each_node_state(n
, N_MEMORY
) {
5262 /* Don't want a node to appear more than once */
5263 if (node_isset(n
, *used_node_mask
))
5266 /* Use the distance array to find the distance */
5267 val
= node_distance(node
, n
);
5269 /* Penalize nodes under us ("prefer the next node") */
5272 /* Give preference to headless and unused nodes */
5273 tmp
= cpumask_of_node(n
);
5274 if (!cpumask_empty(tmp
))
5275 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5277 /* Slight preference for less loaded node */
5278 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5279 val
+= node_load
[n
];
5281 if (val
< min_val
) {
5288 node_set(best_node
, *used_node_mask
);
5295 * Build zonelists ordered by node and zones within node.
5296 * This results in maximum locality--normal zone overflows into local
5297 * DMA zone, if any--but risks exhausting DMA zone.
5299 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5302 struct zoneref
*zonerefs
;
5305 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5307 for (i
= 0; i
< nr_nodes
; i
++) {
5310 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5312 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5313 zonerefs
+= nr_zones
;
5315 zonerefs
->zone
= NULL
;
5316 zonerefs
->zone_idx
= 0;
5320 * Build gfp_thisnode zonelists
5322 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5324 struct zoneref
*zonerefs
;
5327 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5328 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5329 zonerefs
+= nr_zones
;
5330 zonerefs
->zone
= NULL
;
5331 zonerefs
->zone_idx
= 0;
5335 * Build zonelists ordered by zone and nodes within zones.
5336 * This results in conserving DMA zone[s] until all Normal memory is
5337 * exhausted, but results in overflowing to remote node while memory
5338 * may still exist in local DMA zone.
5341 static void build_zonelists(pg_data_t
*pgdat
)
5343 static int node_order
[MAX_NUMNODES
];
5344 int node
, load
, nr_nodes
= 0;
5345 nodemask_t used_mask
;
5346 int local_node
, prev_node
;
5348 /* NUMA-aware ordering of nodes */
5349 local_node
= pgdat
->node_id
;
5350 load
= nr_online_nodes
;
5351 prev_node
= local_node
;
5352 nodes_clear(used_mask
);
5354 memset(node_order
, 0, sizeof(node_order
));
5355 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5357 * We don't want to pressure a particular node.
5358 * So adding penalty to the first node in same
5359 * distance group to make it round-robin.
5361 if (node_distance(local_node
, node
) !=
5362 node_distance(local_node
, prev_node
))
5363 node_load
[node
] = load
;
5365 node_order
[nr_nodes
++] = node
;
5370 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5371 build_thisnode_zonelists(pgdat
);
5374 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5376 * Return node id of node used for "local" allocations.
5377 * I.e., first node id of first zone in arg node's generic zonelist.
5378 * Used for initializing percpu 'numa_mem', which is used primarily
5379 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5381 int local_memory_node(int node
)
5385 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5386 gfp_zone(GFP_KERNEL
),
5388 return zone_to_nid(z
->zone
);
5392 static void setup_min_unmapped_ratio(void);
5393 static void setup_min_slab_ratio(void);
5394 #else /* CONFIG_NUMA */
5396 static void build_zonelists(pg_data_t
*pgdat
)
5398 int node
, local_node
;
5399 struct zoneref
*zonerefs
;
5402 local_node
= pgdat
->node_id
;
5404 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5405 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5406 zonerefs
+= nr_zones
;
5409 * Now we build the zonelist so that it contains the zones
5410 * of all the other nodes.
5411 * We don't want to pressure a particular node, so when
5412 * building the zones for node N, we make sure that the
5413 * zones coming right after the local ones are those from
5414 * node N+1 (modulo N)
5416 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5417 if (!node_online(node
))
5419 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5420 zonerefs
+= nr_zones
;
5422 for (node
= 0; node
< local_node
; node
++) {
5423 if (!node_online(node
))
5425 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5426 zonerefs
+= nr_zones
;
5429 zonerefs
->zone
= NULL
;
5430 zonerefs
->zone_idx
= 0;
5433 #endif /* CONFIG_NUMA */
5436 * Boot pageset table. One per cpu which is going to be used for all
5437 * zones and all nodes. The parameters will be set in such a way
5438 * that an item put on a list will immediately be handed over to
5439 * the buddy list. This is safe since pageset manipulation is done
5440 * with interrupts disabled.
5442 * The boot_pagesets must be kept even after bootup is complete for
5443 * unused processors and/or zones. They do play a role for bootstrapping
5444 * hotplugged processors.
5446 * zoneinfo_show() and maybe other functions do
5447 * not check if the processor is online before following the pageset pointer.
5448 * Other parts of the kernel may not check if the zone is available.
5450 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5451 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5452 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5454 static void __build_all_zonelists(void *data
)
5457 int __maybe_unused cpu
;
5458 pg_data_t
*self
= data
;
5459 static DEFINE_SPINLOCK(lock
);
5464 memset(node_load
, 0, sizeof(node_load
));
5468 * This node is hotadded and no memory is yet present. So just
5469 * building zonelists is fine - no need to touch other nodes.
5471 if (self
&& !node_online(self
->node_id
)) {
5472 build_zonelists(self
);
5474 for_each_online_node(nid
) {
5475 pg_data_t
*pgdat
= NODE_DATA(nid
);
5477 build_zonelists(pgdat
);
5480 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5482 * We now know the "local memory node" for each node--
5483 * i.e., the node of the first zone in the generic zonelist.
5484 * Set up numa_mem percpu variable for on-line cpus. During
5485 * boot, only the boot cpu should be on-line; we'll init the
5486 * secondary cpus' numa_mem as they come on-line. During
5487 * node/memory hotplug, we'll fixup all on-line cpus.
5489 for_each_online_cpu(cpu
)
5490 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5497 static noinline
void __init
5498 build_all_zonelists_init(void)
5502 __build_all_zonelists(NULL
);
5505 * Initialize the boot_pagesets that are going to be used
5506 * for bootstrapping processors. The real pagesets for
5507 * each zone will be allocated later when the per cpu
5508 * allocator is available.
5510 * boot_pagesets are used also for bootstrapping offline
5511 * cpus if the system is already booted because the pagesets
5512 * are needed to initialize allocators on a specific cpu too.
5513 * F.e. the percpu allocator needs the page allocator which
5514 * needs the percpu allocator in order to allocate its pagesets
5515 * (a chicken-egg dilemma).
5517 for_each_possible_cpu(cpu
)
5518 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5520 mminit_verify_zonelist();
5521 cpuset_init_current_mems_allowed();
5525 * unless system_state == SYSTEM_BOOTING.
5527 * __ref due to call of __init annotated helper build_all_zonelists_init
5528 * [protected by SYSTEM_BOOTING].
5530 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5532 if (system_state
== SYSTEM_BOOTING
) {
5533 build_all_zonelists_init();
5535 __build_all_zonelists(pgdat
);
5536 /* cpuset refresh routine should be here */
5538 vm_total_pages
= nr_free_pagecache_pages();
5540 * Disable grouping by mobility if the number of pages in the
5541 * system is too low to allow the mechanism to work. It would be
5542 * more accurate, but expensive to check per-zone. This check is
5543 * made on memory-hotadd so a system can start with mobility
5544 * disabled and enable it later
5546 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5547 page_group_by_mobility_disabled
= 1;
5549 page_group_by_mobility_disabled
= 0;
5551 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5553 page_group_by_mobility_disabled
? "off" : "on",
5556 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5560 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5561 static bool __meminit
5562 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5564 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5565 static struct memblock_region
*r
;
5567 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5568 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5569 for_each_memblock(memory
, r
) {
5570 if (*pfn
< memblock_region_memory_end_pfn(r
))
5574 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5575 memblock_is_mirror(r
)) {
5576 *pfn
= memblock_region_memory_end_pfn(r
);
5585 * Initially all pages are reserved - free ones are freed
5586 * up by memblock_free_all() once the early boot process is
5587 * done. Non-atomic initialization, single-pass.
5589 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5590 unsigned long start_pfn
, enum memmap_context context
,
5591 struct vmem_altmap
*altmap
)
5593 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5596 if (highest_memmap_pfn
< end_pfn
- 1)
5597 highest_memmap_pfn
= end_pfn
- 1;
5599 #ifdef CONFIG_ZONE_DEVICE
5601 * Honor reservation requested by the driver for this ZONE_DEVICE
5602 * memory. We limit the total number of pages to initialize to just
5603 * those that might contain the memory mapping. We will defer the
5604 * ZONE_DEVICE page initialization until after we have released
5607 if (zone
== ZONE_DEVICE
) {
5611 if (start_pfn
== altmap
->base_pfn
)
5612 start_pfn
+= altmap
->reserve
;
5613 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5617 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5619 * There can be holes in boot-time mem_map[]s handed to this
5620 * function. They do not exist on hotplugged memory.
5622 if (context
== MEMMAP_EARLY
) {
5623 if (!early_pfn_valid(pfn
))
5625 if (!early_pfn_in_nid(pfn
, nid
))
5627 if (overlap_memmap_init(zone
, &pfn
))
5629 if (defer_init(nid
, pfn
, end_pfn
))
5633 page
= pfn_to_page(pfn
);
5634 __init_single_page(page
, pfn
, zone
, nid
);
5635 if (context
== MEMMAP_HOTPLUG
)
5636 __SetPageReserved(page
);
5639 * Mark the block movable so that blocks are reserved for
5640 * movable at startup. This will force kernel allocations
5641 * to reserve their blocks rather than leaking throughout
5642 * the address space during boot when many long-lived
5643 * kernel allocations are made.
5645 * bitmap is created for zone's valid pfn range. but memmap
5646 * can be created for invalid pages (for alignment)
5647 * check here not to call set_pageblock_migratetype() against
5650 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5651 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5655 #ifdef CONFIG_SPARSEMEM
5657 * If the zone does not span the rest of the section then
5658 * we should at least initialize those pages. Otherwise we
5659 * could blow up on a poisoned page in some paths which depend
5660 * on full sections being initialized (e.g. memory hotplug).
5662 while (end_pfn
% PAGES_PER_SECTION
) {
5663 __init_single_page(pfn_to_page(end_pfn
), end_pfn
, zone
, nid
);
5669 #ifdef CONFIG_ZONE_DEVICE
5670 void __ref
memmap_init_zone_device(struct zone
*zone
,
5671 unsigned long start_pfn
,
5673 struct dev_pagemap
*pgmap
)
5675 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5676 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5677 unsigned long zone_idx
= zone_idx(zone
);
5678 unsigned long start
= jiffies
;
5679 int nid
= pgdat
->node_id
;
5681 if (WARN_ON_ONCE(!pgmap
|| !is_dev_zone(zone
)))
5685 * The call to memmap_init_zone should have already taken care
5686 * of the pages reserved for the memmap, so we can just jump to
5687 * the end of that region and start processing the device pages.
5689 if (pgmap
->altmap_valid
) {
5690 struct vmem_altmap
*altmap
= &pgmap
->altmap
;
5692 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5693 size
= end_pfn
- start_pfn
;
5696 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5697 struct page
*page
= pfn_to_page(pfn
);
5699 __init_single_page(page
, pfn
, zone_idx
, nid
);
5702 * Mark page reserved as it will need to wait for onlining
5703 * phase for it to be fully associated with a zone.
5705 * We can use the non-atomic __set_bit operation for setting
5706 * the flag as we are still initializing the pages.
5708 __SetPageReserved(page
);
5711 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5712 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5713 * page is ever freed or placed on a driver-private list.
5715 page
->pgmap
= pgmap
;
5719 * Mark the block movable so that blocks are reserved for
5720 * movable at startup. This will force kernel allocations
5721 * to reserve their blocks rather than leaking throughout
5722 * the address space during boot when many long-lived
5723 * kernel allocations are made.
5725 * bitmap is created for zone's valid pfn range. but memmap
5726 * can be created for invalid pages (for alignment)
5727 * check here not to call set_pageblock_migratetype() against
5730 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5731 * because this is done early in sparse_add_one_section
5733 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5734 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5739 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap
->dev
),
5740 size
, jiffies_to_msecs(jiffies
- start
));
5744 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5746 unsigned int order
, t
;
5747 for_each_migratetype_order(order
, t
) {
5748 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5749 zone
->free_area
[order
].nr_free
= 0;
5753 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
5754 unsigned long zone
, unsigned long start_pfn
)
5756 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
, NULL
);
5759 static int zone_batchsize(struct zone
*zone
)
5765 * The per-cpu-pages pools are set to around 1000th of the
5768 batch
= zone_managed_pages(zone
) / 1024;
5769 /* But no more than a meg. */
5770 if (batch
* PAGE_SIZE
> 1024 * 1024)
5771 batch
= (1024 * 1024) / PAGE_SIZE
;
5772 batch
/= 4; /* We effectively *= 4 below */
5777 * Clamp the batch to a 2^n - 1 value. Having a power
5778 * of 2 value was found to be more likely to have
5779 * suboptimal cache aliasing properties in some cases.
5781 * For example if 2 tasks are alternately allocating
5782 * batches of pages, one task can end up with a lot
5783 * of pages of one half of the possible page colors
5784 * and the other with pages of the other colors.
5786 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5791 /* The deferral and batching of frees should be suppressed under NOMMU
5794 * The problem is that NOMMU needs to be able to allocate large chunks
5795 * of contiguous memory as there's no hardware page translation to
5796 * assemble apparent contiguous memory from discontiguous pages.
5798 * Queueing large contiguous runs of pages for batching, however,
5799 * causes the pages to actually be freed in smaller chunks. As there
5800 * can be a significant delay between the individual batches being
5801 * recycled, this leads to the once large chunks of space being
5802 * fragmented and becoming unavailable for high-order allocations.
5809 * pcp->high and pcp->batch values are related and dependent on one another:
5810 * ->batch must never be higher then ->high.
5811 * The following function updates them in a safe manner without read side
5814 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5815 * those fields changing asynchronously (acording the the above rule).
5817 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5818 * outside of boot time (or some other assurance that no concurrent updaters
5821 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5822 unsigned long batch
)
5824 /* start with a fail safe value for batch */
5828 /* Update high, then batch, in order */
5835 /* a companion to pageset_set_high() */
5836 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5838 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5841 static void pageset_init(struct per_cpu_pageset
*p
)
5843 struct per_cpu_pages
*pcp
;
5846 memset(p
, 0, sizeof(*p
));
5849 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5850 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5853 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5856 pageset_set_batch(p
, batch
);
5860 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5861 * to the value high for the pageset p.
5863 static void pageset_set_high(struct per_cpu_pageset
*p
,
5866 unsigned long batch
= max(1UL, high
/ 4);
5867 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5868 batch
= PAGE_SHIFT
* 8;
5870 pageset_update(&p
->pcp
, high
, batch
);
5873 static void pageset_set_high_and_batch(struct zone
*zone
,
5874 struct per_cpu_pageset
*pcp
)
5876 if (percpu_pagelist_fraction
)
5877 pageset_set_high(pcp
,
5878 (zone_managed_pages(zone
) /
5879 percpu_pagelist_fraction
));
5881 pageset_set_batch(pcp
, zone_batchsize(zone
));
5884 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5886 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5889 pageset_set_high_and_batch(zone
, pcp
);
5892 void __meminit
setup_zone_pageset(struct zone
*zone
)
5895 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5896 for_each_possible_cpu(cpu
)
5897 zone_pageset_init(zone
, cpu
);
5901 * Allocate per cpu pagesets and initialize them.
5902 * Before this call only boot pagesets were available.
5904 void __init
setup_per_cpu_pageset(void)
5906 struct pglist_data
*pgdat
;
5909 for_each_populated_zone(zone
)
5910 setup_zone_pageset(zone
);
5912 for_each_online_pgdat(pgdat
)
5913 pgdat
->per_cpu_nodestats
=
5914 alloc_percpu(struct per_cpu_nodestat
);
5917 static __meminit
void zone_pcp_init(struct zone
*zone
)
5920 * per cpu subsystem is not up at this point. The following code
5921 * relies on the ability of the linker to provide the
5922 * offset of a (static) per cpu variable into the per cpu area.
5924 zone
->pageset
= &boot_pageset
;
5926 if (populated_zone(zone
))
5927 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5928 zone
->name
, zone
->present_pages
,
5929 zone_batchsize(zone
));
5932 void __meminit
init_currently_empty_zone(struct zone
*zone
,
5933 unsigned long zone_start_pfn
,
5936 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5937 int zone_idx
= zone_idx(zone
) + 1;
5939 if (zone_idx
> pgdat
->nr_zones
)
5940 pgdat
->nr_zones
= zone_idx
;
5942 zone
->zone_start_pfn
= zone_start_pfn
;
5944 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5945 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5947 (unsigned long)zone_idx(zone
),
5948 zone_start_pfn
, (zone_start_pfn
+ size
));
5950 zone_init_free_lists(zone
);
5951 zone
->initialized
= 1;
5954 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5955 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5958 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5960 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5961 struct mminit_pfnnid_cache
*state
)
5963 unsigned long start_pfn
, end_pfn
;
5966 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5967 return state
->last_nid
;
5969 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5971 state
->last_start
= start_pfn
;
5972 state
->last_end
= end_pfn
;
5973 state
->last_nid
= nid
;
5978 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5981 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5982 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5983 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5985 * If an architecture guarantees that all ranges registered contain no holes
5986 * and may be freed, this this function may be used instead of calling
5987 * memblock_free_early_nid() manually.
5989 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5991 unsigned long start_pfn
, end_pfn
;
5994 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5995 start_pfn
= min(start_pfn
, max_low_pfn
);
5996 end_pfn
= min(end_pfn
, max_low_pfn
);
5998 if (start_pfn
< end_pfn
)
5999 memblock_free_early_nid(PFN_PHYS(start_pfn
),
6000 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
6006 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6007 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6009 * If an architecture guarantees that all ranges registered contain no holes and may
6010 * be freed, this function may be used instead of calling memory_present() manually.
6012 void __init
sparse_memory_present_with_active_regions(int nid
)
6014 unsigned long start_pfn
, end_pfn
;
6017 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
6018 memory_present(this_nid
, start_pfn
, end_pfn
);
6022 * get_pfn_range_for_nid - Return the start and end page frames for a node
6023 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6024 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6025 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6027 * It returns the start and end page frame of a node based on information
6028 * provided by memblock_set_node(). If called for a node
6029 * with no available memory, a warning is printed and the start and end
6032 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
6033 unsigned long *start_pfn
, unsigned long *end_pfn
)
6035 unsigned long this_start_pfn
, this_end_pfn
;
6041 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6042 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6043 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6046 if (*start_pfn
== -1UL)
6051 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6052 * assumption is made that zones within a node are ordered in monotonic
6053 * increasing memory addresses so that the "highest" populated zone is used
6055 static void __init
find_usable_zone_for_movable(void)
6058 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6059 if (zone_index
== ZONE_MOVABLE
)
6062 if (arch_zone_highest_possible_pfn
[zone_index
] >
6063 arch_zone_lowest_possible_pfn
[zone_index
])
6067 VM_BUG_ON(zone_index
== -1);
6068 movable_zone
= zone_index
;
6072 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6073 * because it is sized independent of architecture. Unlike the other zones,
6074 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6075 * in each node depending on the size of each node and how evenly kernelcore
6076 * is distributed. This helper function adjusts the zone ranges
6077 * provided by the architecture for a given node by using the end of the
6078 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6079 * zones within a node are in order of monotonic increases memory addresses
6081 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
6082 unsigned long zone_type
,
6083 unsigned long node_start_pfn
,
6084 unsigned long node_end_pfn
,
6085 unsigned long *zone_start_pfn
,
6086 unsigned long *zone_end_pfn
)
6088 /* Only adjust if ZONE_MOVABLE is on this node */
6089 if (zone_movable_pfn
[nid
]) {
6090 /* Size ZONE_MOVABLE */
6091 if (zone_type
== ZONE_MOVABLE
) {
6092 *zone_start_pfn
= zone_movable_pfn
[nid
];
6093 *zone_end_pfn
= min(node_end_pfn
,
6094 arch_zone_highest_possible_pfn
[movable_zone
]);
6096 /* Adjust for ZONE_MOVABLE starting within this range */
6097 } else if (!mirrored_kernelcore
&&
6098 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6099 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6100 *zone_end_pfn
= zone_movable_pfn
[nid
];
6102 /* Check if this whole range is within ZONE_MOVABLE */
6103 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6104 *zone_start_pfn
= *zone_end_pfn
;
6109 * Return the number of pages a zone spans in a node, including holes
6110 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6112 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
6113 unsigned long zone_type
,
6114 unsigned long node_start_pfn
,
6115 unsigned long node_end_pfn
,
6116 unsigned long *zone_start_pfn
,
6117 unsigned long *zone_end_pfn
,
6118 unsigned long *ignored
)
6120 /* When hotadd a new node from cpu_up(), the node should be empty */
6121 if (!node_start_pfn
&& !node_end_pfn
)
6124 /* Get the start and end of the zone */
6125 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
6126 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
6127 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6128 node_start_pfn
, node_end_pfn
,
6129 zone_start_pfn
, zone_end_pfn
);
6131 /* Check that this node has pages within the zone's required range */
6132 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6135 /* Move the zone boundaries inside the node if necessary */
6136 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6137 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6139 /* Return the spanned pages */
6140 return *zone_end_pfn
- *zone_start_pfn
;
6144 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6145 * then all holes in the requested range will be accounted for.
6147 unsigned long __meminit
__absent_pages_in_range(int nid
,
6148 unsigned long range_start_pfn
,
6149 unsigned long range_end_pfn
)
6151 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6152 unsigned long start_pfn
, end_pfn
;
6155 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6156 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6157 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6158 nr_absent
-= end_pfn
- start_pfn
;
6164 * absent_pages_in_range - Return number of page frames in holes within a range
6165 * @start_pfn: The start PFN to start searching for holes
6166 * @end_pfn: The end PFN to stop searching for holes
6168 * It returns the number of pages frames in memory holes within a range.
6170 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6171 unsigned long end_pfn
)
6173 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6176 /* Return the number of page frames in holes in a zone on a node */
6177 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
6178 unsigned long zone_type
,
6179 unsigned long node_start_pfn
,
6180 unsigned long node_end_pfn
,
6181 unsigned long *ignored
)
6183 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6184 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6185 unsigned long zone_start_pfn
, zone_end_pfn
;
6186 unsigned long nr_absent
;
6188 /* When hotadd a new node from cpu_up(), the node should be empty */
6189 if (!node_start_pfn
&& !node_end_pfn
)
6192 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6193 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6195 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6196 node_start_pfn
, node_end_pfn
,
6197 &zone_start_pfn
, &zone_end_pfn
);
6198 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6201 * ZONE_MOVABLE handling.
6202 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6205 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6206 unsigned long start_pfn
, end_pfn
;
6207 struct memblock_region
*r
;
6209 for_each_memblock(memory
, r
) {
6210 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6211 zone_start_pfn
, zone_end_pfn
);
6212 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6213 zone_start_pfn
, zone_end_pfn
);
6215 if (zone_type
== ZONE_MOVABLE
&&
6216 memblock_is_mirror(r
))
6217 nr_absent
+= end_pfn
- start_pfn
;
6219 if (zone_type
== ZONE_NORMAL
&&
6220 !memblock_is_mirror(r
))
6221 nr_absent
+= end_pfn
- start_pfn
;
6228 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6229 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
6230 unsigned long zone_type
,
6231 unsigned long node_start_pfn
,
6232 unsigned long node_end_pfn
,
6233 unsigned long *zone_start_pfn
,
6234 unsigned long *zone_end_pfn
,
6235 unsigned long *zones_size
)
6239 *zone_start_pfn
= node_start_pfn
;
6240 for (zone
= 0; zone
< zone_type
; zone
++)
6241 *zone_start_pfn
+= zones_size
[zone
];
6243 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
6245 return zones_size
[zone_type
];
6248 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
6249 unsigned long zone_type
,
6250 unsigned long node_start_pfn
,
6251 unsigned long node_end_pfn
,
6252 unsigned long *zholes_size
)
6257 return zholes_size
[zone_type
];
6260 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6262 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
6263 unsigned long node_start_pfn
,
6264 unsigned long node_end_pfn
,
6265 unsigned long *zones_size
,
6266 unsigned long *zholes_size
)
6268 unsigned long realtotalpages
= 0, totalpages
= 0;
6271 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6272 struct zone
*zone
= pgdat
->node_zones
+ i
;
6273 unsigned long zone_start_pfn
, zone_end_pfn
;
6274 unsigned long size
, real_size
;
6276 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6282 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
6283 node_start_pfn
, node_end_pfn
,
6286 zone
->zone_start_pfn
= zone_start_pfn
;
6288 zone
->zone_start_pfn
= 0;
6289 zone
->spanned_pages
= size
;
6290 zone
->present_pages
= real_size
;
6293 realtotalpages
+= real_size
;
6296 pgdat
->node_spanned_pages
= totalpages
;
6297 pgdat
->node_present_pages
= realtotalpages
;
6298 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6302 #ifndef CONFIG_SPARSEMEM
6304 * Calculate the size of the zone->blockflags rounded to an unsigned long
6305 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6306 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6307 * round what is now in bits to nearest long in bits, then return it in
6310 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6312 unsigned long usemapsize
;
6314 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6315 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6316 usemapsize
= usemapsize
>> pageblock_order
;
6317 usemapsize
*= NR_PAGEBLOCK_BITS
;
6318 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6320 return usemapsize
/ 8;
6323 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6325 unsigned long zone_start_pfn
,
6326 unsigned long zonesize
)
6328 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6329 zone
->pageblock_flags
= NULL
;
6331 zone
->pageblock_flags
=
6332 memblock_alloc_node_nopanic(usemapsize
,
6336 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6337 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6338 #endif /* CONFIG_SPARSEMEM */
6340 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6342 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6343 void __init
set_pageblock_order(void)
6347 /* Check that pageblock_nr_pages has not already been setup */
6348 if (pageblock_order
)
6351 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6352 order
= HUGETLB_PAGE_ORDER
;
6354 order
= MAX_ORDER
- 1;
6357 * Assume the largest contiguous order of interest is a huge page.
6358 * This value may be variable depending on boot parameters on IA64 and
6361 pageblock_order
= order
;
6363 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6366 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6367 * is unused as pageblock_order is set at compile-time. See
6368 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6371 void __init
set_pageblock_order(void)
6375 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6377 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6378 unsigned long present_pages
)
6380 unsigned long pages
= spanned_pages
;
6383 * Provide a more accurate estimation if there are holes within
6384 * the zone and SPARSEMEM is in use. If there are holes within the
6385 * zone, each populated memory region may cost us one or two extra
6386 * memmap pages due to alignment because memmap pages for each
6387 * populated regions may not be naturally aligned on page boundary.
6388 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6390 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6391 IS_ENABLED(CONFIG_SPARSEMEM
))
6392 pages
= present_pages
;
6394 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6397 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6398 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6400 spin_lock_init(&pgdat
->split_queue_lock
);
6401 INIT_LIST_HEAD(&pgdat
->split_queue
);
6402 pgdat
->split_queue_len
= 0;
6405 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6408 #ifdef CONFIG_COMPACTION
6409 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6411 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6414 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6417 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6419 pgdat_resize_init(pgdat
);
6421 pgdat_init_split_queue(pgdat
);
6422 pgdat_init_kcompactd(pgdat
);
6424 init_waitqueue_head(&pgdat
->kswapd_wait
);
6425 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6427 pgdat_page_ext_init(pgdat
);
6428 spin_lock_init(&pgdat
->lru_lock
);
6429 lruvec_init(node_lruvec(pgdat
));
6432 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6433 unsigned long remaining_pages
)
6435 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6436 zone_set_nid(zone
, nid
);
6437 zone
->name
= zone_names
[idx
];
6438 zone
->zone_pgdat
= NODE_DATA(nid
);
6439 spin_lock_init(&zone
->lock
);
6440 zone_seqlock_init(zone
);
6441 zone_pcp_init(zone
);
6445 * Set up the zone data structures
6446 * - init pgdat internals
6447 * - init all zones belonging to this node
6449 * NOTE: this function is only called during memory hotplug
6451 #ifdef CONFIG_MEMORY_HOTPLUG
6452 void __ref
free_area_init_core_hotplug(int nid
)
6455 pg_data_t
*pgdat
= NODE_DATA(nid
);
6457 pgdat_init_internals(pgdat
);
6458 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6459 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6464 * Set up the zone data structures:
6465 * - mark all pages reserved
6466 * - mark all memory queues empty
6467 * - clear the memory bitmaps
6469 * NOTE: pgdat should get zeroed by caller.
6470 * NOTE: this function is only called during early init.
6472 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6475 int nid
= pgdat
->node_id
;
6477 pgdat_init_internals(pgdat
);
6478 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6480 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6481 struct zone
*zone
= pgdat
->node_zones
+ j
;
6482 unsigned long size
, freesize
, memmap_pages
;
6483 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6485 size
= zone
->spanned_pages
;
6486 freesize
= zone
->present_pages
;
6489 * Adjust freesize so that it accounts for how much memory
6490 * is used by this zone for memmap. This affects the watermark
6491 * and per-cpu initialisations
6493 memmap_pages
= calc_memmap_size(size
, freesize
);
6494 if (!is_highmem_idx(j
)) {
6495 if (freesize
>= memmap_pages
) {
6496 freesize
-= memmap_pages
;
6499 " %s zone: %lu pages used for memmap\n",
6500 zone_names
[j
], memmap_pages
);
6502 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6503 zone_names
[j
], memmap_pages
, freesize
);
6506 /* Account for reserved pages */
6507 if (j
== 0 && freesize
> dma_reserve
) {
6508 freesize
-= dma_reserve
;
6509 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6510 zone_names
[0], dma_reserve
);
6513 if (!is_highmem_idx(j
))
6514 nr_kernel_pages
+= freesize
;
6515 /* Charge for highmem memmap if there are enough kernel pages */
6516 else if (nr_kernel_pages
> memmap_pages
* 2)
6517 nr_kernel_pages
-= memmap_pages
;
6518 nr_all_pages
+= freesize
;
6521 * Set an approximate value for lowmem here, it will be adjusted
6522 * when the bootmem allocator frees pages into the buddy system.
6523 * And all highmem pages will be managed by the buddy system.
6525 zone_init_internals(zone
, j
, nid
, freesize
);
6530 set_pageblock_order();
6531 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6532 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6533 memmap_init(size
, nid
, j
, zone_start_pfn
);
6537 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6538 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6540 unsigned long __maybe_unused start
= 0;
6541 unsigned long __maybe_unused offset
= 0;
6543 /* Skip empty nodes */
6544 if (!pgdat
->node_spanned_pages
)
6547 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6548 offset
= pgdat
->node_start_pfn
- start
;
6549 /* ia64 gets its own node_mem_map, before this, without bootmem */
6550 if (!pgdat
->node_mem_map
) {
6551 unsigned long size
, end
;
6555 * The zone's endpoints aren't required to be MAX_ORDER
6556 * aligned but the node_mem_map endpoints must be in order
6557 * for the buddy allocator to function correctly.
6559 end
= pgdat_end_pfn(pgdat
);
6560 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6561 size
= (end
- start
) * sizeof(struct page
);
6562 map
= memblock_alloc_node_nopanic(size
, pgdat
->node_id
);
6563 pgdat
->node_mem_map
= map
+ offset
;
6565 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6566 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6567 (unsigned long)pgdat
->node_mem_map
);
6568 #ifndef CONFIG_NEED_MULTIPLE_NODES
6570 * With no DISCONTIG, the global mem_map is just set as node 0's
6572 if (pgdat
== NODE_DATA(0)) {
6573 mem_map
= NODE_DATA(0)->node_mem_map
;
6574 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6575 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6577 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6582 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6583 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6585 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6586 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6589 * We start only with one section of pages, more pages are added as
6590 * needed until the rest of deferred pages are initialized.
6592 pgdat
->static_init_pgcnt
= min_t(unsigned long, PAGES_PER_SECTION
,
6593 pgdat
->node_spanned_pages
);
6594 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6597 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6600 void __init
free_area_init_node(int nid
, unsigned long *zones_size
,
6601 unsigned long node_start_pfn
,
6602 unsigned long *zholes_size
)
6604 pg_data_t
*pgdat
= NODE_DATA(nid
);
6605 unsigned long start_pfn
= 0;
6606 unsigned long end_pfn
= 0;
6608 /* pg_data_t should be reset to zero when it's allocated */
6609 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6611 pgdat
->node_id
= nid
;
6612 pgdat
->node_start_pfn
= node_start_pfn
;
6613 pgdat
->per_cpu_nodestats
= NULL
;
6614 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6615 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6616 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6617 (u64
)start_pfn
<< PAGE_SHIFT
,
6618 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6620 start_pfn
= node_start_pfn
;
6622 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6623 zones_size
, zholes_size
);
6625 alloc_node_mem_map(pgdat
);
6626 pgdat_set_deferred_range(pgdat
);
6628 free_area_init_core(pgdat
);
6631 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6633 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6636 static u64
zero_pfn_range(unsigned long spfn
, unsigned long epfn
)
6641 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6642 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6643 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6644 + pageblock_nr_pages
- 1;
6647 mm_zero_struct_page(pfn_to_page(pfn
));
6655 * Only struct pages that are backed by physical memory are zeroed and
6656 * initialized by going through __init_single_page(). But, there are some
6657 * struct pages which are reserved in memblock allocator and their fields
6658 * may be accessed (for example page_to_pfn() on some configuration accesses
6659 * flags). We must explicitly zero those struct pages.
6661 * This function also addresses a similar issue where struct pages are left
6662 * uninitialized because the physical address range is not covered by
6663 * memblock.memory or memblock.reserved. That could happen when memblock
6664 * layout is manually configured via memmap=.
6666 void __init
zero_resv_unavail(void)
6668 phys_addr_t start
, end
;
6670 phys_addr_t next
= 0;
6673 * Loop through unavailable ranges not covered by memblock.memory.
6676 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6677 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6679 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), PFN_UP(start
));
6682 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), max_pfn
);
6685 * Struct pages that do not have backing memory. This could be because
6686 * firmware is using some of this memory, or for some other reasons.
6689 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
6691 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6693 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6695 #if MAX_NUMNODES > 1
6697 * Figure out the number of possible node ids.
6699 void __init
setup_nr_node_ids(void)
6701 unsigned int highest
;
6703 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6704 nr_node_ids
= highest
+ 1;
6709 * node_map_pfn_alignment - determine the maximum internode alignment
6711 * This function should be called after node map is populated and sorted.
6712 * It calculates the maximum power of two alignment which can distinguish
6715 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6716 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6717 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6718 * shifted, 1GiB is enough and this function will indicate so.
6720 * This is used to test whether pfn -> nid mapping of the chosen memory
6721 * model has fine enough granularity to avoid incorrect mapping for the
6722 * populated node map.
6724 * Returns the determined alignment in pfn's. 0 if there is no alignment
6725 * requirement (single node).
6727 unsigned long __init
node_map_pfn_alignment(void)
6729 unsigned long accl_mask
= 0, last_end
= 0;
6730 unsigned long start
, end
, mask
;
6734 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6735 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6742 * Start with a mask granular enough to pin-point to the
6743 * start pfn and tick off bits one-by-one until it becomes
6744 * too coarse to separate the current node from the last.
6746 mask
= ~((1 << __ffs(start
)) - 1);
6747 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6750 /* accumulate all internode masks */
6754 /* convert mask to number of pages */
6755 return ~accl_mask
+ 1;
6758 /* Find the lowest pfn for a node */
6759 static unsigned long __init
find_min_pfn_for_node(int nid
)
6761 unsigned long min_pfn
= ULONG_MAX
;
6762 unsigned long start_pfn
;
6765 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6766 min_pfn
= min(min_pfn
, start_pfn
);
6768 if (min_pfn
== ULONG_MAX
) {
6769 pr_warn("Could not find start_pfn for node %d\n", nid
);
6777 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6779 * It returns the minimum PFN based on information provided via
6780 * memblock_set_node().
6782 unsigned long __init
find_min_pfn_with_active_regions(void)
6784 return find_min_pfn_for_node(MAX_NUMNODES
);
6788 * early_calculate_totalpages()
6789 * Sum pages in active regions for movable zone.
6790 * Populate N_MEMORY for calculating usable_nodes.
6792 static unsigned long __init
early_calculate_totalpages(void)
6794 unsigned long totalpages
= 0;
6795 unsigned long start_pfn
, end_pfn
;
6798 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6799 unsigned long pages
= end_pfn
- start_pfn
;
6801 totalpages
+= pages
;
6803 node_set_state(nid
, N_MEMORY
);
6809 * Find the PFN the Movable zone begins in each node. Kernel memory
6810 * is spread evenly between nodes as long as the nodes have enough
6811 * memory. When they don't, some nodes will have more kernelcore than
6814 static void __init
find_zone_movable_pfns_for_nodes(void)
6817 unsigned long usable_startpfn
;
6818 unsigned long kernelcore_node
, kernelcore_remaining
;
6819 /* save the state before borrow the nodemask */
6820 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6821 unsigned long totalpages
= early_calculate_totalpages();
6822 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6823 struct memblock_region
*r
;
6825 /* Need to find movable_zone earlier when movable_node is specified. */
6826 find_usable_zone_for_movable();
6829 * If movable_node is specified, ignore kernelcore and movablecore
6832 if (movable_node_is_enabled()) {
6833 for_each_memblock(memory
, r
) {
6834 if (!memblock_is_hotpluggable(r
))
6839 usable_startpfn
= PFN_DOWN(r
->base
);
6840 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6841 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6849 * If kernelcore=mirror is specified, ignore movablecore option
6851 if (mirrored_kernelcore
) {
6852 bool mem_below_4gb_not_mirrored
= false;
6854 for_each_memblock(memory
, r
) {
6855 if (memblock_is_mirror(r
))
6860 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6862 if (usable_startpfn
< 0x100000) {
6863 mem_below_4gb_not_mirrored
= true;
6867 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6868 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6872 if (mem_below_4gb_not_mirrored
)
6873 pr_warn("This configuration results in unmirrored kernel memory.");
6879 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6880 * amount of necessary memory.
6882 if (required_kernelcore_percent
)
6883 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
6885 if (required_movablecore_percent
)
6886 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
6890 * If movablecore= was specified, calculate what size of
6891 * kernelcore that corresponds so that memory usable for
6892 * any allocation type is evenly spread. If both kernelcore
6893 * and movablecore are specified, then the value of kernelcore
6894 * will be used for required_kernelcore if it's greater than
6895 * what movablecore would have allowed.
6897 if (required_movablecore
) {
6898 unsigned long corepages
;
6901 * Round-up so that ZONE_MOVABLE is at least as large as what
6902 * was requested by the user
6904 required_movablecore
=
6905 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6906 required_movablecore
= min(totalpages
, required_movablecore
);
6907 corepages
= totalpages
- required_movablecore
;
6909 required_kernelcore
= max(required_kernelcore
, corepages
);
6913 * If kernelcore was not specified or kernelcore size is larger
6914 * than totalpages, there is no ZONE_MOVABLE.
6916 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6919 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6920 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6923 /* Spread kernelcore memory as evenly as possible throughout nodes */
6924 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6925 for_each_node_state(nid
, N_MEMORY
) {
6926 unsigned long start_pfn
, end_pfn
;
6929 * Recalculate kernelcore_node if the division per node
6930 * now exceeds what is necessary to satisfy the requested
6931 * amount of memory for the kernel
6933 if (required_kernelcore
< kernelcore_node
)
6934 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6937 * As the map is walked, we track how much memory is usable
6938 * by the kernel using kernelcore_remaining. When it is
6939 * 0, the rest of the node is usable by ZONE_MOVABLE
6941 kernelcore_remaining
= kernelcore_node
;
6943 /* Go through each range of PFNs within this node */
6944 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6945 unsigned long size_pages
;
6947 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6948 if (start_pfn
>= end_pfn
)
6951 /* Account for what is only usable for kernelcore */
6952 if (start_pfn
< usable_startpfn
) {
6953 unsigned long kernel_pages
;
6954 kernel_pages
= min(end_pfn
, usable_startpfn
)
6957 kernelcore_remaining
-= min(kernel_pages
,
6958 kernelcore_remaining
);
6959 required_kernelcore
-= min(kernel_pages
,
6960 required_kernelcore
);
6962 /* Continue if range is now fully accounted */
6963 if (end_pfn
<= usable_startpfn
) {
6966 * Push zone_movable_pfn to the end so
6967 * that if we have to rebalance
6968 * kernelcore across nodes, we will
6969 * not double account here
6971 zone_movable_pfn
[nid
] = end_pfn
;
6974 start_pfn
= usable_startpfn
;
6978 * The usable PFN range for ZONE_MOVABLE is from
6979 * start_pfn->end_pfn. Calculate size_pages as the
6980 * number of pages used as kernelcore
6982 size_pages
= end_pfn
- start_pfn
;
6983 if (size_pages
> kernelcore_remaining
)
6984 size_pages
= kernelcore_remaining
;
6985 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6988 * Some kernelcore has been met, update counts and
6989 * break if the kernelcore for this node has been
6992 required_kernelcore
-= min(required_kernelcore
,
6994 kernelcore_remaining
-= size_pages
;
6995 if (!kernelcore_remaining
)
7001 * If there is still required_kernelcore, we do another pass with one
7002 * less node in the count. This will push zone_movable_pfn[nid] further
7003 * along on the nodes that still have memory until kernelcore is
7007 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7011 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7012 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7013 zone_movable_pfn
[nid
] =
7014 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7017 /* restore the node_state */
7018 node_states
[N_MEMORY
] = saved_node_state
;
7021 /* Any regular or high memory on that node ? */
7022 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7024 enum zone_type zone_type
;
7026 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7027 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7028 if (populated_zone(zone
)) {
7029 if (IS_ENABLED(CONFIG_HIGHMEM
))
7030 node_set_state(nid
, N_HIGH_MEMORY
);
7031 if (zone_type
<= ZONE_NORMAL
)
7032 node_set_state(nid
, N_NORMAL_MEMORY
);
7039 * free_area_init_nodes - Initialise all pg_data_t and zone data
7040 * @max_zone_pfn: an array of max PFNs for each zone
7042 * This will call free_area_init_node() for each active node in the system.
7043 * Using the page ranges provided by memblock_set_node(), the size of each
7044 * zone in each node and their holes is calculated. If the maximum PFN
7045 * between two adjacent zones match, it is assumed that the zone is empty.
7046 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7047 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7048 * starts where the previous one ended. For example, ZONE_DMA32 starts
7049 * at arch_max_dma_pfn.
7051 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
7053 unsigned long start_pfn
, end_pfn
;
7056 /* Record where the zone boundaries are */
7057 memset(arch_zone_lowest_possible_pfn
, 0,
7058 sizeof(arch_zone_lowest_possible_pfn
));
7059 memset(arch_zone_highest_possible_pfn
, 0,
7060 sizeof(arch_zone_highest_possible_pfn
));
7062 start_pfn
= find_min_pfn_with_active_regions();
7064 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7065 if (i
== ZONE_MOVABLE
)
7068 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
7069 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
7070 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
7072 start_pfn
= end_pfn
;
7075 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7076 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7077 find_zone_movable_pfns_for_nodes();
7079 /* Print out the zone ranges */
7080 pr_info("Zone ranges:\n");
7081 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7082 if (i
== ZONE_MOVABLE
)
7084 pr_info(" %-8s ", zone_names
[i
]);
7085 if (arch_zone_lowest_possible_pfn
[i
] ==
7086 arch_zone_highest_possible_pfn
[i
])
7089 pr_cont("[mem %#018Lx-%#018Lx]\n",
7090 (u64
)arch_zone_lowest_possible_pfn
[i
]
7092 ((u64
)arch_zone_highest_possible_pfn
[i
]
7093 << PAGE_SHIFT
) - 1);
7096 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7097 pr_info("Movable zone start for each node\n");
7098 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7099 if (zone_movable_pfn
[i
])
7100 pr_info(" Node %d: %#018Lx\n", i
,
7101 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7104 /* Print out the early node map */
7105 pr_info("Early memory node ranges\n");
7106 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
7107 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7108 (u64
)start_pfn
<< PAGE_SHIFT
,
7109 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7111 /* Initialise every node */
7112 mminit_verify_pageflags_layout();
7113 setup_nr_node_ids();
7114 zero_resv_unavail();
7115 for_each_online_node(nid
) {
7116 pg_data_t
*pgdat
= NODE_DATA(nid
);
7117 free_area_init_node(nid
, NULL
,
7118 find_min_pfn_for_node(nid
), NULL
);
7120 /* Any memory on that node */
7121 if (pgdat
->node_present_pages
)
7122 node_set_state(nid
, N_MEMORY
);
7123 check_for_memory(pgdat
, nid
);
7127 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7128 unsigned long *percent
)
7130 unsigned long long coremem
;
7136 /* Value may be a percentage of total memory, otherwise bytes */
7137 coremem
= simple_strtoull(p
, &endptr
, 0);
7138 if (*endptr
== '%') {
7139 /* Paranoid check for percent values greater than 100 */
7140 WARN_ON(coremem
> 100);
7144 coremem
= memparse(p
, &p
);
7145 /* Paranoid check that UL is enough for the coremem value */
7146 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7148 *core
= coremem
>> PAGE_SHIFT
;
7155 * kernelcore=size sets the amount of memory for use for allocations that
7156 * cannot be reclaimed or migrated.
7158 static int __init
cmdline_parse_kernelcore(char *p
)
7160 /* parse kernelcore=mirror */
7161 if (parse_option_str(p
, "mirror")) {
7162 mirrored_kernelcore
= true;
7166 return cmdline_parse_core(p
, &required_kernelcore
,
7167 &required_kernelcore_percent
);
7171 * movablecore=size sets the amount of memory for use for allocations that
7172 * can be reclaimed or migrated.
7174 static int __init
cmdline_parse_movablecore(char *p
)
7176 return cmdline_parse_core(p
, &required_movablecore
,
7177 &required_movablecore_percent
);
7180 early_param("kernelcore", cmdline_parse_kernelcore
);
7181 early_param("movablecore", cmdline_parse_movablecore
);
7183 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7185 void adjust_managed_page_count(struct page
*page
, long count
)
7187 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7188 totalram_pages_add(count
);
7189 #ifdef CONFIG_HIGHMEM
7190 if (PageHighMem(page
))
7191 totalhigh_pages_add(count
);
7194 EXPORT_SYMBOL(adjust_managed_page_count
);
7196 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7199 unsigned long pages
= 0;
7201 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7202 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7203 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7204 struct page
*page
= virt_to_page(pos
);
7205 void *direct_map_addr
;
7208 * 'direct_map_addr' might be different from 'pos'
7209 * because some architectures' virt_to_page()
7210 * work with aliases. Getting the direct map
7211 * address ensures that we get a _writeable_
7212 * alias for the memset().
7214 direct_map_addr
= page_address(page
);
7215 if ((unsigned int)poison
<= 0xFF)
7216 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7218 free_reserved_page(page
);
7222 pr_info("Freeing %s memory: %ldK\n",
7223 s
, pages
<< (PAGE_SHIFT
- 10));
7227 EXPORT_SYMBOL(free_reserved_area
);
7229 #ifdef CONFIG_HIGHMEM
7230 void free_highmem_page(struct page
*page
)
7232 __free_reserved_page(page
);
7233 totalram_pages_inc();
7234 atomic_long_inc(&page_zone(page
)->managed_pages
);
7235 totalhigh_pages_inc();
7240 void __init
mem_init_print_info(const char *str
)
7242 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7243 unsigned long init_code_size
, init_data_size
;
7245 physpages
= get_num_physpages();
7246 codesize
= _etext
- _stext
;
7247 datasize
= _edata
- _sdata
;
7248 rosize
= __end_rodata
- __start_rodata
;
7249 bss_size
= __bss_stop
- __bss_start
;
7250 init_data_size
= __init_end
- __init_begin
;
7251 init_code_size
= _einittext
- _sinittext
;
7254 * Detect special cases and adjust section sizes accordingly:
7255 * 1) .init.* may be embedded into .data sections
7256 * 2) .init.text.* may be out of [__init_begin, __init_end],
7257 * please refer to arch/tile/kernel/vmlinux.lds.S.
7258 * 3) .rodata.* may be embedded into .text or .data sections.
7260 #define adj_init_size(start, end, size, pos, adj) \
7262 if (start <= pos && pos < end && size > adj) \
7266 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7267 _sinittext
, init_code_size
);
7268 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7269 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7270 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7271 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7273 #undef adj_init_size
7275 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7276 #ifdef CONFIG_HIGHMEM
7280 nr_free_pages() << (PAGE_SHIFT
- 10),
7281 physpages
<< (PAGE_SHIFT
- 10),
7282 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7283 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7284 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7285 totalcma_pages
<< (PAGE_SHIFT
- 10),
7286 #ifdef CONFIG_HIGHMEM
7287 totalhigh_pages() << (PAGE_SHIFT
- 10),
7289 str
? ", " : "", str
? str
: "");
7293 * set_dma_reserve - set the specified number of pages reserved in the first zone
7294 * @new_dma_reserve: The number of pages to mark reserved
7296 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7297 * In the DMA zone, a significant percentage may be consumed by kernel image
7298 * and other unfreeable allocations which can skew the watermarks badly. This
7299 * function may optionally be used to account for unfreeable pages in the
7300 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7301 * smaller per-cpu batchsize.
7303 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7305 dma_reserve
= new_dma_reserve
;
7308 void __init
free_area_init(unsigned long *zones_size
)
7310 zero_resv_unavail();
7311 free_area_init_node(0, zones_size
,
7312 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
7315 static int page_alloc_cpu_dead(unsigned int cpu
)
7318 lru_add_drain_cpu(cpu
);
7322 * Spill the event counters of the dead processor
7323 * into the current processors event counters.
7324 * This artificially elevates the count of the current
7327 vm_events_fold_cpu(cpu
);
7330 * Zero the differential counters of the dead processor
7331 * so that the vm statistics are consistent.
7333 * This is only okay since the processor is dead and cannot
7334 * race with what we are doing.
7336 cpu_vm_stats_fold(cpu
);
7340 void __init
page_alloc_init(void)
7344 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7345 "mm/page_alloc:dead", NULL
,
7346 page_alloc_cpu_dead
);
7351 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7352 * or min_free_kbytes changes.
7354 static void calculate_totalreserve_pages(void)
7356 struct pglist_data
*pgdat
;
7357 unsigned long reserve_pages
= 0;
7358 enum zone_type i
, j
;
7360 for_each_online_pgdat(pgdat
) {
7362 pgdat
->totalreserve_pages
= 0;
7364 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7365 struct zone
*zone
= pgdat
->node_zones
+ i
;
7367 unsigned long managed_pages
= zone_managed_pages(zone
);
7369 /* Find valid and maximum lowmem_reserve in the zone */
7370 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7371 if (zone
->lowmem_reserve
[j
] > max
)
7372 max
= zone
->lowmem_reserve
[j
];
7375 /* we treat the high watermark as reserved pages. */
7376 max
+= high_wmark_pages(zone
);
7378 if (max
> managed_pages
)
7379 max
= managed_pages
;
7381 pgdat
->totalreserve_pages
+= max
;
7383 reserve_pages
+= max
;
7386 totalreserve_pages
= reserve_pages
;
7390 * setup_per_zone_lowmem_reserve - called whenever
7391 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7392 * has a correct pages reserved value, so an adequate number of
7393 * pages are left in the zone after a successful __alloc_pages().
7395 static void setup_per_zone_lowmem_reserve(void)
7397 struct pglist_data
*pgdat
;
7398 enum zone_type j
, idx
;
7400 for_each_online_pgdat(pgdat
) {
7401 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7402 struct zone
*zone
= pgdat
->node_zones
+ j
;
7403 unsigned long managed_pages
= zone_managed_pages(zone
);
7405 zone
->lowmem_reserve
[j
] = 0;
7409 struct zone
*lower_zone
;
7412 lower_zone
= pgdat
->node_zones
+ idx
;
7414 if (sysctl_lowmem_reserve_ratio
[idx
] < 1) {
7415 sysctl_lowmem_reserve_ratio
[idx
] = 0;
7416 lower_zone
->lowmem_reserve
[j
] = 0;
7418 lower_zone
->lowmem_reserve
[j
] =
7419 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7421 managed_pages
+= zone_managed_pages(lower_zone
);
7426 /* update totalreserve_pages */
7427 calculate_totalreserve_pages();
7430 static void __setup_per_zone_wmarks(void)
7432 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7433 unsigned long lowmem_pages
= 0;
7435 unsigned long flags
;
7437 /* Calculate total number of !ZONE_HIGHMEM pages */
7438 for_each_zone(zone
) {
7439 if (!is_highmem(zone
))
7440 lowmem_pages
+= zone_managed_pages(zone
);
7443 for_each_zone(zone
) {
7446 spin_lock_irqsave(&zone
->lock
, flags
);
7447 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7448 do_div(tmp
, lowmem_pages
);
7449 if (is_highmem(zone
)) {
7451 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7452 * need highmem pages, so cap pages_min to a small
7455 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7456 * deltas control asynch page reclaim, and so should
7457 * not be capped for highmem.
7459 unsigned long min_pages
;
7461 min_pages
= zone_managed_pages(zone
) / 1024;
7462 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7463 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7466 * If it's a lowmem zone, reserve a number of pages
7467 * proportionate to the zone's size.
7469 zone
->_watermark
[WMARK_MIN
] = tmp
;
7473 * Set the kswapd watermarks distance according to the
7474 * scale factor in proportion to available memory, but
7475 * ensure a minimum size on small systems.
7477 tmp
= max_t(u64
, tmp
>> 2,
7478 mult_frac(zone_managed_pages(zone
),
7479 watermark_scale_factor
, 10000));
7481 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7482 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7483 zone
->watermark_boost
= 0;
7485 spin_unlock_irqrestore(&zone
->lock
, flags
);
7488 /* update totalreserve_pages */
7489 calculate_totalreserve_pages();
7493 * setup_per_zone_wmarks - called when min_free_kbytes changes
7494 * or when memory is hot-{added|removed}
7496 * Ensures that the watermark[min,low,high] values for each zone are set
7497 * correctly with respect to min_free_kbytes.
7499 void setup_per_zone_wmarks(void)
7501 static DEFINE_SPINLOCK(lock
);
7504 __setup_per_zone_wmarks();
7509 * Initialise min_free_kbytes.
7511 * For small machines we want it small (128k min). For large machines
7512 * we want it large (64MB max). But it is not linear, because network
7513 * bandwidth does not increase linearly with machine size. We use
7515 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7516 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7532 int __meminit
init_per_zone_wmark_min(void)
7534 unsigned long lowmem_kbytes
;
7535 int new_min_free_kbytes
;
7537 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7538 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7540 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7541 min_free_kbytes
= new_min_free_kbytes
;
7542 if (min_free_kbytes
< 128)
7543 min_free_kbytes
= 128;
7544 if (min_free_kbytes
> 65536)
7545 min_free_kbytes
= 65536;
7547 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7548 new_min_free_kbytes
, user_min_free_kbytes
);
7550 setup_per_zone_wmarks();
7551 refresh_zone_stat_thresholds();
7552 setup_per_zone_lowmem_reserve();
7555 setup_min_unmapped_ratio();
7556 setup_min_slab_ratio();
7561 core_initcall(init_per_zone_wmark_min
)
7564 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7565 * that we can call two helper functions whenever min_free_kbytes
7568 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7569 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7573 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7578 user_min_free_kbytes
= min_free_kbytes
;
7579 setup_per_zone_wmarks();
7584 int watermark_boost_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7585 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7589 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7596 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7597 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7601 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7606 setup_per_zone_wmarks();
7612 static void setup_min_unmapped_ratio(void)
7617 for_each_online_pgdat(pgdat
)
7618 pgdat
->min_unmapped_pages
= 0;
7621 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
7622 sysctl_min_unmapped_ratio
) / 100;
7626 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7627 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7631 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7635 setup_min_unmapped_ratio();
7640 static void setup_min_slab_ratio(void)
7645 for_each_online_pgdat(pgdat
)
7646 pgdat
->min_slab_pages
= 0;
7649 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
7650 sysctl_min_slab_ratio
) / 100;
7653 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7654 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7658 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7662 setup_min_slab_ratio();
7669 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7670 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7671 * whenever sysctl_lowmem_reserve_ratio changes.
7673 * The reserve ratio obviously has absolutely no relation with the
7674 * minimum watermarks. The lowmem reserve ratio can only make sense
7675 * if in function of the boot time zone sizes.
7677 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7678 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7680 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7681 setup_per_zone_lowmem_reserve();
7686 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7687 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7688 * pagelist can have before it gets flushed back to buddy allocator.
7690 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7691 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7694 int old_percpu_pagelist_fraction
;
7697 mutex_lock(&pcp_batch_high_lock
);
7698 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7700 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7701 if (!write
|| ret
< 0)
7704 /* Sanity checking to avoid pcp imbalance */
7705 if (percpu_pagelist_fraction
&&
7706 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7707 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7713 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7716 for_each_populated_zone(zone
) {
7719 for_each_possible_cpu(cpu
)
7720 pageset_set_high_and_batch(zone
,
7721 per_cpu_ptr(zone
->pageset
, cpu
));
7724 mutex_unlock(&pcp_batch_high_lock
);
7729 int hashdist
= HASHDIST_DEFAULT
;
7731 static int __init
set_hashdist(char *str
)
7735 hashdist
= simple_strtoul(str
, &str
, 0);
7738 __setup("hashdist=", set_hashdist
);
7741 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7743 * Returns the number of pages that arch has reserved but
7744 * is not known to alloc_large_system_hash().
7746 static unsigned long __init
arch_reserved_kernel_pages(void)
7753 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7754 * machines. As memory size is increased the scale is also increased but at
7755 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7756 * quadruples the scale is increased by one, which means the size of hash table
7757 * only doubles, instead of quadrupling as well.
7758 * Because 32-bit systems cannot have large physical memory, where this scaling
7759 * makes sense, it is disabled on such platforms.
7761 #if __BITS_PER_LONG > 32
7762 #define ADAPT_SCALE_BASE (64ul << 30)
7763 #define ADAPT_SCALE_SHIFT 2
7764 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7768 * allocate a large system hash table from bootmem
7769 * - it is assumed that the hash table must contain an exact power-of-2
7770 * quantity of entries
7771 * - limit is the number of hash buckets, not the total allocation size
7773 void *__init
alloc_large_system_hash(const char *tablename
,
7774 unsigned long bucketsize
,
7775 unsigned long numentries
,
7778 unsigned int *_hash_shift
,
7779 unsigned int *_hash_mask
,
7780 unsigned long low_limit
,
7781 unsigned long high_limit
)
7783 unsigned long long max
= high_limit
;
7784 unsigned long log2qty
, size
;
7788 /* allow the kernel cmdline to have a say */
7790 /* round applicable memory size up to nearest megabyte */
7791 numentries
= nr_kernel_pages
;
7792 numentries
-= arch_reserved_kernel_pages();
7794 /* It isn't necessary when PAGE_SIZE >= 1MB */
7795 if (PAGE_SHIFT
< 20)
7796 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7798 #if __BITS_PER_LONG > 32
7800 unsigned long adapt
;
7802 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
7803 adapt
<<= ADAPT_SCALE_SHIFT
)
7808 /* limit to 1 bucket per 2^scale bytes of low memory */
7809 if (scale
> PAGE_SHIFT
)
7810 numentries
>>= (scale
- PAGE_SHIFT
);
7812 numentries
<<= (PAGE_SHIFT
- scale
);
7814 /* Make sure we've got at least a 0-order allocation.. */
7815 if (unlikely(flags
& HASH_SMALL
)) {
7816 /* Makes no sense without HASH_EARLY */
7817 WARN_ON(!(flags
& HASH_EARLY
));
7818 if (!(numentries
>> *_hash_shift
)) {
7819 numentries
= 1UL << *_hash_shift
;
7820 BUG_ON(!numentries
);
7822 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7823 numentries
= PAGE_SIZE
/ bucketsize
;
7825 numentries
= roundup_pow_of_two(numentries
);
7827 /* limit allocation size to 1/16 total memory by default */
7829 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7830 do_div(max
, bucketsize
);
7832 max
= min(max
, 0x80000000ULL
);
7834 if (numentries
< low_limit
)
7835 numentries
= low_limit
;
7836 if (numentries
> max
)
7839 log2qty
= ilog2(numentries
);
7841 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
7843 size
= bucketsize
<< log2qty
;
7844 if (flags
& HASH_EARLY
) {
7845 if (flags
& HASH_ZERO
)
7846 table
= memblock_alloc_nopanic(size
,
7849 table
= memblock_alloc_raw(size
,
7851 } else if (hashdist
) {
7852 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
7855 * If bucketsize is not a power-of-two, we may free
7856 * some pages at the end of hash table which
7857 * alloc_pages_exact() automatically does
7859 if (get_order(size
) < MAX_ORDER
) {
7860 table
= alloc_pages_exact(size
, gfp_flags
);
7861 kmemleak_alloc(table
, size
, 1, gfp_flags
);
7864 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7867 panic("Failed to allocate %s hash table\n", tablename
);
7869 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7870 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7873 *_hash_shift
= log2qty
;
7875 *_hash_mask
= (1 << log2qty
) - 1;
7881 * This function checks whether pageblock includes unmovable pages or not.
7882 * If @count is not zero, it is okay to include less @count unmovable pages
7884 * PageLRU check without isolation or lru_lock could race so that
7885 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7886 * check without lock_page also may miss some movable non-lru pages at
7887 * race condition. So you can't expect this function should be exact.
7889 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7890 int migratetype
, int flags
)
7892 unsigned long pfn
, iter
, found
;
7895 * TODO we could make this much more efficient by not checking every
7896 * page in the range if we know all of them are in MOVABLE_ZONE and
7897 * that the movable zone guarantees that pages are migratable but
7898 * the later is not the case right now unfortunatelly. E.g. movablecore
7899 * can still lead to having bootmem allocations in zone_movable.
7903 * CMA allocations (alloc_contig_range) really need to mark isolate
7904 * CMA pageblocks even when they are not movable in fact so consider
7905 * them movable here.
7907 if (is_migrate_cma(migratetype
) &&
7908 is_migrate_cma(get_pageblock_migratetype(page
)))
7911 pfn
= page_to_pfn(page
);
7912 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7913 unsigned long check
= pfn
+ iter
;
7915 if (!pfn_valid_within(check
))
7918 page
= pfn_to_page(check
);
7920 if (PageReserved(page
))
7924 * If the zone is movable and we have ruled out all reserved
7925 * pages then it should be reasonably safe to assume the rest
7928 if (zone_idx(zone
) == ZONE_MOVABLE
)
7932 * Hugepages are not in LRU lists, but they're movable.
7933 * We need not scan over tail pages bacause we don't
7934 * handle each tail page individually in migration.
7936 if (PageHuge(page
)) {
7937 struct page
*head
= compound_head(page
);
7938 unsigned int skip_pages
;
7940 if (!hugepage_migration_supported(page_hstate(head
)))
7943 skip_pages
= (1 << compound_order(head
)) - (page
- head
);
7944 iter
+= skip_pages
- 1;
7949 * We can't use page_count without pin a page
7950 * because another CPU can free compound page.
7951 * This check already skips compound tails of THP
7952 * because their page->_refcount is zero at all time.
7954 if (!page_ref_count(page
)) {
7955 if (PageBuddy(page
))
7956 iter
+= (1 << page_order(page
)) - 1;
7961 * The HWPoisoned page may be not in buddy system, and
7962 * page_count() is not 0.
7964 if ((flags
& SKIP_HWPOISON
) && PageHWPoison(page
))
7967 if (__PageMovable(page
))
7973 * If there are RECLAIMABLE pages, we need to check
7974 * it. But now, memory offline itself doesn't call
7975 * shrink_node_slabs() and it still to be fixed.
7978 * If the page is not RAM, page_count()should be 0.
7979 * we don't need more check. This is an _used_ not-movable page.
7981 * The problematic thing here is PG_reserved pages. PG_reserved
7982 * is set to both of a memory hole page and a _used_ kernel
7990 WARN_ON_ONCE(zone_idx(zone
) == ZONE_MOVABLE
);
7991 if (flags
& REPORT_FAILURE
)
7992 dump_page(pfn_to_page(pfn
+iter
), "unmovable page");
7996 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7998 static unsigned long pfn_max_align_down(unsigned long pfn
)
8000 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8001 pageblock_nr_pages
) - 1);
8004 static unsigned long pfn_max_align_up(unsigned long pfn
)
8006 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8007 pageblock_nr_pages
));
8010 /* [start, end) must belong to a single zone. */
8011 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8012 unsigned long start
, unsigned long end
)
8014 /* This function is based on compact_zone() from compaction.c. */
8015 unsigned long nr_reclaimed
;
8016 unsigned long pfn
= start
;
8017 unsigned int tries
= 0;
8022 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8023 if (fatal_signal_pending(current
)) {
8028 if (list_empty(&cc
->migratepages
)) {
8029 cc
->nr_migratepages
= 0;
8030 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8036 } else if (++tries
== 5) {
8037 ret
= ret
< 0 ? ret
: -EBUSY
;
8041 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8043 cc
->nr_migratepages
-= nr_reclaimed
;
8045 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
8046 NULL
, 0, cc
->mode
, MR_CONTIG_RANGE
);
8049 putback_movable_pages(&cc
->migratepages
);
8056 * alloc_contig_range() -- tries to allocate given range of pages
8057 * @start: start PFN to allocate
8058 * @end: one-past-the-last PFN to allocate
8059 * @migratetype: migratetype of the underlaying pageblocks (either
8060 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8061 * in range must have the same migratetype and it must
8062 * be either of the two.
8063 * @gfp_mask: GFP mask to use during compaction
8065 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8066 * aligned. The PFN range must belong to a single zone.
8068 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8069 * pageblocks in the range. Once isolated, the pageblocks should not
8070 * be modified by others.
8072 * Returns zero on success or negative error code. On success all
8073 * pages which PFN is in [start, end) are allocated for the caller and
8074 * need to be freed with free_contig_range().
8076 int alloc_contig_range(unsigned long start
, unsigned long end
,
8077 unsigned migratetype
, gfp_t gfp_mask
)
8079 unsigned long outer_start
, outer_end
;
8083 struct compact_control cc
= {
8084 .nr_migratepages
= 0,
8086 .zone
= page_zone(pfn_to_page(start
)),
8087 .mode
= MIGRATE_SYNC
,
8088 .ignore_skip_hint
= true,
8089 .no_set_skip_hint
= true,
8090 .gfp_mask
= current_gfp_context(gfp_mask
),
8092 INIT_LIST_HEAD(&cc
.migratepages
);
8095 * What we do here is we mark all pageblocks in range as
8096 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8097 * have different sizes, and due to the way page allocator
8098 * work, we align the range to biggest of the two pages so
8099 * that page allocator won't try to merge buddies from
8100 * different pageblocks and change MIGRATE_ISOLATE to some
8101 * other migration type.
8103 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8104 * migrate the pages from an unaligned range (ie. pages that
8105 * we are interested in). This will put all the pages in
8106 * range back to page allocator as MIGRATE_ISOLATE.
8108 * When this is done, we take the pages in range from page
8109 * allocator removing them from the buddy system. This way
8110 * page allocator will never consider using them.
8112 * This lets us mark the pageblocks back as
8113 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8114 * aligned range but not in the unaligned, original range are
8115 * put back to page allocator so that buddy can use them.
8118 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8119 pfn_max_align_up(end
), migratetype
, 0);
8124 * In case of -EBUSY, we'd like to know which page causes problem.
8125 * So, just fall through. test_pages_isolated() has a tracepoint
8126 * which will report the busy page.
8128 * It is possible that busy pages could become available before
8129 * the call to test_pages_isolated, and the range will actually be
8130 * allocated. So, if we fall through be sure to clear ret so that
8131 * -EBUSY is not accidentally used or returned to caller.
8133 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8134 if (ret
&& ret
!= -EBUSY
)
8139 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8140 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8141 * more, all pages in [start, end) are free in page allocator.
8142 * What we are going to do is to allocate all pages from
8143 * [start, end) (that is remove them from page allocator).
8145 * The only problem is that pages at the beginning and at the
8146 * end of interesting range may be not aligned with pages that
8147 * page allocator holds, ie. they can be part of higher order
8148 * pages. Because of this, we reserve the bigger range and
8149 * once this is done free the pages we are not interested in.
8151 * We don't have to hold zone->lock here because the pages are
8152 * isolated thus they won't get removed from buddy.
8155 lru_add_drain_all();
8156 drain_all_pages(cc
.zone
);
8159 outer_start
= start
;
8160 while (!PageBuddy(pfn_to_page(outer_start
))) {
8161 if (++order
>= MAX_ORDER
) {
8162 outer_start
= start
;
8165 outer_start
&= ~0UL << order
;
8168 if (outer_start
!= start
) {
8169 order
= page_order(pfn_to_page(outer_start
));
8172 * outer_start page could be small order buddy page and
8173 * it doesn't include start page. Adjust outer_start
8174 * in this case to report failed page properly
8175 * on tracepoint in test_pages_isolated()
8177 if (outer_start
+ (1UL << order
) <= start
)
8178 outer_start
= start
;
8181 /* Make sure the range is really isolated. */
8182 if (test_pages_isolated(outer_start
, end
, false)) {
8183 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8184 __func__
, outer_start
, end
);
8189 /* Grab isolated pages from freelists. */
8190 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8196 /* Free head and tail (if any) */
8197 if (start
!= outer_start
)
8198 free_contig_range(outer_start
, start
- outer_start
);
8199 if (end
!= outer_end
)
8200 free_contig_range(end
, outer_end
- end
);
8203 undo_isolate_page_range(pfn_max_align_down(start
),
8204 pfn_max_align_up(end
), migratetype
);
8208 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
8210 unsigned int count
= 0;
8212 for (; nr_pages
--; pfn
++) {
8213 struct page
*page
= pfn_to_page(pfn
);
8215 count
+= page_count(page
) != 1;
8218 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8222 #ifdef CONFIG_MEMORY_HOTPLUG
8224 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8225 * page high values need to be recalulated.
8227 void __meminit
zone_pcp_update(struct zone
*zone
)
8230 mutex_lock(&pcp_batch_high_lock
);
8231 for_each_possible_cpu(cpu
)
8232 pageset_set_high_and_batch(zone
,
8233 per_cpu_ptr(zone
->pageset
, cpu
));
8234 mutex_unlock(&pcp_batch_high_lock
);
8238 void zone_pcp_reset(struct zone
*zone
)
8240 unsigned long flags
;
8242 struct per_cpu_pageset
*pset
;
8244 /* avoid races with drain_pages() */
8245 local_irq_save(flags
);
8246 if (zone
->pageset
!= &boot_pageset
) {
8247 for_each_online_cpu(cpu
) {
8248 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8249 drain_zonestat(zone
, pset
);
8251 free_percpu(zone
->pageset
);
8252 zone
->pageset
= &boot_pageset
;
8254 local_irq_restore(flags
);
8257 #ifdef CONFIG_MEMORY_HOTREMOVE
8259 * All pages in the range must be in a single zone and isolated
8260 * before calling this.
8263 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8267 unsigned int order
, i
;
8269 unsigned long flags
;
8270 /* find the first valid pfn */
8271 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8276 offline_mem_sections(pfn
, end_pfn
);
8277 zone
= page_zone(pfn_to_page(pfn
));
8278 spin_lock_irqsave(&zone
->lock
, flags
);
8280 while (pfn
< end_pfn
) {
8281 if (!pfn_valid(pfn
)) {
8285 page
= pfn_to_page(pfn
);
8287 * The HWPoisoned page may be not in buddy system, and
8288 * page_count() is not 0.
8290 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8292 SetPageReserved(page
);
8296 BUG_ON(page_count(page
));
8297 BUG_ON(!PageBuddy(page
));
8298 order
= page_order(page
);
8299 #ifdef CONFIG_DEBUG_VM
8300 pr_info("remove from free list %lx %d %lx\n",
8301 pfn
, 1 << order
, end_pfn
);
8303 list_del(&page
->lru
);
8304 rmv_page_order(page
);
8305 zone
->free_area
[order
].nr_free
--;
8306 for (i
= 0; i
< (1 << order
); i
++)
8307 SetPageReserved((page
+i
));
8308 pfn
+= (1 << order
);
8310 spin_unlock_irqrestore(&zone
->lock
, flags
);
8314 bool is_free_buddy_page(struct page
*page
)
8316 struct zone
*zone
= page_zone(page
);
8317 unsigned long pfn
= page_to_pfn(page
);
8318 unsigned long flags
;
8321 spin_lock_irqsave(&zone
->lock
, flags
);
8322 for (order
= 0; order
< MAX_ORDER
; order
++) {
8323 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8325 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8328 spin_unlock_irqrestore(&zone
->lock
, flags
);
8330 return order
< MAX_ORDER
;
8333 #ifdef CONFIG_MEMORY_FAILURE
8335 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8336 * test is performed under the zone lock to prevent a race against page
8339 bool set_hwpoison_free_buddy_page(struct page
*page
)
8341 struct zone
*zone
= page_zone(page
);
8342 unsigned long pfn
= page_to_pfn(page
);
8343 unsigned long flags
;
8345 bool hwpoisoned
= false;
8347 spin_lock_irqsave(&zone
->lock
, flags
);
8348 for (order
= 0; order
< MAX_ORDER
; order
++) {
8349 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8351 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8352 if (!TestSetPageHWPoison(page
))
8357 spin_unlock_irqrestore(&zone
->lock
, flags
);