1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
78 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
79 static DEFINE_MUTEX(pcp_batch_high_lock
);
80 #define MIN_PERCPU_PAGELIST_FRACTION (8)
82 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
83 DEFINE_PER_CPU(int, numa_node
);
84 EXPORT_PER_CPU_SYMBOL(numa_node
);
87 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
89 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
91 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
92 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
93 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
94 * defined in <linux/topology.h>.
96 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
97 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
100 /* work_structs for global per-cpu drains */
103 struct work_struct work
;
105 DEFINE_MUTEX(pcpu_drain_mutex
);
106 DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
108 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
109 volatile unsigned long latent_entropy __latent_entropy
;
110 EXPORT_SYMBOL(latent_entropy
);
114 * Array of node states.
116 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
117 [N_POSSIBLE
] = NODE_MASK_ALL
,
118 [N_ONLINE
] = { { [0] = 1UL } },
120 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
121 #ifdef CONFIG_HIGHMEM
122 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
124 [N_MEMORY
] = { { [0] = 1UL } },
125 [N_CPU
] = { { [0] = 1UL } },
128 EXPORT_SYMBOL(node_states
);
130 atomic_long_t _totalram_pages __read_mostly
;
131 EXPORT_SYMBOL(_totalram_pages
);
132 unsigned long totalreserve_pages __read_mostly
;
133 unsigned long totalcma_pages __read_mostly
;
135 int percpu_pagelist_fraction
;
136 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
137 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
138 DEFINE_STATIC_KEY_TRUE(init_on_alloc
);
140 DEFINE_STATIC_KEY_FALSE(init_on_alloc
);
142 EXPORT_SYMBOL(init_on_alloc
);
144 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
145 DEFINE_STATIC_KEY_TRUE(init_on_free
);
147 DEFINE_STATIC_KEY_FALSE(init_on_free
);
149 EXPORT_SYMBOL(init_on_free
);
151 static int __init
early_init_on_alloc(char *buf
)
158 ret
= kstrtobool(buf
, &bool_result
);
159 if (bool_result
&& page_poisoning_enabled())
160 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
162 static_branch_enable(&init_on_alloc
);
164 static_branch_disable(&init_on_alloc
);
167 early_param("init_on_alloc", early_init_on_alloc
);
169 static int __init
early_init_on_free(char *buf
)
176 ret
= kstrtobool(buf
, &bool_result
);
177 if (bool_result
&& page_poisoning_enabled())
178 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
180 static_branch_enable(&init_on_free
);
182 static_branch_disable(&init_on_free
);
185 early_param("init_on_free", early_init_on_free
);
188 * A cached value of the page's pageblock's migratetype, used when the page is
189 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
190 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
191 * Also the migratetype set in the page does not necessarily match the pcplist
192 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
193 * other index - this ensures that it will be put on the correct CMA freelist.
195 static inline int get_pcppage_migratetype(struct page
*page
)
200 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
202 page
->index
= migratetype
;
205 #ifdef CONFIG_PM_SLEEP
207 * The following functions are used by the suspend/hibernate code to temporarily
208 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
209 * while devices are suspended. To avoid races with the suspend/hibernate code,
210 * they should always be called with system_transition_mutex held
211 * (gfp_allowed_mask also should only be modified with system_transition_mutex
212 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
213 * with that modification).
216 static gfp_t saved_gfp_mask
;
218 void pm_restore_gfp_mask(void)
220 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
221 if (saved_gfp_mask
) {
222 gfp_allowed_mask
= saved_gfp_mask
;
227 void pm_restrict_gfp_mask(void)
229 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
230 WARN_ON(saved_gfp_mask
);
231 saved_gfp_mask
= gfp_allowed_mask
;
232 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
235 bool pm_suspended_storage(void)
237 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
241 #endif /* CONFIG_PM_SLEEP */
243 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
244 unsigned int pageblock_order __read_mostly
;
247 static void __free_pages_ok(struct page
*page
, unsigned int order
);
250 * results with 256, 32 in the lowmem_reserve sysctl:
251 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
252 * 1G machine -> (16M dma, 784M normal, 224M high)
253 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
254 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
255 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
257 * TBD: should special case ZONE_DMA32 machines here - in those we normally
258 * don't need any ZONE_NORMAL reservation
260 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
261 #ifdef CONFIG_ZONE_DMA
264 #ifdef CONFIG_ZONE_DMA32
268 #ifdef CONFIG_HIGHMEM
274 static char * const zone_names
[MAX_NR_ZONES
] = {
275 #ifdef CONFIG_ZONE_DMA
278 #ifdef CONFIG_ZONE_DMA32
282 #ifdef CONFIG_HIGHMEM
286 #ifdef CONFIG_ZONE_DEVICE
291 const char * const migratetype_names
[MIGRATE_TYPES
] = {
299 #ifdef CONFIG_MEMORY_ISOLATION
304 compound_page_dtor
* const compound_page_dtors
[] = {
307 #ifdef CONFIG_HUGETLB_PAGE
310 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
315 int min_free_kbytes
= 1024;
316 int user_min_free_kbytes
= -1;
317 #ifdef CONFIG_DISCONTIGMEM
319 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
320 * are not on separate NUMA nodes. Functionally this works but with
321 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
322 * quite small. By default, do not boost watermarks on discontigmem as in
323 * many cases very high-order allocations like THP are likely to be
324 * unsupported and the premature reclaim offsets the advantage of long-term
325 * fragmentation avoidance.
327 int watermark_boost_factor __read_mostly
;
329 int watermark_boost_factor __read_mostly
= 15000;
331 int watermark_scale_factor
= 10;
333 static unsigned long nr_kernel_pages __initdata
;
334 static unsigned long nr_all_pages __initdata
;
335 static unsigned long dma_reserve __initdata
;
337 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
338 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
339 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
340 static unsigned long required_kernelcore __initdata
;
341 static unsigned long required_kernelcore_percent __initdata
;
342 static unsigned long required_movablecore __initdata
;
343 static unsigned long required_movablecore_percent __initdata
;
344 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
345 static bool mirrored_kernelcore __meminitdata
;
347 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
349 EXPORT_SYMBOL(movable_zone
);
350 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
353 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
354 unsigned int nr_online_nodes __read_mostly
= 1;
355 EXPORT_SYMBOL(nr_node_ids
);
356 EXPORT_SYMBOL(nr_online_nodes
);
359 int page_group_by_mobility_disabled __read_mostly
;
361 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
363 * During boot we initialize deferred pages on-demand, as needed, but once
364 * page_alloc_init_late() has finished, the deferred pages are all initialized,
365 * and we can permanently disable that path.
367 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
370 * Calling kasan_free_pages() only after deferred memory initialization
371 * has completed. Poisoning pages during deferred memory init will greatly
372 * lengthen the process and cause problem in large memory systems as the
373 * deferred pages initialization is done with interrupt disabled.
375 * Assuming that there will be no reference to those newly initialized
376 * pages before they are ever allocated, this should have no effect on
377 * KASAN memory tracking as the poison will be properly inserted at page
378 * allocation time. The only corner case is when pages are allocated by
379 * on-demand allocation and then freed again before the deferred pages
380 * initialization is done, but this is not likely to happen.
382 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
384 if (!static_branch_unlikely(&deferred_pages
))
385 kasan_free_pages(page
, order
);
388 /* Returns true if the struct page for the pfn is uninitialised */
389 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
391 int nid
= early_pfn_to_nid(pfn
);
393 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
400 * Returns true when the remaining initialisation should be deferred until
401 * later in the boot cycle when it can be parallelised.
403 static bool __meminit
404 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
406 static unsigned long prev_end_pfn
, nr_initialised
;
409 * prev_end_pfn static that contains the end of previous zone
410 * No need to protect because called very early in boot before smp_init.
412 if (prev_end_pfn
!= end_pfn
) {
413 prev_end_pfn
= end_pfn
;
417 /* Always populate low zones for address-constrained allocations */
418 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
422 * We start only with one section of pages, more pages are added as
423 * needed until the rest of deferred pages are initialized.
426 if ((nr_initialised
> PAGES_PER_SECTION
) &&
427 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
428 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
434 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
436 static inline bool early_page_uninitialised(unsigned long pfn
)
441 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
447 /* Return a pointer to the bitmap storing bits affecting a block of pages */
448 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
451 #ifdef CONFIG_SPARSEMEM
452 return section_to_usemap(__pfn_to_section(pfn
));
454 return page_zone(page
)->pageblock_flags
;
455 #endif /* CONFIG_SPARSEMEM */
458 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
460 #ifdef CONFIG_SPARSEMEM
461 pfn
&= (PAGES_PER_SECTION
-1);
462 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
464 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
465 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
466 #endif /* CONFIG_SPARSEMEM */
470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
471 * @page: The page within the block of interest
472 * @pfn: The target page frame number
473 * @end_bitidx: The last bit of interest to retrieve
474 * @mask: mask of bits that the caller is interested in
476 * Return: pageblock_bits flags
478 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
480 unsigned long end_bitidx
,
483 unsigned long *bitmap
;
484 unsigned long bitidx
, word_bitidx
;
487 bitmap
= get_pageblock_bitmap(page
, pfn
);
488 bitidx
= pfn_to_bitidx(page
, pfn
);
489 word_bitidx
= bitidx
/ BITS_PER_LONG
;
490 bitidx
&= (BITS_PER_LONG
-1);
492 word
= bitmap
[word_bitidx
];
493 bitidx
+= end_bitidx
;
494 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
497 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
498 unsigned long end_bitidx
,
501 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
504 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
506 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
510 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
511 * @page: The page within the block of interest
512 * @flags: The flags to set
513 * @pfn: The target page frame number
514 * @end_bitidx: The last bit of interest
515 * @mask: mask of bits that the caller is interested in
517 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
519 unsigned long end_bitidx
,
522 unsigned long *bitmap
;
523 unsigned long bitidx
, word_bitidx
;
524 unsigned long old_word
, word
;
526 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
527 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
529 bitmap
= get_pageblock_bitmap(page
, pfn
);
530 bitidx
= pfn_to_bitidx(page
, pfn
);
531 word_bitidx
= bitidx
/ BITS_PER_LONG
;
532 bitidx
&= (BITS_PER_LONG
-1);
534 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
536 bitidx
+= end_bitidx
;
537 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
538 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
540 word
= READ_ONCE(bitmap
[word_bitidx
]);
542 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
543 if (word
== old_word
)
549 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
551 if (unlikely(page_group_by_mobility_disabled
&&
552 migratetype
< MIGRATE_PCPTYPES
))
553 migratetype
= MIGRATE_UNMOVABLE
;
555 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
556 PB_migrate
, PB_migrate_end
);
559 #ifdef CONFIG_DEBUG_VM
560 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
564 unsigned long pfn
= page_to_pfn(page
);
565 unsigned long sp
, start_pfn
;
568 seq
= zone_span_seqbegin(zone
);
569 start_pfn
= zone
->zone_start_pfn
;
570 sp
= zone
->spanned_pages
;
571 if (!zone_spans_pfn(zone
, pfn
))
573 } while (zone_span_seqretry(zone
, seq
));
576 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
577 pfn
, zone_to_nid(zone
), zone
->name
,
578 start_pfn
, start_pfn
+ sp
);
583 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
585 if (!pfn_valid_within(page_to_pfn(page
)))
587 if (zone
!= page_zone(page
))
593 * Temporary debugging check for pages not lying within a given zone.
595 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
597 if (page_outside_zone_boundaries(zone
, page
))
599 if (!page_is_consistent(zone
, page
))
605 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
611 static void bad_page(struct page
*page
, const char *reason
,
612 unsigned long bad_flags
)
614 static unsigned long resume
;
615 static unsigned long nr_shown
;
616 static unsigned long nr_unshown
;
619 * Allow a burst of 60 reports, then keep quiet for that minute;
620 * or allow a steady drip of one report per second.
622 if (nr_shown
== 60) {
623 if (time_before(jiffies
, resume
)) {
629 "BUG: Bad page state: %lu messages suppressed\n",
636 resume
= jiffies
+ 60 * HZ
;
638 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
639 current
->comm
, page_to_pfn(page
));
640 __dump_page(page
, reason
);
641 bad_flags
&= page
->flags
;
643 pr_alert("bad because of flags: %#lx(%pGp)\n",
644 bad_flags
, &bad_flags
);
645 dump_page_owner(page
);
650 /* Leave bad fields for debug, except PageBuddy could make trouble */
651 page_mapcount_reset(page
); /* remove PageBuddy */
652 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
656 * Higher-order pages are called "compound pages". They are structured thusly:
658 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
660 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
661 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
663 * The first tail page's ->compound_dtor holds the offset in array of compound
664 * page destructors. See compound_page_dtors.
666 * The first tail page's ->compound_order holds the order of allocation.
667 * This usage means that zero-order pages may not be compound.
670 void free_compound_page(struct page
*page
)
672 mem_cgroup_uncharge(page
);
673 __free_pages_ok(page
, compound_order(page
));
676 void prep_compound_page(struct page
*page
, unsigned int order
)
679 int nr_pages
= 1 << order
;
681 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
682 set_compound_order(page
, order
);
684 for (i
= 1; i
< nr_pages
; i
++) {
685 struct page
*p
= page
+ i
;
686 set_page_count(p
, 0);
687 p
->mapping
= TAIL_MAPPING
;
688 set_compound_head(p
, page
);
690 atomic_set(compound_mapcount_ptr(page
), -1);
691 if (hpage_pincount_available(page
))
692 atomic_set(compound_pincount_ptr(page
), 0);
695 #ifdef CONFIG_DEBUG_PAGEALLOC
696 unsigned int _debug_guardpage_minorder
;
698 bool _debug_pagealloc_enabled_early __read_mostly
699 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
700 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
701 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
702 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
704 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
706 static int __init
early_debug_pagealloc(char *buf
)
708 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
710 early_param("debug_pagealloc", early_debug_pagealloc
);
712 void init_debug_pagealloc(void)
714 if (!debug_pagealloc_enabled())
717 static_branch_enable(&_debug_pagealloc_enabled
);
719 if (!debug_guardpage_minorder())
722 static_branch_enable(&_debug_guardpage_enabled
);
725 static int __init
debug_guardpage_minorder_setup(char *buf
)
729 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
730 pr_err("Bad debug_guardpage_minorder value\n");
733 _debug_guardpage_minorder
= res
;
734 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
737 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
739 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
740 unsigned int order
, int migratetype
)
742 if (!debug_guardpage_enabled())
745 if (order
>= debug_guardpage_minorder())
748 __SetPageGuard(page
);
749 INIT_LIST_HEAD(&page
->lru
);
750 set_page_private(page
, order
);
751 /* Guard pages are not available for any usage */
752 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
757 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
758 unsigned int order
, int migratetype
)
760 if (!debug_guardpage_enabled())
763 __ClearPageGuard(page
);
765 set_page_private(page
, 0);
766 if (!is_migrate_isolate(migratetype
))
767 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
770 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
771 unsigned int order
, int migratetype
) { return false; }
772 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
773 unsigned int order
, int migratetype
) {}
776 static inline void set_page_order(struct page
*page
, unsigned int order
)
778 set_page_private(page
, order
);
779 __SetPageBuddy(page
);
783 * This function checks whether a page is free && is the buddy
784 * we can coalesce a page and its buddy if
785 * (a) the buddy is not in a hole (check before calling!) &&
786 * (b) the buddy is in the buddy system &&
787 * (c) a page and its buddy have the same order &&
788 * (d) a page and its buddy are in the same zone.
790 * For recording whether a page is in the buddy system, we set PageBuddy.
791 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
793 * For recording page's order, we use page_private(page).
795 static inline bool page_is_buddy(struct page
*page
, struct page
*buddy
,
798 if (!page_is_guard(buddy
) && !PageBuddy(buddy
))
801 if (page_order(buddy
) != order
)
805 * zone check is done late to avoid uselessly calculating
806 * zone/node ids for pages that could never merge.
808 if (page_zone_id(page
) != page_zone_id(buddy
))
811 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
816 #ifdef CONFIG_COMPACTION
817 static inline struct capture_control
*task_capc(struct zone
*zone
)
819 struct capture_control
*capc
= current
->capture_control
;
822 !(current
->flags
& PF_KTHREAD
) &&
824 capc
->cc
->zone
== zone
&&
825 capc
->cc
->direct_compaction
? capc
: NULL
;
829 compaction_capture(struct capture_control
*capc
, struct page
*page
,
830 int order
, int migratetype
)
832 if (!capc
|| order
!= capc
->cc
->order
)
835 /* Do not accidentally pollute CMA or isolated regions*/
836 if (is_migrate_cma(migratetype
) ||
837 is_migrate_isolate(migratetype
))
841 * Do not let lower order allocations polluate a movable pageblock.
842 * This might let an unmovable request use a reclaimable pageblock
843 * and vice-versa but no more than normal fallback logic which can
844 * have trouble finding a high-order free page.
846 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
854 static inline struct capture_control
*task_capc(struct zone
*zone
)
860 compaction_capture(struct capture_control
*capc
, struct page
*page
,
861 int order
, int migratetype
)
865 #endif /* CONFIG_COMPACTION */
868 * Freeing function for a buddy system allocator.
870 * The concept of a buddy system is to maintain direct-mapped table
871 * (containing bit values) for memory blocks of various "orders".
872 * The bottom level table contains the map for the smallest allocatable
873 * units of memory (here, pages), and each level above it describes
874 * pairs of units from the levels below, hence, "buddies".
875 * At a high level, all that happens here is marking the table entry
876 * at the bottom level available, and propagating the changes upward
877 * as necessary, plus some accounting needed to play nicely with other
878 * parts of the VM system.
879 * At each level, we keep a list of pages, which are heads of continuous
880 * free pages of length of (1 << order) and marked with PageBuddy.
881 * Page's order is recorded in page_private(page) field.
882 * So when we are allocating or freeing one, we can derive the state of the
883 * other. That is, if we allocate a small block, and both were
884 * free, the remainder of the region must be split into blocks.
885 * If a block is freed, and its buddy is also free, then this
886 * triggers coalescing into a block of larger size.
891 static inline void __free_one_page(struct page
*page
,
893 struct zone
*zone
, unsigned int order
,
896 unsigned long combined_pfn
;
897 unsigned long uninitialized_var(buddy_pfn
);
899 unsigned int max_order
;
900 struct capture_control
*capc
= task_capc(zone
);
902 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
904 VM_BUG_ON(!zone_is_initialized(zone
));
905 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
907 VM_BUG_ON(migratetype
== -1);
908 if (likely(!is_migrate_isolate(migratetype
)))
909 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
911 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
912 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
915 while (order
< max_order
- 1) {
916 if (compaction_capture(capc
, page
, order
, migratetype
)) {
917 __mod_zone_freepage_state(zone
, -(1 << order
),
921 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
922 buddy
= page
+ (buddy_pfn
- pfn
);
924 if (!pfn_valid_within(buddy_pfn
))
926 if (!page_is_buddy(page
, buddy
, order
))
929 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
930 * merge with it and move up one order.
932 if (page_is_guard(buddy
))
933 clear_page_guard(zone
, buddy
, order
, migratetype
);
935 del_page_from_free_area(buddy
, &zone
->free_area
[order
]);
936 combined_pfn
= buddy_pfn
& pfn
;
937 page
= page
+ (combined_pfn
- pfn
);
941 if (max_order
< MAX_ORDER
) {
942 /* If we are here, it means order is >= pageblock_order.
943 * We want to prevent merge between freepages on isolate
944 * pageblock and normal pageblock. Without this, pageblock
945 * isolation could cause incorrect freepage or CMA accounting.
947 * We don't want to hit this code for the more frequent
950 if (unlikely(has_isolate_pageblock(zone
))) {
953 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
954 buddy
= page
+ (buddy_pfn
- pfn
);
955 buddy_mt
= get_pageblock_migratetype(buddy
);
957 if (migratetype
!= buddy_mt
958 && (is_migrate_isolate(migratetype
) ||
959 is_migrate_isolate(buddy_mt
)))
963 goto continue_merging
;
967 set_page_order(page
, order
);
970 * If this is not the largest possible page, check if the buddy
971 * of the next-highest order is free. If it is, it's possible
972 * that pages are being freed that will coalesce soon. In case,
973 * that is happening, add the free page to the tail of the list
974 * so it's less likely to be used soon and more likely to be merged
975 * as a higher order page
977 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)
978 && !is_shuffle_order(order
)) {
979 struct page
*higher_page
, *higher_buddy
;
980 combined_pfn
= buddy_pfn
& pfn
;
981 higher_page
= page
+ (combined_pfn
- pfn
);
982 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
983 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
984 if (pfn_valid_within(buddy_pfn
) &&
985 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
986 add_to_free_area_tail(page
, &zone
->free_area
[order
],
992 if (is_shuffle_order(order
))
993 add_to_free_area_random(page
, &zone
->free_area
[order
],
996 add_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
1001 * A bad page could be due to a number of fields. Instead of multiple branches,
1002 * try and check multiple fields with one check. The caller must do a detailed
1003 * check if necessary.
1005 static inline bool page_expected_state(struct page
*page
,
1006 unsigned long check_flags
)
1008 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1011 if (unlikely((unsigned long)page
->mapping
|
1012 page_ref_count(page
) |
1014 (unsigned long)page
->mem_cgroup
|
1016 (page
->flags
& check_flags
)))
1022 static void free_pages_check_bad(struct page
*page
)
1024 const char *bad_reason
;
1025 unsigned long bad_flags
;
1030 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1031 bad_reason
= "nonzero mapcount";
1032 if (unlikely(page
->mapping
!= NULL
))
1033 bad_reason
= "non-NULL mapping";
1034 if (unlikely(page_ref_count(page
) != 0))
1035 bad_reason
= "nonzero _refcount";
1036 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
1037 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1038 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
1041 if (unlikely(page
->mem_cgroup
))
1042 bad_reason
= "page still charged to cgroup";
1044 bad_page(page
, bad_reason
, bad_flags
);
1047 static inline int free_pages_check(struct page
*page
)
1049 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1052 /* Something has gone sideways, find it */
1053 free_pages_check_bad(page
);
1057 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1062 * We rely page->lru.next never has bit 0 set, unless the page
1063 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1065 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1067 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1071 switch (page
- head_page
) {
1073 /* the first tail page: ->mapping may be compound_mapcount() */
1074 if (unlikely(compound_mapcount(page
))) {
1075 bad_page(page
, "nonzero compound_mapcount", 0);
1081 * the second tail page: ->mapping is
1082 * deferred_list.next -- ignore value.
1086 if (page
->mapping
!= TAIL_MAPPING
) {
1087 bad_page(page
, "corrupted mapping in tail page", 0);
1092 if (unlikely(!PageTail(page
))) {
1093 bad_page(page
, "PageTail not set", 0);
1096 if (unlikely(compound_head(page
) != head_page
)) {
1097 bad_page(page
, "compound_head not consistent", 0);
1102 page
->mapping
= NULL
;
1103 clear_compound_head(page
);
1107 static void kernel_init_free_pages(struct page
*page
, int numpages
)
1111 for (i
= 0; i
< numpages
; i
++)
1112 clear_highpage(page
+ i
);
1115 static __always_inline
bool free_pages_prepare(struct page
*page
,
1116 unsigned int order
, bool check_free
)
1120 VM_BUG_ON_PAGE(PageTail(page
), page
);
1122 trace_mm_page_free(page
, order
);
1125 * Check tail pages before head page information is cleared to
1126 * avoid checking PageCompound for order-0 pages.
1128 if (unlikely(order
)) {
1129 bool compound
= PageCompound(page
);
1132 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1135 ClearPageDoubleMap(page
);
1136 for (i
= 1; i
< (1 << order
); i
++) {
1138 bad
+= free_tail_pages_check(page
, page
+ i
);
1139 if (unlikely(free_pages_check(page
+ i
))) {
1143 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1146 if (PageMappingFlags(page
))
1147 page
->mapping
= NULL
;
1148 if (memcg_kmem_enabled() && PageKmemcg(page
))
1149 __memcg_kmem_uncharge_page(page
, order
);
1151 bad
+= free_pages_check(page
);
1155 page_cpupid_reset_last(page
);
1156 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1157 reset_page_owner(page
, order
);
1159 if (!PageHighMem(page
)) {
1160 debug_check_no_locks_freed(page_address(page
),
1161 PAGE_SIZE
<< order
);
1162 debug_check_no_obj_freed(page_address(page
),
1163 PAGE_SIZE
<< order
);
1165 if (want_init_on_free())
1166 kernel_init_free_pages(page
, 1 << order
);
1168 kernel_poison_pages(page
, 1 << order
, 0);
1170 * arch_free_page() can make the page's contents inaccessible. s390
1171 * does this. So nothing which can access the page's contents should
1172 * happen after this.
1174 arch_free_page(page
, order
);
1176 if (debug_pagealloc_enabled_static())
1177 kernel_map_pages(page
, 1 << order
, 0);
1179 kasan_free_nondeferred_pages(page
, order
);
1184 #ifdef CONFIG_DEBUG_VM
1186 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1187 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1188 * moved from pcp lists to free lists.
1190 static bool free_pcp_prepare(struct page
*page
)
1192 return free_pages_prepare(page
, 0, true);
1195 static bool bulkfree_pcp_prepare(struct page
*page
)
1197 if (debug_pagealloc_enabled_static())
1198 return free_pages_check(page
);
1204 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1205 * moving from pcp lists to free list in order to reduce overhead. With
1206 * debug_pagealloc enabled, they are checked also immediately when being freed
1209 static bool free_pcp_prepare(struct page
*page
)
1211 if (debug_pagealloc_enabled_static())
1212 return free_pages_prepare(page
, 0, true);
1214 return free_pages_prepare(page
, 0, false);
1217 static bool bulkfree_pcp_prepare(struct page
*page
)
1219 return free_pages_check(page
);
1221 #endif /* CONFIG_DEBUG_VM */
1223 static inline void prefetch_buddy(struct page
*page
)
1225 unsigned long pfn
= page_to_pfn(page
);
1226 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1227 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1233 * Frees a number of pages from the PCP lists
1234 * Assumes all pages on list are in same zone, and of same order.
1235 * count is the number of pages to free.
1237 * If the zone was previously in an "all pages pinned" state then look to
1238 * see if this freeing clears that state.
1240 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1241 * pinned" detection logic.
1243 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1244 struct per_cpu_pages
*pcp
)
1246 int migratetype
= 0;
1248 int prefetch_nr
= 0;
1249 bool isolated_pageblocks
;
1250 struct page
*page
, *tmp
;
1254 struct list_head
*list
;
1257 * Remove pages from lists in a round-robin fashion. A
1258 * batch_free count is maintained that is incremented when an
1259 * empty list is encountered. This is so more pages are freed
1260 * off fuller lists instead of spinning excessively around empty
1265 if (++migratetype
== MIGRATE_PCPTYPES
)
1267 list
= &pcp
->lists
[migratetype
];
1268 } while (list_empty(list
));
1270 /* This is the only non-empty list. Free them all. */
1271 if (batch_free
== MIGRATE_PCPTYPES
)
1275 page
= list_last_entry(list
, struct page
, lru
);
1276 /* must delete to avoid corrupting pcp list */
1277 list_del(&page
->lru
);
1280 if (bulkfree_pcp_prepare(page
))
1283 list_add_tail(&page
->lru
, &head
);
1286 * We are going to put the page back to the global
1287 * pool, prefetch its buddy to speed up later access
1288 * under zone->lock. It is believed the overhead of
1289 * an additional test and calculating buddy_pfn here
1290 * can be offset by reduced memory latency later. To
1291 * avoid excessive prefetching due to large count, only
1292 * prefetch buddy for the first pcp->batch nr of pages.
1294 if (prefetch_nr
++ < pcp
->batch
)
1295 prefetch_buddy(page
);
1296 } while (--count
&& --batch_free
&& !list_empty(list
));
1299 spin_lock(&zone
->lock
);
1300 isolated_pageblocks
= has_isolate_pageblock(zone
);
1303 * Use safe version since after __free_one_page(),
1304 * page->lru.next will not point to original list.
1306 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1307 int mt
= get_pcppage_migratetype(page
);
1308 /* MIGRATE_ISOLATE page should not go to pcplists */
1309 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1310 /* Pageblock could have been isolated meanwhile */
1311 if (unlikely(isolated_pageblocks
))
1312 mt
= get_pageblock_migratetype(page
);
1314 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1315 trace_mm_page_pcpu_drain(page
, 0, mt
);
1317 spin_unlock(&zone
->lock
);
1320 static void free_one_page(struct zone
*zone
,
1321 struct page
*page
, unsigned long pfn
,
1325 spin_lock(&zone
->lock
);
1326 if (unlikely(has_isolate_pageblock(zone
) ||
1327 is_migrate_isolate(migratetype
))) {
1328 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1330 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1331 spin_unlock(&zone
->lock
);
1334 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1335 unsigned long zone
, int nid
)
1337 mm_zero_struct_page(page
);
1338 set_page_links(page
, zone
, nid
, pfn
);
1339 init_page_count(page
);
1340 page_mapcount_reset(page
);
1341 page_cpupid_reset_last(page
);
1342 page_kasan_tag_reset(page
);
1344 INIT_LIST_HEAD(&page
->lru
);
1345 #ifdef WANT_PAGE_VIRTUAL
1346 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1347 if (!is_highmem_idx(zone
))
1348 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1352 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1353 static void __meminit
init_reserved_page(unsigned long pfn
)
1358 if (!early_page_uninitialised(pfn
))
1361 nid
= early_pfn_to_nid(pfn
);
1362 pgdat
= NODE_DATA(nid
);
1364 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1365 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1367 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1370 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1373 static inline void init_reserved_page(unsigned long pfn
)
1376 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1379 * Initialised pages do not have PageReserved set. This function is
1380 * called for each range allocated by the bootmem allocator and
1381 * marks the pages PageReserved. The remaining valid pages are later
1382 * sent to the buddy page allocator.
1384 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1386 unsigned long start_pfn
= PFN_DOWN(start
);
1387 unsigned long end_pfn
= PFN_UP(end
);
1389 for (; start_pfn
< end_pfn
; start_pfn
++) {
1390 if (pfn_valid(start_pfn
)) {
1391 struct page
*page
= pfn_to_page(start_pfn
);
1393 init_reserved_page(start_pfn
);
1395 /* Avoid false-positive PageTail() */
1396 INIT_LIST_HEAD(&page
->lru
);
1399 * no need for atomic set_bit because the struct
1400 * page is not visible yet so nobody should
1403 __SetPageReserved(page
);
1408 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1410 unsigned long flags
;
1412 unsigned long pfn
= page_to_pfn(page
);
1414 if (!free_pages_prepare(page
, order
, true))
1417 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1418 local_irq_save(flags
);
1419 __count_vm_events(PGFREE
, 1 << order
);
1420 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1421 local_irq_restore(flags
);
1424 void __free_pages_core(struct page
*page
, unsigned int order
)
1426 unsigned int nr_pages
= 1 << order
;
1427 struct page
*p
= page
;
1431 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1433 __ClearPageReserved(p
);
1434 set_page_count(p
, 0);
1436 __ClearPageReserved(p
);
1437 set_page_count(p
, 0);
1439 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1440 set_page_refcounted(page
);
1441 __free_pages(page
, order
);
1444 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1445 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1447 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1449 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1451 static DEFINE_SPINLOCK(early_pfn_lock
);
1454 spin_lock(&early_pfn_lock
);
1455 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1457 nid
= first_online_node
;
1458 spin_unlock(&early_pfn_lock
);
1464 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1465 /* Only safe to use early in boot when initialisation is single-threaded */
1466 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1470 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1471 if (nid
>= 0 && nid
!= node
)
1477 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1484 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1487 if (early_page_uninitialised(pfn
))
1489 __free_pages_core(page
, order
);
1493 * Check that the whole (or subset of) a pageblock given by the interval of
1494 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1495 * with the migration of free compaction scanner. The scanners then need to
1496 * use only pfn_valid_within() check for arches that allow holes within
1499 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1501 * It's possible on some configurations to have a setup like node0 node1 node0
1502 * i.e. it's possible that all pages within a zones range of pages do not
1503 * belong to a single zone. We assume that a border between node0 and node1
1504 * can occur within a single pageblock, but not a node0 node1 node0
1505 * interleaving within a single pageblock. It is therefore sufficient to check
1506 * the first and last page of a pageblock and avoid checking each individual
1507 * page in a pageblock.
1509 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1510 unsigned long end_pfn
, struct zone
*zone
)
1512 struct page
*start_page
;
1513 struct page
*end_page
;
1515 /* end_pfn is one past the range we are checking */
1518 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1521 start_page
= pfn_to_online_page(start_pfn
);
1525 if (page_zone(start_page
) != zone
)
1528 end_page
= pfn_to_page(end_pfn
);
1530 /* This gives a shorter code than deriving page_zone(end_page) */
1531 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1537 void set_zone_contiguous(struct zone
*zone
)
1539 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1540 unsigned long block_end_pfn
;
1542 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1543 for (; block_start_pfn
< zone_end_pfn(zone
);
1544 block_start_pfn
= block_end_pfn
,
1545 block_end_pfn
+= pageblock_nr_pages
) {
1547 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1549 if (!__pageblock_pfn_to_page(block_start_pfn
,
1550 block_end_pfn
, zone
))
1554 /* We confirm that there is no hole */
1555 zone
->contiguous
= true;
1558 void clear_zone_contiguous(struct zone
*zone
)
1560 zone
->contiguous
= false;
1563 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1564 static void __init
deferred_free_range(unsigned long pfn
,
1565 unsigned long nr_pages
)
1573 page
= pfn_to_page(pfn
);
1575 /* Free a large naturally-aligned chunk if possible */
1576 if (nr_pages
== pageblock_nr_pages
&&
1577 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1578 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1579 __free_pages_core(page
, pageblock_order
);
1583 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1584 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1585 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1586 __free_pages_core(page
, 0);
1590 /* Completion tracking for deferred_init_memmap() threads */
1591 static atomic_t pgdat_init_n_undone __initdata
;
1592 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1594 static inline void __init
pgdat_init_report_one_done(void)
1596 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1597 complete(&pgdat_init_all_done_comp
);
1601 * Returns true if page needs to be initialized or freed to buddy allocator.
1603 * First we check if pfn is valid on architectures where it is possible to have
1604 * holes within pageblock_nr_pages. On systems where it is not possible, this
1605 * function is optimized out.
1607 * Then, we check if a current large page is valid by only checking the validity
1610 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1612 if (!pfn_valid_within(pfn
))
1614 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1620 * Free pages to buddy allocator. Try to free aligned pages in
1621 * pageblock_nr_pages sizes.
1623 static void __init
deferred_free_pages(unsigned long pfn
,
1624 unsigned long end_pfn
)
1626 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1627 unsigned long nr_free
= 0;
1629 for (; pfn
< end_pfn
; pfn
++) {
1630 if (!deferred_pfn_valid(pfn
)) {
1631 deferred_free_range(pfn
- nr_free
, nr_free
);
1633 } else if (!(pfn
& nr_pgmask
)) {
1634 deferred_free_range(pfn
- nr_free
, nr_free
);
1636 touch_nmi_watchdog();
1641 /* Free the last block of pages to allocator */
1642 deferred_free_range(pfn
- nr_free
, nr_free
);
1646 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1647 * by performing it only once every pageblock_nr_pages.
1648 * Return number of pages initialized.
1650 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1652 unsigned long end_pfn
)
1654 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1655 int nid
= zone_to_nid(zone
);
1656 unsigned long nr_pages
= 0;
1657 int zid
= zone_idx(zone
);
1658 struct page
*page
= NULL
;
1660 for (; pfn
< end_pfn
; pfn
++) {
1661 if (!deferred_pfn_valid(pfn
)) {
1664 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1665 page
= pfn_to_page(pfn
);
1666 touch_nmi_watchdog();
1670 __init_single_page(page
, pfn
, zid
, nid
);
1677 * This function is meant to pre-load the iterator for the zone init.
1678 * Specifically it walks through the ranges until we are caught up to the
1679 * first_init_pfn value and exits there. If we never encounter the value we
1680 * return false indicating there are no valid ranges left.
1683 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1684 unsigned long *spfn
, unsigned long *epfn
,
1685 unsigned long first_init_pfn
)
1690 * Start out by walking through the ranges in this zone that have
1691 * already been initialized. We don't need to do anything with them
1692 * so we just need to flush them out of the system.
1694 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1695 if (*epfn
<= first_init_pfn
)
1697 if (*spfn
< first_init_pfn
)
1698 *spfn
= first_init_pfn
;
1707 * Initialize and free pages. We do it in two loops: first we initialize
1708 * struct page, then free to buddy allocator, because while we are
1709 * freeing pages we can access pages that are ahead (computing buddy
1710 * page in __free_one_page()).
1712 * In order to try and keep some memory in the cache we have the loop
1713 * broken along max page order boundaries. This way we will not cause
1714 * any issues with the buddy page computation.
1716 static unsigned long __init
1717 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1718 unsigned long *end_pfn
)
1720 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1721 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1722 unsigned long nr_pages
= 0;
1725 /* First we loop through and initialize the page values */
1726 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1729 if (mo_pfn
<= *start_pfn
)
1732 t
= min(mo_pfn
, *end_pfn
);
1733 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1735 if (mo_pfn
< *end_pfn
) {
1736 *start_pfn
= mo_pfn
;
1741 /* Reset values and now loop through freeing pages as needed */
1744 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1750 t
= min(mo_pfn
, epfn
);
1751 deferred_free_pages(spfn
, t
);
1760 /* Initialise remaining memory on a node */
1761 static int __init
deferred_init_memmap(void *data
)
1763 pg_data_t
*pgdat
= data
;
1764 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1765 unsigned long spfn
= 0, epfn
= 0, nr_pages
= 0;
1766 unsigned long first_init_pfn
, flags
;
1767 unsigned long start
= jiffies
;
1772 /* Bind memory initialisation thread to a local node if possible */
1773 if (!cpumask_empty(cpumask
))
1774 set_cpus_allowed_ptr(current
, cpumask
);
1776 pgdat_resize_lock(pgdat
, &flags
);
1777 first_init_pfn
= pgdat
->first_deferred_pfn
;
1778 if (first_init_pfn
== ULONG_MAX
) {
1779 pgdat_resize_unlock(pgdat
, &flags
);
1780 pgdat_init_report_one_done();
1784 /* Sanity check boundaries */
1785 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1786 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1787 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1789 /* Only the highest zone is deferred so find it */
1790 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1791 zone
= pgdat
->node_zones
+ zid
;
1792 if (first_init_pfn
< zone_end_pfn(zone
))
1796 /* If the zone is empty somebody else may have cleared out the zone */
1797 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1802 * Initialize and free pages in MAX_ORDER sized increments so
1803 * that we can avoid introducing any issues with the buddy
1807 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1809 pgdat_resize_unlock(pgdat
, &flags
);
1811 /* Sanity check that the next zone really is unpopulated */
1812 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1814 pr_info("node %d initialised, %lu pages in %ums\n",
1815 pgdat
->node_id
, nr_pages
, jiffies_to_msecs(jiffies
- start
));
1817 pgdat_init_report_one_done();
1822 * If this zone has deferred pages, try to grow it by initializing enough
1823 * deferred pages to satisfy the allocation specified by order, rounded up to
1824 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1825 * of SECTION_SIZE bytes by initializing struct pages in increments of
1826 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1828 * Return true when zone was grown, otherwise return false. We return true even
1829 * when we grow less than requested, to let the caller decide if there are
1830 * enough pages to satisfy the allocation.
1832 * Note: We use noinline because this function is needed only during boot, and
1833 * it is called from a __ref function _deferred_grow_zone. This way we are
1834 * making sure that it is not inlined into permanent text section.
1836 static noinline
bool __init
1837 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1839 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1840 pg_data_t
*pgdat
= zone
->zone_pgdat
;
1841 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1842 unsigned long spfn
, epfn
, flags
;
1843 unsigned long nr_pages
= 0;
1846 /* Only the last zone may have deferred pages */
1847 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1850 pgdat_resize_lock(pgdat
, &flags
);
1853 * If deferred pages have been initialized while we were waiting for
1854 * the lock, return true, as the zone was grown. The caller will retry
1855 * this zone. We won't return to this function since the caller also
1856 * has this static branch.
1858 if (!static_branch_unlikely(&deferred_pages
)) {
1859 pgdat_resize_unlock(pgdat
, &flags
);
1864 * If someone grew this zone while we were waiting for spinlock, return
1865 * true, as there might be enough pages already.
1867 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1868 pgdat_resize_unlock(pgdat
, &flags
);
1872 /* If the zone is empty somebody else may have cleared out the zone */
1873 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1874 first_deferred_pfn
)) {
1875 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1876 pgdat_resize_unlock(pgdat
, &flags
);
1877 /* Retry only once. */
1878 return first_deferred_pfn
!= ULONG_MAX
;
1882 * Initialize and free pages in MAX_ORDER sized increments so
1883 * that we can avoid introducing any issues with the buddy
1886 while (spfn
< epfn
) {
1887 /* update our first deferred PFN for this section */
1888 first_deferred_pfn
= spfn
;
1890 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1892 /* We should only stop along section boundaries */
1893 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
1896 /* If our quota has been met we can stop here */
1897 if (nr_pages
>= nr_pages_needed
)
1901 pgdat
->first_deferred_pfn
= spfn
;
1902 pgdat_resize_unlock(pgdat
, &flags
);
1904 return nr_pages
> 0;
1908 * deferred_grow_zone() is __init, but it is called from
1909 * get_page_from_freelist() during early boot until deferred_pages permanently
1910 * disables this call. This is why we have refdata wrapper to avoid warning,
1911 * and to ensure that the function body gets unloaded.
1914 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1916 return deferred_grow_zone(zone
, order
);
1919 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1921 void __init
page_alloc_init_late(void)
1926 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1928 /* There will be num_node_state(N_MEMORY) threads */
1929 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1930 for_each_node_state(nid
, N_MEMORY
) {
1931 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1934 /* Block until all are initialised */
1935 wait_for_completion(&pgdat_init_all_done_comp
);
1938 * The number of managed pages has changed due to the initialisation
1939 * so the pcpu batch and high limits needs to be updated or the limits
1940 * will be artificially small.
1942 for_each_populated_zone(zone
)
1943 zone_pcp_update(zone
);
1946 * We initialized the rest of the deferred pages. Permanently disable
1947 * on-demand struct page initialization.
1949 static_branch_disable(&deferred_pages
);
1951 /* Reinit limits that are based on free pages after the kernel is up */
1952 files_maxfiles_init();
1955 /* Discard memblock private memory */
1958 for_each_node_state(nid
, N_MEMORY
)
1959 shuffle_free_memory(NODE_DATA(nid
));
1961 for_each_populated_zone(zone
)
1962 set_zone_contiguous(zone
);
1966 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1967 void __init
init_cma_reserved_pageblock(struct page
*page
)
1969 unsigned i
= pageblock_nr_pages
;
1970 struct page
*p
= page
;
1973 __ClearPageReserved(p
);
1974 set_page_count(p
, 0);
1977 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1979 if (pageblock_order
>= MAX_ORDER
) {
1980 i
= pageblock_nr_pages
;
1983 set_page_refcounted(p
);
1984 __free_pages(p
, MAX_ORDER
- 1);
1985 p
+= MAX_ORDER_NR_PAGES
;
1986 } while (i
-= MAX_ORDER_NR_PAGES
);
1988 set_page_refcounted(page
);
1989 __free_pages(page
, pageblock_order
);
1992 adjust_managed_page_count(page
, pageblock_nr_pages
);
1997 * The order of subdivision here is critical for the IO subsystem.
1998 * Please do not alter this order without good reasons and regression
1999 * testing. Specifically, as large blocks of memory are subdivided,
2000 * the order in which smaller blocks are delivered depends on the order
2001 * they're subdivided in this function. This is the primary factor
2002 * influencing the order in which pages are delivered to the IO
2003 * subsystem according to empirical testing, and this is also justified
2004 * by considering the behavior of a buddy system containing a single
2005 * large block of memory acted on by a series of small allocations.
2006 * This behavior is a critical factor in sglist merging's success.
2010 static inline void expand(struct zone
*zone
, struct page
*page
,
2011 int low
, int high
, struct free_area
*area
,
2014 unsigned long size
= 1 << high
;
2016 while (high
> low
) {
2020 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2023 * Mark as guard pages (or page), that will allow to
2024 * merge back to allocator when buddy will be freed.
2025 * Corresponding page table entries will not be touched,
2026 * pages will stay not present in virtual address space
2028 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2031 add_to_free_area(&page
[size
], area
, migratetype
);
2032 set_page_order(&page
[size
], high
);
2036 static void check_new_page_bad(struct page
*page
)
2038 const char *bad_reason
= NULL
;
2039 unsigned long bad_flags
= 0;
2041 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
2042 bad_reason
= "nonzero mapcount";
2043 if (unlikely(page
->mapping
!= NULL
))
2044 bad_reason
= "non-NULL mapping";
2045 if (unlikely(page_ref_count(page
) != 0))
2046 bad_reason
= "nonzero _refcount";
2047 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2048 bad_reason
= "HWPoisoned (hardware-corrupted)";
2049 bad_flags
= __PG_HWPOISON
;
2050 /* Don't complain about hwpoisoned pages */
2051 page_mapcount_reset(page
); /* remove PageBuddy */
2054 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
2055 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
2056 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
2059 if (unlikely(page
->mem_cgroup
))
2060 bad_reason
= "page still charged to cgroup";
2062 bad_page(page
, bad_reason
, bad_flags
);
2066 * This page is about to be returned from the page allocator
2068 static inline int check_new_page(struct page
*page
)
2070 if (likely(page_expected_state(page
,
2071 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2074 check_new_page_bad(page
);
2078 static inline bool free_pages_prezeroed(void)
2080 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
2081 page_poisoning_enabled()) || want_init_on_free();
2084 #ifdef CONFIG_DEBUG_VM
2086 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2087 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2088 * also checked when pcp lists are refilled from the free lists.
2090 static inline bool check_pcp_refill(struct page
*page
)
2092 if (debug_pagealloc_enabled_static())
2093 return check_new_page(page
);
2098 static inline bool check_new_pcp(struct page
*page
)
2100 return check_new_page(page
);
2104 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2105 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2106 * enabled, they are also checked when being allocated from the pcp lists.
2108 static inline bool check_pcp_refill(struct page
*page
)
2110 return check_new_page(page
);
2112 static inline bool check_new_pcp(struct page
*page
)
2114 if (debug_pagealloc_enabled_static())
2115 return check_new_page(page
);
2119 #endif /* CONFIG_DEBUG_VM */
2121 static bool check_new_pages(struct page
*page
, unsigned int order
)
2124 for (i
= 0; i
< (1 << order
); i
++) {
2125 struct page
*p
= page
+ i
;
2127 if (unlikely(check_new_page(p
)))
2134 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2137 set_page_private(page
, 0);
2138 set_page_refcounted(page
);
2140 arch_alloc_page(page
, order
);
2141 if (debug_pagealloc_enabled_static())
2142 kernel_map_pages(page
, 1 << order
, 1);
2143 kasan_alloc_pages(page
, order
);
2144 kernel_poison_pages(page
, 1 << order
, 1);
2145 set_page_owner(page
, order
, gfp_flags
);
2148 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2149 unsigned int alloc_flags
)
2151 post_alloc_hook(page
, order
, gfp_flags
);
2153 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags
))
2154 kernel_init_free_pages(page
, 1 << order
);
2156 if (order
&& (gfp_flags
& __GFP_COMP
))
2157 prep_compound_page(page
, order
);
2160 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2161 * allocate the page. The expectation is that the caller is taking
2162 * steps that will free more memory. The caller should avoid the page
2163 * being used for !PFMEMALLOC purposes.
2165 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2166 set_page_pfmemalloc(page
);
2168 clear_page_pfmemalloc(page
);
2172 * Go through the free lists for the given migratetype and remove
2173 * the smallest available page from the freelists
2175 static __always_inline
2176 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2179 unsigned int current_order
;
2180 struct free_area
*area
;
2183 /* Find a page of the appropriate size in the preferred list */
2184 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2185 area
= &(zone
->free_area
[current_order
]);
2186 page
= get_page_from_free_area(area
, migratetype
);
2189 del_page_from_free_area(page
, area
);
2190 expand(zone
, page
, order
, current_order
, area
, migratetype
);
2191 set_pcppage_migratetype(page
, migratetype
);
2200 * This array describes the order lists are fallen back to when
2201 * the free lists for the desirable migrate type are depleted
2203 static int fallbacks
[MIGRATE_TYPES
][4] = {
2204 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2205 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2206 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2208 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2210 #ifdef CONFIG_MEMORY_ISOLATION
2211 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2216 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2219 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2222 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2223 unsigned int order
) { return NULL
; }
2227 * Move the free pages in a range to the free lists of the requested type.
2228 * Note that start_page and end_pages are not aligned on a pageblock
2229 * boundary. If alignment is required, use move_freepages_block()
2231 static int move_freepages(struct zone
*zone
,
2232 struct page
*start_page
, struct page
*end_page
,
2233 int migratetype
, int *num_movable
)
2237 int pages_moved
= 0;
2239 for (page
= start_page
; page
<= end_page
;) {
2240 if (!pfn_valid_within(page_to_pfn(page
))) {
2245 if (!PageBuddy(page
)) {
2247 * We assume that pages that could be isolated for
2248 * migration are movable. But we don't actually try
2249 * isolating, as that would be expensive.
2252 (PageLRU(page
) || __PageMovable(page
)))
2259 /* Make sure we are not inadvertently changing nodes */
2260 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2261 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2263 order
= page_order(page
);
2264 move_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
2266 pages_moved
+= 1 << order
;
2272 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2273 int migratetype
, int *num_movable
)
2275 unsigned long start_pfn
, end_pfn
;
2276 struct page
*start_page
, *end_page
;
2281 start_pfn
= page_to_pfn(page
);
2282 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2283 start_page
= pfn_to_page(start_pfn
);
2284 end_page
= start_page
+ pageblock_nr_pages
- 1;
2285 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2287 /* Do not cross zone boundaries */
2288 if (!zone_spans_pfn(zone
, start_pfn
))
2290 if (!zone_spans_pfn(zone
, end_pfn
))
2293 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2297 static void change_pageblock_range(struct page
*pageblock_page
,
2298 int start_order
, int migratetype
)
2300 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2302 while (nr_pageblocks
--) {
2303 set_pageblock_migratetype(pageblock_page
, migratetype
);
2304 pageblock_page
+= pageblock_nr_pages
;
2309 * When we are falling back to another migratetype during allocation, try to
2310 * steal extra free pages from the same pageblocks to satisfy further
2311 * allocations, instead of polluting multiple pageblocks.
2313 * If we are stealing a relatively large buddy page, it is likely there will
2314 * be more free pages in the pageblock, so try to steal them all. For
2315 * reclaimable and unmovable allocations, we steal regardless of page size,
2316 * as fragmentation caused by those allocations polluting movable pageblocks
2317 * is worse than movable allocations stealing from unmovable and reclaimable
2320 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2323 * Leaving this order check is intended, although there is
2324 * relaxed order check in next check. The reason is that
2325 * we can actually steal whole pageblock if this condition met,
2326 * but, below check doesn't guarantee it and that is just heuristic
2327 * so could be changed anytime.
2329 if (order
>= pageblock_order
)
2332 if (order
>= pageblock_order
/ 2 ||
2333 start_mt
== MIGRATE_RECLAIMABLE
||
2334 start_mt
== MIGRATE_UNMOVABLE
||
2335 page_group_by_mobility_disabled
)
2341 static inline void boost_watermark(struct zone
*zone
)
2343 unsigned long max_boost
;
2345 if (!watermark_boost_factor
)
2348 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2349 watermark_boost_factor
, 10000);
2352 * high watermark may be uninitialised if fragmentation occurs
2353 * very early in boot so do not boost. We do not fall
2354 * through and boost by pageblock_nr_pages as failing
2355 * allocations that early means that reclaim is not going
2356 * to help and it may even be impossible to reclaim the
2357 * boosted watermark resulting in a hang.
2362 max_boost
= max(pageblock_nr_pages
, max_boost
);
2364 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2369 * This function implements actual steal behaviour. If order is large enough,
2370 * we can steal whole pageblock. If not, we first move freepages in this
2371 * pageblock to our migratetype and determine how many already-allocated pages
2372 * are there in the pageblock with a compatible migratetype. If at least half
2373 * of pages are free or compatible, we can change migratetype of the pageblock
2374 * itself, so pages freed in the future will be put on the correct free list.
2376 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2377 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2379 unsigned int current_order
= page_order(page
);
2380 struct free_area
*area
;
2381 int free_pages
, movable_pages
, alike_pages
;
2384 old_block_type
= get_pageblock_migratetype(page
);
2387 * This can happen due to races and we want to prevent broken
2388 * highatomic accounting.
2390 if (is_migrate_highatomic(old_block_type
))
2393 /* Take ownership for orders >= pageblock_order */
2394 if (current_order
>= pageblock_order
) {
2395 change_pageblock_range(page
, current_order
, start_type
);
2400 * Boost watermarks to increase reclaim pressure to reduce the
2401 * likelihood of future fallbacks. Wake kswapd now as the node
2402 * may be balanced overall and kswapd will not wake naturally.
2404 boost_watermark(zone
);
2405 if (alloc_flags
& ALLOC_KSWAPD
)
2406 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2408 /* We are not allowed to try stealing from the whole block */
2412 free_pages
= move_freepages_block(zone
, page
, start_type
,
2415 * Determine how many pages are compatible with our allocation.
2416 * For movable allocation, it's the number of movable pages which
2417 * we just obtained. For other types it's a bit more tricky.
2419 if (start_type
== MIGRATE_MOVABLE
) {
2420 alike_pages
= movable_pages
;
2423 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2424 * to MOVABLE pageblock, consider all non-movable pages as
2425 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2426 * vice versa, be conservative since we can't distinguish the
2427 * exact migratetype of non-movable pages.
2429 if (old_block_type
== MIGRATE_MOVABLE
)
2430 alike_pages
= pageblock_nr_pages
2431 - (free_pages
+ movable_pages
);
2436 /* moving whole block can fail due to zone boundary conditions */
2441 * If a sufficient number of pages in the block are either free or of
2442 * comparable migratability as our allocation, claim the whole block.
2444 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2445 page_group_by_mobility_disabled
)
2446 set_pageblock_migratetype(page
, start_type
);
2451 area
= &zone
->free_area
[current_order
];
2452 move_to_free_area(page
, area
, start_type
);
2456 * Check whether there is a suitable fallback freepage with requested order.
2457 * If only_stealable is true, this function returns fallback_mt only if
2458 * we can steal other freepages all together. This would help to reduce
2459 * fragmentation due to mixed migratetype pages in one pageblock.
2461 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2462 int migratetype
, bool only_stealable
, bool *can_steal
)
2467 if (area
->nr_free
== 0)
2472 fallback_mt
= fallbacks
[migratetype
][i
];
2473 if (fallback_mt
== MIGRATE_TYPES
)
2476 if (free_area_empty(area
, fallback_mt
))
2479 if (can_steal_fallback(order
, migratetype
))
2482 if (!only_stealable
)
2493 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2494 * there are no empty page blocks that contain a page with a suitable order
2496 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2497 unsigned int alloc_order
)
2500 unsigned long max_managed
, flags
;
2503 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2504 * Check is race-prone but harmless.
2506 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2507 if (zone
->nr_reserved_highatomic
>= max_managed
)
2510 spin_lock_irqsave(&zone
->lock
, flags
);
2512 /* Recheck the nr_reserved_highatomic limit under the lock */
2513 if (zone
->nr_reserved_highatomic
>= max_managed
)
2517 mt
= get_pageblock_migratetype(page
);
2518 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2519 && !is_migrate_cma(mt
)) {
2520 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2521 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2522 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2526 spin_unlock_irqrestore(&zone
->lock
, flags
);
2530 * Used when an allocation is about to fail under memory pressure. This
2531 * potentially hurts the reliability of high-order allocations when under
2532 * intense memory pressure but failed atomic allocations should be easier
2533 * to recover from than an OOM.
2535 * If @force is true, try to unreserve a pageblock even though highatomic
2536 * pageblock is exhausted.
2538 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2541 struct zonelist
*zonelist
= ac
->zonelist
;
2542 unsigned long flags
;
2549 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2552 * Preserve at least one pageblock unless memory pressure
2555 if (!force
&& zone
->nr_reserved_highatomic
<=
2559 spin_lock_irqsave(&zone
->lock
, flags
);
2560 for (order
= 0; order
< MAX_ORDER
; order
++) {
2561 struct free_area
*area
= &(zone
->free_area
[order
]);
2563 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2568 * In page freeing path, migratetype change is racy so
2569 * we can counter several free pages in a pageblock
2570 * in this loop althoug we changed the pageblock type
2571 * from highatomic to ac->migratetype. So we should
2572 * adjust the count once.
2574 if (is_migrate_highatomic_page(page
)) {
2576 * It should never happen but changes to
2577 * locking could inadvertently allow a per-cpu
2578 * drain to add pages to MIGRATE_HIGHATOMIC
2579 * while unreserving so be safe and watch for
2582 zone
->nr_reserved_highatomic
-= min(
2584 zone
->nr_reserved_highatomic
);
2588 * Convert to ac->migratetype and avoid the normal
2589 * pageblock stealing heuristics. Minimally, the caller
2590 * is doing the work and needs the pages. More
2591 * importantly, if the block was always converted to
2592 * MIGRATE_UNMOVABLE or another type then the number
2593 * of pageblocks that cannot be completely freed
2596 set_pageblock_migratetype(page
, ac
->migratetype
);
2597 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2600 spin_unlock_irqrestore(&zone
->lock
, flags
);
2604 spin_unlock_irqrestore(&zone
->lock
, flags
);
2611 * Try finding a free buddy page on the fallback list and put it on the free
2612 * list of requested migratetype, possibly along with other pages from the same
2613 * block, depending on fragmentation avoidance heuristics. Returns true if
2614 * fallback was found so that __rmqueue_smallest() can grab it.
2616 * The use of signed ints for order and current_order is a deliberate
2617 * deviation from the rest of this file, to make the for loop
2618 * condition simpler.
2620 static __always_inline
bool
2621 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2622 unsigned int alloc_flags
)
2624 struct free_area
*area
;
2626 int min_order
= order
;
2632 * Do not steal pages from freelists belonging to other pageblocks
2633 * i.e. orders < pageblock_order. If there are no local zones free,
2634 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2636 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2637 min_order
= pageblock_order
;
2640 * Find the largest available free page in the other list. This roughly
2641 * approximates finding the pageblock with the most free pages, which
2642 * would be too costly to do exactly.
2644 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2646 area
= &(zone
->free_area
[current_order
]);
2647 fallback_mt
= find_suitable_fallback(area
, current_order
,
2648 start_migratetype
, false, &can_steal
);
2649 if (fallback_mt
== -1)
2653 * We cannot steal all free pages from the pageblock and the
2654 * requested migratetype is movable. In that case it's better to
2655 * steal and split the smallest available page instead of the
2656 * largest available page, because even if the next movable
2657 * allocation falls back into a different pageblock than this
2658 * one, it won't cause permanent fragmentation.
2660 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2661 && current_order
> order
)
2670 for (current_order
= order
; current_order
< MAX_ORDER
;
2672 area
= &(zone
->free_area
[current_order
]);
2673 fallback_mt
= find_suitable_fallback(area
, current_order
,
2674 start_migratetype
, false, &can_steal
);
2675 if (fallback_mt
!= -1)
2680 * This should not happen - we already found a suitable fallback
2681 * when looking for the largest page.
2683 VM_BUG_ON(current_order
== MAX_ORDER
);
2686 page
= get_page_from_free_area(area
, fallback_mt
);
2688 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2691 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2692 start_migratetype
, fallback_mt
);
2699 * Do the hard work of removing an element from the buddy allocator.
2700 * Call me with the zone->lock already held.
2702 static __always_inline
struct page
*
2703 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2704 unsigned int alloc_flags
)
2709 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2710 if (unlikely(!page
)) {
2711 if (migratetype
== MIGRATE_MOVABLE
)
2712 page
= __rmqueue_cma_fallback(zone
, order
);
2714 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2719 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2724 * Obtain a specified number of elements from the buddy allocator, all under
2725 * a single hold of the lock, for efficiency. Add them to the supplied list.
2726 * Returns the number of new pages which were placed at *list.
2728 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2729 unsigned long count
, struct list_head
*list
,
2730 int migratetype
, unsigned int alloc_flags
)
2734 spin_lock(&zone
->lock
);
2735 for (i
= 0; i
< count
; ++i
) {
2736 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2738 if (unlikely(page
== NULL
))
2741 if (unlikely(check_pcp_refill(page
)))
2745 * Split buddy pages returned by expand() are received here in
2746 * physical page order. The page is added to the tail of
2747 * caller's list. From the callers perspective, the linked list
2748 * is ordered by page number under some conditions. This is
2749 * useful for IO devices that can forward direction from the
2750 * head, thus also in the physical page order. This is useful
2751 * for IO devices that can merge IO requests if the physical
2752 * pages are ordered properly.
2754 list_add_tail(&page
->lru
, list
);
2756 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2757 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2762 * i pages were removed from the buddy list even if some leak due
2763 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2764 * on i. Do not confuse with 'alloced' which is the number of
2765 * pages added to the pcp list.
2767 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2768 spin_unlock(&zone
->lock
);
2774 * Called from the vmstat counter updater to drain pagesets of this
2775 * currently executing processor on remote nodes after they have
2778 * Note that this function must be called with the thread pinned to
2779 * a single processor.
2781 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2783 unsigned long flags
;
2784 int to_drain
, batch
;
2786 local_irq_save(flags
);
2787 batch
= READ_ONCE(pcp
->batch
);
2788 to_drain
= min(pcp
->count
, batch
);
2790 free_pcppages_bulk(zone
, to_drain
, pcp
);
2791 local_irq_restore(flags
);
2796 * Drain pcplists of the indicated processor and zone.
2798 * The processor must either be the current processor and the
2799 * thread pinned to the current processor or a processor that
2802 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2804 unsigned long flags
;
2805 struct per_cpu_pageset
*pset
;
2806 struct per_cpu_pages
*pcp
;
2808 local_irq_save(flags
);
2809 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2813 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2814 local_irq_restore(flags
);
2818 * Drain pcplists of all zones on the indicated processor.
2820 * The processor must either be the current processor and the
2821 * thread pinned to the current processor or a processor that
2824 static void drain_pages(unsigned int cpu
)
2828 for_each_populated_zone(zone
) {
2829 drain_pages_zone(cpu
, zone
);
2834 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2836 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2837 * the single zone's pages.
2839 void drain_local_pages(struct zone
*zone
)
2841 int cpu
= smp_processor_id();
2844 drain_pages_zone(cpu
, zone
);
2849 static void drain_local_pages_wq(struct work_struct
*work
)
2851 struct pcpu_drain
*drain
;
2853 drain
= container_of(work
, struct pcpu_drain
, work
);
2856 * drain_all_pages doesn't use proper cpu hotplug protection so
2857 * we can race with cpu offline when the WQ can move this from
2858 * a cpu pinned worker to an unbound one. We can operate on a different
2859 * cpu which is allright but we also have to make sure to not move to
2863 drain_local_pages(drain
->zone
);
2868 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2870 * When zone parameter is non-NULL, spill just the single zone's pages.
2872 * Note that this can be extremely slow as the draining happens in a workqueue.
2874 void drain_all_pages(struct zone
*zone
)
2879 * Allocate in the BSS so we wont require allocation in
2880 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2882 static cpumask_t cpus_with_pcps
;
2885 * Make sure nobody triggers this path before mm_percpu_wq is fully
2888 if (WARN_ON_ONCE(!mm_percpu_wq
))
2892 * Do not drain if one is already in progress unless it's specific to
2893 * a zone. Such callers are primarily CMA and memory hotplug and need
2894 * the drain to be complete when the call returns.
2896 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2899 mutex_lock(&pcpu_drain_mutex
);
2903 * We don't care about racing with CPU hotplug event
2904 * as offline notification will cause the notified
2905 * cpu to drain that CPU pcps and on_each_cpu_mask
2906 * disables preemption as part of its processing
2908 for_each_online_cpu(cpu
) {
2909 struct per_cpu_pageset
*pcp
;
2911 bool has_pcps
= false;
2914 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2918 for_each_populated_zone(z
) {
2919 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2920 if (pcp
->pcp
.count
) {
2928 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2930 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2933 for_each_cpu(cpu
, &cpus_with_pcps
) {
2934 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
2937 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
2938 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
2940 for_each_cpu(cpu
, &cpus_with_pcps
)
2941 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
2943 mutex_unlock(&pcpu_drain_mutex
);
2946 #ifdef CONFIG_HIBERNATION
2949 * Touch the watchdog for every WD_PAGE_COUNT pages.
2951 #define WD_PAGE_COUNT (128*1024)
2953 void mark_free_pages(struct zone
*zone
)
2955 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2956 unsigned long flags
;
2957 unsigned int order
, t
;
2960 if (zone_is_empty(zone
))
2963 spin_lock_irqsave(&zone
->lock
, flags
);
2965 max_zone_pfn
= zone_end_pfn(zone
);
2966 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2967 if (pfn_valid(pfn
)) {
2968 page
= pfn_to_page(pfn
);
2970 if (!--page_count
) {
2971 touch_nmi_watchdog();
2972 page_count
= WD_PAGE_COUNT
;
2975 if (page_zone(page
) != zone
)
2978 if (!swsusp_page_is_forbidden(page
))
2979 swsusp_unset_page_free(page
);
2982 for_each_migratetype_order(order
, t
) {
2983 list_for_each_entry(page
,
2984 &zone
->free_area
[order
].free_list
[t
], lru
) {
2987 pfn
= page_to_pfn(page
);
2988 for (i
= 0; i
< (1UL << order
); i
++) {
2989 if (!--page_count
) {
2990 touch_nmi_watchdog();
2991 page_count
= WD_PAGE_COUNT
;
2993 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2997 spin_unlock_irqrestore(&zone
->lock
, flags
);
2999 #endif /* CONFIG_PM */
3001 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
3005 if (!free_pcp_prepare(page
))
3008 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3009 set_pcppage_migratetype(page
, migratetype
);
3013 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
3015 struct zone
*zone
= page_zone(page
);
3016 struct per_cpu_pages
*pcp
;
3019 migratetype
= get_pcppage_migratetype(page
);
3020 __count_vm_event(PGFREE
);
3023 * We only track unmovable, reclaimable and movable on pcp lists.
3024 * Free ISOLATE pages back to the allocator because they are being
3025 * offlined but treat HIGHATOMIC as movable pages so we can get those
3026 * areas back if necessary. Otherwise, we may have to free
3027 * excessively into the page allocator
3029 if (migratetype
>= MIGRATE_PCPTYPES
) {
3030 if (unlikely(is_migrate_isolate(migratetype
))) {
3031 free_one_page(zone
, page
, pfn
, 0, migratetype
);
3034 migratetype
= MIGRATE_MOVABLE
;
3037 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3038 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
3040 if (pcp
->count
>= pcp
->high
) {
3041 unsigned long batch
= READ_ONCE(pcp
->batch
);
3042 free_pcppages_bulk(zone
, batch
, pcp
);
3047 * Free a 0-order page
3049 void free_unref_page(struct page
*page
)
3051 unsigned long flags
;
3052 unsigned long pfn
= page_to_pfn(page
);
3054 if (!free_unref_page_prepare(page
, pfn
))
3057 local_irq_save(flags
);
3058 free_unref_page_commit(page
, pfn
);
3059 local_irq_restore(flags
);
3063 * Free a list of 0-order pages
3065 void free_unref_page_list(struct list_head
*list
)
3067 struct page
*page
, *next
;
3068 unsigned long flags
, pfn
;
3069 int batch_count
= 0;
3071 /* Prepare pages for freeing */
3072 list_for_each_entry_safe(page
, next
, list
, lru
) {
3073 pfn
= page_to_pfn(page
);
3074 if (!free_unref_page_prepare(page
, pfn
))
3075 list_del(&page
->lru
);
3076 set_page_private(page
, pfn
);
3079 local_irq_save(flags
);
3080 list_for_each_entry_safe(page
, next
, list
, lru
) {
3081 unsigned long pfn
= page_private(page
);
3083 set_page_private(page
, 0);
3084 trace_mm_page_free_batched(page
);
3085 free_unref_page_commit(page
, pfn
);
3088 * Guard against excessive IRQ disabled times when we get
3089 * a large list of pages to free.
3091 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3092 local_irq_restore(flags
);
3094 local_irq_save(flags
);
3097 local_irq_restore(flags
);
3101 * split_page takes a non-compound higher-order page, and splits it into
3102 * n (1<<order) sub-pages: page[0..n]
3103 * Each sub-page must be freed individually.
3105 * Note: this is probably too low level an operation for use in drivers.
3106 * Please consult with lkml before using this in your driver.
3108 void split_page(struct page
*page
, unsigned int order
)
3112 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3113 VM_BUG_ON_PAGE(!page_count(page
), page
);
3115 for (i
= 1; i
< (1 << order
); i
++)
3116 set_page_refcounted(page
+ i
);
3117 split_page_owner(page
, order
);
3119 EXPORT_SYMBOL_GPL(split_page
);
3121 int __isolate_free_page(struct page
*page
, unsigned int order
)
3123 struct free_area
*area
= &page_zone(page
)->free_area
[order
];
3124 unsigned long watermark
;
3128 BUG_ON(!PageBuddy(page
));
3130 zone
= page_zone(page
);
3131 mt
= get_pageblock_migratetype(page
);
3133 if (!is_migrate_isolate(mt
)) {
3135 * Obey watermarks as if the page was being allocated. We can
3136 * emulate a high-order watermark check with a raised order-0
3137 * watermark, because we already know our high-order page
3140 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3141 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3144 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3147 /* Remove page from free list */
3149 del_page_from_free_area(page
, area
);
3152 * Set the pageblock if the isolated page is at least half of a
3155 if (order
>= pageblock_order
- 1) {
3156 struct page
*endpage
= page
+ (1 << order
) - 1;
3157 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3158 int mt
= get_pageblock_migratetype(page
);
3159 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3160 && !is_migrate_highatomic(mt
))
3161 set_pageblock_migratetype(page
,
3167 return 1UL << order
;
3171 * Update NUMA hit/miss statistics
3173 * Must be called with interrupts disabled.
3175 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
3178 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3180 /* skip numa counters update if numa stats is disabled */
3181 if (!static_branch_likely(&vm_numa_stat_key
))
3184 if (zone_to_nid(z
) != numa_node_id())
3185 local_stat
= NUMA_OTHER
;
3187 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3188 __inc_numa_state(z
, NUMA_HIT
);
3190 __inc_numa_state(z
, NUMA_MISS
);
3191 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3193 __inc_numa_state(z
, local_stat
);
3197 /* Remove page from the per-cpu list, caller must protect the list */
3198 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3199 unsigned int alloc_flags
,
3200 struct per_cpu_pages
*pcp
,
3201 struct list_head
*list
)
3206 if (list_empty(list
)) {
3207 pcp
->count
+= rmqueue_bulk(zone
, 0,
3209 migratetype
, alloc_flags
);
3210 if (unlikely(list_empty(list
)))
3214 page
= list_first_entry(list
, struct page
, lru
);
3215 list_del(&page
->lru
);
3217 } while (check_new_pcp(page
));
3222 /* Lock and remove page from the per-cpu list */
3223 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3224 struct zone
*zone
, gfp_t gfp_flags
,
3225 int migratetype
, unsigned int alloc_flags
)
3227 struct per_cpu_pages
*pcp
;
3228 struct list_head
*list
;
3230 unsigned long flags
;
3232 local_irq_save(flags
);
3233 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3234 list
= &pcp
->lists
[migratetype
];
3235 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3237 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3238 zone_statistics(preferred_zone
, zone
);
3240 local_irq_restore(flags
);
3245 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3248 struct page
*rmqueue(struct zone
*preferred_zone
,
3249 struct zone
*zone
, unsigned int order
,
3250 gfp_t gfp_flags
, unsigned int alloc_flags
,
3253 unsigned long flags
;
3256 if (likely(order
== 0)) {
3257 page
= rmqueue_pcplist(preferred_zone
, zone
, gfp_flags
,
3258 migratetype
, alloc_flags
);
3263 * We most definitely don't want callers attempting to
3264 * allocate greater than order-1 page units with __GFP_NOFAIL.
3266 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3267 spin_lock_irqsave(&zone
->lock
, flags
);
3271 if (alloc_flags
& ALLOC_HARDER
) {
3272 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3274 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3277 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3278 } while (page
&& check_new_pages(page
, order
));
3279 spin_unlock(&zone
->lock
);
3282 __mod_zone_freepage_state(zone
, -(1 << order
),
3283 get_pcppage_migratetype(page
));
3285 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3286 zone_statistics(preferred_zone
, zone
);
3287 local_irq_restore(flags
);
3290 /* Separate test+clear to avoid unnecessary atomics */
3291 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3292 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3293 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3296 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3300 local_irq_restore(flags
);
3304 #ifdef CONFIG_FAIL_PAGE_ALLOC
3307 struct fault_attr attr
;
3309 bool ignore_gfp_highmem
;
3310 bool ignore_gfp_reclaim
;
3312 } fail_page_alloc
= {
3313 .attr
= FAULT_ATTR_INITIALIZER
,
3314 .ignore_gfp_reclaim
= true,
3315 .ignore_gfp_highmem
= true,
3319 static int __init
setup_fail_page_alloc(char *str
)
3321 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3323 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3325 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3327 if (order
< fail_page_alloc
.min_order
)
3329 if (gfp_mask
& __GFP_NOFAIL
)
3331 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3333 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3334 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3337 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3340 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3342 static int __init
fail_page_alloc_debugfs(void)
3344 umode_t mode
= S_IFREG
| 0600;
3347 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3348 &fail_page_alloc
.attr
);
3350 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3351 &fail_page_alloc
.ignore_gfp_reclaim
);
3352 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3353 &fail_page_alloc
.ignore_gfp_highmem
);
3354 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3359 late_initcall(fail_page_alloc_debugfs
);
3361 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3363 #else /* CONFIG_FAIL_PAGE_ALLOC */
3365 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3370 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3372 static noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3374 return __should_fail_alloc_page(gfp_mask
, order
);
3376 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3379 * Return true if free base pages are above 'mark'. For high-order checks it
3380 * will return true of the order-0 watermark is reached and there is at least
3381 * one free page of a suitable size. Checking now avoids taking the zone lock
3382 * to check in the allocation paths if no pages are free.
3384 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3385 int classzone_idx
, unsigned int alloc_flags
,
3390 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3392 /* free_pages may go negative - that's OK */
3393 free_pages
-= (1 << order
) - 1;
3395 if (alloc_flags
& ALLOC_HIGH
)
3399 * If the caller does not have rights to ALLOC_HARDER then subtract
3400 * the high-atomic reserves. This will over-estimate the size of the
3401 * atomic reserve but it avoids a search.
3403 if (likely(!alloc_harder
)) {
3404 free_pages
-= z
->nr_reserved_highatomic
;
3407 * OOM victims can try even harder than normal ALLOC_HARDER
3408 * users on the grounds that it's definitely going to be in
3409 * the exit path shortly and free memory. Any allocation it
3410 * makes during the free path will be small and short-lived.
3412 if (alloc_flags
& ALLOC_OOM
)
3420 /* If allocation can't use CMA areas don't use free CMA pages */
3421 if (!(alloc_flags
& ALLOC_CMA
))
3422 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3426 * Check watermarks for an order-0 allocation request. If these
3427 * are not met, then a high-order request also cannot go ahead
3428 * even if a suitable page happened to be free.
3430 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3433 /* If this is an order-0 request then the watermark is fine */
3437 /* For a high-order request, check at least one suitable page is free */
3438 for (o
= order
; o
< MAX_ORDER
; o
++) {
3439 struct free_area
*area
= &z
->free_area
[o
];
3445 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3446 if (!free_area_empty(area
, mt
))
3451 if ((alloc_flags
& ALLOC_CMA
) &&
3452 !free_area_empty(area
, MIGRATE_CMA
)) {
3456 if (alloc_harder
&& !free_area_empty(area
, MIGRATE_HIGHATOMIC
))
3462 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3463 int classzone_idx
, unsigned int alloc_flags
)
3465 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3466 zone_page_state(z
, NR_FREE_PAGES
));
3469 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3470 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3472 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3476 /* If allocation can't use CMA areas don't use free CMA pages */
3477 if (!(alloc_flags
& ALLOC_CMA
))
3478 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3482 * Fast check for order-0 only. If this fails then the reserves
3483 * need to be calculated. There is a corner case where the check
3484 * passes but only the high-order atomic reserve are free. If
3485 * the caller is !atomic then it'll uselessly search the free
3486 * list. That corner case is then slower but it is harmless.
3488 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3491 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3495 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3496 unsigned long mark
, int classzone_idx
)
3498 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3500 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3501 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3503 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3508 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3510 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3511 node_reclaim_distance
;
3513 #else /* CONFIG_NUMA */
3514 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3518 #endif /* CONFIG_NUMA */
3521 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3522 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3523 * premature use of a lower zone may cause lowmem pressure problems that
3524 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3525 * probably too small. It only makes sense to spread allocations to avoid
3526 * fragmentation between the Normal and DMA32 zones.
3528 static inline unsigned int
3529 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3531 unsigned int alloc_flags
;
3534 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3537 alloc_flags
= (__force
int) (gfp_mask
& __GFP_KSWAPD_RECLAIM
);
3539 #ifdef CONFIG_ZONE_DMA32
3543 if (zone_idx(zone
) != ZONE_NORMAL
)
3547 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3548 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3549 * on UMA that if Normal is populated then so is DMA32.
3551 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3552 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3555 alloc_flags
|= ALLOC_NOFRAGMENT
;
3556 #endif /* CONFIG_ZONE_DMA32 */
3561 * get_page_from_freelist goes through the zonelist trying to allocate
3564 static struct page
*
3565 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3566 const struct alloc_context
*ac
)
3570 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3575 * Scan zonelist, looking for a zone with enough free.
3576 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3578 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3579 z
= ac
->preferred_zoneref
;
3580 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3585 if (cpusets_enabled() &&
3586 (alloc_flags
& ALLOC_CPUSET
) &&
3587 !__cpuset_zone_allowed(zone
, gfp_mask
))
3590 * When allocating a page cache page for writing, we
3591 * want to get it from a node that is within its dirty
3592 * limit, such that no single node holds more than its
3593 * proportional share of globally allowed dirty pages.
3594 * The dirty limits take into account the node's
3595 * lowmem reserves and high watermark so that kswapd
3596 * should be able to balance it without having to
3597 * write pages from its LRU list.
3599 * XXX: For now, allow allocations to potentially
3600 * exceed the per-node dirty limit in the slowpath
3601 * (spread_dirty_pages unset) before going into reclaim,
3602 * which is important when on a NUMA setup the allowed
3603 * nodes are together not big enough to reach the
3604 * global limit. The proper fix for these situations
3605 * will require awareness of nodes in the
3606 * dirty-throttling and the flusher threads.
3608 if (ac
->spread_dirty_pages
) {
3609 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3612 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3613 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3618 if (no_fallback
&& nr_online_nodes
> 1 &&
3619 zone
!= ac
->preferred_zoneref
->zone
) {
3623 * If moving to a remote node, retry but allow
3624 * fragmenting fallbacks. Locality is more important
3625 * than fragmentation avoidance.
3627 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3628 if (zone_to_nid(zone
) != local_nid
) {
3629 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3634 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3635 if (!zone_watermark_fast(zone
, order
, mark
,
3636 ac_classzone_idx(ac
), alloc_flags
)) {
3639 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3641 * Watermark failed for this zone, but see if we can
3642 * grow this zone if it contains deferred pages.
3644 if (static_branch_unlikely(&deferred_pages
)) {
3645 if (_deferred_grow_zone(zone
, order
))
3649 /* Checked here to keep the fast path fast */
3650 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3651 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3654 if (node_reclaim_mode
== 0 ||
3655 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3658 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3660 case NODE_RECLAIM_NOSCAN
:
3663 case NODE_RECLAIM_FULL
:
3664 /* scanned but unreclaimable */
3667 /* did we reclaim enough */
3668 if (zone_watermark_ok(zone
, order
, mark
,
3669 ac_classzone_idx(ac
), alloc_flags
))
3677 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3678 gfp_mask
, alloc_flags
, ac
->migratetype
);
3680 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3683 * If this is a high-order atomic allocation then check
3684 * if the pageblock should be reserved for the future
3686 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3687 reserve_highatomic_pageblock(page
, zone
, order
);
3691 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3692 /* Try again if zone has deferred pages */
3693 if (static_branch_unlikely(&deferred_pages
)) {
3694 if (_deferred_grow_zone(zone
, order
))
3702 * It's possible on a UMA machine to get through all zones that are
3703 * fragmented. If avoiding fragmentation, reset and try again.
3706 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3713 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3715 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3718 * This documents exceptions given to allocations in certain
3719 * contexts that are allowed to allocate outside current's set
3722 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3723 if (tsk_is_oom_victim(current
) ||
3724 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3725 filter
&= ~SHOW_MEM_FILTER_NODES
;
3726 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3727 filter
&= ~SHOW_MEM_FILTER_NODES
;
3729 show_mem(filter
, nodemask
);
3732 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3734 struct va_format vaf
;
3736 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
3738 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3741 va_start(args
, fmt
);
3744 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3745 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3746 nodemask_pr_args(nodemask
));
3749 cpuset_print_current_mems_allowed();
3752 warn_alloc_show_mem(gfp_mask
, nodemask
);
3755 static inline struct page
*
3756 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3757 unsigned int alloc_flags
,
3758 const struct alloc_context
*ac
)
3762 page
= get_page_from_freelist(gfp_mask
, order
,
3763 alloc_flags
|ALLOC_CPUSET
, ac
);
3765 * fallback to ignore cpuset restriction if our nodes
3769 page
= get_page_from_freelist(gfp_mask
, order
,
3775 static inline struct page
*
3776 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3777 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3779 struct oom_control oc
= {
3780 .zonelist
= ac
->zonelist
,
3781 .nodemask
= ac
->nodemask
,
3783 .gfp_mask
= gfp_mask
,
3788 *did_some_progress
= 0;
3791 * Acquire the oom lock. If that fails, somebody else is
3792 * making progress for us.
3794 if (!mutex_trylock(&oom_lock
)) {
3795 *did_some_progress
= 1;
3796 schedule_timeout_uninterruptible(1);
3801 * Go through the zonelist yet one more time, keep very high watermark
3802 * here, this is only to catch a parallel oom killing, we must fail if
3803 * we're still under heavy pressure. But make sure that this reclaim
3804 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3805 * allocation which will never fail due to oom_lock already held.
3807 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3808 ~__GFP_DIRECT_RECLAIM
, order
,
3809 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3813 /* Coredumps can quickly deplete all memory reserves */
3814 if (current
->flags
& PF_DUMPCORE
)
3816 /* The OOM killer will not help higher order allocs */
3817 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3820 * We have already exhausted all our reclaim opportunities without any
3821 * success so it is time to admit defeat. We will skip the OOM killer
3822 * because it is very likely that the caller has a more reasonable
3823 * fallback than shooting a random task.
3825 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3827 /* The OOM killer does not needlessly kill tasks for lowmem */
3828 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3830 if (pm_suspended_storage())
3833 * XXX: GFP_NOFS allocations should rather fail than rely on
3834 * other request to make a forward progress.
3835 * We are in an unfortunate situation where out_of_memory cannot
3836 * do much for this context but let's try it to at least get
3837 * access to memory reserved if the current task is killed (see
3838 * out_of_memory). Once filesystems are ready to handle allocation
3839 * failures more gracefully we should just bail out here.
3842 /* The OOM killer may not free memory on a specific node */
3843 if (gfp_mask
& __GFP_THISNODE
)
3846 /* Exhausted what can be done so it's blame time */
3847 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3848 *did_some_progress
= 1;
3851 * Help non-failing allocations by giving them access to memory
3854 if (gfp_mask
& __GFP_NOFAIL
)
3855 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3856 ALLOC_NO_WATERMARKS
, ac
);
3859 mutex_unlock(&oom_lock
);
3864 * Maximum number of compaction retries wit a progress before OOM
3865 * killer is consider as the only way to move forward.
3867 #define MAX_COMPACT_RETRIES 16
3869 #ifdef CONFIG_COMPACTION
3870 /* Try memory compaction for high-order allocations before reclaim */
3871 static struct page
*
3872 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3873 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3874 enum compact_priority prio
, enum compact_result
*compact_result
)
3876 struct page
*page
= NULL
;
3877 unsigned long pflags
;
3878 unsigned int noreclaim_flag
;
3883 psi_memstall_enter(&pflags
);
3884 noreclaim_flag
= memalloc_noreclaim_save();
3886 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3889 memalloc_noreclaim_restore(noreclaim_flag
);
3890 psi_memstall_leave(&pflags
);
3893 * At least in one zone compaction wasn't deferred or skipped, so let's
3894 * count a compaction stall
3896 count_vm_event(COMPACTSTALL
);
3898 /* Prep a captured page if available */
3900 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3902 /* Try get a page from the freelist if available */
3904 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3907 struct zone
*zone
= page_zone(page
);
3909 zone
->compact_blockskip_flush
= false;
3910 compaction_defer_reset(zone
, order
, true);
3911 count_vm_event(COMPACTSUCCESS
);
3916 * It's bad if compaction run occurs and fails. The most likely reason
3917 * is that pages exist, but not enough to satisfy watermarks.
3919 count_vm_event(COMPACTFAIL
);
3927 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3928 enum compact_result compact_result
,
3929 enum compact_priority
*compact_priority
,
3930 int *compaction_retries
)
3932 int max_retries
= MAX_COMPACT_RETRIES
;
3935 int retries
= *compaction_retries
;
3936 enum compact_priority priority
= *compact_priority
;
3941 if (compaction_made_progress(compact_result
))
3942 (*compaction_retries
)++;
3945 * compaction considers all the zone as desperately out of memory
3946 * so it doesn't really make much sense to retry except when the
3947 * failure could be caused by insufficient priority
3949 if (compaction_failed(compact_result
))
3950 goto check_priority
;
3953 * compaction was skipped because there are not enough order-0 pages
3954 * to work with, so we retry only if it looks like reclaim can help.
3956 if (compaction_needs_reclaim(compact_result
)) {
3957 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3962 * make sure the compaction wasn't deferred or didn't bail out early
3963 * due to locks contention before we declare that we should give up.
3964 * But the next retry should use a higher priority if allowed, so
3965 * we don't just keep bailing out endlessly.
3967 if (compaction_withdrawn(compact_result
)) {
3968 goto check_priority
;
3972 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3973 * costly ones because they are de facto nofail and invoke OOM
3974 * killer to move on while costly can fail and users are ready
3975 * to cope with that. 1/4 retries is rather arbitrary but we
3976 * would need much more detailed feedback from compaction to
3977 * make a better decision.
3979 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3981 if (*compaction_retries
<= max_retries
) {
3987 * Make sure there are attempts at the highest priority if we exhausted
3988 * all retries or failed at the lower priorities.
3991 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3992 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3994 if (*compact_priority
> min_priority
) {
3995 (*compact_priority
)--;
3996 *compaction_retries
= 0;
4000 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4004 static inline struct page
*
4005 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4006 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4007 enum compact_priority prio
, enum compact_result
*compact_result
)
4009 *compact_result
= COMPACT_SKIPPED
;
4014 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4015 enum compact_result compact_result
,
4016 enum compact_priority
*compact_priority
,
4017 int *compaction_retries
)
4022 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4026 * There are setups with compaction disabled which would prefer to loop
4027 * inside the allocator rather than hit the oom killer prematurely.
4028 * Let's give them a good hope and keep retrying while the order-0
4029 * watermarks are OK.
4031 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4033 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4034 ac_classzone_idx(ac
), alloc_flags
))
4039 #endif /* CONFIG_COMPACTION */
4041 #ifdef CONFIG_LOCKDEP
4042 static struct lockdep_map __fs_reclaim_map
=
4043 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4045 static bool __need_fs_reclaim(gfp_t gfp_mask
)
4047 gfp_mask
= current_gfp_context(gfp_mask
);
4049 /* no reclaim without waiting on it */
4050 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4053 /* this guy won't enter reclaim */
4054 if (current
->flags
& PF_MEMALLOC
)
4057 /* We're only interested __GFP_FS allocations for now */
4058 if (!(gfp_mask
& __GFP_FS
))
4061 if (gfp_mask
& __GFP_NOLOCKDEP
)
4067 void __fs_reclaim_acquire(void)
4069 lock_map_acquire(&__fs_reclaim_map
);
4072 void __fs_reclaim_release(void)
4074 lock_map_release(&__fs_reclaim_map
);
4077 void fs_reclaim_acquire(gfp_t gfp_mask
)
4079 if (__need_fs_reclaim(gfp_mask
))
4080 __fs_reclaim_acquire();
4082 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4084 void fs_reclaim_release(gfp_t gfp_mask
)
4086 if (__need_fs_reclaim(gfp_mask
))
4087 __fs_reclaim_release();
4089 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4092 /* Perform direct synchronous page reclaim */
4094 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4095 const struct alloc_context
*ac
)
4098 unsigned int noreclaim_flag
;
4099 unsigned long pflags
;
4103 /* We now go into synchronous reclaim */
4104 cpuset_memory_pressure_bump();
4105 psi_memstall_enter(&pflags
);
4106 fs_reclaim_acquire(gfp_mask
);
4107 noreclaim_flag
= memalloc_noreclaim_save();
4109 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4112 memalloc_noreclaim_restore(noreclaim_flag
);
4113 fs_reclaim_release(gfp_mask
);
4114 psi_memstall_leave(&pflags
);
4121 /* The really slow allocator path where we enter direct reclaim */
4122 static inline struct page
*
4123 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4124 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4125 unsigned long *did_some_progress
)
4127 struct page
*page
= NULL
;
4128 bool drained
= false;
4130 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4131 if (unlikely(!(*did_some_progress
)))
4135 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4138 * If an allocation failed after direct reclaim, it could be because
4139 * pages are pinned on the per-cpu lists or in high alloc reserves.
4140 * Shrink them them and try again
4142 if (!page
&& !drained
) {
4143 unreserve_highatomic_pageblock(ac
, false);
4144 drain_all_pages(NULL
);
4152 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4153 const struct alloc_context
*ac
)
4157 pg_data_t
*last_pgdat
= NULL
;
4158 enum zone_type high_zoneidx
= ac
->high_zoneidx
;
4160 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, high_zoneidx
,
4162 if (last_pgdat
!= zone
->zone_pgdat
)
4163 wakeup_kswapd(zone
, gfp_mask
, order
, high_zoneidx
);
4164 last_pgdat
= zone
->zone_pgdat
;
4168 static inline unsigned int
4169 gfp_to_alloc_flags(gfp_t gfp_mask
)
4171 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4174 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4175 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4176 * to save two branches.
4178 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4179 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM
!= (__force gfp_t
) ALLOC_KSWAPD
);
4182 * The caller may dip into page reserves a bit more if the caller
4183 * cannot run direct reclaim, or if the caller has realtime scheduling
4184 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4185 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4187 alloc_flags
|= (__force
int)
4188 (gfp_mask
& (__GFP_HIGH
| __GFP_KSWAPD_RECLAIM
));
4190 if (gfp_mask
& __GFP_ATOMIC
) {
4192 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4193 * if it can't schedule.
4195 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4196 alloc_flags
|= ALLOC_HARDER
;
4198 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4199 * comment for __cpuset_node_allowed().
4201 alloc_flags
&= ~ALLOC_CPUSET
;
4202 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4203 alloc_flags
|= ALLOC_HARDER
;
4206 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
4207 alloc_flags
|= ALLOC_CMA
;
4212 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4214 if (!tsk_is_oom_victim(tsk
))
4218 * !MMU doesn't have oom reaper so give access to memory reserves
4219 * only to the thread with TIF_MEMDIE set
4221 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4228 * Distinguish requests which really need access to full memory
4229 * reserves from oom victims which can live with a portion of it
4231 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4233 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4235 if (gfp_mask
& __GFP_MEMALLOC
)
4236 return ALLOC_NO_WATERMARKS
;
4237 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4238 return ALLOC_NO_WATERMARKS
;
4239 if (!in_interrupt()) {
4240 if (current
->flags
& PF_MEMALLOC
)
4241 return ALLOC_NO_WATERMARKS
;
4242 else if (oom_reserves_allowed(current
))
4249 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4251 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4255 * Checks whether it makes sense to retry the reclaim to make a forward progress
4256 * for the given allocation request.
4258 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4259 * without success, or when we couldn't even meet the watermark if we
4260 * reclaimed all remaining pages on the LRU lists.
4262 * Returns true if a retry is viable or false to enter the oom path.
4265 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4266 struct alloc_context
*ac
, int alloc_flags
,
4267 bool did_some_progress
, int *no_progress_loops
)
4274 * Costly allocations might have made a progress but this doesn't mean
4275 * their order will become available due to high fragmentation so
4276 * always increment the no progress counter for them
4278 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4279 *no_progress_loops
= 0;
4281 (*no_progress_loops
)++;
4284 * Make sure we converge to OOM if we cannot make any progress
4285 * several times in the row.
4287 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4288 /* Before OOM, exhaust highatomic_reserve */
4289 return unreserve_highatomic_pageblock(ac
, true);
4293 * Keep reclaiming pages while there is a chance this will lead
4294 * somewhere. If none of the target zones can satisfy our allocation
4295 * request even if all reclaimable pages are considered then we are
4296 * screwed and have to go OOM.
4298 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4300 unsigned long available
;
4301 unsigned long reclaimable
;
4302 unsigned long min_wmark
= min_wmark_pages(zone
);
4305 available
= reclaimable
= zone_reclaimable_pages(zone
);
4306 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4309 * Would the allocation succeed if we reclaimed all
4310 * reclaimable pages?
4312 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4313 ac_classzone_idx(ac
), alloc_flags
, available
);
4314 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4315 available
, min_wmark
, *no_progress_loops
, wmark
);
4318 * If we didn't make any progress and have a lot of
4319 * dirty + writeback pages then we should wait for
4320 * an IO to complete to slow down the reclaim and
4321 * prevent from pre mature OOM
4323 if (!did_some_progress
) {
4324 unsigned long write_pending
;
4326 write_pending
= zone_page_state_snapshot(zone
,
4327 NR_ZONE_WRITE_PENDING
);
4329 if (2 * write_pending
> reclaimable
) {
4330 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4342 * Memory allocation/reclaim might be called from a WQ context and the
4343 * current implementation of the WQ concurrency control doesn't
4344 * recognize that a particular WQ is congested if the worker thread is
4345 * looping without ever sleeping. Therefore we have to do a short sleep
4346 * here rather than calling cond_resched().
4348 if (current
->flags
& PF_WQ_WORKER
)
4349 schedule_timeout_uninterruptible(1);
4356 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4359 * It's possible that cpuset's mems_allowed and the nodemask from
4360 * mempolicy don't intersect. This should be normally dealt with by
4361 * policy_nodemask(), but it's possible to race with cpuset update in
4362 * such a way the check therein was true, and then it became false
4363 * before we got our cpuset_mems_cookie here.
4364 * This assumes that for all allocations, ac->nodemask can come only
4365 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4366 * when it does not intersect with the cpuset restrictions) or the
4367 * caller can deal with a violated nodemask.
4369 if (cpusets_enabled() && ac
->nodemask
&&
4370 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4371 ac
->nodemask
= NULL
;
4376 * When updating a task's mems_allowed or mempolicy nodemask, it is
4377 * possible to race with parallel threads in such a way that our
4378 * allocation can fail while the mask is being updated. If we are about
4379 * to fail, check if the cpuset changed during allocation and if so,
4382 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4388 static inline struct page
*
4389 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4390 struct alloc_context
*ac
)
4392 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4393 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4394 struct page
*page
= NULL
;
4395 unsigned int alloc_flags
;
4396 unsigned long did_some_progress
;
4397 enum compact_priority compact_priority
;
4398 enum compact_result compact_result
;
4399 int compaction_retries
;
4400 int no_progress_loops
;
4401 unsigned int cpuset_mems_cookie
;
4405 * We also sanity check to catch abuse of atomic reserves being used by
4406 * callers that are not in atomic context.
4408 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4409 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4410 gfp_mask
&= ~__GFP_ATOMIC
;
4413 compaction_retries
= 0;
4414 no_progress_loops
= 0;
4415 compact_priority
= DEF_COMPACT_PRIORITY
;
4416 cpuset_mems_cookie
= read_mems_allowed_begin();
4419 * The fast path uses conservative alloc_flags to succeed only until
4420 * kswapd needs to be woken up, and to avoid the cost of setting up
4421 * alloc_flags precisely. So we do that now.
4423 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4426 * We need to recalculate the starting point for the zonelist iterator
4427 * because we might have used different nodemask in the fast path, or
4428 * there was a cpuset modification and we are retrying - otherwise we
4429 * could end up iterating over non-eligible zones endlessly.
4431 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4432 ac
->high_zoneidx
, ac
->nodemask
);
4433 if (!ac
->preferred_zoneref
->zone
)
4436 if (alloc_flags
& ALLOC_KSWAPD
)
4437 wake_all_kswapds(order
, gfp_mask
, ac
);
4440 * The adjusted alloc_flags might result in immediate success, so try
4443 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4448 * For costly allocations, try direct compaction first, as it's likely
4449 * that we have enough base pages and don't need to reclaim. For non-
4450 * movable high-order allocations, do that as well, as compaction will
4451 * try prevent permanent fragmentation by migrating from blocks of the
4453 * Don't try this for allocations that are allowed to ignore
4454 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4456 if (can_direct_reclaim
&&
4458 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4459 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4460 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4462 INIT_COMPACT_PRIORITY
,
4468 * Checks for costly allocations with __GFP_NORETRY, which
4469 * includes some THP page fault allocations
4471 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4473 * If allocating entire pageblock(s) and compaction
4474 * failed because all zones are below low watermarks
4475 * or is prohibited because it recently failed at this
4476 * order, fail immediately unless the allocator has
4477 * requested compaction and reclaim retry.
4480 * - potentially very expensive because zones are far
4481 * below their low watermarks or this is part of very
4482 * bursty high order allocations,
4483 * - not guaranteed to help because isolate_freepages()
4484 * may not iterate over freed pages as part of its
4486 * - unlikely to make entire pageblocks free on its
4489 if (compact_result
== COMPACT_SKIPPED
||
4490 compact_result
== COMPACT_DEFERRED
)
4494 * Looks like reclaim/compaction is worth trying, but
4495 * sync compaction could be very expensive, so keep
4496 * using async compaction.
4498 compact_priority
= INIT_COMPACT_PRIORITY
;
4503 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4504 if (alloc_flags
& ALLOC_KSWAPD
)
4505 wake_all_kswapds(order
, gfp_mask
, ac
);
4507 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4509 alloc_flags
= reserve_flags
;
4512 * Reset the nodemask and zonelist iterators if memory policies can be
4513 * ignored. These allocations are high priority and system rather than
4516 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4517 ac
->nodemask
= NULL
;
4518 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4519 ac
->high_zoneidx
, ac
->nodemask
);
4522 /* Attempt with potentially adjusted zonelist and alloc_flags */
4523 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4527 /* Caller is not willing to reclaim, we can't balance anything */
4528 if (!can_direct_reclaim
)
4531 /* Avoid recursion of direct reclaim */
4532 if (current
->flags
& PF_MEMALLOC
)
4535 /* Try direct reclaim and then allocating */
4536 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4537 &did_some_progress
);
4541 /* Try direct compaction and then allocating */
4542 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4543 compact_priority
, &compact_result
);
4547 /* Do not loop if specifically requested */
4548 if (gfp_mask
& __GFP_NORETRY
)
4552 * Do not retry costly high order allocations unless they are
4553 * __GFP_RETRY_MAYFAIL
4555 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4558 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4559 did_some_progress
> 0, &no_progress_loops
))
4563 * It doesn't make any sense to retry for the compaction if the order-0
4564 * reclaim is not able to make any progress because the current
4565 * implementation of the compaction depends on the sufficient amount
4566 * of free memory (see __compaction_suitable)
4568 if (did_some_progress
> 0 &&
4569 should_compact_retry(ac
, order
, alloc_flags
,
4570 compact_result
, &compact_priority
,
4571 &compaction_retries
))
4575 /* Deal with possible cpuset update races before we start OOM killing */
4576 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4579 /* Reclaim has failed us, start killing things */
4580 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4584 /* Avoid allocations with no watermarks from looping endlessly */
4585 if (tsk_is_oom_victim(current
) &&
4586 (alloc_flags
== ALLOC_OOM
||
4587 (gfp_mask
& __GFP_NOMEMALLOC
)))
4590 /* Retry as long as the OOM killer is making progress */
4591 if (did_some_progress
) {
4592 no_progress_loops
= 0;
4597 /* Deal with possible cpuset update races before we fail */
4598 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4602 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4605 if (gfp_mask
& __GFP_NOFAIL
) {
4607 * All existing users of the __GFP_NOFAIL are blockable, so warn
4608 * of any new users that actually require GFP_NOWAIT
4610 if (WARN_ON_ONCE(!can_direct_reclaim
))
4614 * PF_MEMALLOC request from this context is rather bizarre
4615 * because we cannot reclaim anything and only can loop waiting
4616 * for somebody to do a work for us
4618 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4621 * non failing costly orders are a hard requirement which we
4622 * are not prepared for much so let's warn about these users
4623 * so that we can identify them and convert them to something
4626 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4629 * Help non-failing allocations by giving them access to memory
4630 * reserves but do not use ALLOC_NO_WATERMARKS because this
4631 * could deplete whole memory reserves which would just make
4632 * the situation worse
4634 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4642 warn_alloc(gfp_mask
, ac
->nodemask
,
4643 "page allocation failure: order:%u", order
);
4648 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4649 int preferred_nid
, nodemask_t
*nodemask
,
4650 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4651 unsigned int *alloc_flags
)
4653 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4654 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4655 ac
->nodemask
= nodemask
;
4656 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4658 if (cpusets_enabled()) {
4659 *alloc_mask
|= __GFP_HARDWALL
;
4661 ac
->nodemask
= &cpuset_current_mems_allowed
;
4663 *alloc_flags
|= ALLOC_CPUSET
;
4666 fs_reclaim_acquire(gfp_mask
);
4667 fs_reclaim_release(gfp_mask
);
4669 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4671 if (should_fail_alloc_page(gfp_mask
, order
))
4674 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4675 *alloc_flags
|= ALLOC_CMA
;
4680 /* Determine whether to spread dirty pages and what the first usable zone */
4681 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4683 /* Dirty zone balancing only done in the fast path */
4684 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4687 * The preferred zone is used for statistics but crucially it is
4688 * also used as the starting point for the zonelist iterator. It
4689 * may get reset for allocations that ignore memory policies.
4691 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4692 ac
->high_zoneidx
, ac
->nodemask
);
4696 * This is the 'heart' of the zoned buddy allocator.
4699 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4700 nodemask_t
*nodemask
)
4703 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4704 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4705 struct alloc_context ac
= { };
4708 * There are several places where we assume that the order value is sane
4709 * so bail out early if the request is out of bound.
4711 if (unlikely(order
>= MAX_ORDER
)) {
4712 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4716 gfp_mask
&= gfp_allowed_mask
;
4717 alloc_mask
= gfp_mask
;
4718 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4721 finalise_ac(gfp_mask
, &ac
);
4724 * Forbid the first pass from falling back to types that fragment
4725 * memory until all local zones are considered.
4727 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4729 /* First allocation attempt */
4730 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4735 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4736 * resp. GFP_NOIO which has to be inherited for all allocation requests
4737 * from a particular context which has been marked by
4738 * memalloc_no{fs,io}_{save,restore}.
4740 alloc_mask
= current_gfp_context(gfp_mask
);
4741 ac
.spread_dirty_pages
= false;
4744 * Restore the original nodemask if it was potentially replaced with
4745 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4747 ac
.nodemask
= nodemask
;
4749 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4752 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4753 unlikely(__memcg_kmem_charge_page(page
, gfp_mask
, order
) != 0)) {
4754 __free_pages(page
, order
);
4758 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4762 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4765 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4766 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4767 * you need to access high mem.
4769 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4773 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4776 return (unsigned long) page_address(page
);
4778 EXPORT_SYMBOL(__get_free_pages
);
4780 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4782 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4784 EXPORT_SYMBOL(get_zeroed_page
);
4786 static inline void free_the_page(struct page
*page
, unsigned int order
)
4788 if (order
== 0) /* Via pcp? */
4789 free_unref_page(page
);
4791 __free_pages_ok(page
, order
);
4794 void __free_pages(struct page
*page
, unsigned int order
)
4796 if (put_page_testzero(page
))
4797 free_the_page(page
, order
);
4799 EXPORT_SYMBOL(__free_pages
);
4801 void free_pages(unsigned long addr
, unsigned int order
)
4804 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4805 __free_pages(virt_to_page((void *)addr
), order
);
4809 EXPORT_SYMBOL(free_pages
);
4813 * An arbitrary-length arbitrary-offset area of memory which resides
4814 * within a 0 or higher order page. Multiple fragments within that page
4815 * are individually refcounted, in the page's reference counter.
4817 * The page_frag functions below provide a simple allocation framework for
4818 * page fragments. This is used by the network stack and network device
4819 * drivers to provide a backing region of memory for use as either an
4820 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4822 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4825 struct page
*page
= NULL
;
4826 gfp_t gfp
= gfp_mask
;
4828 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4829 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4831 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4832 PAGE_FRAG_CACHE_MAX_ORDER
);
4833 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4835 if (unlikely(!page
))
4836 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4838 nc
->va
= page
? page_address(page
) : NULL
;
4843 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4845 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4847 if (page_ref_sub_and_test(page
, count
))
4848 free_the_page(page
, compound_order(page
));
4850 EXPORT_SYMBOL(__page_frag_cache_drain
);
4852 void *page_frag_alloc(struct page_frag_cache
*nc
,
4853 unsigned int fragsz
, gfp_t gfp_mask
)
4855 unsigned int size
= PAGE_SIZE
;
4859 if (unlikely(!nc
->va
)) {
4861 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4865 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4866 /* if size can vary use size else just use PAGE_SIZE */
4869 /* Even if we own the page, we do not use atomic_set().
4870 * This would break get_page_unless_zero() users.
4872 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
4874 /* reset page count bias and offset to start of new frag */
4875 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4876 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4880 offset
= nc
->offset
- fragsz
;
4881 if (unlikely(offset
< 0)) {
4882 page
= virt_to_page(nc
->va
);
4884 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4887 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4888 /* if size can vary use size else just use PAGE_SIZE */
4891 /* OK, page count is 0, we can safely set it */
4892 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
4894 /* reset page count bias and offset to start of new frag */
4895 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4896 offset
= size
- fragsz
;
4900 nc
->offset
= offset
;
4902 return nc
->va
+ offset
;
4904 EXPORT_SYMBOL(page_frag_alloc
);
4907 * Frees a page fragment allocated out of either a compound or order 0 page.
4909 void page_frag_free(void *addr
)
4911 struct page
*page
= virt_to_head_page(addr
);
4913 if (unlikely(put_page_testzero(page
)))
4914 free_the_page(page
, compound_order(page
));
4916 EXPORT_SYMBOL(page_frag_free
);
4918 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4922 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4923 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4925 split_page(virt_to_page((void *)addr
), order
);
4926 while (used
< alloc_end
) {
4931 return (void *)addr
;
4935 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4936 * @size: the number of bytes to allocate
4937 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4939 * This function is similar to alloc_pages(), except that it allocates the
4940 * minimum number of pages to satisfy the request. alloc_pages() can only
4941 * allocate memory in power-of-two pages.
4943 * This function is also limited by MAX_ORDER.
4945 * Memory allocated by this function must be released by free_pages_exact().
4947 * Return: pointer to the allocated area or %NULL in case of error.
4949 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4951 unsigned int order
= get_order(size
);
4954 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
4955 gfp_mask
&= ~__GFP_COMP
;
4957 addr
= __get_free_pages(gfp_mask
, order
);
4958 return make_alloc_exact(addr
, order
, size
);
4960 EXPORT_SYMBOL(alloc_pages_exact
);
4963 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4965 * @nid: the preferred node ID where memory should be allocated
4966 * @size: the number of bytes to allocate
4967 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4969 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4972 * Return: pointer to the allocated area or %NULL in case of error.
4974 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4976 unsigned int order
= get_order(size
);
4979 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
4980 gfp_mask
&= ~__GFP_COMP
;
4982 p
= alloc_pages_node(nid
, gfp_mask
, order
);
4985 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4989 * free_pages_exact - release memory allocated via alloc_pages_exact()
4990 * @virt: the value returned by alloc_pages_exact.
4991 * @size: size of allocation, same value as passed to alloc_pages_exact().
4993 * Release the memory allocated by a previous call to alloc_pages_exact.
4995 void free_pages_exact(void *virt
, size_t size
)
4997 unsigned long addr
= (unsigned long)virt
;
4998 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5000 while (addr
< end
) {
5005 EXPORT_SYMBOL(free_pages_exact
);
5008 * nr_free_zone_pages - count number of pages beyond high watermark
5009 * @offset: The zone index of the highest zone
5011 * nr_free_zone_pages() counts the number of pages which are beyond the
5012 * high watermark within all zones at or below a given zone index. For each
5013 * zone, the number of pages is calculated as:
5015 * nr_free_zone_pages = managed_pages - high_pages
5017 * Return: number of pages beyond high watermark.
5019 static unsigned long nr_free_zone_pages(int offset
)
5024 /* Just pick one node, since fallback list is circular */
5025 unsigned long sum
= 0;
5027 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5029 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5030 unsigned long size
= zone_managed_pages(zone
);
5031 unsigned long high
= high_wmark_pages(zone
);
5040 * nr_free_buffer_pages - count number of pages beyond high watermark
5042 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5043 * watermark within ZONE_DMA and ZONE_NORMAL.
5045 * Return: number of pages beyond high watermark within ZONE_DMA and
5048 unsigned long nr_free_buffer_pages(void)
5050 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5052 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5055 * nr_free_pagecache_pages - count number of pages beyond high watermark
5057 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5058 * high watermark within all zones.
5060 * Return: number of pages beyond high watermark within all zones.
5062 unsigned long nr_free_pagecache_pages(void)
5064 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5067 static inline void show_node(struct zone
*zone
)
5069 if (IS_ENABLED(CONFIG_NUMA
))
5070 printk("Node %d ", zone_to_nid(zone
));
5073 long si_mem_available(void)
5076 unsigned long pagecache
;
5077 unsigned long wmark_low
= 0;
5078 unsigned long pages
[NR_LRU_LISTS
];
5079 unsigned long reclaimable
;
5083 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5084 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5087 wmark_low
+= low_wmark_pages(zone
);
5090 * Estimate the amount of memory available for userspace allocations,
5091 * without causing swapping.
5093 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5096 * Not all the page cache can be freed, otherwise the system will
5097 * start swapping. Assume at least half of the page cache, or the
5098 * low watermark worth of cache, needs to stay.
5100 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5101 pagecache
-= min(pagecache
/ 2, wmark_low
);
5102 available
+= pagecache
;
5105 * Part of the reclaimable slab and other kernel memory consists of
5106 * items that are in use, and cannot be freed. Cap this estimate at the
5109 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
5110 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5111 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5117 EXPORT_SYMBOL_GPL(si_mem_available
);
5119 void si_meminfo(struct sysinfo
*val
)
5121 val
->totalram
= totalram_pages();
5122 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5123 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5124 val
->bufferram
= nr_blockdev_pages();
5125 val
->totalhigh
= totalhigh_pages();
5126 val
->freehigh
= nr_free_highpages();
5127 val
->mem_unit
= PAGE_SIZE
;
5130 EXPORT_SYMBOL(si_meminfo
);
5133 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5135 int zone_type
; /* needs to be signed */
5136 unsigned long managed_pages
= 0;
5137 unsigned long managed_highpages
= 0;
5138 unsigned long free_highpages
= 0;
5139 pg_data_t
*pgdat
= NODE_DATA(nid
);
5141 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5142 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5143 val
->totalram
= managed_pages
;
5144 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5145 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5146 #ifdef CONFIG_HIGHMEM
5147 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5148 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5150 if (is_highmem(zone
)) {
5151 managed_highpages
+= zone_managed_pages(zone
);
5152 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5155 val
->totalhigh
= managed_highpages
;
5156 val
->freehigh
= free_highpages
;
5158 val
->totalhigh
= managed_highpages
;
5159 val
->freehigh
= free_highpages
;
5161 val
->mem_unit
= PAGE_SIZE
;
5166 * Determine whether the node should be displayed or not, depending on whether
5167 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5169 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5171 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5175 * no node mask - aka implicit memory numa policy. Do not bother with
5176 * the synchronization - read_mems_allowed_begin - because we do not
5177 * have to be precise here.
5180 nodemask
= &cpuset_current_mems_allowed
;
5182 return !node_isset(nid
, *nodemask
);
5185 #define K(x) ((x) << (PAGE_SHIFT-10))
5187 static void show_migration_types(unsigned char type
)
5189 static const char types
[MIGRATE_TYPES
] = {
5190 [MIGRATE_UNMOVABLE
] = 'U',
5191 [MIGRATE_MOVABLE
] = 'M',
5192 [MIGRATE_RECLAIMABLE
] = 'E',
5193 [MIGRATE_HIGHATOMIC
] = 'H',
5195 [MIGRATE_CMA
] = 'C',
5197 #ifdef CONFIG_MEMORY_ISOLATION
5198 [MIGRATE_ISOLATE
] = 'I',
5201 char tmp
[MIGRATE_TYPES
+ 1];
5205 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5206 if (type
& (1 << i
))
5211 printk(KERN_CONT
"(%s) ", tmp
);
5215 * Show free area list (used inside shift_scroll-lock stuff)
5216 * We also calculate the percentage fragmentation. We do this by counting the
5217 * memory on each free list with the exception of the first item on the list.
5220 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5223 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5225 unsigned long free_pcp
= 0;
5230 for_each_populated_zone(zone
) {
5231 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5234 for_each_online_cpu(cpu
)
5235 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5238 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5239 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5240 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5241 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5242 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5243 " free:%lu free_pcp:%lu free_cma:%lu\n",
5244 global_node_page_state(NR_ACTIVE_ANON
),
5245 global_node_page_state(NR_INACTIVE_ANON
),
5246 global_node_page_state(NR_ISOLATED_ANON
),
5247 global_node_page_state(NR_ACTIVE_FILE
),
5248 global_node_page_state(NR_INACTIVE_FILE
),
5249 global_node_page_state(NR_ISOLATED_FILE
),
5250 global_node_page_state(NR_UNEVICTABLE
),
5251 global_node_page_state(NR_FILE_DIRTY
),
5252 global_node_page_state(NR_WRITEBACK
),
5253 global_node_page_state(NR_UNSTABLE_NFS
),
5254 global_node_page_state(NR_SLAB_RECLAIMABLE
),
5255 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
5256 global_node_page_state(NR_FILE_MAPPED
),
5257 global_node_page_state(NR_SHMEM
),
5258 global_zone_page_state(NR_PAGETABLE
),
5259 global_zone_page_state(NR_BOUNCE
),
5260 global_zone_page_state(NR_FREE_PAGES
),
5262 global_zone_page_state(NR_FREE_CMA_PAGES
));
5264 for_each_online_pgdat(pgdat
) {
5265 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5269 " active_anon:%lukB"
5270 " inactive_anon:%lukB"
5271 " active_file:%lukB"
5272 " inactive_file:%lukB"
5273 " unevictable:%lukB"
5274 " isolated(anon):%lukB"
5275 " isolated(file):%lukB"
5280 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5282 " shmem_pmdmapped: %lukB"
5285 " writeback_tmp:%lukB"
5287 " all_unreclaimable? %s"
5290 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5291 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5292 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5293 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5294 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5295 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5296 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5297 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5298 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5299 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5300 K(node_page_state(pgdat
, NR_SHMEM
)),
5301 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5302 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5303 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5305 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5307 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5308 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
5309 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5313 for_each_populated_zone(zone
) {
5316 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5320 for_each_online_cpu(cpu
)
5321 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5330 " reserved_highatomic:%luKB"
5331 " active_anon:%lukB"
5332 " inactive_anon:%lukB"
5333 " active_file:%lukB"
5334 " inactive_file:%lukB"
5335 " unevictable:%lukB"
5336 " writepending:%lukB"
5340 " kernel_stack:%lukB"
5348 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5349 K(min_wmark_pages(zone
)),
5350 K(low_wmark_pages(zone
)),
5351 K(high_wmark_pages(zone
)),
5352 K(zone
->nr_reserved_highatomic
),
5353 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5354 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5355 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5356 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5357 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5358 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5359 K(zone
->present_pages
),
5360 K(zone_managed_pages(zone
)),
5361 K(zone_page_state(zone
, NR_MLOCK
)),
5362 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5363 K(zone_page_state(zone
, NR_PAGETABLE
)),
5364 K(zone_page_state(zone
, NR_BOUNCE
)),
5366 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5367 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5368 printk("lowmem_reserve[]:");
5369 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5370 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5371 printk(KERN_CONT
"\n");
5374 for_each_populated_zone(zone
) {
5376 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5377 unsigned char types
[MAX_ORDER
];
5379 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5382 printk(KERN_CONT
"%s: ", zone
->name
);
5384 spin_lock_irqsave(&zone
->lock
, flags
);
5385 for (order
= 0; order
< MAX_ORDER
; order
++) {
5386 struct free_area
*area
= &zone
->free_area
[order
];
5389 nr
[order
] = area
->nr_free
;
5390 total
+= nr
[order
] << order
;
5393 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5394 if (!free_area_empty(area
, type
))
5395 types
[order
] |= 1 << type
;
5398 spin_unlock_irqrestore(&zone
->lock
, flags
);
5399 for (order
= 0; order
< MAX_ORDER
; order
++) {
5400 printk(KERN_CONT
"%lu*%lukB ",
5401 nr
[order
], K(1UL) << order
);
5403 show_migration_types(types
[order
]);
5405 printk(KERN_CONT
"= %lukB\n", K(total
));
5408 hugetlb_show_meminfo();
5410 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5412 show_swap_cache_info();
5415 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5417 zoneref
->zone
= zone
;
5418 zoneref
->zone_idx
= zone_idx(zone
);
5422 * Builds allocation fallback zone lists.
5424 * Add all populated zones of a node to the zonelist.
5426 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5429 enum zone_type zone_type
= MAX_NR_ZONES
;
5434 zone
= pgdat
->node_zones
+ zone_type
;
5435 if (managed_zone(zone
)) {
5436 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5437 check_highest_zone(zone_type
);
5439 } while (zone_type
);
5446 static int __parse_numa_zonelist_order(char *s
)
5449 * We used to support different zonlists modes but they turned
5450 * out to be just not useful. Let's keep the warning in place
5451 * if somebody still use the cmd line parameter so that we do
5452 * not fail it silently
5454 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5455 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5461 static __init
int setup_numa_zonelist_order(char *s
)
5466 return __parse_numa_zonelist_order(s
);
5468 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
5470 char numa_zonelist_order
[] = "Node";
5473 * sysctl handler for numa_zonelist_order
5475 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5476 void __user
*buffer
, size_t *length
,
5483 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5484 str
= memdup_user_nul(buffer
, 16);
5486 return PTR_ERR(str
);
5488 ret
= __parse_numa_zonelist_order(str
);
5494 #define MAX_NODE_LOAD (nr_online_nodes)
5495 static int node_load
[MAX_NUMNODES
];
5498 * find_next_best_node - find the next node that should appear in a given node's fallback list
5499 * @node: node whose fallback list we're appending
5500 * @used_node_mask: nodemask_t of already used nodes
5502 * We use a number of factors to determine which is the next node that should
5503 * appear on a given node's fallback list. The node should not have appeared
5504 * already in @node's fallback list, and it should be the next closest node
5505 * according to the distance array (which contains arbitrary distance values
5506 * from each node to each node in the system), and should also prefer nodes
5507 * with no CPUs, since presumably they'll have very little allocation pressure
5508 * on them otherwise.
5510 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5512 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5515 int min_val
= INT_MAX
;
5516 int best_node
= NUMA_NO_NODE
;
5517 const struct cpumask
*tmp
= cpumask_of_node(0);
5519 /* Use the local node if we haven't already */
5520 if (!node_isset(node
, *used_node_mask
)) {
5521 node_set(node
, *used_node_mask
);
5525 for_each_node_state(n
, N_MEMORY
) {
5527 /* Don't want a node to appear more than once */
5528 if (node_isset(n
, *used_node_mask
))
5531 /* Use the distance array to find the distance */
5532 val
= node_distance(node
, n
);
5534 /* Penalize nodes under us ("prefer the next node") */
5537 /* Give preference to headless and unused nodes */
5538 tmp
= cpumask_of_node(n
);
5539 if (!cpumask_empty(tmp
))
5540 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5542 /* Slight preference for less loaded node */
5543 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5544 val
+= node_load
[n
];
5546 if (val
< min_val
) {
5553 node_set(best_node
, *used_node_mask
);
5560 * Build zonelists ordered by node and zones within node.
5561 * This results in maximum locality--normal zone overflows into local
5562 * DMA zone, if any--but risks exhausting DMA zone.
5564 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5567 struct zoneref
*zonerefs
;
5570 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5572 for (i
= 0; i
< nr_nodes
; i
++) {
5575 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5577 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5578 zonerefs
+= nr_zones
;
5580 zonerefs
->zone
= NULL
;
5581 zonerefs
->zone_idx
= 0;
5585 * Build gfp_thisnode zonelists
5587 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5589 struct zoneref
*zonerefs
;
5592 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5593 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5594 zonerefs
+= nr_zones
;
5595 zonerefs
->zone
= NULL
;
5596 zonerefs
->zone_idx
= 0;
5600 * Build zonelists ordered by zone and nodes within zones.
5601 * This results in conserving DMA zone[s] until all Normal memory is
5602 * exhausted, but results in overflowing to remote node while memory
5603 * may still exist in local DMA zone.
5606 static void build_zonelists(pg_data_t
*pgdat
)
5608 static int node_order
[MAX_NUMNODES
];
5609 int node
, load
, nr_nodes
= 0;
5610 nodemask_t used_mask
;
5611 int local_node
, prev_node
;
5613 /* NUMA-aware ordering of nodes */
5614 local_node
= pgdat
->node_id
;
5615 load
= nr_online_nodes
;
5616 prev_node
= local_node
;
5617 nodes_clear(used_mask
);
5619 memset(node_order
, 0, sizeof(node_order
));
5620 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5622 * We don't want to pressure a particular node.
5623 * So adding penalty to the first node in same
5624 * distance group to make it round-robin.
5626 if (node_distance(local_node
, node
) !=
5627 node_distance(local_node
, prev_node
))
5628 node_load
[node
] = load
;
5630 node_order
[nr_nodes
++] = node
;
5635 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5636 build_thisnode_zonelists(pgdat
);
5639 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5641 * Return node id of node used for "local" allocations.
5642 * I.e., first node id of first zone in arg node's generic zonelist.
5643 * Used for initializing percpu 'numa_mem', which is used primarily
5644 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5646 int local_memory_node(int node
)
5650 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5651 gfp_zone(GFP_KERNEL
),
5653 return zone_to_nid(z
->zone
);
5657 static void setup_min_unmapped_ratio(void);
5658 static void setup_min_slab_ratio(void);
5659 #else /* CONFIG_NUMA */
5661 static void build_zonelists(pg_data_t
*pgdat
)
5663 int node
, local_node
;
5664 struct zoneref
*zonerefs
;
5667 local_node
= pgdat
->node_id
;
5669 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5670 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5671 zonerefs
+= nr_zones
;
5674 * Now we build the zonelist so that it contains the zones
5675 * of all the other nodes.
5676 * We don't want to pressure a particular node, so when
5677 * building the zones for node N, we make sure that the
5678 * zones coming right after the local ones are those from
5679 * node N+1 (modulo N)
5681 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5682 if (!node_online(node
))
5684 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5685 zonerefs
+= nr_zones
;
5687 for (node
= 0; node
< local_node
; node
++) {
5688 if (!node_online(node
))
5690 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5691 zonerefs
+= nr_zones
;
5694 zonerefs
->zone
= NULL
;
5695 zonerefs
->zone_idx
= 0;
5698 #endif /* CONFIG_NUMA */
5701 * Boot pageset table. One per cpu which is going to be used for all
5702 * zones and all nodes. The parameters will be set in such a way
5703 * that an item put on a list will immediately be handed over to
5704 * the buddy list. This is safe since pageset manipulation is done
5705 * with interrupts disabled.
5707 * The boot_pagesets must be kept even after bootup is complete for
5708 * unused processors and/or zones. They do play a role for bootstrapping
5709 * hotplugged processors.
5711 * zoneinfo_show() and maybe other functions do
5712 * not check if the processor is online before following the pageset pointer.
5713 * Other parts of the kernel may not check if the zone is available.
5715 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5716 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5717 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5719 static void __build_all_zonelists(void *data
)
5722 int __maybe_unused cpu
;
5723 pg_data_t
*self
= data
;
5724 static DEFINE_SPINLOCK(lock
);
5729 memset(node_load
, 0, sizeof(node_load
));
5733 * This node is hotadded and no memory is yet present. So just
5734 * building zonelists is fine - no need to touch other nodes.
5736 if (self
&& !node_online(self
->node_id
)) {
5737 build_zonelists(self
);
5739 for_each_online_node(nid
) {
5740 pg_data_t
*pgdat
= NODE_DATA(nid
);
5742 build_zonelists(pgdat
);
5745 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5747 * We now know the "local memory node" for each node--
5748 * i.e., the node of the first zone in the generic zonelist.
5749 * Set up numa_mem percpu variable for on-line cpus. During
5750 * boot, only the boot cpu should be on-line; we'll init the
5751 * secondary cpus' numa_mem as they come on-line. During
5752 * node/memory hotplug, we'll fixup all on-line cpus.
5754 for_each_online_cpu(cpu
)
5755 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5762 static noinline
void __init
5763 build_all_zonelists_init(void)
5767 __build_all_zonelists(NULL
);
5770 * Initialize the boot_pagesets that are going to be used
5771 * for bootstrapping processors. The real pagesets for
5772 * each zone will be allocated later when the per cpu
5773 * allocator is available.
5775 * boot_pagesets are used also for bootstrapping offline
5776 * cpus if the system is already booted because the pagesets
5777 * are needed to initialize allocators on a specific cpu too.
5778 * F.e. the percpu allocator needs the page allocator which
5779 * needs the percpu allocator in order to allocate its pagesets
5780 * (a chicken-egg dilemma).
5782 for_each_possible_cpu(cpu
)
5783 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5785 mminit_verify_zonelist();
5786 cpuset_init_current_mems_allowed();
5790 * unless system_state == SYSTEM_BOOTING.
5792 * __ref due to call of __init annotated helper build_all_zonelists_init
5793 * [protected by SYSTEM_BOOTING].
5795 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5797 if (system_state
== SYSTEM_BOOTING
) {
5798 build_all_zonelists_init();
5800 __build_all_zonelists(pgdat
);
5801 /* cpuset refresh routine should be here */
5803 vm_total_pages
= nr_free_pagecache_pages();
5805 * Disable grouping by mobility if the number of pages in the
5806 * system is too low to allow the mechanism to work. It would be
5807 * more accurate, but expensive to check per-zone. This check is
5808 * made on memory-hotadd so a system can start with mobility
5809 * disabled and enable it later
5811 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5812 page_group_by_mobility_disabled
= 1;
5814 page_group_by_mobility_disabled
= 0;
5816 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5818 page_group_by_mobility_disabled
? "off" : "on",
5821 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5825 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5826 static bool __meminit
5827 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5829 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5830 static struct memblock_region
*r
;
5832 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5833 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5834 for_each_memblock(memory
, r
) {
5835 if (*pfn
< memblock_region_memory_end_pfn(r
))
5839 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5840 memblock_is_mirror(r
)) {
5841 *pfn
= memblock_region_memory_end_pfn(r
);
5849 #ifdef CONFIG_SPARSEMEM
5850 /* Skip PFNs that belong to non-present sections */
5851 static inline __meminit
unsigned long next_pfn(unsigned long pfn
)
5853 const unsigned long section_nr
= pfn_to_section_nr(++pfn
);
5855 if (present_section_nr(section_nr
))
5857 return section_nr_to_pfn(next_present_section_nr(section_nr
));
5860 static inline __meminit
unsigned long next_pfn(unsigned long pfn
)
5867 * Initially all pages are reserved - free ones are freed
5868 * up by memblock_free_all() once the early boot process is
5869 * done. Non-atomic initialization, single-pass.
5871 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5872 unsigned long start_pfn
, enum memmap_context context
,
5873 struct vmem_altmap
*altmap
)
5875 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5878 if (highest_memmap_pfn
< end_pfn
- 1)
5879 highest_memmap_pfn
= end_pfn
- 1;
5881 #ifdef CONFIG_ZONE_DEVICE
5883 * Honor reservation requested by the driver for this ZONE_DEVICE
5884 * memory. We limit the total number of pages to initialize to just
5885 * those that might contain the memory mapping. We will defer the
5886 * ZONE_DEVICE page initialization until after we have released
5889 if (zone
== ZONE_DEVICE
) {
5893 if (start_pfn
== altmap
->base_pfn
)
5894 start_pfn
+= altmap
->reserve
;
5895 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5899 for (pfn
= start_pfn
; pfn
< end_pfn
; ) {
5901 * There can be holes in boot-time mem_map[]s handed to this
5902 * function. They do not exist on hotplugged memory.
5904 if (context
== MEMMAP_EARLY
) {
5905 if (!early_pfn_valid(pfn
)) {
5906 pfn
= next_pfn(pfn
);
5909 if (!early_pfn_in_nid(pfn
, nid
)) {
5913 if (overlap_memmap_init(zone
, &pfn
))
5915 if (defer_init(nid
, pfn
, end_pfn
))
5919 page
= pfn_to_page(pfn
);
5920 __init_single_page(page
, pfn
, zone
, nid
);
5921 if (context
== MEMMAP_HOTPLUG
)
5922 __SetPageReserved(page
);
5925 * Mark the block movable so that blocks are reserved for
5926 * movable at startup. This will force kernel allocations
5927 * to reserve their blocks rather than leaking throughout
5928 * the address space during boot when many long-lived
5929 * kernel allocations are made.
5931 * bitmap is created for zone's valid pfn range. but memmap
5932 * can be created for invalid pages (for alignment)
5933 * check here not to call set_pageblock_migratetype() against
5936 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5937 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5944 #ifdef CONFIG_ZONE_DEVICE
5945 void __ref
memmap_init_zone_device(struct zone
*zone
,
5946 unsigned long start_pfn
,
5947 unsigned long nr_pages
,
5948 struct dev_pagemap
*pgmap
)
5950 unsigned long pfn
, end_pfn
= start_pfn
+ nr_pages
;
5951 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5952 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
5953 unsigned long zone_idx
= zone_idx(zone
);
5954 unsigned long start
= jiffies
;
5955 int nid
= pgdat
->node_id
;
5957 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
5961 * The call to memmap_init_zone should have already taken care
5962 * of the pages reserved for the memmap, so we can just jump to
5963 * the end of that region and start processing the device pages.
5966 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5967 nr_pages
= end_pfn
- start_pfn
;
5970 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5971 struct page
*page
= pfn_to_page(pfn
);
5973 __init_single_page(page
, pfn
, zone_idx
, nid
);
5976 * Mark page reserved as it will need to wait for onlining
5977 * phase for it to be fully associated with a zone.
5979 * We can use the non-atomic __set_bit operation for setting
5980 * the flag as we are still initializing the pages.
5982 __SetPageReserved(page
);
5985 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5986 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5987 * ever freed or placed on a driver-private list.
5989 page
->pgmap
= pgmap
;
5990 page
->zone_device_data
= NULL
;
5993 * Mark the block movable so that blocks are reserved for
5994 * movable at startup. This will force kernel allocations
5995 * to reserve their blocks rather than leaking throughout
5996 * the address space during boot when many long-lived
5997 * kernel allocations are made.
5999 * bitmap is created for zone's valid pfn range. but memmap
6000 * can be created for invalid pages (for alignment)
6001 * check here not to call set_pageblock_migratetype() against
6004 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6005 * because this is done early in section_activate()
6007 if (!(pfn
& (pageblock_nr_pages
- 1))) {
6008 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6013 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6014 nr_pages
, jiffies_to_msecs(jiffies
- start
));
6018 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6020 unsigned int order
, t
;
6021 for_each_migratetype_order(order
, t
) {
6022 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6023 zone
->free_area
[order
].nr_free
= 0;
6027 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
6028 unsigned long zone
, unsigned long start_pfn
)
6030 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
, NULL
);
6033 static int zone_batchsize(struct zone
*zone
)
6039 * The per-cpu-pages pools are set to around 1000th of the
6042 batch
= zone_managed_pages(zone
) / 1024;
6043 /* But no more than a meg. */
6044 if (batch
* PAGE_SIZE
> 1024 * 1024)
6045 batch
= (1024 * 1024) / PAGE_SIZE
;
6046 batch
/= 4; /* We effectively *= 4 below */
6051 * Clamp the batch to a 2^n - 1 value. Having a power
6052 * of 2 value was found to be more likely to have
6053 * suboptimal cache aliasing properties in some cases.
6055 * For example if 2 tasks are alternately allocating
6056 * batches of pages, one task can end up with a lot
6057 * of pages of one half of the possible page colors
6058 * and the other with pages of the other colors.
6060 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6065 /* The deferral and batching of frees should be suppressed under NOMMU
6068 * The problem is that NOMMU needs to be able to allocate large chunks
6069 * of contiguous memory as there's no hardware page translation to
6070 * assemble apparent contiguous memory from discontiguous pages.
6072 * Queueing large contiguous runs of pages for batching, however,
6073 * causes the pages to actually be freed in smaller chunks. As there
6074 * can be a significant delay between the individual batches being
6075 * recycled, this leads to the once large chunks of space being
6076 * fragmented and becoming unavailable for high-order allocations.
6083 * pcp->high and pcp->batch values are related and dependent on one another:
6084 * ->batch must never be higher then ->high.
6085 * The following function updates them in a safe manner without read side
6088 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6089 * those fields changing asynchronously (acording the the above rule).
6091 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6092 * outside of boot time (or some other assurance that no concurrent updaters
6095 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6096 unsigned long batch
)
6098 /* start with a fail safe value for batch */
6102 /* Update high, then batch, in order */
6109 /* a companion to pageset_set_high() */
6110 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
6112 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
6115 static void pageset_init(struct per_cpu_pageset
*p
)
6117 struct per_cpu_pages
*pcp
;
6120 memset(p
, 0, sizeof(*p
));
6123 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
6124 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
6127 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
6130 pageset_set_batch(p
, batch
);
6134 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6135 * to the value high for the pageset p.
6137 static void pageset_set_high(struct per_cpu_pageset
*p
,
6140 unsigned long batch
= max(1UL, high
/ 4);
6141 if ((high
/ 4) > (PAGE_SHIFT
* 8))
6142 batch
= PAGE_SHIFT
* 8;
6144 pageset_update(&p
->pcp
, high
, batch
);
6147 static void pageset_set_high_and_batch(struct zone
*zone
,
6148 struct per_cpu_pageset
*pcp
)
6150 if (percpu_pagelist_fraction
)
6151 pageset_set_high(pcp
,
6152 (zone_managed_pages(zone
) /
6153 percpu_pagelist_fraction
));
6155 pageset_set_batch(pcp
, zone_batchsize(zone
));
6158 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
6160 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
6163 pageset_set_high_and_batch(zone
, pcp
);
6166 void __meminit
setup_zone_pageset(struct zone
*zone
)
6169 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
6170 for_each_possible_cpu(cpu
)
6171 zone_pageset_init(zone
, cpu
);
6175 * Allocate per cpu pagesets and initialize them.
6176 * Before this call only boot pagesets were available.
6178 void __init
setup_per_cpu_pageset(void)
6180 struct pglist_data
*pgdat
;
6183 for_each_populated_zone(zone
)
6184 setup_zone_pageset(zone
);
6186 for_each_online_pgdat(pgdat
)
6187 pgdat
->per_cpu_nodestats
=
6188 alloc_percpu(struct per_cpu_nodestat
);
6191 static __meminit
void zone_pcp_init(struct zone
*zone
)
6194 * per cpu subsystem is not up at this point. The following code
6195 * relies on the ability of the linker to provide the
6196 * offset of a (static) per cpu variable into the per cpu area.
6198 zone
->pageset
= &boot_pageset
;
6200 if (populated_zone(zone
))
6201 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
6202 zone
->name
, zone
->present_pages
,
6203 zone_batchsize(zone
));
6206 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6207 unsigned long zone_start_pfn
,
6210 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6211 int zone_idx
= zone_idx(zone
) + 1;
6213 if (zone_idx
> pgdat
->nr_zones
)
6214 pgdat
->nr_zones
= zone_idx
;
6216 zone
->zone_start_pfn
= zone_start_pfn
;
6218 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6219 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6221 (unsigned long)zone_idx(zone
),
6222 zone_start_pfn
, (zone_start_pfn
+ size
));
6224 zone_init_free_lists(zone
);
6225 zone
->initialized
= 1;
6228 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6229 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6232 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6234 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
6235 struct mminit_pfnnid_cache
*state
)
6237 unsigned long start_pfn
, end_pfn
;
6240 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
6241 return state
->last_nid
;
6243 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
6244 if (nid
!= NUMA_NO_NODE
) {
6245 state
->last_start
= start_pfn
;
6246 state
->last_end
= end_pfn
;
6247 state
->last_nid
= nid
;
6252 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6255 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6256 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6257 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6259 * If an architecture guarantees that all ranges registered contain no holes
6260 * and may be freed, this this function may be used instead of calling
6261 * memblock_free_early_nid() manually.
6263 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
6265 unsigned long start_pfn
, end_pfn
;
6268 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
6269 start_pfn
= min(start_pfn
, max_low_pfn
);
6270 end_pfn
= min(end_pfn
, max_low_pfn
);
6272 if (start_pfn
< end_pfn
)
6273 memblock_free_early_nid(PFN_PHYS(start_pfn
),
6274 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
6280 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6281 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6283 * If an architecture guarantees that all ranges registered contain no holes and may
6284 * be freed, this function may be used instead of calling memory_present() manually.
6286 void __init
sparse_memory_present_with_active_regions(int nid
)
6288 unsigned long start_pfn
, end_pfn
;
6291 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
6292 memory_present(this_nid
, start_pfn
, end_pfn
);
6296 * get_pfn_range_for_nid - Return the start and end page frames for a node
6297 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6298 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6299 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6301 * It returns the start and end page frame of a node based on information
6302 * provided by memblock_set_node(). If called for a node
6303 * with no available memory, a warning is printed and the start and end
6306 void __init
get_pfn_range_for_nid(unsigned int nid
,
6307 unsigned long *start_pfn
, unsigned long *end_pfn
)
6309 unsigned long this_start_pfn
, this_end_pfn
;
6315 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6316 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6317 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6320 if (*start_pfn
== -1UL)
6325 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6326 * assumption is made that zones within a node are ordered in monotonic
6327 * increasing memory addresses so that the "highest" populated zone is used
6329 static void __init
find_usable_zone_for_movable(void)
6332 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6333 if (zone_index
== ZONE_MOVABLE
)
6336 if (arch_zone_highest_possible_pfn
[zone_index
] >
6337 arch_zone_lowest_possible_pfn
[zone_index
])
6341 VM_BUG_ON(zone_index
== -1);
6342 movable_zone
= zone_index
;
6346 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6347 * because it is sized independent of architecture. Unlike the other zones,
6348 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6349 * in each node depending on the size of each node and how evenly kernelcore
6350 * is distributed. This helper function adjusts the zone ranges
6351 * provided by the architecture for a given node by using the end of the
6352 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6353 * zones within a node are in order of monotonic increases memory addresses
6355 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6356 unsigned long zone_type
,
6357 unsigned long node_start_pfn
,
6358 unsigned long node_end_pfn
,
6359 unsigned long *zone_start_pfn
,
6360 unsigned long *zone_end_pfn
)
6362 /* Only adjust if ZONE_MOVABLE is on this node */
6363 if (zone_movable_pfn
[nid
]) {
6364 /* Size ZONE_MOVABLE */
6365 if (zone_type
== ZONE_MOVABLE
) {
6366 *zone_start_pfn
= zone_movable_pfn
[nid
];
6367 *zone_end_pfn
= min(node_end_pfn
,
6368 arch_zone_highest_possible_pfn
[movable_zone
]);
6370 /* Adjust for ZONE_MOVABLE starting within this range */
6371 } else if (!mirrored_kernelcore
&&
6372 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6373 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6374 *zone_end_pfn
= zone_movable_pfn
[nid
];
6376 /* Check if this whole range is within ZONE_MOVABLE */
6377 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6378 *zone_start_pfn
= *zone_end_pfn
;
6383 * Return the number of pages a zone spans in a node, including holes
6384 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6386 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6387 unsigned long zone_type
,
6388 unsigned long node_start_pfn
,
6389 unsigned long node_end_pfn
,
6390 unsigned long *zone_start_pfn
,
6391 unsigned long *zone_end_pfn
,
6392 unsigned long *ignored
)
6394 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6395 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6396 /* When hotadd a new node from cpu_up(), the node should be empty */
6397 if (!node_start_pfn
&& !node_end_pfn
)
6400 /* Get the start and end of the zone */
6401 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6402 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6403 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6404 node_start_pfn
, node_end_pfn
,
6405 zone_start_pfn
, zone_end_pfn
);
6407 /* Check that this node has pages within the zone's required range */
6408 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6411 /* Move the zone boundaries inside the node if necessary */
6412 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6413 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6415 /* Return the spanned pages */
6416 return *zone_end_pfn
- *zone_start_pfn
;
6420 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6421 * then all holes in the requested range will be accounted for.
6423 unsigned long __init
__absent_pages_in_range(int nid
,
6424 unsigned long range_start_pfn
,
6425 unsigned long range_end_pfn
)
6427 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6428 unsigned long start_pfn
, end_pfn
;
6431 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6432 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6433 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6434 nr_absent
-= end_pfn
- start_pfn
;
6440 * absent_pages_in_range - Return number of page frames in holes within a range
6441 * @start_pfn: The start PFN to start searching for holes
6442 * @end_pfn: The end PFN to stop searching for holes
6444 * Return: the number of pages frames in memory holes within a range.
6446 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6447 unsigned long end_pfn
)
6449 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6452 /* Return the number of page frames in holes in a zone on a node */
6453 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6454 unsigned long zone_type
,
6455 unsigned long node_start_pfn
,
6456 unsigned long node_end_pfn
,
6457 unsigned long *ignored
)
6459 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6460 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6461 unsigned long zone_start_pfn
, zone_end_pfn
;
6462 unsigned long nr_absent
;
6464 /* When hotadd a new node from cpu_up(), the node should be empty */
6465 if (!node_start_pfn
&& !node_end_pfn
)
6468 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6469 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6471 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6472 node_start_pfn
, node_end_pfn
,
6473 &zone_start_pfn
, &zone_end_pfn
);
6474 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6477 * ZONE_MOVABLE handling.
6478 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6481 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6482 unsigned long start_pfn
, end_pfn
;
6483 struct memblock_region
*r
;
6485 for_each_memblock(memory
, r
) {
6486 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6487 zone_start_pfn
, zone_end_pfn
);
6488 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6489 zone_start_pfn
, zone_end_pfn
);
6491 if (zone_type
== ZONE_MOVABLE
&&
6492 memblock_is_mirror(r
))
6493 nr_absent
+= end_pfn
- start_pfn
;
6495 if (zone_type
== ZONE_NORMAL
&&
6496 !memblock_is_mirror(r
))
6497 nr_absent
+= end_pfn
- start_pfn
;
6504 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6505 static inline unsigned long __init
zone_spanned_pages_in_node(int nid
,
6506 unsigned long zone_type
,
6507 unsigned long node_start_pfn
,
6508 unsigned long node_end_pfn
,
6509 unsigned long *zone_start_pfn
,
6510 unsigned long *zone_end_pfn
,
6511 unsigned long *zones_size
)
6515 *zone_start_pfn
= node_start_pfn
;
6516 for (zone
= 0; zone
< zone_type
; zone
++)
6517 *zone_start_pfn
+= zones_size
[zone
];
6519 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
6521 return zones_size
[zone_type
];
6524 static inline unsigned long __init
zone_absent_pages_in_node(int nid
,
6525 unsigned long zone_type
,
6526 unsigned long node_start_pfn
,
6527 unsigned long node_end_pfn
,
6528 unsigned long *zholes_size
)
6533 return zholes_size
[zone_type
];
6536 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6538 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6539 unsigned long node_start_pfn
,
6540 unsigned long node_end_pfn
,
6541 unsigned long *zones_size
,
6542 unsigned long *zholes_size
)
6544 unsigned long realtotalpages
= 0, totalpages
= 0;
6547 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6548 struct zone
*zone
= pgdat
->node_zones
+ i
;
6549 unsigned long zone_start_pfn
, zone_end_pfn
;
6550 unsigned long size
, real_size
;
6552 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6558 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
6559 node_start_pfn
, node_end_pfn
,
6562 zone
->zone_start_pfn
= zone_start_pfn
;
6564 zone
->zone_start_pfn
= 0;
6565 zone
->spanned_pages
= size
;
6566 zone
->present_pages
= real_size
;
6569 realtotalpages
+= real_size
;
6572 pgdat
->node_spanned_pages
= totalpages
;
6573 pgdat
->node_present_pages
= realtotalpages
;
6574 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6578 #ifndef CONFIG_SPARSEMEM
6580 * Calculate the size of the zone->blockflags rounded to an unsigned long
6581 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6582 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6583 * round what is now in bits to nearest long in bits, then return it in
6586 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6588 unsigned long usemapsize
;
6590 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6591 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6592 usemapsize
= usemapsize
>> pageblock_order
;
6593 usemapsize
*= NR_PAGEBLOCK_BITS
;
6594 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6596 return usemapsize
/ 8;
6599 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6601 unsigned long zone_start_pfn
,
6602 unsigned long zonesize
)
6604 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6605 zone
->pageblock_flags
= NULL
;
6607 zone
->pageblock_flags
=
6608 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
6610 if (!zone
->pageblock_flags
)
6611 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6612 usemapsize
, zone
->name
, pgdat
->node_id
);
6616 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6617 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6618 #endif /* CONFIG_SPARSEMEM */
6620 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6622 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6623 void __init
set_pageblock_order(void)
6627 /* Check that pageblock_nr_pages has not already been setup */
6628 if (pageblock_order
)
6631 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6632 order
= HUGETLB_PAGE_ORDER
;
6634 order
= MAX_ORDER
- 1;
6637 * Assume the largest contiguous order of interest is a huge page.
6638 * This value may be variable depending on boot parameters on IA64 and
6641 pageblock_order
= order
;
6643 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6646 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6647 * is unused as pageblock_order is set at compile-time. See
6648 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6651 void __init
set_pageblock_order(void)
6655 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6657 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6658 unsigned long present_pages
)
6660 unsigned long pages
= spanned_pages
;
6663 * Provide a more accurate estimation if there are holes within
6664 * the zone and SPARSEMEM is in use. If there are holes within the
6665 * zone, each populated memory region may cost us one or two extra
6666 * memmap pages due to alignment because memmap pages for each
6667 * populated regions may not be naturally aligned on page boundary.
6668 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6670 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6671 IS_ENABLED(CONFIG_SPARSEMEM
))
6672 pages
= present_pages
;
6674 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6677 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6678 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6680 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
6682 spin_lock_init(&ds_queue
->split_queue_lock
);
6683 INIT_LIST_HEAD(&ds_queue
->split_queue
);
6684 ds_queue
->split_queue_len
= 0;
6687 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6690 #ifdef CONFIG_COMPACTION
6691 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6693 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6696 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6699 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6701 pgdat_resize_init(pgdat
);
6703 pgdat_init_split_queue(pgdat
);
6704 pgdat_init_kcompactd(pgdat
);
6706 init_waitqueue_head(&pgdat
->kswapd_wait
);
6707 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6709 pgdat_page_ext_init(pgdat
);
6710 spin_lock_init(&pgdat
->lru_lock
);
6711 lruvec_init(&pgdat
->__lruvec
);
6714 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6715 unsigned long remaining_pages
)
6717 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6718 zone_set_nid(zone
, nid
);
6719 zone
->name
= zone_names
[idx
];
6720 zone
->zone_pgdat
= NODE_DATA(nid
);
6721 spin_lock_init(&zone
->lock
);
6722 zone_seqlock_init(zone
);
6723 zone_pcp_init(zone
);
6727 * Set up the zone data structures
6728 * - init pgdat internals
6729 * - init all zones belonging to this node
6731 * NOTE: this function is only called during memory hotplug
6733 #ifdef CONFIG_MEMORY_HOTPLUG
6734 void __ref
free_area_init_core_hotplug(int nid
)
6737 pg_data_t
*pgdat
= NODE_DATA(nid
);
6739 pgdat_init_internals(pgdat
);
6740 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6741 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6746 * Set up the zone data structures:
6747 * - mark all pages reserved
6748 * - mark all memory queues empty
6749 * - clear the memory bitmaps
6751 * NOTE: pgdat should get zeroed by caller.
6752 * NOTE: this function is only called during early init.
6754 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6757 int nid
= pgdat
->node_id
;
6759 pgdat_init_internals(pgdat
);
6760 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6762 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6763 struct zone
*zone
= pgdat
->node_zones
+ j
;
6764 unsigned long size
, freesize
, memmap_pages
;
6765 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6767 size
= zone
->spanned_pages
;
6768 freesize
= zone
->present_pages
;
6771 * Adjust freesize so that it accounts for how much memory
6772 * is used by this zone for memmap. This affects the watermark
6773 * and per-cpu initialisations
6775 memmap_pages
= calc_memmap_size(size
, freesize
);
6776 if (!is_highmem_idx(j
)) {
6777 if (freesize
>= memmap_pages
) {
6778 freesize
-= memmap_pages
;
6781 " %s zone: %lu pages used for memmap\n",
6782 zone_names
[j
], memmap_pages
);
6784 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6785 zone_names
[j
], memmap_pages
, freesize
);
6788 /* Account for reserved pages */
6789 if (j
== 0 && freesize
> dma_reserve
) {
6790 freesize
-= dma_reserve
;
6791 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6792 zone_names
[0], dma_reserve
);
6795 if (!is_highmem_idx(j
))
6796 nr_kernel_pages
+= freesize
;
6797 /* Charge for highmem memmap if there are enough kernel pages */
6798 else if (nr_kernel_pages
> memmap_pages
* 2)
6799 nr_kernel_pages
-= memmap_pages
;
6800 nr_all_pages
+= freesize
;
6803 * Set an approximate value for lowmem here, it will be adjusted
6804 * when the bootmem allocator frees pages into the buddy system.
6805 * And all highmem pages will be managed by the buddy system.
6807 zone_init_internals(zone
, j
, nid
, freesize
);
6812 set_pageblock_order();
6813 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6814 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6815 memmap_init(size
, nid
, j
, zone_start_pfn
);
6819 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6820 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6822 unsigned long __maybe_unused start
= 0;
6823 unsigned long __maybe_unused offset
= 0;
6825 /* Skip empty nodes */
6826 if (!pgdat
->node_spanned_pages
)
6829 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6830 offset
= pgdat
->node_start_pfn
- start
;
6831 /* ia64 gets its own node_mem_map, before this, without bootmem */
6832 if (!pgdat
->node_mem_map
) {
6833 unsigned long size
, end
;
6837 * The zone's endpoints aren't required to be MAX_ORDER
6838 * aligned but the node_mem_map endpoints must be in order
6839 * for the buddy allocator to function correctly.
6841 end
= pgdat_end_pfn(pgdat
);
6842 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6843 size
= (end
- start
) * sizeof(struct page
);
6844 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
6847 panic("Failed to allocate %ld bytes for node %d memory map\n",
6848 size
, pgdat
->node_id
);
6849 pgdat
->node_mem_map
= map
+ offset
;
6851 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6852 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6853 (unsigned long)pgdat
->node_mem_map
);
6854 #ifndef CONFIG_NEED_MULTIPLE_NODES
6856 * With no DISCONTIG, the global mem_map is just set as node 0's
6858 if (pgdat
== NODE_DATA(0)) {
6859 mem_map
= NODE_DATA(0)->node_mem_map
;
6860 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6861 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6863 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6868 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6869 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6871 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6872 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6874 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6877 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6880 void __init
free_area_init_node(int nid
, unsigned long *zones_size
,
6881 unsigned long node_start_pfn
,
6882 unsigned long *zholes_size
)
6884 pg_data_t
*pgdat
= NODE_DATA(nid
);
6885 unsigned long start_pfn
= 0;
6886 unsigned long end_pfn
= 0;
6888 /* pg_data_t should be reset to zero when it's allocated */
6889 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6891 pgdat
->node_id
= nid
;
6892 pgdat
->node_start_pfn
= node_start_pfn
;
6893 pgdat
->per_cpu_nodestats
= NULL
;
6894 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6895 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6896 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6897 (u64
)start_pfn
<< PAGE_SHIFT
,
6898 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6900 start_pfn
= node_start_pfn
;
6902 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6903 zones_size
, zholes_size
);
6905 alloc_node_mem_map(pgdat
);
6906 pgdat_set_deferred_range(pgdat
);
6908 free_area_init_core(pgdat
);
6911 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6913 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6914 * PageReserved(). Return the number of struct pages that were initialized.
6916 static u64 __init
init_unavailable_range(unsigned long spfn
, unsigned long epfn
)
6921 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6922 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6923 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6924 + pageblock_nr_pages
- 1;
6928 * Use a fake node/zone (0) for now. Some of these pages
6929 * (in memblock.reserved but not in memblock.memory) will
6930 * get re-initialized via reserve_bootmem_region() later.
6932 __init_single_page(pfn_to_page(pfn
), pfn
, 0, 0);
6933 __SetPageReserved(pfn_to_page(pfn
));
6941 * Only struct pages that are backed by physical memory are zeroed and
6942 * initialized by going through __init_single_page(). But, there are some
6943 * struct pages which are reserved in memblock allocator and their fields
6944 * may be accessed (for example page_to_pfn() on some configuration accesses
6945 * flags). We must explicitly initialize those struct pages.
6947 * This function also addresses a similar issue where struct pages are left
6948 * uninitialized because the physical address range is not covered by
6949 * memblock.memory or memblock.reserved. That could happen when memblock
6950 * layout is manually configured via memmap=, or when the highest physical
6951 * address (max_pfn) does not end on a section boundary.
6953 static void __init
init_unavailable_mem(void)
6955 phys_addr_t start
, end
;
6957 phys_addr_t next
= 0;
6960 * Loop through unavailable ranges not covered by memblock.memory.
6963 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6964 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6966 pgcnt
+= init_unavailable_range(PFN_DOWN(next
),
6972 * Early sections always have a fully populated memmap for the whole
6973 * section - see pfn_valid(). If the last section has holes at the
6974 * end and that section is marked "online", the memmap will be
6975 * considered initialized. Make sure that memmap has a well defined
6978 pgcnt
+= init_unavailable_range(PFN_DOWN(next
),
6979 round_up(max_pfn
, PAGES_PER_SECTION
));
6982 * Struct pages that do not have backing memory. This could be because
6983 * firmware is using some of this memory, or for some other reasons.
6986 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
6989 static inline void __init
init_unavailable_mem(void)
6992 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6994 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6996 #if MAX_NUMNODES > 1
6998 * Figure out the number of possible node ids.
7000 void __init
setup_nr_node_ids(void)
7002 unsigned int highest
;
7004 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
7005 nr_node_ids
= highest
+ 1;
7010 * node_map_pfn_alignment - determine the maximum internode alignment
7012 * This function should be called after node map is populated and sorted.
7013 * It calculates the maximum power of two alignment which can distinguish
7016 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7017 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7018 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7019 * shifted, 1GiB is enough and this function will indicate so.
7021 * This is used to test whether pfn -> nid mapping of the chosen memory
7022 * model has fine enough granularity to avoid incorrect mapping for the
7023 * populated node map.
7025 * Return: the determined alignment in pfn's. 0 if there is no alignment
7026 * requirement (single node).
7028 unsigned long __init
node_map_pfn_alignment(void)
7030 unsigned long accl_mask
= 0, last_end
= 0;
7031 unsigned long start
, end
, mask
;
7032 int last_nid
= NUMA_NO_NODE
;
7035 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
7036 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
7043 * Start with a mask granular enough to pin-point to the
7044 * start pfn and tick off bits one-by-one until it becomes
7045 * too coarse to separate the current node from the last.
7047 mask
= ~((1 << __ffs(start
)) - 1);
7048 while (mask
&& last_end
<= (start
& (mask
<< 1)))
7051 /* accumulate all internode masks */
7055 /* convert mask to number of pages */
7056 return ~accl_mask
+ 1;
7059 /* Find the lowest pfn for a node */
7060 static unsigned long __init
find_min_pfn_for_node(int nid
)
7062 unsigned long min_pfn
= ULONG_MAX
;
7063 unsigned long start_pfn
;
7066 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
7067 min_pfn
= min(min_pfn
, start_pfn
);
7069 if (min_pfn
== ULONG_MAX
) {
7070 pr_warn("Could not find start_pfn for node %d\n", nid
);
7078 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7080 * Return: the minimum PFN based on information provided via
7081 * memblock_set_node().
7083 unsigned long __init
find_min_pfn_with_active_regions(void)
7085 return find_min_pfn_for_node(MAX_NUMNODES
);
7089 * early_calculate_totalpages()
7090 * Sum pages in active regions for movable zone.
7091 * Populate N_MEMORY for calculating usable_nodes.
7093 static unsigned long __init
early_calculate_totalpages(void)
7095 unsigned long totalpages
= 0;
7096 unsigned long start_pfn
, end_pfn
;
7099 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7100 unsigned long pages
= end_pfn
- start_pfn
;
7102 totalpages
+= pages
;
7104 node_set_state(nid
, N_MEMORY
);
7110 * Find the PFN the Movable zone begins in each node. Kernel memory
7111 * is spread evenly between nodes as long as the nodes have enough
7112 * memory. When they don't, some nodes will have more kernelcore than
7115 static void __init
find_zone_movable_pfns_for_nodes(void)
7118 unsigned long usable_startpfn
;
7119 unsigned long kernelcore_node
, kernelcore_remaining
;
7120 /* save the state before borrow the nodemask */
7121 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
7122 unsigned long totalpages
= early_calculate_totalpages();
7123 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
7124 struct memblock_region
*r
;
7126 /* Need to find movable_zone earlier when movable_node is specified. */
7127 find_usable_zone_for_movable();
7130 * If movable_node is specified, ignore kernelcore and movablecore
7133 if (movable_node_is_enabled()) {
7134 for_each_memblock(memory
, r
) {
7135 if (!memblock_is_hotpluggable(r
))
7140 usable_startpfn
= PFN_DOWN(r
->base
);
7141 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7142 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7150 * If kernelcore=mirror is specified, ignore movablecore option
7152 if (mirrored_kernelcore
) {
7153 bool mem_below_4gb_not_mirrored
= false;
7155 for_each_memblock(memory
, r
) {
7156 if (memblock_is_mirror(r
))
7161 usable_startpfn
= memblock_region_memory_base_pfn(r
);
7163 if (usable_startpfn
< 0x100000) {
7164 mem_below_4gb_not_mirrored
= true;
7168 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7169 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7173 if (mem_below_4gb_not_mirrored
)
7174 pr_warn("This configuration results in unmirrored kernel memory.");
7180 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7181 * amount of necessary memory.
7183 if (required_kernelcore_percent
)
7184 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
7186 if (required_movablecore_percent
)
7187 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
7191 * If movablecore= was specified, calculate what size of
7192 * kernelcore that corresponds so that memory usable for
7193 * any allocation type is evenly spread. If both kernelcore
7194 * and movablecore are specified, then the value of kernelcore
7195 * will be used for required_kernelcore if it's greater than
7196 * what movablecore would have allowed.
7198 if (required_movablecore
) {
7199 unsigned long corepages
;
7202 * Round-up so that ZONE_MOVABLE is at least as large as what
7203 * was requested by the user
7205 required_movablecore
=
7206 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
7207 required_movablecore
= min(totalpages
, required_movablecore
);
7208 corepages
= totalpages
- required_movablecore
;
7210 required_kernelcore
= max(required_kernelcore
, corepages
);
7214 * If kernelcore was not specified or kernelcore size is larger
7215 * than totalpages, there is no ZONE_MOVABLE.
7217 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
7220 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7221 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
7224 /* Spread kernelcore memory as evenly as possible throughout nodes */
7225 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7226 for_each_node_state(nid
, N_MEMORY
) {
7227 unsigned long start_pfn
, end_pfn
;
7230 * Recalculate kernelcore_node if the division per node
7231 * now exceeds what is necessary to satisfy the requested
7232 * amount of memory for the kernel
7234 if (required_kernelcore
< kernelcore_node
)
7235 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7238 * As the map is walked, we track how much memory is usable
7239 * by the kernel using kernelcore_remaining. When it is
7240 * 0, the rest of the node is usable by ZONE_MOVABLE
7242 kernelcore_remaining
= kernelcore_node
;
7244 /* Go through each range of PFNs within this node */
7245 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7246 unsigned long size_pages
;
7248 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
7249 if (start_pfn
>= end_pfn
)
7252 /* Account for what is only usable for kernelcore */
7253 if (start_pfn
< usable_startpfn
) {
7254 unsigned long kernel_pages
;
7255 kernel_pages
= min(end_pfn
, usable_startpfn
)
7258 kernelcore_remaining
-= min(kernel_pages
,
7259 kernelcore_remaining
);
7260 required_kernelcore
-= min(kernel_pages
,
7261 required_kernelcore
);
7263 /* Continue if range is now fully accounted */
7264 if (end_pfn
<= usable_startpfn
) {
7267 * Push zone_movable_pfn to the end so
7268 * that if we have to rebalance
7269 * kernelcore across nodes, we will
7270 * not double account here
7272 zone_movable_pfn
[nid
] = end_pfn
;
7275 start_pfn
= usable_startpfn
;
7279 * The usable PFN range for ZONE_MOVABLE is from
7280 * start_pfn->end_pfn. Calculate size_pages as the
7281 * number of pages used as kernelcore
7283 size_pages
= end_pfn
- start_pfn
;
7284 if (size_pages
> kernelcore_remaining
)
7285 size_pages
= kernelcore_remaining
;
7286 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7289 * Some kernelcore has been met, update counts and
7290 * break if the kernelcore for this node has been
7293 required_kernelcore
-= min(required_kernelcore
,
7295 kernelcore_remaining
-= size_pages
;
7296 if (!kernelcore_remaining
)
7302 * If there is still required_kernelcore, we do another pass with one
7303 * less node in the count. This will push zone_movable_pfn[nid] further
7304 * along on the nodes that still have memory until kernelcore is
7308 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7312 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7313 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7314 zone_movable_pfn
[nid
] =
7315 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7318 /* restore the node_state */
7319 node_states
[N_MEMORY
] = saved_node_state
;
7322 /* Any regular or high memory on that node ? */
7323 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7325 enum zone_type zone_type
;
7327 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7328 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7329 if (populated_zone(zone
)) {
7330 if (IS_ENABLED(CONFIG_HIGHMEM
))
7331 node_set_state(nid
, N_HIGH_MEMORY
);
7332 if (zone_type
<= ZONE_NORMAL
)
7333 node_set_state(nid
, N_NORMAL_MEMORY
);
7340 * free_area_init_nodes - Initialise all pg_data_t and zone data
7341 * @max_zone_pfn: an array of max PFNs for each zone
7343 * This will call free_area_init_node() for each active node in the system.
7344 * Using the page ranges provided by memblock_set_node(), the size of each
7345 * zone in each node and their holes is calculated. If the maximum PFN
7346 * between two adjacent zones match, it is assumed that the zone is empty.
7347 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7348 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7349 * starts where the previous one ended. For example, ZONE_DMA32 starts
7350 * at arch_max_dma_pfn.
7352 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
7354 unsigned long start_pfn
, end_pfn
;
7357 /* Record where the zone boundaries are */
7358 memset(arch_zone_lowest_possible_pfn
, 0,
7359 sizeof(arch_zone_lowest_possible_pfn
));
7360 memset(arch_zone_highest_possible_pfn
, 0,
7361 sizeof(arch_zone_highest_possible_pfn
));
7363 start_pfn
= find_min_pfn_with_active_regions();
7365 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7366 if (i
== ZONE_MOVABLE
)
7369 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
7370 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
7371 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
7373 start_pfn
= end_pfn
;
7376 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7377 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7378 find_zone_movable_pfns_for_nodes();
7380 /* Print out the zone ranges */
7381 pr_info("Zone ranges:\n");
7382 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7383 if (i
== ZONE_MOVABLE
)
7385 pr_info(" %-8s ", zone_names
[i
]);
7386 if (arch_zone_lowest_possible_pfn
[i
] ==
7387 arch_zone_highest_possible_pfn
[i
])
7390 pr_cont("[mem %#018Lx-%#018Lx]\n",
7391 (u64
)arch_zone_lowest_possible_pfn
[i
]
7393 ((u64
)arch_zone_highest_possible_pfn
[i
]
7394 << PAGE_SHIFT
) - 1);
7397 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7398 pr_info("Movable zone start for each node\n");
7399 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7400 if (zone_movable_pfn
[i
])
7401 pr_info(" Node %d: %#018Lx\n", i
,
7402 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7406 * Print out the early node map, and initialize the
7407 * subsection-map relative to active online memory ranges to
7408 * enable future "sub-section" extensions of the memory map.
7410 pr_info("Early memory node ranges\n");
7411 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7412 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7413 (u64
)start_pfn
<< PAGE_SHIFT
,
7414 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7415 subsection_map_init(start_pfn
, end_pfn
- start_pfn
);
7418 /* Initialise every node */
7419 mminit_verify_pageflags_layout();
7420 setup_nr_node_ids();
7421 init_unavailable_mem();
7422 for_each_online_node(nid
) {
7423 pg_data_t
*pgdat
= NODE_DATA(nid
);
7424 free_area_init_node(nid
, NULL
,
7425 find_min_pfn_for_node(nid
), NULL
);
7427 /* Any memory on that node */
7428 if (pgdat
->node_present_pages
)
7429 node_set_state(nid
, N_MEMORY
);
7430 check_for_memory(pgdat
, nid
);
7434 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7435 unsigned long *percent
)
7437 unsigned long long coremem
;
7443 /* Value may be a percentage of total memory, otherwise bytes */
7444 coremem
= simple_strtoull(p
, &endptr
, 0);
7445 if (*endptr
== '%') {
7446 /* Paranoid check for percent values greater than 100 */
7447 WARN_ON(coremem
> 100);
7451 coremem
= memparse(p
, &p
);
7452 /* Paranoid check that UL is enough for the coremem value */
7453 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7455 *core
= coremem
>> PAGE_SHIFT
;
7462 * kernelcore=size sets the amount of memory for use for allocations that
7463 * cannot be reclaimed or migrated.
7465 static int __init
cmdline_parse_kernelcore(char *p
)
7467 /* parse kernelcore=mirror */
7468 if (parse_option_str(p
, "mirror")) {
7469 mirrored_kernelcore
= true;
7473 return cmdline_parse_core(p
, &required_kernelcore
,
7474 &required_kernelcore_percent
);
7478 * movablecore=size sets the amount of memory for use for allocations that
7479 * can be reclaimed or migrated.
7481 static int __init
cmdline_parse_movablecore(char *p
)
7483 return cmdline_parse_core(p
, &required_movablecore
,
7484 &required_movablecore_percent
);
7487 early_param("kernelcore", cmdline_parse_kernelcore
);
7488 early_param("movablecore", cmdline_parse_movablecore
);
7490 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7492 void adjust_managed_page_count(struct page
*page
, long count
)
7494 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7495 totalram_pages_add(count
);
7496 #ifdef CONFIG_HIGHMEM
7497 if (PageHighMem(page
))
7498 totalhigh_pages_add(count
);
7501 EXPORT_SYMBOL(adjust_managed_page_count
);
7503 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7506 unsigned long pages
= 0;
7508 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7509 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7510 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7511 struct page
*page
= virt_to_page(pos
);
7512 void *direct_map_addr
;
7515 * 'direct_map_addr' might be different from 'pos'
7516 * because some architectures' virt_to_page()
7517 * work with aliases. Getting the direct map
7518 * address ensures that we get a _writeable_
7519 * alias for the memset().
7521 direct_map_addr
= page_address(page
);
7522 if ((unsigned int)poison
<= 0xFF)
7523 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7525 free_reserved_page(page
);
7529 pr_info("Freeing %s memory: %ldK\n",
7530 s
, pages
<< (PAGE_SHIFT
- 10));
7535 #ifdef CONFIG_HIGHMEM
7536 void free_highmem_page(struct page
*page
)
7538 __free_reserved_page(page
);
7539 totalram_pages_inc();
7540 atomic_long_inc(&page_zone(page
)->managed_pages
);
7541 totalhigh_pages_inc();
7546 void __init
mem_init_print_info(const char *str
)
7548 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7549 unsigned long init_code_size
, init_data_size
;
7551 physpages
= get_num_physpages();
7552 codesize
= _etext
- _stext
;
7553 datasize
= _edata
- _sdata
;
7554 rosize
= __end_rodata
- __start_rodata
;
7555 bss_size
= __bss_stop
- __bss_start
;
7556 init_data_size
= __init_end
- __init_begin
;
7557 init_code_size
= _einittext
- _sinittext
;
7560 * Detect special cases and adjust section sizes accordingly:
7561 * 1) .init.* may be embedded into .data sections
7562 * 2) .init.text.* may be out of [__init_begin, __init_end],
7563 * please refer to arch/tile/kernel/vmlinux.lds.S.
7564 * 3) .rodata.* may be embedded into .text or .data sections.
7566 #define adj_init_size(start, end, size, pos, adj) \
7568 if (start <= pos && pos < end && size > adj) \
7572 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7573 _sinittext
, init_code_size
);
7574 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7575 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7576 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7577 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7579 #undef adj_init_size
7581 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7582 #ifdef CONFIG_HIGHMEM
7586 nr_free_pages() << (PAGE_SHIFT
- 10),
7587 physpages
<< (PAGE_SHIFT
- 10),
7588 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7589 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7590 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7591 totalcma_pages
<< (PAGE_SHIFT
- 10),
7592 #ifdef CONFIG_HIGHMEM
7593 totalhigh_pages() << (PAGE_SHIFT
- 10),
7595 str
? ", " : "", str
? str
: "");
7599 * set_dma_reserve - set the specified number of pages reserved in the first zone
7600 * @new_dma_reserve: The number of pages to mark reserved
7602 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7603 * In the DMA zone, a significant percentage may be consumed by kernel image
7604 * and other unfreeable allocations which can skew the watermarks badly. This
7605 * function may optionally be used to account for unfreeable pages in the
7606 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7607 * smaller per-cpu batchsize.
7609 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7611 dma_reserve
= new_dma_reserve
;
7614 void __init
free_area_init(unsigned long *zones_size
)
7616 init_unavailable_mem();
7617 free_area_init_node(0, zones_size
,
7618 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
7621 static int page_alloc_cpu_dead(unsigned int cpu
)
7624 lru_add_drain_cpu(cpu
);
7628 * Spill the event counters of the dead processor
7629 * into the current processors event counters.
7630 * This artificially elevates the count of the current
7633 vm_events_fold_cpu(cpu
);
7636 * Zero the differential counters of the dead processor
7637 * so that the vm statistics are consistent.
7639 * This is only okay since the processor is dead and cannot
7640 * race with what we are doing.
7642 cpu_vm_stats_fold(cpu
);
7647 int hashdist
= HASHDIST_DEFAULT
;
7649 static int __init
set_hashdist(char *str
)
7653 hashdist
= simple_strtoul(str
, &str
, 0);
7656 __setup("hashdist=", set_hashdist
);
7659 void __init
page_alloc_init(void)
7664 if (num_node_state(N_MEMORY
) == 1)
7668 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7669 "mm/page_alloc:dead", NULL
,
7670 page_alloc_cpu_dead
);
7675 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7676 * or min_free_kbytes changes.
7678 static void calculate_totalreserve_pages(void)
7680 struct pglist_data
*pgdat
;
7681 unsigned long reserve_pages
= 0;
7682 enum zone_type i
, j
;
7684 for_each_online_pgdat(pgdat
) {
7686 pgdat
->totalreserve_pages
= 0;
7688 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7689 struct zone
*zone
= pgdat
->node_zones
+ i
;
7691 unsigned long managed_pages
= zone_managed_pages(zone
);
7693 /* Find valid and maximum lowmem_reserve in the zone */
7694 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7695 if (zone
->lowmem_reserve
[j
] > max
)
7696 max
= zone
->lowmem_reserve
[j
];
7699 /* we treat the high watermark as reserved pages. */
7700 max
+= high_wmark_pages(zone
);
7702 if (max
> managed_pages
)
7703 max
= managed_pages
;
7705 pgdat
->totalreserve_pages
+= max
;
7707 reserve_pages
+= max
;
7710 totalreserve_pages
= reserve_pages
;
7714 * setup_per_zone_lowmem_reserve - called whenever
7715 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7716 * has a correct pages reserved value, so an adequate number of
7717 * pages are left in the zone after a successful __alloc_pages().
7719 static void setup_per_zone_lowmem_reserve(void)
7721 struct pglist_data
*pgdat
;
7722 enum zone_type j
, idx
;
7724 for_each_online_pgdat(pgdat
) {
7725 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7726 struct zone
*zone
= pgdat
->node_zones
+ j
;
7727 unsigned long managed_pages
= zone_managed_pages(zone
);
7729 zone
->lowmem_reserve
[j
] = 0;
7733 struct zone
*lower_zone
;
7736 lower_zone
= pgdat
->node_zones
+ idx
;
7738 if (sysctl_lowmem_reserve_ratio
[idx
] < 1) {
7739 sysctl_lowmem_reserve_ratio
[idx
] = 0;
7740 lower_zone
->lowmem_reserve
[j
] = 0;
7742 lower_zone
->lowmem_reserve
[j
] =
7743 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7745 managed_pages
+= zone_managed_pages(lower_zone
);
7750 /* update totalreserve_pages */
7751 calculate_totalreserve_pages();
7754 static void __setup_per_zone_wmarks(void)
7756 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7757 unsigned long lowmem_pages
= 0;
7759 unsigned long flags
;
7761 /* Calculate total number of !ZONE_HIGHMEM pages */
7762 for_each_zone(zone
) {
7763 if (!is_highmem(zone
))
7764 lowmem_pages
+= zone_managed_pages(zone
);
7767 for_each_zone(zone
) {
7770 spin_lock_irqsave(&zone
->lock
, flags
);
7771 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7772 do_div(tmp
, lowmem_pages
);
7773 if (is_highmem(zone
)) {
7775 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7776 * need highmem pages, so cap pages_min to a small
7779 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7780 * deltas control async page reclaim, and so should
7781 * not be capped for highmem.
7783 unsigned long min_pages
;
7785 min_pages
= zone_managed_pages(zone
) / 1024;
7786 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7787 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7790 * If it's a lowmem zone, reserve a number of pages
7791 * proportionate to the zone's size.
7793 zone
->_watermark
[WMARK_MIN
] = tmp
;
7797 * Set the kswapd watermarks distance according to the
7798 * scale factor in proportion to available memory, but
7799 * ensure a minimum size on small systems.
7801 tmp
= max_t(u64
, tmp
>> 2,
7802 mult_frac(zone_managed_pages(zone
),
7803 watermark_scale_factor
, 10000));
7805 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7806 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7807 zone
->watermark_boost
= 0;
7809 spin_unlock_irqrestore(&zone
->lock
, flags
);
7812 /* update totalreserve_pages */
7813 calculate_totalreserve_pages();
7817 * setup_per_zone_wmarks - called when min_free_kbytes changes
7818 * or when memory is hot-{added|removed}
7820 * Ensures that the watermark[min,low,high] values for each zone are set
7821 * correctly with respect to min_free_kbytes.
7823 void setup_per_zone_wmarks(void)
7825 static DEFINE_SPINLOCK(lock
);
7828 __setup_per_zone_wmarks();
7833 * Initialise min_free_kbytes.
7835 * For small machines we want it small (128k min). For large machines
7836 * we want it large (64MB max). But it is not linear, because network
7837 * bandwidth does not increase linearly with machine size. We use
7839 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7840 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7856 int __meminit
init_per_zone_wmark_min(void)
7858 unsigned long lowmem_kbytes
;
7859 int new_min_free_kbytes
;
7861 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7862 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7864 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7865 min_free_kbytes
= new_min_free_kbytes
;
7866 if (min_free_kbytes
< 128)
7867 min_free_kbytes
= 128;
7868 if (min_free_kbytes
> 262144)
7869 min_free_kbytes
= 262144;
7871 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7872 new_min_free_kbytes
, user_min_free_kbytes
);
7874 setup_per_zone_wmarks();
7875 refresh_zone_stat_thresholds();
7876 setup_per_zone_lowmem_reserve();
7879 setup_min_unmapped_ratio();
7880 setup_min_slab_ratio();
7885 core_initcall(init_per_zone_wmark_min
)
7888 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7889 * that we can call two helper functions whenever min_free_kbytes
7892 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7893 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7897 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7902 user_min_free_kbytes
= min_free_kbytes
;
7903 setup_per_zone_wmarks();
7908 int watermark_boost_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7909 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7913 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7920 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7921 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7925 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7930 setup_per_zone_wmarks();
7936 static void setup_min_unmapped_ratio(void)
7941 for_each_online_pgdat(pgdat
)
7942 pgdat
->min_unmapped_pages
= 0;
7945 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
7946 sysctl_min_unmapped_ratio
) / 100;
7950 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7951 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7955 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7959 setup_min_unmapped_ratio();
7964 static void setup_min_slab_ratio(void)
7969 for_each_online_pgdat(pgdat
)
7970 pgdat
->min_slab_pages
= 0;
7973 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
7974 sysctl_min_slab_ratio
) / 100;
7977 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7978 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7982 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7986 setup_min_slab_ratio();
7993 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7994 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7995 * whenever sysctl_lowmem_reserve_ratio changes.
7997 * The reserve ratio obviously has absolutely no relation with the
7998 * minimum watermarks. The lowmem reserve ratio can only make sense
7999 * if in function of the boot time zone sizes.
8001 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8002 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
8004 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8005 setup_per_zone_lowmem_reserve();
8009 static void __zone_pcp_update(struct zone
*zone
)
8013 for_each_possible_cpu(cpu
)
8014 pageset_set_high_and_batch(zone
,
8015 per_cpu_ptr(zone
->pageset
, cpu
));
8019 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8020 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8021 * pagelist can have before it gets flushed back to buddy allocator.
8023 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
8024 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
8027 int old_percpu_pagelist_fraction
;
8030 mutex_lock(&pcp_batch_high_lock
);
8031 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
8033 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8034 if (!write
|| ret
< 0)
8037 /* Sanity checking to avoid pcp imbalance */
8038 if (percpu_pagelist_fraction
&&
8039 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
8040 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
8046 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
8049 for_each_populated_zone(zone
)
8050 __zone_pcp_update(zone
);
8052 mutex_unlock(&pcp_batch_high_lock
);
8056 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8058 * Returns the number of pages that arch has reserved but
8059 * is not known to alloc_large_system_hash().
8061 static unsigned long __init
arch_reserved_kernel_pages(void)
8068 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8069 * machines. As memory size is increased the scale is also increased but at
8070 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8071 * quadruples the scale is increased by one, which means the size of hash table
8072 * only doubles, instead of quadrupling as well.
8073 * Because 32-bit systems cannot have large physical memory, where this scaling
8074 * makes sense, it is disabled on such platforms.
8076 #if __BITS_PER_LONG > 32
8077 #define ADAPT_SCALE_BASE (64ul << 30)
8078 #define ADAPT_SCALE_SHIFT 2
8079 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8083 * allocate a large system hash table from bootmem
8084 * - it is assumed that the hash table must contain an exact power-of-2
8085 * quantity of entries
8086 * - limit is the number of hash buckets, not the total allocation size
8088 void *__init
alloc_large_system_hash(const char *tablename
,
8089 unsigned long bucketsize
,
8090 unsigned long numentries
,
8093 unsigned int *_hash_shift
,
8094 unsigned int *_hash_mask
,
8095 unsigned long low_limit
,
8096 unsigned long high_limit
)
8098 unsigned long long max
= high_limit
;
8099 unsigned long log2qty
, size
;
8104 /* allow the kernel cmdline to have a say */
8106 /* round applicable memory size up to nearest megabyte */
8107 numentries
= nr_kernel_pages
;
8108 numentries
-= arch_reserved_kernel_pages();
8110 /* It isn't necessary when PAGE_SIZE >= 1MB */
8111 if (PAGE_SHIFT
< 20)
8112 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
8114 #if __BITS_PER_LONG > 32
8116 unsigned long adapt
;
8118 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
8119 adapt
<<= ADAPT_SCALE_SHIFT
)
8124 /* limit to 1 bucket per 2^scale bytes of low memory */
8125 if (scale
> PAGE_SHIFT
)
8126 numentries
>>= (scale
- PAGE_SHIFT
);
8128 numentries
<<= (PAGE_SHIFT
- scale
);
8130 /* Make sure we've got at least a 0-order allocation.. */
8131 if (unlikely(flags
& HASH_SMALL
)) {
8132 /* Makes no sense without HASH_EARLY */
8133 WARN_ON(!(flags
& HASH_EARLY
));
8134 if (!(numentries
>> *_hash_shift
)) {
8135 numentries
= 1UL << *_hash_shift
;
8136 BUG_ON(!numentries
);
8138 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
8139 numentries
= PAGE_SIZE
/ bucketsize
;
8141 numentries
= roundup_pow_of_two(numentries
);
8143 /* limit allocation size to 1/16 total memory by default */
8145 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
8146 do_div(max
, bucketsize
);
8148 max
= min(max
, 0x80000000ULL
);
8150 if (numentries
< low_limit
)
8151 numentries
= low_limit
;
8152 if (numentries
> max
)
8155 log2qty
= ilog2(numentries
);
8157 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
8160 size
= bucketsize
<< log2qty
;
8161 if (flags
& HASH_EARLY
) {
8162 if (flags
& HASH_ZERO
)
8163 table
= memblock_alloc(size
, SMP_CACHE_BYTES
);
8165 table
= memblock_alloc_raw(size
,
8167 } else if (get_order(size
) >= MAX_ORDER
|| hashdist
) {
8168 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
8172 * If bucketsize is not a power-of-two, we may free
8173 * some pages at the end of hash table which
8174 * alloc_pages_exact() automatically does
8176 table
= alloc_pages_exact(size
, gfp_flags
);
8177 kmemleak_alloc(table
, size
, 1, gfp_flags
);
8179 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
8182 panic("Failed to allocate %s hash table\n", tablename
);
8184 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8185 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
,
8186 virt
? "vmalloc" : "linear");
8189 *_hash_shift
= log2qty
;
8191 *_hash_mask
= (1 << log2qty
) - 1;
8197 * This function checks whether pageblock includes unmovable pages or not.
8199 * PageLRU check without isolation or lru_lock could race so that
8200 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8201 * check without lock_page also may miss some movable non-lru pages at
8202 * race condition. So you can't expect this function should be exact.
8204 * Returns a page without holding a reference. If the caller wants to
8205 * dereference that page (e.g., dumping), it has to make sure that that it
8206 * cannot get removed (e.g., via memory unplug) concurrently.
8209 struct page
*has_unmovable_pages(struct zone
*zone
, struct page
*page
,
8210 int migratetype
, int flags
)
8212 unsigned long iter
= 0;
8213 unsigned long pfn
= page_to_pfn(page
);
8216 * TODO we could make this much more efficient by not checking every
8217 * page in the range if we know all of them are in MOVABLE_ZONE and
8218 * that the movable zone guarantees that pages are migratable but
8219 * the later is not the case right now unfortunatelly. E.g. movablecore
8220 * can still lead to having bootmem allocations in zone_movable.
8223 if (is_migrate_cma_page(page
)) {
8225 * CMA allocations (alloc_contig_range) really need to mark
8226 * isolate CMA pageblocks even when they are not movable in fact
8227 * so consider them movable here.
8229 if (is_migrate_cma(migratetype
))
8235 for (; iter
< pageblock_nr_pages
; iter
++) {
8236 if (!pfn_valid_within(pfn
+ iter
))
8239 page
= pfn_to_page(pfn
+ iter
);
8241 if (PageReserved(page
))
8245 * If the zone is movable and we have ruled out all reserved
8246 * pages then it should be reasonably safe to assume the rest
8249 if (zone_idx(zone
) == ZONE_MOVABLE
)
8253 * Hugepages are not in LRU lists, but they're movable.
8254 * THPs are on the LRU, but need to be counted as #small pages.
8255 * We need not scan over tail pages because we don't
8256 * handle each tail page individually in migration.
8258 if (PageHuge(page
) || PageTransCompound(page
)) {
8259 struct page
*head
= compound_head(page
);
8260 unsigned int skip_pages
;
8262 if (PageHuge(page
)) {
8263 if (!hugepage_migration_supported(page_hstate(head
)))
8265 } else if (!PageLRU(head
) && !__PageMovable(head
)) {
8269 skip_pages
= compound_nr(head
) - (page
- head
);
8270 iter
+= skip_pages
- 1;
8275 * We can't use page_count without pin a page
8276 * because another CPU can free compound page.
8277 * This check already skips compound tails of THP
8278 * because their page->_refcount is zero at all time.
8280 if (!page_ref_count(page
)) {
8281 if (PageBuddy(page
))
8282 iter
+= (1 << page_order(page
)) - 1;
8287 * The HWPoisoned page may be not in buddy system, and
8288 * page_count() is not 0.
8290 if ((flags
& MEMORY_OFFLINE
) && PageHWPoison(page
))
8293 if (__PageMovable(page
) || PageLRU(page
))
8297 * If there are RECLAIMABLE pages, we need to check
8298 * it. But now, memory offline itself doesn't call
8299 * shrink_node_slabs() and it still to be fixed.
8302 * If the page is not RAM, page_count()should be 0.
8303 * we don't need more check. This is an _used_ not-movable page.
8305 * The problematic thing here is PG_reserved pages. PG_reserved
8306 * is set to both of a memory hole page and a _used_ kernel
8314 #ifdef CONFIG_CONTIG_ALLOC
8315 static unsigned long pfn_max_align_down(unsigned long pfn
)
8317 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8318 pageblock_nr_pages
) - 1);
8321 static unsigned long pfn_max_align_up(unsigned long pfn
)
8323 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8324 pageblock_nr_pages
));
8327 /* [start, end) must belong to a single zone. */
8328 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8329 unsigned long start
, unsigned long end
)
8331 /* This function is based on compact_zone() from compaction.c. */
8332 unsigned long nr_reclaimed
;
8333 unsigned long pfn
= start
;
8334 unsigned int tries
= 0;
8339 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8340 if (fatal_signal_pending(current
)) {
8345 if (list_empty(&cc
->migratepages
)) {
8346 cc
->nr_migratepages
= 0;
8347 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8353 } else if (++tries
== 5) {
8354 ret
= ret
< 0 ? ret
: -EBUSY
;
8358 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8360 cc
->nr_migratepages
-= nr_reclaimed
;
8362 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
8363 NULL
, 0, cc
->mode
, MR_CONTIG_RANGE
);
8366 putback_movable_pages(&cc
->migratepages
);
8373 * alloc_contig_range() -- tries to allocate given range of pages
8374 * @start: start PFN to allocate
8375 * @end: one-past-the-last PFN to allocate
8376 * @migratetype: migratetype of the underlaying pageblocks (either
8377 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8378 * in range must have the same migratetype and it must
8379 * be either of the two.
8380 * @gfp_mask: GFP mask to use during compaction
8382 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8383 * aligned. The PFN range must belong to a single zone.
8385 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8386 * pageblocks in the range. Once isolated, the pageblocks should not
8387 * be modified by others.
8389 * Return: zero on success or negative error code. On success all
8390 * pages which PFN is in [start, end) are allocated for the caller and
8391 * need to be freed with free_contig_range().
8393 int alloc_contig_range(unsigned long start
, unsigned long end
,
8394 unsigned migratetype
, gfp_t gfp_mask
)
8396 unsigned long outer_start
, outer_end
;
8400 struct compact_control cc
= {
8401 .nr_migratepages
= 0,
8403 .zone
= page_zone(pfn_to_page(start
)),
8404 .mode
= MIGRATE_SYNC
,
8405 .ignore_skip_hint
= true,
8406 .no_set_skip_hint
= true,
8407 .gfp_mask
= current_gfp_context(gfp_mask
),
8408 .alloc_contig
= true,
8410 INIT_LIST_HEAD(&cc
.migratepages
);
8413 * What we do here is we mark all pageblocks in range as
8414 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8415 * have different sizes, and due to the way page allocator
8416 * work, we align the range to biggest of the two pages so
8417 * that page allocator won't try to merge buddies from
8418 * different pageblocks and change MIGRATE_ISOLATE to some
8419 * other migration type.
8421 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8422 * migrate the pages from an unaligned range (ie. pages that
8423 * we are interested in). This will put all the pages in
8424 * range back to page allocator as MIGRATE_ISOLATE.
8426 * When this is done, we take the pages in range from page
8427 * allocator removing them from the buddy system. This way
8428 * page allocator will never consider using them.
8430 * This lets us mark the pageblocks back as
8431 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8432 * aligned range but not in the unaligned, original range are
8433 * put back to page allocator so that buddy can use them.
8436 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8437 pfn_max_align_up(end
), migratetype
, 0);
8442 * In case of -EBUSY, we'd like to know which page causes problem.
8443 * So, just fall through. test_pages_isolated() has a tracepoint
8444 * which will report the busy page.
8446 * It is possible that busy pages could become available before
8447 * the call to test_pages_isolated, and the range will actually be
8448 * allocated. So, if we fall through be sure to clear ret so that
8449 * -EBUSY is not accidentally used or returned to caller.
8451 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8452 if (ret
&& ret
!= -EBUSY
)
8457 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8458 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8459 * more, all pages in [start, end) are free in page allocator.
8460 * What we are going to do is to allocate all pages from
8461 * [start, end) (that is remove them from page allocator).
8463 * The only problem is that pages at the beginning and at the
8464 * end of interesting range may be not aligned with pages that
8465 * page allocator holds, ie. they can be part of higher order
8466 * pages. Because of this, we reserve the bigger range and
8467 * once this is done free the pages we are not interested in.
8469 * We don't have to hold zone->lock here because the pages are
8470 * isolated thus they won't get removed from buddy.
8473 lru_add_drain_all();
8476 outer_start
= start
;
8477 while (!PageBuddy(pfn_to_page(outer_start
))) {
8478 if (++order
>= MAX_ORDER
) {
8479 outer_start
= start
;
8482 outer_start
&= ~0UL << order
;
8485 if (outer_start
!= start
) {
8486 order
= page_order(pfn_to_page(outer_start
));
8489 * outer_start page could be small order buddy page and
8490 * it doesn't include start page. Adjust outer_start
8491 * in this case to report failed page properly
8492 * on tracepoint in test_pages_isolated()
8494 if (outer_start
+ (1UL << order
) <= start
)
8495 outer_start
= start
;
8498 /* Make sure the range is really isolated. */
8499 if (test_pages_isolated(outer_start
, end
, 0)) {
8500 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8501 __func__
, outer_start
, end
);
8506 /* Grab isolated pages from freelists. */
8507 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8513 /* Free head and tail (if any) */
8514 if (start
!= outer_start
)
8515 free_contig_range(outer_start
, start
- outer_start
);
8516 if (end
!= outer_end
)
8517 free_contig_range(end
, outer_end
- end
);
8520 undo_isolate_page_range(pfn_max_align_down(start
),
8521 pfn_max_align_up(end
), migratetype
);
8525 static int __alloc_contig_pages(unsigned long start_pfn
,
8526 unsigned long nr_pages
, gfp_t gfp_mask
)
8528 unsigned long end_pfn
= start_pfn
+ nr_pages
;
8530 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
8534 static bool pfn_range_valid_contig(struct zone
*z
, unsigned long start_pfn
,
8535 unsigned long nr_pages
)
8537 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
8540 for (i
= start_pfn
; i
< end_pfn
; i
++) {
8541 page
= pfn_to_online_page(i
);
8545 if (page_zone(page
) != z
)
8548 if (PageReserved(page
))
8551 if (page_count(page
) > 0)
8560 static bool zone_spans_last_pfn(const struct zone
*zone
,
8561 unsigned long start_pfn
, unsigned long nr_pages
)
8563 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
8565 return zone_spans_pfn(zone
, last_pfn
);
8569 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8570 * @nr_pages: Number of contiguous pages to allocate
8571 * @gfp_mask: GFP mask to limit search and used during compaction
8573 * @nodemask: Mask for other possible nodes
8575 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8576 * on an applicable zonelist to find a contiguous pfn range which can then be
8577 * tried for allocation with alloc_contig_range(). This routine is intended
8578 * for allocation requests which can not be fulfilled with the buddy allocator.
8580 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8581 * power of two then the alignment is guaranteed to be to the given nr_pages
8582 * (e.g. 1GB request would be aligned to 1GB).
8584 * Allocated pages can be freed with free_contig_range() or by manually calling
8585 * __free_page() on each allocated page.
8587 * Return: pointer to contiguous pages on success, or NULL if not successful.
8589 struct page
*alloc_contig_pages(unsigned long nr_pages
, gfp_t gfp_mask
,
8590 int nid
, nodemask_t
*nodemask
)
8592 unsigned long ret
, pfn
, flags
;
8593 struct zonelist
*zonelist
;
8597 zonelist
= node_zonelist(nid
, gfp_mask
);
8598 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
8599 gfp_zone(gfp_mask
), nodemask
) {
8600 spin_lock_irqsave(&zone
->lock
, flags
);
8602 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
8603 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
8604 if (pfn_range_valid_contig(zone
, pfn
, nr_pages
)) {
8606 * We release the zone lock here because
8607 * alloc_contig_range() will also lock the zone
8608 * at some point. If there's an allocation
8609 * spinning on this lock, it may win the race
8610 * and cause alloc_contig_range() to fail...
8612 spin_unlock_irqrestore(&zone
->lock
, flags
);
8613 ret
= __alloc_contig_pages(pfn
, nr_pages
,
8616 return pfn_to_page(pfn
);
8617 spin_lock_irqsave(&zone
->lock
, flags
);
8621 spin_unlock_irqrestore(&zone
->lock
, flags
);
8625 #endif /* CONFIG_CONTIG_ALLOC */
8627 void free_contig_range(unsigned long pfn
, unsigned int nr_pages
)
8629 unsigned int count
= 0;
8631 for (; nr_pages
--; pfn
++) {
8632 struct page
*page
= pfn_to_page(pfn
);
8634 count
+= page_count(page
) != 1;
8637 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8641 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8642 * page high values need to be recalulated.
8644 void __meminit
zone_pcp_update(struct zone
*zone
)
8646 mutex_lock(&pcp_batch_high_lock
);
8647 __zone_pcp_update(zone
);
8648 mutex_unlock(&pcp_batch_high_lock
);
8651 void zone_pcp_reset(struct zone
*zone
)
8653 unsigned long flags
;
8655 struct per_cpu_pageset
*pset
;
8657 /* avoid races with drain_pages() */
8658 local_irq_save(flags
);
8659 if (zone
->pageset
!= &boot_pageset
) {
8660 for_each_online_cpu(cpu
) {
8661 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8662 drain_zonestat(zone
, pset
);
8664 free_percpu(zone
->pageset
);
8665 zone
->pageset
= &boot_pageset
;
8667 local_irq_restore(flags
);
8670 #ifdef CONFIG_MEMORY_HOTREMOVE
8672 * All pages in the range must be in a single zone and isolated
8673 * before calling this.
8676 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8682 unsigned long flags
;
8683 unsigned long offlined_pages
= 0;
8685 /* find the first valid pfn */
8686 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8690 return offlined_pages
;
8692 offline_mem_sections(pfn
, end_pfn
);
8693 zone
= page_zone(pfn_to_page(pfn
));
8694 spin_lock_irqsave(&zone
->lock
, flags
);
8696 while (pfn
< end_pfn
) {
8697 if (!pfn_valid(pfn
)) {
8701 page
= pfn_to_page(pfn
);
8703 * The HWPoisoned page may be not in buddy system, and
8704 * page_count() is not 0.
8706 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8712 BUG_ON(page_count(page
));
8713 BUG_ON(!PageBuddy(page
));
8714 order
= page_order(page
);
8715 offlined_pages
+= 1 << order
;
8716 del_page_from_free_area(page
, &zone
->free_area
[order
]);
8717 pfn
+= (1 << order
);
8719 spin_unlock_irqrestore(&zone
->lock
, flags
);
8721 return offlined_pages
;
8725 bool is_free_buddy_page(struct page
*page
)
8727 struct zone
*zone
= page_zone(page
);
8728 unsigned long pfn
= page_to_pfn(page
);
8729 unsigned long flags
;
8732 spin_lock_irqsave(&zone
->lock
, flags
);
8733 for (order
= 0; order
< MAX_ORDER
; order
++) {
8734 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8736 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8739 spin_unlock_irqrestore(&zone
->lock
, flags
);
8741 return order
< MAX_ORDER
;
8744 #ifdef CONFIG_MEMORY_FAILURE
8746 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8747 * test is performed under the zone lock to prevent a race against page
8750 bool set_hwpoison_free_buddy_page(struct page
*page
)
8752 struct zone
*zone
= page_zone(page
);
8753 unsigned long pfn
= page_to_pfn(page
);
8754 unsigned long flags
;
8756 bool hwpoisoned
= false;
8758 spin_lock_irqsave(&zone
->lock
, flags
);
8759 for (order
= 0; order
< MAX_ORDER
; order
++) {
8760 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8762 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8763 if (!TestSetPageHWPoison(page
))
8768 spin_unlock_irqrestore(&zone
->lock
, flags
);