2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/highmem.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/psi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock
);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node
);
82 EXPORT_PER_CPU_SYMBOL(numa_node
);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
96 int _node_numa_mem_
[MAX_NUMNODES
];
99 /* work_structs for global per-cpu drains */
102 struct work_struct work
;
104 DEFINE_MUTEX(pcpu_drain_mutex
);
105 DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
107 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
108 volatile unsigned long latent_entropy __latent_entropy
;
109 EXPORT_SYMBOL(latent_entropy
);
113 * Array of node states.
115 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
116 [N_POSSIBLE
] = NODE_MASK_ALL
,
117 [N_ONLINE
] = { { [0] = 1UL } },
119 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
120 #ifdef CONFIG_HIGHMEM
121 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
123 [N_MEMORY
] = { { [0] = 1UL } },
124 [N_CPU
] = { { [0] = 1UL } },
127 EXPORT_SYMBOL(node_states
);
129 atomic_long_t _totalram_pages __read_mostly
;
130 EXPORT_SYMBOL(_totalram_pages
);
131 unsigned long totalreserve_pages __read_mostly
;
132 unsigned long totalcma_pages __read_mostly
;
134 int percpu_pagelist_fraction
;
135 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
138 * A cached value of the page's pageblock's migratetype, used when the page is
139 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
140 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
141 * Also the migratetype set in the page does not necessarily match the pcplist
142 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
143 * other index - this ensures that it will be put on the correct CMA freelist.
145 static inline int get_pcppage_migratetype(struct page
*page
)
150 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
152 page
->index
= migratetype
;
155 #ifdef CONFIG_PM_SLEEP
157 * The following functions are used by the suspend/hibernate code to temporarily
158 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
159 * while devices are suspended. To avoid races with the suspend/hibernate code,
160 * they should always be called with system_transition_mutex held
161 * (gfp_allowed_mask also should only be modified with system_transition_mutex
162 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
163 * with that modification).
166 static gfp_t saved_gfp_mask
;
168 void pm_restore_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
171 if (saved_gfp_mask
) {
172 gfp_allowed_mask
= saved_gfp_mask
;
177 void pm_restrict_gfp_mask(void)
179 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
180 WARN_ON(saved_gfp_mask
);
181 saved_gfp_mask
= gfp_allowed_mask
;
182 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
185 bool pm_suspended_storage(void)
187 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
191 #endif /* CONFIG_PM_SLEEP */
193 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
194 unsigned int pageblock_order __read_mostly
;
197 static void __free_pages_ok(struct page
*page
, unsigned int order
);
200 * results with 256, 32 in the lowmem_reserve sysctl:
201 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
202 * 1G machine -> (16M dma, 784M normal, 224M high)
203 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
204 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
205 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
207 * TBD: should special case ZONE_DMA32 machines here - in those we normally
208 * don't need any ZONE_NORMAL reservation
210 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
211 #ifdef CONFIG_ZONE_DMA
214 #ifdef CONFIG_ZONE_DMA32
218 #ifdef CONFIG_HIGHMEM
224 EXPORT_SYMBOL(totalram_pages
);
226 static char * const zone_names
[MAX_NR_ZONES
] = {
227 #ifdef CONFIG_ZONE_DMA
230 #ifdef CONFIG_ZONE_DMA32
234 #ifdef CONFIG_HIGHMEM
238 #ifdef CONFIG_ZONE_DEVICE
243 const char * const migratetype_names
[MIGRATE_TYPES
] = {
251 #ifdef CONFIG_MEMORY_ISOLATION
256 compound_page_dtor
* const compound_page_dtors
[] = {
259 #ifdef CONFIG_HUGETLB_PAGE
262 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
267 int min_free_kbytes
= 1024;
268 int user_min_free_kbytes
= -1;
269 int watermark_boost_factor __read_mostly
= 15000;
270 int watermark_scale_factor
= 10;
272 static unsigned long nr_kernel_pages __initdata
;
273 static unsigned long nr_all_pages __initdata
;
274 static unsigned long dma_reserve __initdata
;
276 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
277 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
278 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
279 static unsigned long required_kernelcore __initdata
;
280 static unsigned long required_kernelcore_percent __initdata
;
281 static unsigned long required_movablecore __initdata
;
282 static unsigned long required_movablecore_percent __initdata
;
283 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
284 static bool mirrored_kernelcore __meminitdata
;
286 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
288 EXPORT_SYMBOL(movable_zone
);
289 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
292 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
293 int nr_online_nodes __read_mostly
= 1;
294 EXPORT_SYMBOL(nr_node_ids
);
295 EXPORT_SYMBOL(nr_online_nodes
);
298 int page_group_by_mobility_disabled __read_mostly
;
300 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
302 * During boot we initialize deferred pages on-demand, as needed, but once
303 * page_alloc_init_late() has finished, the deferred pages are all initialized,
304 * and we can permanently disable that path.
306 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
309 * Calling kasan_free_pages() only after deferred memory initialization
310 * has completed. Poisoning pages during deferred memory init will greatly
311 * lengthen the process and cause problem in large memory systems as the
312 * deferred pages initialization is done with interrupt disabled.
314 * Assuming that there will be no reference to those newly initialized
315 * pages before they are ever allocated, this should have no effect on
316 * KASAN memory tracking as the poison will be properly inserted at page
317 * allocation time. The only corner case is when pages are allocated by
318 * on-demand allocation and then freed again before the deferred pages
319 * initialization is done, but this is not likely to happen.
321 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
323 if (!static_branch_unlikely(&deferred_pages
))
324 kasan_free_pages(page
, order
);
327 /* Returns true if the struct page for the pfn is uninitialised */
328 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
330 int nid
= early_pfn_to_nid(pfn
);
332 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
339 * Returns true when the remaining initialisation should be deferred until
340 * later in the boot cycle when it can be parallelised.
342 static bool __meminit
343 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
345 static unsigned long prev_end_pfn
, nr_initialised
;
348 * prev_end_pfn static that contains the end of previous zone
349 * No need to protect because called very early in boot before smp_init.
351 if (prev_end_pfn
!= end_pfn
) {
352 prev_end_pfn
= end_pfn
;
356 /* Always populate low zones for address-constrained allocations */
357 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
361 * We start only with one section of pages, more pages are added as
362 * needed until the rest of deferred pages are initialized.
365 if ((nr_initialised
> PAGES_PER_SECTION
) &&
366 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
367 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
373 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
375 static inline bool early_page_uninitialised(unsigned long pfn
)
380 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
386 /* Return a pointer to the bitmap storing bits affecting a block of pages */
387 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
390 #ifdef CONFIG_SPARSEMEM
391 return __pfn_to_section(pfn
)->pageblock_flags
;
393 return page_zone(page
)->pageblock_flags
;
394 #endif /* CONFIG_SPARSEMEM */
397 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
399 #ifdef CONFIG_SPARSEMEM
400 pfn
&= (PAGES_PER_SECTION
-1);
401 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
403 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
404 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
405 #endif /* CONFIG_SPARSEMEM */
409 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
410 * @page: The page within the block of interest
411 * @pfn: The target page frame number
412 * @end_bitidx: The last bit of interest to retrieve
413 * @mask: mask of bits that the caller is interested in
415 * Return: pageblock_bits flags
417 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
419 unsigned long end_bitidx
,
422 unsigned long *bitmap
;
423 unsigned long bitidx
, word_bitidx
;
426 bitmap
= get_pageblock_bitmap(page
, pfn
);
427 bitidx
= pfn_to_bitidx(page
, pfn
);
428 word_bitidx
= bitidx
/ BITS_PER_LONG
;
429 bitidx
&= (BITS_PER_LONG
-1);
431 word
= bitmap
[word_bitidx
];
432 bitidx
+= end_bitidx
;
433 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
436 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
437 unsigned long end_bitidx
,
440 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
443 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
445 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
449 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
450 * @page: The page within the block of interest
451 * @flags: The flags to set
452 * @pfn: The target page frame number
453 * @end_bitidx: The last bit of interest
454 * @mask: mask of bits that the caller is interested in
456 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
458 unsigned long end_bitidx
,
461 unsigned long *bitmap
;
462 unsigned long bitidx
, word_bitidx
;
463 unsigned long old_word
, word
;
465 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
466 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
468 bitmap
= get_pageblock_bitmap(page
, pfn
);
469 bitidx
= pfn_to_bitidx(page
, pfn
);
470 word_bitidx
= bitidx
/ BITS_PER_LONG
;
471 bitidx
&= (BITS_PER_LONG
-1);
473 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
475 bitidx
+= end_bitidx
;
476 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
477 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
479 word
= READ_ONCE(bitmap
[word_bitidx
]);
481 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
482 if (word
== old_word
)
488 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
490 if (unlikely(page_group_by_mobility_disabled
&&
491 migratetype
< MIGRATE_PCPTYPES
))
492 migratetype
= MIGRATE_UNMOVABLE
;
494 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
495 PB_migrate
, PB_migrate_end
);
498 #ifdef CONFIG_DEBUG_VM
499 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
503 unsigned long pfn
= page_to_pfn(page
);
504 unsigned long sp
, start_pfn
;
507 seq
= zone_span_seqbegin(zone
);
508 start_pfn
= zone
->zone_start_pfn
;
509 sp
= zone
->spanned_pages
;
510 if (!zone_spans_pfn(zone
, pfn
))
512 } while (zone_span_seqretry(zone
, seq
));
515 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
516 pfn
, zone_to_nid(zone
), zone
->name
,
517 start_pfn
, start_pfn
+ sp
);
522 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
524 if (!pfn_valid_within(page_to_pfn(page
)))
526 if (zone
!= page_zone(page
))
532 * Temporary debugging check for pages not lying within a given zone.
534 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
536 if (page_outside_zone_boundaries(zone
, page
))
538 if (!page_is_consistent(zone
, page
))
544 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
550 static void bad_page(struct page
*page
, const char *reason
,
551 unsigned long bad_flags
)
553 static unsigned long resume
;
554 static unsigned long nr_shown
;
555 static unsigned long nr_unshown
;
558 * Allow a burst of 60 reports, then keep quiet for that minute;
559 * or allow a steady drip of one report per second.
561 if (nr_shown
== 60) {
562 if (time_before(jiffies
, resume
)) {
568 "BUG: Bad page state: %lu messages suppressed\n",
575 resume
= jiffies
+ 60 * HZ
;
577 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
578 current
->comm
, page_to_pfn(page
));
579 __dump_page(page
, reason
);
580 bad_flags
&= page
->flags
;
582 pr_alert("bad because of flags: %#lx(%pGp)\n",
583 bad_flags
, &bad_flags
);
584 dump_page_owner(page
);
589 /* Leave bad fields for debug, except PageBuddy could make trouble */
590 page_mapcount_reset(page
); /* remove PageBuddy */
591 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
595 * Higher-order pages are called "compound pages". They are structured thusly:
597 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
599 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
600 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
602 * The first tail page's ->compound_dtor holds the offset in array of compound
603 * page destructors. See compound_page_dtors.
605 * The first tail page's ->compound_order holds the order of allocation.
606 * This usage means that zero-order pages may not be compound.
609 void free_compound_page(struct page
*page
)
611 __free_pages_ok(page
, compound_order(page
));
614 void prep_compound_page(struct page
*page
, unsigned int order
)
617 int nr_pages
= 1 << order
;
619 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
620 set_compound_order(page
, order
);
622 for (i
= 1; i
< nr_pages
; i
++) {
623 struct page
*p
= page
+ i
;
624 set_page_count(p
, 0);
625 p
->mapping
= TAIL_MAPPING
;
626 set_compound_head(p
, page
);
628 atomic_set(compound_mapcount_ptr(page
), -1);
631 #ifdef CONFIG_DEBUG_PAGEALLOC
632 unsigned int _debug_guardpage_minorder
;
633 bool _debug_pagealloc_enabled __read_mostly
634 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
635 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
636 bool _debug_guardpage_enabled __read_mostly
;
638 static int __init
early_debug_pagealloc(char *buf
)
642 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
644 early_param("debug_pagealloc", early_debug_pagealloc
);
646 static bool need_debug_guardpage(void)
648 /* If we don't use debug_pagealloc, we don't need guard page */
649 if (!debug_pagealloc_enabled())
652 if (!debug_guardpage_minorder())
658 static void init_debug_guardpage(void)
660 if (!debug_pagealloc_enabled())
663 if (!debug_guardpage_minorder())
666 _debug_guardpage_enabled
= true;
669 struct page_ext_operations debug_guardpage_ops
= {
670 .need
= need_debug_guardpage
,
671 .init
= init_debug_guardpage
,
674 static int __init
debug_guardpage_minorder_setup(char *buf
)
678 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
679 pr_err("Bad debug_guardpage_minorder value\n");
682 _debug_guardpage_minorder
= res
;
683 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
686 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
688 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
689 unsigned int order
, int migratetype
)
691 struct page_ext
*page_ext
;
693 if (!debug_guardpage_enabled())
696 if (order
>= debug_guardpage_minorder())
699 page_ext
= lookup_page_ext(page
);
700 if (unlikely(!page_ext
))
703 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
705 INIT_LIST_HEAD(&page
->lru
);
706 set_page_private(page
, order
);
707 /* Guard pages are not available for any usage */
708 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
713 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
714 unsigned int order
, int migratetype
)
716 struct page_ext
*page_ext
;
718 if (!debug_guardpage_enabled())
721 page_ext
= lookup_page_ext(page
);
722 if (unlikely(!page_ext
))
725 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
727 set_page_private(page
, 0);
728 if (!is_migrate_isolate(migratetype
))
729 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
732 struct page_ext_operations debug_guardpage_ops
;
733 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
734 unsigned int order
, int migratetype
) { return false; }
735 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
736 unsigned int order
, int migratetype
) {}
739 static inline void set_page_order(struct page
*page
, unsigned int order
)
741 set_page_private(page
, order
);
742 __SetPageBuddy(page
);
745 static inline void rmv_page_order(struct page
*page
)
747 __ClearPageBuddy(page
);
748 set_page_private(page
, 0);
752 * This function checks whether a page is free && is the buddy
753 * we can coalesce a page and its buddy if
754 * (a) the buddy is not in a hole (check before calling!) &&
755 * (b) the buddy is in the buddy system &&
756 * (c) a page and its buddy have the same order &&
757 * (d) a page and its buddy are in the same zone.
759 * For recording whether a page is in the buddy system, we set PageBuddy.
760 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
762 * For recording page's order, we use page_private(page).
764 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
767 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
768 if (page_zone_id(page
) != page_zone_id(buddy
))
771 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
776 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
778 * zone check is done late to avoid uselessly
779 * calculating zone/node ids for pages that could
782 if (page_zone_id(page
) != page_zone_id(buddy
))
785 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
793 * Freeing function for a buddy system allocator.
795 * The concept of a buddy system is to maintain direct-mapped table
796 * (containing bit values) for memory blocks of various "orders".
797 * The bottom level table contains the map for the smallest allocatable
798 * units of memory (here, pages), and each level above it describes
799 * pairs of units from the levels below, hence, "buddies".
800 * At a high level, all that happens here is marking the table entry
801 * at the bottom level available, and propagating the changes upward
802 * as necessary, plus some accounting needed to play nicely with other
803 * parts of the VM system.
804 * At each level, we keep a list of pages, which are heads of continuous
805 * free pages of length of (1 << order) and marked with PageBuddy.
806 * Page's order is recorded in page_private(page) field.
807 * So when we are allocating or freeing one, we can derive the state of the
808 * other. That is, if we allocate a small block, and both were
809 * free, the remainder of the region must be split into blocks.
810 * If a block is freed, and its buddy is also free, then this
811 * triggers coalescing into a block of larger size.
816 static inline void __free_one_page(struct page
*page
,
818 struct zone
*zone
, unsigned int order
,
821 unsigned long combined_pfn
;
822 unsigned long uninitialized_var(buddy_pfn
);
824 unsigned int max_order
;
826 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
828 VM_BUG_ON(!zone_is_initialized(zone
));
829 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
831 VM_BUG_ON(migratetype
== -1);
832 if (likely(!is_migrate_isolate(migratetype
)))
833 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
835 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
836 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
839 while (order
< max_order
- 1) {
840 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
841 buddy
= page
+ (buddy_pfn
- pfn
);
843 if (!pfn_valid_within(buddy_pfn
))
845 if (!page_is_buddy(page
, buddy
, order
))
848 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
849 * merge with it and move up one order.
851 if (page_is_guard(buddy
)) {
852 clear_page_guard(zone
, buddy
, order
, migratetype
);
854 list_del(&buddy
->lru
);
855 zone
->free_area
[order
].nr_free
--;
856 rmv_page_order(buddy
);
858 combined_pfn
= buddy_pfn
& pfn
;
859 page
= page
+ (combined_pfn
- pfn
);
863 if (max_order
< MAX_ORDER
) {
864 /* If we are here, it means order is >= pageblock_order.
865 * We want to prevent merge between freepages on isolate
866 * pageblock and normal pageblock. Without this, pageblock
867 * isolation could cause incorrect freepage or CMA accounting.
869 * We don't want to hit this code for the more frequent
872 if (unlikely(has_isolate_pageblock(zone
))) {
875 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
876 buddy
= page
+ (buddy_pfn
- pfn
);
877 buddy_mt
= get_pageblock_migratetype(buddy
);
879 if (migratetype
!= buddy_mt
880 && (is_migrate_isolate(migratetype
) ||
881 is_migrate_isolate(buddy_mt
)))
885 goto continue_merging
;
889 set_page_order(page
, order
);
892 * If this is not the largest possible page, check if the buddy
893 * of the next-highest order is free. If it is, it's possible
894 * that pages are being freed that will coalesce soon. In case,
895 * that is happening, add the free page to the tail of the list
896 * so it's less likely to be used soon and more likely to be merged
897 * as a higher order page
899 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
900 struct page
*higher_page
, *higher_buddy
;
901 combined_pfn
= buddy_pfn
& pfn
;
902 higher_page
= page
+ (combined_pfn
- pfn
);
903 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
904 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
905 if (pfn_valid_within(buddy_pfn
) &&
906 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
907 list_add_tail(&page
->lru
,
908 &zone
->free_area
[order
].free_list
[migratetype
]);
913 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
915 zone
->free_area
[order
].nr_free
++;
919 * A bad page could be due to a number of fields. Instead of multiple branches,
920 * try and check multiple fields with one check. The caller must do a detailed
921 * check if necessary.
923 static inline bool page_expected_state(struct page
*page
,
924 unsigned long check_flags
)
926 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
929 if (unlikely((unsigned long)page
->mapping
|
930 page_ref_count(page
) |
932 (unsigned long)page
->mem_cgroup
|
934 (page
->flags
& check_flags
)))
940 static void free_pages_check_bad(struct page
*page
)
942 const char *bad_reason
;
943 unsigned long bad_flags
;
948 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
949 bad_reason
= "nonzero mapcount";
950 if (unlikely(page
->mapping
!= NULL
))
951 bad_reason
= "non-NULL mapping";
952 if (unlikely(page_ref_count(page
) != 0))
953 bad_reason
= "nonzero _refcount";
954 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
955 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
956 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
959 if (unlikely(page
->mem_cgroup
))
960 bad_reason
= "page still charged to cgroup";
962 bad_page(page
, bad_reason
, bad_flags
);
965 static inline int free_pages_check(struct page
*page
)
967 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
970 /* Something has gone sideways, find it */
971 free_pages_check_bad(page
);
975 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
980 * We rely page->lru.next never has bit 0 set, unless the page
981 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
983 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
985 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
989 switch (page
- head_page
) {
991 /* the first tail page: ->mapping may be compound_mapcount() */
992 if (unlikely(compound_mapcount(page
))) {
993 bad_page(page
, "nonzero compound_mapcount", 0);
999 * the second tail page: ->mapping is
1000 * deferred_list.next -- ignore value.
1004 if (page
->mapping
!= TAIL_MAPPING
) {
1005 bad_page(page
, "corrupted mapping in tail page", 0);
1010 if (unlikely(!PageTail(page
))) {
1011 bad_page(page
, "PageTail not set", 0);
1014 if (unlikely(compound_head(page
) != head_page
)) {
1015 bad_page(page
, "compound_head not consistent", 0);
1020 page
->mapping
= NULL
;
1021 clear_compound_head(page
);
1025 static __always_inline
bool free_pages_prepare(struct page
*page
,
1026 unsigned int order
, bool check_free
)
1030 VM_BUG_ON_PAGE(PageTail(page
), page
);
1032 trace_mm_page_free(page
, order
);
1035 * Check tail pages before head page information is cleared to
1036 * avoid checking PageCompound for order-0 pages.
1038 if (unlikely(order
)) {
1039 bool compound
= PageCompound(page
);
1042 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1045 ClearPageDoubleMap(page
);
1046 for (i
= 1; i
< (1 << order
); i
++) {
1048 bad
+= free_tail_pages_check(page
, page
+ i
);
1049 if (unlikely(free_pages_check(page
+ i
))) {
1053 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1056 if (PageMappingFlags(page
))
1057 page
->mapping
= NULL
;
1058 if (memcg_kmem_enabled() && PageKmemcg(page
))
1059 memcg_kmem_uncharge(page
, order
);
1061 bad
+= free_pages_check(page
);
1065 page_cpupid_reset_last(page
);
1066 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1067 reset_page_owner(page
, order
);
1069 if (!PageHighMem(page
)) {
1070 debug_check_no_locks_freed(page_address(page
),
1071 PAGE_SIZE
<< order
);
1072 debug_check_no_obj_freed(page_address(page
),
1073 PAGE_SIZE
<< order
);
1075 arch_free_page(page
, order
);
1076 kernel_poison_pages(page
, 1 << order
, 0);
1077 kernel_map_pages(page
, 1 << order
, 0);
1078 kasan_free_nondeferred_pages(page
, order
);
1083 #ifdef CONFIG_DEBUG_VM
1084 static inline bool free_pcp_prepare(struct page
*page
)
1086 return free_pages_prepare(page
, 0, true);
1089 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1094 static bool free_pcp_prepare(struct page
*page
)
1096 return free_pages_prepare(page
, 0, false);
1099 static bool bulkfree_pcp_prepare(struct page
*page
)
1101 return free_pages_check(page
);
1103 #endif /* CONFIG_DEBUG_VM */
1105 static inline void prefetch_buddy(struct page
*page
)
1107 unsigned long pfn
= page_to_pfn(page
);
1108 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1109 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1115 * Frees a number of pages from the PCP lists
1116 * Assumes all pages on list are in same zone, and of same order.
1117 * count is the number of pages to free.
1119 * If the zone was previously in an "all pages pinned" state then look to
1120 * see if this freeing clears that state.
1122 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1123 * pinned" detection logic.
1125 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1126 struct per_cpu_pages
*pcp
)
1128 int migratetype
= 0;
1130 int prefetch_nr
= 0;
1131 bool isolated_pageblocks
;
1132 struct page
*page
, *tmp
;
1136 struct list_head
*list
;
1139 * Remove pages from lists in a round-robin fashion. A
1140 * batch_free count is maintained that is incremented when an
1141 * empty list is encountered. This is so more pages are freed
1142 * off fuller lists instead of spinning excessively around empty
1147 if (++migratetype
== MIGRATE_PCPTYPES
)
1149 list
= &pcp
->lists
[migratetype
];
1150 } while (list_empty(list
));
1152 /* This is the only non-empty list. Free them all. */
1153 if (batch_free
== MIGRATE_PCPTYPES
)
1157 page
= list_last_entry(list
, struct page
, lru
);
1158 /* must delete to avoid corrupting pcp list */
1159 list_del(&page
->lru
);
1162 if (bulkfree_pcp_prepare(page
))
1165 list_add_tail(&page
->lru
, &head
);
1168 * We are going to put the page back to the global
1169 * pool, prefetch its buddy to speed up later access
1170 * under zone->lock. It is believed the overhead of
1171 * an additional test and calculating buddy_pfn here
1172 * can be offset by reduced memory latency later. To
1173 * avoid excessive prefetching due to large count, only
1174 * prefetch buddy for the first pcp->batch nr of pages.
1176 if (prefetch_nr
++ < pcp
->batch
)
1177 prefetch_buddy(page
);
1178 } while (--count
&& --batch_free
&& !list_empty(list
));
1181 spin_lock(&zone
->lock
);
1182 isolated_pageblocks
= has_isolate_pageblock(zone
);
1185 * Use safe version since after __free_one_page(),
1186 * page->lru.next will not point to original list.
1188 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1189 int mt
= get_pcppage_migratetype(page
);
1190 /* MIGRATE_ISOLATE page should not go to pcplists */
1191 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1192 /* Pageblock could have been isolated meanwhile */
1193 if (unlikely(isolated_pageblocks
))
1194 mt
= get_pageblock_migratetype(page
);
1196 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1197 trace_mm_page_pcpu_drain(page
, 0, mt
);
1199 spin_unlock(&zone
->lock
);
1202 static void free_one_page(struct zone
*zone
,
1203 struct page
*page
, unsigned long pfn
,
1207 spin_lock(&zone
->lock
);
1208 if (unlikely(has_isolate_pageblock(zone
) ||
1209 is_migrate_isolate(migratetype
))) {
1210 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1212 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1213 spin_unlock(&zone
->lock
);
1216 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1217 unsigned long zone
, int nid
)
1219 mm_zero_struct_page(page
);
1220 set_page_links(page
, zone
, nid
, pfn
);
1221 init_page_count(page
);
1222 page_mapcount_reset(page
);
1223 page_cpupid_reset_last(page
);
1224 page_kasan_tag_reset(page
);
1226 INIT_LIST_HEAD(&page
->lru
);
1227 #ifdef WANT_PAGE_VIRTUAL
1228 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1229 if (!is_highmem_idx(zone
))
1230 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1234 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1235 static void __meminit
init_reserved_page(unsigned long pfn
)
1240 if (!early_page_uninitialised(pfn
))
1243 nid
= early_pfn_to_nid(pfn
);
1244 pgdat
= NODE_DATA(nid
);
1246 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1247 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1249 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1252 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1255 static inline void init_reserved_page(unsigned long pfn
)
1258 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1261 * Initialised pages do not have PageReserved set. This function is
1262 * called for each range allocated by the bootmem allocator and
1263 * marks the pages PageReserved. The remaining valid pages are later
1264 * sent to the buddy page allocator.
1266 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1268 unsigned long start_pfn
= PFN_DOWN(start
);
1269 unsigned long end_pfn
= PFN_UP(end
);
1271 for (; start_pfn
< end_pfn
; start_pfn
++) {
1272 if (pfn_valid(start_pfn
)) {
1273 struct page
*page
= pfn_to_page(start_pfn
);
1275 init_reserved_page(start_pfn
);
1277 /* Avoid false-positive PageTail() */
1278 INIT_LIST_HEAD(&page
->lru
);
1281 * no need for atomic set_bit because the struct
1282 * page is not visible yet so nobody should
1285 __SetPageReserved(page
);
1290 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1292 unsigned long flags
;
1294 unsigned long pfn
= page_to_pfn(page
);
1296 if (!free_pages_prepare(page
, order
, true))
1299 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1300 local_irq_save(flags
);
1301 __count_vm_events(PGFREE
, 1 << order
);
1302 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1303 local_irq_restore(flags
);
1306 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1308 unsigned int nr_pages
= 1 << order
;
1309 struct page
*p
= page
;
1313 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1315 __ClearPageReserved(p
);
1316 set_page_count(p
, 0);
1318 __ClearPageReserved(p
);
1319 set_page_count(p
, 0);
1321 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1322 set_page_refcounted(page
);
1323 __free_pages(page
, order
);
1326 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1327 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1329 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1331 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1333 static DEFINE_SPINLOCK(early_pfn_lock
);
1336 spin_lock(&early_pfn_lock
);
1337 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1339 nid
= first_online_node
;
1340 spin_unlock(&early_pfn_lock
);
1346 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1347 static inline bool __meminit __maybe_unused
1348 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1349 struct mminit_pfnnid_cache
*state
)
1353 nid
= __early_pfn_to_nid(pfn
, state
);
1354 if (nid
>= 0 && nid
!= node
)
1359 /* Only safe to use early in boot when initialisation is single-threaded */
1360 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1362 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1367 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1371 static inline bool __meminit __maybe_unused
1372 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1373 struct mminit_pfnnid_cache
*state
)
1380 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1383 if (early_page_uninitialised(pfn
))
1385 return __free_pages_boot_core(page
, order
);
1389 * Check that the whole (or subset of) a pageblock given by the interval of
1390 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1391 * with the migration of free compaction scanner. The scanners then need to
1392 * use only pfn_valid_within() check for arches that allow holes within
1395 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1397 * It's possible on some configurations to have a setup like node0 node1 node0
1398 * i.e. it's possible that all pages within a zones range of pages do not
1399 * belong to a single zone. We assume that a border between node0 and node1
1400 * can occur within a single pageblock, but not a node0 node1 node0
1401 * interleaving within a single pageblock. It is therefore sufficient to check
1402 * the first and last page of a pageblock and avoid checking each individual
1403 * page in a pageblock.
1405 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1406 unsigned long end_pfn
, struct zone
*zone
)
1408 struct page
*start_page
;
1409 struct page
*end_page
;
1411 /* end_pfn is one past the range we are checking */
1414 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1417 start_page
= pfn_to_online_page(start_pfn
);
1421 if (page_zone(start_page
) != zone
)
1424 end_page
= pfn_to_page(end_pfn
);
1426 /* This gives a shorter code than deriving page_zone(end_page) */
1427 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1433 void set_zone_contiguous(struct zone
*zone
)
1435 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1436 unsigned long block_end_pfn
;
1438 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1439 for (; block_start_pfn
< zone_end_pfn(zone
);
1440 block_start_pfn
= block_end_pfn
,
1441 block_end_pfn
+= pageblock_nr_pages
) {
1443 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1445 if (!__pageblock_pfn_to_page(block_start_pfn
,
1446 block_end_pfn
, zone
))
1450 /* We confirm that there is no hole */
1451 zone
->contiguous
= true;
1454 void clear_zone_contiguous(struct zone
*zone
)
1456 zone
->contiguous
= false;
1459 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1460 static void __init
deferred_free_range(unsigned long pfn
,
1461 unsigned long nr_pages
)
1469 page
= pfn_to_page(pfn
);
1471 /* Free a large naturally-aligned chunk if possible */
1472 if (nr_pages
== pageblock_nr_pages
&&
1473 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1474 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1475 __free_pages_boot_core(page
, pageblock_order
);
1479 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1480 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1481 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1482 __free_pages_boot_core(page
, 0);
1486 /* Completion tracking for deferred_init_memmap() threads */
1487 static atomic_t pgdat_init_n_undone __initdata
;
1488 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1490 static inline void __init
pgdat_init_report_one_done(void)
1492 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1493 complete(&pgdat_init_all_done_comp
);
1497 * Returns true if page needs to be initialized or freed to buddy allocator.
1499 * First we check if pfn is valid on architectures where it is possible to have
1500 * holes within pageblock_nr_pages. On systems where it is not possible, this
1501 * function is optimized out.
1503 * Then, we check if a current large page is valid by only checking the validity
1506 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1507 * within a node: a pfn is between start and end of a node, but does not belong
1508 * to this memory node.
1510 static inline bool __init
1511 deferred_pfn_valid(int nid
, unsigned long pfn
,
1512 struct mminit_pfnnid_cache
*nid_init_state
)
1514 if (!pfn_valid_within(pfn
))
1516 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1518 if (!meminit_pfn_in_nid(pfn
, nid
, nid_init_state
))
1524 * Free pages to buddy allocator. Try to free aligned pages in
1525 * pageblock_nr_pages sizes.
1527 static void __init
deferred_free_pages(int nid
, int zid
, unsigned long pfn
,
1528 unsigned long end_pfn
)
1530 struct mminit_pfnnid_cache nid_init_state
= { };
1531 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1532 unsigned long nr_free
= 0;
1534 for (; pfn
< end_pfn
; pfn
++) {
1535 if (!deferred_pfn_valid(nid
, pfn
, &nid_init_state
)) {
1536 deferred_free_range(pfn
- nr_free
, nr_free
);
1538 } else if (!(pfn
& nr_pgmask
)) {
1539 deferred_free_range(pfn
- nr_free
, nr_free
);
1541 touch_nmi_watchdog();
1546 /* Free the last block of pages to allocator */
1547 deferred_free_range(pfn
- nr_free
, nr_free
);
1551 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1552 * by performing it only once every pageblock_nr_pages.
1553 * Return number of pages initialized.
1555 static unsigned long __init
deferred_init_pages(int nid
, int zid
,
1557 unsigned long end_pfn
)
1559 struct mminit_pfnnid_cache nid_init_state
= { };
1560 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1561 unsigned long nr_pages
= 0;
1562 struct page
*page
= NULL
;
1564 for (; pfn
< end_pfn
; pfn
++) {
1565 if (!deferred_pfn_valid(nid
, pfn
, &nid_init_state
)) {
1568 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1569 page
= pfn_to_page(pfn
);
1570 touch_nmi_watchdog();
1574 __init_single_page(page
, pfn
, zid
, nid
);
1580 /* Initialise remaining memory on a node */
1581 static int __init
deferred_init_memmap(void *data
)
1583 pg_data_t
*pgdat
= data
;
1584 int nid
= pgdat
->node_id
;
1585 unsigned long start
= jiffies
;
1586 unsigned long nr_pages
= 0;
1587 unsigned long spfn
, epfn
, first_init_pfn
, flags
;
1588 phys_addr_t spa
, epa
;
1591 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1594 /* Bind memory initialisation thread to a local node if possible */
1595 if (!cpumask_empty(cpumask
))
1596 set_cpus_allowed_ptr(current
, cpumask
);
1598 pgdat_resize_lock(pgdat
, &flags
);
1599 first_init_pfn
= pgdat
->first_deferred_pfn
;
1600 if (first_init_pfn
== ULONG_MAX
) {
1601 pgdat_resize_unlock(pgdat
, &flags
);
1602 pgdat_init_report_one_done();
1606 /* Sanity check boundaries */
1607 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1608 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1609 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1611 /* Only the highest zone is deferred so find it */
1612 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1613 zone
= pgdat
->node_zones
+ zid
;
1614 if (first_init_pfn
< zone_end_pfn(zone
))
1617 first_init_pfn
= max(zone
->zone_start_pfn
, first_init_pfn
);
1620 * Initialize and free pages. We do it in two loops: first we initialize
1621 * struct page, than free to buddy allocator, because while we are
1622 * freeing pages we can access pages that are ahead (computing buddy
1623 * page in __free_one_page()).
1625 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1626 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1627 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1628 nr_pages
+= deferred_init_pages(nid
, zid
, spfn
, epfn
);
1630 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1631 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1632 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1633 deferred_free_pages(nid
, zid
, spfn
, epfn
);
1635 pgdat_resize_unlock(pgdat
, &flags
);
1637 /* Sanity check that the next zone really is unpopulated */
1638 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1640 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1641 jiffies_to_msecs(jiffies
- start
));
1643 pgdat_init_report_one_done();
1648 * If this zone has deferred pages, try to grow it by initializing enough
1649 * deferred pages to satisfy the allocation specified by order, rounded up to
1650 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1651 * of SECTION_SIZE bytes by initializing struct pages in increments of
1652 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1654 * Return true when zone was grown, otherwise return false. We return true even
1655 * when we grow less than requested, to let the caller decide if there are
1656 * enough pages to satisfy the allocation.
1658 * Note: We use noinline because this function is needed only during boot, and
1659 * it is called from a __ref function _deferred_grow_zone. This way we are
1660 * making sure that it is not inlined into permanent text section.
1662 static noinline
bool __init
1663 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1665 int zid
= zone_idx(zone
);
1666 int nid
= zone_to_nid(zone
);
1667 pg_data_t
*pgdat
= NODE_DATA(nid
);
1668 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1669 unsigned long nr_pages
= 0;
1670 unsigned long first_init_pfn
, spfn
, epfn
, t
, flags
;
1671 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1672 phys_addr_t spa
, epa
;
1675 /* Only the last zone may have deferred pages */
1676 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1679 pgdat_resize_lock(pgdat
, &flags
);
1682 * If deferred pages have been initialized while we were waiting for
1683 * the lock, return true, as the zone was grown. The caller will retry
1684 * this zone. We won't return to this function since the caller also
1685 * has this static branch.
1687 if (!static_branch_unlikely(&deferred_pages
)) {
1688 pgdat_resize_unlock(pgdat
, &flags
);
1693 * If someone grew this zone while we were waiting for spinlock, return
1694 * true, as there might be enough pages already.
1696 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1697 pgdat_resize_unlock(pgdat
, &flags
);
1701 first_init_pfn
= max(zone
->zone_start_pfn
, first_deferred_pfn
);
1703 if (first_init_pfn
>= pgdat_end_pfn(pgdat
)) {
1704 pgdat_resize_unlock(pgdat
, &flags
);
1708 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1709 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1710 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1712 while (spfn
< epfn
&& nr_pages
< nr_pages_needed
) {
1713 t
= ALIGN(spfn
+ PAGES_PER_SECTION
, PAGES_PER_SECTION
);
1714 first_deferred_pfn
= min(t
, epfn
);
1715 nr_pages
+= deferred_init_pages(nid
, zid
, spfn
,
1716 first_deferred_pfn
);
1717 spfn
= first_deferred_pfn
;
1720 if (nr_pages
>= nr_pages_needed
)
1724 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1725 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1726 epfn
= min_t(unsigned long, first_deferred_pfn
, PFN_DOWN(epa
));
1727 deferred_free_pages(nid
, zid
, spfn
, epfn
);
1729 if (first_deferred_pfn
== epfn
)
1732 pgdat
->first_deferred_pfn
= first_deferred_pfn
;
1733 pgdat_resize_unlock(pgdat
, &flags
);
1735 return nr_pages
> 0;
1739 * deferred_grow_zone() is __init, but it is called from
1740 * get_page_from_freelist() during early boot until deferred_pages permanently
1741 * disables this call. This is why we have refdata wrapper to avoid warning,
1742 * and to ensure that the function body gets unloaded.
1745 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1747 return deferred_grow_zone(zone
, order
);
1750 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1752 void __init
page_alloc_init_late(void)
1756 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1759 /* There will be num_node_state(N_MEMORY) threads */
1760 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1761 for_each_node_state(nid
, N_MEMORY
) {
1762 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1765 /* Block until all are initialised */
1766 wait_for_completion(&pgdat_init_all_done_comp
);
1769 * We initialized the rest of the deferred pages. Permanently disable
1770 * on-demand struct page initialization.
1772 static_branch_disable(&deferred_pages
);
1774 /* Reinit limits that are based on free pages after the kernel is up */
1775 files_maxfiles_init();
1777 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1778 /* Discard memblock private memory */
1782 for_each_populated_zone(zone
)
1783 set_zone_contiguous(zone
);
1787 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1788 void __init
init_cma_reserved_pageblock(struct page
*page
)
1790 unsigned i
= pageblock_nr_pages
;
1791 struct page
*p
= page
;
1794 __ClearPageReserved(p
);
1795 set_page_count(p
, 0);
1798 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1800 if (pageblock_order
>= MAX_ORDER
) {
1801 i
= pageblock_nr_pages
;
1804 set_page_refcounted(p
);
1805 __free_pages(p
, MAX_ORDER
- 1);
1806 p
+= MAX_ORDER_NR_PAGES
;
1807 } while (i
-= MAX_ORDER_NR_PAGES
);
1809 set_page_refcounted(page
);
1810 __free_pages(page
, pageblock_order
);
1813 adjust_managed_page_count(page
, pageblock_nr_pages
);
1818 * The order of subdivision here is critical for the IO subsystem.
1819 * Please do not alter this order without good reasons and regression
1820 * testing. Specifically, as large blocks of memory are subdivided,
1821 * the order in which smaller blocks are delivered depends on the order
1822 * they're subdivided in this function. This is the primary factor
1823 * influencing the order in which pages are delivered to the IO
1824 * subsystem according to empirical testing, and this is also justified
1825 * by considering the behavior of a buddy system containing a single
1826 * large block of memory acted on by a series of small allocations.
1827 * This behavior is a critical factor in sglist merging's success.
1831 static inline void expand(struct zone
*zone
, struct page
*page
,
1832 int low
, int high
, struct free_area
*area
,
1835 unsigned long size
= 1 << high
;
1837 while (high
> low
) {
1841 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1844 * Mark as guard pages (or page), that will allow to
1845 * merge back to allocator when buddy will be freed.
1846 * Corresponding page table entries will not be touched,
1847 * pages will stay not present in virtual address space
1849 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1852 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1854 set_page_order(&page
[size
], high
);
1858 static void check_new_page_bad(struct page
*page
)
1860 const char *bad_reason
= NULL
;
1861 unsigned long bad_flags
= 0;
1863 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1864 bad_reason
= "nonzero mapcount";
1865 if (unlikely(page
->mapping
!= NULL
))
1866 bad_reason
= "non-NULL mapping";
1867 if (unlikely(page_ref_count(page
) != 0))
1868 bad_reason
= "nonzero _count";
1869 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1870 bad_reason
= "HWPoisoned (hardware-corrupted)";
1871 bad_flags
= __PG_HWPOISON
;
1872 /* Don't complain about hwpoisoned pages */
1873 page_mapcount_reset(page
); /* remove PageBuddy */
1876 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1877 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1878 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1881 if (unlikely(page
->mem_cgroup
))
1882 bad_reason
= "page still charged to cgroup";
1884 bad_page(page
, bad_reason
, bad_flags
);
1888 * This page is about to be returned from the page allocator
1890 static inline int check_new_page(struct page
*page
)
1892 if (likely(page_expected_state(page
,
1893 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1896 check_new_page_bad(page
);
1900 static inline bool free_pages_prezeroed(void)
1902 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1903 page_poisoning_enabled();
1906 #ifdef CONFIG_DEBUG_VM
1907 static bool check_pcp_refill(struct page
*page
)
1912 static bool check_new_pcp(struct page
*page
)
1914 return check_new_page(page
);
1917 static bool check_pcp_refill(struct page
*page
)
1919 return check_new_page(page
);
1921 static bool check_new_pcp(struct page
*page
)
1925 #endif /* CONFIG_DEBUG_VM */
1927 static bool check_new_pages(struct page
*page
, unsigned int order
)
1930 for (i
= 0; i
< (1 << order
); i
++) {
1931 struct page
*p
= page
+ i
;
1933 if (unlikely(check_new_page(p
)))
1940 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1943 set_page_private(page
, 0);
1944 set_page_refcounted(page
);
1946 arch_alloc_page(page
, order
);
1947 kernel_map_pages(page
, 1 << order
, 1);
1948 kernel_poison_pages(page
, 1 << order
, 1);
1949 kasan_alloc_pages(page
, order
);
1950 set_page_owner(page
, order
, gfp_flags
);
1953 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1954 unsigned int alloc_flags
)
1958 post_alloc_hook(page
, order
, gfp_flags
);
1960 if (!free_pages_prezeroed() && (gfp_flags
& __GFP_ZERO
))
1961 for (i
= 0; i
< (1 << order
); i
++)
1962 clear_highpage(page
+ i
);
1964 if (order
&& (gfp_flags
& __GFP_COMP
))
1965 prep_compound_page(page
, order
);
1968 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1969 * allocate the page. The expectation is that the caller is taking
1970 * steps that will free more memory. The caller should avoid the page
1971 * being used for !PFMEMALLOC purposes.
1973 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1974 set_page_pfmemalloc(page
);
1976 clear_page_pfmemalloc(page
);
1980 * Go through the free lists for the given migratetype and remove
1981 * the smallest available page from the freelists
1983 static __always_inline
1984 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1987 unsigned int current_order
;
1988 struct free_area
*area
;
1991 /* Find a page of the appropriate size in the preferred list */
1992 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1993 area
= &(zone
->free_area
[current_order
]);
1994 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1998 list_del(&page
->lru
);
1999 rmv_page_order(page
);
2001 expand(zone
, page
, order
, current_order
, area
, migratetype
);
2002 set_pcppage_migratetype(page
, migratetype
);
2011 * This array describes the order lists are fallen back to when
2012 * the free lists for the desirable migrate type are depleted
2014 static int fallbacks
[MIGRATE_TYPES
][4] = {
2015 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2016 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2017 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2019 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2021 #ifdef CONFIG_MEMORY_ISOLATION
2022 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2027 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2030 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2033 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2034 unsigned int order
) { return NULL
; }
2038 * Move the free pages in a range to the free lists of the requested type.
2039 * Note that start_page and end_pages are not aligned on a pageblock
2040 * boundary. If alignment is required, use move_freepages_block()
2042 static int move_freepages(struct zone
*zone
,
2043 struct page
*start_page
, struct page
*end_page
,
2044 int migratetype
, int *num_movable
)
2048 int pages_moved
= 0;
2050 #ifndef CONFIG_HOLES_IN_ZONE
2052 * page_zone is not safe to call in this context when
2053 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2054 * anyway as we check zone boundaries in move_freepages_block().
2055 * Remove at a later date when no bug reports exist related to
2056 * grouping pages by mobility
2058 VM_BUG_ON(pfn_valid(page_to_pfn(start_page
)) &&
2059 pfn_valid(page_to_pfn(end_page
)) &&
2060 page_zone(start_page
) != page_zone(end_page
));
2062 for (page
= start_page
; page
<= end_page
;) {
2063 if (!pfn_valid_within(page_to_pfn(page
))) {
2068 /* Make sure we are not inadvertently changing nodes */
2069 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2071 if (!PageBuddy(page
)) {
2073 * We assume that pages that could be isolated for
2074 * migration are movable. But we don't actually try
2075 * isolating, as that would be expensive.
2078 (PageLRU(page
) || __PageMovable(page
)))
2085 order
= page_order(page
);
2086 list_move(&page
->lru
,
2087 &zone
->free_area
[order
].free_list
[migratetype
]);
2089 pages_moved
+= 1 << order
;
2095 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2096 int migratetype
, int *num_movable
)
2098 unsigned long start_pfn
, end_pfn
;
2099 struct page
*start_page
, *end_page
;
2104 start_pfn
= page_to_pfn(page
);
2105 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2106 start_page
= pfn_to_page(start_pfn
);
2107 end_page
= start_page
+ pageblock_nr_pages
- 1;
2108 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2110 /* Do not cross zone boundaries */
2111 if (!zone_spans_pfn(zone
, start_pfn
))
2113 if (!zone_spans_pfn(zone
, end_pfn
))
2116 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2120 static void change_pageblock_range(struct page
*pageblock_page
,
2121 int start_order
, int migratetype
)
2123 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2125 while (nr_pageblocks
--) {
2126 set_pageblock_migratetype(pageblock_page
, migratetype
);
2127 pageblock_page
+= pageblock_nr_pages
;
2132 * When we are falling back to another migratetype during allocation, try to
2133 * steal extra free pages from the same pageblocks to satisfy further
2134 * allocations, instead of polluting multiple pageblocks.
2136 * If we are stealing a relatively large buddy page, it is likely there will
2137 * be more free pages in the pageblock, so try to steal them all. For
2138 * reclaimable and unmovable allocations, we steal regardless of page size,
2139 * as fragmentation caused by those allocations polluting movable pageblocks
2140 * is worse than movable allocations stealing from unmovable and reclaimable
2143 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2146 * Leaving this order check is intended, although there is
2147 * relaxed order check in next check. The reason is that
2148 * we can actually steal whole pageblock if this condition met,
2149 * but, below check doesn't guarantee it and that is just heuristic
2150 * so could be changed anytime.
2152 if (order
>= pageblock_order
)
2155 if (order
>= pageblock_order
/ 2 ||
2156 start_mt
== MIGRATE_RECLAIMABLE
||
2157 start_mt
== MIGRATE_UNMOVABLE
||
2158 page_group_by_mobility_disabled
)
2164 static inline void boost_watermark(struct zone
*zone
)
2166 unsigned long max_boost
;
2168 if (!watermark_boost_factor
)
2171 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2172 watermark_boost_factor
, 10000);
2173 max_boost
= max(pageblock_nr_pages
, max_boost
);
2175 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2180 * This function implements actual steal behaviour. If order is large enough,
2181 * we can steal whole pageblock. If not, we first move freepages in this
2182 * pageblock to our migratetype and determine how many already-allocated pages
2183 * are there in the pageblock with a compatible migratetype. If at least half
2184 * of pages are free or compatible, we can change migratetype of the pageblock
2185 * itself, so pages freed in the future will be put on the correct free list.
2187 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2188 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2190 unsigned int current_order
= page_order(page
);
2191 struct free_area
*area
;
2192 int free_pages
, movable_pages
, alike_pages
;
2195 old_block_type
= get_pageblock_migratetype(page
);
2198 * This can happen due to races and we want to prevent broken
2199 * highatomic accounting.
2201 if (is_migrate_highatomic(old_block_type
))
2204 /* Take ownership for orders >= pageblock_order */
2205 if (current_order
>= pageblock_order
) {
2206 change_pageblock_range(page
, current_order
, start_type
);
2211 * Boost watermarks to increase reclaim pressure to reduce the
2212 * likelihood of future fallbacks. Wake kswapd now as the node
2213 * may be balanced overall and kswapd will not wake naturally.
2215 boost_watermark(zone
);
2216 if (alloc_flags
& ALLOC_KSWAPD
)
2217 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2219 /* We are not allowed to try stealing from the whole block */
2223 free_pages
= move_freepages_block(zone
, page
, start_type
,
2226 * Determine how many pages are compatible with our allocation.
2227 * For movable allocation, it's the number of movable pages which
2228 * we just obtained. For other types it's a bit more tricky.
2230 if (start_type
== MIGRATE_MOVABLE
) {
2231 alike_pages
= movable_pages
;
2234 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2235 * to MOVABLE pageblock, consider all non-movable pages as
2236 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2237 * vice versa, be conservative since we can't distinguish the
2238 * exact migratetype of non-movable pages.
2240 if (old_block_type
== MIGRATE_MOVABLE
)
2241 alike_pages
= pageblock_nr_pages
2242 - (free_pages
+ movable_pages
);
2247 /* moving whole block can fail due to zone boundary conditions */
2252 * If a sufficient number of pages in the block are either free or of
2253 * comparable migratability as our allocation, claim the whole block.
2255 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2256 page_group_by_mobility_disabled
)
2257 set_pageblock_migratetype(page
, start_type
);
2262 area
= &zone
->free_area
[current_order
];
2263 list_move(&page
->lru
, &area
->free_list
[start_type
]);
2267 * Check whether there is a suitable fallback freepage with requested order.
2268 * If only_stealable is true, this function returns fallback_mt only if
2269 * we can steal other freepages all together. This would help to reduce
2270 * fragmentation due to mixed migratetype pages in one pageblock.
2272 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2273 int migratetype
, bool only_stealable
, bool *can_steal
)
2278 if (area
->nr_free
== 0)
2283 fallback_mt
= fallbacks
[migratetype
][i
];
2284 if (fallback_mt
== MIGRATE_TYPES
)
2287 if (list_empty(&area
->free_list
[fallback_mt
]))
2290 if (can_steal_fallback(order
, migratetype
))
2293 if (!only_stealable
)
2304 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2305 * there are no empty page blocks that contain a page with a suitable order
2307 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2308 unsigned int alloc_order
)
2311 unsigned long max_managed
, flags
;
2314 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2315 * Check is race-prone but harmless.
2317 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2318 if (zone
->nr_reserved_highatomic
>= max_managed
)
2321 spin_lock_irqsave(&zone
->lock
, flags
);
2323 /* Recheck the nr_reserved_highatomic limit under the lock */
2324 if (zone
->nr_reserved_highatomic
>= max_managed
)
2328 mt
= get_pageblock_migratetype(page
);
2329 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2330 && !is_migrate_cma(mt
)) {
2331 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2332 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2333 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2337 spin_unlock_irqrestore(&zone
->lock
, flags
);
2341 * Used when an allocation is about to fail under memory pressure. This
2342 * potentially hurts the reliability of high-order allocations when under
2343 * intense memory pressure but failed atomic allocations should be easier
2344 * to recover from than an OOM.
2346 * If @force is true, try to unreserve a pageblock even though highatomic
2347 * pageblock is exhausted.
2349 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2352 struct zonelist
*zonelist
= ac
->zonelist
;
2353 unsigned long flags
;
2360 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2363 * Preserve at least one pageblock unless memory pressure
2366 if (!force
&& zone
->nr_reserved_highatomic
<=
2370 spin_lock_irqsave(&zone
->lock
, flags
);
2371 for (order
= 0; order
< MAX_ORDER
; order
++) {
2372 struct free_area
*area
= &(zone
->free_area
[order
]);
2374 page
= list_first_entry_or_null(
2375 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2381 * In page freeing path, migratetype change is racy so
2382 * we can counter several free pages in a pageblock
2383 * in this loop althoug we changed the pageblock type
2384 * from highatomic to ac->migratetype. So we should
2385 * adjust the count once.
2387 if (is_migrate_highatomic_page(page
)) {
2389 * It should never happen but changes to
2390 * locking could inadvertently allow a per-cpu
2391 * drain to add pages to MIGRATE_HIGHATOMIC
2392 * while unreserving so be safe and watch for
2395 zone
->nr_reserved_highatomic
-= min(
2397 zone
->nr_reserved_highatomic
);
2401 * Convert to ac->migratetype and avoid the normal
2402 * pageblock stealing heuristics. Minimally, the caller
2403 * is doing the work and needs the pages. More
2404 * importantly, if the block was always converted to
2405 * MIGRATE_UNMOVABLE or another type then the number
2406 * of pageblocks that cannot be completely freed
2409 set_pageblock_migratetype(page
, ac
->migratetype
);
2410 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2413 spin_unlock_irqrestore(&zone
->lock
, flags
);
2417 spin_unlock_irqrestore(&zone
->lock
, flags
);
2424 * Try finding a free buddy page on the fallback list and put it on the free
2425 * list of requested migratetype, possibly along with other pages from the same
2426 * block, depending on fragmentation avoidance heuristics. Returns true if
2427 * fallback was found so that __rmqueue_smallest() can grab it.
2429 * The use of signed ints for order and current_order is a deliberate
2430 * deviation from the rest of this file, to make the for loop
2431 * condition simpler.
2433 static __always_inline
bool
2434 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2435 unsigned int alloc_flags
)
2437 struct free_area
*area
;
2439 int min_order
= order
;
2445 * Do not steal pages from freelists belonging to other pageblocks
2446 * i.e. orders < pageblock_order. If there are no local zones free,
2447 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2449 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2450 min_order
= pageblock_order
;
2453 * Find the largest available free page in the other list. This roughly
2454 * approximates finding the pageblock with the most free pages, which
2455 * would be too costly to do exactly.
2457 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2459 area
= &(zone
->free_area
[current_order
]);
2460 fallback_mt
= find_suitable_fallback(area
, current_order
,
2461 start_migratetype
, false, &can_steal
);
2462 if (fallback_mt
== -1)
2466 * We cannot steal all free pages from the pageblock and the
2467 * requested migratetype is movable. In that case it's better to
2468 * steal and split the smallest available page instead of the
2469 * largest available page, because even if the next movable
2470 * allocation falls back into a different pageblock than this
2471 * one, it won't cause permanent fragmentation.
2473 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2474 && current_order
> order
)
2483 for (current_order
= order
; current_order
< MAX_ORDER
;
2485 area
= &(zone
->free_area
[current_order
]);
2486 fallback_mt
= find_suitable_fallback(area
, current_order
,
2487 start_migratetype
, false, &can_steal
);
2488 if (fallback_mt
!= -1)
2493 * This should not happen - we already found a suitable fallback
2494 * when looking for the largest page.
2496 VM_BUG_ON(current_order
== MAX_ORDER
);
2499 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2502 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2505 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2506 start_migratetype
, fallback_mt
);
2513 * Do the hard work of removing an element from the buddy allocator.
2514 * Call me with the zone->lock already held.
2516 static __always_inline
struct page
*
2517 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2518 unsigned int alloc_flags
)
2523 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2524 if (unlikely(!page
)) {
2525 if (migratetype
== MIGRATE_MOVABLE
)
2526 page
= __rmqueue_cma_fallback(zone
, order
);
2528 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2533 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2538 * Obtain a specified number of elements from the buddy allocator, all under
2539 * a single hold of the lock, for efficiency. Add them to the supplied list.
2540 * Returns the number of new pages which were placed at *list.
2542 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2543 unsigned long count
, struct list_head
*list
,
2544 int migratetype
, unsigned int alloc_flags
)
2548 spin_lock(&zone
->lock
);
2549 for (i
= 0; i
< count
; ++i
) {
2550 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2552 if (unlikely(page
== NULL
))
2555 if (unlikely(check_pcp_refill(page
)))
2559 * Split buddy pages returned by expand() are received here in
2560 * physical page order. The page is added to the tail of
2561 * caller's list. From the callers perspective, the linked list
2562 * is ordered by page number under some conditions. This is
2563 * useful for IO devices that can forward direction from the
2564 * head, thus also in the physical page order. This is useful
2565 * for IO devices that can merge IO requests if the physical
2566 * pages are ordered properly.
2568 list_add_tail(&page
->lru
, list
);
2570 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2571 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2576 * i pages were removed from the buddy list even if some leak due
2577 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2578 * on i. Do not confuse with 'alloced' which is the number of
2579 * pages added to the pcp list.
2581 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2582 spin_unlock(&zone
->lock
);
2588 * Called from the vmstat counter updater to drain pagesets of this
2589 * currently executing processor on remote nodes after they have
2592 * Note that this function must be called with the thread pinned to
2593 * a single processor.
2595 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2597 unsigned long flags
;
2598 int to_drain
, batch
;
2600 local_irq_save(flags
);
2601 batch
= READ_ONCE(pcp
->batch
);
2602 to_drain
= min(pcp
->count
, batch
);
2604 free_pcppages_bulk(zone
, to_drain
, pcp
);
2605 local_irq_restore(flags
);
2610 * Drain pcplists of the indicated processor and zone.
2612 * The processor must either be the current processor and the
2613 * thread pinned to the current processor or a processor that
2616 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2618 unsigned long flags
;
2619 struct per_cpu_pageset
*pset
;
2620 struct per_cpu_pages
*pcp
;
2622 local_irq_save(flags
);
2623 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2627 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2628 local_irq_restore(flags
);
2632 * Drain pcplists of all zones on the indicated processor.
2634 * The processor must either be the current processor and the
2635 * thread pinned to the current processor or a processor that
2638 static void drain_pages(unsigned int cpu
)
2642 for_each_populated_zone(zone
) {
2643 drain_pages_zone(cpu
, zone
);
2648 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2650 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2651 * the single zone's pages.
2653 void drain_local_pages(struct zone
*zone
)
2655 int cpu
= smp_processor_id();
2658 drain_pages_zone(cpu
, zone
);
2663 static void drain_local_pages_wq(struct work_struct
*work
)
2665 struct pcpu_drain
*drain
;
2667 drain
= container_of(work
, struct pcpu_drain
, work
);
2670 * drain_all_pages doesn't use proper cpu hotplug protection so
2671 * we can race with cpu offline when the WQ can move this from
2672 * a cpu pinned worker to an unbound one. We can operate on a different
2673 * cpu which is allright but we also have to make sure to not move to
2677 drain_local_pages(drain
->zone
);
2682 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2684 * When zone parameter is non-NULL, spill just the single zone's pages.
2686 * Note that this can be extremely slow as the draining happens in a workqueue.
2688 void drain_all_pages(struct zone
*zone
)
2693 * Allocate in the BSS so we wont require allocation in
2694 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2696 static cpumask_t cpus_with_pcps
;
2699 * Make sure nobody triggers this path before mm_percpu_wq is fully
2702 if (WARN_ON_ONCE(!mm_percpu_wq
))
2706 * Do not drain if one is already in progress unless it's specific to
2707 * a zone. Such callers are primarily CMA and memory hotplug and need
2708 * the drain to be complete when the call returns.
2710 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2713 mutex_lock(&pcpu_drain_mutex
);
2717 * We don't care about racing with CPU hotplug event
2718 * as offline notification will cause the notified
2719 * cpu to drain that CPU pcps and on_each_cpu_mask
2720 * disables preemption as part of its processing
2722 for_each_online_cpu(cpu
) {
2723 struct per_cpu_pageset
*pcp
;
2725 bool has_pcps
= false;
2728 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2732 for_each_populated_zone(z
) {
2733 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2734 if (pcp
->pcp
.count
) {
2742 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2744 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2747 for_each_cpu(cpu
, &cpus_with_pcps
) {
2748 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
2751 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
2752 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
2754 for_each_cpu(cpu
, &cpus_with_pcps
)
2755 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
2757 mutex_unlock(&pcpu_drain_mutex
);
2760 #ifdef CONFIG_HIBERNATION
2763 * Touch the watchdog for every WD_PAGE_COUNT pages.
2765 #define WD_PAGE_COUNT (128*1024)
2767 void mark_free_pages(struct zone
*zone
)
2769 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2770 unsigned long flags
;
2771 unsigned int order
, t
;
2774 if (zone_is_empty(zone
))
2777 spin_lock_irqsave(&zone
->lock
, flags
);
2779 max_zone_pfn
= zone_end_pfn(zone
);
2780 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2781 if (pfn_valid(pfn
)) {
2782 page
= pfn_to_page(pfn
);
2784 if (!--page_count
) {
2785 touch_nmi_watchdog();
2786 page_count
= WD_PAGE_COUNT
;
2789 if (page_zone(page
) != zone
)
2792 if (!swsusp_page_is_forbidden(page
))
2793 swsusp_unset_page_free(page
);
2796 for_each_migratetype_order(order
, t
) {
2797 list_for_each_entry(page
,
2798 &zone
->free_area
[order
].free_list
[t
], lru
) {
2801 pfn
= page_to_pfn(page
);
2802 for (i
= 0; i
< (1UL << order
); i
++) {
2803 if (!--page_count
) {
2804 touch_nmi_watchdog();
2805 page_count
= WD_PAGE_COUNT
;
2807 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2811 spin_unlock_irqrestore(&zone
->lock
, flags
);
2813 #endif /* CONFIG_PM */
2815 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
2819 if (!free_pcp_prepare(page
))
2822 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2823 set_pcppage_migratetype(page
, migratetype
);
2827 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
2829 struct zone
*zone
= page_zone(page
);
2830 struct per_cpu_pages
*pcp
;
2833 migratetype
= get_pcppage_migratetype(page
);
2834 __count_vm_event(PGFREE
);
2837 * We only track unmovable, reclaimable and movable on pcp lists.
2838 * Free ISOLATE pages back to the allocator because they are being
2839 * offlined but treat HIGHATOMIC as movable pages so we can get those
2840 * areas back if necessary. Otherwise, we may have to free
2841 * excessively into the page allocator
2843 if (migratetype
>= MIGRATE_PCPTYPES
) {
2844 if (unlikely(is_migrate_isolate(migratetype
))) {
2845 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2848 migratetype
= MIGRATE_MOVABLE
;
2851 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2852 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2854 if (pcp
->count
>= pcp
->high
) {
2855 unsigned long batch
= READ_ONCE(pcp
->batch
);
2856 free_pcppages_bulk(zone
, batch
, pcp
);
2861 * Free a 0-order page
2863 void free_unref_page(struct page
*page
)
2865 unsigned long flags
;
2866 unsigned long pfn
= page_to_pfn(page
);
2868 if (!free_unref_page_prepare(page
, pfn
))
2871 local_irq_save(flags
);
2872 free_unref_page_commit(page
, pfn
);
2873 local_irq_restore(flags
);
2877 * Free a list of 0-order pages
2879 void free_unref_page_list(struct list_head
*list
)
2881 struct page
*page
, *next
;
2882 unsigned long flags
, pfn
;
2883 int batch_count
= 0;
2885 /* Prepare pages for freeing */
2886 list_for_each_entry_safe(page
, next
, list
, lru
) {
2887 pfn
= page_to_pfn(page
);
2888 if (!free_unref_page_prepare(page
, pfn
))
2889 list_del(&page
->lru
);
2890 set_page_private(page
, pfn
);
2893 local_irq_save(flags
);
2894 list_for_each_entry_safe(page
, next
, list
, lru
) {
2895 unsigned long pfn
= page_private(page
);
2897 set_page_private(page
, 0);
2898 trace_mm_page_free_batched(page
);
2899 free_unref_page_commit(page
, pfn
);
2902 * Guard against excessive IRQ disabled times when we get
2903 * a large list of pages to free.
2905 if (++batch_count
== SWAP_CLUSTER_MAX
) {
2906 local_irq_restore(flags
);
2908 local_irq_save(flags
);
2911 local_irq_restore(flags
);
2915 * split_page takes a non-compound higher-order page, and splits it into
2916 * n (1<<order) sub-pages: page[0..n]
2917 * Each sub-page must be freed individually.
2919 * Note: this is probably too low level an operation for use in drivers.
2920 * Please consult with lkml before using this in your driver.
2922 void split_page(struct page
*page
, unsigned int order
)
2926 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2927 VM_BUG_ON_PAGE(!page_count(page
), page
);
2929 for (i
= 1; i
< (1 << order
); i
++)
2930 set_page_refcounted(page
+ i
);
2931 split_page_owner(page
, order
);
2933 EXPORT_SYMBOL_GPL(split_page
);
2935 int __isolate_free_page(struct page
*page
, unsigned int order
)
2937 unsigned long watermark
;
2941 BUG_ON(!PageBuddy(page
));
2943 zone
= page_zone(page
);
2944 mt
= get_pageblock_migratetype(page
);
2946 if (!is_migrate_isolate(mt
)) {
2948 * Obey watermarks as if the page was being allocated. We can
2949 * emulate a high-order watermark check with a raised order-0
2950 * watermark, because we already know our high-order page
2953 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2954 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2957 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2960 /* Remove page from free list */
2961 list_del(&page
->lru
);
2962 zone
->free_area
[order
].nr_free
--;
2963 rmv_page_order(page
);
2966 * Set the pageblock if the isolated page is at least half of a
2969 if (order
>= pageblock_order
- 1) {
2970 struct page
*endpage
= page
+ (1 << order
) - 1;
2971 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2972 int mt
= get_pageblock_migratetype(page
);
2973 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2974 && !is_migrate_highatomic(mt
))
2975 set_pageblock_migratetype(page
,
2981 return 1UL << order
;
2985 * Update NUMA hit/miss statistics
2987 * Must be called with interrupts disabled.
2989 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2992 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2994 /* skip numa counters update if numa stats is disabled */
2995 if (!static_branch_likely(&vm_numa_stat_key
))
2998 if (zone_to_nid(z
) != numa_node_id())
2999 local_stat
= NUMA_OTHER
;
3001 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3002 __inc_numa_state(z
, NUMA_HIT
);
3004 __inc_numa_state(z
, NUMA_MISS
);
3005 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3007 __inc_numa_state(z
, local_stat
);
3011 /* Remove page from the per-cpu list, caller must protect the list */
3012 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3013 unsigned int alloc_flags
,
3014 struct per_cpu_pages
*pcp
,
3015 struct list_head
*list
)
3020 if (list_empty(list
)) {
3021 pcp
->count
+= rmqueue_bulk(zone
, 0,
3023 migratetype
, alloc_flags
);
3024 if (unlikely(list_empty(list
)))
3028 page
= list_first_entry(list
, struct page
, lru
);
3029 list_del(&page
->lru
);
3031 } while (check_new_pcp(page
));
3036 /* Lock and remove page from the per-cpu list */
3037 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3038 struct zone
*zone
, unsigned int order
,
3039 gfp_t gfp_flags
, int migratetype
,
3040 unsigned int alloc_flags
)
3042 struct per_cpu_pages
*pcp
;
3043 struct list_head
*list
;
3045 unsigned long flags
;
3047 local_irq_save(flags
);
3048 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3049 list
= &pcp
->lists
[migratetype
];
3050 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3052 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3053 zone_statistics(preferred_zone
, zone
);
3055 local_irq_restore(flags
);
3060 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3063 struct page
*rmqueue(struct zone
*preferred_zone
,
3064 struct zone
*zone
, unsigned int order
,
3065 gfp_t gfp_flags
, unsigned int alloc_flags
,
3068 unsigned long flags
;
3071 if (likely(order
== 0)) {
3072 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
3073 gfp_flags
, migratetype
, alloc_flags
);
3078 * We most definitely don't want callers attempting to
3079 * allocate greater than order-1 page units with __GFP_NOFAIL.
3081 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3082 spin_lock_irqsave(&zone
->lock
, flags
);
3086 if (alloc_flags
& ALLOC_HARDER
) {
3087 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3089 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3092 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3093 } while (page
&& check_new_pages(page
, order
));
3094 spin_unlock(&zone
->lock
);
3097 __mod_zone_freepage_state(zone
, -(1 << order
),
3098 get_pcppage_migratetype(page
));
3100 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3101 zone_statistics(preferred_zone
, zone
);
3102 local_irq_restore(flags
);
3105 /* Separate test+clear to avoid unnecessary atomics */
3106 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3107 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3108 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3111 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3115 local_irq_restore(flags
);
3119 #ifdef CONFIG_FAIL_PAGE_ALLOC
3122 struct fault_attr attr
;
3124 bool ignore_gfp_highmem
;
3125 bool ignore_gfp_reclaim
;
3127 } fail_page_alloc
= {
3128 .attr
= FAULT_ATTR_INITIALIZER
,
3129 .ignore_gfp_reclaim
= true,
3130 .ignore_gfp_highmem
= true,
3134 static int __init
setup_fail_page_alloc(char *str
)
3136 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3138 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3140 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3142 if (order
< fail_page_alloc
.min_order
)
3144 if (gfp_mask
& __GFP_NOFAIL
)
3146 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3148 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3149 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3152 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3155 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3157 static int __init
fail_page_alloc_debugfs(void)
3159 umode_t mode
= S_IFREG
| 0600;
3162 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3163 &fail_page_alloc
.attr
);
3165 return PTR_ERR(dir
);
3167 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3168 &fail_page_alloc
.ignore_gfp_reclaim
))
3170 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3171 &fail_page_alloc
.ignore_gfp_highmem
))
3173 if (!debugfs_create_u32("min-order", mode
, dir
,
3174 &fail_page_alloc
.min_order
))
3179 debugfs_remove_recursive(dir
);
3184 late_initcall(fail_page_alloc_debugfs
);
3186 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3188 #else /* CONFIG_FAIL_PAGE_ALLOC */
3190 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3195 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3197 static noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3199 return __should_fail_alloc_page(gfp_mask
, order
);
3201 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3204 * Return true if free base pages are above 'mark'. For high-order checks it
3205 * will return true of the order-0 watermark is reached and there is at least
3206 * one free page of a suitable size. Checking now avoids taking the zone lock
3207 * to check in the allocation paths if no pages are free.
3209 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3210 int classzone_idx
, unsigned int alloc_flags
,
3215 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3217 /* free_pages may go negative - that's OK */
3218 free_pages
-= (1 << order
) - 1;
3220 if (alloc_flags
& ALLOC_HIGH
)
3224 * If the caller does not have rights to ALLOC_HARDER then subtract
3225 * the high-atomic reserves. This will over-estimate the size of the
3226 * atomic reserve but it avoids a search.
3228 if (likely(!alloc_harder
)) {
3229 free_pages
-= z
->nr_reserved_highatomic
;
3232 * OOM victims can try even harder than normal ALLOC_HARDER
3233 * users on the grounds that it's definitely going to be in
3234 * the exit path shortly and free memory. Any allocation it
3235 * makes during the free path will be small and short-lived.
3237 if (alloc_flags
& ALLOC_OOM
)
3245 /* If allocation can't use CMA areas don't use free CMA pages */
3246 if (!(alloc_flags
& ALLOC_CMA
))
3247 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3251 * Check watermarks for an order-0 allocation request. If these
3252 * are not met, then a high-order request also cannot go ahead
3253 * even if a suitable page happened to be free.
3255 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3258 /* If this is an order-0 request then the watermark is fine */
3262 /* For a high-order request, check at least one suitable page is free */
3263 for (o
= order
; o
< MAX_ORDER
; o
++) {
3264 struct free_area
*area
= &z
->free_area
[o
];
3270 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3271 if (!list_empty(&area
->free_list
[mt
]))
3276 if ((alloc_flags
& ALLOC_CMA
) &&
3277 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
3282 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3288 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3289 int classzone_idx
, unsigned int alloc_flags
)
3291 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3292 zone_page_state(z
, NR_FREE_PAGES
));
3295 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3296 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3298 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3302 /* If allocation can't use CMA areas don't use free CMA pages */
3303 if (!(alloc_flags
& ALLOC_CMA
))
3304 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3308 * Fast check for order-0 only. If this fails then the reserves
3309 * need to be calculated. There is a corner case where the check
3310 * passes but only the high-order atomic reserve are free. If
3311 * the caller is !atomic then it'll uselessly search the free
3312 * list. That corner case is then slower but it is harmless.
3314 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3317 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3321 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3322 unsigned long mark
, int classzone_idx
)
3324 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3326 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3327 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3329 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3334 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3336 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3339 #else /* CONFIG_NUMA */
3340 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3344 #endif /* CONFIG_NUMA */
3347 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3348 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3349 * premature use of a lower zone may cause lowmem pressure problems that
3350 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3351 * probably too small. It only makes sense to spread allocations to avoid
3352 * fragmentation between the Normal and DMA32 zones.
3354 static inline unsigned int
3355 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3357 unsigned int alloc_flags
= 0;
3359 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3360 alloc_flags
|= ALLOC_KSWAPD
;
3362 #ifdef CONFIG_ZONE_DMA32
3363 if (zone_idx(zone
) != ZONE_NORMAL
)
3367 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3368 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3369 * on UMA that if Normal is populated then so is DMA32.
3371 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3372 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3376 #endif /* CONFIG_ZONE_DMA32 */
3381 * get_page_from_freelist goes through the zonelist trying to allocate
3384 static struct page
*
3385 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3386 const struct alloc_context
*ac
)
3390 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3395 * Scan zonelist, looking for a zone with enough free.
3396 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3398 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3399 z
= ac
->preferred_zoneref
;
3400 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3405 if (cpusets_enabled() &&
3406 (alloc_flags
& ALLOC_CPUSET
) &&
3407 !__cpuset_zone_allowed(zone
, gfp_mask
))
3410 * When allocating a page cache page for writing, we
3411 * want to get it from a node that is within its dirty
3412 * limit, such that no single node holds more than its
3413 * proportional share of globally allowed dirty pages.
3414 * The dirty limits take into account the node's
3415 * lowmem reserves and high watermark so that kswapd
3416 * should be able to balance it without having to
3417 * write pages from its LRU list.
3419 * XXX: For now, allow allocations to potentially
3420 * exceed the per-node dirty limit in the slowpath
3421 * (spread_dirty_pages unset) before going into reclaim,
3422 * which is important when on a NUMA setup the allowed
3423 * nodes are together not big enough to reach the
3424 * global limit. The proper fix for these situations
3425 * will require awareness of nodes in the
3426 * dirty-throttling and the flusher threads.
3428 if (ac
->spread_dirty_pages
) {
3429 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3432 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3433 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3438 if (no_fallback
&& nr_online_nodes
> 1 &&
3439 zone
!= ac
->preferred_zoneref
->zone
) {
3443 * If moving to a remote node, retry but allow
3444 * fragmenting fallbacks. Locality is more important
3445 * than fragmentation avoidance.
3447 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3448 if (zone_to_nid(zone
) != local_nid
) {
3449 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3454 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3455 if (!zone_watermark_fast(zone
, order
, mark
,
3456 ac_classzone_idx(ac
), alloc_flags
)) {
3459 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3461 * Watermark failed for this zone, but see if we can
3462 * grow this zone if it contains deferred pages.
3464 if (static_branch_unlikely(&deferred_pages
)) {
3465 if (_deferred_grow_zone(zone
, order
))
3469 /* Checked here to keep the fast path fast */
3470 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3471 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3474 if (node_reclaim_mode
== 0 ||
3475 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3478 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3480 case NODE_RECLAIM_NOSCAN
:
3483 case NODE_RECLAIM_FULL
:
3484 /* scanned but unreclaimable */
3487 /* did we reclaim enough */
3488 if (zone_watermark_ok(zone
, order
, mark
,
3489 ac_classzone_idx(ac
), alloc_flags
))
3497 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3498 gfp_mask
, alloc_flags
, ac
->migratetype
);
3500 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3503 * If this is a high-order atomic allocation then check
3504 * if the pageblock should be reserved for the future
3506 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3507 reserve_highatomic_pageblock(page
, zone
, order
);
3511 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3512 /* Try again if zone has deferred pages */
3513 if (static_branch_unlikely(&deferred_pages
)) {
3514 if (_deferred_grow_zone(zone
, order
))
3522 * It's possible on a UMA machine to get through all zones that are
3523 * fragmented. If avoiding fragmentation, reset and try again.
3526 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3533 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3535 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3536 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3538 if (!__ratelimit(&show_mem_rs
))
3542 * This documents exceptions given to allocations in certain
3543 * contexts that are allowed to allocate outside current's set
3546 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3547 if (tsk_is_oom_victim(current
) ||
3548 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3549 filter
&= ~SHOW_MEM_FILTER_NODES
;
3550 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3551 filter
&= ~SHOW_MEM_FILTER_NODES
;
3553 show_mem(filter
, nodemask
);
3556 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3558 struct va_format vaf
;
3560 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3561 DEFAULT_RATELIMIT_BURST
);
3563 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3566 va_start(args
, fmt
);
3569 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3570 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3571 nodemask_pr_args(nodemask
));
3574 cpuset_print_current_mems_allowed();
3577 warn_alloc_show_mem(gfp_mask
, nodemask
);
3580 static inline struct page
*
3581 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3582 unsigned int alloc_flags
,
3583 const struct alloc_context
*ac
)
3587 page
= get_page_from_freelist(gfp_mask
, order
,
3588 alloc_flags
|ALLOC_CPUSET
, ac
);
3590 * fallback to ignore cpuset restriction if our nodes
3594 page
= get_page_from_freelist(gfp_mask
, order
,
3600 static inline struct page
*
3601 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3602 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3604 struct oom_control oc
= {
3605 .zonelist
= ac
->zonelist
,
3606 .nodemask
= ac
->nodemask
,
3608 .gfp_mask
= gfp_mask
,
3613 *did_some_progress
= 0;
3616 * Acquire the oom lock. If that fails, somebody else is
3617 * making progress for us.
3619 if (!mutex_trylock(&oom_lock
)) {
3620 *did_some_progress
= 1;
3621 schedule_timeout_uninterruptible(1);
3626 * Go through the zonelist yet one more time, keep very high watermark
3627 * here, this is only to catch a parallel oom killing, we must fail if
3628 * we're still under heavy pressure. But make sure that this reclaim
3629 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3630 * allocation which will never fail due to oom_lock already held.
3632 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3633 ~__GFP_DIRECT_RECLAIM
, order
,
3634 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3638 /* Coredumps can quickly deplete all memory reserves */
3639 if (current
->flags
& PF_DUMPCORE
)
3641 /* The OOM killer will not help higher order allocs */
3642 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3645 * We have already exhausted all our reclaim opportunities without any
3646 * success so it is time to admit defeat. We will skip the OOM killer
3647 * because it is very likely that the caller has a more reasonable
3648 * fallback than shooting a random task.
3650 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3652 /* The OOM killer does not needlessly kill tasks for lowmem */
3653 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3655 if (pm_suspended_storage())
3658 * XXX: GFP_NOFS allocations should rather fail than rely on
3659 * other request to make a forward progress.
3660 * We are in an unfortunate situation where out_of_memory cannot
3661 * do much for this context but let's try it to at least get
3662 * access to memory reserved if the current task is killed (see
3663 * out_of_memory). Once filesystems are ready to handle allocation
3664 * failures more gracefully we should just bail out here.
3667 /* The OOM killer may not free memory on a specific node */
3668 if (gfp_mask
& __GFP_THISNODE
)
3671 /* Exhausted what can be done so it's blame time */
3672 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3673 *did_some_progress
= 1;
3676 * Help non-failing allocations by giving them access to memory
3679 if (gfp_mask
& __GFP_NOFAIL
)
3680 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3681 ALLOC_NO_WATERMARKS
, ac
);
3684 mutex_unlock(&oom_lock
);
3689 * Maximum number of compaction retries wit a progress before OOM
3690 * killer is consider as the only way to move forward.
3692 #define MAX_COMPACT_RETRIES 16
3694 #ifdef CONFIG_COMPACTION
3695 /* Try memory compaction for high-order allocations before reclaim */
3696 static struct page
*
3697 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3698 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3699 enum compact_priority prio
, enum compact_result
*compact_result
)
3702 unsigned long pflags
;
3703 unsigned int noreclaim_flag
;
3708 psi_memstall_enter(&pflags
);
3709 noreclaim_flag
= memalloc_noreclaim_save();
3711 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3714 memalloc_noreclaim_restore(noreclaim_flag
);
3715 psi_memstall_leave(&pflags
);
3717 if (*compact_result
<= COMPACT_INACTIVE
)
3721 * At least in one zone compaction wasn't deferred or skipped, so let's
3722 * count a compaction stall
3724 count_vm_event(COMPACTSTALL
);
3726 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3729 struct zone
*zone
= page_zone(page
);
3731 zone
->compact_blockskip_flush
= false;
3732 compaction_defer_reset(zone
, order
, true);
3733 count_vm_event(COMPACTSUCCESS
);
3738 * It's bad if compaction run occurs and fails. The most likely reason
3739 * is that pages exist, but not enough to satisfy watermarks.
3741 count_vm_event(COMPACTFAIL
);
3749 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3750 enum compact_result compact_result
,
3751 enum compact_priority
*compact_priority
,
3752 int *compaction_retries
)
3754 int max_retries
= MAX_COMPACT_RETRIES
;
3757 int retries
= *compaction_retries
;
3758 enum compact_priority priority
= *compact_priority
;
3763 if (compaction_made_progress(compact_result
))
3764 (*compaction_retries
)++;
3767 * compaction considers all the zone as desperately out of memory
3768 * so it doesn't really make much sense to retry except when the
3769 * failure could be caused by insufficient priority
3771 if (compaction_failed(compact_result
))
3772 goto check_priority
;
3775 * make sure the compaction wasn't deferred or didn't bail out early
3776 * due to locks contention before we declare that we should give up.
3777 * But do not retry if the given zonelist is not suitable for
3780 if (compaction_withdrawn(compact_result
)) {
3781 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3786 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3787 * costly ones because they are de facto nofail and invoke OOM
3788 * killer to move on while costly can fail and users are ready
3789 * to cope with that. 1/4 retries is rather arbitrary but we
3790 * would need much more detailed feedback from compaction to
3791 * make a better decision.
3793 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3795 if (*compaction_retries
<= max_retries
) {
3801 * Make sure there are attempts at the highest priority if we exhausted
3802 * all retries or failed at the lower priorities.
3805 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3806 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3808 if (*compact_priority
> min_priority
) {
3809 (*compact_priority
)--;
3810 *compaction_retries
= 0;
3814 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3818 static inline struct page
*
3819 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3820 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3821 enum compact_priority prio
, enum compact_result
*compact_result
)
3823 *compact_result
= COMPACT_SKIPPED
;
3828 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3829 enum compact_result compact_result
,
3830 enum compact_priority
*compact_priority
,
3831 int *compaction_retries
)
3836 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3840 * There are setups with compaction disabled which would prefer to loop
3841 * inside the allocator rather than hit the oom killer prematurely.
3842 * Let's give them a good hope and keep retrying while the order-0
3843 * watermarks are OK.
3845 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3847 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3848 ac_classzone_idx(ac
), alloc_flags
))
3853 #endif /* CONFIG_COMPACTION */
3855 #ifdef CONFIG_LOCKDEP
3856 static struct lockdep_map __fs_reclaim_map
=
3857 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3859 static bool __need_fs_reclaim(gfp_t gfp_mask
)
3861 gfp_mask
= current_gfp_context(gfp_mask
);
3863 /* no reclaim without waiting on it */
3864 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3867 /* this guy won't enter reclaim */
3868 if (current
->flags
& PF_MEMALLOC
)
3871 /* We're only interested __GFP_FS allocations for now */
3872 if (!(gfp_mask
& __GFP_FS
))
3875 if (gfp_mask
& __GFP_NOLOCKDEP
)
3881 void __fs_reclaim_acquire(void)
3883 lock_map_acquire(&__fs_reclaim_map
);
3886 void __fs_reclaim_release(void)
3888 lock_map_release(&__fs_reclaim_map
);
3891 void fs_reclaim_acquire(gfp_t gfp_mask
)
3893 if (__need_fs_reclaim(gfp_mask
))
3894 __fs_reclaim_acquire();
3896 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3898 void fs_reclaim_release(gfp_t gfp_mask
)
3900 if (__need_fs_reclaim(gfp_mask
))
3901 __fs_reclaim_release();
3903 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3906 /* Perform direct synchronous page reclaim */
3908 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3909 const struct alloc_context
*ac
)
3911 struct reclaim_state reclaim_state
;
3913 unsigned int noreclaim_flag
;
3914 unsigned long pflags
;
3918 /* We now go into synchronous reclaim */
3919 cpuset_memory_pressure_bump();
3920 psi_memstall_enter(&pflags
);
3921 fs_reclaim_acquire(gfp_mask
);
3922 noreclaim_flag
= memalloc_noreclaim_save();
3923 reclaim_state
.reclaimed_slab
= 0;
3924 current
->reclaim_state
= &reclaim_state
;
3926 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3929 current
->reclaim_state
= NULL
;
3930 memalloc_noreclaim_restore(noreclaim_flag
);
3931 fs_reclaim_release(gfp_mask
);
3932 psi_memstall_leave(&pflags
);
3939 /* The really slow allocator path where we enter direct reclaim */
3940 static inline struct page
*
3941 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3942 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3943 unsigned long *did_some_progress
)
3945 struct page
*page
= NULL
;
3946 bool drained
= false;
3948 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3949 if (unlikely(!(*did_some_progress
)))
3953 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3956 * If an allocation failed after direct reclaim, it could be because
3957 * pages are pinned on the per-cpu lists or in high alloc reserves.
3958 * Shrink them them and try again
3960 if (!page
&& !drained
) {
3961 unreserve_highatomic_pageblock(ac
, false);
3962 drain_all_pages(NULL
);
3970 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
3971 const struct alloc_context
*ac
)
3975 pg_data_t
*last_pgdat
= NULL
;
3976 enum zone_type high_zoneidx
= ac
->high_zoneidx
;
3978 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, high_zoneidx
,
3980 if (last_pgdat
!= zone
->zone_pgdat
)
3981 wakeup_kswapd(zone
, gfp_mask
, order
, high_zoneidx
);
3982 last_pgdat
= zone
->zone_pgdat
;
3986 static inline unsigned int
3987 gfp_to_alloc_flags(gfp_t gfp_mask
)
3989 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3991 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3992 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3995 * The caller may dip into page reserves a bit more if the caller
3996 * cannot run direct reclaim, or if the caller has realtime scheduling
3997 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3998 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4000 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
4002 if (gfp_mask
& __GFP_ATOMIC
) {
4004 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4005 * if it can't schedule.
4007 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4008 alloc_flags
|= ALLOC_HARDER
;
4010 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4011 * comment for __cpuset_node_allowed().
4013 alloc_flags
&= ~ALLOC_CPUSET
;
4014 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4015 alloc_flags
|= ALLOC_HARDER
;
4017 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4018 alloc_flags
|= ALLOC_KSWAPD
;
4021 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
4022 alloc_flags
|= ALLOC_CMA
;
4027 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4029 if (!tsk_is_oom_victim(tsk
))
4033 * !MMU doesn't have oom reaper so give access to memory reserves
4034 * only to the thread with TIF_MEMDIE set
4036 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4043 * Distinguish requests which really need access to full memory
4044 * reserves from oom victims which can live with a portion of it
4046 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4048 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4050 if (gfp_mask
& __GFP_MEMALLOC
)
4051 return ALLOC_NO_WATERMARKS
;
4052 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4053 return ALLOC_NO_WATERMARKS
;
4054 if (!in_interrupt()) {
4055 if (current
->flags
& PF_MEMALLOC
)
4056 return ALLOC_NO_WATERMARKS
;
4057 else if (oom_reserves_allowed(current
))
4064 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4066 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4070 * Checks whether it makes sense to retry the reclaim to make a forward progress
4071 * for the given allocation request.
4073 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4074 * without success, or when we couldn't even meet the watermark if we
4075 * reclaimed all remaining pages on the LRU lists.
4077 * Returns true if a retry is viable or false to enter the oom path.
4080 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4081 struct alloc_context
*ac
, int alloc_flags
,
4082 bool did_some_progress
, int *no_progress_loops
)
4089 * Costly allocations might have made a progress but this doesn't mean
4090 * their order will become available due to high fragmentation so
4091 * always increment the no progress counter for them
4093 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4094 *no_progress_loops
= 0;
4096 (*no_progress_loops
)++;
4099 * Make sure we converge to OOM if we cannot make any progress
4100 * several times in the row.
4102 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4103 /* Before OOM, exhaust highatomic_reserve */
4104 return unreserve_highatomic_pageblock(ac
, true);
4108 * Keep reclaiming pages while there is a chance this will lead
4109 * somewhere. If none of the target zones can satisfy our allocation
4110 * request even if all reclaimable pages are considered then we are
4111 * screwed and have to go OOM.
4113 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4115 unsigned long available
;
4116 unsigned long reclaimable
;
4117 unsigned long min_wmark
= min_wmark_pages(zone
);
4120 available
= reclaimable
= zone_reclaimable_pages(zone
);
4121 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4124 * Would the allocation succeed if we reclaimed all
4125 * reclaimable pages?
4127 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4128 ac_classzone_idx(ac
), alloc_flags
, available
);
4129 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4130 available
, min_wmark
, *no_progress_loops
, wmark
);
4133 * If we didn't make any progress and have a lot of
4134 * dirty + writeback pages then we should wait for
4135 * an IO to complete to slow down the reclaim and
4136 * prevent from pre mature OOM
4138 if (!did_some_progress
) {
4139 unsigned long write_pending
;
4141 write_pending
= zone_page_state_snapshot(zone
,
4142 NR_ZONE_WRITE_PENDING
);
4144 if (2 * write_pending
> reclaimable
) {
4145 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4157 * Memory allocation/reclaim might be called from a WQ context and the
4158 * current implementation of the WQ concurrency control doesn't
4159 * recognize that a particular WQ is congested if the worker thread is
4160 * looping without ever sleeping. Therefore we have to do a short sleep
4161 * here rather than calling cond_resched().
4163 if (current
->flags
& PF_WQ_WORKER
)
4164 schedule_timeout_uninterruptible(1);
4171 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4174 * It's possible that cpuset's mems_allowed and the nodemask from
4175 * mempolicy don't intersect. This should be normally dealt with by
4176 * policy_nodemask(), but it's possible to race with cpuset update in
4177 * such a way the check therein was true, and then it became false
4178 * before we got our cpuset_mems_cookie here.
4179 * This assumes that for all allocations, ac->nodemask can come only
4180 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4181 * when it does not intersect with the cpuset restrictions) or the
4182 * caller can deal with a violated nodemask.
4184 if (cpusets_enabled() && ac
->nodemask
&&
4185 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4186 ac
->nodemask
= NULL
;
4191 * When updating a task's mems_allowed or mempolicy nodemask, it is
4192 * possible to race with parallel threads in such a way that our
4193 * allocation can fail while the mask is being updated. If we are about
4194 * to fail, check if the cpuset changed during allocation and if so,
4197 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4203 static inline struct page
*
4204 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4205 struct alloc_context
*ac
)
4207 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4208 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4209 struct page
*page
= NULL
;
4210 unsigned int alloc_flags
;
4211 unsigned long did_some_progress
;
4212 enum compact_priority compact_priority
;
4213 enum compact_result compact_result
;
4214 int compaction_retries
;
4215 int no_progress_loops
;
4216 unsigned int cpuset_mems_cookie
;
4220 * We also sanity check to catch abuse of atomic reserves being used by
4221 * callers that are not in atomic context.
4223 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4224 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4225 gfp_mask
&= ~__GFP_ATOMIC
;
4228 compaction_retries
= 0;
4229 no_progress_loops
= 0;
4230 compact_priority
= DEF_COMPACT_PRIORITY
;
4231 cpuset_mems_cookie
= read_mems_allowed_begin();
4234 * The fast path uses conservative alloc_flags to succeed only until
4235 * kswapd needs to be woken up, and to avoid the cost of setting up
4236 * alloc_flags precisely. So we do that now.
4238 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4241 * We need to recalculate the starting point for the zonelist iterator
4242 * because we might have used different nodemask in the fast path, or
4243 * there was a cpuset modification and we are retrying - otherwise we
4244 * could end up iterating over non-eligible zones endlessly.
4246 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4247 ac
->high_zoneidx
, ac
->nodemask
);
4248 if (!ac
->preferred_zoneref
->zone
)
4251 if (alloc_flags
& ALLOC_KSWAPD
)
4252 wake_all_kswapds(order
, gfp_mask
, ac
);
4255 * The adjusted alloc_flags might result in immediate success, so try
4258 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4263 * For costly allocations, try direct compaction first, as it's likely
4264 * that we have enough base pages and don't need to reclaim. For non-
4265 * movable high-order allocations, do that as well, as compaction will
4266 * try prevent permanent fragmentation by migrating from blocks of the
4268 * Don't try this for allocations that are allowed to ignore
4269 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4271 if (can_direct_reclaim
&&
4273 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4274 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4275 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4277 INIT_COMPACT_PRIORITY
,
4283 * Checks for costly allocations with __GFP_NORETRY, which
4284 * includes THP page fault allocations
4286 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4288 * If compaction is deferred for high-order allocations,
4289 * it is because sync compaction recently failed. If
4290 * this is the case and the caller requested a THP
4291 * allocation, we do not want to heavily disrupt the
4292 * system, so we fail the allocation instead of entering
4295 if (compact_result
== COMPACT_DEFERRED
)
4299 * Looks like reclaim/compaction is worth trying, but
4300 * sync compaction could be very expensive, so keep
4301 * using async compaction.
4303 compact_priority
= INIT_COMPACT_PRIORITY
;
4308 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4309 if (alloc_flags
& ALLOC_KSWAPD
)
4310 wake_all_kswapds(order
, gfp_mask
, ac
);
4312 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4314 alloc_flags
= reserve_flags
;
4317 * Reset the nodemask and zonelist iterators if memory policies can be
4318 * ignored. These allocations are high priority and system rather than
4321 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4322 ac
->nodemask
= NULL
;
4323 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4324 ac
->high_zoneidx
, ac
->nodemask
);
4327 /* Attempt with potentially adjusted zonelist and alloc_flags */
4328 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4332 /* Caller is not willing to reclaim, we can't balance anything */
4333 if (!can_direct_reclaim
)
4336 /* Avoid recursion of direct reclaim */
4337 if (current
->flags
& PF_MEMALLOC
)
4340 /* Try direct reclaim and then allocating */
4341 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4342 &did_some_progress
);
4346 /* Try direct compaction and then allocating */
4347 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4348 compact_priority
, &compact_result
);
4352 /* Do not loop if specifically requested */
4353 if (gfp_mask
& __GFP_NORETRY
)
4357 * Do not retry costly high order allocations unless they are
4358 * __GFP_RETRY_MAYFAIL
4360 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4363 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4364 did_some_progress
> 0, &no_progress_loops
))
4368 * It doesn't make any sense to retry for the compaction if the order-0
4369 * reclaim is not able to make any progress because the current
4370 * implementation of the compaction depends on the sufficient amount
4371 * of free memory (see __compaction_suitable)
4373 if (did_some_progress
> 0 &&
4374 should_compact_retry(ac
, order
, alloc_flags
,
4375 compact_result
, &compact_priority
,
4376 &compaction_retries
))
4380 /* Deal with possible cpuset update races before we start OOM killing */
4381 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4384 /* Reclaim has failed us, start killing things */
4385 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4389 /* Avoid allocations with no watermarks from looping endlessly */
4390 if (tsk_is_oom_victim(current
) &&
4391 (alloc_flags
== ALLOC_OOM
||
4392 (gfp_mask
& __GFP_NOMEMALLOC
)))
4395 /* Retry as long as the OOM killer is making progress */
4396 if (did_some_progress
) {
4397 no_progress_loops
= 0;
4402 /* Deal with possible cpuset update races before we fail */
4403 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4407 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4410 if (gfp_mask
& __GFP_NOFAIL
) {
4412 * All existing users of the __GFP_NOFAIL are blockable, so warn
4413 * of any new users that actually require GFP_NOWAIT
4415 if (WARN_ON_ONCE(!can_direct_reclaim
))
4419 * PF_MEMALLOC request from this context is rather bizarre
4420 * because we cannot reclaim anything and only can loop waiting
4421 * for somebody to do a work for us
4423 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4426 * non failing costly orders are a hard requirement which we
4427 * are not prepared for much so let's warn about these users
4428 * so that we can identify them and convert them to something
4431 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4434 * Help non-failing allocations by giving them access to memory
4435 * reserves but do not use ALLOC_NO_WATERMARKS because this
4436 * could deplete whole memory reserves which would just make
4437 * the situation worse
4439 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4447 warn_alloc(gfp_mask
, ac
->nodemask
,
4448 "page allocation failure: order:%u", order
);
4453 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4454 int preferred_nid
, nodemask_t
*nodemask
,
4455 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4456 unsigned int *alloc_flags
)
4458 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4459 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4460 ac
->nodemask
= nodemask
;
4461 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4463 if (cpusets_enabled()) {
4464 *alloc_mask
|= __GFP_HARDWALL
;
4466 ac
->nodemask
= &cpuset_current_mems_allowed
;
4468 *alloc_flags
|= ALLOC_CPUSET
;
4471 fs_reclaim_acquire(gfp_mask
);
4472 fs_reclaim_release(gfp_mask
);
4474 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4476 if (should_fail_alloc_page(gfp_mask
, order
))
4479 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4480 *alloc_flags
|= ALLOC_CMA
;
4485 /* Determine whether to spread dirty pages and what the first usable zone */
4486 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4488 /* Dirty zone balancing only done in the fast path */
4489 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4492 * The preferred zone is used for statistics but crucially it is
4493 * also used as the starting point for the zonelist iterator. It
4494 * may get reset for allocations that ignore memory policies.
4496 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4497 ac
->high_zoneidx
, ac
->nodemask
);
4501 * This is the 'heart' of the zoned buddy allocator.
4504 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4505 nodemask_t
*nodemask
)
4508 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4509 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4510 struct alloc_context ac
= { };
4513 * There are several places where we assume that the order value is sane
4514 * so bail out early if the request is out of bound.
4516 if (unlikely(order
>= MAX_ORDER
)) {
4517 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4521 gfp_mask
&= gfp_allowed_mask
;
4522 alloc_mask
= gfp_mask
;
4523 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4526 finalise_ac(gfp_mask
, &ac
);
4529 * Forbid the first pass from falling back to types that fragment
4530 * memory until all local zones are considered.
4532 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4534 /* First allocation attempt */
4535 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4540 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4541 * resp. GFP_NOIO which has to be inherited for all allocation requests
4542 * from a particular context which has been marked by
4543 * memalloc_no{fs,io}_{save,restore}.
4545 alloc_mask
= current_gfp_context(gfp_mask
);
4546 ac
.spread_dirty_pages
= false;
4549 * Restore the original nodemask if it was potentially replaced with
4550 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4552 if (unlikely(ac
.nodemask
!= nodemask
))
4553 ac
.nodemask
= nodemask
;
4555 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4558 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4559 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4560 __free_pages(page
, order
);
4564 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4568 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4571 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4572 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4573 * you need to access high mem.
4575 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4579 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4582 return (unsigned long) page_address(page
);
4584 EXPORT_SYMBOL(__get_free_pages
);
4586 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4588 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4590 EXPORT_SYMBOL(get_zeroed_page
);
4592 static inline void free_the_page(struct page
*page
, unsigned int order
)
4594 if (order
== 0) /* Via pcp? */
4595 free_unref_page(page
);
4597 __free_pages_ok(page
, order
);
4600 void __free_pages(struct page
*page
, unsigned int order
)
4602 if (put_page_testzero(page
))
4603 free_the_page(page
, order
);
4605 EXPORT_SYMBOL(__free_pages
);
4607 void free_pages(unsigned long addr
, unsigned int order
)
4610 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4611 __free_pages(virt_to_page((void *)addr
), order
);
4615 EXPORT_SYMBOL(free_pages
);
4619 * An arbitrary-length arbitrary-offset area of memory which resides
4620 * within a 0 or higher order page. Multiple fragments within that page
4621 * are individually refcounted, in the page's reference counter.
4623 * The page_frag functions below provide a simple allocation framework for
4624 * page fragments. This is used by the network stack and network device
4625 * drivers to provide a backing region of memory for use as either an
4626 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4628 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4631 struct page
*page
= NULL
;
4632 gfp_t gfp
= gfp_mask
;
4634 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4635 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4637 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4638 PAGE_FRAG_CACHE_MAX_ORDER
);
4639 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4641 if (unlikely(!page
))
4642 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4644 nc
->va
= page
? page_address(page
) : NULL
;
4649 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4651 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4653 if (page_ref_sub_and_test(page
, count
))
4654 free_the_page(page
, compound_order(page
));
4656 EXPORT_SYMBOL(__page_frag_cache_drain
);
4658 void *page_frag_alloc(struct page_frag_cache
*nc
,
4659 unsigned int fragsz
, gfp_t gfp_mask
)
4661 unsigned int size
= PAGE_SIZE
;
4665 if (unlikely(!nc
->va
)) {
4667 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4671 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4672 /* if size can vary use size else just use PAGE_SIZE */
4675 /* Even if we own the page, we do not use atomic_set().
4676 * This would break get_page_unless_zero() users.
4678 page_ref_add(page
, size
);
4680 /* reset page count bias and offset to start of new frag */
4681 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4682 nc
->pagecnt_bias
= size
+ 1;
4686 offset
= nc
->offset
- fragsz
;
4687 if (unlikely(offset
< 0)) {
4688 page
= virt_to_page(nc
->va
);
4690 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4693 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4694 /* if size can vary use size else just use PAGE_SIZE */
4697 /* OK, page count is 0, we can safely set it */
4698 set_page_count(page
, size
+ 1);
4700 /* reset page count bias and offset to start of new frag */
4701 nc
->pagecnt_bias
= size
+ 1;
4702 offset
= size
- fragsz
;
4706 nc
->offset
= offset
;
4708 return nc
->va
+ offset
;
4710 EXPORT_SYMBOL(page_frag_alloc
);
4713 * Frees a page fragment allocated out of either a compound or order 0 page.
4715 void page_frag_free(void *addr
)
4717 struct page
*page
= virt_to_head_page(addr
);
4719 if (unlikely(put_page_testzero(page
)))
4720 free_the_page(page
, compound_order(page
));
4722 EXPORT_SYMBOL(page_frag_free
);
4724 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4728 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4729 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4731 split_page(virt_to_page((void *)addr
), order
);
4732 while (used
< alloc_end
) {
4737 return (void *)addr
;
4741 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4742 * @size: the number of bytes to allocate
4743 * @gfp_mask: GFP flags for the allocation
4745 * This function is similar to alloc_pages(), except that it allocates the
4746 * minimum number of pages to satisfy the request. alloc_pages() can only
4747 * allocate memory in power-of-two pages.
4749 * This function is also limited by MAX_ORDER.
4751 * Memory allocated by this function must be released by free_pages_exact().
4753 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4755 unsigned int order
= get_order(size
);
4758 addr
= __get_free_pages(gfp_mask
, order
);
4759 return make_alloc_exact(addr
, order
, size
);
4761 EXPORT_SYMBOL(alloc_pages_exact
);
4764 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4766 * @nid: the preferred node ID where memory should be allocated
4767 * @size: the number of bytes to allocate
4768 * @gfp_mask: GFP flags for the allocation
4770 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4773 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4775 unsigned int order
= get_order(size
);
4776 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4779 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4783 * free_pages_exact - release memory allocated via alloc_pages_exact()
4784 * @virt: the value returned by alloc_pages_exact.
4785 * @size: size of allocation, same value as passed to alloc_pages_exact().
4787 * Release the memory allocated by a previous call to alloc_pages_exact.
4789 void free_pages_exact(void *virt
, size_t size
)
4791 unsigned long addr
= (unsigned long)virt
;
4792 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4794 while (addr
< end
) {
4799 EXPORT_SYMBOL(free_pages_exact
);
4802 * nr_free_zone_pages - count number of pages beyond high watermark
4803 * @offset: The zone index of the highest zone
4805 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4806 * high watermark within all zones at or below a given zone index. For each
4807 * zone, the number of pages is calculated as:
4809 * nr_free_zone_pages = managed_pages - high_pages
4811 static unsigned long nr_free_zone_pages(int offset
)
4816 /* Just pick one node, since fallback list is circular */
4817 unsigned long sum
= 0;
4819 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4821 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4822 unsigned long size
= zone_managed_pages(zone
);
4823 unsigned long high
= high_wmark_pages(zone
);
4832 * nr_free_buffer_pages - count number of pages beyond high watermark
4834 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4835 * watermark within ZONE_DMA and ZONE_NORMAL.
4837 unsigned long nr_free_buffer_pages(void)
4839 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4841 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4844 * nr_free_pagecache_pages - count number of pages beyond high watermark
4846 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4847 * high watermark within all zones.
4849 unsigned long nr_free_pagecache_pages(void)
4851 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4854 static inline void show_node(struct zone
*zone
)
4856 if (IS_ENABLED(CONFIG_NUMA
))
4857 printk("Node %d ", zone_to_nid(zone
));
4860 long si_mem_available(void)
4863 unsigned long pagecache
;
4864 unsigned long wmark_low
= 0;
4865 unsigned long pages
[NR_LRU_LISTS
];
4866 unsigned long reclaimable
;
4870 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4871 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4874 wmark_low
+= low_wmark_pages(zone
);
4877 * Estimate the amount of memory available for userspace allocations,
4878 * without causing swapping.
4880 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4883 * Not all the page cache can be freed, otherwise the system will
4884 * start swapping. Assume at least half of the page cache, or the
4885 * low watermark worth of cache, needs to stay.
4887 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4888 pagecache
-= min(pagecache
/ 2, wmark_low
);
4889 available
+= pagecache
;
4892 * Part of the reclaimable slab and other kernel memory consists of
4893 * items that are in use, and cannot be freed. Cap this estimate at the
4896 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
4897 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
4898 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
4904 EXPORT_SYMBOL_GPL(si_mem_available
);
4906 void si_meminfo(struct sysinfo
*val
)
4908 val
->totalram
= totalram_pages();
4909 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4910 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
4911 val
->bufferram
= nr_blockdev_pages();
4912 val
->totalhigh
= totalhigh_pages();
4913 val
->freehigh
= nr_free_highpages();
4914 val
->mem_unit
= PAGE_SIZE
;
4917 EXPORT_SYMBOL(si_meminfo
);
4920 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4922 int zone_type
; /* needs to be signed */
4923 unsigned long managed_pages
= 0;
4924 unsigned long managed_highpages
= 0;
4925 unsigned long free_highpages
= 0;
4926 pg_data_t
*pgdat
= NODE_DATA(nid
);
4928 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4929 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
4930 val
->totalram
= managed_pages
;
4931 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4932 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4933 #ifdef CONFIG_HIGHMEM
4934 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4935 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4937 if (is_highmem(zone
)) {
4938 managed_highpages
+= zone_managed_pages(zone
);
4939 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4942 val
->totalhigh
= managed_highpages
;
4943 val
->freehigh
= free_highpages
;
4945 val
->totalhigh
= managed_highpages
;
4946 val
->freehigh
= free_highpages
;
4948 val
->mem_unit
= PAGE_SIZE
;
4953 * Determine whether the node should be displayed or not, depending on whether
4954 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4956 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4958 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4962 * no node mask - aka implicit memory numa policy. Do not bother with
4963 * the synchronization - read_mems_allowed_begin - because we do not
4964 * have to be precise here.
4967 nodemask
= &cpuset_current_mems_allowed
;
4969 return !node_isset(nid
, *nodemask
);
4972 #define K(x) ((x) << (PAGE_SHIFT-10))
4974 static void show_migration_types(unsigned char type
)
4976 static const char types
[MIGRATE_TYPES
] = {
4977 [MIGRATE_UNMOVABLE
] = 'U',
4978 [MIGRATE_MOVABLE
] = 'M',
4979 [MIGRATE_RECLAIMABLE
] = 'E',
4980 [MIGRATE_HIGHATOMIC
] = 'H',
4982 [MIGRATE_CMA
] = 'C',
4984 #ifdef CONFIG_MEMORY_ISOLATION
4985 [MIGRATE_ISOLATE
] = 'I',
4988 char tmp
[MIGRATE_TYPES
+ 1];
4992 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4993 if (type
& (1 << i
))
4998 printk(KERN_CONT
"(%s) ", tmp
);
5002 * Show free area list (used inside shift_scroll-lock stuff)
5003 * We also calculate the percentage fragmentation. We do this by counting the
5004 * memory on each free list with the exception of the first item on the list.
5007 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5010 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5012 unsigned long free_pcp
= 0;
5017 for_each_populated_zone(zone
) {
5018 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5021 for_each_online_cpu(cpu
)
5022 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5025 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5026 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5027 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5028 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5029 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5030 " free:%lu free_pcp:%lu free_cma:%lu\n",
5031 global_node_page_state(NR_ACTIVE_ANON
),
5032 global_node_page_state(NR_INACTIVE_ANON
),
5033 global_node_page_state(NR_ISOLATED_ANON
),
5034 global_node_page_state(NR_ACTIVE_FILE
),
5035 global_node_page_state(NR_INACTIVE_FILE
),
5036 global_node_page_state(NR_ISOLATED_FILE
),
5037 global_node_page_state(NR_UNEVICTABLE
),
5038 global_node_page_state(NR_FILE_DIRTY
),
5039 global_node_page_state(NR_WRITEBACK
),
5040 global_node_page_state(NR_UNSTABLE_NFS
),
5041 global_node_page_state(NR_SLAB_RECLAIMABLE
),
5042 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
5043 global_node_page_state(NR_FILE_MAPPED
),
5044 global_node_page_state(NR_SHMEM
),
5045 global_zone_page_state(NR_PAGETABLE
),
5046 global_zone_page_state(NR_BOUNCE
),
5047 global_zone_page_state(NR_FREE_PAGES
),
5049 global_zone_page_state(NR_FREE_CMA_PAGES
));
5051 for_each_online_pgdat(pgdat
) {
5052 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5056 " active_anon:%lukB"
5057 " inactive_anon:%lukB"
5058 " active_file:%lukB"
5059 " inactive_file:%lukB"
5060 " unevictable:%lukB"
5061 " isolated(anon):%lukB"
5062 " isolated(file):%lukB"
5067 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5069 " shmem_pmdmapped: %lukB"
5072 " writeback_tmp:%lukB"
5074 " all_unreclaimable? %s"
5077 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5078 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5079 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5080 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5081 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5082 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5083 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5084 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5085 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5086 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5087 K(node_page_state(pgdat
, NR_SHMEM
)),
5088 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5089 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5090 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5092 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5094 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5095 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
5096 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5100 for_each_populated_zone(zone
) {
5103 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5107 for_each_online_cpu(cpu
)
5108 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5117 " active_anon:%lukB"
5118 " inactive_anon:%lukB"
5119 " active_file:%lukB"
5120 " inactive_file:%lukB"
5121 " unevictable:%lukB"
5122 " writepending:%lukB"
5126 " kernel_stack:%lukB"
5134 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5135 K(min_wmark_pages(zone
)),
5136 K(low_wmark_pages(zone
)),
5137 K(high_wmark_pages(zone
)),
5138 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5139 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5140 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5141 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5142 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5143 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5144 K(zone
->present_pages
),
5145 K(zone_managed_pages(zone
)),
5146 K(zone_page_state(zone
, NR_MLOCK
)),
5147 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5148 K(zone_page_state(zone
, NR_PAGETABLE
)),
5149 K(zone_page_state(zone
, NR_BOUNCE
)),
5151 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5152 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5153 printk("lowmem_reserve[]:");
5154 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5155 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5156 printk(KERN_CONT
"\n");
5159 for_each_populated_zone(zone
) {
5161 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5162 unsigned char types
[MAX_ORDER
];
5164 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5167 printk(KERN_CONT
"%s: ", zone
->name
);
5169 spin_lock_irqsave(&zone
->lock
, flags
);
5170 for (order
= 0; order
< MAX_ORDER
; order
++) {
5171 struct free_area
*area
= &zone
->free_area
[order
];
5174 nr
[order
] = area
->nr_free
;
5175 total
+= nr
[order
] << order
;
5178 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5179 if (!list_empty(&area
->free_list
[type
]))
5180 types
[order
] |= 1 << type
;
5183 spin_unlock_irqrestore(&zone
->lock
, flags
);
5184 for (order
= 0; order
< MAX_ORDER
; order
++) {
5185 printk(KERN_CONT
"%lu*%lukB ",
5186 nr
[order
], K(1UL) << order
);
5188 show_migration_types(types
[order
]);
5190 printk(KERN_CONT
"= %lukB\n", K(total
));
5193 hugetlb_show_meminfo();
5195 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5197 show_swap_cache_info();
5200 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5202 zoneref
->zone
= zone
;
5203 zoneref
->zone_idx
= zone_idx(zone
);
5207 * Builds allocation fallback zone lists.
5209 * Add all populated zones of a node to the zonelist.
5211 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5214 enum zone_type zone_type
= MAX_NR_ZONES
;
5219 zone
= pgdat
->node_zones
+ zone_type
;
5220 if (managed_zone(zone
)) {
5221 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5222 check_highest_zone(zone_type
);
5224 } while (zone_type
);
5231 static int __parse_numa_zonelist_order(char *s
)
5234 * We used to support different zonlists modes but they turned
5235 * out to be just not useful. Let's keep the warning in place
5236 * if somebody still use the cmd line parameter so that we do
5237 * not fail it silently
5239 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5240 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5246 static __init
int setup_numa_zonelist_order(char *s
)
5251 return __parse_numa_zonelist_order(s
);
5253 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
5255 char numa_zonelist_order
[] = "Node";
5258 * sysctl handler for numa_zonelist_order
5260 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5261 void __user
*buffer
, size_t *length
,
5268 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5269 str
= memdup_user_nul(buffer
, 16);
5271 return PTR_ERR(str
);
5273 ret
= __parse_numa_zonelist_order(str
);
5279 #define MAX_NODE_LOAD (nr_online_nodes)
5280 static int node_load
[MAX_NUMNODES
];
5283 * find_next_best_node - find the next node that should appear in a given node's fallback list
5284 * @node: node whose fallback list we're appending
5285 * @used_node_mask: nodemask_t of already used nodes
5287 * We use a number of factors to determine which is the next node that should
5288 * appear on a given node's fallback list. The node should not have appeared
5289 * already in @node's fallback list, and it should be the next closest node
5290 * according to the distance array (which contains arbitrary distance values
5291 * from each node to each node in the system), and should also prefer nodes
5292 * with no CPUs, since presumably they'll have very little allocation pressure
5293 * on them otherwise.
5294 * It returns -1 if no node is found.
5296 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5299 int min_val
= INT_MAX
;
5300 int best_node
= NUMA_NO_NODE
;
5301 const struct cpumask
*tmp
= cpumask_of_node(0);
5303 /* Use the local node if we haven't already */
5304 if (!node_isset(node
, *used_node_mask
)) {
5305 node_set(node
, *used_node_mask
);
5309 for_each_node_state(n
, N_MEMORY
) {
5311 /* Don't want a node to appear more than once */
5312 if (node_isset(n
, *used_node_mask
))
5315 /* Use the distance array to find the distance */
5316 val
= node_distance(node
, n
);
5318 /* Penalize nodes under us ("prefer the next node") */
5321 /* Give preference to headless and unused nodes */
5322 tmp
= cpumask_of_node(n
);
5323 if (!cpumask_empty(tmp
))
5324 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5326 /* Slight preference for less loaded node */
5327 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5328 val
+= node_load
[n
];
5330 if (val
< min_val
) {
5337 node_set(best_node
, *used_node_mask
);
5344 * Build zonelists ordered by node and zones within node.
5345 * This results in maximum locality--normal zone overflows into local
5346 * DMA zone, if any--but risks exhausting DMA zone.
5348 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5351 struct zoneref
*zonerefs
;
5354 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5356 for (i
= 0; i
< nr_nodes
; i
++) {
5359 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5361 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5362 zonerefs
+= nr_zones
;
5364 zonerefs
->zone
= NULL
;
5365 zonerefs
->zone_idx
= 0;
5369 * Build gfp_thisnode zonelists
5371 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5373 struct zoneref
*zonerefs
;
5376 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5377 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5378 zonerefs
+= nr_zones
;
5379 zonerefs
->zone
= NULL
;
5380 zonerefs
->zone_idx
= 0;
5384 * Build zonelists ordered by zone and nodes within zones.
5385 * This results in conserving DMA zone[s] until all Normal memory is
5386 * exhausted, but results in overflowing to remote node while memory
5387 * may still exist in local DMA zone.
5390 static void build_zonelists(pg_data_t
*pgdat
)
5392 static int node_order
[MAX_NUMNODES
];
5393 int node
, load
, nr_nodes
= 0;
5394 nodemask_t used_mask
;
5395 int local_node
, prev_node
;
5397 /* NUMA-aware ordering of nodes */
5398 local_node
= pgdat
->node_id
;
5399 load
= nr_online_nodes
;
5400 prev_node
= local_node
;
5401 nodes_clear(used_mask
);
5403 memset(node_order
, 0, sizeof(node_order
));
5404 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5406 * We don't want to pressure a particular node.
5407 * So adding penalty to the first node in same
5408 * distance group to make it round-robin.
5410 if (node_distance(local_node
, node
) !=
5411 node_distance(local_node
, prev_node
))
5412 node_load
[node
] = load
;
5414 node_order
[nr_nodes
++] = node
;
5419 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5420 build_thisnode_zonelists(pgdat
);
5423 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5425 * Return node id of node used for "local" allocations.
5426 * I.e., first node id of first zone in arg node's generic zonelist.
5427 * Used for initializing percpu 'numa_mem', which is used primarily
5428 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5430 int local_memory_node(int node
)
5434 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5435 gfp_zone(GFP_KERNEL
),
5437 return zone_to_nid(z
->zone
);
5441 static void setup_min_unmapped_ratio(void);
5442 static void setup_min_slab_ratio(void);
5443 #else /* CONFIG_NUMA */
5445 static void build_zonelists(pg_data_t
*pgdat
)
5447 int node
, local_node
;
5448 struct zoneref
*zonerefs
;
5451 local_node
= pgdat
->node_id
;
5453 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5454 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5455 zonerefs
+= nr_zones
;
5458 * Now we build the zonelist so that it contains the zones
5459 * of all the other nodes.
5460 * We don't want to pressure a particular node, so when
5461 * building the zones for node N, we make sure that the
5462 * zones coming right after the local ones are those from
5463 * node N+1 (modulo N)
5465 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5466 if (!node_online(node
))
5468 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5469 zonerefs
+= nr_zones
;
5471 for (node
= 0; node
< local_node
; node
++) {
5472 if (!node_online(node
))
5474 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5475 zonerefs
+= nr_zones
;
5478 zonerefs
->zone
= NULL
;
5479 zonerefs
->zone_idx
= 0;
5482 #endif /* CONFIG_NUMA */
5485 * Boot pageset table. One per cpu which is going to be used for all
5486 * zones and all nodes. The parameters will be set in such a way
5487 * that an item put on a list will immediately be handed over to
5488 * the buddy list. This is safe since pageset manipulation is done
5489 * with interrupts disabled.
5491 * The boot_pagesets must be kept even after bootup is complete for
5492 * unused processors and/or zones. They do play a role for bootstrapping
5493 * hotplugged processors.
5495 * zoneinfo_show() and maybe other functions do
5496 * not check if the processor is online before following the pageset pointer.
5497 * Other parts of the kernel may not check if the zone is available.
5499 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5500 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5501 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5503 static void __build_all_zonelists(void *data
)
5506 int __maybe_unused cpu
;
5507 pg_data_t
*self
= data
;
5508 static DEFINE_SPINLOCK(lock
);
5513 memset(node_load
, 0, sizeof(node_load
));
5517 * This node is hotadded and no memory is yet present. So just
5518 * building zonelists is fine - no need to touch other nodes.
5520 if (self
&& !node_online(self
->node_id
)) {
5521 build_zonelists(self
);
5523 for_each_online_node(nid
) {
5524 pg_data_t
*pgdat
= NODE_DATA(nid
);
5526 build_zonelists(pgdat
);
5529 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5531 * We now know the "local memory node" for each node--
5532 * i.e., the node of the first zone in the generic zonelist.
5533 * Set up numa_mem percpu variable for on-line cpus. During
5534 * boot, only the boot cpu should be on-line; we'll init the
5535 * secondary cpus' numa_mem as they come on-line. During
5536 * node/memory hotplug, we'll fixup all on-line cpus.
5538 for_each_online_cpu(cpu
)
5539 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5546 static noinline
void __init
5547 build_all_zonelists_init(void)
5551 __build_all_zonelists(NULL
);
5554 * Initialize the boot_pagesets that are going to be used
5555 * for bootstrapping processors. The real pagesets for
5556 * each zone will be allocated later when the per cpu
5557 * allocator is available.
5559 * boot_pagesets are used also for bootstrapping offline
5560 * cpus if the system is already booted because the pagesets
5561 * are needed to initialize allocators on a specific cpu too.
5562 * F.e. the percpu allocator needs the page allocator which
5563 * needs the percpu allocator in order to allocate its pagesets
5564 * (a chicken-egg dilemma).
5566 for_each_possible_cpu(cpu
)
5567 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5569 mminit_verify_zonelist();
5570 cpuset_init_current_mems_allowed();
5574 * unless system_state == SYSTEM_BOOTING.
5576 * __ref due to call of __init annotated helper build_all_zonelists_init
5577 * [protected by SYSTEM_BOOTING].
5579 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5581 if (system_state
== SYSTEM_BOOTING
) {
5582 build_all_zonelists_init();
5584 __build_all_zonelists(pgdat
);
5585 /* cpuset refresh routine should be here */
5587 vm_total_pages
= nr_free_pagecache_pages();
5589 * Disable grouping by mobility if the number of pages in the
5590 * system is too low to allow the mechanism to work. It would be
5591 * more accurate, but expensive to check per-zone. This check is
5592 * made on memory-hotadd so a system can start with mobility
5593 * disabled and enable it later
5595 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5596 page_group_by_mobility_disabled
= 1;
5598 page_group_by_mobility_disabled
= 0;
5600 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5602 page_group_by_mobility_disabled
? "off" : "on",
5605 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5609 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5610 static bool __meminit
5611 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5613 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5614 static struct memblock_region
*r
;
5616 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5617 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5618 for_each_memblock(memory
, r
) {
5619 if (*pfn
< memblock_region_memory_end_pfn(r
))
5623 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5624 memblock_is_mirror(r
)) {
5625 *pfn
= memblock_region_memory_end_pfn(r
);
5634 * Initially all pages are reserved - free ones are freed
5635 * up by memblock_free_all() once the early boot process is
5636 * done. Non-atomic initialization, single-pass.
5638 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5639 unsigned long start_pfn
, enum memmap_context context
,
5640 struct vmem_altmap
*altmap
)
5642 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5645 if (highest_memmap_pfn
< end_pfn
- 1)
5646 highest_memmap_pfn
= end_pfn
- 1;
5648 #ifdef CONFIG_ZONE_DEVICE
5650 * Honor reservation requested by the driver for this ZONE_DEVICE
5651 * memory. We limit the total number of pages to initialize to just
5652 * those that might contain the memory mapping. We will defer the
5653 * ZONE_DEVICE page initialization until after we have released
5656 if (zone
== ZONE_DEVICE
) {
5660 if (start_pfn
== altmap
->base_pfn
)
5661 start_pfn
+= altmap
->reserve
;
5662 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5666 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5668 * There can be holes in boot-time mem_map[]s handed to this
5669 * function. They do not exist on hotplugged memory.
5671 if (context
== MEMMAP_EARLY
) {
5672 if (!early_pfn_valid(pfn
))
5674 if (!early_pfn_in_nid(pfn
, nid
))
5676 if (overlap_memmap_init(zone
, &pfn
))
5678 if (defer_init(nid
, pfn
, end_pfn
))
5682 page
= pfn_to_page(pfn
);
5683 __init_single_page(page
, pfn
, zone
, nid
);
5684 if (context
== MEMMAP_HOTPLUG
)
5685 __SetPageReserved(page
);
5688 * Mark the block movable so that blocks are reserved for
5689 * movable at startup. This will force kernel allocations
5690 * to reserve their blocks rather than leaking throughout
5691 * the address space during boot when many long-lived
5692 * kernel allocations are made.
5694 * bitmap is created for zone's valid pfn range. but memmap
5695 * can be created for invalid pages (for alignment)
5696 * check here not to call set_pageblock_migratetype() against
5699 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5700 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5706 #ifdef CONFIG_ZONE_DEVICE
5707 void __ref
memmap_init_zone_device(struct zone
*zone
,
5708 unsigned long start_pfn
,
5710 struct dev_pagemap
*pgmap
)
5712 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5713 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5714 unsigned long zone_idx
= zone_idx(zone
);
5715 unsigned long start
= jiffies
;
5716 int nid
= pgdat
->node_id
;
5718 if (WARN_ON_ONCE(!pgmap
|| !is_dev_zone(zone
)))
5722 * The call to memmap_init_zone should have already taken care
5723 * of the pages reserved for the memmap, so we can just jump to
5724 * the end of that region and start processing the device pages.
5726 if (pgmap
->altmap_valid
) {
5727 struct vmem_altmap
*altmap
= &pgmap
->altmap
;
5729 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5730 size
= end_pfn
- start_pfn
;
5733 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5734 struct page
*page
= pfn_to_page(pfn
);
5736 __init_single_page(page
, pfn
, zone_idx
, nid
);
5739 * Mark page reserved as it will need to wait for onlining
5740 * phase for it to be fully associated with a zone.
5742 * We can use the non-atomic __set_bit operation for setting
5743 * the flag as we are still initializing the pages.
5745 __SetPageReserved(page
);
5748 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5749 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5750 * page is ever freed or placed on a driver-private list.
5752 page
->pgmap
= pgmap
;
5756 * Mark the block movable so that blocks are reserved for
5757 * movable at startup. This will force kernel allocations
5758 * to reserve their blocks rather than leaking throughout
5759 * the address space during boot when many long-lived
5760 * kernel allocations are made.
5762 * bitmap is created for zone's valid pfn range. but memmap
5763 * can be created for invalid pages (for alignment)
5764 * check here not to call set_pageblock_migratetype() against
5767 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5768 * because this is done early in sparse_add_one_section
5770 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5771 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5776 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap
->dev
),
5777 size
, jiffies_to_msecs(jiffies
- start
));
5781 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5783 unsigned int order
, t
;
5784 for_each_migratetype_order(order
, t
) {
5785 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5786 zone
->free_area
[order
].nr_free
= 0;
5790 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
5791 unsigned long zone
, unsigned long start_pfn
)
5793 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
, NULL
);
5796 static int zone_batchsize(struct zone
*zone
)
5802 * The per-cpu-pages pools are set to around 1000th of the
5805 batch
= zone_managed_pages(zone
) / 1024;
5806 /* But no more than a meg. */
5807 if (batch
* PAGE_SIZE
> 1024 * 1024)
5808 batch
= (1024 * 1024) / PAGE_SIZE
;
5809 batch
/= 4; /* We effectively *= 4 below */
5814 * Clamp the batch to a 2^n - 1 value. Having a power
5815 * of 2 value was found to be more likely to have
5816 * suboptimal cache aliasing properties in some cases.
5818 * For example if 2 tasks are alternately allocating
5819 * batches of pages, one task can end up with a lot
5820 * of pages of one half of the possible page colors
5821 * and the other with pages of the other colors.
5823 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5828 /* The deferral and batching of frees should be suppressed under NOMMU
5831 * The problem is that NOMMU needs to be able to allocate large chunks
5832 * of contiguous memory as there's no hardware page translation to
5833 * assemble apparent contiguous memory from discontiguous pages.
5835 * Queueing large contiguous runs of pages for batching, however,
5836 * causes the pages to actually be freed in smaller chunks. As there
5837 * can be a significant delay between the individual batches being
5838 * recycled, this leads to the once large chunks of space being
5839 * fragmented and becoming unavailable for high-order allocations.
5846 * pcp->high and pcp->batch values are related and dependent on one another:
5847 * ->batch must never be higher then ->high.
5848 * The following function updates them in a safe manner without read side
5851 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5852 * those fields changing asynchronously (acording the the above rule).
5854 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5855 * outside of boot time (or some other assurance that no concurrent updaters
5858 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5859 unsigned long batch
)
5861 /* start with a fail safe value for batch */
5865 /* Update high, then batch, in order */
5872 /* a companion to pageset_set_high() */
5873 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5875 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5878 static void pageset_init(struct per_cpu_pageset
*p
)
5880 struct per_cpu_pages
*pcp
;
5883 memset(p
, 0, sizeof(*p
));
5886 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5887 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5890 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5893 pageset_set_batch(p
, batch
);
5897 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5898 * to the value high for the pageset p.
5900 static void pageset_set_high(struct per_cpu_pageset
*p
,
5903 unsigned long batch
= max(1UL, high
/ 4);
5904 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5905 batch
= PAGE_SHIFT
* 8;
5907 pageset_update(&p
->pcp
, high
, batch
);
5910 static void pageset_set_high_and_batch(struct zone
*zone
,
5911 struct per_cpu_pageset
*pcp
)
5913 if (percpu_pagelist_fraction
)
5914 pageset_set_high(pcp
,
5915 (zone_managed_pages(zone
) /
5916 percpu_pagelist_fraction
));
5918 pageset_set_batch(pcp
, zone_batchsize(zone
));
5921 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5923 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5926 pageset_set_high_and_batch(zone
, pcp
);
5929 void __meminit
setup_zone_pageset(struct zone
*zone
)
5932 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5933 for_each_possible_cpu(cpu
)
5934 zone_pageset_init(zone
, cpu
);
5938 * Allocate per cpu pagesets and initialize them.
5939 * Before this call only boot pagesets were available.
5941 void __init
setup_per_cpu_pageset(void)
5943 struct pglist_data
*pgdat
;
5946 for_each_populated_zone(zone
)
5947 setup_zone_pageset(zone
);
5949 for_each_online_pgdat(pgdat
)
5950 pgdat
->per_cpu_nodestats
=
5951 alloc_percpu(struct per_cpu_nodestat
);
5954 static __meminit
void zone_pcp_init(struct zone
*zone
)
5957 * per cpu subsystem is not up at this point. The following code
5958 * relies on the ability of the linker to provide the
5959 * offset of a (static) per cpu variable into the per cpu area.
5961 zone
->pageset
= &boot_pageset
;
5963 if (populated_zone(zone
))
5964 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5965 zone
->name
, zone
->present_pages
,
5966 zone_batchsize(zone
));
5969 void __meminit
init_currently_empty_zone(struct zone
*zone
,
5970 unsigned long zone_start_pfn
,
5973 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5974 int zone_idx
= zone_idx(zone
) + 1;
5976 if (zone_idx
> pgdat
->nr_zones
)
5977 pgdat
->nr_zones
= zone_idx
;
5979 zone
->zone_start_pfn
= zone_start_pfn
;
5981 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5982 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5984 (unsigned long)zone_idx(zone
),
5985 zone_start_pfn
, (zone_start_pfn
+ size
));
5987 zone_init_free_lists(zone
);
5988 zone
->initialized
= 1;
5991 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5992 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5995 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5997 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5998 struct mminit_pfnnid_cache
*state
)
6000 unsigned long start_pfn
, end_pfn
;
6003 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
6004 return state
->last_nid
;
6006 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
6008 state
->last_start
= start_pfn
;
6009 state
->last_end
= end_pfn
;
6010 state
->last_nid
= nid
;
6015 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6018 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6019 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6020 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6022 * If an architecture guarantees that all ranges registered contain no holes
6023 * and may be freed, this this function may be used instead of calling
6024 * memblock_free_early_nid() manually.
6026 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
6028 unsigned long start_pfn
, end_pfn
;
6031 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
6032 start_pfn
= min(start_pfn
, max_low_pfn
);
6033 end_pfn
= min(end_pfn
, max_low_pfn
);
6035 if (start_pfn
< end_pfn
)
6036 memblock_free_early_nid(PFN_PHYS(start_pfn
),
6037 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
6043 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6044 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6046 * If an architecture guarantees that all ranges registered contain no holes and may
6047 * be freed, this function may be used instead of calling memory_present() manually.
6049 void __init
sparse_memory_present_with_active_regions(int nid
)
6051 unsigned long start_pfn
, end_pfn
;
6054 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
6055 memory_present(this_nid
, start_pfn
, end_pfn
);
6059 * get_pfn_range_for_nid - Return the start and end page frames for a node
6060 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6061 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6062 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6064 * It returns the start and end page frame of a node based on information
6065 * provided by memblock_set_node(). If called for a node
6066 * with no available memory, a warning is printed and the start and end
6069 void __init
get_pfn_range_for_nid(unsigned int nid
,
6070 unsigned long *start_pfn
, unsigned long *end_pfn
)
6072 unsigned long this_start_pfn
, this_end_pfn
;
6078 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6079 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6080 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6083 if (*start_pfn
== -1UL)
6088 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6089 * assumption is made that zones within a node are ordered in monotonic
6090 * increasing memory addresses so that the "highest" populated zone is used
6092 static void __init
find_usable_zone_for_movable(void)
6095 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6096 if (zone_index
== ZONE_MOVABLE
)
6099 if (arch_zone_highest_possible_pfn
[zone_index
] >
6100 arch_zone_lowest_possible_pfn
[zone_index
])
6104 VM_BUG_ON(zone_index
== -1);
6105 movable_zone
= zone_index
;
6109 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6110 * because it is sized independent of architecture. Unlike the other zones,
6111 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6112 * in each node depending on the size of each node and how evenly kernelcore
6113 * is distributed. This helper function adjusts the zone ranges
6114 * provided by the architecture for a given node by using the end of the
6115 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6116 * zones within a node are in order of monotonic increases memory addresses
6118 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6119 unsigned long zone_type
,
6120 unsigned long node_start_pfn
,
6121 unsigned long node_end_pfn
,
6122 unsigned long *zone_start_pfn
,
6123 unsigned long *zone_end_pfn
)
6125 /* Only adjust if ZONE_MOVABLE is on this node */
6126 if (zone_movable_pfn
[nid
]) {
6127 /* Size ZONE_MOVABLE */
6128 if (zone_type
== ZONE_MOVABLE
) {
6129 *zone_start_pfn
= zone_movable_pfn
[nid
];
6130 *zone_end_pfn
= min(node_end_pfn
,
6131 arch_zone_highest_possible_pfn
[movable_zone
]);
6133 /* Adjust for ZONE_MOVABLE starting within this range */
6134 } else if (!mirrored_kernelcore
&&
6135 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6136 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6137 *zone_end_pfn
= zone_movable_pfn
[nid
];
6139 /* Check if this whole range is within ZONE_MOVABLE */
6140 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6141 *zone_start_pfn
= *zone_end_pfn
;
6146 * Return the number of pages a zone spans in a node, including holes
6147 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6149 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6150 unsigned long zone_type
,
6151 unsigned long node_start_pfn
,
6152 unsigned long node_end_pfn
,
6153 unsigned long *zone_start_pfn
,
6154 unsigned long *zone_end_pfn
,
6155 unsigned long *ignored
)
6157 /* When hotadd a new node from cpu_up(), the node should be empty */
6158 if (!node_start_pfn
&& !node_end_pfn
)
6161 /* Get the start and end of the zone */
6162 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
6163 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
6164 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6165 node_start_pfn
, node_end_pfn
,
6166 zone_start_pfn
, zone_end_pfn
);
6168 /* Check that this node has pages within the zone's required range */
6169 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6172 /* Move the zone boundaries inside the node if necessary */
6173 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6174 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6176 /* Return the spanned pages */
6177 return *zone_end_pfn
- *zone_start_pfn
;
6181 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6182 * then all holes in the requested range will be accounted for.
6184 unsigned long __init
__absent_pages_in_range(int nid
,
6185 unsigned long range_start_pfn
,
6186 unsigned long range_end_pfn
)
6188 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6189 unsigned long start_pfn
, end_pfn
;
6192 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6193 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6194 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6195 nr_absent
-= end_pfn
- start_pfn
;
6201 * absent_pages_in_range - Return number of page frames in holes within a range
6202 * @start_pfn: The start PFN to start searching for holes
6203 * @end_pfn: The end PFN to stop searching for holes
6205 * It returns the number of pages frames in memory holes within a range.
6207 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6208 unsigned long end_pfn
)
6210 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6213 /* Return the number of page frames in holes in a zone on a node */
6214 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6215 unsigned long zone_type
,
6216 unsigned long node_start_pfn
,
6217 unsigned long node_end_pfn
,
6218 unsigned long *ignored
)
6220 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6221 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6222 unsigned long zone_start_pfn
, zone_end_pfn
;
6223 unsigned long nr_absent
;
6225 /* When hotadd a new node from cpu_up(), the node should be empty */
6226 if (!node_start_pfn
&& !node_end_pfn
)
6229 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6230 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6232 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6233 node_start_pfn
, node_end_pfn
,
6234 &zone_start_pfn
, &zone_end_pfn
);
6235 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6238 * ZONE_MOVABLE handling.
6239 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6242 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6243 unsigned long start_pfn
, end_pfn
;
6244 struct memblock_region
*r
;
6246 for_each_memblock(memory
, r
) {
6247 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6248 zone_start_pfn
, zone_end_pfn
);
6249 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6250 zone_start_pfn
, zone_end_pfn
);
6252 if (zone_type
== ZONE_MOVABLE
&&
6253 memblock_is_mirror(r
))
6254 nr_absent
+= end_pfn
- start_pfn
;
6256 if (zone_type
== ZONE_NORMAL
&&
6257 !memblock_is_mirror(r
))
6258 nr_absent
+= end_pfn
- start_pfn
;
6265 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6266 static inline unsigned long __init
zone_spanned_pages_in_node(int nid
,
6267 unsigned long zone_type
,
6268 unsigned long node_start_pfn
,
6269 unsigned long node_end_pfn
,
6270 unsigned long *zone_start_pfn
,
6271 unsigned long *zone_end_pfn
,
6272 unsigned long *zones_size
)
6276 *zone_start_pfn
= node_start_pfn
;
6277 for (zone
= 0; zone
< zone_type
; zone
++)
6278 *zone_start_pfn
+= zones_size
[zone
];
6280 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
6282 return zones_size
[zone_type
];
6285 static inline unsigned long __init
zone_absent_pages_in_node(int nid
,
6286 unsigned long zone_type
,
6287 unsigned long node_start_pfn
,
6288 unsigned long node_end_pfn
,
6289 unsigned long *zholes_size
)
6294 return zholes_size
[zone_type
];
6297 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6299 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6300 unsigned long node_start_pfn
,
6301 unsigned long node_end_pfn
,
6302 unsigned long *zones_size
,
6303 unsigned long *zholes_size
)
6305 unsigned long realtotalpages
= 0, totalpages
= 0;
6308 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6309 struct zone
*zone
= pgdat
->node_zones
+ i
;
6310 unsigned long zone_start_pfn
, zone_end_pfn
;
6311 unsigned long size
, real_size
;
6313 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6319 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
6320 node_start_pfn
, node_end_pfn
,
6323 zone
->zone_start_pfn
= zone_start_pfn
;
6325 zone
->zone_start_pfn
= 0;
6326 zone
->spanned_pages
= size
;
6327 zone
->present_pages
= real_size
;
6330 realtotalpages
+= real_size
;
6333 pgdat
->node_spanned_pages
= totalpages
;
6334 pgdat
->node_present_pages
= realtotalpages
;
6335 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6339 #ifndef CONFIG_SPARSEMEM
6341 * Calculate the size of the zone->blockflags rounded to an unsigned long
6342 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6343 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6344 * round what is now in bits to nearest long in bits, then return it in
6347 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6349 unsigned long usemapsize
;
6351 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6352 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6353 usemapsize
= usemapsize
>> pageblock_order
;
6354 usemapsize
*= NR_PAGEBLOCK_BITS
;
6355 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6357 return usemapsize
/ 8;
6360 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6362 unsigned long zone_start_pfn
,
6363 unsigned long zonesize
)
6365 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6366 zone
->pageblock_flags
= NULL
;
6368 zone
->pageblock_flags
=
6369 memblock_alloc_node_nopanic(usemapsize
,
6373 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6374 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6375 #endif /* CONFIG_SPARSEMEM */
6377 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6379 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6380 void __init
set_pageblock_order(void)
6384 /* Check that pageblock_nr_pages has not already been setup */
6385 if (pageblock_order
)
6388 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6389 order
= HUGETLB_PAGE_ORDER
;
6391 order
= MAX_ORDER
- 1;
6394 * Assume the largest contiguous order of interest is a huge page.
6395 * This value may be variable depending on boot parameters on IA64 and
6398 pageblock_order
= order
;
6400 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6403 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6404 * is unused as pageblock_order is set at compile-time. See
6405 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6408 void __init
set_pageblock_order(void)
6412 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6414 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6415 unsigned long present_pages
)
6417 unsigned long pages
= spanned_pages
;
6420 * Provide a more accurate estimation if there are holes within
6421 * the zone and SPARSEMEM is in use. If there are holes within the
6422 * zone, each populated memory region may cost us one or two extra
6423 * memmap pages due to alignment because memmap pages for each
6424 * populated regions may not be naturally aligned on page boundary.
6425 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6427 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6428 IS_ENABLED(CONFIG_SPARSEMEM
))
6429 pages
= present_pages
;
6431 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6434 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6435 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6437 spin_lock_init(&pgdat
->split_queue_lock
);
6438 INIT_LIST_HEAD(&pgdat
->split_queue
);
6439 pgdat
->split_queue_len
= 0;
6442 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6445 #ifdef CONFIG_COMPACTION
6446 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6448 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6451 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6454 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6456 pgdat_resize_init(pgdat
);
6458 pgdat_init_split_queue(pgdat
);
6459 pgdat_init_kcompactd(pgdat
);
6461 init_waitqueue_head(&pgdat
->kswapd_wait
);
6462 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6464 pgdat_page_ext_init(pgdat
);
6465 spin_lock_init(&pgdat
->lru_lock
);
6466 lruvec_init(node_lruvec(pgdat
));
6469 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6470 unsigned long remaining_pages
)
6472 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6473 zone_set_nid(zone
, nid
);
6474 zone
->name
= zone_names
[idx
];
6475 zone
->zone_pgdat
= NODE_DATA(nid
);
6476 spin_lock_init(&zone
->lock
);
6477 zone_seqlock_init(zone
);
6478 zone_pcp_init(zone
);
6482 * Set up the zone data structures
6483 * - init pgdat internals
6484 * - init all zones belonging to this node
6486 * NOTE: this function is only called during memory hotplug
6488 #ifdef CONFIG_MEMORY_HOTPLUG
6489 void __ref
free_area_init_core_hotplug(int nid
)
6492 pg_data_t
*pgdat
= NODE_DATA(nid
);
6494 pgdat_init_internals(pgdat
);
6495 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6496 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6501 * Set up the zone data structures:
6502 * - mark all pages reserved
6503 * - mark all memory queues empty
6504 * - clear the memory bitmaps
6506 * NOTE: pgdat should get zeroed by caller.
6507 * NOTE: this function is only called during early init.
6509 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6512 int nid
= pgdat
->node_id
;
6514 pgdat_init_internals(pgdat
);
6515 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6517 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6518 struct zone
*zone
= pgdat
->node_zones
+ j
;
6519 unsigned long size
, freesize
, memmap_pages
;
6520 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6522 size
= zone
->spanned_pages
;
6523 freesize
= zone
->present_pages
;
6526 * Adjust freesize so that it accounts for how much memory
6527 * is used by this zone for memmap. This affects the watermark
6528 * and per-cpu initialisations
6530 memmap_pages
= calc_memmap_size(size
, freesize
);
6531 if (!is_highmem_idx(j
)) {
6532 if (freesize
>= memmap_pages
) {
6533 freesize
-= memmap_pages
;
6536 " %s zone: %lu pages used for memmap\n",
6537 zone_names
[j
], memmap_pages
);
6539 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6540 zone_names
[j
], memmap_pages
, freesize
);
6543 /* Account for reserved pages */
6544 if (j
== 0 && freesize
> dma_reserve
) {
6545 freesize
-= dma_reserve
;
6546 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6547 zone_names
[0], dma_reserve
);
6550 if (!is_highmem_idx(j
))
6551 nr_kernel_pages
+= freesize
;
6552 /* Charge for highmem memmap if there are enough kernel pages */
6553 else if (nr_kernel_pages
> memmap_pages
* 2)
6554 nr_kernel_pages
-= memmap_pages
;
6555 nr_all_pages
+= freesize
;
6558 * Set an approximate value for lowmem here, it will be adjusted
6559 * when the bootmem allocator frees pages into the buddy system.
6560 * And all highmem pages will be managed by the buddy system.
6562 zone_init_internals(zone
, j
, nid
, freesize
);
6567 set_pageblock_order();
6568 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6569 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6570 memmap_init(size
, nid
, j
, zone_start_pfn
);
6574 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6575 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6577 unsigned long __maybe_unused start
= 0;
6578 unsigned long __maybe_unused offset
= 0;
6580 /* Skip empty nodes */
6581 if (!pgdat
->node_spanned_pages
)
6584 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6585 offset
= pgdat
->node_start_pfn
- start
;
6586 /* ia64 gets its own node_mem_map, before this, without bootmem */
6587 if (!pgdat
->node_mem_map
) {
6588 unsigned long size
, end
;
6592 * The zone's endpoints aren't required to be MAX_ORDER
6593 * aligned but the node_mem_map endpoints must be in order
6594 * for the buddy allocator to function correctly.
6596 end
= pgdat_end_pfn(pgdat
);
6597 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6598 size
= (end
- start
) * sizeof(struct page
);
6599 map
= memblock_alloc_node_nopanic(size
, pgdat
->node_id
);
6600 pgdat
->node_mem_map
= map
+ offset
;
6602 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6603 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6604 (unsigned long)pgdat
->node_mem_map
);
6605 #ifndef CONFIG_NEED_MULTIPLE_NODES
6607 * With no DISCONTIG, the global mem_map is just set as node 0's
6609 if (pgdat
== NODE_DATA(0)) {
6610 mem_map
= NODE_DATA(0)->node_mem_map
;
6611 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6612 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6614 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6619 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6620 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6622 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6623 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6625 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6628 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6631 void __init
free_area_init_node(int nid
, unsigned long *zones_size
,
6632 unsigned long node_start_pfn
,
6633 unsigned long *zholes_size
)
6635 pg_data_t
*pgdat
= NODE_DATA(nid
);
6636 unsigned long start_pfn
= 0;
6637 unsigned long end_pfn
= 0;
6639 /* pg_data_t should be reset to zero when it's allocated */
6640 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6642 pgdat
->node_id
= nid
;
6643 pgdat
->node_start_pfn
= node_start_pfn
;
6644 pgdat
->per_cpu_nodestats
= NULL
;
6645 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6646 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6647 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6648 (u64
)start_pfn
<< PAGE_SHIFT
,
6649 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6651 start_pfn
= node_start_pfn
;
6653 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6654 zones_size
, zholes_size
);
6656 alloc_node_mem_map(pgdat
);
6657 pgdat_set_deferred_range(pgdat
);
6659 free_area_init_core(pgdat
);
6662 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6664 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6667 static u64
zero_pfn_range(unsigned long spfn
, unsigned long epfn
)
6672 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6673 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6674 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6675 + pageblock_nr_pages
- 1;
6678 mm_zero_struct_page(pfn_to_page(pfn
));
6686 * Only struct pages that are backed by physical memory are zeroed and
6687 * initialized by going through __init_single_page(). But, there are some
6688 * struct pages which are reserved in memblock allocator and their fields
6689 * may be accessed (for example page_to_pfn() on some configuration accesses
6690 * flags). We must explicitly zero those struct pages.
6692 * This function also addresses a similar issue where struct pages are left
6693 * uninitialized because the physical address range is not covered by
6694 * memblock.memory or memblock.reserved. That could happen when memblock
6695 * layout is manually configured via memmap=.
6697 void __init
zero_resv_unavail(void)
6699 phys_addr_t start
, end
;
6701 phys_addr_t next
= 0;
6704 * Loop through unavailable ranges not covered by memblock.memory.
6707 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6708 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6710 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), PFN_UP(start
));
6713 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), max_pfn
);
6716 * Struct pages that do not have backing memory. This could be because
6717 * firmware is using some of this memory, or for some other reasons.
6720 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
6722 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6724 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6726 #if MAX_NUMNODES > 1
6728 * Figure out the number of possible node ids.
6730 void __init
setup_nr_node_ids(void)
6732 unsigned int highest
;
6734 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6735 nr_node_ids
= highest
+ 1;
6740 * node_map_pfn_alignment - determine the maximum internode alignment
6742 * This function should be called after node map is populated and sorted.
6743 * It calculates the maximum power of two alignment which can distinguish
6746 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6747 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6748 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6749 * shifted, 1GiB is enough and this function will indicate so.
6751 * This is used to test whether pfn -> nid mapping of the chosen memory
6752 * model has fine enough granularity to avoid incorrect mapping for the
6753 * populated node map.
6755 * Returns the determined alignment in pfn's. 0 if there is no alignment
6756 * requirement (single node).
6758 unsigned long __init
node_map_pfn_alignment(void)
6760 unsigned long accl_mask
= 0, last_end
= 0;
6761 unsigned long start
, end
, mask
;
6765 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6766 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6773 * Start with a mask granular enough to pin-point to the
6774 * start pfn and tick off bits one-by-one until it becomes
6775 * too coarse to separate the current node from the last.
6777 mask
= ~((1 << __ffs(start
)) - 1);
6778 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6781 /* accumulate all internode masks */
6785 /* convert mask to number of pages */
6786 return ~accl_mask
+ 1;
6789 /* Find the lowest pfn for a node */
6790 static unsigned long __init
find_min_pfn_for_node(int nid
)
6792 unsigned long min_pfn
= ULONG_MAX
;
6793 unsigned long start_pfn
;
6796 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6797 min_pfn
= min(min_pfn
, start_pfn
);
6799 if (min_pfn
== ULONG_MAX
) {
6800 pr_warn("Could not find start_pfn for node %d\n", nid
);
6808 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6810 * It returns the minimum PFN based on information provided via
6811 * memblock_set_node().
6813 unsigned long __init
find_min_pfn_with_active_regions(void)
6815 return find_min_pfn_for_node(MAX_NUMNODES
);
6819 * early_calculate_totalpages()
6820 * Sum pages in active regions for movable zone.
6821 * Populate N_MEMORY for calculating usable_nodes.
6823 static unsigned long __init
early_calculate_totalpages(void)
6825 unsigned long totalpages
= 0;
6826 unsigned long start_pfn
, end_pfn
;
6829 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6830 unsigned long pages
= end_pfn
- start_pfn
;
6832 totalpages
+= pages
;
6834 node_set_state(nid
, N_MEMORY
);
6840 * Find the PFN the Movable zone begins in each node. Kernel memory
6841 * is spread evenly between nodes as long as the nodes have enough
6842 * memory. When they don't, some nodes will have more kernelcore than
6845 static void __init
find_zone_movable_pfns_for_nodes(void)
6848 unsigned long usable_startpfn
;
6849 unsigned long kernelcore_node
, kernelcore_remaining
;
6850 /* save the state before borrow the nodemask */
6851 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6852 unsigned long totalpages
= early_calculate_totalpages();
6853 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6854 struct memblock_region
*r
;
6856 /* Need to find movable_zone earlier when movable_node is specified. */
6857 find_usable_zone_for_movable();
6860 * If movable_node is specified, ignore kernelcore and movablecore
6863 if (movable_node_is_enabled()) {
6864 for_each_memblock(memory
, r
) {
6865 if (!memblock_is_hotpluggable(r
))
6870 usable_startpfn
= PFN_DOWN(r
->base
);
6871 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6872 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6880 * If kernelcore=mirror is specified, ignore movablecore option
6882 if (mirrored_kernelcore
) {
6883 bool mem_below_4gb_not_mirrored
= false;
6885 for_each_memblock(memory
, r
) {
6886 if (memblock_is_mirror(r
))
6891 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6893 if (usable_startpfn
< 0x100000) {
6894 mem_below_4gb_not_mirrored
= true;
6898 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6899 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6903 if (mem_below_4gb_not_mirrored
)
6904 pr_warn("This configuration results in unmirrored kernel memory.");
6910 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6911 * amount of necessary memory.
6913 if (required_kernelcore_percent
)
6914 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
6916 if (required_movablecore_percent
)
6917 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
6921 * If movablecore= was specified, calculate what size of
6922 * kernelcore that corresponds so that memory usable for
6923 * any allocation type is evenly spread. If both kernelcore
6924 * and movablecore are specified, then the value of kernelcore
6925 * will be used for required_kernelcore if it's greater than
6926 * what movablecore would have allowed.
6928 if (required_movablecore
) {
6929 unsigned long corepages
;
6932 * Round-up so that ZONE_MOVABLE is at least as large as what
6933 * was requested by the user
6935 required_movablecore
=
6936 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6937 required_movablecore
= min(totalpages
, required_movablecore
);
6938 corepages
= totalpages
- required_movablecore
;
6940 required_kernelcore
= max(required_kernelcore
, corepages
);
6944 * If kernelcore was not specified or kernelcore size is larger
6945 * than totalpages, there is no ZONE_MOVABLE.
6947 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6950 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6951 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6954 /* Spread kernelcore memory as evenly as possible throughout nodes */
6955 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6956 for_each_node_state(nid
, N_MEMORY
) {
6957 unsigned long start_pfn
, end_pfn
;
6960 * Recalculate kernelcore_node if the division per node
6961 * now exceeds what is necessary to satisfy the requested
6962 * amount of memory for the kernel
6964 if (required_kernelcore
< kernelcore_node
)
6965 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6968 * As the map is walked, we track how much memory is usable
6969 * by the kernel using kernelcore_remaining. When it is
6970 * 0, the rest of the node is usable by ZONE_MOVABLE
6972 kernelcore_remaining
= kernelcore_node
;
6974 /* Go through each range of PFNs within this node */
6975 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6976 unsigned long size_pages
;
6978 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6979 if (start_pfn
>= end_pfn
)
6982 /* Account for what is only usable for kernelcore */
6983 if (start_pfn
< usable_startpfn
) {
6984 unsigned long kernel_pages
;
6985 kernel_pages
= min(end_pfn
, usable_startpfn
)
6988 kernelcore_remaining
-= min(kernel_pages
,
6989 kernelcore_remaining
);
6990 required_kernelcore
-= min(kernel_pages
,
6991 required_kernelcore
);
6993 /* Continue if range is now fully accounted */
6994 if (end_pfn
<= usable_startpfn
) {
6997 * Push zone_movable_pfn to the end so
6998 * that if we have to rebalance
6999 * kernelcore across nodes, we will
7000 * not double account here
7002 zone_movable_pfn
[nid
] = end_pfn
;
7005 start_pfn
= usable_startpfn
;
7009 * The usable PFN range for ZONE_MOVABLE is from
7010 * start_pfn->end_pfn. Calculate size_pages as the
7011 * number of pages used as kernelcore
7013 size_pages
= end_pfn
- start_pfn
;
7014 if (size_pages
> kernelcore_remaining
)
7015 size_pages
= kernelcore_remaining
;
7016 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7019 * Some kernelcore has been met, update counts and
7020 * break if the kernelcore for this node has been
7023 required_kernelcore
-= min(required_kernelcore
,
7025 kernelcore_remaining
-= size_pages
;
7026 if (!kernelcore_remaining
)
7032 * If there is still required_kernelcore, we do another pass with one
7033 * less node in the count. This will push zone_movable_pfn[nid] further
7034 * along on the nodes that still have memory until kernelcore is
7038 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7042 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7043 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7044 zone_movable_pfn
[nid
] =
7045 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7048 /* restore the node_state */
7049 node_states
[N_MEMORY
] = saved_node_state
;
7052 /* Any regular or high memory on that node ? */
7053 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7055 enum zone_type zone_type
;
7057 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7058 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7059 if (populated_zone(zone
)) {
7060 if (IS_ENABLED(CONFIG_HIGHMEM
))
7061 node_set_state(nid
, N_HIGH_MEMORY
);
7062 if (zone_type
<= ZONE_NORMAL
)
7063 node_set_state(nid
, N_NORMAL_MEMORY
);
7070 * free_area_init_nodes - Initialise all pg_data_t and zone data
7071 * @max_zone_pfn: an array of max PFNs for each zone
7073 * This will call free_area_init_node() for each active node in the system.
7074 * Using the page ranges provided by memblock_set_node(), the size of each
7075 * zone in each node and their holes is calculated. If the maximum PFN
7076 * between two adjacent zones match, it is assumed that the zone is empty.
7077 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7078 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7079 * starts where the previous one ended. For example, ZONE_DMA32 starts
7080 * at arch_max_dma_pfn.
7082 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
7084 unsigned long start_pfn
, end_pfn
;
7087 /* Record where the zone boundaries are */
7088 memset(arch_zone_lowest_possible_pfn
, 0,
7089 sizeof(arch_zone_lowest_possible_pfn
));
7090 memset(arch_zone_highest_possible_pfn
, 0,
7091 sizeof(arch_zone_highest_possible_pfn
));
7093 start_pfn
= find_min_pfn_with_active_regions();
7095 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7096 if (i
== ZONE_MOVABLE
)
7099 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
7100 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
7101 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
7103 start_pfn
= end_pfn
;
7106 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7107 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7108 find_zone_movable_pfns_for_nodes();
7110 /* Print out the zone ranges */
7111 pr_info("Zone ranges:\n");
7112 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7113 if (i
== ZONE_MOVABLE
)
7115 pr_info(" %-8s ", zone_names
[i
]);
7116 if (arch_zone_lowest_possible_pfn
[i
] ==
7117 arch_zone_highest_possible_pfn
[i
])
7120 pr_cont("[mem %#018Lx-%#018Lx]\n",
7121 (u64
)arch_zone_lowest_possible_pfn
[i
]
7123 ((u64
)arch_zone_highest_possible_pfn
[i
]
7124 << PAGE_SHIFT
) - 1);
7127 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7128 pr_info("Movable zone start for each node\n");
7129 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7130 if (zone_movable_pfn
[i
])
7131 pr_info(" Node %d: %#018Lx\n", i
,
7132 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7135 /* Print out the early node map */
7136 pr_info("Early memory node ranges\n");
7137 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
7138 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7139 (u64
)start_pfn
<< PAGE_SHIFT
,
7140 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7142 /* Initialise every node */
7143 mminit_verify_pageflags_layout();
7144 setup_nr_node_ids();
7145 zero_resv_unavail();
7146 for_each_online_node(nid
) {
7147 pg_data_t
*pgdat
= NODE_DATA(nid
);
7148 free_area_init_node(nid
, NULL
,
7149 find_min_pfn_for_node(nid
), NULL
);
7151 /* Any memory on that node */
7152 if (pgdat
->node_present_pages
)
7153 node_set_state(nid
, N_MEMORY
);
7154 check_for_memory(pgdat
, nid
);
7158 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7159 unsigned long *percent
)
7161 unsigned long long coremem
;
7167 /* Value may be a percentage of total memory, otherwise bytes */
7168 coremem
= simple_strtoull(p
, &endptr
, 0);
7169 if (*endptr
== '%') {
7170 /* Paranoid check for percent values greater than 100 */
7171 WARN_ON(coremem
> 100);
7175 coremem
= memparse(p
, &p
);
7176 /* Paranoid check that UL is enough for the coremem value */
7177 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7179 *core
= coremem
>> PAGE_SHIFT
;
7186 * kernelcore=size sets the amount of memory for use for allocations that
7187 * cannot be reclaimed or migrated.
7189 static int __init
cmdline_parse_kernelcore(char *p
)
7191 /* parse kernelcore=mirror */
7192 if (parse_option_str(p
, "mirror")) {
7193 mirrored_kernelcore
= true;
7197 return cmdline_parse_core(p
, &required_kernelcore
,
7198 &required_kernelcore_percent
);
7202 * movablecore=size sets the amount of memory for use for allocations that
7203 * can be reclaimed or migrated.
7205 static int __init
cmdline_parse_movablecore(char *p
)
7207 return cmdline_parse_core(p
, &required_movablecore
,
7208 &required_movablecore_percent
);
7211 early_param("kernelcore", cmdline_parse_kernelcore
);
7212 early_param("movablecore", cmdline_parse_movablecore
);
7214 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7216 void adjust_managed_page_count(struct page
*page
, long count
)
7218 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7219 totalram_pages_add(count
);
7220 #ifdef CONFIG_HIGHMEM
7221 if (PageHighMem(page
))
7222 totalhigh_pages_add(count
);
7225 EXPORT_SYMBOL(adjust_managed_page_count
);
7227 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7230 unsigned long pages
= 0;
7232 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7233 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7234 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7235 struct page
*page
= virt_to_page(pos
);
7236 void *direct_map_addr
;
7239 * 'direct_map_addr' might be different from 'pos'
7240 * because some architectures' virt_to_page()
7241 * work with aliases. Getting the direct map
7242 * address ensures that we get a _writeable_
7243 * alias for the memset().
7245 direct_map_addr
= page_address(page
);
7246 if ((unsigned int)poison
<= 0xFF)
7247 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7249 free_reserved_page(page
);
7253 pr_info("Freeing %s memory: %ldK\n",
7254 s
, pages
<< (PAGE_SHIFT
- 10));
7258 EXPORT_SYMBOL(free_reserved_area
);
7260 #ifdef CONFIG_HIGHMEM
7261 void free_highmem_page(struct page
*page
)
7263 __free_reserved_page(page
);
7264 totalram_pages_inc();
7265 atomic_long_inc(&page_zone(page
)->managed_pages
);
7266 totalhigh_pages_inc();
7271 void __init
mem_init_print_info(const char *str
)
7273 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7274 unsigned long init_code_size
, init_data_size
;
7276 physpages
= get_num_physpages();
7277 codesize
= _etext
- _stext
;
7278 datasize
= _edata
- _sdata
;
7279 rosize
= __end_rodata
- __start_rodata
;
7280 bss_size
= __bss_stop
- __bss_start
;
7281 init_data_size
= __init_end
- __init_begin
;
7282 init_code_size
= _einittext
- _sinittext
;
7285 * Detect special cases and adjust section sizes accordingly:
7286 * 1) .init.* may be embedded into .data sections
7287 * 2) .init.text.* may be out of [__init_begin, __init_end],
7288 * please refer to arch/tile/kernel/vmlinux.lds.S.
7289 * 3) .rodata.* may be embedded into .text or .data sections.
7291 #define adj_init_size(start, end, size, pos, adj) \
7293 if (start <= pos && pos < end && size > adj) \
7297 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7298 _sinittext
, init_code_size
);
7299 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7300 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7301 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7302 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7304 #undef adj_init_size
7306 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7307 #ifdef CONFIG_HIGHMEM
7311 nr_free_pages() << (PAGE_SHIFT
- 10),
7312 physpages
<< (PAGE_SHIFT
- 10),
7313 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7314 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7315 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7316 totalcma_pages
<< (PAGE_SHIFT
- 10),
7317 #ifdef CONFIG_HIGHMEM
7318 totalhigh_pages() << (PAGE_SHIFT
- 10),
7320 str
? ", " : "", str
? str
: "");
7324 * set_dma_reserve - set the specified number of pages reserved in the first zone
7325 * @new_dma_reserve: The number of pages to mark reserved
7327 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7328 * In the DMA zone, a significant percentage may be consumed by kernel image
7329 * and other unfreeable allocations which can skew the watermarks badly. This
7330 * function may optionally be used to account for unfreeable pages in the
7331 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7332 * smaller per-cpu batchsize.
7334 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7336 dma_reserve
= new_dma_reserve
;
7339 void __init
free_area_init(unsigned long *zones_size
)
7341 zero_resv_unavail();
7342 free_area_init_node(0, zones_size
,
7343 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
7346 static int page_alloc_cpu_dead(unsigned int cpu
)
7349 lru_add_drain_cpu(cpu
);
7353 * Spill the event counters of the dead processor
7354 * into the current processors event counters.
7355 * This artificially elevates the count of the current
7358 vm_events_fold_cpu(cpu
);
7361 * Zero the differential counters of the dead processor
7362 * so that the vm statistics are consistent.
7364 * This is only okay since the processor is dead and cannot
7365 * race with what we are doing.
7367 cpu_vm_stats_fold(cpu
);
7371 void __init
page_alloc_init(void)
7375 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7376 "mm/page_alloc:dead", NULL
,
7377 page_alloc_cpu_dead
);
7382 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7383 * or min_free_kbytes changes.
7385 static void calculate_totalreserve_pages(void)
7387 struct pglist_data
*pgdat
;
7388 unsigned long reserve_pages
= 0;
7389 enum zone_type i
, j
;
7391 for_each_online_pgdat(pgdat
) {
7393 pgdat
->totalreserve_pages
= 0;
7395 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7396 struct zone
*zone
= pgdat
->node_zones
+ i
;
7398 unsigned long managed_pages
= zone_managed_pages(zone
);
7400 /* Find valid and maximum lowmem_reserve in the zone */
7401 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7402 if (zone
->lowmem_reserve
[j
] > max
)
7403 max
= zone
->lowmem_reserve
[j
];
7406 /* we treat the high watermark as reserved pages. */
7407 max
+= high_wmark_pages(zone
);
7409 if (max
> managed_pages
)
7410 max
= managed_pages
;
7412 pgdat
->totalreserve_pages
+= max
;
7414 reserve_pages
+= max
;
7417 totalreserve_pages
= reserve_pages
;
7421 * setup_per_zone_lowmem_reserve - called whenever
7422 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7423 * has a correct pages reserved value, so an adequate number of
7424 * pages are left in the zone after a successful __alloc_pages().
7426 static void setup_per_zone_lowmem_reserve(void)
7428 struct pglist_data
*pgdat
;
7429 enum zone_type j
, idx
;
7431 for_each_online_pgdat(pgdat
) {
7432 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7433 struct zone
*zone
= pgdat
->node_zones
+ j
;
7434 unsigned long managed_pages
= zone_managed_pages(zone
);
7436 zone
->lowmem_reserve
[j
] = 0;
7440 struct zone
*lower_zone
;
7443 lower_zone
= pgdat
->node_zones
+ idx
;
7445 if (sysctl_lowmem_reserve_ratio
[idx
] < 1) {
7446 sysctl_lowmem_reserve_ratio
[idx
] = 0;
7447 lower_zone
->lowmem_reserve
[j
] = 0;
7449 lower_zone
->lowmem_reserve
[j
] =
7450 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7452 managed_pages
+= zone_managed_pages(lower_zone
);
7457 /* update totalreserve_pages */
7458 calculate_totalreserve_pages();
7461 static void __setup_per_zone_wmarks(void)
7463 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7464 unsigned long lowmem_pages
= 0;
7466 unsigned long flags
;
7468 /* Calculate total number of !ZONE_HIGHMEM pages */
7469 for_each_zone(zone
) {
7470 if (!is_highmem(zone
))
7471 lowmem_pages
+= zone_managed_pages(zone
);
7474 for_each_zone(zone
) {
7477 spin_lock_irqsave(&zone
->lock
, flags
);
7478 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7479 do_div(tmp
, lowmem_pages
);
7480 if (is_highmem(zone
)) {
7482 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7483 * need highmem pages, so cap pages_min to a small
7486 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7487 * deltas control asynch page reclaim, and so should
7488 * not be capped for highmem.
7490 unsigned long min_pages
;
7492 min_pages
= zone_managed_pages(zone
) / 1024;
7493 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7494 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7497 * If it's a lowmem zone, reserve a number of pages
7498 * proportionate to the zone's size.
7500 zone
->_watermark
[WMARK_MIN
] = tmp
;
7504 * Set the kswapd watermarks distance according to the
7505 * scale factor in proportion to available memory, but
7506 * ensure a minimum size on small systems.
7508 tmp
= max_t(u64
, tmp
>> 2,
7509 mult_frac(zone_managed_pages(zone
),
7510 watermark_scale_factor
, 10000));
7512 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7513 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7514 zone
->watermark_boost
= 0;
7516 spin_unlock_irqrestore(&zone
->lock
, flags
);
7519 /* update totalreserve_pages */
7520 calculate_totalreserve_pages();
7524 * setup_per_zone_wmarks - called when min_free_kbytes changes
7525 * or when memory is hot-{added|removed}
7527 * Ensures that the watermark[min,low,high] values for each zone are set
7528 * correctly with respect to min_free_kbytes.
7530 void setup_per_zone_wmarks(void)
7532 static DEFINE_SPINLOCK(lock
);
7535 __setup_per_zone_wmarks();
7540 * Initialise min_free_kbytes.
7542 * For small machines we want it small (128k min). For large machines
7543 * we want it large (64MB max). But it is not linear, because network
7544 * bandwidth does not increase linearly with machine size. We use
7546 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7547 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7563 int __meminit
init_per_zone_wmark_min(void)
7565 unsigned long lowmem_kbytes
;
7566 int new_min_free_kbytes
;
7568 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7569 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7571 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7572 min_free_kbytes
= new_min_free_kbytes
;
7573 if (min_free_kbytes
< 128)
7574 min_free_kbytes
= 128;
7575 if (min_free_kbytes
> 65536)
7576 min_free_kbytes
= 65536;
7578 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7579 new_min_free_kbytes
, user_min_free_kbytes
);
7581 setup_per_zone_wmarks();
7582 refresh_zone_stat_thresholds();
7583 setup_per_zone_lowmem_reserve();
7586 setup_min_unmapped_ratio();
7587 setup_min_slab_ratio();
7592 core_initcall(init_per_zone_wmark_min
)
7595 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7596 * that we can call two helper functions whenever min_free_kbytes
7599 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7600 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7604 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7609 user_min_free_kbytes
= min_free_kbytes
;
7610 setup_per_zone_wmarks();
7615 int watermark_boost_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7616 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7620 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7627 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7628 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7632 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7637 setup_per_zone_wmarks();
7643 static void setup_min_unmapped_ratio(void)
7648 for_each_online_pgdat(pgdat
)
7649 pgdat
->min_unmapped_pages
= 0;
7652 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
7653 sysctl_min_unmapped_ratio
) / 100;
7657 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7658 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7662 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7666 setup_min_unmapped_ratio();
7671 static void setup_min_slab_ratio(void)
7676 for_each_online_pgdat(pgdat
)
7677 pgdat
->min_slab_pages
= 0;
7680 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
7681 sysctl_min_slab_ratio
) / 100;
7684 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7685 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7689 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7693 setup_min_slab_ratio();
7700 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7701 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7702 * whenever sysctl_lowmem_reserve_ratio changes.
7704 * The reserve ratio obviously has absolutely no relation with the
7705 * minimum watermarks. The lowmem reserve ratio can only make sense
7706 * if in function of the boot time zone sizes.
7708 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7709 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7711 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7712 setup_per_zone_lowmem_reserve();
7717 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7718 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7719 * pagelist can have before it gets flushed back to buddy allocator.
7721 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7722 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7725 int old_percpu_pagelist_fraction
;
7728 mutex_lock(&pcp_batch_high_lock
);
7729 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7731 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7732 if (!write
|| ret
< 0)
7735 /* Sanity checking to avoid pcp imbalance */
7736 if (percpu_pagelist_fraction
&&
7737 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7738 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7744 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7747 for_each_populated_zone(zone
) {
7750 for_each_possible_cpu(cpu
)
7751 pageset_set_high_and_batch(zone
,
7752 per_cpu_ptr(zone
->pageset
, cpu
));
7755 mutex_unlock(&pcp_batch_high_lock
);
7760 int hashdist
= HASHDIST_DEFAULT
;
7762 static int __init
set_hashdist(char *str
)
7766 hashdist
= simple_strtoul(str
, &str
, 0);
7769 __setup("hashdist=", set_hashdist
);
7772 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7774 * Returns the number of pages that arch has reserved but
7775 * is not known to alloc_large_system_hash().
7777 static unsigned long __init
arch_reserved_kernel_pages(void)
7784 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7785 * machines. As memory size is increased the scale is also increased but at
7786 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7787 * quadruples the scale is increased by one, which means the size of hash table
7788 * only doubles, instead of quadrupling as well.
7789 * Because 32-bit systems cannot have large physical memory, where this scaling
7790 * makes sense, it is disabled on such platforms.
7792 #if __BITS_PER_LONG > 32
7793 #define ADAPT_SCALE_BASE (64ul << 30)
7794 #define ADAPT_SCALE_SHIFT 2
7795 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7799 * allocate a large system hash table from bootmem
7800 * - it is assumed that the hash table must contain an exact power-of-2
7801 * quantity of entries
7802 * - limit is the number of hash buckets, not the total allocation size
7804 void *__init
alloc_large_system_hash(const char *tablename
,
7805 unsigned long bucketsize
,
7806 unsigned long numentries
,
7809 unsigned int *_hash_shift
,
7810 unsigned int *_hash_mask
,
7811 unsigned long low_limit
,
7812 unsigned long high_limit
)
7814 unsigned long long max
= high_limit
;
7815 unsigned long log2qty
, size
;
7819 /* allow the kernel cmdline to have a say */
7821 /* round applicable memory size up to nearest megabyte */
7822 numentries
= nr_kernel_pages
;
7823 numentries
-= arch_reserved_kernel_pages();
7825 /* It isn't necessary when PAGE_SIZE >= 1MB */
7826 if (PAGE_SHIFT
< 20)
7827 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7829 #if __BITS_PER_LONG > 32
7831 unsigned long adapt
;
7833 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
7834 adapt
<<= ADAPT_SCALE_SHIFT
)
7839 /* limit to 1 bucket per 2^scale bytes of low memory */
7840 if (scale
> PAGE_SHIFT
)
7841 numentries
>>= (scale
- PAGE_SHIFT
);
7843 numentries
<<= (PAGE_SHIFT
- scale
);
7845 /* Make sure we've got at least a 0-order allocation.. */
7846 if (unlikely(flags
& HASH_SMALL
)) {
7847 /* Makes no sense without HASH_EARLY */
7848 WARN_ON(!(flags
& HASH_EARLY
));
7849 if (!(numentries
>> *_hash_shift
)) {
7850 numentries
= 1UL << *_hash_shift
;
7851 BUG_ON(!numentries
);
7853 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7854 numentries
= PAGE_SIZE
/ bucketsize
;
7856 numentries
= roundup_pow_of_two(numentries
);
7858 /* limit allocation size to 1/16 total memory by default */
7860 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7861 do_div(max
, bucketsize
);
7863 max
= min(max
, 0x80000000ULL
);
7865 if (numentries
< low_limit
)
7866 numentries
= low_limit
;
7867 if (numentries
> max
)
7870 log2qty
= ilog2(numentries
);
7872 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
7874 size
= bucketsize
<< log2qty
;
7875 if (flags
& HASH_EARLY
) {
7876 if (flags
& HASH_ZERO
)
7877 table
= memblock_alloc_nopanic(size
,
7880 table
= memblock_alloc_raw(size
,
7882 } else if (hashdist
) {
7883 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
7886 * If bucketsize is not a power-of-two, we may free
7887 * some pages at the end of hash table which
7888 * alloc_pages_exact() automatically does
7890 if (get_order(size
) < MAX_ORDER
) {
7891 table
= alloc_pages_exact(size
, gfp_flags
);
7892 kmemleak_alloc(table
, size
, 1, gfp_flags
);
7895 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7898 panic("Failed to allocate %s hash table\n", tablename
);
7900 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7901 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7904 *_hash_shift
= log2qty
;
7906 *_hash_mask
= (1 << log2qty
) - 1;
7912 * This function checks whether pageblock includes unmovable pages or not.
7913 * If @count is not zero, it is okay to include less @count unmovable pages
7915 * PageLRU check without isolation or lru_lock could race so that
7916 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7917 * check without lock_page also may miss some movable non-lru pages at
7918 * race condition. So you can't expect this function should be exact.
7920 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7921 int migratetype
, int flags
)
7923 unsigned long pfn
, iter
, found
;
7926 * TODO we could make this much more efficient by not checking every
7927 * page in the range if we know all of them are in MOVABLE_ZONE and
7928 * that the movable zone guarantees that pages are migratable but
7929 * the later is not the case right now unfortunatelly. E.g. movablecore
7930 * can still lead to having bootmem allocations in zone_movable.
7934 * CMA allocations (alloc_contig_range) really need to mark isolate
7935 * CMA pageblocks even when they are not movable in fact so consider
7936 * them movable here.
7938 if (is_migrate_cma(migratetype
) &&
7939 is_migrate_cma(get_pageblock_migratetype(page
)))
7942 pfn
= page_to_pfn(page
);
7943 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7944 unsigned long check
= pfn
+ iter
;
7946 if (!pfn_valid_within(check
))
7949 page
= pfn_to_page(check
);
7951 if (PageReserved(page
))
7955 * If the zone is movable and we have ruled out all reserved
7956 * pages then it should be reasonably safe to assume the rest
7959 if (zone_idx(zone
) == ZONE_MOVABLE
)
7963 * Hugepages are not in LRU lists, but they're movable.
7964 * We need not scan over tail pages bacause we don't
7965 * handle each tail page individually in migration.
7967 if (PageHuge(page
)) {
7968 struct page
*head
= compound_head(page
);
7969 unsigned int skip_pages
;
7971 if (!hugepage_migration_supported(page_hstate(head
)))
7974 skip_pages
= (1 << compound_order(head
)) - (page
- head
);
7975 iter
+= skip_pages
- 1;
7980 * We can't use page_count without pin a page
7981 * because another CPU can free compound page.
7982 * This check already skips compound tails of THP
7983 * because their page->_refcount is zero at all time.
7985 if (!page_ref_count(page
)) {
7986 if (PageBuddy(page
))
7987 iter
+= (1 << page_order(page
)) - 1;
7992 * The HWPoisoned page may be not in buddy system, and
7993 * page_count() is not 0.
7995 if ((flags
& SKIP_HWPOISON
) && PageHWPoison(page
))
7998 if (__PageMovable(page
))
8004 * If there are RECLAIMABLE pages, we need to check
8005 * it. But now, memory offline itself doesn't call
8006 * shrink_node_slabs() and it still to be fixed.
8009 * If the page is not RAM, page_count()should be 0.
8010 * we don't need more check. This is an _used_ not-movable page.
8012 * The problematic thing here is PG_reserved pages. PG_reserved
8013 * is set to both of a memory hole page and a _used_ kernel
8021 WARN_ON_ONCE(zone_idx(zone
) == ZONE_MOVABLE
);
8022 if (flags
& REPORT_FAILURE
)
8023 dump_page(pfn_to_page(pfn
+iter
), "unmovable page");
8027 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
8029 static unsigned long pfn_max_align_down(unsigned long pfn
)
8031 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8032 pageblock_nr_pages
) - 1);
8035 static unsigned long pfn_max_align_up(unsigned long pfn
)
8037 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8038 pageblock_nr_pages
));
8041 /* [start, end) must belong to a single zone. */
8042 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8043 unsigned long start
, unsigned long end
)
8045 /* This function is based on compact_zone() from compaction.c. */
8046 unsigned long nr_reclaimed
;
8047 unsigned long pfn
= start
;
8048 unsigned int tries
= 0;
8053 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8054 if (fatal_signal_pending(current
)) {
8059 if (list_empty(&cc
->migratepages
)) {
8060 cc
->nr_migratepages
= 0;
8061 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8067 } else if (++tries
== 5) {
8068 ret
= ret
< 0 ? ret
: -EBUSY
;
8072 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8074 cc
->nr_migratepages
-= nr_reclaimed
;
8076 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
8077 NULL
, 0, cc
->mode
, MR_CONTIG_RANGE
);
8080 putback_movable_pages(&cc
->migratepages
);
8087 * alloc_contig_range() -- tries to allocate given range of pages
8088 * @start: start PFN to allocate
8089 * @end: one-past-the-last PFN to allocate
8090 * @migratetype: migratetype of the underlaying pageblocks (either
8091 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8092 * in range must have the same migratetype and it must
8093 * be either of the two.
8094 * @gfp_mask: GFP mask to use during compaction
8096 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8097 * aligned. The PFN range must belong to a single zone.
8099 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8100 * pageblocks in the range. Once isolated, the pageblocks should not
8101 * be modified by others.
8103 * Returns zero on success or negative error code. On success all
8104 * pages which PFN is in [start, end) are allocated for the caller and
8105 * need to be freed with free_contig_range().
8107 int alloc_contig_range(unsigned long start
, unsigned long end
,
8108 unsigned migratetype
, gfp_t gfp_mask
)
8110 unsigned long outer_start
, outer_end
;
8114 struct compact_control cc
= {
8115 .nr_migratepages
= 0,
8117 .zone
= page_zone(pfn_to_page(start
)),
8118 .mode
= MIGRATE_SYNC
,
8119 .ignore_skip_hint
= true,
8120 .no_set_skip_hint
= true,
8121 .gfp_mask
= current_gfp_context(gfp_mask
),
8123 INIT_LIST_HEAD(&cc
.migratepages
);
8126 * What we do here is we mark all pageblocks in range as
8127 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8128 * have different sizes, and due to the way page allocator
8129 * work, we align the range to biggest of the two pages so
8130 * that page allocator won't try to merge buddies from
8131 * different pageblocks and change MIGRATE_ISOLATE to some
8132 * other migration type.
8134 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8135 * migrate the pages from an unaligned range (ie. pages that
8136 * we are interested in). This will put all the pages in
8137 * range back to page allocator as MIGRATE_ISOLATE.
8139 * When this is done, we take the pages in range from page
8140 * allocator removing them from the buddy system. This way
8141 * page allocator will never consider using them.
8143 * This lets us mark the pageblocks back as
8144 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8145 * aligned range but not in the unaligned, original range are
8146 * put back to page allocator so that buddy can use them.
8149 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8150 pfn_max_align_up(end
), migratetype
, 0);
8155 * In case of -EBUSY, we'd like to know which page causes problem.
8156 * So, just fall through. test_pages_isolated() has a tracepoint
8157 * which will report the busy page.
8159 * It is possible that busy pages could become available before
8160 * the call to test_pages_isolated, and the range will actually be
8161 * allocated. So, if we fall through be sure to clear ret so that
8162 * -EBUSY is not accidentally used or returned to caller.
8164 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8165 if (ret
&& ret
!= -EBUSY
)
8170 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8171 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8172 * more, all pages in [start, end) are free in page allocator.
8173 * What we are going to do is to allocate all pages from
8174 * [start, end) (that is remove them from page allocator).
8176 * The only problem is that pages at the beginning and at the
8177 * end of interesting range may be not aligned with pages that
8178 * page allocator holds, ie. they can be part of higher order
8179 * pages. Because of this, we reserve the bigger range and
8180 * once this is done free the pages we are not interested in.
8182 * We don't have to hold zone->lock here because the pages are
8183 * isolated thus they won't get removed from buddy.
8186 lru_add_drain_all();
8187 drain_all_pages(cc
.zone
);
8190 outer_start
= start
;
8191 while (!PageBuddy(pfn_to_page(outer_start
))) {
8192 if (++order
>= MAX_ORDER
) {
8193 outer_start
= start
;
8196 outer_start
&= ~0UL << order
;
8199 if (outer_start
!= start
) {
8200 order
= page_order(pfn_to_page(outer_start
));
8203 * outer_start page could be small order buddy page and
8204 * it doesn't include start page. Adjust outer_start
8205 * in this case to report failed page properly
8206 * on tracepoint in test_pages_isolated()
8208 if (outer_start
+ (1UL << order
) <= start
)
8209 outer_start
= start
;
8212 /* Make sure the range is really isolated. */
8213 if (test_pages_isolated(outer_start
, end
, false)) {
8214 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8215 __func__
, outer_start
, end
);
8220 /* Grab isolated pages from freelists. */
8221 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8227 /* Free head and tail (if any) */
8228 if (start
!= outer_start
)
8229 free_contig_range(outer_start
, start
- outer_start
);
8230 if (end
!= outer_end
)
8231 free_contig_range(end
, outer_end
- end
);
8234 undo_isolate_page_range(pfn_max_align_down(start
),
8235 pfn_max_align_up(end
), migratetype
);
8239 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
8241 unsigned int count
= 0;
8243 for (; nr_pages
--; pfn
++) {
8244 struct page
*page
= pfn_to_page(pfn
);
8246 count
+= page_count(page
) != 1;
8249 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8253 #ifdef CONFIG_MEMORY_HOTPLUG
8255 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8256 * page high values need to be recalulated.
8258 void __meminit
zone_pcp_update(struct zone
*zone
)
8261 mutex_lock(&pcp_batch_high_lock
);
8262 for_each_possible_cpu(cpu
)
8263 pageset_set_high_and_batch(zone
,
8264 per_cpu_ptr(zone
->pageset
, cpu
));
8265 mutex_unlock(&pcp_batch_high_lock
);
8269 void zone_pcp_reset(struct zone
*zone
)
8271 unsigned long flags
;
8273 struct per_cpu_pageset
*pset
;
8275 /* avoid races with drain_pages() */
8276 local_irq_save(flags
);
8277 if (zone
->pageset
!= &boot_pageset
) {
8278 for_each_online_cpu(cpu
) {
8279 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8280 drain_zonestat(zone
, pset
);
8282 free_percpu(zone
->pageset
);
8283 zone
->pageset
= &boot_pageset
;
8285 local_irq_restore(flags
);
8288 #ifdef CONFIG_MEMORY_HOTREMOVE
8290 * All pages in the range must be in a single zone and isolated
8291 * before calling this.
8294 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8298 unsigned int order
, i
;
8300 unsigned long flags
;
8301 /* find the first valid pfn */
8302 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8307 offline_mem_sections(pfn
, end_pfn
);
8308 zone
= page_zone(pfn_to_page(pfn
));
8309 spin_lock_irqsave(&zone
->lock
, flags
);
8311 while (pfn
< end_pfn
) {
8312 if (!pfn_valid(pfn
)) {
8316 page
= pfn_to_page(pfn
);
8318 * The HWPoisoned page may be not in buddy system, and
8319 * page_count() is not 0.
8321 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8323 SetPageReserved(page
);
8327 BUG_ON(page_count(page
));
8328 BUG_ON(!PageBuddy(page
));
8329 order
= page_order(page
);
8330 #ifdef CONFIG_DEBUG_VM
8331 pr_info("remove from free list %lx %d %lx\n",
8332 pfn
, 1 << order
, end_pfn
);
8334 list_del(&page
->lru
);
8335 rmv_page_order(page
);
8336 zone
->free_area
[order
].nr_free
--;
8337 for (i
= 0; i
< (1 << order
); i
++)
8338 SetPageReserved((page
+i
));
8339 pfn
+= (1 << order
);
8341 spin_unlock_irqrestore(&zone
->lock
, flags
);
8345 bool is_free_buddy_page(struct page
*page
)
8347 struct zone
*zone
= page_zone(page
);
8348 unsigned long pfn
= page_to_pfn(page
);
8349 unsigned long flags
;
8352 spin_lock_irqsave(&zone
->lock
, flags
);
8353 for (order
= 0; order
< MAX_ORDER
; order
++) {
8354 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8356 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8359 spin_unlock_irqrestore(&zone
->lock
, flags
);
8361 return order
< MAX_ORDER
;
8364 #ifdef CONFIG_MEMORY_FAILURE
8366 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8367 * test is performed under the zone lock to prevent a race against page
8370 bool set_hwpoison_free_buddy_page(struct page
*page
)
8372 struct zone
*zone
= page_zone(page
);
8373 unsigned long pfn
= page_to_pfn(page
);
8374 unsigned long flags
;
8376 bool hwpoisoned
= false;
8378 spin_lock_irqsave(&zone
->lock
, flags
);
8379 for (order
= 0; order
< MAX_ORDER
; order
++) {
8380 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8382 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8383 if (!TestSetPageHWPoison(page
))
8388 spin_unlock_irqrestore(&zone
->lock
, flags
);