2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/memblock.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kasan.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/ratelimit.h>
33 #include <linux/oom.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/vmstat.h>
42 #include <linux/mempolicy.h>
43 #include <linux/memremap.h>
44 #include <linux/stop_machine.h>
45 #include <linux/sort.h>
46 #include <linux/pfn.h>
47 #include <linux/backing-dev.h>
48 #include <linux/fault-inject.h>
49 #include <linux/page-isolation.h>
50 #include <linux/page_ext.h>
51 #include <linux/debugobjects.h>
52 #include <linux/kmemleak.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <trace/events/oom.h>
56 #include <linux/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/hugetlb.h>
60 #include <linux/sched/rt.h>
61 #include <linux/sched/mm.h>
62 #include <linux/page_owner.h>
63 #include <linux/kthread.h>
64 #include <linux/memcontrol.h>
65 #include <linux/ftrace.h>
66 #include <linux/lockdep.h>
67 #include <linux/nmi.h>
68 #include <linux/psi.h>
70 #include <asm/sections.h>
71 #include <asm/tlbflush.h>
72 #include <asm/div64.h>
75 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
76 static DEFINE_MUTEX(pcp_batch_high_lock
);
77 #define MIN_PERCPU_PAGELIST_FRACTION (8)
79 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
80 DEFINE_PER_CPU(int, numa_node
);
81 EXPORT_PER_CPU_SYMBOL(numa_node
);
84 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
86 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
88 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
89 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
90 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
91 * defined in <linux/topology.h>.
93 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
94 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
95 int _node_numa_mem_
[MAX_NUMNODES
];
98 /* work_structs for global per-cpu drains */
99 DEFINE_MUTEX(pcpu_drain_mutex
);
100 DEFINE_PER_CPU(struct work_struct
, pcpu_drain
);
102 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
103 volatile unsigned long latent_entropy __latent_entropy
;
104 EXPORT_SYMBOL(latent_entropy
);
108 * Array of node states.
110 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
111 [N_POSSIBLE
] = NODE_MASK_ALL
,
112 [N_ONLINE
] = { { [0] = 1UL } },
114 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
115 #ifdef CONFIG_HIGHMEM
116 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
118 [N_MEMORY
] = { { [0] = 1UL } },
119 [N_CPU
] = { { [0] = 1UL } },
122 EXPORT_SYMBOL(node_states
);
124 /* Protect totalram_pages and zone->managed_pages */
125 static DEFINE_SPINLOCK(managed_page_count_lock
);
127 unsigned long totalram_pages __read_mostly
;
128 unsigned long totalreserve_pages __read_mostly
;
129 unsigned long totalcma_pages __read_mostly
;
131 int percpu_pagelist_fraction
;
132 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
135 * A cached value of the page's pageblock's migratetype, used when the page is
136 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
137 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
138 * Also the migratetype set in the page does not necessarily match the pcplist
139 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
140 * other index - this ensures that it will be put on the correct CMA freelist.
142 static inline int get_pcppage_migratetype(struct page
*page
)
147 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
149 page
->index
= migratetype
;
152 #ifdef CONFIG_PM_SLEEP
154 * The following functions are used by the suspend/hibernate code to temporarily
155 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
156 * while devices are suspended. To avoid races with the suspend/hibernate code,
157 * they should always be called with system_transition_mutex held
158 * (gfp_allowed_mask also should only be modified with system_transition_mutex
159 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
160 * with that modification).
163 static gfp_t saved_gfp_mask
;
165 void pm_restore_gfp_mask(void)
167 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
168 if (saved_gfp_mask
) {
169 gfp_allowed_mask
= saved_gfp_mask
;
174 void pm_restrict_gfp_mask(void)
176 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
177 WARN_ON(saved_gfp_mask
);
178 saved_gfp_mask
= gfp_allowed_mask
;
179 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
182 bool pm_suspended_storage(void)
184 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
188 #endif /* CONFIG_PM_SLEEP */
190 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
191 unsigned int pageblock_order __read_mostly
;
194 static void __free_pages_ok(struct page
*page
, unsigned int order
);
197 * results with 256, 32 in the lowmem_reserve sysctl:
198 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
199 * 1G machine -> (16M dma, 784M normal, 224M high)
200 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
201 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
202 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
204 * TBD: should special case ZONE_DMA32 machines here - in those we normally
205 * don't need any ZONE_NORMAL reservation
207 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
208 #ifdef CONFIG_ZONE_DMA
211 #ifdef CONFIG_ZONE_DMA32
215 #ifdef CONFIG_HIGHMEM
221 EXPORT_SYMBOL(totalram_pages
);
223 static char * const zone_names
[MAX_NR_ZONES
] = {
224 #ifdef CONFIG_ZONE_DMA
227 #ifdef CONFIG_ZONE_DMA32
231 #ifdef CONFIG_HIGHMEM
235 #ifdef CONFIG_ZONE_DEVICE
240 char * const migratetype_names
[MIGRATE_TYPES
] = {
248 #ifdef CONFIG_MEMORY_ISOLATION
253 compound_page_dtor
* const compound_page_dtors
[] = {
256 #ifdef CONFIG_HUGETLB_PAGE
259 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
264 int min_free_kbytes
= 1024;
265 int user_min_free_kbytes
= -1;
266 int watermark_scale_factor
= 10;
268 static unsigned long nr_kernel_pages __meminitdata
;
269 static unsigned long nr_all_pages __meminitdata
;
270 static unsigned long dma_reserve __meminitdata
;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __meminitdata
;
274 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __meminitdata
;
275 static unsigned long required_kernelcore __initdata
;
276 static unsigned long required_kernelcore_percent __initdata
;
277 static unsigned long required_movablecore __initdata
;
278 static unsigned long required_movablecore_percent __initdata
;
279 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __meminitdata
;
280 static bool mirrored_kernelcore __meminitdata
;
282 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
284 EXPORT_SYMBOL(movable_zone
);
285 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
288 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
289 int nr_online_nodes __read_mostly
= 1;
290 EXPORT_SYMBOL(nr_node_ids
);
291 EXPORT_SYMBOL(nr_online_nodes
);
294 int page_group_by_mobility_disabled __read_mostly
;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
298 * During boot we initialize deferred pages on-demand, as needed, but once
299 * page_alloc_init_late() has finished, the deferred pages are all initialized,
300 * and we can permanently disable that path.
302 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
305 * Calling kasan_free_pages() only after deferred memory initialization
306 * has completed. Poisoning pages during deferred memory init will greatly
307 * lengthen the process and cause problem in large memory systems as the
308 * deferred pages initialization is done with interrupt disabled.
310 * Assuming that there will be no reference to those newly initialized
311 * pages before they are ever allocated, this should have no effect on
312 * KASAN memory tracking as the poison will be properly inserted at page
313 * allocation time. The only corner case is when pages are allocated by
314 * on-demand allocation and then freed again before the deferred pages
315 * initialization is done, but this is not likely to happen.
317 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
319 if (!static_branch_unlikely(&deferred_pages
))
320 kasan_free_pages(page
, order
);
323 /* Returns true if the struct page for the pfn is uninitialised */
324 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
326 int nid
= early_pfn_to_nid(pfn
);
328 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
335 * Returns true when the remaining initialisation should be deferred until
336 * later in the boot cycle when it can be parallelised.
338 static bool __meminit
339 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
341 static unsigned long prev_end_pfn
, nr_initialised
;
344 * prev_end_pfn static that contains the end of previous zone
345 * No need to protect because called very early in boot before smp_init.
347 if (prev_end_pfn
!= end_pfn
) {
348 prev_end_pfn
= end_pfn
;
352 /* Always populate low zones for address-constrained allocations */
353 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
356 if ((nr_initialised
> NODE_DATA(nid
)->static_init_pgcnt
) &&
357 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
358 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
364 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
366 static inline bool early_page_uninitialised(unsigned long pfn
)
371 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
377 /* Return a pointer to the bitmap storing bits affecting a block of pages */
378 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
381 #ifdef CONFIG_SPARSEMEM
382 return __pfn_to_section(pfn
)->pageblock_flags
;
384 return page_zone(page
)->pageblock_flags
;
385 #endif /* CONFIG_SPARSEMEM */
388 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
390 #ifdef CONFIG_SPARSEMEM
391 pfn
&= (PAGES_PER_SECTION
-1);
392 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
394 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
395 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
396 #endif /* CONFIG_SPARSEMEM */
400 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
401 * @page: The page within the block of interest
402 * @pfn: The target page frame number
403 * @end_bitidx: The last bit of interest to retrieve
404 * @mask: mask of bits that the caller is interested in
406 * Return: pageblock_bits flags
408 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
410 unsigned long end_bitidx
,
413 unsigned long *bitmap
;
414 unsigned long bitidx
, word_bitidx
;
417 bitmap
= get_pageblock_bitmap(page
, pfn
);
418 bitidx
= pfn_to_bitidx(page
, pfn
);
419 word_bitidx
= bitidx
/ BITS_PER_LONG
;
420 bitidx
&= (BITS_PER_LONG
-1);
422 word
= bitmap
[word_bitidx
];
423 bitidx
+= end_bitidx
;
424 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
427 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
428 unsigned long end_bitidx
,
431 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
434 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
436 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
440 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
441 * @page: The page within the block of interest
442 * @flags: The flags to set
443 * @pfn: The target page frame number
444 * @end_bitidx: The last bit of interest
445 * @mask: mask of bits that the caller is interested in
447 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
449 unsigned long end_bitidx
,
452 unsigned long *bitmap
;
453 unsigned long bitidx
, word_bitidx
;
454 unsigned long old_word
, word
;
456 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
458 bitmap
= get_pageblock_bitmap(page
, pfn
);
459 bitidx
= pfn_to_bitidx(page
, pfn
);
460 word_bitidx
= bitidx
/ BITS_PER_LONG
;
461 bitidx
&= (BITS_PER_LONG
-1);
463 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
465 bitidx
+= end_bitidx
;
466 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
467 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
469 word
= READ_ONCE(bitmap
[word_bitidx
]);
471 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
472 if (word
== old_word
)
478 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
480 if (unlikely(page_group_by_mobility_disabled
&&
481 migratetype
< MIGRATE_PCPTYPES
))
482 migratetype
= MIGRATE_UNMOVABLE
;
484 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
485 PB_migrate
, PB_migrate_end
);
488 #ifdef CONFIG_DEBUG_VM
489 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
493 unsigned long pfn
= page_to_pfn(page
);
494 unsigned long sp
, start_pfn
;
497 seq
= zone_span_seqbegin(zone
);
498 start_pfn
= zone
->zone_start_pfn
;
499 sp
= zone
->spanned_pages
;
500 if (!zone_spans_pfn(zone
, pfn
))
502 } while (zone_span_seqretry(zone
, seq
));
505 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
506 pfn
, zone_to_nid(zone
), zone
->name
,
507 start_pfn
, start_pfn
+ sp
);
512 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
514 if (!pfn_valid_within(page_to_pfn(page
)))
516 if (zone
!= page_zone(page
))
522 * Temporary debugging check for pages not lying within a given zone.
524 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
526 if (page_outside_zone_boundaries(zone
, page
))
528 if (!page_is_consistent(zone
, page
))
534 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
540 static void bad_page(struct page
*page
, const char *reason
,
541 unsigned long bad_flags
)
543 static unsigned long resume
;
544 static unsigned long nr_shown
;
545 static unsigned long nr_unshown
;
548 * Allow a burst of 60 reports, then keep quiet for that minute;
549 * or allow a steady drip of one report per second.
551 if (nr_shown
== 60) {
552 if (time_before(jiffies
, resume
)) {
558 "BUG: Bad page state: %lu messages suppressed\n",
565 resume
= jiffies
+ 60 * HZ
;
567 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
568 current
->comm
, page_to_pfn(page
));
569 __dump_page(page
, reason
);
570 bad_flags
&= page
->flags
;
572 pr_alert("bad because of flags: %#lx(%pGp)\n",
573 bad_flags
, &bad_flags
);
574 dump_page_owner(page
);
579 /* Leave bad fields for debug, except PageBuddy could make trouble */
580 page_mapcount_reset(page
); /* remove PageBuddy */
581 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
585 * Higher-order pages are called "compound pages". They are structured thusly:
587 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
589 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
590 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
592 * The first tail page's ->compound_dtor holds the offset in array of compound
593 * page destructors. See compound_page_dtors.
595 * The first tail page's ->compound_order holds the order of allocation.
596 * This usage means that zero-order pages may not be compound.
599 void free_compound_page(struct page
*page
)
601 __free_pages_ok(page
, compound_order(page
));
604 void prep_compound_page(struct page
*page
, unsigned int order
)
607 int nr_pages
= 1 << order
;
609 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
610 set_compound_order(page
, order
);
612 for (i
= 1; i
< nr_pages
; i
++) {
613 struct page
*p
= page
+ i
;
614 set_page_count(p
, 0);
615 p
->mapping
= TAIL_MAPPING
;
616 set_compound_head(p
, page
);
618 atomic_set(compound_mapcount_ptr(page
), -1);
621 #ifdef CONFIG_DEBUG_PAGEALLOC
622 unsigned int _debug_guardpage_minorder
;
623 bool _debug_pagealloc_enabled __read_mostly
624 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
625 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
626 bool _debug_guardpage_enabled __read_mostly
;
628 static int __init
early_debug_pagealloc(char *buf
)
632 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
634 early_param("debug_pagealloc", early_debug_pagealloc
);
636 static bool need_debug_guardpage(void)
638 /* If we don't use debug_pagealloc, we don't need guard page */
639 if (!debug_pagealloc_enabled())
642 if (!debug_guardpage_minorder())
648 static void init_debug_guardpage(void)
650 if (!debug_pagealloc_enabled())
653 if (!debug_guardpage_minorder())
656 _debug_guardpage_enabled
= true;
659 struct page_ext_operations debug_guardpage_ops
= {
660 .need
= need_debug_guardpage
,
661 .init
= init_debug_guardpage
,
664 static int __init
debug_guardpage_minorder_setup(char *buf
)
668 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
669 pr_err("Bad debug_guardpage_minorder value\n");
672 _debug_guardpage_minorder
= res
;
673 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
676 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
678 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
679 unsigned int order
, int migratetype
)
681 struct page_ext
*page_ext
;
683 if (!debug_guardpage_enabled())
686 if (order
>= debug_guardpage_minorder())
689 page_ext
= lookup_page_ext(page
);
690 if (unlikely(!page_ext
))
693 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
695 INIT_LIST_HEAD(&page
->lru
);
696 set_page_private(page
, order
);
697 /* Guard pages are not available for any usage */
698 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
703 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
704 unsigned int order
, int migratetype
)
706 struct page_ext
*page_ext
;
708 if (!debug_guardpage_enabled())
711 page_ext
= lookup_page_ext(page
);
712 if (unlikely(!page_ext
))
715 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
717 set_page_private(page
, 0);
718 if (!is_migrate_isolate(migratetype
))
719 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
722 struct page_ext_operations debug_guardpage_ops
;
723 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
724 unsigned int order
, int migratetype
) { return false; }
725 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
726 unsigned int order
, int migratetype
) {}
729 static inline void set_page_order(struct page
*page
, unsigned int order
)
731 set_page_private(page
, order
);
732 __SetPageBuddy(page
);
735 static inline void rmv_page_order(struct page
*page
)
737 __ClearPageBuddy(page
);
738 set_page_private(page
, 0);
742 * This function checks whether a page is free && is the buddy
743 * we can coalesce a page and its buddy if
744 * (a) the buddy is not in a hole (check before calling!) &&
745 * (b) the buddy is in the buddy system &&
746 * (c) a page and its buddy have the same order &&
747 * (d) a page and its buddy are in the same zone.
749 * For recording whether a page is in the buddy system, we set PageBuddy.
750 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
752 * For recording page's order, we use page_private(page).
754 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
757 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
758 if (page_zone_id(page
) != page_zone_id(buddy
))
761 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
766 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
768 * zone check is done late to avoid uselessly
769 * calculating zone/node ids for pages that could
772 if (page_zone_id(page
) != page_zone_id(buddy
))
775 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
783 * Freeing function for a buddy system allocator.
785 * The concept of a buddy system is to maintain direct-mapped table
786 * (containing bit values) for memory blocks of various "orders".
787 * The bottom level table contains the map for the smallest allocatable
788 * units of memory (here, pages), and each level above it describes
789 * pairs of units from the levels below, hence, "buddies".
790 * At a high level, all that happens here is marking the table entry
791 * at the bottom level available, and propagating the changes upward
792 * as necessary, plus some accounting needed to play nicely with other
793 * parts of the VM system.
794 * At each level, we keep a list of pages, which are heads of continuous
795 * free pages of length of (1 << order) and marked with PageBuddy.
796 * Page's order is recorded in page_private(page) field.
797 * So when we are allocating or freeing one, we can derive the state of the
798 * other. That is, if we allocate a small block, and both were
799 * free, the remainder of the region must be split into blocks.
800 * If a block is freed, and its buddy is also free, then this
801 * triggers coalescing into a block of larger size.
806 static inline void __free_one_page(struct page
*page
,
808 struct zone
*zone
, unsigned int order
,
811 unsigned long combined_pfn
;
812 unsigned long uninitialized_var(buddy_pfn
);
814 unsigned int max_order
;
816 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
818 VM_BUG_ON(!zone_is_initialized(zone
));
819 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
821 VM_BUG_ON(migratetype
== -1);
822 if (likely(!is_migrate_isolate(migratetype
)))
823 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
825 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
826 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
829 while (order
< max_order
- 1) {
830 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
831 buddy
= page
+ (buddy_pfn
- pfn
);
833 if (!pfn_valid_within(buddy_pfn
))
835 if (!page_is_buddy(page
, buddy
, order
))
838 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
839 * merge with it and move up one order.
841 if (page_is_guard(buddy
)) {
842 clear_page_guard(zone
, buddy
, order
, migratetype
);
844 list_del(&buddy
->lru
);
845 zone
->free_area
[order
].nr_free
--;
846 rmv_page_order(buddy
);
848 combined_pfn
= buddy_pfn
& pfn
;
849 page
= page
+ (combined_pfn
- pfn
);
853 if (max_order
< MAX_ORDER
) {
854 /* If we are here, it means order is >= pageblock_order.
855 * We want to prevent merge between freepages on isolate
856 * pageblock and normal pageblock. Without this, pageblock
857 * isolation could cause incorrect freepage or CMA accounting.
859 * We don't want to hit this code for the more frequent
862 if (unlikely(has_isolate_pageblock(zone
))) {
865 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
866 buddy
= page
+ (buddy_pfn
- pfn
);
867 buddy_mt
= get_pageblock_migratetype(buddy
);
869 if (migratetype
!= buddy_mt
870 && (is_migrate_isolate(migratetype
) ||
871 is_migrate_isolate(buddy_mt
)))
875 goto continue_merging
;
879 set_page_order(page
, order
);
882 * If this is not the largest possible page, check if the buddy
883 * of the next-highest order is free. If it is, it's possible
884 * that pages are being freed that will coalesce soon. In case,
885 * that is happening, add the free page to the tail of the list
886 * so it's less likely to be used soon and more likely to be merged
887 * as a higher order page
889 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
890 struct page
*higher_page
, *higher_buddy
;
891 combined_pfn
= buddy_pfn
& pfn
;
892 higher_page
= page
+ (combined_pfn
- pfn
);
893 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
894 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
895 if (pfn_valid_within(buddy_pfn
) &&
896 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
897 list_add_tail(&page
->lru
,
898 &zone
->free_area
[order
].free_list
[migratetype
]);
903 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
905 zone
->free_area
[order
].nr_free
++;
909 * A bad page could be due to a number of fields. Instead of multiple branches,
910 * try and check multiple fields with one check. The caller must do a detailed
911 * check if necessary.
913 static inline bool page_expected_state(struct page
*page
,
914 unsigned long check_flags
)
916 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
919 if (unlikely((unsigned long)page
->mapping
|
920 page_ref_count(page
) |
922 (unsigned long)page
->mem_cgroup
|
924 (page
->flags
& check_flags
)))
930 static void free_pages_check_bad(struct page
*page
)
932 const char *bad_reason
;
933 unsigned long bad_flags
;
938 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
939 bad_reason
= "nonzero mapcount";
940 if (unlikely(page
->mapping
!= NULL
))
941 bad_reason
= "non-NULL mapping";
942 if (unlikely(page_ref_count(page
) != 0))
943 bad_reason
= "nonzero _refcount";
944 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
945 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
946 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
949 if (unlikely(page
->mem_cgroup
))
950 bad_reason
= "page still charged to cgroup";
952 bad_page(page
, bad_reason
, bad_flags
);
955 static inline int free_pages_check(struct page
*page
)
957 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
960 /* Something has gone sideways, find it */
961 free_pages_check_bad(page
);
965 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
970 * We rely page->lru.next never has bit 0 set, unless the page
971 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
973 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
975 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
979 switch (page
- head_page
) {
981 /* the first tail page: ->mapping may be compound_mapcount() */
982 if (unlikely(compound_mapcount(page
))) {
983 bad_page(page
, "nonzero compound_mapcount", 0);
989 * the second tail page: ->mapping is
990 * deferred_list.next -- ignore value.
994 if (page
->mapping
!= TAIL_MAPPING
) {
995 bad_page(page
, "corrupted mapping in tail page", 0);
1000 if (unlikely(!PageTail(page
))) {
1001 bad_page(page
, "PageTail not set", 0);
1004 if (unlikely(compound_head(page
) != head_page
)) {
1005 bad_page(page
, "compound_head not consistent", 0);
1010 page
->mapping
= NULL
;
1011 clear_compound_head(page
);
1015 static __always_inline
bool free_pages_prepare(struct page
*page
,
1016 unsigned int order
, bool check_free
)
1020 VM_BUG_ON_PAGE(PageTail(page
), page
);
1022 trace_mm_page_free(page
, order
);
1025 * Check tail pages before head page information is cleared to
1026 * avoid checking PageCompound for order-0 pages.
1028 if (unlikely(order
)) {
1029 bool compound
= PageCompound(page
);
1032 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1035 ClearPageDoubleMap(page
);
1036 for (i
= 1; i
< (1 << order
); i
++) {
1038 bad
+= free_tail_pages_check(page
, page
+ i
);
1039 if (unlikely(free_pages_check(page
+ i
))) {
1043 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1046 if (PageMappingFlags(page
))
1047 page
->mapping
= NULL
;
1048 if (memcg_kmem_enabled() && PageKmemcg(page
))
1049 memcg_kmem_uncharge(page
, order
);
1051 bad
+= free_pages_check(page
);
1055 page_cpupid_reset_last(page
);
1056 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1057 reset_page_owner(page
, order
);
1059 if (!PageHighMem(page
)) {
1060 debug_check_no_locks_freed(page_address(page
),
1061 PAGE_SIZE
<< order
);
1062 debug_check_no_obj_freed(page_address(page
),
1063 PAGE_SIZE
<< order
);
1065 arch_free_page(page
, order
);
1066 kernel_poison_pages(page
, 1 << order
, 0);
1067 kernel_map_pages(page
, 1 << order
, 0);
1068 kasan_free_nondeferred_pages(page
, order
);
1073 #ifdef CONFIG_DEBUG_VM
1074 static inline bool free_pcp_prepare(struct page
*page
)
1076 return free_pages_prepare(page
, 0, true);
1079 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1084 static bool free_pcp_prepare(struct page
*page
)
1086 return free_pages_prepare(page
, 0, false);
1089 static bool bulkfree_pcp_prepare(struct page
*page
)
1091 return free_pages_check(page
);
1093 #endif /* CONFIG_DEBUG_VM */
1095 static inline void prefetch_buddy(struct page
*page
)
1097 unsigned long pfn
= page_to_pfn(page
);
1098 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1099 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1105 * Frees a number of pages from the PCP lists
1106 * Assumes all pages on list are in same zone, and of same order.
1107 * count is the number of pages to free.
1109 * If the zone was previously in an "all pages pinned" state then look to
1110 * see if this freeing clears that state.
1112 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1113 * pinned" detection logic.
1115 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1116 struct per_cpu_pages
*pcp
)
1118 int migratetype
= 0;
1120 int prefetch_nr
= 0;
1121 bool isolated_pageblocks
;
1122 struct page
*page
, *tmp
;
1126 struct list_head
*list
;
1129 * Remove pages from lists in a round-robin fashion. A
1130 * batch_free count is maintained that is incremented when an
1131 * empty list is encountered. This is so more pages are freed
1132 * off fuller lists instead of spinning excessively around empty
1137 if (++migratetype
== MIGRATE_PCPTYPES
)
1139 list
= &pcp
->lists
[migratetype
];
1140 } while (list_empty(list
));
1142 /* This is the only non-empty list. Free them all. */
1143 if (batch_free
== MIGRATE_PCPTYPES
)
1147 page
= list_last_entry(list
, struct page
, lru
);
1148 /* must delete to avoid corrupting pcp list */
1149 list_del(&page
->lru
);
1152 if (bulkfree_pcp_prepare(page
))
1155 list_add_tail(&page
->lru
, &head
);
1158 * We are going to put the page back to the global
1159 * pool, prefetch its buddy to speed up later access
1160 * under zone->lock. It is believed the overhead of
1161 * an additional test and calculating buddy_pfn here
1162 * can be offset by reduced memory latency later. To
1163 * avoid excessive prefetching due to large count, only
1164 * prefetch buddy for the first pcp->batch nr of pages.
1166 if (prefetch_nr
++ < pcp
->batch
)
1167 prefetch_buddy(page
);
1168 } while (--count
&& --batch_free
&& !list_empty(list
));
1171 spin_lock(&zone
->lock
);
1172 isolated_pageblocks
= has_isolate_pageblock(zone
);
1175 * Use safe version since after __free_one_page(),
1176 * page->lru.next will not point to original list.
1178 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1179 int mt
= get_pcppage_migratetype(page
);
1180 /* MIGRATE_ISOLATE page should not go to pcplists */
1181 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1182 /* Pageblock could have been isolated meanwhile */
1183 if (unlikely(isolated_pageblocks
))
1184 mt
= get_pageblock_migratetype(page
);
1186 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1187 trace_mm_page_pcpu_drain(page
, 0, mt
);
1189 spin_unlock(&zone
->lock
);
1192 static void free_one_page(struct zone
*zone
,
1193 struct page
*page
, unsigned long pfn
,
1197 spin_lock(&zone
->lock
);
1198 if (unlikely(has_isolate_pageblock(zone
) ||
1199 is_migrate_isolate(migratetype
))) {
1200 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1202 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1203 spin_unlock(&zone
->lock
);
1206 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1207 unsigned long zone
, int nid
)
1209 mm_zero_struct_page(page
);
1210 set_page_links(page
, zone
, nid
, pfn
);
1211 init_page_count(page
);
1212 page_mapcount_reset(page
);
1213 page_cpupid_reset_last(page
);
1215 INIT_LIST_HEAD(&page
->lru
);
1216 #ifdef WANT_PAGE_VIRTUAL
1217 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1218 if (!is_highmem_idx(zone
))
1219 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1223 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1224 static void __meminit
init_reserved_page(unsigned long pfn
)
1229 if (!early_page_uninitialised(pfn
))
1232 nid
= early_pfn_to_nid(pfn
);
1233 pgdat
= NODE_DATA(nid
);
1235 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1236 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1238 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1241 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1244 static inline void init_reserved_page(unsigned long pfn
)
1247 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1250 * Initialised pages do not have PageReserved set. This function is
1251 * called for each range allocated by the bootmem allocator and
1252 * marks the pages PageReserved. The remaining valid pages are later
1253 * sent to the buddy page allocator.
1255 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1257 unsigned long start_pfn
= PFN_DOWN(start
);
1258 unsigned long end_pfn
= PFN_UP(end
);
1260 for (; start_pfn
< end_pfn
; start_pfn
++) {
1261 if (pfn_valid(start_pfn
)) {
1262 struct page
*page
= pfn_to_page(start_pfn
);
1264 init_reserved_page(start_pfn
);
1266 /* Avoid false-positive PageTail() */
1267 INIT_LIST_HEAD(&page
->lru
);
1270 * no need for atomic set_bit because the struct
1271 * page is not visible yet so nobody should
1274 __SetPageReserved(page
);
1279 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1281 unsigned long flags
;
1283 unsigned long pfn
= page_to_pfn(page
);
1285 if (!free_pages_prepare(page
, order
, true))
1288 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1289 local_irq_save(flags
);
1290 __count_vm_events(PGFREE
, 1 << order
);
1291 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1292 local_irq_restore(flags
);
1295 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1297 unsigned int nr_pages
= 1 << order
;
1298 struct page
*p
= page
;
1302 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1304 __ClearPageReserved(p
);
1305 set_page_count(p
, 0);
1307 __ClearPageReserved(p
);
1308 set_page_count(p
, 0);
1310 page_zone(page
)->managed_pages
+= nr_pages
;
1311 set_page_refcounted(page
);
1312 __free_pages(page
, order
);
1315 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1316 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1318 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1320 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1322 static DEFINE_SPINLOCK(early_pfn_lock
);
1325 spin_lock(&early_pfn_lock
);
1326 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1328 nid
= first_online_node
;
1329 spin_unlock(&early_pfn_lock
);
1335 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1336 static inline bool __meminit __maybe_unused
1337 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1338 struct mminit_pfnnid_cache
*state
)
1342 nid
= __early_pfn_to_nid(pfn
, state
);
1343 if (nid
>= 0 && nid
!= node
)
1348 /* Only safe to use early in boot when initialisation is single-threaded */
1349 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1351 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1356 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1360 static inline bool __meminit __maybe_unused
1361 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1362 struct mminit_pfnnid_cache
*state
)
1369 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1372 if (early_page_uninitialised(pfn
))
1374 return __free_pages_boot_core(page
, order
);
1378 * Check that the whole (or subset of) a pageblock given by the interval of
1379 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1380 * with the migration of free compaction scanner. The scanners then need to
1381 * use only pfn_valid_within() check for arches that allow holes within
1384 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1386 * It's possible on some configurations to have a setup like node0 node1 node0
1387 * i.e. it's possible that all pages within a zones range of pages do not
1388 * belong to a single zone. We assume that a border between node0 and node1
1389 * can occur within a single pageblock, but not a node0 node1 node0
1390 * interleaving within a single pageblock. It is therefore sufficient to check
1391 * the first and last page of a pageblock and avoid checking each individual
1392 * page in a pageblock.
1394 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1395 unsigned long end_pfn
, struct zone
*zone
)
1397 struct page
*start_page
;
1398 struct page
*end_page
;
1400 /* end_pfn is one past the range we are checking */
1403 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1406 start_page
= pfn_to_online_page(start_pfn
);
1410 if (page_zone(start_page
) != zone
)
1413 end_page
= pfn_to_page(end_pfn
);
1415 /* This gives a shorter code than deriving page_zone(end_page) */
1416 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1422 void set_zone_contiguous(struct zone
*zone
)
1424 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1425 unsigned long block_end_pfn
;
1427 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1428 for (; block_start_pfn
< zone_end_pfn(zone
);
1429 block_start_pfn
= block_end_pfn
,
1430 block_end_pfn
+= pageblock_nr_pages
) {
1432 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1434 if (!__pageblock_pfn_to_page(block_start_pfn
,
1435 block_end_pfn
, zone
))
1439 /* We confirm that there is no hole */
1440 zone
->contiguous
= true;
1443 void clear_zone_contiguous(struct zone
*zone
)
1445 zone
->contiguous
= false;
1448 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1449 static void __init
deferred_free_range(unsigned long pfn
,
1450 unsigned long nr_pages
)
1458 page
= pfn_to_page(pfn
);
1460 /* Free a large naturally-aligned chunk if possible */
1461 if (nr_pages
== pageblock_nr_pages
&&
1462 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1463 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1464 __free_pages_boot_core(page
, pageblock_order
);
1468 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1469 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1470 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1471 __free_pages_boot_core(page
, 0);
1475 /* Completion tracking for deferred_init_memmap() threads */
1476 static atomic_t pgdat_init_n_undone __initdata
;
1477 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1479 static inline void __init
pgdat_init_report_one_done(void)
1481 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1482 complete(&pgdat_init_all_done_comp
);
1486 * Returns true if page needs to be initialized or freed to buddy allocator.
1488 * First we check if pfn is valid on architectures where it is possible to have
1489 * holes within pageblock_nr_pages. On systems where it is not possible, this
1490 * function is optimized out.
1492 * Then, we check if a current large page is valid by only checking the validity
1495 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1496 * within a node: a pfn is between start and end of a node, but does not belong
1497 * to this memory node.
1499 static inline bool __init
1500 deferred_pfn_valid(int nid
, unsigned long pfn
,
1501 struct mminit_pfnnid_cache
*nid_init_state
)
1503 if (!pfn_valid_within(pfn
))
1505 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1507 if (!meminit_pfn_in_nid(pfn
, nid
, nid_init_state
))
1513 * Free pages to buddy allocator. Try to free aligned pages in
1514 * pageblock_nr_pages sizes.
1516 static void __init
deferred_free_pages(int nid
, int zid
, unsigned long pfn
,
1517 unsigned long end_pfn
)
1519 struct mminit_pfnnid_cache nid_init_state
= { };
1520 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1521 unsigned long nr_free
= 0;
1523 for (; pfn
< end_pfn
; pfn
++) {
1524 if (!deferred_pfn_valid(nid
, pfn
, &nid_init_state
)) {
1525 deferred_free_range(pfn
- nr_free
, nr_free
);
1527 } else if (!(pfn
& nr_pgmask
)) {
1528 deferred_free_range(pfn
- nr_free
, nr_free
);
1530 touch_nmi_watchdog();
1535 /* Free the last block of pages to allocator */
1536 deferred_free_range(pfn
- nr_free
, nr_free
);
1540 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1541 * by performing it only once every pageblock_nr_pages.
1542 * Return number of pages initialized.
1544 static unsigned long __init
deferred_init_pages(int nid
, int zid
,
1546 unsigned long end_pfn
)
1548 struct mminit_pfnnid_cache nid_init_state
= { };
1549 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1550 unsigned long nr_pages
= 0;
1551 struct page
*page
= NULL
;
1553 for (; pfn
< end_pfn
; pfn
++) {
1554 if (!deferred_pfn_valid(nid
, pfn
, &nid_init_state
)) {
1557 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1558 page
= pfn_to_page(pfn
);
1559 touch_nmi_watchdog();
1563 __init_single_page(page
, pfn
, zid
, nid
);
1569 /* Initialise remaining memory on a node */
1570 static int __init
deferred_init_memmap(void *data
)
1572 pg_data_t
*pgdat
= data
;
1573 int nid
= pgdat
->node_id
;
1574 unsigned long start
= jiffies
;
1575 unsigned long nr_pages
= 0;
1576 unsigned long spfn
, epfn
, first_init_pfn
, flags
;
1577 phys_addr_t spa
, epa
;
1580 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1583 /* Bind memory initialisation thread to a local node if possible */
1584 if (!cpumask_empty(cpumask
))
1585 set_cpus_allowed_ptr(current
, cpumask
);
1587 pgdat_resize_lock(pgdat
, &flags
);
1588 first_init_pfn
= pgdat
->first_deferred_pfn
;
1589 if (first_init_pfn
== ULONG_MAX
) {
1590 pgdat_resize_unlock(pgdat
, &flags
);
1591 pgdat_init_report_one_done();
1595 /* Sanity check boundaries */
1596 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1597 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1598 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1600 /* Only the highest zone is deferred so find it */
1601 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1602 zone
= pgdat
->node_zones
+ zid
;
1603 if (first_init_pfn
< zone_end_pfn(zone
))
1606 first_init_pfn
= max(zone
->zone_start_pfn
, first_init_pfn
);
1609 * Initialize and free pages. We do it in two loops: first we initialize
1610 * struct page, than free to buddy allocator, because while we are
1611 * freeing pages we can access pages that are ahead (computing buddy
1612 * page in __free_one_page()).
1614 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1615 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1616 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1617 nr_pages
+= deferred_init_pages(nid
, zid
, spfn
, epfn
);
1619 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1620 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1621 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1622 deferred_free_pages(nid
, zid
, spfn
, epfn
);
1624 pgdat_resize_unlock(pgdat
, &flags
);
1626 /* Sanity check that the next zone really is unpopulated */
1627 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1629 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1630 jiffies_to_msecs(jiffies
- start
));
1632 pgdat_init_report_one_done();
1637 * If this zone has deferred pages, try to grow it by initializing enough
1638 * deferred pages to satisfy the allocation specified by order, rounded up to
1639 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1640 * of SECTION_SIZE bytes by initializing struct pages in increments of
1641 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1643 * Return true when zone was grown, otherwise return false. We return true even
1644 * when we grow less than requested, to let the caller decide if there are
1645 * enough pages to satisfy the allocation.
1647 * Note: We use noinline because this function is needed only during boot, and
1648 * it is called from a __ref function _deferred_grow_zone. This way we are
1649 * making sure that it is not inlined into permanent text section.
1651 static noinline
bool __init
1652 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1654 int zid
= zone_idx(zone
);
1655 int nid
= zone_to_nid(zone
);
1656 pg_data_t
*pgdat
= NODE_DATA(nid
);
1657 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1658 unsigned long nr_pages
= 0;
1659 unsigned long first_init_pfn
, spfn
, epfn
, t
, flags
;
1660 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1661 phys_addr_t spa
, epa
;
1664 /* Only the last zone may have deferred pages */
1665 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1668 pgdat_resize_lock(pgdat
, &flags
);
1671 * If deferred pages have been initialized while we were waiting for
1672 * the lock, return true, as the zone was grown. The caller will retry
1673 * this zone. We won't return to this function since the caller also
1674 * has this static branch.
1676 if (!static_branch_unlikely(&deferred_pages
)) {
1677 pgdat_resize_unlock(pgdat
, &flags
);
1682 * If someone grew this zone while we were waiting for spinlock, return
1683 * true, as there might be enough pages already.
1685 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1686 pgdat_resize_unlock(pgdat
, &flags
);
1690 first_init_pfn
= max(zone
->zone_start_pfn
, first_deferred_pfn
);
1692 if (first_init_pfn
>= pgdat_end_pfn(pgdat
)) {
1693 pgdat_resize_unlock(pgdat
, &flags
);
1697 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1698 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1699 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1701 while (spfn
< epfn
&& nr_pages
< nr_pages_needed
) {
1702 t
= ALIGN(spfn
+ PAGES_PER_SECTION
, PAGES_PER_SECTION
);
1703 first_deferred_pfn
= min(t
, epfn
);
1704 nr_pages
+= deferred_init_pages(nid
, zid
, spfn
,
1705 first_deferred_pfn
);
1706 spfn
= first_deferred_pfn
;
1709 if (nr_pages
>= nr_pages_needed
)
1713 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1714 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1715 epfn
= min_t(unsigned long, first_deferred_pfn
, PFN_DOWN(epa
));
1716 deferred_free_pages(nid
, zid
, spfn
, epfn
);
1718 if (first_deferred_pfn
== epfn
)
1721 pgdat
->first_deferred_pfn
= first_deferred_pfn
;
1722 pgdat_resize_unlock(pgdat
, &flags
);
1724 return nr_pages
> 0;
1728 * deferred_grow_zone() is __init, but it is called from
1729 * get_page_from_freelist() during early boot until deferred_pages permanently
1730 * disables this call. This is why we have refdata wrapper to avoid warning,
1731 * and to ensure that the function body gets unloaded.
1734 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1736 return deferred_grow_zone(zone
, order
);
1739 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1741 void __init
page_alloc_init_late(void)
1745 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1748 /* There will be num_node_state(N_MEMORY) threads */
1749 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1750 for_each_node_state(nid
, N_MEMORY
) {
1751 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1754 /* Block until all are initialised */
1755 wait_for_completion(&pgdat_init_all_done_comp
);
1758 * We initialized the rest of the deferred pages. Permanently disable
1759 * on-demand struct page initialization.
1761 static_branch_disable(&deferred_pages
);
1763 /* Reinit limits that are based on free pages after the kernel is up */
1764 files_maxfiles_init();
1766 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1767 /* Discard memblock private memory */
1771 for_each_populated_zone(zone
)
1772 set_zone_contiguous(zone
);
1776 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1777 void __init
init_cma_reserved_pageblock(struct page
*page
)
1779 unsigned i
= pageblock_nr_pages
;
1780 struct page
*p
= page
;
1783 __ClearPageReserved(p
);
1784 set_page_count(p
, 0);
1787 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1789 if (pageblock_order
>= MAX_ORDER
) {
1790 i
= pageblock_nr_pages
;
1793 set_page_refcounted(p
);
1794 __free_pages(p
, MAX_ORDER
- 1);
1795 p
+= MAX_ORDER_NR_PAGES
;
1796 } while (i
-= MAX_ORDER_NR_PAGES
);
1798 set_page_refcounted(page
);
1799 __free_pages(page
, pageblock_order
);
1802 adjust_managed_page_count(page
, pageblock_nr_pages
);
1807 * The order of subdivision here is critical for the IO subsystem.
1808 * Please do not alter this order without good reasons and regression
1809 * testing. Specifically, as large blocks of memory are subdivided,
1810 * the order in which smaller blocks are delivered depends on the order
1811 * they're subdivided in this function. This is the primary factor
1812 * influencing the order in which pages are delivered to the IO
1813 * subsystem according to empirical testing, and this is also justified
1814 * by considering the behavior of a buddy system containing a single
1815 * large block of memory acted on by a series of small allocations.
1816 * This behavior is a critical factor in sglist merging's success.
1820 static inline void expand(struct zone
*zone
, struct page
*page
,
1821 int low
, int high
, struct free_area
*area
,
1824 unsigned long size
= 1 << high
;
1826 while (high
> low
) {
1830 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1833 * Mark as guard pages (or page), that will allow to
1834 * merge back to allocator when buddy will be freed.
1835 * Corresponding page table entries will not be touched,
1836 * pages will stay not present in virtual address space
1838 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1841 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1843 set_page_order(&page
[size
], high
);
1847 static void check_new_page_bad(struct page
*page
)
1849 const char *bad_reason
= NULL
;
1850 unsigned long bad_flags
= 0;
1852 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1853 bad_reason
= "nonzero mapcount";
1854 if (unlikely(page
->mapping
!= NULL
))
1855 bad_reason
= "non-NULL mapping";
1856 if (unlikely(page_ref_count(page
) != 0))
1857 bad_reason
= "nonzero _count";
1858 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1859 bad_reason
= "HWPoisoned (hardware-corrupted)";
1860 bad_flags
= __PG_HWPOISON
;
1861 /* Don't complain about hwpoisoned pages */
1862 page_mapcount_reset(page
); /* remove PageBuddy */
1865 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1866 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1867 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1870 if (unlikely(page
->mem_cgroup
))
1871 bad_reason
= "page still charged to cgroup";
1873 bad_page(page
, bad_reason
, bad_flags
);
1877 * This page is about to be returned from the page allocator
1879 static inline int check_new_page(struct page
*page
)
1881 if (likely(page_expected_state(page
,
1882 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1885 check_new_page_bad(page
);
1889 static inline bool free_pages_prezeroed(void)
1891 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1892 page_poisoning_enabled();
1895 #ifdef CONFIG_DEBUG_VM
1896 static bool check_pcp_refill(struct page
*page
)
1901 static bool check_new_pcp(struct page
*page
)
1903 return check_new_page(page
);
1906 static bool check_pcp_refill(struct page
*page
)
1908 return check_new_page(page
);
1910 static bool check_new_pcp(struct page
*page
)
1914 #endif /* CONFIG_DEBUG_VM */
1916 static bool check_new_pages(struct page
*page
, unsigned int order
)
1919 for (i
= 0; i
< (1 << order
); i
++) {
1920 struct page
*p
= page
+ i
;
1922 if (unlikely(check_new_page(p
)))
1929 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1932 set_page_private(page
, 0);
1933 set_page_refcounted(page
);
1935 arch_alloc_page(page
, order
);
1936 kernel_map_pages(page
, 1 << order
, 1);
1937 kernel_poison_pages(page
, 1 << order
, 1);
1938 kasan_alloc_pages(page
, order
);
1939 set_page_owner(page
, order
, gfp_flags
);
1942 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1943 unsigned int alloc_flags
)
1947 post_alloc_hook(page
, order
, gfp_flags
);
1949 if (!free_pages_prezeroed() && (gfp_flags
& __GFP_ZERO
))
1950 for (i
= 0; i
< (1 << order
); i
++)
1951 clear_highpage(page
+ i
);
1953 if (order
&& (gfp_flags
& __GFP_COMP
))
1954 prep_compound_page(page
, order
);
1957 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1958 * allocate the page. The expectation is that the caller is taking
1959 * steps that will free more memory. The caller should avoid the page
1960 * being used for !PFMEMALLOC purposes.
1962 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1963 set_page_pfmemalloc(page
);
1965 clear_page_pfmemalloc(page
);
1969 * Go through the free lists for the given migratetype and remove
1970 * the smallest available page from the freelists
1972 static __always_inline
1973 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1976 unsigned int current_order
;
1977 struct free_area
*area
;
1980 /* Find a page of the appropriate size in the preferred list */
1981 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1982 area
= &(zone
->free_area
[current_order
]);
1983 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1987 list_del(&page
->lru
);
1988 rmv_page_order(page
);
1990 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1991 set_pcppage_migratetype(page
, migratetype
);
2000 * This array describes the order lists are fallen back to when
2001 * the free lists for the desirable migrate type are depleted
2003 static int fallbacks
[MIGRATE_TYPES
][4] = {
2004 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2005 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2006 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2008 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2010 #ifdef CONFIG_MEMORY_ISOLATION
2011 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2016 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2019 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2022 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2023 unsigned int order
) { return NULL
; }
2027 * Move the free pages in a range to the free lists of the requested type.
2028 * Note that start_page and end_pages are not aligned on a pageblock
2029 * boundary. If alignment is required, use move_freepages_block()
2031 static int move_freepages(struct zone
*zone
,
2032 struct page
*start_page
, struct page
*end_page
,
2033 int migratetype
, int *num_movable
)
2037 int pages_moved
= 0;
2039 #ifndef CONFIG_HOLES_IN_ZONE
2041 * page_zone is not safe to call in this context when
2042 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2043 * anyway as we check zone boundaries in move_freepages_block().
2044 * Remove at a later date when no bug reports exist related to
2045 * grouping pages by mobility
2047 VM_BUG_ON(pfn_valid(page_to_pfn(start_page
)) &&
2048 pfn_valid(page_to_pfn(end_page
)) &&
2049 page_zone(start_page
) != page_zone(end_page
));
2051 for (page
= start_page
; page
<= end_page
;) {
2052 if (!pfn_valid_within(page_to_pfn(page
))) {
2057 /* Make sure we are not inadvertently changing nodes */
2058 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2060 if (!PageBuddy(page
)) {
2062 * We assume that pages that could be isolated for
2063 * migration are movable. But we don't actually try
2064 * isolating, as that would be expensive.
2067 (PageLRU(page
) || __PageMovable(page
)))
2074 order
= page_order(page
);
2075 list_move(&page
->lru
,
2076 &zone
->free_area
[order
].free_list
[migratetype
]);
2078 pages_moved
+= 1 << order
;
2084 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2085 int migratetype
, int *num_movable
)
2087 unsigned long start_pfn
, end_pfn
;
2088 struct page
*start_page
, *end_page
;
2093 start_pfn
= page_to_pfn(page
);
2094 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2095 start_page
= pfn_to_page(start_pfn
);
2096 end_page
= start_page
+ pageblock_nr_pages
- 1;
2097 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2099 /* Do not cross zone boundaries */
2100 if (!zone_spans_pfn(zone
, start_pfn
))
2102 if (!zone_spans_pfn(zone
, end_pfn
))
2105 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2109 static void change_pageblock_range(struct page
*pageblock_page
,
2110 int start_order
, int migratetype
)
2112 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2114 while (nr_pageblocks
--) {
2115 set_pageblock_migratetype(pageblock_page
, migratetype
);
2116 pageblock_page
+= pageblock_nr_pages
;
2121 * When we are falling back to another migratetype during allocation, try to
2122 * steal extra free pages from the same pageblocks to satisfy further
2123 * allocations, instead of polluting multiple pageblocks.
2125 * If we are stealing a relatively large buddy page, it is likely there will
2126 * be more free pages in the pageblock, so try to steal them all. For
2127 * reclaimable and unmovable allocations, we steal regardless of page size,
2128 * as fragmentation caused by those allocations polluting movable pageblocks
2129 * is worse than movable allocations stealing from unmovable and reclaimable
2132 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2135 * Leaving this order check is intended, although there is
2136 * relaxed order check in next check. The reason is that
2137 * we can actually steal whole pageblock if this condition met,
2138 * but, below check doesn't guarantee it and that is just heuristic
2139 * so could be changed anytime.
2141 if (order
>= pageblock_order
)
2144 if (order
>= pageblock_order
/ 2 ||
2145 start_mt
== MIGRATE_RECLAIMABLE
||
2146 start_mt
== MIGRATE_UNMOVABLE
||
2147 page_group_by_mobility_disabled
)
2154 * This function implements actual steal behaviour. If order is large enough,
2155 * we can steal whole pageblock. If not, we first move freepages in this
2156 * pageblock to our migratetype and determine how many already-allocated pages
2157 * are there in the pageblock with a compatible migratetype. If at least half
2158 * of pages are free or compatible, we can change migratetype of the pageblock
2159 * itself, so pages freed in the future will be put on the correct free list.
2161 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2162 int start_type
, bool whole_block
)
2164 unsigned int current_order
= page_order(page
);
2165 struct free_area
*area
;
2166 int free_pages
, movable_pages
, alike_pages
;
2169 old_block_type
= get_pageblock_migratetype(page
);
2172 * This can happen due to races and we want to prevent broken
2173 * highatomic accounting.
2175 if (is_migrate_highatomic(old_block_type
))
2178 /* Take ownership for orders >= pageblock_order */
2179 if (current_order
>= pageblock_order
) {
2180 change_pageblock_range(page
, current_order
, start_type
);
2184 /* We are not allowed to try stealing from the whole block */
2188 free_pages
= move_freepages_block(zone
, page
, start_type
,
2191 * Determine how many pages are compatible with our allocation.
2192 * For movable allocation, it's the number of movable pages which
2193 * we just obtained. For other types it's a bit more tricky.
2195 if (start_type
== MIGRATE_MOVABLE
) {
2196 alike_pages
= movable_pages
;
2199 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2200 * to MOVABLE pageblock, consider all non-movable pages as
2201 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2202 * vice versa, be conservative since we can't distinguish the
2203 * exact migratetype of non-movable pages.
2205 if (old_block_type
== MIGRATE_MOVABLE
)
2206 alike_pages
= pageblock_nr_pages
2207 - (free_pages
+ movable_pages
);
2212 /* moving whole block can fail due to zone boundary conditions */
2217 * If a sufficient number of pages in the block are either free or of
2218 * comparable migratability as our allocation, claim the whole block.
2220 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2221 page_group_by_mobility_disabled
)
2222 set_pageblock_migratetype(page
, start_type
);
2227 area
= &zone
->free_area
[current_order
];
2228 list_move(&page
->lru
, &area
->free_list
[start_type
]);
2232 * Check whether there is a suitable fallback freepage with requested order.
2233 * If only_stealable is true, this function returns fallback_mt only if
2234 * we can steal other freepages all together. This would help to reduce
2235 * fragmentation due to mixed migratetype pages in one pageblock.
2237 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2238 int migratetype
, bool only_stealable
, bool *can_steal
)
2243 if (area
->nr_free
== 0)
2248 fallback_mt
= fallbacks
[migratetype
][i
];
2249 if (fallback_mt
== MIGRATE_TYPES
)
2252 if (list_empty(&area
->free_list
[fallback_mt
]))
2255 if (can_steal_fallback(order
, migratetype
))
2258 if (!only_stealable
)
2269 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2270 * there are no empty page blocks that contain a page with a suitable order
2272 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2273 unsigned int alloc_order
)
2276 unsigned long max_managed
, flags
;
2279 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2280 * Check is race-prone but harmless.
2282 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2283 if (zone
->nr_reserved_highatomic
>= max_managed
)
2286 spin_lock_irqsave(&zone
->lock
, flags
);
2288 /* Recheck the nr_reserved_highatomic limit under the lock */
2289 if (zone
->nr_reserved_highatomic
>= max_managed
)
2293 mt
= get_pageblock_migratetype(page
);
2294 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2295 && !is_migrate_cma(mt
)) {
2296 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2297 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2298 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2302 spin_unlock_irqrestore(&zone
->lock
, flags
);
2306 * Used when an allocation is about to fail under memory pressure. This
2307 * potentially hurts the reliability of high-order allocations when under
2308 * intense memory pressure but failed atomic allocations should be easier
2309 * to recover from than an OOM.
2311 * If @force is true, try to unreserve a pageblock even though highatomic
2312 * pageblock is exhausted.
2314 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2317 struct zonelist
*zonelist
= ac
->zonelist
;
2318 unsigned long flags
;
2325 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2328 * Preserve at least one pageblock unless memory pressure
2331 if (!force
&& zone
->nr_reserved_highatomic
<=
2335 spin_lock_irqsave(&zone
->lock
, flags
);
2336 for (order
= 0; order
< MAX_ORDER
; order
++) {
2337 struct free_area
*area
= &(zone
->free_area
[order
]);
2339 page
= list_first_entry_or_null(
2340 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2346 * In page freeing path, migratetype change is racy so
2347 * we can counter several free pages in a pageblock
2348 * in this loop althoug we changed the pageblock type
2349 * from highatomic to ac->migratetype. So we should
2350 * adjust the count once.
2352 if (is_migrate_highatomic_page(page
)) {
2354 * It should never happen but changes to
2355 * locking could inadvertently allow a per-cpu
2356 * drain to add pages to MIGRATE_HIGHATOMIC
2357 * while unreserving so be safe and watch for
2360 zone
->nr_reserved_highatomic
-= min(
2362 zone
->nr_reserved_highatomic
);
2366 * Convert to ac->migratetype and avoid the normal
2367 * pageblock stealing heuristics. Minimally, the caller
2368 * is doing the work and needs the pages. More
2369 * importantly, if the block was always converted to
2370 * MIGRATE_UNMOVABLE or another type then the number
2371 * of pageblocks that cannot be completely freed
2374 set_pageblock_migratetype(page
, ac
->migratetype
);
2375 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2378 spin_unlock_irqrestore(&zone
->lock
, flags
);
2382 spin_unlock_irqrestore(&zone
->lock
, flags
);
2389 * Try finding a free buddy page on the fallback list and put it on the free
2390 * list of requested migratetype, possibly along with other pages from the same
2391 * block, depending on fragmentation avoidance heuristics. Returns true if
2392 * fallback was found so that __rmqueue_smallest() can grab it.
2394 * The use of signed ints for order and current_order is a deliberate
2395 * deviation from the rest of this file, to make the for loop
2396 * condition simpler.
2398 static __always_inline
bool
2399 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
2401 struct free_area
*area
;
2408 * Find the largest available free page in the other list. This roughly
2409 * approximates finding the pageblock with the most free pages, which
2410 * would be too costly to do exactly.
2412 for (current_order
= MAX_ORDER
- 1; current_order
>= order
;
2414 area
= &(zone
->free_area
[current_order
]);
2415 fallback_mt
= find_suitable_fallback(area
, current_order
,
2416 start_migratetype
, false, &can_steal
);
2417 if (fallback_mt
== -1)
2421 * We cannot steal all free pages from the pageblock and the
2422 * requested migratetype is movable. In that case it's better to
2423 * steal and split the smallest available page instead of the
2424 * largest available page, because even if the next movable
2425 * allocation falls back into a different pageblock than this
2426 * one, it won't cause permanent fragmentation.
2428 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2429 && current_order
> order
)
2438 for (current_order
= order
; current_order
< MAX_ORDER
;
2440 area
= &(zone
->free_area
[current_order
]);
2441 fallback_mt
= find_suitable_fallback(area
, current_order
,
2442 start_migratetype
, false, &can_steal
);
2443 if (fallback_mt
!= -1)
2448 * This should not happen - we already found a suitable fallback
2449 * when looking for the largest page.
2451 VM_BUG_ON(current_order
== MAX_ORDER
);
2454 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2457 steal_suitable_fallback(zone
, page
, start_migratetype
, can_steal
);
2459 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2460 start_migratetype
, fallback_mt
);
2467 * Do the hard work of removing an element from the buddy allocator.
2468 * Call me with the zone->lock already held.
2470 static __always_inline
struct page
*
2471 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
)
2476 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2477 if (unlikely(!page
)) {
2478 if (migratetype
== MIGRATE_MOVABLE
)
2479 page
= __rmqueue_cma_fallback(zone
, order
);
2481 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
))
2485 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2490 * Obtain a specified number of elements from the buddy allocator, all under
2491 * a single hold of the lock, for efficiency. Add them to the supplied list.
2492 * Returns the number of new pages which were placed at *list.
2494 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2495 unsigned long count
, struct list_head
*list
,
2500 spin_lock(&zone
->lock
);
2501 for (i
= 0; i
< count
; ++i
) {
2502 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2503 if (unlikely(page
== NULL
))
2506 if (unlikely(check_pcp_refill(page
)))
2510 * Split buddy pages returned by expand() are received here in
2511 * physical page order. The page is added to the tail of
2512 * caller's list. From the callers perspective, the linked list
2513 * is ordered by page number under some conditions. This is
2514 * useful for IO devices that can forward direction from the
2515 * head, thus also in the physical page order. This is useful
2516 * for IO devices that can merge IO requests if the physical
2517 * pages are ordered properly.
2519 list_add_tail(&page
->lru
, list
);
2521 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2522 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2527 * i pages were removed from the buddy list even if some leak due
2528 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2529 * on i. Do not confuse with 'alloced' which is the number of
2530 * pages added to the pcp list.
2532 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2533 spin_unlock(&zone
->lock
);
2539 * Called from the vmstat counter updater to drain pagesets of this
2540 * currently executing processor on remote nodes after they have
2543 * Note that this function must be called with the thread pinned to
2544 * a single processor.
2546 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2548 unsigned long flags
;
2549 int to_drain
, batch
;
2551 local_irq_save(flags
);
2552 batch
= READ_ONCE(pcp
->batch
);
2553 to_drain
= min(pcp
->count
, batch
);
2555 free_pcppages_bulk(zone
, to_drain
, pcp
);
2556 local_irq_restore(flags
);
2561 * Drain pcplists of the indicated processor and zone.
2563 * The processor must either be the current processor and the
2564 * thread pinned to the current processor or a processor that
2567 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2569 unsigned long flags
;
2570 struct per_cpu_pageset
*pset
;
2571 struct per_cpu_pages
*pcp
;
2573 local_irq_save(flags
);
2574 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2578 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2579 local_irq_restore(flags
);
2583 * Drain pcplists of all zones on the indicated processor.
2585 * The processor must either be the current processor and the
2586 * thread pinned to the current processor or a processor that
2589 static void drain_pages(unsigned int cpu
)
2593 for_each_populated_zone(zone
) {
2594 drain_pages_zone(cpu
, zone
);
2599 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2601 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2602 * the single zone's pages.
2604 void drain_local_pages(struct zone
*zone
)
2606 int cpu
= smp_processor_id();
2609 drain_pages_zone(cpu
, zone
);
2614 static void drain_local_pages_wq(struct work_struct
*work
)
2617 * drain_all_pages doesn't use proper cpu hotplug protection so
2618 * we can race with cpu offline when the WQ can move this from
2619 * a cpu pinned worker to an unbound one. We can operate on a different
2620 * cpu which is allright but we also have to make sure to not move to
2624 drain_local_pages(NULL
);
2629 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2631 * When zone parameter is non-NULL, spill just the single zone's pages.
2633 * Note that this can be extremely slow as the draining happens in a workqueue.
2635 void drain_all_pages(struct zone
*zone
)
2640 * Allocate in the BSS so we wont require allocation in
2641 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2643 static cpumask_t cpus_with_pcps
;
2646 * Make sure nobody triggers this path before mm_percpu_wq is fully
2649 if (WARN_ON_ONCE(!mm_percpu_wq
))
2653 * Do not drain if one is already in progress unless it's specific to
2654 * a zone. Such callers are primarily CMA and memory hotplug and need
2655 * the drain to be complete when the call returns.
2657 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2660 mutex_lock(&pcpu_drain_mutex
);
2664 * We don't care about racing with CPU hotplug event
2665 * as offline notification will cause the notified
2666 * cpu to drain that CPU pcps and on_each_cpu_mask
2667 * disables preemption as part of its processing
2669 for_each_online_cpu(cpu
) {
2670 struct per_cpu_pageset
*pcp
;
2672 bool has_pcps
= false;
2675 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2679 for_each_populated_zone(z
) {
2680 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2681 if (pcp
->pcp
.count
) {
2689 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2691 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2694 for_each_cpu(cpu
, &cpus_with_pcps
) {
2695 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2696 INIT_WORK(work
, drain_local_pages_wq
);
2697 queue_work_on(cpu
, mm_percpu_wq
, work
);
2699 for_each_cpu(cpu
, &cpus_with_pcps
)
2700 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2702 mutex_unlock(&pcpu_drain_mutex
);
2705 #ifdef CONFIG_HIBERNATION
2708 * Touch the watchdog for every WD_PAGE_COUNT pages.
2710 #define WD_PAGE_COUNT (128*1024)
2712 void mark_free_pages(struct zone
*zone
)
2714 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2715 unsigned long flags
;
2716 unsigned int order
, t
;
2719 if (zone_is_empty(zone
))
2722 spin_lock_irqsave(&zone
->lock
, flags
);
2724 max_zone_pfn
= zone_end_pfn(zone
);
2725 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2726 if (pfn_valid(pfn
)) {
2727 page
= pfn_to_page(pfn
);
2729 if (!--page_count
) {
2730 touch_nmi_watchdog();
2731 page_count
= WD_PAGE_COUNT
;
2734 if (page_zone(page
) != zone
)
2737 if (!swsusp_page_is_forbidden(page
))
2738 swsusp_unset_page_free(page
);
2741 for_each_migratetype_order(order
, t
) {
2742 list_for_each_entry(page
,
2743 &zone
->free_area
[order
].free_list
[t
], lru
) {
2746 pfn
= page_to_pfn(page
);
2747 for (i
= 0; i
< (1UL << order
); i
++) {
2748 if (!--page_count
) {
2749 touch_nmi_watchdog();
2750 page_count
= WD_PAGE_COUNT
;
2752 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2756 spin_unlock_irqrestore(&zone
->lock
, flags
);
2758 #endif /* CONFIG_PM */
2760 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
2764 if (!free_pcp_prepare(page
))
2767 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2768 set_pcppage_migratetype(page
, migratetype
);
2772 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
2774 struct zone
*zone
= page_zone(page
);
2775 struct per_cpu_pages
*pcp
;
2778 migratetype
= get_pcppage_migratetype(page
);
2779 __count_vm_event(PGFREE
);
2782 * We only track unmovable, reclaimable and movable on pcp lists.
2783 * Free ISOLATE pages back to the allocator because they are being
2784 * offlined but treat HIGHATOMIC as movable pages so we can get those
2785 * areas back if necessary. Otherwise, we may have to free
2786 * excessively into the page allocator
2788 if (migratetype
>= MIGRATE_PCPTYPES
) {
2789 if (unlikely(is_migrate_isolate(migratetype
))) {
2790 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2793 migratetype
= MIGRATE_MOVABLE
;
2796 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2797 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2799 if (pcp
->count
>= pcp
->high
) {
2800 unsigned long batch
= READ_ONCE(pcp
->batch
);
2801 free_pcppages_bulk(zone
, batch
, pcp
);
2806 * Free a 0-order page
2808 void free_unref_page(struct page
*page
)
2810 unsigned long flags
;
2811 unsigned long pfn
= page_to_pfn(page
);
2813 if (!free_unref_page_prepare(page
, pfn
))
2816 local_irq_save(flags
);
2817 free_unref_page_commit(page
, pfn
);
2818 local_irq_restore(flags
);
2822 * Free a list of 0-order pages
2824 void free_unref_page_list(struct list_head
*list
)
2826 struct page
*page
, *next
;
2827 unsigned long flags
, pfn
;
2828 int batch_count
= 0;
2830 /* Prepare pages for freeing */
2831 list_for_each_entry_safe(page
, next
, list
, lru
) {
2832 pfn
= page_to_pfn(page
);
2833 if (!free_unref_page_prepare(page
, pfn
))
2834 list_del(&page
->lru
);
2835 set_page_private(page
, pfn
);
2838 local_irq_save(flags
);
2839 list_for_each_entry_safe(page
, next
, list
, lru
) {
2840 unsigned long pfn
= page_private(page
);
2842 set_page_private(page
, 0);
2843 trace_mm_page_free_batched(page
);
2844 free_unref_page_commit(page
, pfn
);
2847 * Guard against excessive IRQ disabled times when we get
2848 * a large list of pages to free.
2850 if (++batch_count
== SWAP_CLUSTER_MAX
) {
2851 local_irq_restore(flags
);
2853 local_irq_save(flags
);
2856 local_irq_restore(flags
);
2860 * split_page takes a non-compound higher-order page, and splits it into
2861 * n (1<<order) sub-pages: page[0..n]
2862 * Each sub-page must be freed individually.
2864 * Note: this is probably too low level an operation for use in drivers.
2865 * Please consult with lkml before using this in your driver.
2867 void split_page(struct page
*page
, unsigned int order
)
2871 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2872 VM_BUG_ON_PAGE(!page_count(page
), page
);
2874 for (i
= 1; i
< (1 << order
); i
++)
2875 set_page_refcounted(page
+ i
);
2876 split_page_owner(page
, order
);
2878 EXPORT_SYMBOL_GPL(split_page
);
2880 int __isolate_free_page(struct page
*page
, unsigned int order
)
2882 unsigned long watermark
;
2886 BUG_ON(!PageBuddy(page
));
2888 zone
= page_zone(page
);
2889 mt
= get_pageblock_migratetype(page
);
2891 if (!is_migrate_isolate(mt
)) {
2893 * Obey watermarks as if the page was being allocated. We can
2894 * emulate a high-order watermark check with a raised order-0
2895 * watermark, because we already know our high-order page
2898 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2899 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2902 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2905 /* Remove page from free list */
2906 list_del(&page
->lru
);
2907 zone
->free_area
[order
].nr_free
--;
2908 rmv_page_order(page
);
2911 * Set the pageblock if the isolated page is at least half of a
2914 if (order
>= pageblock_order
- 1) {
2915 struct page
*endpage
= page
+ (1 << order
) - 1;
2916 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2917 int mt
= get_pageblock_migratetype(page
);
2918 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2919 && !is_migrate_highatomic(mt
))
2920 set_pageblock_migratetype(page
,
2926 return 1UL << order
;
2930 * Update NUMA hit/miss statistics
2932 * Must be called with interrupts disabled.
2934 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2937 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2939 /* skip numa counters update if numa stats is disabled */
2940 if (!static_branch_likely(&vm_numa_stat_key
))
2943 if (zone_to_nid(z
) != numa_node_id())
2944 local_stat
= NUMA_OTHER
;
2946 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
2947 __inc_numa_state(z
, NUMA_HIT
);
2949 __inc_numa_state(z
, NUMA_MISS
);
2950 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
2952 __inc_numa_state(z
, local_stat
);
2956 /* Remove page from the per-cpu list, caller must protect the list */
2957 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2958 struct per_cpu_pages
*pcp
,
2959 struct list_head
*list
)
2964 if (list_empty(list
)) {
2965 pcp
->count
+= rmqueue_bulk(zone
, 0,
2968 if (unlikely(list_empty(list
)))
2972 page
= list_first_entry(list
, struct page
, lru
);
2973 list_del(&page
->lru
);
2975 } while (check_new_pcp(page
));
2980 /* Lock and remove page from the per-cpu list */
2981 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2982 struct zone
*zone
, unsigned int order
,
2983 gfp_t gfp_flags
, int migratetype
)
2985 struct per_cpu_pages
*pcp
;
2986 struct list_head
*list
;
2988 unsigned long flags
;
2990 local_irq_save(flags
);
2991 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2992 list
= &pcp
->lists
[migratetype
];
2993 page
= __rmqueue_pcplist(zone
, migratetype
, pcp
, list
);
2995 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2996 zone_statistics(preferred_zone
, zone
);
2998 local_irq_restore(flags
);
3003 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3006 struct page
*rmqueue(struct zone
*preferred_zone
,
3007 struct zone
*zone
, unsigned int order
,
3008 gfp_t gfp_flags
, unsigned int alloc_flags
,
3011 unsigned long flags
;
3014 if (likely(order
== 0)) {
3015 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
3016 gfp_flags
, migratetype
);
3021 * We most definitely don't want callers attempting to
3022 * allocate greater than order-1 page units with __GFP_NOFAIL.
3024 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3025 spin_lock_irqsave(&zone
->lock
, flags
);
3029 if (alloc_flags
& ALLOC_HARDER
) {
3030 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3032 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3035 page
= __rmqueue(zone
, order
, migratetype
);
3036 } while (page
&& check_new_pages(page
, order
));
3037 spin_unlock(&zone
->lock
);
3040 __mod_zone_freepage_state(zone
, -(1 << order
),
3041 get_pcppage_migratetype(page
));
3043 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3044 zone_statistics(preferred_zone
, zone
);
3045 local_irq_restore(flags
);
3048 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3052 local_irq_restore(flags
);
3056 #ifdef CONFIG_FAIL_PAGE_ALLOC
3059 struct fault_attr attr
;
3061 bool ignore_gfp_highmem
;
3062 bool ignore_gfp_reclaim
;
3064 } fail_page_alloc
= {
3065 .attr
= FAULT_ATTR_INITIALIZER
,
3066 .ignore_gfp_reclaim
= true,
3067 .ignore_gfp_highmem
= true,
3071 static int __init
setup_fail_page_alloc(char *str
)
3073 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3075 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3077 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3079 if (order
< fail_page_alloc
.min_order
)
3081 if (gfp_mask
& __GFP_NOFAIL
)
3083 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3085 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3086 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3089 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3092 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3094 static int __init
fail_page_alloc_debugfs(void)
3096 umode_t mode
= S_IFREG
| 0600;
3099 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3100 &fail_page_alloc
.attr
);
3102 return PTR_ERR(dir
);
3104 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3105 &fail_page_alloc
.ignore_gfp_reclaim
))
3107 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3108 &fail_page_alloc
.ignore_gfp_highmem
))
3110 if (!debugfs_create_u32("min-order", mode
, dir
,
3111 &fail_page_alloc
.min_order
))
3116 debugfs_remove_recursive(dir
);
3121 late_initcall(fail_page_alloc_debugfs
);
3123 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3125 #else /* CONFIG_FAIL_PAGE_ALLOC */
3127 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3132 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3135 * Return true if free base pages are above 'mark'. For high-order checks it
3136 * will return true of the order-0 watermark is reached and there is at least
3137 * one free page of a suitable size. Checking now avoids taking the zone lock
3138 * to check in the allocation paths if no pages are free.
3140 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3141 int classzone_idx
, unsigned int alloc_flags
,
3146 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3148 /* free_pages may go negative - that's OK */
3149 free_pages
-= (1 << order
) - 1;
3151 if (alloc_flags
& ALLOC_HIGH
)
3155 * If the caller does not have rights to ALLOC_HARDER then subtract
3156 * the high-atomic reserves. This will over-estimate the size of the
3157 * atomic reserve but it avoids a search.
3159 if (likely(!alloc_harder
)) {
3160 free_pages
-= z
->nr_reserved_highatomic
;
3163 * OOM victims can try even harder than normal ALLOC_HARDER
3164 * users on the grounds that it's definitely going to be in
3165 * the exit path shortly and free memory. Any allocation it
3166 * makes during the free path will be small and short-lived.
3168 if (alloc_flags
& ALLOC_OOM
)
3176 /* If allocation can't use CMA areas don't use free CMA pages */
3177 if (!(alloc_flags
& ALLOC_CMA
))
3178 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3182 * Check watermarks for an order-0 allocation request. If these
3183 * are not met, then a high-order request also cannot go ahead
3184 * even if a suitable page happened to be free.
3186 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3189 /* If this is an order-0 request then the watermark is fine */
3193 /* For a high-order request, check at least one suitable page is free */
3194 for (o
= order
; o
< MAX_ORDER
; o
++) {
3195 struct free_area
*area
= &z
->free_area
[o
];
3201 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3202 if (!list_empty(&area
->free_list
[mt
]))
3207 if ((alloc_flags
& ALLOC_CMA
) &&
3208 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
3213 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3219 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3220 int classzone_idx
, unsigned int alloc_flags
)
3222 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3223 zone_page_state(z
, NR_FREE_PAGES
));
3226 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3227 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3229 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3233 /* If allocation can't use CMA areas don't use free CMA pages */
3234 if (!(alloc_flags
& ALLOC_CMA
))
3235 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3239 * Fast check for order-0 only. If this fails then the reserves
3240 * need to be calculated. There is a corner case where the check
3241 * passes but only the high-order atomic reserve are free. If
3242 * the caller is !atomic then it'll uselessly search the free
3243 * list. That corner case is then slower but it is harmless.
3245 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3248 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3252 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3253 unsigned long mark
, int classzone_idx
)
3255 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3257 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3258 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3260 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3265 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3267 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3270 #else /* CONFIG_NUMA */
3271 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3275 #endif /* CONFIG_NUMA */
3278 * get_page_from_freelist goes through the zonelist trying to allocate
3281 static struct page
*
3282 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3283 const struct alloc_context
*ac
)
3285 struct zoneref
*z
= ac
->preferred_zoneref
;
3287 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3290 * Scan zonelist, looking for a zone with enough free.
3291 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3293 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3298 if (cpusets_enabled() &&
3299 (alloc_flags
& ALLOC_CPUSET
) &&
3300 !__cpuset_zone_allowed(zone
, gfp_mask
))
3303 * When allocating a page cache page for writing, we
3304 * want to get it from a node that is within its dirty
3305 * limit, such that no single node holds more than its
3306 * proportional share of globally allowed dirty pages.
3307 * The dirty limits take into account the node's
3308 * lowmem reserves and high watermark so that kswapd
3309 * should be able to balance it without having to
3310 * write pages from its LRU list.
3312 * XXX: For now, allow allocations to potentially
3313 * exceed the per-node dirty limit in the slowpath
3314 * (spread_dirty_pages unset) before going into reclaim,
3315 * which is important when on a NUMA setup the allowed
3316 * nodes are together not big enough to reach the
3317 * global limit. The proper fix for these situations
3318 * will require awareness of nodes in the
3319 * dirty-throttling and the flusher threads.
3321 if (ac
->spread_dirty_pages
) {
3322 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3325 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3326 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3331 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
3332 if (!zone_watermark_fast(zone
, order
, mark
,
3333 ac_classzone_idx(ac
), alloc_flags
)) {
3336 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3338 * Watermark failed for this zone, but see if we can
3339 * grow this zone if it contains deferred pages.
3341 if (static_branch_unlikely(&deferred_pages
)) {
3342 if (_deferred_grow_zone(zone
, order
))
3346 /* Checked here to keep the fast path fast */
3347 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3348 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3351 if (node_reclaim_mode
== 0 ||
3352 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3355 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3357 case NODE_RECLAIM_NOSCAN
:
3360 case NODE_RECLAIM_FULL
:
3361 /* scanned but unreclaimable */
3364 /* did we reclaim enough */
3365 if (zone_watermark_ok(zone
, order
, mark
,
3366 ac_classzone_idx(ac
), alloc_flags
))
3374 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3375 gfp_mask
, alloc_flags
, ac
->migratetype
);
3377 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3380 * If this is a high-order atomic allocation then check
3381 * if the pageblock should be reserved for the future
3383 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3384 reserve_highatomic_pageblock(page
, zone
, order
);
3388 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3389 /* Try again if zone has deferred pages */
3390 if (static_branch_unlikely(&deferred_pages
)) {
3391 if (_deferred_grow_zone(zone
, order
))
3401 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3403 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3404 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3406 if (!__ratelimit(&show_mem_rs
))
3410 * This documents exceptions given to allocations in certain
3411 * contexts that are allowed to allocate outside current's set
3414 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3415 if (tsk_is_oom_victim(current
) ||
3416 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3417 filter
&= ~SHOW_MEM_FILTER_NODES
;
3418 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3419 filter
&= ~SHOW_MEM_FILTER_NODES
;
3421 show_mem(filter
, nodemask
);
3424 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3426 struct va_format vaf
;
3428 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3429 DEFAULT_RATELIMIT_BURST
);
3431 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3434 va_start(args
, fmt
);
3437 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3438 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3439 nodemask_pr_args(nodemask
));
3442 cpuset_print_current_mems_allowed();
3445 warn_alloc_show_mem(gfp_mask
, nodemask
);
3448 static inline struct page
*
3449 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3450 unsigned int alloc_flags
,
3451 const struct alloc_context
*ac
)
3455 page
= get_page_from_freelist(gfp_mask
, order
,
3456 alloc_flags
|ALLOC_CPUSET
, ac
);
3458 * fallback to ignore cpuset restriction if our nodes
3462 page
= get_page_from_freelist(gfp_mask
, order
,
3468 static inline struct page
*
3469 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3470 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3472 struct oom_control oc
= {
3473 .zonelist
= ac
->zonelist
,
3474 .nodemask
= ac
->nodemask
,
3476 .gfp_mask
= gfp_mask
,
3481 *did_some_progress
= 0;
3484 * Acquire the oom lock. If that fails, somebody else is
3485 * making progress for us.
3487 if (!mutex_trylock(&oom_lock
)) {
3488 *did_some_progress
= 1;
3489 schedule_timeout_uninterruptible(1);
3494 * Go through the zonelist yet one more time, keep very high watermark
3495 * here, this is only to catch a parallel oom killing, we must fail if
3496 * we're still under heavy pressure. But make sure that this reclaim
3497 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3498 * allocation which will never fail due to oom_lock already held.
3500 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3501 ~__GFP_DIRECT_RECLAIM
, order
,
3502 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3506 /* Coredumps can quickly deplete all memory reserves */
3507 if (current
->flags
& PF_DUMPCORE
)
3509 /* The OOM killer will not help higher order allocs */
3510 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3513 * We have already exhausted all our reclaim opportunities without any
3514 * success so it is time to admit defeat. We will skip the OOM killer
3515 * because it is very likely that the caller has a more reasonable
3516 * fallback than shooting a random task.
3518 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3520 /* The OOM killer does not needlessly kill tasks for lowmem */
3521 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3523 if (pm_suspended_storage())
3526 * XXX: GFP_NOFS allocations should rather fail than rely on
3527 * other request to make a forward progress.
3528 * We are in an unfortunate situation where out_of_memory cannot
3529 * do much for this context but let's try it to at least get
3530 * access to memory reserved if the current task is killed (see
3531 * out_of_memory). Once filesystems are ready to handle allocation
3532 * failures more gracefully we should just bail out here.
3535 /* The OOM killer may not free memory on a specific node */
3536 if (gfp_mask
& __GFP_THISNODE
)
3539 /* Exhausted what can be done so it's blame time */
3540 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3541 *did_some_progress
= 1;
3544 * Help non-failing allocations by giving them access to memory
3547 if (gfp_mask
& __GFP_NOFAIL
)
3548 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3549 ALLOC_NO_WATERMARKS
, ac
);
3552 mutex_unlock(&oom_lock
);
3557 * Maximum number of compaction retries wit a progress before OOM
3558 * killer is consider as the only way to move forward.
3560 #define MAX_COMPACT_RETRIES 16
3562 #ifdef CONFIG_COMPACTION
3563 /* Try memory compaction for high-order allocations before reclaim */
3564 static struct page
*
3565 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3566 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3567 enum compact_priority prio
, enum compact_result
*compact_result
)
3570 unsigned long pflags
;
3571 unsigned int noreclaim_flag
;
3576 psi_memstall_enter(&pflags
);
3577 noreclaim_flag
= memalloc_noreclaim_save();
3579 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3582 memalloc_noreclaim_restore(noreclaim_flag
);
3583 psi_memstall_leave(&pflags
);
3585 if (*compact_result
<= COMPACT_INACTIVE
)
3589 * At least in one zone compaction wasn't deferred or skipped, so let's
3590 * count a compaction stall
3592 count_vm_event(COMPACTSTALL
);
3594 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3597 struct zone
*zone
= page_zone(page
);
3599 zone
->compact_blockskip_flush
= false;
3600 compaction_defer_reset(zone
, order
, true);
3601 count_vm_event(COMPACTSUCCESS
);
3606 * It's bad if compaction run occurs and fails. The most likely reason
3607 * is that pages exist, but not enough to satisfy watermarks.
3609 count_vm_event(COMPACTFAIL
);
3617 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3618 enum compact_result compact_result
,
3619 enum compact_priority
*compact_priority
,
3620 int *compaction_retries
)
3622 int max_retries
= MAX_COMPACT_RETRIES
;
3625 int retries
= *compaction_retries
;
3626 enum compact_priority priority
= *compact_priority
;
3631 if (compaction_made_progress(compact_result
))
3632 (*compaction_retries
)++;
3635 * compaction considers all the zone as desperately out of memory
3636 * so it doesn't really make much sense to retry except when the
3637 * failure could be caused by insufficient priority
3639 if (compaction_failed(compact_result
))
3640 goto check_priority
;
3643 * make sure the compaction wasn't deferred or didn't bail out early
3644 * due to locks contention before we declare that we should give up.
3645 * But do not retry if the given zonelist is not suitable for
3648 if (compaction_withdrawn(compact_result
)) {
3649 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3654 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3655 * costly ones because they are de facto nofail and invoke OOM
3656 * killer to move on while costly can fail and users are ready
3657 * to cope with that. 1/4 retries is rather arbitrary but we
3658 * would need much more detailed feedback from compaction to
3659 * make a better decision.
3661 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3663 if (*compaction_retries
<= max_retries
) {
3669 * Make sure there are attempts at the highest priority if we exhausted
3670 * all retries or failed at the lower priorities.
3673 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3674 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3676 if (*compact_priority
> min_priority
) {
3677 (*compact_priority
)--;
3678 *compaction_retries
= 0;
3682 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3686 static inline struct page
*
3687 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3688 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3689 enum compact_priority prio
, enum compact_result
*compact_result
)
3691 *compact_result
= COMPACT_SKIPPED
;
3696 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3697 enum compact_result compact_result
,
3698 enum compact_priority
*compact_priority
,
3699 int *compaction_retries
)
3704 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3708 * There are setups with compaction disabled which would prefer to loop
3709 * inside the allocator rather than hit the oom killer prematurely.
3710 * Let's give them a good hope and keep retrying while the order-0
3711 * watermarks are OK.
3713 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3715 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3716 ac_classzone_idx(ac
), alloc_flags
))
3721 #endif /* CONFIG_COMPACTION */
3723 #ifdef CONFIG_LOCKDEP
3724 static struct lockdep_map __fs_reclaim_map
=
3725 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3727 static bool __need_fs_reclaim(gfp_t gfp_mask
)
3729 gfp_mask
= current_gfp_context(gfp_mask
);
3731 /* no reclaim without waiting on it */
3732 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3735 /* this guy won't enter reclaim */
3736 if (current
->flags
& PF_MEMALLOC
)
3739 /* We're only interested __GFP_FS allocations for now */
3740 if (!(gfp_mask
& __GFP_FS
))
3743 if (gfp_mask
& __GFP_NOLOCKDEP
)
3749 void __fs_reclaim_acquire(void)
3751 lock_map_acquire(&__fs_reclaim_map
);
3754 void __fs_reclaim_release(void)
3756 lock_map_release(&__fs_reclaim_map
);
3759 void fs_reclaim_acquire(gfp_t gfp_mask
)
3761 if (__need_fs_reclaim(gfp_mask
))
3762 __fs_reclaim_acquire();
3764 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3766 void fs_reclaim_release(gfp_t gfp_mask
)
3768 if (__need_fs_reclaim(gfp_mask
))
3769 __fs_reclaim_release();
3771 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3774 /* Perform direct synchronous page reclaim */
3776 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3777 const struct alloc_context
*ac
)
3779 struct reclaim_state reclaim_state
;
3781 unsigned int noreclaim_flag
;
3782 unsigned long pflags
;
3786 /* We now go into synchronous reclaim */
3787 cpuset_memory_pressure_bump();
3788 psi_memstall_enter(&pflags
);
3789 fs_reclaim_acquire(gfp_mask
);
3790 noreclaim_flag
= memalloc_noreclaim_save();
3791 reclaim_state
.reclaimed_slab
= 0;
3792 current
->reclaim_state
= &reclaim_state
;
3794 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3797 current
->reclaim_state
= NULL
;
3798 memalloc_noreclaim_restore(noreclaim_flag
);
3799 fs_reclaim_release(gfp_mask
);
3800 psi_memstall_leave(&pflags
);
3807 /* The really slow allocator path where we enter direct reclaim */
3808 static inline struct page
*
3809 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3810 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3811 unsigned long *did_some_progress
)
3813 struct page
*page
= NULL
;
3814 bool drained
= false;
3816 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3817 if (unlikely(!(*did_some_progress
)))
3821 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3824 * If an allocation failed after direct reclaim, it could be because
3825 * pages are pinned on the per-cpu lists or in high alloc reserves.
3826 * Shrink them them and try again
3828 if (!page
&& !drained
) {
3829 unreserve_highatomic_pageblock(ac
, false);
3830 drain_all_pages(NULL
);
3838 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
3839 const struct alloc_context
*ac
)
3843 pg_data_t
*last_pgdat
= NULL
;
3844 enum zone_type high_zoneidx
= ac
->high_zoneidx
;
3846 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, high_zoneidx
,
3848 if (last_pgdat
!= zone
->zone_pgdat
)
3849 wakeup_kswapd(zone
, gfp_mask
, order
, high_zoneidx
);
3850 last_pgdat
= zone
->zone_pgdat
;
3854 static inline unsigned int
3855 gfp_to_alloc_flags(gfp_t gfp_mask
)
3857 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3859 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3860 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3863 * The caller may dip into page reserves a bit more if the caller
3864 * cannot run direct reclaim, or if the caller has realtime scheduling
3865 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3866 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3868 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3870 if (gfp_mask
& __GFP_ATOMIC
) {
3872 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3873 * if it can't schedule.
3875 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3876 alloc_flags
|= ALLOC_HARDER
;
3878 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3879 * comment for __cpuset_node_allowed().
3881 alloc_flags
&= ~ALLOC_CPUSET
;
3882 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3883 alloc_flags
|= ALLOC_HARDER
;
3886 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3887 alloc_flags
|= ALLOC_CMA
;
3892 static bool oom_reserves_allowed(struct task_struct
*tsk
)
3894 if (!tsk_is_oom_victim(tsk
))
3898 * !MMU doesn't have oom reaper so give access to memory reserves
3899 * only to the thread with TIF_MEMDIE set
3901 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
3908 * Distinguish requests which really need access to full memory
3909 * reserves from oom victims which can live with a portion of it
3911 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
3913 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3915 if (gfp_mask
& __GFP_MEMALLOC
)
3916 return ALLOC_NO_WATERMARKS
;
3917 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3918 return ALLOC_NO_WATERMARKS
;
3919 if (!in_interrupt()) {
3920 if (current
->flags
& PF_MEMALLOC
)
3921 return ALLOC_NO_WATERMARKS
;
3922 else if (oom_reserves_allowed(current
))
3929 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3931 return !!__gfp_pfmemalloc_flags(gfp_mask
);
3935 * Checks whether it makes sense to retry the reclaim to make a forward progress
3936 * for the given allocation request.
3938 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3939 * without success, or when we couldn't even meet the watermark if we
3940 * reclaimed all remaining pages on the LRU lists.
3942 * Returns true if a retry is viable or false to enter the oom path.
3945 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3946 struct alloc_context
*ac
, int alloc_flags
,
3947 bool did_some_progress
, int *no_progress_loops
)
3954 * Costly allocations might have made a progress but this doesn't mean
3955 * their order will become available due to high fragmentation so
3956 * always increment the no progress counter for them
3958 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3959 *no_progress_loops
= 0;
3961 (*no_progress_loops
)++;
3964 * Make sure we converge to OOM if we cannot make any progress
3965 * several times in the row.
3967 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3968 /* Before OOM, exhaust highatomic_reserve */
3969 return unreserve_highatomic_pageblock(ac
, true);
3973 * Keep reclaiming pages while there is a chance this will lead
3974 * somewhere. If none of the target zones can satisfy our allocation
3975 * request even if all reclaimable pages are considered then we are
3976 * screwed and have to go OOM.
3978 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3980 unsigned long available
;
3981 unsigned long reclaimable
;
3982 unsigned long min_wmark
= min_wmark_pages(zone
);
3985 available
= reclaimable
= zone_reclaimable_pages(zone
);
3986 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3989 * Would the allocation succeed if we reclaimed all
3990 * reclaimable pages?
3992 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3993 ac_classzone_idx(ac
), alloc_flags
, available
);
3994 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3995 available
, min_wmark
, *no_progress_loops
, wmark
);
3998 * If we didn't make any progress and have a lot of
3999 * dirty + writeback pages then we should wait for
4000 * an IO to complete to slow down the reclaim and
4001 * prevent from pre mature OOM
4003 if (!did_some_progress
) {
4004 unsigned long write_pending
;
4006 write_pending
= zone_page_state_snapshot(zone
,
4007 NR_ZONE_WRITE_PENDING
);
4009 if (2 * write_pending
> reclaimable
) {
4010 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4022 * Memory allocation/reclaim might be called from a WQ context and the
4023 * current implementation of the WQ concurrency control doesn't
4024 * recognize that a particular WQ is congested if the worker thread is
4025 * looping without ever sleeping. Therefore we have to do a short sleep
4026 * here rather than calling cond_resched().
4028 if (current
->flags
& PF_WQ_WORKER
)
4029 schedule_timeout_uninterruptible(1);
4036 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4039 * It's possible that cpuset's mems_allowed and the nodemask from
4040 * mempolicy don't intersect. This should be normally dealt with by
4041 * policy_nodemask(), but it's possible to race with cpuset update in
4042 * such a way the check therein was true, and then it became false
4043 * before we got our cpuset_mems_cookie here.
4044 * This assumes that for all allocations, ac->nodemask can come only
4045 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4046 * when it does not intersect with the cpuset restrictions) or the
4047 * caller can deal with a violated nodemask.
4049 if (cpusets_enabled() && ac
->nodemask
&&
4050 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4051 ac
->nodemask
= NULL
;
4056 * When updating a task's mems_allowed or mempolicy nodemask, it is
4057 * possible to race with parallel threads in such a way that our
4058 * allocation can fail while the mask is being updated. If we are about
4059 * to fail, check if the cpuset changed during allocation and if so,
4062 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4068 static inline struct page
*
4069 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4070 struct alloc_context
*ac
)
4072 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4073 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4074 struct page
*page
= NULL
;
4075 unsigned int alloc_flags
;
4076 unsigned long did_some_progress
;
4077 enum compact_priority compact_priority
;
4078 enum compact_result compact_result
;
4079 int compaction_retries
;
4080 int no_progress_loops
;
4081 unsigned int cpuset_mems_cookie
;
4085 * We also sanity check to catch abuse of atomic reserves being used by
4086 * callers that are not in atomic context.
4088 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4089 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4090 gfp_mask
&= ~__GFP_ATOMIC
;
4093 compaction_retries
= 0;
4094 no_progress_loops
= 0;
4095 compact_priority
= DEF_COMPACT_PRIORITY
;
4096 cpuset_mems_cookie
= read_mems_allowed_begin();
4099 * The fast path uses conservative alloc_flags to succeed only until
4100 * kswapd needs to be woken up, and to avoid the cost of setting up
4101 * alloc_flags precisely. So we do that now.
4103 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4106 * We need to recalculate the starting point for the zonelist iterator
4107 * because we might have used different nodemask in the fast path, or
4108 * there was a cpuset modification and we are retrying - otherwise we
4109 * could end up iterating over non-eligible zones endlessly.
4111 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4112 ac
->high_zoneidx
, ac
->nodemask
);
4113 if (!ac
->preferred_zoneref
->zone
)
4116 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4117 wake_all_kswapds(order
, gfp_mask
, ac
);
4120 * The adjusted alloc_flags might result in immediate success, so try
4123 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4128 * For costly allocations, try direct compaction first, as it's likely
4129 * that we have enough base pages and don't need to reclaim. For non-
4130 * movable high-order allocations, do that as well, as compaction will
4131 * try prevent permanent fragmentation by migrating from blocks of the
4133 * Don't try this for allocations that are allowed to ignore
4134 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4136 if (can_direct_reclaim
&&
4138 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4139 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4140 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4142 INIT_COMPACT_PRIORITY
,
4148 * Checks for costly allocations with __GFP_NORETRY, which
4149 * includes THP page fault allocations
4151 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4153 * If compaction is deferred for high-order allocations,
4154 * it is because sync compaction recently failed. If
4155 * this is the case and the caller requested a THP
4156 * allocation, we do not want to heavily disrupt the
4157 * system, so we fail the allocation instead of entering
4160 if (compact_result
== COMPACT_DEFERRED
)
4164 * Looks like reclaim/compaction is worth trying, but
4165 * sync compaction could be very expensive, so keep
4166 * using async compaction.
4168 compact_priority
= INIT_COMPACT_PRIORITY
;
4173 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4174 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4175 wake_all_kswapds(order
, gfp_mask
, ac
);
4177 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4179 alloc_flags
= reserve_flags
;
4182 * Reset the nodemask and zonelist iterators if memory policies can be
4183 * ignored. These allocations are high priority and system rather than
4186 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4187 ac
->nodemask
= NULL
;
4188 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4189 ac
->high_zoneidx
, ac
->nodemask
);
4192 /* Attempt with potentially adjusted zonelist and alloc_flags */
4193 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4197 /* Caller is not willing to reclaim, we can't balance anything */
4198 if (!can_direct_reclaim
)
4201 /* Avoid recursion of direct reclaim */
4202 if (current
->flags
& PF_MEMALLOC
)
4205 /* Try direct reclaim and then allocating */
4206 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4207 &did_some_progress
);
4211 /* Try direct compaction and then allocating */
4212 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4213 compact_priority
, &compact_result
);
4217 /* Do not loop if specifically requested */
4218 if (gfp_mask
& __GFP_NORETRY
)
4222 * Do not retry costly high order allocations unless they are
4223 * __GFP_RETRY_MAYFAIL
4225 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4228 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4229 did_some_progress
> 0, &no_progress_loops
))
4233 * It doesn't make any sense to retry for the compaction if the order-0
4234 * reclaim is not able to make any progress because the current
4235 * implementation of the compaction depends on the sufficient amount
4236 * of free memory (see __compaction_suitable)
4238 if (did_some_progress
> 0 &&
4239 should_compact_retry(ac
, order
, alloc_flags
,
4240 compact_result
, &compact_priority
,
4241 &compaction_retries
))
4245 /* Deal with possible cpuset update races before we start OOM killing */
4246 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4249 /* Reclaim has failed us, start killing things */
4250 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4254 /* Avoid allocations with no watermarks from looping endlessly */
4255 if (tsk_is_oom_victim(current
) &&
4256 (alloc_flags
== ALLOC_OOM
||
4257 (gfp_mask
& __GFP_NOMEMALLOC
)))
4260 /* Retry as long as the OOM killer is making progress */
4261 if (did_some_progress
) {
4262 no_progress_loops
= 0;
4267 /* Deal with possible cpuset update races before we fail */
4268 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4272 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4275 if (gfp_mask
& __GFP_NOFAIL
) {
4277 * All existing users of the __GFP_NOFAIL are blockable, so warn
4278 * of any new users that actually require GFP_NOWAIT
4280 if (WARN_ON_ONCE(!can_direct_reclaim
))
4284 * PF_MEMALLOC request from this context is rather bizarre
4285 * because we cannot reclaim anything and only can loop waiting
4286 * for somebody to do a work for us
4288 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4291 * non failing costly orders are a hard requirement which we
4292 * are not prepared for much so let's warn about these users
4293 * so that we can identify them and convert them to something
4296 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4299 * Help non-failing allocations by giving them access to memory
4300 * reserves but do not use ALLOC_NO_WATERMARKS because this
4301 * could deplete whole memory reserves which would just make
4302 * the situation worse
4304 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4312 warn_alloc(gfp_mask
, ac
->nodemask
,
4313 "page allocation failure: order:%u", order
);
4318 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4319 int preferred_nid
, nodemask_t
*nodemask
,
4320 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4321 unsigned int *alloc_flags
)
4323 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4324 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4325 ac
->nodemask
= nodemask
;
4326 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4328 if (cpusets_enabled()) {
4329 *alloc_mask
|= __GFP_HARDWALL
;
4331 ac
->nodemask
= &cpuset_current_mems_allowed
;
4333 *alloc_flags
|= ALLOC_CPUSET
;
4336 fs_reclaim_acquire(gfp_mask
);
4337 fs_reclaim_release(gfp_mask
);
4339 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4341 if (should_fail_alloc_page(gfp_mask
, order
))
4344 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4345 *alloc_flags
|= ALLOC_CMA
;
4350 /* Determine whether to spread dirty pages and what the first usable zone */
4351 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4353 /* Dirty zone balancing only done in the fast path */
4354 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4357 * The preferred zone is used for statistics but crucially it is
4358 * also used as the starting point for the zonelist iterator. It
4359 * may get reset for allocations that ignore memory policies.
4361 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4362 ac
->high_zoneidx
, ac
->nodemask
);
4366 * This is the 'heart' of the zoned buddy allocator.
4369 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4370 nodemask_t
*nodemask
)
4373 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4374 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4375 struct alloc_context ac
= { };
4378 * There are several places where we assume that the order value is sane
4379 * so bail out early if the request is out of bound.
4381 if (unlikely(order
>= MAX_ORDER
)) {
4382 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4386 gfp_mask
&= gfp_allowed_mask
;
4387 alloc_mask
= gfp_mask
;
4388 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4391 finalise_ac(gfp_mask
, &ac
);
4393 /* First allocation attempt */
4394 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4399 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4400 * resp. GFP_NOIO which has to be inherited for all allocation requests
4401 * from a particular context which has been marked by
4402 * memalloc_no{fs,io}_{save,restore}.
4404 alloc_mask
= current_gfp_context(gfp_mask
);
4405 ac
.spread_dirty_pages
= false;
4408 * Restore the original nodemask if it was potentially replaced with
4409 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4411 if (unlikely(ac
.nodemask
!= nodemask
))
4412 ac
.nodemask
= nodemask
;
4414 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4417 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4418 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4419 __free_pages(page
, order
);
4423 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4427 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4430 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4431 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4432 * you need to access high mem.
4434 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4438 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4441 return (unsigned long) page_address(page
);
4443 EXPORT_SYMBOL(__get_free_pages
);
4445 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4447 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4449 EXPORT_SYMBOL(get_zeroed_page
);
4451 void __free_pages(struct page
*page
, unsigned int order
)
4453 if (put_page_testzero(page
)) {
4455 free_unref_page(page
);
4457 __free_pages_ok(page
, order
);
4461 EXPORT_SYMBOL(__free_pages
);
4463 void free_pages(unsigned long addr
, unsigned int order
)
4466 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4467 __free_pages(virt_to_page((void *)addr
), order
);
4471 EXPORT_SYMBOL(free_pages
);
4475 * An arbitrary-length arbitrary-offset area of memory which resides
4476 * within a 0 or higher order page. Multiple fragments within that page
4477 * are individually refcounted, in the page's reference counter.
4479 * The page_frag functions below provide a simple allocation framework for
4480 * page fragments. This is used by the network stack and network device
4481 * drivers to provide a backing region of memory for use as either an
4482 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4484 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4487 struct page
*page
= NULL
;
4488 gfp_t gfp
= gfp_mask
;
4490 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4491 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4493 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4494 PAGE_FRAG_CACHE_MAX_ORDER
);
4495 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4497 if (unlikely(!page
))
4498 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4500 nc
->va
= page
? page_address(page
) : NULL
;
4505 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4507 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4509 if (page_ref_sub_and_test(page
, count
)) {
4510 unsigned int order
= compound_order(page
);
4513 free_unref_page(page
);
4515 __free_pages_ok(page
, order
);
4518 EXPORT_SYMBOL(__page_frag_cache_drain
);
4520 void *page_frag_alloc(struct page_frag_cache
*nc
,
4521 unsigned int fragsz
, gfp_t gfp_mask
)
4523 unsigned int size
= PAGE_SIZE
;
4527 if (unlikely(!nc
->va
)) {
4529 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4533 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4534 /* if size can vary use size else just use PAGE_SIZE */
4537 /* Even if we own the page, we do not use atomic_set().
4538 * This would break get_page_unless_zero() users.
4540 page_ref_add(page
, size
- 1);
4542 /* reset page count bias and offset to start of new frag */
4543 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4544 nc
->pagecnt_bias
= size
;
4548 offset
= nc
->offset
- fragsz
;
4549 if (unlikely(offset
< 0)) {
4550 page
= virt_to_page(nc
->va
);
4552 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4555 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4556 /* if size can vary use size else just use PAGE_SIZE */
4559 /* OK, page count is 0, we can safely set it */
4560 set_page_count(page
, size
);
4562 /* reset page count bias and offset to start of new frag */
4563 nc
->pagecnt_bias
= size
;
4564 offset
= size
- fragsz
;
4568 nc
->offset
= offset
;
4570 return nc
->va
+ offset
;
4572 EXPORT_SYMBOL(page_frag_alloc
);
4575 * Frees a page fragment allocated out of either a compound or order 0 page.
4577 void page_frag_free(void *addr
)
4579 struct page
*page
= virt_to_head_page(addr
);
4581 if (unlikely(put_page_testzero(page
)))
4582 __free_pages_ok(page
, compound_order(page
));
4584 EXPORT_SYMBOL(page_frag_free
);
4586 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4590 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4591 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4593 split_page(virt_to_page((void *)addr
), order
);
4594 while (used
< alloc_end
) {
4599 return (void *)addr
;
4603 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4604 * @size: the number of bytes to allocate
4605 * @gfp_mask: GFP flags for the allocation
4607 * This function is similar to alloc_pages(), except that it allocates the
4608 * minimum number of pages to satisfy the request. alloc_pages() can only
4609 * allocate memory in power-of-two pages.
4611 * This function is also limited by MAX_ORDER.
4613 * Memory allocated by this function must be released by free_pages_exact().
4615 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4617 unsigned int order
= get_order(size
);
4620 addr
= __get_free_pages(gfp_mask
, order
);
4621 return make_alloc_exact(addr
, order
, size
);
4623 EXPORT_SYMBOL(alloc_pages_exact
);
4626 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4628 * @nid: the preferred node ID where memory should be allocated
4629 * @size: the number of bytes to allocate
4630 * @gfp_mask: GFP flags for the allocation
4632 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4635 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4637 unsigned int order
= get_order(size
);
4638 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4641 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4645 * free_pages_exact - release memory allocated via alloc_pages_exact()
4646 * @virt: the value returned by alloc_pages_exact.
4647 * @size: size of allocation, same value as passed to alloc_pages_exact().
4649 * Release the memory allocated by a previous call to alloc_pages_exact.
4651 void free_pages_exact(void *virt
, size_t size
)
4653 unsigned long addr
= (unsigned long)virt
;
4654 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4656 while (addr
< end
) {
4661 EXPORT_SYMBOL(free_pages_exact
);
4664 * nr_free_zone_pages - count number of pages beyond high watermark
4665 * @offset: The zone index of the highest zone
4667 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4668 * high watermark within all zones at or below a given zone index. For each
4669 * zone, the number of pages is calculated as:
4671 * nr_free_zone_pages = managed_pages - high_pages
4673 static unsigned long nr_free_zone_pages(int offset
)
4678 /* Just pick one node, since fallback list is circular */
4679 unsigned long sum
= 0;
4681 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4683 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4684 unsigned long size
= zone
->managed_pages
;
4685 unsigned long high
= high_wmark_pages(zone
);
4694 * nr_free_buffer_pages - count number of pages beyond high watermark
4696 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4697 * watermark within ZONE_DMA and ZONE_NORMAL.
4699 unsigned long nr_free_buffer_pages(void)
4701 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4703 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4706 * nr_free_pagecache_pages - count number of pages beyond high watermark
4708 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4709 * high watermark within all zones.
4711 unsigned long nr_free_pagecache_pages(void)
4713 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4716 static inline void show_node(struct zone
*zone
)
4718 if (IS_ENABLED(CONFIG_NUMA
))
4719 printk("Node %d ", zone_to_nid(zone
));
4722 long si_mem_available(void)
4725 unsigned long pagecache
;
4726 unsigned long wmark_low
= 0;
4727 unsigned long pages
[NR_LRU_LISTS
];
4728 unsigned long reclaimable
;
4732 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4733 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4736 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4739 * Estimate the amount of memory available for userspace allocations,
4740 * without causing swapping.
4742 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4745 * Not all the page cache can be freed, otherwise the system will
4746 * start swapping. Assume at least half of the page cache, or the
4747 * low watermark worth of cache, needs to stay.
4749 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4750 pagecache
-= min(pagecache
/ 2, wmark_low
);
4751 available
+= pagecache
;
4754 * Part of the reclaimable slab and other kernel memory consists of
4755 * items that are in use, and cannot be freed. Cap this estimate at the
4758 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
4759 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
4760 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
4766 EXPORT_SYMBOL_GPL(si_mem_available
);
4768 void si_meminfo(struct sysinfo
*val
)
4770 val
->totalram
= totalram_pages
;
4771 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4772 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
4773 val
->bufferram
= nr_blockdev_pages();
4774 val
->totalhigh
= totalhigh_pages
;
4775 val
->freehigh
= nr_free_highpages();
4776 val
->mem_unit
= PAGE_SIZE
;
4779 EXPORT_SYMBOL(si_meminfo
);
4782 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4784 int zone_type
; /* needs to be signed */
4785 unsigned long managed_pages
= 0;
4786 unsigned long managed_highpages
= 0;
4787 unsigned long free_highpages
= 0;
4788 pg_data_t
*pgdat
= NODE_DATA(nid
);
4790 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4791 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4792 val
->totalram
= managed_pages
;
4793 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4794 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4795 #ifdef CONFIG_HIGHMEM
4796 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4797 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4799 if (is_highmem(zone
)) {
4800 managed_highpages
+= zone
->managed_pages
;
4801 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4804 val
->totalhigh
= managed_highpages
;
4805 val
->freehigh
= free_highpages
;
4807 val
->totalhigh
= managed_highpages
;
4808 val
->freehigh
= free_highpages
;
4810 val
->mem_unit
= PAGE_SIZE
;
4815 * Determine whether the node should be displayed or not, depending on whether
4816 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4818 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4820 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4824 * no node mask - aka implicit memory numa policy. Do not bother with
4825 * the synchronization - read_mems_allowed_begin - because we do not
4826 * have to be precise here.
4829 nodemask
= &cpuset_current_mems_allowed
;
4831 return !node_isset(nid
, *nodemask
);
4834 #define K(x) ((x) << (PAGE_SHIFT-10))
4836 static void show_migration_types(unsigned char type
)
4838 static const char types
[MIGRATE_TYPES
] = {
4839 [MIGRATE_UNMOVABLE
] = 'U',
4840 [MIGRATE_MOVABLE
] = 'M',
4841 [MIGRATE_RECLAIMABLE
] = 'E',
4842 [MIGRATE_HIGHATOMIC
] = 'H',
4844 [MIGRATE_CMA
] = 'C',
4846 #ifdef CONFIG_MEMORY_ISOLATION
4847 [MIGRATE_ISOLATE
] = 'I',
4850 char tmp
[MIGRATE_TYPES
+ 1];
4854 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4855 if (type
& (1 << i
))
4860 printk(KERN_CONT
"(%s) ", tmp
);
4864 * Show free area list (used inside shift_scroll-lock stuff)
4865 * We also calculate the percentage fragmentation. We do this by counting the
4866 * memory on each free list with the exception of the first item on the list.
4869 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4872 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4874 unsigned long free_pcp
= 0;
4879 for_each_populated_zone(zone
) {
4880 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4883 for_each_online_cpu(cpu
)
4884 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4887 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4888 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4889 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4890 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4891 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4892 " free:%lu free_pcp:%lu free_cma:%lu\n",
4893 global_node_page_state(NR_ACTIVE_ANON
),
4894 global_node_page_state(NR_INACTIVE_ANON
),
4895 global_node_page_state(NR_ISOLATED_ANON
),
4896 global_node_page_state(NR_ACTIVE_FILE
),
4897 global_node_page_state(NR_INACTIVE_FILE
),
4898 global_node_page_state(NR_ISOLATED_FILE
),
4899 global_node_page_state(NR_UNEVICTABLE
),
4900 global_node_page_state(NR_FILE_DIRTY
),
4901 global_node_page_state(NR_WRITEBACK
),
4902 global_node_page_state(NR_UNSTABLE_NFS
),
4903 global_node_page_state(NR_SLAB_RECLAIMABLE
),
4904 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
4905 global_node_page_state(NR_FILE_MAPPED
),
4906 global_node_page_state(NR_SHMEM
),
4907 global_zone_page_state(NR_PAGETABLE
),
4908 global_zone_page_state(NR_BOUNCE
),
4909 global_zone_page_state(NR_FREE_PAGES
),
4911 global_zone_page_state(NR_FREE_CMA_PAGES
));
4913 for_each_online_pgdat(pgdat
) {
4914 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
4918 " active_anon:%lukB"
4919 " inactive_anon:%lukB"
4920 " active_file:%lukB"
4921 " inactive_file:%lukB"
4922 " unevictable:%lukB"
4923 " isolated(anon):%lukB"
4924 " isolated(file):%lukB"
4929 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4931 " shmem_pmdmapped: %lukB"
4934 " writeback_tmp:%lukB"
4936 " all_unreclaimable? %s"
4939 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4940 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4941 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4942 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4943 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4944 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4945 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4946 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4947 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4948 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4949 K(node_page_state(pgdat
, NR_SHMEM
)),
4950 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4951 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4952 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4954 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4956 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4957 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4958 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
4962 for_each_populated_zone(zone
) {
4965 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4969 for_each_online_cpu(cpu
)
4970 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4979 " active_anon:%lukB"
4980 " inactive_anon:%lukB"
4981 " active_file:%lukB"
4982 " inactive_file:%lukB"
4983 " unevictable:%lukB"
4984 " writepending:%lukB"
4988 " kernel_stack:%lukB"
4996 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4997 K(min_wmark_pages(zone
)),
4998 K(low_wmark_pages(zone
)),
4999 K(high_wmark_pages(zone
)),
5000 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5001 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5002 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5003 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5004 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5005 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5006 K(zone
->present_pages
),
5007 K(zone
->managed_pages
),
5008 K(zone_page_state(zone
, NR_MLOCK
)),
5009 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5010 K(zone_page_state(zone
, NR_PAGETABLE
)),
5011 K(zone_page_state(zone
, NR_BOUNCE
)),
5013 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5014 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5015 printk("lowmem_reserve[]:");
5016 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5017 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5018 printk(KERN_CONT
"\n");
5021 for_each_populated_zone(zone
) {
5023 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5024 unsigned char types
[MAX_ORDER
];
5026 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5029 printk(KERN_CONT
"%s: ", zone
->name
);
5031 spin_lock_irqsave(&zone
->lock
, flags
);
5032 for (order
= 0; order
< MAX_ORDER
; order
++) {
5033 struct free_area
*area
= &zone
->free_area
[order
];
5036 nr
[order
] = area
->nr_free
;
5037 total
+= nr
[order
] << order
;
5040 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5041 if (!list_empty(&area
->free_list
[type
]))
5042 types
[order
] |= 1 << type
;
5045 spin_unlock_irqrestore(&zone
->lock
, flags
);
5046 for (order
= 0; order
< MAX_ORDER
; order
++) {
5047 printk(KERN_CONT
"%lu*%lukB ",
5048 nr
[order
], K(1UL) << order
);
5050 show_migration_types(types
[order
]);
5052 printk(KERN_CONT
"= %lukB\n", K(total
));
5055 hugetlb_show_meminfo();
5057 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5059 show_swap_cache_info();
5062 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5064 zoneref
->zone
= zone
;
5065 zoneref
->zone_idx
= zone_idx(zone
);
5069 * Builds allocation fallback zone lists.
5071 * Add all populated zones of a node to the zonelist.
5073 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5076 enum zone_type zone_type
= MAX_NR_ZONES
;
5081 zone
= pgdat
->node_zones
+ zone_type
;
5082 if (managed_zone(zone
)) {
5083 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5084 check_highest_zone(zone_type
);
5086 } while (zone_type
);
5093 static int __parse_numa_zonelist_order(char *s
)
5096 * We used to support different zonlists modes but they turned
5097 * out to be just not useful. Let's keep the warning in place
5098 * if somebody still use the cmd line parameter so that we do
5099 * not fail it silently
5101 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5102 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5108 static __init
int setup_numa_zonelist_order(char *s
)
5113 return __parse_numa_zonelist_order(s
);
5115 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
5117 char numa_zonelist_order
[] = "Node";
5120 * sysctl handler for numa_zonelist_order
5122 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5123 void __user
*buffer
, size_t *length
,
5130 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5131 str
= memdup_user_nul(buffer
, 16);
5133 return PTR_ERR(str
);
5135 ret
= __parse_numa_zonelist_order(str
);
5141 #define MAX_NODE_LOAD (nr_online_nodes)
5142 static int node_load
[MAX_NUMNODES
];
5145 * find_next_best_node - find the next node that should appear in a given node's fallback list
5146 * @node: node whose fallback list we're appending
5147 * @used_node_mask: nodemask_t of already used nodes
5149 * We use a number of factors to determine which is the next node that should
5150 * appear on a given node's fallback list. The node should not have appeared
5151 * already in @node's fallback list, and it should be the next closest node
5152 * according to the distance array (which contains arbitrary distance values
5153 * from each node to each node in the system), and should also prefer nodes
5154 * with no CPUs, since presumably they'll have very little allocation pressure
5155 * on them otherwise.
5156 * It returns -1 if no node is found.
5158 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5161 int min_val
= INT_MAX
;
5162 int best_node
= NUMA_NO_NODE
;
5163 const struct cpumask
*tmp
= cpumask_of_node(0);
5165 /* Use the local node if we haven't already */
5166 if (!node_isset(node
, *used_node_mask
)) {
5167 node_set(node
, *used_node_mask
);
5171 for_each_node_state(n
, N_MEMORY
) {
5173 /* Don't want a node to appear more than once */
5174 if (node_isset(n
, *used_node_mask
))
5177 /* Use the distance array to find the distance */
5178 val
= node_distance(node
, n
);
5180 /* Penalize nodes under us ("prefer the next node") */
5183 /* Give preference to headless and unused nodes */
5184 tmp
= cpumask_of_node(n
);
5185 if (!cpumask_empty(tmp
))
5186 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5188 /* Slight preference for less loaded node */
5189 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5190 val
+= node_load
[n
];
5192 if (val
< min_val
) {
5199 node_set(best_node
, *used_node_mask
);
5206 * Build zonelists ordered by node and zones within node.
5207 * This results in maximum locality--normal zone overflows into local
5208 * DMA zone, if any--but risks exhausting DMA zone.
5210 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5213 struct zoneref
*zonerefs
;
5216 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5218 for (i
= 0; i
< nr_nodes
; i
++) {
5221 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5223 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5224 zonerefs
+= nr_zones
;
5226 zonerefs
->zone
= NULL
;
5227 zonerefs
->zone_idx
= 0;
5231 * Build gfp_thisnode zonelists
5233 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5235 struct zoneref
*zonerefs
;
5238 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5239 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5240 zonerefs
+= nr_zones
;
5241 zonerefs
->zone
= NULL
;
5242 zonerefs
->zone_idx
= 0;
5246 * Build zonelists ordered by zone and nodes within zones.
5247 * This results in conserving DMA zone[s] until all Normal memory is
5248 * exhausted, but results in overflowing to remote node while memory
5249 * may still exist in local DMA zone.
5252 static void build_zonelists(pg_data_t
*pgdat
)
5254 static int node_order
[MAX_NUMNODES
];
5255 int node
, load
, nr_nodes
= 0;
5256 nodemask_t used_mask
;
5257 int local_node
, prev_node
;
5259 /* NUMA-aware ordering of nodes */
5260 local_node
= pgdat
->node_id
;
5261 load
= nr_online_nodes
;
5262 prev_node
= local_node
;
5263 nodes_clear(used_mask
);
5265 memset(node_order
, 0, sizeof(node_order
));
5266 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5268 * We don't want to pressure a particular node.
5269 * So adding penalty to the first node in same
5270 * distance group to make it round-robin.
5272 if (node_distance(local_node
, node
) !=
5273 node_distance(local_node
, prev_node
))
5274 node_load
[node
] = load
;
5276 node_order
[nr_nodes
++] = node
;
5281 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5282 build_thisnode_zonelists(pgdat
);
5285 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5287 * Return node id of node used for "local" allocations.
5288 * I.e., first node id of first zone in arg node's generic zonelist.
5289 * Used for initializing percpu 'numa_mem', which is used primarily
5290 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5292 int local_memory_node(int node
)
5296 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5297 gfp_zone(GFP_KERNEL
),
5299 return zone_to_nid(z
->zone
);
5303 static void setup_min_unmapped_ratio(void);
5304 static void setup_min_slab_ratio(void);
5305 #else /* CONFIG_NUMA */
5307 static void build_zonelists(pg_data_t
*pgdat
)
5309 int node
, local_node
;
5310 struct zoneref
*zonerefs
;
5313 local_node
= pgdat
->node_id
;
5315 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5316 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5317 zonerefs
+= nr_zones
;
5320 * Now we build the zonelist so that it contains the zones
5321 * of all the other nodes.
5322 * We don't want to pressure a particular node, so when
5323 * building the zones for node N, we make sure that the
5324 * zones coming right after the local ones are those from
5325 * node N+1 (modulo N)
5327 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5328 if (!node_online(node
))
5330 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5331 zonerefs
+= nr_zones
;
5333 for (node
= 0; node
< local_node
; node
++) {
5334 if (!node_online(node
))
5336 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5337 zonerefs
+= nr_zones
;
5340 zonerefs
->zone
= NULL
;
5341 zonerefs
->zone_idx
= 0;
5344 #endif /* CONFIG_NUMA */
5347 * Boot pageset table. One per cpu which is going to be used for all
5348 * zones and all nodes. The parameters will be set in such a way
5349 * that an item put on a list will immediately be handed over to
5350 * the buddy list. This is safe since pageset manipulation is done
5351 * with interrupts disabled.
5353 * The boot_pagesets must be kept even after bootup is complete for
5354 * unused processors and/or zones. They do play a role for bootstrapping
5355 * hotplugged processors.
5357 * zoneinfo_show() and maybe other functions do
5358 * not check if the processor is online before following the pageset pointer.
5359 * Other parts of the kernel may not check if the zone is available.
5361 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5362 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5363 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5365 static void __build_all_zonelists(void *data
)
5368 int __maybe_unused cpu
;
5369 pg_data_t
*self
= data
;
5370 static DEFINE_SPINLOCK(lock
);
5375 memset(node_load
, 0, sizeof(node_load
));
5379 * This node is hotadded and no memory is yet present. So just
5380 * building zonelists is fine - no need to touch other nodes.
5382 if (self
&& !node_online(self
->node_id
)) {
5383 build_zonelists(self
);
5385 for_each_online_node(nid
) {
5386 pg_data_t
*pgdat
= NODE_DATA(nid
);
5388 build_zonelists(pgdat
);
5391 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5393 * We now know the "local memory node" for each node--
5394 * i.e., the node of the first zone in the generic zonelist.
5395 * Set up numa_mem percpu variable for on-line cpus. During
5396 * boot, only the boot cpu should be on-line; we'll init the
5397 * secondary cpus' numa_mem as they come on-line. During
5398 * node/memory hotplug, we'll fixup all on-line cpus.
5400 for_each_online_cpu(cpu
)
5401 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5408 static noinline
void __init
5409 build_all_zonelists_init(void)
5413 __build_all_zonelists(NULL
);
5416 * Initialize the boot_pagesets that are going to be used
5417 * for bootstrapping processors. The real pagesets for
5418 * each zone will be allocated later when the per cpu
5419 * allocator is available.
5421 * boot_pagesets are used also for bootstrapping offline
5422 * cpus if the system is already booted because the pagesets
5423 * are needed to initialize allocators on a specific cpu too.
5424 * F.e. the percpu allocator needs the page allocator which
5425 * needs the percpu allocator in order to allocate its pagesets
5426 * (a chicken-egg dilemma).
5428 for_each_possible_cpu(cpu
)
5429 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5431 mminit_verify_zonelist();
5432 cpuset_init_current_mems_allowed();
5436 * unless system_state == SYSTEM_BOOTING.
5438 * __ref due to call of __init annotated helper build_all_zonelists_init
5439 * [protected by SYSTEM_BOOTING].
5441 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5443 if (system_state
== SYSTEM_BOOTING
) {
5444 build_all_zonelists_init();
5446 __build_all_zonelists(pgdat
);
5447 /* cpuset refresh routine should be here */
5449 vm_total_pages
= nr_free_pagecache_pages();
5451 * Disable grouping by mobility if the number of pages in the
5452 * system is too low to allow the mechanism to work. It would be
5453 * more accurate, but expensive to check per-zone. This check is
5454 * made on memory-hotadd so a system can start with mobility
5455 * disabled and enable it later
5457 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5458 page_group_by_mobility_disabled
= 1;
5460 page_group_by_mobility_disabled
= 0;
5462 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5464 page_group_by_mobility_disabled
? "off" : "on",
5467 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5471 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5472 static bool __meminit
5473 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5475 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5476 static struct memblock_region
*r
;
5478 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5479 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5480 for_each_memblock(memory
, r
) {
5481 if (*pfn
< memblock_region_memory_end_pfn(r
))
5485 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5486 memblock_is_mirror(r
)) {
5487 *pfn
= memblock_region_memory_end_pfn(r
);
5496 * Initially all pages are reserved - free ones are freed
5497 * up by memblock_free_all() once the early boot process is
5498 * done. Non-atomic initialization, single-pass.
5500 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5501 unsigned long start_pfn
, enum memmap_context context
,
5502 struct vmem_altmap
*altmap
)
5504 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5507 if (highest_memmap_pfn
< end_pfn
- 1)
5508 highest_memmap_pfn
= end_pfn
- 1;
5510 #ifdef CONFIG_ZONE_DEVICE
5512 * Honor reservation requested by the driver for this ZONE_DEVICE
5513 * memory. We limit the total number of pages to initialize to just
5514 * those that might contain the memory mapping. We will defer the
5515 * ZONE_DEVICE page initialization until after we have released
5518 if (zone
== ZONE_DEVICE
) {
5522 if (start_pfn
== altmap
->base_pfn
)
5523 start_pfn
+= altmap
->reserve
;
5524 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5528 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5530 * There can be holes in boot-time mem_map[]s handed to this
5531 * function. They do not exist on hotplugged memory.
5533 if (context
== MEMMAP_EARLY
) {
5534 if (!early_pfn_valid(pfn
))
5536 if (!early_pfn_in_nid(pfn
, nid
))
5538 if (overlap_memmap_init(zone
, &pfn
))
5540 if (defer_init(nid
, pfn
, end_pfn
))
5544 page
= pfn_to_page(pfn
);
5545 __init_single_page(page
, pfn
, zone
, nid
);
5546 if (context
== MEMMAP_HOTPLUG
)
5547 __SetPageReserved(page
);
5550 * Mark the block movable so that blocks are reserved for
5551 * movable at startup. This will force kernel allocations
5552 * to reserve their blocks rather than leaking throughout
5553 * the address space during boot when many long-lived
5554 * kernel allocations are made.
5556 * bitmap is created for zone's valid pfn range. but memmap
5557 * can be created for invalid pages (for alignment)
5558 * check here not to call set_pageblock_migratetype() against
5561 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5562 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5568 #ifdef CONFIG_ZONE_DEVICE
5569 void __ref
memmap_init_zone_device(struct zone
*zone
,
5570 unsigned long start_pfn
,
5572 struct dev_pagemap
*pgmap
)
5574 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5575 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5576 unsigned long zone_idx
= zone_idx(zone
);
5577 unsigned long start
= jiffies
;
5578 int nid
= pgdat
->node_id
;
5580 if (WARN_ON_ONCE(!pgmap
|| !is_dev_zone(zone
)))
5584 * The call to memmap_init_zone should have already taken care
5585 * of the pages reserved for the memmap, so we can just jump to
5586 * the end of that region and start processing the device pages.
5588 if (pgmap
->altmap_valid
) {
5589 struct vmem_altmap
*altmap
= &pgmap
->altmap
;
5591 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5592 size
= end_pfn
- start_pfn
;
5595 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5596 struct page
*page
= pfn_to_page(pfn
);
5598 __init_single_page(page
, pfn
, zone_idx
, nid
);
5601 * Mark page reserved as it will need to wait for onlining
5602 * phase for it to be fully associated with a zone.
5604 * We can use the non-atomic __set_bit operation for setting
5605 * the flag as we are still initializing the pages.
5607 __SetPageReserved(page
);
5610 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5611 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5612 * page is ever freed or placed on a driver-private list.
5614 page
->pgmap
= pgmap
;
5618 * Mark the block movable so that blocks are reserved for
5619 * movable at startup. This will force kernel allocations
5620 * to reserve their blocks rather than leaking throughout
5621 * the address space during boot when many long-lived
5622 * kernel allocations are made.
5624 * bitmap is created for zone's valid pfn range. but memmap
5625 * can be created for invalid pages (for alignment)
5626 * check here not to call set_pageblock_migratetype() against
5629 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5630 * because this is done early in sparse_add_one_section
5632 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5633 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5638 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap
->dev
),
5639 size
, jiffies_to_msecs(jiffies
- start
));
5643 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5645 unsigned int order
, t
;
5646 for_each_migratetype_order(order
, t
) {
5647 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5648 zone
->free_area
[order
].nr_free
= 0;
5652 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
5653 unsigned long zone
, unsigned long start_pfn
)
5655 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
, NULL
);
5658 static int zone_batchsize(struct zone
*zone
)
5664 * The per-cpu-pages pools are set to around 1000th of the
5667 batch
= zone
->managed_pages
/ 1024;
5668 /* But no more than a meg. */
5669 if (batch
* PAGE_SIZE
> 1024 * 1024)
5670 batch
= (1024 * 1024) / PAGE_SIZE
;
5671 batch
/= 4; /* We effectively *= 4 below */
5676 * Clamp the batch to a 2^n - 1 value. Having a power
5677 * of 2 value was found to be more likely to have
5678 * suboptimal cache aliasing properties in some cases.
5680 * For example if 2 tasks are alternately allocating
5681 * batches of pages, one task can end up with a lot
5682 * of pages of one half of the possible page colors
5683 * and the other with pages of the other colors.
5685 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5690 /* The deferral and batching of frees should be suppressed under NOMMU
5693 * The problem is that NOMMU needs to be able to allocate large chunks
5694 * of contiguous memory as there's no hardware page translation to
5695 * assemble apparent contiguous memory from discontiguous pages.
5697 * Queueing large contiguous runs of pages for batching, however,
5698 * causes the pages to actually be freed in smaller chunks. As there
5699 * can be a significant delay between the individual batches being
5700 * recycled, this leads to the once large chunks of space being
5701 * fragmented and becoming unavailable for high-order allocations.
5708 * pcp->high and pcp->batch values are related and dependent on one another:
5709 * ->batch must never be higher then ->high.
5710 * The following function updates them in a safe manner without read side
5713 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5714 * those fields changing asynchronously (acording the the above rule).
5716 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5717 * outside of boot time (or some other assurance that no concurrent updaters
5720 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5721 unsigned long batch
)
5723 /* start with a fail safe value for batch */
5727 /* Update high, then batch, in order */
5734 /* a companion to pageset_set_high() */
5735 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5737 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5740 static void pageset_init(struct per_cpu_pageset
*p
)
5742 struct per_cpu_pages
*pcp
;
5745 memset(p
, 0, sizeof(*p
));
5749 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5750 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5753 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5756 pageset_set_batch(p
, batch
);
5760 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5761 * to the value high for the pageset p.
5763 static void pageset_set_high(struct per_cpu_pageset
*p
,
5766 unsigned long batch
= max(1UL, high
/ 4);
5767 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5768 batch
= PAGE_SHIFT
* 8;
5770 pageset_update(&p
->pcp
, high
, batch
);
5773 static void pageset_set_high_and_batch(struct zone
*zone
,
5774 struct per_cpu_pageset
*pcp
)
5776 if (percpu_pagelist_fraction
)
5777 pageset_set_high(pcp
,
5778 (zone
->managed_pages
/
5779 percpu_pagelist_fraction
));
5781 pageset_set_batch(pcp
, zone_batchsize(zone
));
5784 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5786 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5789 pageset_set_high_and_batch(zone
, pcp
);
5792 void __meminit
setup_zone_pageset(struct zone
*zone
)
5795 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5796 for_each_possible_cpu(cpu
)
5797 zone_pageset_init(zone
, cpu
);
5801 * Allocate per cpu pagesets and initialize them.
5802 * Before this call only boot pagesets were available.
5804 void __init
setup_per_cpu_pageset(void)
5806 struct pglist_data
*pgdat
;
5809 for_each_populated_zone(zone
)
5810 setup_zone_pageset(zone
);
5812 for_each_online_pgdat(pgdat
)
5813 pgdat
->per_cpu_nodestats
=
5814 alloc_percpu(struct per_cpu_nodestat
);
5817 static __meminit
void zone_pcp_init(struct zone
*zone
)
5820 * per cpu subsystem is not up at this point. The following code
5821 * relies on the ability of the linker to provide the
5822 * offset of a (static) per cpu variable into the per cpu area.
5824 zone
->pageset
= &boot_pageset
;
5826 if (populated_zone(zone
))
5827 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5828 zone
->name
, zone
->present_pages
,
5829 zone_batchsize(zone
));
5832 void __meminit
init_currently_empty_zone(struct zone
*zone
,
5833 unsigned long zone_start_pfn
,
5836 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5837 int zone_idx
= zone_idx(zone
) + 1;
5839 if (zone_idx
> pgdat
->nr_zones
)
5840 pgdat
->nr_zones
= zone_idx
;
5842 zone
->zone_start_pfn
= zone_start_pfn
;
5844 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5845 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5847 (unsigned long)zone_idx(zone
),
5848 zone_start_pfn
, (zone_start_pfn
+ size
));
5850 zone_init_free_lists(zone
);
5851 zone
->initialized
= 1;
5854 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5855 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5858 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5860 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5861 struct mminit_pfnnid_cache
*state
)
5863 unsigned long start_pfn
, end_pfn
;
5866 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5867 return state
->last_nid
;
5869 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5871 state
->last_start
= start_pfn
;
5872 state
->last_end
= end_pfn
;
5873 state
->last_nid
= nid
;
5878 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5881 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5882 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5883 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5885 * If an architecture guarantees that all ranges registered contain no holes
5886 * and may be freed, this this function may be used instead of calling
5887 * memblock_free_early_nid() manually.
5889 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5891 unsigned long start_pfn
, end_pfn
;
5894 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5895 start_pfn
= min(start_pfn
, max_low_pfn
);
5896 end_pfn
= min(end_pfn
, max_low_pfn
);
5898 if (start_pfn
< end_pfn
)
5899 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5900 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5906 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5907 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5909 * If an architecture guarantees that all ranges registered contain no holes and may
5910 * be freed, this function may be used instead of calling memory_present() manually.
5912 void __init
sparse_memory_present_with_active_regions(int nid
)
5914 unsigned long start_pfn
, end_pfn
;
5917 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5918 memory_present(this_nid
, start_pfn
, end_pfn
);
5922 * get_pfn_range_for_nid - Return the start and end page frames for a node
5923 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5924 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5925 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5927 * It returns the start and end page frame of a node based on information
5928 * provided by memblock_set_node(). If called for a node
5929 * with no available memory, a warning is printed and the start and end
5932 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5933 unsigned long *start_pfn
, unsigned long *end_pfn
)
5935 unsigned long this_start_pfn
, this_end_pfn
;
5941 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5942 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5943 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5946 if (*start_pfn
== -1UL)
5951 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5952 * assumption is made that zones within a node are ordered in monotonic
5953 * increasing memory addresses so that the "highest" populated zone is used
5955 static void __init
find_usable_zone_for_movable(void)
5958 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5959 if (zone_index
== ZONE_MOVABLE
)
5962 if (arch_zone_highest_possible_pfn
[zone_index
] >
5963 arch_zone_lowest_possible_pfn
[zone_index
])
5967 VM_BUG_ON(zone_index
== -1);
5968 movable_zone
= zone_index
;
5972 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5973 * because it is sized independent of architecture. Unlike the other zones,
5974 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5975 * in each node depending on the size of each node and how evenly kernelcore
5976 * is distributed. This helper function adjusts the zone ranges
5977 * provided by the architecture for a given node by using the end of the
5978 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5979 * zones within a node are in order of monotonic increases memory addresses
5981 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5982 unsigned long zone_type
,
5983 unsigned long node_start_pfn
,
5984 unsigned long node_end_pfn
,
5985 unsigned long *zone_start_pfn
,
5986 unsigned long *zone_end_pfn
)
5988 /* Only adjust if ZONE_MOVABLE is on this node */
5989 if (zone_movable_pfn
[nid
]) {
5990 /* Size ZONE_MOVABLE */
5991 if (zone_type
== ZONE_MOVABLE
) {
5992 *zone_start_pfn
= zone_movable_pfn
[nid
];
5993 *zone_end_pfn
= min(node_end_pfn
,
5994 arch_zone_highest_possible_pfn
[movable_zone
]);
5996 /* Adjust for ZONE_MOVABLE starting within this range */
5997 } else if (!mirrored_kernelcore
&&
5998 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5999 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6000 *zone_end_pfn
= zone_movable_pfn
[nid
];
6002 /* Check if this whole range is within ZONE_MOVABLE */
6003 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6004 *zone_start_pfn
= *zone_end_pfn
;
6009 * Return the number of pages a zone spans in a node, including holes
6010 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6012 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
6013 unsigned long zone_type
,
6014 unsigned long node_start_pfn
,
6015 unsigned long node_end_pfn
,
6016 unsigned long *zone_start_pfn
,
6017 unsigned long *zone_end_pfn
,
6018 unsigned long *ignored
)
6020 /* When hotadd a new node from cpu_up(), the node should be empty */
6021 if (!node_start_pfn
&& !node_end_pfn
)
6024 /* Get the start and end of the zone */
6025 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
6026 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
6027 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6028 node_start_pfn
, node_end_pfn
,
6029 zone_start_pfn
, zone_end_pfn
);
6031 /* Check that this node has pages within the zone's required range */
6032 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6035 /* Move the zone boundaries inside the node if necessary */
6036 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6037 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6039 /* Return the spanned pages */
6040 return *zone_end_pfn
- *zone_start_pfn
;
6044 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6045 * then all holes in the requested range will be accounted for.
6047 unsigned long __meminit
__absent_pages_in_range(int nid
,
6048 unsigned long range_start_pfn
,
6049 unsigned long range_end_pfn
)
6051 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6052 unsigned long start_pfn
, end_pfn
;
6055 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6056 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6057 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6058 nr_absent
-= end_pfn
- start_pfn
;
6064 * absent_pages_in_range - Return number of page frames in holes within a range
6065 * @start_pfn: The start PFN to start searching for holes
6066 * @end_pfn: The end PFN to stop searching for holes
6068 * It returns the number of pages frames in memory holes within a range.
6070 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6071 unsigned long end_pfn
)
6073 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6076 /* Return the number of page frames in holes in a zone on a node */
6077 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
6078 unsigned long zone_type
,
6079 unsigned long node_start_pfn
,
6080 unsigned long node_end_pfn
,
6081 unsigned long *ignored
)
6083 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6084 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6085 unsigned long zone_start_pfn
, zone_end_pfn
;
6086 unsigned long nr_absent
;
6088 /* When hotadd a new node from cpu_up(), the node should be empty */
6089 if (!node_start_pfn
&& !node_end_pfn
)
6092 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6093 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6095 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6096 node_start_pfn
, node_end_pfn
,
6097 &zone_start_pfn
, &zone_end_pfn
);
6098 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6101 * ZONE_MOVABLE handling.
6102 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6105 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6106 unsigned long start_pfn
, end_pfn
;
6107 struct memblock_region
*r
;
6109 for_each_memblock(memory
, r
) {
6110 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6111 zone_start_pfn
, zone_end_pfn
);
6112 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6113 zone_start_pfn
, zone_end_pfn
);
6115 if (zone_type
== ZONE_MOVABLE
&&
6116 memblock_is_mirror(r
))
6117 nr_absent
+= end_pfn
- start_pfn
;
6119 if (zone_type
== ZONE_NORMAL
&&
6120 !memblock_is_mirror(r
))
6121 nr_absent
+= end_pfn
- start_pfn
;
6128 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6129 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
6130 unsigned long zone_type
,
6131 unsigned long node_start_pfn
,
6132 unsigned long node_end_pfn
,
6133 unsigned long *zone_start_pfn
,
6134 unsigned long *zone_end_pfn
,
6135 unsigned long *zones_size
)
6139 *zone_start_pfn
= node_start_pfn
;
6140 for (zone
= 0; zone
< zone_type
; zone
++)
6141 *zone_start_pfn
+= zones_size
[zone
];
6143 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
6145 return zones_size
[zone_type
];
6148 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
6149 unsigned long zone_type
,
6150 unsigned long node_start_pfn
,
6151 unsigned long node_end_pfn
,
6152 unsigned long *zholes_size
)
6157 return zholes_size
[zone_type
];
6160 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6162 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
6163 unsigned long node_start_pfn
,
6164 unsigned long node_end_pfn
,
6165 unsigned long *zones_size
,
6166 unsigned long *zholes_size
)
6168 unsigned long realtotalpages
= 0, totalpages
= 0;
6171 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6172 struct zone
*zone
= pgdat
->node_zones
+ i
;
6173 unsigned long zone_start_pfn
, zone_end_pfn
;
6174 unsigned long size
, real_size
;
6176 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6182 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
6183 node_start_pfn
, node_end_pfn
,
6186 zone
->zone_start_pfn
= zone_start_pfn
;
6188 zone
->zone_start_pfn
= 0;
6189 zone
->spanned_pages
= size
;
6190 zone
->present_pages
= real_size
;
6193 realtotalpages
+= real_size
;
6196 pgdat
->node_spanned_pages
= totalpages
;
6197 pgdat
->node_present_pages
= realtotalpages
;
6198 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6202 #ifndef CONFIG_SPARSEMEM
6204 * Calculate the size of the zone->blockflags rounded to an unsigned long
6205 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6206 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6207 * round what is now in bits to nearest long in bits, then return it in
6210 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6212 unsigned long usemapsize
;
6214 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6215 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6216 usemapsize
= usemapsize
>> pageblock_order
;
6217 usemapsize
*= NR_PAGEBLOCK_BITS
;
6218 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6220 return usemapsize
/ 8;
6223 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6225 unsigned long zone_start_pfn
,
6226 unsigned long zonesize
)
6228 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6229 zone
->pageblock_flags
= NULL
;
6231 zone
->pageblock_flags
=
6232 memblock_alloc_node_nopanic(usemapsize
,
6236 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6237 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6238 #endif /* CONFIG_SPARSEMEM */
6240 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6242 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6243 void __init
set_pageblock_order(void)
6247 /* Check that pageblock_nr_pages has not already been setup */
6248 if (pageblock_order
)
6251 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6252 order
= HUGETLB_PAGE_ORDER
;
6254 order
= MAX_ORDER
- 1;
6257 * Assume the largest contiguous order of interest is a huge page.
6258 * This value may be variable depending on boot parameters on IA64 and
6261 pageblock_order
= order
;
6263 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6266 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6267 * is unused as pageblock_order is set at compile-time. See
6268 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6271 void __init
set_pageblock_order(void)
6275 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6277 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6278 unsigned long present_pages
)
6280 unsigned long pages
= spanned_pages
;
6283 * Provide a more accurate estimation if there are holes within
6284 * the zone and SPARSEMEM is in use. If there are holes within the
6285 * zone, each populated memory region may cost us one or two extra
6286 * memmap pages due to alignment because memmap pages for each
6287 * populated regions may not be naturally aligned on page boundary.
6288 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6290 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6291 IS_ENABLED(CONFIG_SPARSEMEM
))
6292 pages
= present_pages
;
6294 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6297 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6298 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6300 spin_lock_init(&pgdat
->split_queue_lock
);
6301 INIT_LIST_HEAD(&pgdat
->split_queue
);
6302 pgdat
->split_queue_len
= 0;
6305 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6308 #ifdef CONFIG_COMPACTION
6309 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6311 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6314 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6317 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6319 pgdat_resize_init(pgdat
);
6321 pgdat_init_split_queue(pgdat
);
6322 pgdat_init_kcompactd(pgdat
);
6324 init_waitqueue_head(&pgdat
->kswapd_wait
);
6325 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6327 pgdat_page_ext_init(pgdat
);
6328 spin_lock_init(&pgdat
->lru_lock
);
6329 lruvec_init(node_lruvec(pgdat
));
6332 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6333 unsigned long remaining_pages
)
6335 zone
->managed_pages
= remaining_pages
;
6336 zone_set_nid(zone
, nid
);
6337 zone
->name
= zone_names
[idx
];
6338 zone
->zone_pgdat
= NODE_DATA(nid
);
6339 spin_lock_init(&zone
->lock
);
6340 zone_seqlock_init(zone
);
6341 zone_pcp_init(zone
);
6345 * Set up the zone data structures
6346 * - init pgdat internals
6347 * - init all zones belonging to this node
6349 * NOTE: this function is only called during memory hotplug
6351 #ifdef CONFIG_MEMORY_HOTPLUG
6352 void __ref
free_area_init_core_hotplug(int nid
)
6355 pg_data_t
*pgdat
= NODE_DATA(nid
);
6357 pgdat_init_internals(pgdat
);
6358 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6359 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6364 * Set up the zone data structures:
6365 * - mark all pages reserved
6366 * - mark all memory queues empty
6367 * - clear the memory bitmaps
6369 * NOTE: pgdat should get zeroed by caller.
6370 * NOTE: this function is only called during early init.
6372 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6375 int nid
= pgdat
->node_id
;
6377 pgdat_init_internals(pgdat
);
6378 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6380 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6381 struct zone
*zone
= pgdat
->node_zones
+ j
;
6382 unsigned long size
, freesize
, memmap_pages
;
6383 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6385 size
= zone
->spanned_pages
;
6386 freesize
= zone
->present_pages
;
6389 * Adjust freesize so that it accounts for how much memory
6390 * is used by this zone for memmap. This affects the watermark
6391 * and per-cpu initialisations
6393 memmap_pages
= calc_memmap_size(size
, freesize
);
6394 if (!is_highmem_idx(j
)) {
6395 if (freesize
>= memmap_pages
) {
6396 freesize
-= memmap_pages
;
6399 " %s zone: %lu pages used for memmap\n",
6400 zone_names
[j
], memmap_pages
);
6402 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6403 zone_names
[j
], memmap_pages
, freesize
);
6406 /* Account for reserved pages */
6407 if (j
== 0 && freesize
> dma_reserve
) {
6408 freesize
-= dma_reserve
;
6409 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6410 zone_names
[0], dma_reserve
);
6413 if (!is_highmem_idx(j
))
6414 nr_kernel_pages
+= freesize
;
6415 /* Charge for highmem memmap if there are enough kernel pages */
6416 else if (nr_kernel_pages
> memmap_pages
* 2)
6417 nr_kernel_pages
-= memmap_pages
;
6418 nr_all_pages
+= freesize
;
6421 * Set an approximate value for lowmem here, it will be adjusted
6422 * when the bootmem allocator frees pages into the buddy system.
6423 * And all highmem pages will be managed by the buddy system.
6425 zone_init_internals(zone
, j
, nid
, freesize
);
6430 set_pageblock_order();
6431 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6432 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6433 memmap_init(size
, nid
, j
, zone_start_pfn
);
6437 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6438 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6440 unsigned long __maybe_unused start
= 0;
6441 unsigned long __maybe_unused offset
= 0;
6443 /* Skip empty nodes */
6444 if (!pgdat
->node_spanned_pages
)
6447 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6448 offset
= pgdat
->node_start_pfn
- start
;
6449 /* ia64 gets its own node_mem_map, before this, without bootmem */
6450 if (!pgdat
->node_mem_map
) {
6451 unsigned long size
, end
;
6455 * The zone's endpoints aren't required to be MAX_ORDER
6456 * aligned but the node_mem_map endpoints must be in order
6457 * for the buddy allocator to function correctly.
6459 end
= pgdat_end_pfn(pgdat
);
6460 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6461 size
= (end
- start
) * sizeof(struct page
);
6462 map
= memblock_alloc_node_nopanic(size
, pgdat
->node_id
);
6463 pgdat
->node_mem_map
= map
+ offset
;
6465 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6466 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6467 (unsigned long)pgdat
->node_mem_map
);
6468 #ifndef CONFIG_NEED_MULTIPLE_NODES
6470 * With no DISCONTIG, the global mem_map is just set as node 0's
6472 if (pgdat
== NODE_DATA(0)) {
6473 mem_map
= NODE_DATA(0)->node_mem_map
;
6474 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6475 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6477 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6482 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6483 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6485 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6486 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6489 * We start only with one section of pages, more pages are added as
6490 * needed until the rest of deferred pages are initialized.
6492 pgdat
->static_init_pgcnt
= min_t(unsigned long, PAGES_PER_SECTION
,
6493 pgdat
->node_spanned_pages
);
6494 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6497 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6500 void __init
free_area_init_node(int nid
, unsigned long *zones_size
,
6501 unsigned long node_start_pfn
,
6502 unsigned long *zholes_size
)
6504 pg_data_t
*pgdat
= NODE_DATA(nid
);
6505 unsigned long start_pfn
= 0;
6506 unsigned long end_pfn
= 0;
6508 /* pg_data_t should be reset to zero when it's allocated */
6509 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6511 pgdat
->node_id
= nid
;
6512 pgdat
->node_start_pfn
= node_start_pfn
;
6513 pgdat
->per_cpu_nodestats
= NULL
;
6514 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6515 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6516 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6517 (u64
)start_pfn
<< PAGE_SHIFT
,
6518 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6520 start_pfn
= node_start_pfn
;
6522 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6523 zones_size
, zholes_size
);
6525 alloc_node_mem_map(pgdat
);
6526 pgdat_set_deferred_range(pgdat
);
6528 free_area_init_core(pgdat
);
6531 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6533 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6536 static u64
zero_pfn_range(unsigned long spfn
, unsigned long epfn
)
6541 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6542 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6543 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6544 + pageblock_nr_pages
- 1;
6547 mm_zero_struct_page(pfn_to_page(pfn
));
6555 * Only struct pages that are backed by physical memory are zeroed and
6556 * initialized by going through __init_single_page(). But, there are some
6557 * struct pages which are reserved in memblock allocator and their fields
6558 * may be accessed (for example page_to_pfn() on some configuration accesses
6559 * flags). We must explicitly zero those struct pages.
6561 * This function also addresses a similar issue where struct pages are left
6562 * uninitialized because the physical address range is not covered by
6563 * memblock.memory or memblock.reserved. That could happen when memblock
6564 * layout is manually configured via memmap=.
6566 void __init
zero_resv_unavail(void)
6568 phys_addr_t start
, end
;
6570 phys_addr_t next
= 0;
6573 * Loop through unavailable ranges not covered by memblock.memory.
6576 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6577 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6579 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), PFN_UP(start
));
6582 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), max_pfn
);
6585 * Struct pages that do not have backing memory. This could be because
6586 * firmware is using some of this memory, or for some other reasons.
6589 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
6591 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6593 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6595 #if MAX_NUMNODES > 1
6597 * Figure out the number of possible node ids.
6599 void __init
setup_nr_node_ids(void)
6601 unsigned int highest
;
6603 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6604 nr_node_ids
= highest
+ 1;
6609 * node_map_pfn_alignment - determine the maximum internode alignment
6611 * This function should be called after node map is populated and sorted.
6612 * It calculates the maximum power of two alignment which can distinguish
6615 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6616 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6617 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6618 * shifted, 1GiB is enough and this function will indicate so.
6620 * This is used to test whether pfn -> nid mapping of the chosen memory
6621 * model has fine enough granularity to avoid incorrect mapping for the
6622 * populated node map.
6624 * Returns the determined alignment in pfn's. 0 if there is no alignment
6625 * requirement (single node).
6627 unsigned long __init
node_map_pfn_alignment(void)
6629 unsigned long accl_mask
= 0, last_end
= 0;
6630 unsigned long start
, end
, mask
;
6634 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6635 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6642 * Start with a mask granular enough to pin-point to the
6643 * start pfn and tick off bits one-by-one until it becomes
6644 * too coarse to separate the current node from the last.
6646 mask
= ~((1 << __ffs(start
)) - 1);
6647 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6650 /* accumulate all internode masks */
6654 /* convert mask to number of pages */
6655 return ~accl_mask
+ 1;
6658 /* Find the lowest pfn for a node */
6659 static unsigned long __init
find_min_pfn_for_node(int nid
)
6661 unsigned long min_pfn
= ULONG_MAX
;
6662 unsigned long start_pfn
;
6665 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6666 min_pfn
= min(min_pfn
, start_pfn
);
6668 if (min_pfn
== ULONG_MAX
) {
6669 pr_warn("Could not find start_pfn for node %d\n", nid
);
6677 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6679 * It returns the minimum PFN based on information provided via
6680 * memblock_set_node().
6682 unsigned long __init
find_min_pfn_with_active_regions(void)
6684 return find_min_pfn_for_node(MAX_NUMNODES
);
6688 * early_calculate_totalpages()
6689 * Sum pages in active regions for movable zone.
6690 * Populate N_MEMORY for calculating usable_nodes.
6692 static unsigned long __init
early_calculate_totalpages(void)
6694 unsigned long totalpages
= 0;
6695 unsigned long start_pfn
, end_pfn
;
6698 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6699 unsigned long pages
= end_pfn
- start_pfn
;
6701 totalpages
+= pages
;
6703 node_set_state(nid
, N_MEMORY
);
6709 * Find the PFN the Movable zone begins in each node. Kernel memory
6710 * is spread evenly between nodes as long as the nodes have enough
6711 * memory. When they don't, some nodes will have more kernelcore than
6714 static void __init
find_zone_movable_pfns_for_nodes(void)
6717 unsigned long usable_startpfn
;
6718 unsigned long kernelcore_node
, kernelcore_remaining
;
6719 /* save the state before borrow the nodemask */
6720 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6721 unsigned long totalpages
= early_calculate_totalpages();
6722 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6723 struct memblock_region
*r
;
6725 /* Need to find movable_zone earlier when movable_node is specified. */
6726 find_usable_zone_for_movable();
6729 * If movable_node is specified, ignore kernelcore and movablecore
6732 if (movable_node_is_enabled()) {
6733 for_each_memblock(memory
, r
) {
6734 if (!memblock_is_hotpluggable(r
))
6739 usable_startpfn
= PFN_DOWN(r
->base
);
6740 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6741 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6749 * If kernelcore=mirror is specified, ignore movablecore option
6751 if (mirrored_kernelcore
) {
6752 bool mem_below_4gb_not_mirrored
= false;
6754 for_each_memblock(memory
, r
) {
6755 if (memblock_is_mirror(r
))
6760 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6762 if (usable_startpfn
< 0x100000) {
6763 mem_below_4gb_not_mirrored
= true;
6767 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6768 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6772 if (mem_below_4gb_not_mirrored
)
6773 pr_warn("This configuration results in unmirrored kernel memory.");
6779 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6780 * amount of necessary memory.
6782 if (required_kernelcore_percent
)
6783 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
6785 if (required_movablecore_percent
)
6786 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
6790 * If movablecore= was specified, calculate what size of
6791 * kernelcore that corresponds so that memory usable for
6792 * any allocation type is evenly spread. If both kernelcore
6793 * and movablecore are specified, then the value of kernelcore
6794 * will be used for required_kernelcore if it's greater than
6795 * what movablecore would have allowed.
6797 if (required_movablecore
) {
6798 unsigned long corepages
;
6801 * Round-up so that ZONE_MOVABLE is at least as large as what
6802 * was requested by the user
6804 required_movablecore
=
6805 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6806 required_movablecore
= min(totalpages
, required_movablecore
);
6807 corepages
= totalpages
- required_movablecore
;
6809 required_kernelcore
= max(required_kernelcore
, corepages
);
6813 * If kernelcore was not specified or kernelcore size is larger
6814 * than totalpages, there is no ZONE_MOVABLE.
6816 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6819 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6820 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6823 /* Spread kernelcore memory as evenly as possible throughout nodes */
6824 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6825 for_each_node_state(nid
, N_MEMORY
) {
6826 unsigned long start_pfn
, end_pfn
;
6829 * Recalculate kernelcore_node if the division per node
6830 * now exceeds what is necessary to satisfy the requested
6831 * amount of memory for the kernel
6833 if (required_kernelcore
< kernelcore_node
)
6834 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6837 * As the map is walked, we track how much memory is usable
6838 * by the kernel using kernelcore_remaining. When it is
6839 * 0, the rest of the node is usable by ZONE_MOVABLE
6841 kernelcore_remaining
= kernelcore_node
;
6843 /* Go through each range of PFNs within this node */
6844 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6845 unsigned long size_pages
;
6847 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6848 if (start_pfn
>= end_pfn
)
6851 /* Account for what is only usable for kernelcore */
6852 if (start_pfn
< usable_startpfn
) {
6853 unsigned long kernel_pages
;
6854 kernel_pages
= min(end_pfn
, usable_startpfn
)
6857 kernelcore_remaining
-= min(kernel_pages
,
6858 kernelcore_remaining
);
6859 required_kernelcore
-= min(kernel_pages
,
6860 required_kernelcore
);
6862 /* Continue if range is now fully accounted */
6863 if (end_pfn
<= usable_startpfn
) {
6866 * Push zone_movable_pfn to the end so
6867 * that if we have to rebalance
6868 * kernelcore across nodes, we will
6869 * not double account here
6871 zone_movable_pfn
[nid
] = end_pfn
;
6874 start_pfn
= usable_startpfn
;
6878 * The usable PFN range for ZONE_MOVABLE is from
6879 * start_pfn->end_pfn. Calculate size_pages as the
6880 * number of pages used as kernelcore
6882 size_pages
= end_pfn
- start_pfn
;
6883 if (size_pages
> kernelcore_remaining
)
6884 size_pages
= kernelcore_remaining
;
6885 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6888 * Some kernelcore has been met, update counts and
6889 * break if the kernelcore for this node has been
6892 required_kernelcore
-= min(required_kernelcore
,
6894 kernelcore_remaining
-= size_pages
;
6895 if (!kernelcore_remaining
)
6901 * If there is still required_kernelcore, we do another pass with one
6902 * less node in the count. This will push zone_movable_pfn[nid] further
6903 * along on the nodes that still have memory until kernelcore is
6907 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6911 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6912 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6913 zone_movable_pfn
[nid
] =
6914 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6917 /* restore the node_state */
6918 node_states
[N_MEMORY
] = saved_node_state
;
6921 /* Any regular or high memory on that node ? */
6922 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6924 enum zone_type zone_type
;
6926 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6927 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6928 if (populated_zone(zone
)) {
6929 if (IS_ENABLED(CONFIG_HIGHMEM
))
6930 node_set_state(nid
, N_HIGH_MEMORY
);
6931 if (zone_type
<= ZONE_NORMAL
)
6932 node_set_state(nid
, N_NORMAL_MEMORY
);
6939 * free_area_init_nodes - Initialise all pg_data_t and zone data
6940 * @max_zone_pfn: an array of max PFNs for each zone
6942 * This will call free_area_init_node() for each active node in the system.
6943 * Using the page ranges provided by memblock_set_node(), the size of each
6944 * zone in each node and their holes is calculated. If the maximum PFN
6945 * between two adjacent zones match, it is assumed that the zone is empty.
6946 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6947 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6948 * starts where the previous one ended. For example, ZONE_DMA32 starts
6949 * at arch_max_dma_pfn.
6951 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6953 unsigned long start_pfn
, end_pfn
;
6956 /* Record where the zone boundaries are */
6957 memset(arch_zone_lowest_possible_pfn
, 0,
6958 sizeof(arch_zone_lowest_possible_pfn
));
6959 memset(arch_zone_highest_possible_pfn
, 0,
6960 sizeof(arch_zone_highest_possible_pfn
));
6962 start_pfn
= find_min_pfn_with_active_regions();
6964 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6965 if (i
== ZONE_MOVABLE
)
6968 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6969 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6970 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6972 start_pfn
= end_pfn
;
6975 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6976 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6977 find_zone_movable_pfns_for_nodes();
6979 /* Print out the zone ranges */
6980 pr_info("Zone ranges:\n");
6981 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6982 if (i
== ZONE_MOVABLE
)
6984 pr_info(" %-8s ", zone_names
[i
]);
6985 if (arch_zone_lowest_possible_pfn
[i
] ==
6986 arch_zone_highest_possible_pfn
[i
])
6989 pr_cont("[mem %#018Lx-%#018Lx]\n",
6990 (u64
)arch_zone_lowest_possible_pfn
[i
]
6992 ((u64
)arch_zone_highest_possible_pfn
[i
]
6993 << PAGE_SHIFT
) - 1);
6996 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6997 pr_info("Movable zone start for each node\n");
6998 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6999 if (zone_movable_pfn
[i
])
7000 pr_info(" Node %d: %#018Lx\n", i
,
7001 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7004 /* Print out the early node map */
7005 pr_info("Early memory node ranges\n");
7006 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
7007 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7008 (u64
)start_pfn
<< PAGE_SHIFT
,
7009 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7011 /* Initialise every node */
7012 mminit_verify_pageflags_layout();
7013 setup_nr_node_ids();
7014 zero_resv_unavail();
7015 for_each_online_node(nid
) {
7016 pg_data_t
*pgdat
= NODE_DATA(nid
);
7017 free_area_init_node(nid
, NULL
,
7018 find_min_pfn_for_node(nid
), NULL
);
7020 /* Any memory on that node */
7021 if (pgdat
->node_present_pages
)
7022 node_set_state(nid
, N_MEMORY
);
7023 check_for_memory(pgdat
, nid
);
7027 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7028 unsigned long *percent
)
7030 unsigned long long coremem
;
7036 /* Value may be a percentage of total memory, otherwise bytes */
7037 coremem
= simple_strtoull(p
, &endptr
, 0);
7038 if (*endptr
== '%') {
7039 /* Paranoid check for percent values greater than 100 */
7040 WARN_ON(coremem
> 100);
7044 coremem
= memparse(p
, &p
);
7045 /* Paranoid check that UL is enough for the coremem value */
7046 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7048 *core
= coremem
>> PAGE_SHIFT
;
7055 * kernelcore=size sets the amount of memory for use for allocations that
7056 * cannot be reclaimed or migrated.
7058 static int __init
cmdline_parse_kernelcore(char *p
)
7060 /* parse kernelcore=mirror */
7061 if (parse_option_str(p
, "mirror")) {
7062 mirrored_kernelcore
= true;
7066 return cmdline_parse_core(p
, &required_kernelcore
,
7067 &required_kernelcore_percent
);
7071 * movablecore=size sets the amount of memory for use for allocations that
7072 * can be reclaimed or migrated.
7074 static int __init
cmdline_parse_movablecore(char *p
)
7076 return cmdline_parse_core(p
, &required_movablecore
,
7077 &required_movablecore_percent
);
7080 early_param("kernelcore", cmdline_parse_kernelcore
);
7081 early_param("movablecore", cmdline_parse_movablecore
);
7083 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7085 void adjust_managed_page_count(struct page
*page
, long count
)
7087 spin_lock(&managed_page_count_lock
);
7088 page_zone(page
)->managed_pages
+= count
;
7089 totalram_pages
+= count
;
7090 #ifdef CONFIG_HIGHMEM
7091 if (PageHighMem(page
))
7092 totalhigh_pages
+= count
;
7094 spin_unlock(&managed_page_count_lock
);
7096 EXPORT_SYMBOL(adjust_managed_page_count
);
7098 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
7101 unsigned long pages
= 0;
7103 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7104 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7105 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7106 struct page
*page
= virt_to_page(pos
);
7107 void *direct_map_addr
;
7110 * 'direct_map_addr' might be different from 'pos'
7111 * because some architectures' virt_to_page()
7112 * work with aliases. Getting the direct map
7113 * address ensures that we get a _writeable_
7114 * alias for the memset().
7116 direct_map_addr
= page_address(page
);
7117 if ((unsigned int)poison
<= 0xFF)
7118 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7120 free_reserved_page(page
);
7124 pr_info("Freeing %s memory: %ldK\n",
7125 s
, pages
<< (PAGE_SHIFT
- 10));
7129 EXPORT_SYMBOL(free_reserved_area
);
7131 #ifdef CONFIG_HIGHMEM
7132 void free_highmem_page(struct page
*page
)
7134 __free_reserved_page(page
);
7136 page_zone(page
)->managed_pages
++;
7142 void __init
mem_init_print_info(const char *str
)
7144 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7145 unsigned long init_code_size
, init_data_size
;
7147 physpages
= get_num_physpages();
7148 codesize
= _etext
- _stext
;
7149 datasize
= _edata
- _sdata
;
7150 rosize
= __end_rodata
- __start_rodata
;
7151 bss_size
= __bss_stop
- __bss_start
;
7152 init_data_size
= __init_end
- __init_begin
;
7153 init_code_size
= _einittext
- _sinittext
;
7156 * Detect special cases and adjust section sizes accordingly:
7157 * 1) .init.* may be embedded into .data sections
7158 * 2) .init.text.* may be out of [__init_begin, __init_end],
7159 * please refer to arch/tile/kernel/vmlinux.lds.S.
7160 * 3) .rodata.* may be embedded into .text or .data sections.
7162 #define adj_init_size(start, end, size, pos, adj) \
7164 if (start <= pos && pos < end && size > adj) \
7168 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7169 _sinittext
, init_code_size
);
7170 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7171 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7172 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7173 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7175 #undef adj_init_size
7177 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7178 #ifdef CONFIG_HIGHMEM
7182 nr_free_pages() << (PAGE_SHIFT
- 10),
7183 physpages
<< (PAGE_SHIFT
- 10),
7184 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7185 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7186 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
7187 totalcma_pages
<< (PAGE_SHIFT
- 10),
7188 #ifdef CONFIG_HIGHMEM
7189 totalhigh_pages
<< (PAGE_SHIFT
- 10),
7191 str
? ", " : "", str
? str
: "");
7195 * set_dma_reserve - set the specified number of pages reserved in the first zone
7196 * @new_dma_reserve: The number of pages to mark reserved
7198 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7199 * In the DMA zone, a significant percentage may be consumed by kernel image
7200 * and other unfreeable allocations which can skew the watermarks badly. This
7201 * function may optionally be used to account for unfreeable pages in the
7202 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7203 * smaller per-cpu batchsize.
7205 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7207 dma_reserve
= new_dma_reserve
;
7210 void __init
free_area_init(unsigned long *zones_size
)
7212 zero_resv_unavail();
7213 free_area_init_node(0, zones_size
,
7214 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
7217 static int page_alloc_cpu_dead(unsigned int cpu
)
7220 lru_add_drain_cpu(cpu
);
7224 * Spill the event counters of the dead processor
7225 * into the current processors event counters.
7226 * This artificially elevates the count of the current
7229 vm_events_fold_cpu(cpu
);
7232 * Zero the differential counters of the dead processor
7233 * so that the vm statistics are consistent.
7235 * This is only okay since the processor is dead and cannot
7236 * race with what we are doing.
7238 cpu_vm_stats_fold(cpu
);
7242 void __init
page_alloc_init(void)
7246 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7247 "mm/page_alloc:dead", NULL
,
7248 page_alloc_cpu_dead
);
7253 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7254 * or min_free_kbytes changes.
7256 static void calculate_totalreserve_pages(void)
7258 struct pglist_data
*pgdat
;
7259 unsigned long reserve_pages
= 0;
7260 enum zone_type i
, j
;
7262 for_each_online_pgdat(pgdat
) {
7264 pgdat
->totalreserve_pages
= 0;
7266 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7267 struct zone
*zone
= pgdat
->node_zones
+ i
;
7270 /* Find valid and maximum lowmem_reserve in the zone */
7271 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7272 if (zone
->lowmem_reserve
[j
] > max
)
7273 max
= zone
->lowmem_reserve
[j
];
7276 /* we treat the high watermark as reserved pages. */
7277 max
+= high_wmark_pages(zone
);
7279 if (max
> zone
->managed_pages
)
7280 max
= zone
->managed_pages
;
7282 pgdat
->totalreserve_pages
+= max
;
7284 reserve_pages
+= max
;
7287 totalreserve_pages
= reserve_pages
;
7291 * setup_per_zone_lowmem_reserve - called whenever
7292 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7293 * has a correct pages reserved value, so an adequate number of
7294 * pages are left in the zone after a successful __alloc_pages().
7296 static void setup_per_zone_lowmem_reserve(void)
7298 struct pglist_data
*pgdat
;
7299 enum zone_type j
, idx
;
7301 for_each_online_pgdat(pgdat
) {
7302 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7303 struct zone
*zone
= pgdat
->node_zones
+ j
;
7304 unsigned long managed_pages
= zone
->managed_pages
;
7306 zone
->lowmem_reserve
[j
] = 0;
7310 struct zone
*lower_zone
;
7313 lower_zone
= pgdat
->node_zones
+ idx
;
7315 if (sysctl_lowmem_reserve_ratio
[idx
] < 1) {
7316 sysctl_lowmem_reserve_ratio
[idx
] = 0;
7317 lower_zone
->lowmem_reserve
[j
] = 0;
7319 lower_zone
->lowmem_reserve
[j
] =
7320 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7322 managed_pages
+= lower_zone
->managed_pages
;
7327 /* update totalreserve_pages */
7328 calculate_totalreserve_pages();
7331 static void __setup_per_zone_wmarks(void)
7333 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7334 unsigned long lowmem_pages
= 0;
7336 unsigned long flags
;
7338 /* Calculate total number of !ZONE_HIGHMEM pages */
7339 for_each_zone(zone
) {
7340 if (!is_highmem(zone
))
7341 lowmem_pages
+= zone
->managed_pages
;
7344 for_each_zone(zone
) {
7347 spin_lock_irqsave(&zone
->lock
, flags
);
7348 tmp
= (u64
)pages_min
* zone
->managed_pages
;
7349 do_div(tmp
, lowmem_pages
);
7350 if (is_highmem(zone
)) {
7352 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7353 * need highmem pages, so cap pages_min to a small
7356 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7357 * deltas control asynch page reclaim, and so should
7358 * not be capped for highmem.
7360 unsigned long min_pages
;
7362 min_pages
= zone
->managed_pages
/ 1024;
7363 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7364 zone
->watermark
[WMARK_MIN
] = min_pages
;
7367 * If it's a lowmem zone, reserve a number of pages
7368 * proportionate to the zone's size.
7370 zone
->watermark
[WMARK_MIN
] = tmp
;
7374 * Set the kswapd watermarks distance according to the
7375 * scale factor in proportion to available memory, but
7376 * ensure a minimum size on small systems.
7378 tmp
= max_t(u64
, tmp
>> 2,
7379 mult_frac(zone
->managed_pages
,
7380 watermark_scale_factor
, 10000));
7382 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7383 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7385 spin_unlock_irqrestore(&zone
->lock
, flags
);
7388 /* update totalreserve_pages */
7389 calculate_totalreserve_pages();
7393 * setup_per_zone_wmarks - called when min_free_kbytes changes
7394 * or when memory is hot-{added|removed}
7396 * Ensures that the watermark[min,low,high] values for each zone are set
7397 * correctly with respect to min_free_kbytes.
7399 void setup_per_zone_wmarks(void)
7401 static DEFINE_SPINLOCK(lock
);
7404 __setup_per_zone_wmarks();
7409 * Initialise min_free_kbytes.
7411 * For small machines we want it small (128k min). For large machines
7412 * we want it large (64MB max). But it is not linear, because network
7413 * bandwidth does not increase linearly with machine size. We use
7415 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7416 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7432 int __meminit
init_per_zone_wmark_min(void)
7434 unsigned long lowmem_kbytes
;
7435 int new_min_free_kbytes
;
7437 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7438 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7440 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7441 min_free_kbytes
= new_min_free_kbytes
;
7442 if (min_free_kbytes
< 128)
7443 min_free_kbytes
= 128;
7444 if (min_free_kbytes
> 65536)
7445 min_free_kbytes
= 65536;
7447 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7448 new_min_free_kbytes
, user_min_free_kbytes
);
7450 setup_per_zone_wmarks();
7451 refresh_zone_stat_thresholds();
7452 setup_per_zone_lowmem_reserve();
7455 setup_min_unmapped_ratio();
7456 setup_min_slab_ratio();
7461 core_initcall(init_per_zone_wmark_min
)
7464 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7465 * that we can call two helper functions whenever min_free_kbytes
7468 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7469 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7473 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7478 user_min_free_kbytes
= min_free_kbytes
;
7479 setup_per_zone_wmarks();
7484 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7485 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7489 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7494 setup_per_zone_wmarks();
7500 static void setup_min_unmapped_ratio(void)
7505 for_each_online_pgdat(pgdat
)
7506 pgdat
->min_unmapped_pages
= 0;
7509 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
7510 sysctl_min_unmapped_ratio
) / 100;
7514 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7515 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7519 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7523 setup_min_unmapped_ratio();
7528 static void setup_min_slab_ratio(void)
7533 for_each_online_pgdat(pgdat
)
7534 pgdat
->min_slab_pages
= 0;
7537 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
7538 sysctl_min_slab_ratio
) / 100;
7541 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7542 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7546 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7550 setup_min_slab_ratio();
7557 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7558 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7559 * whenever sysctl_lowmem_reserve_ratio changes.
7561 * The reserve ratio obviously has absolutely no relation with the
7562 * minimum watermarks. The lowmem reserve ratio can only make sense
7563 * if in function of the boot time zone sizes.
7565 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7566 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7568 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7569 setup_per_zone_lowmem_reserve();
7574 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7575 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7576 * pagelist can have before it gets flushed back to buddy allocator.
7578 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7579 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7582 int old_percpu_pagelist_fraction
;
7585 mutex_lock(&pcp_batch_high_lock
);
7586 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7588 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7589 if (!write
|| ret
< 0)
7592 /* Sanity checking to avoid pcp imbalance */
7593 if (percpu_pagelist_fraction
&&
7594 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7595 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7601 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7604 for_each_populated_zone(zone
) {
7607 for_each_possible_cpu(cpu
)
7608 pageset_set_high_and_batch(zone
,
7609 per_cpu_ptr(zone
->pageset
, cpu
));
7612 mutex_unlock(&pcp_batch_high_lock
);
7617 int hashdist
= HASHDIST_DEFAULT
;
7619 static int __init
set_hashdist(char *str
)
7623 hashdist
= simple_strtoul(str
, &str
, 0);
7626 __setup("hashdist=", set_hashdist
);
7629 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7631 * Returns the number of pages that arch has reserved but
7632 * is not known to alloc_large_system_hash().
7634 static unsigned long __init
arch_reserved_kernel_pages(void)
7641 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7642 * machines. As memory size is increased the scale is also increased but at
7643 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7644 * quadruples the scale is increased by one, which means the size of hash table
7645 * only doubles, instead of quadrupling as well.
7646 * Because 32-bit systems cannot have large physical memory, where this scaling
7647 * makes sense, it is disabled on such platforms.
7649 #if __BITS_PER_LONG > 32
7650 #define ADAPT_SCALE_BASE (64ul << 30)
7651 #define ADAPT_SCALE_SHIFT 2
7652 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7656 * allocate a large system hash table from bootmem
7657 * - it is assumed that the hash table must contain an exact power-of-2
7658 * quantity of entries
7659 * - limit is the number of hash buckets, not the total allocation size
7661 void *__init
alloc_large_system_hash(const char *tablename
,
7662 unsigned long bucketsize
,
7663 unsigned long numentries
,
7666 unsigned int *_hash_shift
,
7667 unsigned int *_hash_mask
,
7668 unsigned long low_limit
,
7669 unsigned long high_limit
)
7671 unsigned long long max
= high_limit
;
7672 unsigned long log2qty
, size
;
7676 /* allow the kernel cmdline to have a say */
7678 /* round applicable memory size up to nearest megabyte */
7679 numentries
= nr_kernel_pages
;
7680 numentries
-= arch_reserved_kernel_pages();
7682 /* It isn't necessary when PAGE_SIZE >= 1MB */
7683 if (PAGE_SHIFT
< 20)
7684 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7686 #if __BITS_PER_LONG > 32
7688 unsigned long adapt
;
7690 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
7691 adapt
<<= ADAPT_SCALE_SHIFT
)
7696 /* limit to 1 bucket per 2^scale bytes of low memory */
7697 if (scale
> PAGE_SHIFT
)
7698 numentries
>>= (scale
- PAGE_SHIFT
);
7700 numentries
<<= (PAGE_SHIFT
- scale
);
7702 /* Make sure we've got at least a 0-order allocation.. */
7703 if (unlikely(flags
& HASH_SMALL
)) {
7704 /* Makes no sense without HASH_EARLY */
7705 WARN_ON(!(flags
& HASH_EARLY
));
7706 if (!(numentries
>> *_hash_shift
)) {
7707 numentries
= 1UL << *_hash_shift
;
7708 BUG_ON(!numentries
);
7710 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7711 numentries
= PAGE_SIZE
/ bucketsize
;
7713 numentries
= roundup_pow_of_two(numentries
);
7715 /* limit allocation size to 1/16 total memory by default */
7717 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7718 do_div(max
, bucketsize
);
7720 max
= min(max
, 0x80000000ULL
);
7722 if (numentries
< low_limit
)
7723 numentries
= low_limit
;
7724 if (numentries
> max
)
7727 log2qty
= ilog2(numentries
);
7729 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
7731 size
= bucketsize
<< log2qty
;
7732 if (flags
& HASH_EARLY
) {
7733 if (flags
& HASH_ZERO
)
7734 table
= memblock_alloc_nopanic(size
,
7737 table
= memblock_alloc_raw(size
,
7739 } else if (hashdist
) {
7740 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
7743 * If bucketsize is not a power-of-two, we may free
7744 * some pages at the end of hash table which
7745 * alloc_pages_exact() automatically does
7747 if (get_order(size
) < MAX_ORDER
) {
7748 table
= alloc_pages_exact(size
, gfp_flags
);
7749 kmemleak_alloc(table
, size
, 1, gfp_flags
);
7752 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7755 panic("Failed to allocate %s hash table\n", tablename
);
7757 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7758 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7761 *_hash_shift
= log2qty
;
7763 *_hash_mask
= (1 << log2qty
) - 1;
7769 * This function checks whether pageblock includes unmovable pages or not.
7770 * If @count is not zero, it is okay to include less @count unmovable pages
7772 * PageLRU check without isolation or lru_lock could race so that
7773 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7774 * check without lock_page also may miss some movable non-lru pages at
7775 * race condition. So you can't expect this function should be exact.
7777 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7779 bool skip_hwpoisoned_pages
)
7781 unsigned long pfn
, iter
, found
;
7784 * TODO we could make this much more efficient by not checking every
7785 * page in the range if we know all of them are in MOVABLE_ZONE and
7786 * that the movable zone guarantees that pages are migratable but
7787 * the later is not the case right now unfortunatelly. E.g. movablecore
7788 * can still lead to having bootmem allocations in zone_movable.
7792 * CMA allocations (alloc_contig_range) really need to mark isolate
7793 * CMA pageblocks even when they are not movable in fact so consider
7794 * them movable here.
7796 if (is_migrate_cma(migratetype
) &&
7797 is_migrate_cma(get_pageblock_migratetype(page
)))
7800 pfn
= page_to_pfn(page
);
7801 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7802 unsigned long check
= pfn
+ iter
;
7804 if (!pfn_valid_within(check
))
7807 page
= pfn_to_page(check
);
7809 if (PageReserved(page
))
7813 * If the zone is movable and we have ruled out all reserved
7814 * pages then it should be reasonably safe to assume the rest
7817 if (zone_idx(zone
) == ZONE_MOVABLE
)
7821 * Hugepages are not in LRU lists, but they're movable.
7822 * We need not scan over tail pages bacause we don't
7823 * handle each tail page individually in migration.
7825 if (PageHuge(page
)) {
7826 struct page
*head
= compound_head(page
);
7827 unsigned int skip_pages
;
7829 if (!hugepage_migration_supported(page_hstate(head
)))
7832 skip_pages
= (1 << compound_order(head
)) - (page
- head
);
7833 iter
+= skip_pages
- 1;
7838 * We can't use page_count without pin a page
7839 * because another CPU can free compound page.
7840 * This check already skips compound tails of THP
7841 * because their page->_refcount is zero at all time.
7843 if (!page_ref_count(page
)) {
7844 if (PageBuddy(page
))
7845 iter
+= (1 << page_order(page
)) - 1;
7850 * The HWPoisoned page may be not in buddy system, and
7851 * page_count() is not 0.
7853 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7856 if (__PageMovable(page
))
7862 * If there are RECLAIMABLE pages, we need to check
7863 * it. But now, memory offline itself doesn't call
7864 * shrink_node_slabs() and it still to be fixed.
7867 * If the page is not RAM, page_count()should be 0.
7868 * we don't need more check. This is an _used_ not-movable page.
7870 * The problematic thing here is PG_reserved pages. PG_reserved
7871 * is set to both of a memory hole page and a _used_ kernel
7879 WARN_ON_ONCE(zone_idx(zone
) == ZONE_MOVABLE
);
7883 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7885 static unsigned long pfn_max_align_down(unsigned long pfn
)
7887 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7888 pageblock_nr_pages
) - 1);
7891 static unsigned long pfn_max_align_up(unsigned long pfn
)
7893 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7894 pageblock_nr_pages
));
7897 /* [start, end) must belong to a single zone. */
7898 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7899 unsigned long start
, unsigned long end
)
7901 /* This function is based on compact_zone() from compaction.c. */
7902 unsigned long nr_reclaimed
;
7903 unsigned long pfn
= start
;
7904 unsigned int tries
= 0;
7909 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7910 if (fatal_signal_pending(current
)) {
7915 if (list_empty(&cc
->migratepages
)) {
7916 cc
->nr_migratepages
= 0;
7917 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7923 } else if (++tries
== 5) {
7924 ret
= ret
< 0 ? ret
: -EBUSY
;
7928 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7930 cc
->nr_migratepages
-= nr_reclaimed
;
7932 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7933 NULL
, 0, cc
->mode
, MR_CONTIG_RANGE
);
7936 putback_movable_pages(&cc
->migratepages
);
7943 * alloc_contig_range() -- tries to allocate given range of pages
7944 * @start: start PFN to allocate
7945 * @end: one-past-the-last PFN to allocate
7946 * @migratetype: migratetype of the underlaying pageblocks (either
7947 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7948 * in range must have the same migratetype and it must
7949 * be either of the two.
7950 * @gfp_mask: GFP mask to use during compaction
7952 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7953 * aligned. The PFN range must belong to a single zone.
7955 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7956 * pageblocks in the range. Once isolated, the pageblocks should not
7957 * be modified by others.
7959 * Returns zero on success or negative error code. On success all
7960 * pages which PFN is in [start, end) are allocated for the caller and
7961 * need to be freed with free_contig_range().
7963 int alloc_contig_range(unsigned long start
, unsigned long end
,
7964 unsigned migratetype
, gfp_t gfp_mask
)
7966 unsigned long outer_start
, outer_end
;
7970 struct compact_control cc
= {
7971 .nr_migratepages
= 0,
7973 .zone
= page_zone(pfn_to_page(start
)),
7974 .mode
= MIGRATE_SYNC
,
7975 .ignore_skip_hint
= true,
7976 .no_set_skip_hint
= true,
7977 .gfp_mask
= current_gfp_context(gfp_mask
),
7979 INIT_LIST_HEAD(&cc
.migratepages
);
7982 * What we do here is we mark all pageblocks in range as
7983 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7984 * have different sizes, and due to the way page allocator
7985 * work, we align the range to biggest of the two pages so
7986 * that page allocator won't try to merge buddies from
7987 * different pageblocks and change MIGRATE_ISOLATE to some
7988 * other migration type.
7990 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7991 * migrate the pages from an unaligned range (ie. pages that
7992 * we are interested in). This will put all the pages in
7993 * range back to page allocator as MIGRATE_ISOLATE.
7995 * When this is done, we take the pages in range from page
7996 * allocator removing them from the buddy system. This way
7997 * page allocator will never consider using them.
7999 * This lets us mark the pageblocks back as
8000 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8001 * aligned range but not in the unaligned, original range are
8002 * put back to page allocator so that buddy can use them.
8005 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8006 pfn_max_align_up(end
), migratetype
,
8012 * In case of -EBUSY, we'd like to know which page causes problem.
8013 * So, just fall through. test_pages_isolated() has a tracepoint
8014 * which will report the busy page.
8016 * It is possible that busy pages could become available before
8017 * the call to test_pages_isolated, and the range will actually be
8018 * allocated. So, if we fall through be sure to clear ret so that
8019 * -EBUSY is not accidentally used or returned to caller.
8021 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8022 if (ret
&& ret
!= -EBUSY
)
8027 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8028 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8029 * more, all pages in [start, end) are free in page allocator.
8030 * What we are going to do is to allocate all pages from
8031 * [start, end) (that is remove them from page allocator).
8033 * The only problem is that pages at the beginning and at the
8034 * end of interesting range may be not aligned with pages that
8035 * page allocator holds, ie. they can be part of higher order
8036 * pages. Because of this, we reserve the bigger range and
8037 * once this is done free the pages we are not interested in.
8039 * We don't have to hold zone->lock here because the pages are
8040 * isolated thus they won't get removed from buddy.
8043 lru_add_drain_all();
8044 drain_all_pages(cc
.zone
);
8047 outer_start
= start
;
8048 while (!PageBuddy(pfn_to_page(outer_start
))) {
8049 if (++order
>= MAX_ORDER
) {
8050 outer_start
= start
;
8053 outer_start
&= ~0UL << order
;
8056 if (outer_start
!= start
) {
8057 order
= page_order(pfn_to_page(outer_start
));
8060 * outer_start page could be small order buddy page and
8061 * it doesn't include start page. Adjust outer_start
8062 * in this case to report failed page properly
8063 * on tracepoint in test_pages_isolated()
8065 if (outer_start
+ (1UL << order
) <= start
)
8066 outer_start
= start
;
8069 /* Make sure the range is really isolated. */
8070 if (test_pages_isolated(outer_start
, end
, false)) {
8071 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8072 __func__
, outer_start
, end
);
8077 /* Grab isolated pages from freelists. */
8078 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8084 /* Free head and tail (if any) */
8085 if (start
!= outer_start
)
8086 free_contig_range(outer_start
, start
- outer_start
);
8087 if (end
!= outer_end
)
8088 free_contig_range(end
, outer_end
- end
);
8091 undo_isolate_page_range(pfn_max_align_down(start
),
8092 pfn_max_align_up(end
), migratetype
);
8096 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
8098 unsigned int count
= 0;
8100 for (; nr_pages
--; pfn
++) {
8101 struct page
*page
= pfn_to_page(pfn
);
8103 count
+= page_count(page
) != 1;
8106 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8110 #ifdef CONFIG_MEMORY_HOTPLUG
8112 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8113 * page high values need to be recalulated.
8115 void __meminit
zone_pcp_update(struct zone
*zone
)
8118 mutex_lock(&pcp_batch_high_lock
);
8119 for_each_possible_cpu(cpu
)
8120 pageset_set_high_and_batch(zone
,
8121 per_cpu_ptr(zone
->pageset
, cpu
));
8122 mutex_unlock(&pcp_batch_high_lock
);
8126 void zone_pcp_reset(struct zone
*zone
)
8128 unsigned long flags
;
8130 struct per_cpu_pageset
*pset
;
8132 /* avoid races with drain_pages() */
8133 local_irq_save(flags
);
8134 if (zone
->pageset
!= &boot_pageset
) {
8135 for_each_online_cpu(cpu
) {
8136 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8137 drain_zonestat(zone
, pset
);
8139 free_percpu(zone
->pageset
);
8140 zone
->pageset
= &boot_pageset
;
8142 local_irq_restore(flags
);
8145 #ifdef CONFIG_MEMORY_HOTREMOVE
8147 * All pages in the range must be in a single zone and isolated
8148 * before calling this.
8151 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8155 unsigned int order
, i
;
8157 unsigned long flags
;
8158 /* find the first valid pfn */
8159 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8164 offline_mem_sections(pfn
, end_pfn
);
8165 zone
= page_zone(pfn_to_page(pfn
));
8166 spin_lock_irqsave(&zone
->lock
, flags
);
8168 while (pfn
< end_pfn
) {
8169 if (!pfn_valid(pfn
)) {
8173 page
= pfn_to_page(pfn
);
8175 * The HWPoisoned page may be not in buddy system, and
8176 * page_count() is not 0.
8178 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8180 SetPageReserved(page
);
8184 BUG_ON(page_count(page
));
8185 BUG_ON(!PageBuddy(page
));
8186 order
= page_order(page
);
8187 #ifdef CONFIG_DEBUG_VM
8188 pr_info("remove from free list %lx %d %lx\n",
8189 pfn
, 1 << order
, end_pfn
);
8191 list_del(&page
->lru
);
8192 rmv_page_order(page
);
8193 zone
->free_area
[order
].nr_free
--;
8194 for (i
= 0; i
< (1 << order
); i
++)
8195 SetPageReserved((page
+i
));
8196 pfn
+= (1 << order
);
8198 spin_unlock_irqrestore(&zone
->lock
, flags
);
8202 bool is_free_buddy_page(struct page
*page
)
8204 struct zone
*zone
= page_zone(page
);
8205 unsigned long pfn
= page_to_pfn(page
);
8206 unsigned long flags
;
8209 spin_lock_irqsave(&zone
->lock
, flags
);
8210 for (order
= 0; order
< MAX_ORDER
; order
++) {
8211 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8213 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8216 spin_unlock_irqrestore(&zone
->lock
, flags
);
8218 return order
< MAX_ORDER
;
8221 #ifdef CONFIG_MEMORY_FAILURE
8223 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8224 * test is performed under the zone lock to prevent a race against page
8227 bool set_hwpoison_free_buddy_page(struct page
*page
)
8229 struct zone
*zone
= page_zone(page
);
8230 unsigned long pfn
= page_to_pfn(page
);
8231 unsigned long flags
;
8233 bool hwpoisoned
= false;
8235 spin_lock_irqsave(&zone
->lock
, flags
);
8236 for (order
= 0; order
< MAX_ORDER
; order
++) {
8237 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8239 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8240 if (!TestSetPageHWPoison(page
))
8245 spin_unlock_irqrestore(&zone
->lock
, flags
);