1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
71 #include <linux/padata.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
75 #include <asm/div64.h>
78 #include "page_reporting.h"
80 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
81 static DEFINE_MUTEX(pcp_batch_high_lock
);
82 #define MIN_PERCPU_PAGELIST_FRACTION (8)
84 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
85 DEFINE_PER_CPU(int, numa_node
);
86 EXPORT_PER_CPU_SYMBOL(numa_node
);
89 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
91 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
93 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
94 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
95 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
96 * defined in <linux/topology.h>.
98 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
99 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
102 /* work_structs for global per-cpu drains */
105 struct work_struct work
;
107 static DEFINE_MUTEX(pcpu_drain_mutex
);
108 static DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
110 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
111 volatile unsigned long latent_entropy __latent_entropy
;
112 EXPORT_SYMBOL(latent_entropy
);
116 * Array of node states.
118 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
119 [N_POSSIBLE
] = NODE_MASK_ALL
,
120 [N_ONLINE
] = { { [0] = 1UL } },
122 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
123 #ifdef CONFIG_HIGHMEM
124 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
126 [N_MEMORY
] = { { [0] = 1UL } },
127 [N_CPU
] = { { [0] = 1UL } },
130 EXPORT_SYMBOL(node_states
);
132 atomic_long_t _totalram_pages __read_mostly
;
133 EXPORT_SYMBOL(_totalram_pages
);
134 unsigned long totalreserve_pages __read_mostly
;
135 unsigned long totalcma_pages __read_mostly
;
137 int percpu_pagelist_fraction
;
138 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
139 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
140 DEFINE_STATIC_KEY_TRUE(init_on_alloc
);
142 DEFINE_STATIC_KEY_FALSE(init_on_alloc
);
144 EXPORT_SYMBOL(init_on_alloc
);
146 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
147 DEFINE_STATIC_KEY_TRUE(init_on_free
);
149 DEFINE_STATIC_KEY_FALSE(init_on_free
);
151 EXPORT_SYMBOL(init_on_free
);
153 static int __init
early_init_on_alloc(char *buf
)
160 ret
= kstrtobool(buf
, &bool_result
);
161 if (bool_result
&& page_poisoning_enabled())
162 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
164 static_branch_enable(&init_on_alloc
);
166 static_branch_disable(&init_on_alloc
);
169 early_param("init_on_alloc", early_init_on_alloc
);
171 static int __init
early_init_on_free(char *buf
)
178 ret
= kstrtobool(buf
, &bool_result
);
179 if (bool_result
&& page_poisoning_enabled())
180 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
182 static_branch_enable(&init_on_free
);
184 static_branch_disable(&init_on_free
);
187 early_param("init_on_free", early_init_on_free
);
190 * A cached value of the page's pageblock's migratetype, used when the page is
191 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
192 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
193 * Also the migratetype set in the page does not necessarily match the pcplist
194 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
195 * other index - this ensures that it will be put on the correct CMA freelist.
197 static inline int get_pcppage_migratetype(struct page
*page
)
202 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
204 page
->index
= migratetype
;
207 #ifdef CONFIG_PM_SLEEP
209 * The following functions are used by the suspend/hibernate code to temporarily
210 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
211 * while devices are suspended. To avoid races with the suspend/hibernate code,
212 * they should always be called with system_transition_mutex held
213 * (gfp_allowed_mask also should only be modified with system_transition_mutex
214 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
215 * with that modification).
218 static gfp_t saved_gfp_mask
;
220 void pm_restore_gfp_mask(void)
222 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
223 if (saved_gfp_mask
) {
224 gfp_allowed_mask
= saved_gfp_mask
;
229 void pm_restrict_gfp_mask(void)
231 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
232 WARN_ON(saved_gfp_mask
);
233 saved_gfp_mask
= gfp_allowed_mask
;
234 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
237 bool pm_suspended_storage(void)
239 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
243 #endif /* CONFIG_PM_SLEEP */
245 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
246 unsigned int pageblock_order __read_mostly
;
249 static void __free_pages_ok(struct page
*page
, unsigned int order
);
252 * results with 256, 32 in the lowmem_reserve sysctl:
253 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
254 * 1G machine -> (16M dma, 784M normal, 224M high)
255 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
256 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
257 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
259 * TBD: should special case ZONE_DMA32 machines here - in those we normally
260 * don't need any ZONE_NORMAL reservation
262 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
263 #ifdef CONFIG_ZONE_DMA
266 #ifdef CONFIG_ZONE_DMA32
270 #ifdef CONFIG_HIGHMEM
276 static char * const zone_names
[MAX_NR_ZONES
] = {
277 #ifdef CONFIG_ZONE_DMA
280 #ifdef CONFIG_ZONE_DMA32
284 #ifdef CONFIG_HIGHMEM
288 #ifdef CONFIG_ZONE_DEVICE
293 const char * const migratetype_names
[MIGRATE_TYPES
] = {
301 #ifdef CONFIG_MEMORY_ISOLATION
306 compound_page_dtor
* const compound_page_dtors
[NR_COMPOUND_DTORS
] = {
307 [NULL_COMPOUND_DTOR
] = NULL
,
308 [COMPOUND_PAGE_DTOR
] = free_compound_page
,
309 #ifdef CONFIG_HUGETLB_PAGE
310 [HUGETLB_PAGE_DTOR
] = free_huge_page
,
312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
313 [TRANSHUGE_PAGE_DTOR
] = free_transhuge_page
,
317 int min_free_kbytes
= 1024;
318 int user_min_free_kbytes
= -1;
319 #ifdef CONFIG_DISCONTIGMEM
321 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
322 * are not on separate NUMA nodes. Functionally this works but with
323 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
324 * quite small. By default, do not boost watermarks on discontigmem as in
325 * many cases very high-order allocations like THP are likely to be
326 * unsupported and the premature reclaim offsets the advantage of long-term
327 * fragmentation avoidance.
329 int watermark_boost_factor __read_mostly
;
331 int watermark_boost_factor __read_mostly
= 15000;
333 int watermark_scale_factor
= 10;
335 static unsigned long nr_kernel_pages __initdata
;
336 static unsigned long nr_all_pages __initdata
;
337 static unsigned long dma_reserve __initdata
;
339 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
340 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
341 static unsigned long required_kernelcore __initdata
;
342 static unsigned long required_kernelcore_percent __initdata
;
343 static unsigned long required_movablecore __initdata
;
344 static unsigned long required_movablecore_percent __initdata
;
345 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
346 static bool mirrored_kernelcore __meminitdata
;
348 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
350 EXPORT_SYMBOL(movable_zone
);
353 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
354 unsigned int nr_online_nodes __read_mostly
= 1;
355 EXPORT_SYMBOL(nr_node_ids
);
356 EXPORT_SYMBOL(nr_online_nodes
);
359 int page_group_by_mobility_disabled __read_mostly
;
361 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
363 * During boot we initialize deferred pages on-demand, as needed, but once
364 * page_alloc_init_late() has finished, the deferred pages are all initialized,
365 * and we can permanently disable that path.
367 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
370 * Calling kasan_free_pages() only after deferred memory initialization
371 * has completed. Poisoning pages during deferred memory init will greatly
372 * lengthen the process and cause problem in large memory systems as the
373 * deferred pages initialization is done with interrupt disabled.
375 * Assuming that there will be no reference to those newly initialized
376 * pages before they are ever allocated, this should have no effect on
377 * KASAN memory tracking as the poison will be properly inserted at page
378 * allocation time. The only corner case is when pages are allocated by
379 * on-demand allocation and then freed again before the deferred pages
380 * initialization is done, but this is not likely to happen.
382 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
384 if (!static_branch_unlikely(&deferred_pages
))
385 kasan_free_pages(page
, order
);
388 /* Returns true if the struct page for the pfn is uninitialised */
389 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
391 int nid
= early_pfn_to_nid(pfn
);
393 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
400 * Returns true when the remaining initialisation should be deferred until
401 * later in the boot cycle when it can be parallelised.
403 static bool __meminit
404 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
406 static unsigned long prev_end_pfn
, nr_initialised
;
409 * prev_end_pfn static that contains the end of previous zone
410 * No need to protect because called very early in boot before smp_init.
412 if (prev_end_pfn
!= end_pfn
) {
413 prev_end_pfn
= end_pfn
;
417 /* Always populate low zones for address-constrained allocations */
418 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
422 * We start only with one section of pages, more pages are added as
423 * needed until the rest of deferred pages are initialized.
426 if ((nr_initialised
> PAGES_PER_SECTION
) &&
427 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
428 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
434 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
436 static inline bool early_page_uninitialised(unsigned long pfn
)
441 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
447 /* Return a pointer to the bitmap storing bits affecting a block of pages */
448 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
451 #ifdef CONFIG_SPARSEMEM
452 return section_to_usemap(__pfn_to_section(pfn
));
454 return page_zone(page
)->pageblock_flags
;
455 #endif /* CONFIG_SPARSEMEM */
458 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
460 #ifdef CONFIG_SPARSEMEM
461 pfn
&= (PAGES_PER_SECTION
-1);
462 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
464 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
465 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
466 #endif /* CONFIG_SPARSEMEM */
470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
471 * @page: The page within the block of interest
472 * @pfn: The target page frame number
473 * @end_bitidx: The last bit of interest to retrieve
474 * @mask: mask of bits that the caller is interested in
476 * Return: pageblock_bits flags
478 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
480 unsigned long end_bitidx
,
483 unsigned long *bitmap
;
484 unsigned long bitidx
, word_bitidx
;
487 bitmap
= get_pageblock_bitmap(page
, pfn
);
488 bitidx
= pfn_to_bitidx(page
, pfn
);
489 word_bitidx
= bitidx
/ BITS_PER_LONG
;
490 bitidx
&= (BITS_PER_LONG
-1);
492 word
= bitmap
[word_bitidx
];
493 bitidx
+= end_bitidx
;
494 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
497 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
498 unsigned long end_bitidx
,
501 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
504 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
506 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
510 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
511 * @page: The page within the block of interest
512 * @flags: The flags to set
513 * @pfn: The target page frame number
514 * @end_bitidx: The last bit of interest
515 * @mask: mask of bits that the caller is interested in
517 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
519 unsigned long end_bitidx
,
522 unsigned long *bitmap
;
523 unsigned long bitidx
, word_bitidx
;
524 unsigned long old_word
, word
;
526 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
527 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
529 bitmap
= get_pageblock_bitmap(page
, pfn
);
530 bitidx
= pfn_to_bitidx(page
, pfn
);
531 word_bitidx
= bitidx
/ BITS_PER_LONG
;
532 bitidx
&= (BITS_PER_LONG
-1);
534 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
536 bitidx
+= end_bitidx
;
537 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
538 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
540 word
= READ_ONCE(bitmap
[word_bitidx
]);
542 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
543 if (word
== old_word
)
549 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
551 if (unlikely(page_group_by_mobility_disabled
&&
552 migratetype
< MIGRATE_PCPTYPES
))
553 migratetype
= MIGRATE_UNMOVABLE
;
555 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
556 PB_migrate
, PB_migrate_end
);
559 #ifdef CONFIG_DEBUG_VM
560 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
564 unsigned long pfn
= page_to_pfn(page
);
565 unsigned long sp
, start_pfn
;
568 seq
= zone_span_seqbegin(zone
);
569 start_pfn
= zone
->zone_start_pfn
;
570 sp
= zone
->spanned_pages
;
571 if (!zone_spans_pfn(zone
, pfn
))
573 } while (zone_span_seqretry(zone
, seq
));
576 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
577 pfn
, zone_to_nid(zone
), zone
->name
,
578 start_pfn
, start_pfn
+ sp
);
583 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
585 if (!pfn_valid_within(page_to_pfn(page
)))
587 if (zone
!= page_zone(page
))
593 * Temporary debugging check for pages not lying within a given zone.
595 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
597 if (page_outside_zone_boundaries(zone
, page
))
599 if (!page_is_consistent(zone
, page
))
605 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
611 static void bad_page(struct page
*page
, const char *reason
)
613 static unsigned long resume
;
614 static unsigned long nr_shown
;
615 static unsigned long nr_unshown
;
618 * Allow a burst of 60 reports, then keep quiet for that minute;
619 * or allow a steady drip of one report per second.
621 if (nr_shown
== 60) {
622 if (time_before(jiffies
, resume
)) {
628 "BUG: Bad page state: %lu messages suppressed\n",
635 resume
= jiffies
+ 60 * HZ
;
637 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
638 current
->comm
, page_to_pfn(page
));
639 __dump_page(page
, reason
);
640 dump_page_owner(page
);
645 /* Leave bad fields for debug, except PageBuddy could make trouble */
646 page_mapcount_reset(page
); /* remove PageBuddy */
647 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
651 * Higher-order pages are called "compound pages". They are structured thusly:
653 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
655 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
656 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
658 * The first tail page's ->compound_dtor holds the offset in array of compound
659 * page destructors. See compound_page_dtors.
661 * The first tail page's ->compound_order holds the order of allocation.
662 * This usage means that zero-order pages may not be compound.
665 void free_compound_page(struct page
*page
)
667 mem_cgroup_uncharge(page
);
668 __free_pages_ok(page
, compound_order(page
));
671 void prep_compound_page(struct page
*page
, unsigned int order
)
674 int nr_pages
= 1 << order
;
676 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
677 set_compound_order(page
, order
);
679 for (i
= 1; i
< nr_pages
; i
++) {
680 struct page
*p
= page
+ i
;
681 set_page_count(p
, 0);
682 p
->mapping
= TAIL_MAPPING
;
683 set_compound_head(p
, page
);
685 atomic_set(compound_mapcount_ptr(page
), -1);
686 if (hpage_pincount_available(page
))
687 atomic_set(compound_pincount_ptr(page
), 0);
690 #ifdef CONFIG_DEBUG_PAGEALLOC
691 unsigned int _debug_guardpage_minorder
;
693 bool _debug_pagealloc_enabled_early __read_mostly
694 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
695 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
696 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
697 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
699 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
701 static int __init
early_debug_pagealloc(char *buf
)
703 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
705 early_param("debug_pagealloc", early_debug_pagealloc
);
707 void init_debug_pagealloc(void)
709 if (!debug_pagealloc_enabled())
712 static_branch_enable(&_debug_pagealloc_enabled
);
714 if (!debug_guardpage_minorder())
717 static_branch_enable(&_debug_guardpage_enabled
);
720 static int __init
debug_guardpage_minorder_setup(char *buf
)
724 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
725 pr_err("Bad debug_guardpage_minorder value\n");
728 _debug_guardpage_minorder
= res
;
729 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
732 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
734 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
735 unsigned int order
, int migratetype
)
737 if (!debug_guardpage_enabled())
740 if (order
>= debug_guardpage_minorder())
743 __SetPageGuard(page
);
744 INIT_LIST_HEAD(&page
->lru
);
745 set_page_private(page
, order
);
746 /* Guard pages are not available for any usage */
747 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
752 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
753 unsigned int order
, int migratetype
)
755 if (!debug_guardpage_enabled())
758 __ClearPageGuard(page
);
760 set_page_private(page
, 0);
761 if (!is_migrate_isolate(migratetype
))
762 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
765 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
766 unsigned int order
, int migratetype
) { return false; }
767 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
768 unsigned int order
, int migratetype
) {}
771 static inline void set_page_order(struct page
*page
, unsigned int order
)
773 set_page_private(page
, order
);
774 __SetPageBuddy(page
);
778 * This function checks whether a page is free && is the buddy
779 * we can coalesce a page and its buddy if
780 * (a) the buddy is not in a hole (check before calling!) &&
781 * (b) the buddy is in the buddy system &&
782 * (c) a page and its buddy have the same order &&
783 * (d) a page and its buddy are in the same zone.
785 * For recording whether a page is in the buddy system, we set PageBuddy.
786 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
788 * For recording page's order, we use page_private(page).
790 static inline bool page_is_buddy(struct page
*page
, struct page
*buddy
,
793 if (!page_is_guard(buddy
) && !PageBuddy(buddy
))
796 if (page_order(buddy
) != order
)
800 * zone check is done late to avoid uselessly calculating
801 * zone/node ids for pages that could never merge.
803 if (page_zone_id(page
) != page_zone_id(buddy
))
806 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
811 #ifdef CONFIG_COMPACTION
812 static inline struct capture_control
*task_capc(struct zone
*zone
)
814 struct capture_control
*capc
= current
->capture_control
;
817 !(current
->flags
& PF_KTHREAD
) &&
819 capc
->cc
->zone
== zone
&&
820 capc
->cc
->direct_compaction
? capc
: NULL
;
824 compaction_capture(struct capture_control
*capc
, struct page
*page
,
825 int order
, int migratetype
)
827 if (!capc
|| order
!= capc
->cc
->order
)
830 /* Do not accidentally pollute CMA or isolated regions*/
831 if (is_migrate_cma(migratetype
) ||
832 is_migrate_isolate(migratetype
))
836 * Do not let lower order allocations polluate a movable pageblock.
837 * This might let an unmovable request use a reclaimable pageblock
838 * and vice-versa but no more than normal fallback logic which can
839 * have trouble finding a high-order free page.
841 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
849 static inline struct capture_control
*task_capc(struct zone
*zone
)
855 compaction_capture(struct capture_control
*capc
, struct page
*page
,
856 int order
, int migratetype
)
860 #endif /* CONFIG_COMPACTION */
862 /* Used for pages not on another list */
863 static inline void add_to_free_list(struct page
*page
, struct zone
*zone
,
864 unsigned int order
, int migratetype
)
866 struct free_area
*area
= &zone
->free_area
[order
];
868 list_add(&page
->lru
, &area
->free_list
[migratetype
]);
872 /* Used for pages not on another list */
873 static inline void add_to_free_list_tail(struct page
*page
, struct zone
*zone
,
874 unsigned int order
, int migratetype
)
876 struct free_area
*area
= &zone
->free_area
[order
];
878 list_add_tail(&page
->lru
, &area
->free_list
[migratetype
]);
882 /* Used for pages which are on another list */
883 static inline void move_to_free_list(struct page
*page
, struct zone
*zone
,
884 unsigned int order
, int migratetype
)
886 struct free_area
*area
= &zone
->free_area
[order
];
888 list_move(&page
->lru
, &area
->free_list
[migratetype
]);
891 static inline void del_page_from_free_list(struct page
*page
, struct zone
*zone
,
894 /* clear reported state and update reported page count */
895 if (page_reported(page
))
896 __ClearPageReported(page
);
898 list_del(&page
->lru
);
899 __ClearPageBuddy(page
);
900 set_page_private(page
, 0);
901 zone
->free_area
[order
].nr_free
--;
905 * If this is not the largest possible page, check if the buddy
906 * of the next-highest order is free. If it is, it's possible
907 * that pages are being freed that will coalesce soon. In case,
908 * that is happening, add the free page to the tail of the list
909 * so it's less likely to be used soon and more likely to be merged
910 * as a higher order page
913 buddy_merge_likely(unsigned long pfn
, unsigned long buddy_pfn
,
914 struct page
*page
, unsigned int order
)
916 struct page
*higher_page
, *higher_buddy
;
917 unsigned long combined_pfn
;
919 if (order
>= MAX_ORDER
- 2)
922 if (!pfn_valid_within(buddy_pfn
))
925 combined_pfn
= buddy_pfn
& pfn
;
926 higher_page
= page
+ (combined_pfn
- pfn
);
927 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
928 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
930 return pfn_valid_within(buddy_pfn
) &&
931 page_is_buddy(higher_page
, higher_buddy
, order
+ 1);
935 * Freeing function for a buddy system allocator.
937 * The concept of a buddy system is to maintain direct-mapped table
938 * (containing bit values) for memory blocks of various "orders".
939 * The bottom level table contains the map for the smallest allocatable
940 * units of memory (here, pages), and each level above it describes
941 * pairs of units from the levels below, hence, "buddies".
942 * At a high level, all that happens here is marking the table entry
943 * at the bottom level available, and propagating the changes upward
944 * as necessary, plus some accounting needed to play nicely with other
945 * parts of the VM system.
946 * At each level, we keep a list of pages, which are heads of continuous
947 * free pages of length of (1 << order) and marked with PageBuddy.
948 * Page's order is recorded in page_private(page) field.
949 * So when we are allocating or freeing one, we can derive the state of the
950 * other. That is, if we allocate a small block, and both were
951 * free, the remainder of the region must be split into blocks.
952 * If a block is freed, and its buddy is also free, then this
953 * triggers coalescing into a block of larger size.
958 static inline void __free_one_page(struct page
*page
,
960 struct zone
*zone
, unsigned int order
,
961 int migratetype
, bool report
)
963 struct capture_control
*capc
= task_capc(zone
);
964 unsigned long uninitialized_var(buddy_pfn
);
965 unsigned long combined_pfn
;
966 unsigned int max_order
;
970 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
972 VM_BUG_ON(!zone_is_initialized(zone
));
973 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
975 VM_BUG_ON(migratetype
== -1);
976 if (likely(!is_migrate_isolate(migratetype
)))
977 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
979 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
980 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
983 while (order
< max_order
- 1) {
984 if (compaction_capture(capc
, page
, order
, migratetype
)) {
985 __mod_zone_freepage_state(zone
, -(1 << order
),
989 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
990 buddy
= page
+ (buddy_pfn
- pfn
);
992 if (!pfn_valid_within(buddy_pfn
))
994 if (!page_is_buddy(page
, buddy
, order
))
997 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
998 * merge with it and move up one order.
1000 if (page_is_guard(buddy
))
1001 clear_page_guard(zone
, buddy
, order
, migratetype
);
1003 del_page_from_free_list(buddy
, zone
, order
);
1004 combined_pfn
= buddy_pfn
& pfn
;
1005 page
= page
+ (combined_pfn
- pfn
);
1009 if (max_order
< MAX_ORDER
) {
1010 /* If we are here, it means order is >= pageblock_order.
1011 * We want to prevent merge between freepages on isolate
1012 * pageblock and normal pageblock. Without this, pageblock
1013 * isolation could cause incorrect freepage or CMA accounting.
1015 * We don't want to hit this code for the more frequent
1016 * low-order merging.
1018 if (unlikely(has_isolate_pageblock(zone
))) {
1021 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
1022 buddy
= page
+ (buddy_pfn
- pfn
);
1023 buddy_mt
= get_pageblock_migratetype(buddy
);
1025 if (migratetype
!= buddy_mt
1026 && (is_migrate_isolate(migratetype
) ||
1027 is_migrate_isolate(buddy_mt
)))
1031 goto continue_merging
;
1035 set_page_order(page
, order
);
1037 if (is_shuffle_order(order
))
1038 to_tail
= shuffle_pick_tail();
1040 to_tail
= buddy_merge_likely(pfn
, buddy_pfn
, page
, order
);
1043 add_to_free_list_tail(page
, zone
, order
, migratetype
);
1045 add_to_free_list(page
, zone
, order
, migratetype
);
1047 /* Notify page reporting subsystem of freed page */
1049 page_reporting_notify_free(order
);
1053 * A bad page could be due to a number of fields. Instead of multiple branches,
1054 * try and check multiple fields with one check. The caller must do a detailed
1055 * check if necessary.
1057 static inline bool page_expected_state(struct page
*page
,
1058 unsigned long check_flags
)
1060 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1063 if (unlikely((unsigned long)page
->mapping
|
1064 page_ref_count(page
) |
1066 (unsigned long)page
->mem_cgroup
|
1068 (page
->flags
& check_flags
)))
1074 static const char *page_bad_reason(struct page
*page
, unsigned long flags
)
1076 const char *bad_reason
= NULL
;
1078 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1079 bad_reason
= "nonzero mapcount";
1080 if (unlikely(page
->mapping
!= NULL
))
1081 bad_reason
= "non-NULL mapping";
1082 if (unlikely(page_ref_count(page
) != 0))
1083 bad_reason
= "nonzero _refcount";
1084 if (unlikely(page
->flags
& flags
)) {
1085 if (flags
== PAGE_FLAGS_CHECK_AT_PREP
)
1086 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1088 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1091 if (unlikely(page
->mem_cgroup
))
1092 bad_reason
= "page still charged to cgroup";
1097 static void check_free_page_bad(struct page
*page
)
1100 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_FREE
));
1103 static inline int check_free_page(struct page
*page
)
1105 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1108 /* Something has gone sideways, find it */
1109 check_free_page_bad(page
);
1113 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1118 * We rely page->lru.next never has bit 0 set, unless the page
1119 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1121 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1123 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1127 switch (page
- head_page
) {
1129 /* the first tail page: ->mapping may be compound_mapcount() */
1130 if (unlikely(compound_mapcount(page
))) {
1131 bad_page(page
, "nonzero compound_mapcount");
1137 * the second tail page: ->mapping is
1138 * deferred_list.next -- ignore value.
1142 if (page
->mapping
!= TAIL_MAPPING
) {
1143 bad_page(page
, "corrupted mapping in tail page");
1148 if (unlikely(!PageTail(page
))) {
1149 bad_page(page
, "PageTail not set");
1152 if (unlikely(compound_head(page
) != head_page
)) {
1153 bad_page(page
, "compound_head not consistent");
1158 page
->mapping
= NULL
;
1159 clear_compound_head(page
);
1163 static void kernel_init_free_pages(struct page
*page
, int numpages
)
1167 for (i
= 0; i
< numpages
; i
++)
1168 clear_highpage(page
+ i
);
1171 static __always_inline
bool free_pages_prepare(struct page
*page
,
1172 unsigned int order
, bool check_free
)
1176 VM_BUG_ON_PAGE(PageTail(page
), page
);
1178 trace_mm_page_free(page
, order
);
1181 * Check tail pages before head page information is cleared to
1182 * avoid checking PageCompound for order-0 pages.
1184 if (unlikely(order
)) {
1185 bool compound
= PageCompound(page
);
1188 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1191 ClearPageDoubleMap(page
);
1192 for (i
= 1; i
< (1 << order
); i
++) {
1194 bad
+= free_tail_pages_check(page
, page
+ i
);
1195 if (unlikely(check_free_page(page
+ i
))) {
1199 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1202 if (PageMappingFlags(page
))
1203 page
->mapping
= NULL
;
1204 if (memcg_kmem_enabled() && PageKmemcg(page
))
1205 __memcg_kmem_uncharge_page(page
, order
);
1207 bad
+= check_free_page(page
);
1211 page_cpupid_reset_last(page
);
1212 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1213 reset_page_owner(page
, order
);
1215 if (!PageHighMem(page
)) {
1216 debug_check_no_locks_freed(page_address(page
),
1217 PAGE_SIZE
<< order
);
1218 debug_check_no_obj_freed(page_address(page
),
1219 PAGE_SIZE
<< order
);
1221 if (want_init_on_free())
1222 kernel_init_free_pages(page
, 1 << order
);
1224 kernel_poison_pages(page
, 1 << order
, 0);
1226 * arch_free_page() can make the page's contents inaccessible. s390
1227 * does this. So nothing which can access the page's contents should
1228 * happen after this.
1230 arch_free_page(page
, order
);
1232 if (debug_pagealloc_enabled_static())
1233 kernel_map_pages(page
, 1 << order
, 0);
1235 kasan_free_nondeferred_pages(page
, order
);
1240 #ifdef CONFIG_DEBUG_VM
1242 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1243 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1244 * moved from pcp lists to free lists.
1246 static bool free_pcp_prepare(struct page
*page
)
1248 return free_pages_prepare(page
, 0, true);
1251 static bool bulkfree_pcp_prepare(struct page
*page
)
1253 if (debug_pagealloc_enabled_static())
1254 return check_free_page(page
);
1260 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1261 * moving from pcp lists to free list in order to reduce overhead. With
1262 * debug_pagealloc enabled, they are checked also immediately when being freed
1265 static bool free_pcp_prepare(struct page
*page
)
1267 if (debug_pagealloc_enabled_static())
1268 return free_pages_prepare(page
, 0, true);
1270 return free_pages_prepare(page
, 0, false);
1273 static bool bulkfree_pcp_prepare(struct page
*page
)
1275 return check_free_page(page
);
1277 #endif /* CONFIG_DEBUG_VM */
1279 static inline void prefetch_buddy(struct page
*page
)
1281 unsigned long pfn
= page_to_pfn(page
);
1282 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1283 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1289 * Frees a number of pages from the PCP lists
1290 * Assumes all pages on list are in same zone, and of same order.
1291 * count is the number of pages to free.
1293 * If the zone was previously in an "all pages pinned" state then look to
1294 * see if this freeing clears that state.
1296 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1297 * pinned" detection logic.
1299 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1300 struct per_cpu_pages
*pcp
)
1302 int migratetype
= 0;
1304 int prefetch_nr
= 0;
1305 bool isolated_pageblocks
;
1306 struct page
*page
, *tmp
;
1310 struct list_head
*list
;
1313 * Remove pages from lists in a round-robin fashion. A
1314 * batch_free count is maintained that is incremented when an
1315 * empty list is encountered. This is so more pages are freed
1316 * off fuller lists instead of spinning excessively around empty
1321 if (++migratetype
== MIGRATE_PCPTYPES
)
1323 list
= &pcp
->lists
[migratetype
];
1324 } while (list_empty(list
));
1326 /* This is the only non-empty list. Free them all. */
1327 if (batch_free
== MIGRATE_PCPTYPES
)
1331 page
= list_last_entry(list
, struct page
, lru
);
1332 /* must delete to avoid corrupting pcp list */
1333 list_del(&page
->lru
);
1336 if (bulkfree_pcp_prepare(page
))
1339 list_add_tail(&page
->lru
, &head
);
1342 * We are going to put the page back to the global
1343 * pool, prefetch its buddy to speed up later access
1344 * under zone->lock. It is believed the overhead of
1345 * an additional test and calculating buddy_pfn here
1346 * can be offset by reduced memory latency later. To
1347 * avoid excessive prefetching due to large count, only
1348 * prefetch buddy for the first pcp->batch nr of pages.
1350 if (prefetch_nr
++ < pcp
->batch
)
1351 prefetch_buddy(page
);
1352 } while (--count
&& --batch_free
&& !list_empty(list
));
1355 spin_lock(&zone
->lock
);
1356 isolated_pageblocks
= has_isolate_pageblock(zone
);
1359 * Use safe version since after __free_one_page(),
1360 * page->lru.next will not point to original list.
1362 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1363 int mt
= get_pcppage_migratetype(page
);
1364 /* MIGRATE_ISOLATE page should not go to pcplists */
1365 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1366 /* Pageblock could have been isolated meanwhile */
1367 if (unlikely(isolated_pageblocks
))
1368 mt
= get_pageblock_migratetype(page
);
1370 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
, true);
1371 trace_mm_page_pcpu_drain(page
, 0, mt
);
1373 spin_unlock(&zone
->lock
);
1376 static void free_one_page(struct zone
*zone
,
1377 struct page
*page
, unsigned long pfn
,
1381 spin_lock(&zone
->lock
);
1382 if (unlikely(has_isolate_pageblock(zone
) ||
1383 is_migrate_isolate(migratetype
))) {
1384 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1386 __free_one_page(page
, pfn
, zone
, order
, migratetype
, true);
1387 spin_unlock(&zone
->lock
);
1390 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1391 unsigned long zone
, int nid
)
1393 mm_zero_struct_page(page
);
1394 set_page_links(page
, zone
, nid
, pfn
);
1395 init_page_count(page
);
1396 page_mapcount_reset(page
);
1397 page_cpupid_reset_last(page
);
1398 page_kasan_tag_reset(page
);
1400 INIT_LIST_HEAD(&page
->lru
);
1401 #ifdef WANT_PAGE_VIRTUAL
1402 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1403 if (!is_highmem_idx(zone
))
1404 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1408 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1409 static void __meminit
init_reserved_page(unsigned long pfn
)
1414 if (!early_page_uninitialised(pfn
))
1417 nid
= early_pfn_to_nid(pfn
);
1418 pgdat
= NODE_DATA(nid
);
1420 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1421 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1423 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1426 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1429 static inline void init_reserved_page(unsigned long pfn
)
1432 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1435 * Initialised pages do not have PageReserved set. This function is
1436 * called for each range allocated by the bootmem allocator and
1437 * marks the pages PageReserved. The remaining valid pages are later
1438 * sent to the buddy page allocator.
1440 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1442 unsigned long start_pfn
= PFN_DOWN(start
);
1443 unsigned long end_pfn
= PFN_UP(end
);
1445 for (; start_pfn
< end_pfn
; start_pfn
++) {
1446 if (pfn_valid(start_pfn
)) {
1447 struct page
*page
= pfn_to_page(start_pfn
);
1449 init_reserved_page(start_pfn
);
1451 /* Avoid false-positive PageTail() */
1452 INIT_LIST_HEAD(&page
->lru
);
1455 * no need for atomic set_bit because the struct
1456 * page is not visible yet so nobody should
1459 __SetPageReserved(page
);
1464 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1466 unsigned long flags
;
1468 unsigned long pfn
= page_to_pfn(page
);
1470 if (!free_pages_prepare(page
, order
, true))
1473 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1474 local_irq_save(flags
);
1475 __count_vm_events(PGFREE
, 1 << order
);
1476 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1477 local_irq_restore(flags
);
1480 void __free_pages_core(struct page
*page
, unsigned int order
)
1482 unsigned int nr_pages
= 1 << order
;
1483 struct page
*p
= page
;
1487 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1489 __ClearPageReserved(p
);
1490 set_page_count(p
, 0);
1492 __ClearPageReserved(p
);
1493 set_page_count(p
, 0);
1495 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1496 set_page_refcounted(page
);
1497 __free_pages(page
, order
);
1500 #ifdef CONFIG_NEED_MULTIPLE_NODES
1502 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1504 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1507 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1509 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
1510 struct mminit_pfnnid_cache
*state
)
1512 unsigned long start_pfn
, end_pfn
;
1515 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
1516 return state
->last_nid
;
1518 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
1519 if (nid
!= NUMA_NO_NODE
) {
1520 state
->last_start
= start_pfn
;
1521 state
->last_end
= end_pfn
;
1522 state
->last_nid
= nid
;
1527 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1529 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1531 static DEFINE_SPINLOCK(early_pfn_lock
);
1534 spin_lock(&early_pfn_lock
);
1535 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1537 nid
= first_online_node
;
1538 spin_unlock(&early_pfn_lock
);
1542 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1544 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1547 if (early_page_uninitialised(pfn
))
1549 __free_pages_core(page
, order
);
1553 * Check that the whole (or subset of) a pageblock given by the interval of
1554 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1555 * with the migration of free compaction scanner. The scanners then need to
1556 * use only pfn_valid_within() check for arches that allow holes within
1559 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1561 * It's possible on some configurations to have a setup like node0 node1 node0
1562 * i.e. it's possible that all pages within a zones range of pages do not
1563 * belong to a single zone. We assume that a border between node0 and node1
1564 * can occur within a single pageblock, but not a node0 node1 node0
1565 * interleaving within a single pageblock. It is therefore sufficient to check
1566 * the first and last page of a pageblock and avoid checking each individual
1567 * page in a pageblock.
1569 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1570 unsigned long end_pfn
, struct zone
*zone
)
1572 struct page
*start_page
;
1573 struct page
*end_page
;
1575 /* end_pfn is one past the range we are checking */
1578 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1581 start_page
= pfn_to_online_page(start_pfn
);
1585 if (page_zone(start_page
) != zone
)
1588 end_page
= pfn_to_page(end_pfn
);
1590 /* This gives a shorter code than deriving page_zone(end_page) */
1591 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1597 void set_zone_contiguous(struct zone
*zone
)
1599 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1600 unsigned long block_end_pfn
;
1602 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1603 for (; block_start_pfn
< zone_end_pfn(zone
);
1604 block_start_pfn
= block_end_pfn
,
1605 block_end_pfn
+= pageblock_nr_pages
) {
1607 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1609 if (!__pageblock_pfn_to_page(block_start_pfn
,
1610 block_end_pfn
, zone
))
1615 /* We confirm that there is no hole */
1616 zone
->contiguous
= true;
1619 void clear_zone_contiguous(struct zone
*zone
)
1621 zone
->contiguous
= false;
1624 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1625 static void __init
deferred_free_range(unsigned long pfn
,
1626 unsigned long nr_pages
)
1634 page
= pfn_to_page(pfn
);
1636 /* Free a large naturally-aligned chunk if possible */
1637 if (nr_pages
== pageblock_nr_pages
&&
1638 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1639 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1640 __free_pages_core(page
, pageblock_order
);
1644 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1645 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1646 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1647 __free_pages_core(page
, 0);
1651 /* Completion tracking for deferred_init_memmap() threads */
1652 static atomic_t pgdat_init_n_undone __initdata
;
1653 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1655 static inline void __init
pgdat_init_report_one_done(void)
1657 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1658 complete(&pgdat_init_all_done_comp
);
1662 * Returns true if page needs to be initialized or freed to buddy allocator.
1664 * First we check if pfn is valid on architectures where it is possible to have
1665 * holes within pageblock_nr_pages. On systems where it is not possible, this
1666 * function is optimized out.
1668 * Then, we check if a current large page is valid by only checking the validity
1671 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1673 if (!pfn_valid_within(pfn
))
1675 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1681 * Free pages to buddy allocator. Try to free aligned pages in
1682 * pageblock_nr_pages sizes.
1684 static void __init
deferred_free_pages(unsigned long pfn
,
1685 unsigned long end_pfn
)
1687 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1688 unsigned long nr_free
= 0;
1690 for (; pfn
< end_pfn
; pfn
++) {
1691 if (!deferred_pfn_valid(pfn
)) {
1692 deferred_free_range(pfn
- nr_free
, nr_free
);
1694 } else if (!(pfn
& nr_pgmask
)) {
1695 deferred_free_range(pfn
- nr_free
, nr_free
);
1701 /* Free the last block of pages to allocator */
1702 deferred_free_range(pfn
- nr_free
, nr_free
);
1706 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1707 * by performing it only once every pageblock_nr_pages.
1708 * Return number of pages initialized.
1710 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1712 unsigned long end_pfn
)
1714 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1715 int nid
= zone_to_nid(zone
);
1716 unsigned long nr_pages
= 0;
1717 int zid
= zone_idx(zone
);
1718 struct page
*page
= NULL
;
1720 for (; pfn
< end_pfn
; pfn
++) {
1721 if (!deferred_pfn_valid(pfn
)) {
1724 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1725 page
= pfn_to_page(pfn
);
1729 __init_single_page(page
, pfn
, zid
, nid
);
1736 * This function is meant to pre-load the iterator for the zone init.
1737 * Specifically it walks through the ranges until we are caught up to the
1738 * first_init_pfn value and exits there. If we never encounter the value we
1739 * return false indicating there are no valid ranges left.
1742 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1743 unsigned long *spfn
, unsigned long *epfn
,
1744 unsigned long first_init_pfn
)
1749 * Start out by walking through the ranges in this zone that have
1750 * already been initialized. We don't need to do anything with them
1751 * so we just need to flush them out of the system.
1753 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1754 if (*epfn
<= first_init_pfn
)
1756 if (*spfn
< first_init_pfn
)
1757 *spfn
= first_init_pfn
;
1766 * Initialize and free pages. We do it in two loops: first we initialize
1767 * struct page, then free to buddy allocator, because while we are
1768 * freeing pages we can access pages that are ahead (computing buddy
1769 * page in __free_one_page()).
1771 * In order to try and keep some memory in the cache we have the loop
1772 * broken along max page order boundaries. This way we will not cause
1773 * any issues with the buddy page computation.
1775 static unsigned long __init
1776 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1777 unsigned long *end_pfn
)
1779 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1780 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1781 unsigned long nr_pages
= 0;
1784 /* First we loop through and initialize the page values */
1785 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1788 if (mo_pfn
<= *start_pfn
)
1791 t
= min(mo_pfn
, *end_pfn
);
1792 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1794 if (mo_pfn
< *end_pfn
) {
1795 *start_pfn
= mo_pfn
;
1800 /* Reset values and now loop through freeing pages as needed */
1803 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1809 t
= min(mo_pfn
, epfn
);
1810 deferred_free_pages(spfn
, t
);
1820 deferred_init_memmap_chunk(unsigned long start_pfn
, unsigned long end_pfn
,
1823 unsigned long spfn
, epfn
;
1824 struct zone
*zone
= arg
;
1827 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
, start_pfn
);
1830 * Initialize and free pages in MAX_ORDER sized increments so that we
1831 * can avoid introducing any issues with the buddy allocator.
1833 while (spfn
< end_pfn
) {
1834 deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1839 /* An arch may override for more concurrency. */
1841 deferred_page_init_max_threads(const struct cpumask
*node_cpumask
)
1846 /* Initialise remaining memory on a node */
1847 static int __init
deferred_init_memmap(void *data
)
1849 pg_data_t
*pgdat
= data
;
1850 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1851 unsigned long spfn
= 0, epfn
= 0;
1852 unsigned long first_init_pfn
, flags
;
1853 unsigned long start
= jiffies
;
1855 int zid
, max_threads
;
1858 /* Bind memory initialisation thread to a local node if possible */
1859 if (!cpumask_empty(cpumask
))
1860 set_cpus_allowed_ptr(current
, cpumask
);
1862 pgdat_resize_lock(pgdat
, &flags
);
1863 first_init_pfn
= pgdat
->first_deferred_pfn
;
1864 if (first_init_pfn
== ULONG_MAX
) {
1865 pgdat_resize_unlock(pgdat
, &flags
);
1866 pgdat_init_report_one_done();
1870 /* Sanity check boundaries */
1871 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1872 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1873 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1876 * Once we unlock here, the zone cannot be grown anymore, thus if an
1877 * interrupt thread must allocate this early in boot, zone must be
1878 * pre-grown prior to start of deferred page initialization.
1880 pgdat_resize_unlock(pgdat
, &flags
);
1882 /* Only the highest zone is deferred so find it */
1883 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1884 zone
= pgdat
->node_zones
+ zid
;
1885 if (first_init_pfn
< zone_end_pfn(zone
))
1889 /* If the zone is empty somebody else may have cleared out the zone */
1890 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1894 max_threads
= deferred_page_init_max_threads(cpumask
);
1896 while (spfn
< epfn
) {
1897 unsigned long epfn_align
= ALIGN(epfn
, PAGES_PER_SECTION
);
1898 struct padata_mt_job job
= {
1899 .thread_fn
= deferred_init_memmap_chunk
,
1902 .size
= epfn_align
- spfn
,
1903 .align
= PAGES_PER_SECTION
,
1904 .min_chunk
= PAGES_PER_SECTION
,
1905 .max_threads
= max_threads
,
1908 padata_do_multithreaded(&job
);
1909 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1913 /* Sanity check that the next zone really is unpopulated */
1914 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1916 pr_info("node %d deferred pages initialised in %ums\n",
1917 pgdat
->node_id
, jiffies_to_msecs(jiffies
- start
));
1919 pgdat_init_report_one_done();
1924 * If this zone has deferred pages, try to grow it by initializing enough
1925 * deferred pages to satisfy the allocation specified by order, rounded up to
1926 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1927 * of SECTION_SIZE bytes by initializing struct pages in increments of
1928 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1930 * Return true when zone was grown, otherwise return false. We return true even
1931 * when we grow less than requested, to let the caller decide if there are
1932 * enough pages to satisfy the allocation.
1934 * Note: We use noinline because this function is needed only during boot, and
1935 * it is called from a __ref function _deferred_grow_zone. This way we are
1936 * making sure that it is not inlined into permanent text section.
1938 static noinline
bool __init
1939 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1941 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1942 pg_data_t
*pgdat
= zone
->zone_pgdat
;
1943 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1944 unsigned long spfn
, epfn
, flags
;
1945 unsigned long nr_pages
= 0;
1948 /* Only the last zone may have deferred pages */
1949 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1952 pgdat_resize_lock(pgdat
, &flags
);
1955 * If someone grew this zone while we were waiting for spinlock, return
1956 * true, as there might be enough pages already.
1958 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1959 pgdat_resize_unlock(pgdat
, &flags
);
1963 /* If the zone is empty somebody else may have cleared out the zone */
1964 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1965 first_deferred_pfn
)) {
1966 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1967 pgdat_resize_unlock(pgdat
, &flags
);
1968 /* Retry only once. */
1969 return first_deferred_pfn
!= ULONG_MAX
;
1973 * Initialize and free pages in MAX_ORDER sized increments so
1974 * that we can avoid introducing any issues with the buddy
1977 while (spfn
< epfn
) {
1978 /* update our first deferred PFN for this section */
1979 first_deferred_pfn
= spfn
;
1981 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1982 touch_nmi_watchdog();
1984 /* We should only stop along section boundaries */
1985 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
1988 /* If our quota has been met we can stop here */
1989 if (nr_pages
>= nr_pages_needed
)
1993 pgdat
->first_deferred_pfn
= spfn
;
1994 pgdat_resize_unlock(pgdat
, &flags
);
1996 return nr_pages
> 0;
2000 * deferred_grow_zone() is __init, but it is called from
2001 * get_page_from_freelist() during early boot until deferred_pages permanently
2002 * disables this call. This is why we have refdata wrapper to avoid warning,
2003 * and to ensure that the function body gets unloaded.
2006 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
2008 return deferred_grow_zone(zone
, order
);
2011 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2013 void __init
page_alloc_init_late(void)
2018 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2020 /* There will be num_node_state(N_MEMORY) threads */
2021 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
2022 for_each_node_state(nid
, N_MEMORY
) {
2023 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
2026 /* Block until all are initialised */
2027 wait_for_completion(&pgdat_init_all_done_comp
);
2030 * The number of managed pages has changed due to the initialisation
2031 * so the pcpu batch and high limits needs to be updated or the limits
2032 * will be artificially small.
2034 for_each_populated_zone(zone
)
2035 zone_pcp_update(zone
);
2038 * We initialized the rest of the deferred pages. Permanently disable
2039 * on-demand struct page initialization.
2041 static_branch_disable(&deferred_pages
);
2043 /* Reinit limits that are based on free pages after the kernel is up */
2044 files_maxfiles_init();
2047 /* Discard memblock private memory */
2050 for_each_node_state(nid
, N_MEMORY
)
2051 shuffle_free_memory(NODE_DATA(nid
));
2053 for_each_populated_zone(zone
)
2054 set_zone_contiguous(zone
);
2058 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2059 void __init
init_cma_reserved_pageblock(struct page
*page
)
2061 unsigned i
= pageblock_nr_pages
;
2062 struct page
*p
= page
;
2065 __ClearPageReserved(p
);
2066 set_page_count(p
, 0);
2069 set_pageblock_migratetype(page
, MIGRATE_CMA
);
2071 if (pageblock_order
>= MAX_ORDER
) {
2072 i
= pageblock_nr_pages
;
2075 set_page_refcounted(p
);
2076 __free_pages(p
, MAX_ORDER
- 1);
2077 p
+= MAX_ORDER_NR_PAGES
;
2078 } while (i
-= MAX_ORDER_NR_PAGES
);
2080 set_page_refcounted(page
);
2081 __free_pages(page
, pageblock_order
);
2084 adjust_managed_page_count(page
, pageblock_nr_pages
);
2089 * The order of subdivision here is critical for the IO subsystem.
2090 * Please do not alter this order without good reasons and regression
2091 * testing. Specifically, as large blocks of memory are subdivided,
2092 * the order in which smaller blocks are delivered depends on the order
2093 * they're subdivided in this function. This is the primary factor
2094 * influencing the order in which pages are delivered to the IO
2095 * subsystem according to empirical testing, and this is also justified
2096 * by considering the behavior of a buddy system containing a single
2097 * large block of memory acted on by a series of small allocations.
2098 * This behavior is a critical factor in sglist merging's success.
2102 static inline void expand(struct zone
*zone
, struct page
*page
,
2103 int low
, int high
, int migratetype
)
2105 unsigned long size
= 1 << high
;
2107 while (high
> low
) {
2110 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2113 * Mark as guard pages (or page), that will allow to
2114 * merge back to allocator when buddy will be freed.
2115 * Corresponding page table entries will not be touched,
2116 * pages will stay not present in virtual address space
2118 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2121 add_to_free_list(&page
[size
], zone
, high
, migratetype
);
2122 set_page_order(&page
[size
], high
);
2126 static void check_new_page_bad(struct page
*page
)
2128 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2129 /* Don't complain about hwpoisoned pages */
2130 page_mapcount_reset(page
); /* remove PageBuddy */
2135 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_PREP
));
2139 * This page is about to be returned from the page allocator
2141 static inline int check_new_page(struct page
*page
)
2143 if (likely(page_expected_state(page
,
2144 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2147 check_new_page_bad(page
);
2151 static inline bool free_pages_prezeroed(void)
2153 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
2154 page_poisoning_enabled()) || want_init_on_free();
2157 #ifdef CONFIG_DEBUG_VM
2159 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2160 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2161 * also checked when pcp lists are refilled from the free lists.
2163 static inline bool check_pcp_refill(struct page
*page
)
2165 if (debug_pagealloc_enabled_static())
2166 return check_new_page(page
);
2171 static inline bool check_new_pcp(struct page
*page
)
2173 return check_new_page(page
);
2177 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2178 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2179 * enabled, they are also checked when being allocated from the pcp lists.
2181 static inline bool check_pcp_refill(struct page
*page
)
2183 return check_new_page(page
);
2185 static inline bool check_new_pcp(struct page
*page
)
2187 if (debug_pagealloc_enabled_static())
2188 return check_new_page(page
);
2192 #endif /* CONFIG_DEBUG_VM */
2194 static bool check_new_pages(struct page
*page
, unsigned int order
)
2197 for (i
= 0; i
< (1 << order
); i
++) {
2198 struct page
*p
= page
+ i
;
2200 if (unlikely(check_new_page(p
)))
2207 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2210 set_page_private(page
, 0);
2211 set_page_refcounted(page
);
2213 arch_alloc_page(page
, order
);
2214 if (debug_pagealloc_enabled_static())
2215 kernel_map_pages(page
, 1 << order
, 1);
2216 kasan_alloc_pages(page
, order
);
2217 kernel_poison_pages(page
, 1 << order
, 1);
2218 set_page_owner(page
, order
, gfp_flags
);
2221 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2222 unsigned int alloc_flags
)
2224 post_alloc_hook(page
, order
, gfp_flags
);
2226 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags
))
2227 kernel_init_free_pages(page
, 1 << order
);
2229 if (order
&& (gfp_flags
& __GFP_COMP
))
2230 prep_compound_page(page
, order
);
2233 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2234 * allocate the page. The expectation is that the caller is taking
2235 * steps that will free more memory. The caller should avoid the page
2236 * being used for !PFMEMALLOC purposes.
2238 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2239 set_page_pfmemalloc(page
);
2241 clear_page_pfmemalloc(page
);
2245 * Go through the free lists for the given migratetype and remove
2246 * the smallest available page from the freelists
2248 static __always_inline
2249 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2252 unsigned int current_order
;
2253 struct free_area
*area
;
2256 /* Find a page of the appropriate size in the preferred list */
2257 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2258 area
= &(zone
->free_area
[current_order
]);
2259 page
= get_page_from_free_area(area
, migratetype
);
2262 del_page_from_free_list(page
, zone
, current_order
);
2263 expand(zone
, page
, order
, current_order
, migratetype
);
2264 set_pcppage_migratetype(page
, migratetype
);
2273 * This array describes the order lists are fallen back to when
2274 * the free lists for the desirable migrate type are depleted
2276 static int fallbacks
[MIGRATE_TYPES
][4] = {
2277 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2278 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2279 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2281 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2283 #ifdef CONFIG_MEMORY_ISOLATION
2284 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2289 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2292 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2295 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2296 unsigned int order
) { return NULL
; }
2300 * Move the free pages in a range to the free lists of the requested type.
2301 * Note that start_page and end_pages are not aligned on a pageblock
2302 * boundary. If alignment is required, use move_freepages_block()
2304 static int move_freepages(struct zone
*zone
,
2305 struct page
*start_page
, struct page
*end_page
,
2306 int migratetype
, int *num_movable
)
2310 int pages_moved
= 0;
2312 for (page
= start_page
; page
<= end_page
;) {
2313 if (!pfn_valid_within(page_to_pfn(page
))) {
2318 if (!PageBuddy(page
)) {
2320 * We assume that pages that could be isolated for
2321 * migration are movable. But we don't actually try
2322 * isolating, as that would be expensive.
2325 (PageLRU(page
) || __PageMovable(page
)))
2332 /* Make sure we are not inadvertently changing nodes */
2333 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2334 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2336 order
= page_order(page
);
2337 move_to_free_list(page
, zone
, order
, migratetype
);
2339 pages_moved
+= 1 << order
;
2345 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2346 int migratetype
, int *num_movable
)
2348 unsigned long start_pfn
, end_pfn
;
2349 struct page
*start_page
, *end_page
;
2354 start_pfn
= page_to_pfn(page
);
2355 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2356 start_page
= pfn_to_page(start_pfn
);
2357 end_page
= start_page
+ pageblock_nr_pages
- 1;
2358 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2360 /* Do not cross zone boundaries */
2361 if (!zone_spans_pfn(zone
, start_pfn
))
2363 if (!zone_spans_pfn(zone
, end_pfn
))
2366 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2370 static void change_pageblock_range(struct page
*pageblock_page
,
2371 int start_order
, int migratetype
)
2373 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2375 while (nr_pageblocks
--) {
2376 set_pageblock_migratetype(pageblock_page
, migratetype
);
2377 pageblock_page
+= pageblock_nr_pages
;
2382 * When we are falling back to another migratetype during allocation, try to
2383 * steal extra free pages from the same pageblocks to satisfy further
2384 * allocations, instead of polluting multiple pageblocks.
2386 * If we are stealing a relatively large buddy page, it is likely there will
2387 * be more free pages in the pageblock, so try to steal them all. For
2388 * reclaimable and unmovable allocations, we steal regardless of page size,
2389 * as fragmentation caused by those allocations polluting movable pageblocks
2390 * is worse than movable allocations stealing from unmovable and reclaimable
2393 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2396 * Leaving this order check is intended, although there is
2397 * relaxed order check in next check. The reason is that
2398 * we can actually steal whole pageblock if this condition met,
2399 * but, below check doesn't guarantee it and that is just heuristic
2400 * so could be changed anytime.
2402 if (order
>= pageblock_order
)
2405 if (order
>= pageblock_order
/ 2 ||
2406 start_mt
== MIGRATE_RECLAIMABLE
||
2407 start_mt
== MIGRATE_UNMOVABLE
||
2408 page_group_by_mobility_disabled
)
2414 static inline void boost_watermark(struct zone
*zone
)
2416 unsigned long max_boost
;
2418 if (!watermark_boost_factor
)
2421 * Don't bother in zones that are unlikely to produce results.
2422 * On small machines, including kdump capture kernels running
2423 * in a small area, boosting the watermark can cause an out of
2424 * memory situation immediately.
2426 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
2429 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2430 watermark_boost_factor
, 10000);
2433 * high watermark may be uninitialised if fragmentation occurs
2434 * very early in boot so do not boost. We do not fall
2435 * through and boost by pageblock_nr_pages as failing
2436 * allocations that early means that reclaim is not going
2437 * to help and it may even be impossible to reclaim the
2438 * boosted watermark resulting in a hang.
2443 max_boost
= max(pageblock_nr_pages
, max_boost
);
2445 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2450 * This function implements actual steal behaviour. If order is large enough,
2451 * we can steal whole pageblock. If not, we first move freepages in this
2452 * pageblock to our migratetype and determine how many already-allocated pages
2453 * are there in the pageblock with a compatible migratetype. If at least half
2454 * of pages are free or compatible, we can change migratetype of the pageblock
2455 * itself, so pages freed in the future will be put on the correct free list.
2457 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2458 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2460 unsigned int current_order
= page_order(page
);
2461 int free_pages
, movable_pages
, alike_pages
;
2464 old_block_type
= get_pageblock_migratetype(page
);
2467 * This can happen due to races and we want to prevent broken
2468 * highatomic accounting.
2470 if (is_migrate_highatomic(old_block_type
))
2473 /* Take ownership for orders >= pageblock_order */
2474 if (current_order
>= pageblock_order
) {
2475 change_pageblock_range(page
, current_order
, start_type
);
2480 * Boost watermarks to increase reclaim pressure to reduce the
2481 * likelihood of future fallbacks. Wake kswapd now as the node
2482 * may be balanced overall and kswapd will not wake naturally.
2484 boost_watermark(zone
);
2485 if (alloc_flags
& ALLOC_KSWAPD
)
2486 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2488 /* We are not allowed to try stealing from the whole block */
2492 free_pages
= move_freepages_block(zone
, page
, start_type
,
2495 * Determine how many pages are compatible with our allocation.
2496 * For movable allocation, it's the number of movable pages which
2497 * we just obtained. For other types it's a bit more tricky.
2499 if (start_type
== MIGRATE_MOVABLE
) {
2500 alike_pages
= movable_pages
;
2503 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2504 * to MOVABLE pageblock, consider all non-movable pages as
2505 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2506 * vice versa, be conservative since we can't distinguish the
2507 * exact migratetype of non-movable pages.
2509 if (old_block_type
== MIGRATE_MOVABLE
)
2510 alike_pages
= pageblock_nr_pages
2511 - (free_pages
+ movable_pages
);
2516 /* moving whole block can fail due to zone boundary conditions */
2521 * If a sufficient number of pages in the block are either free or of
2522 * comparable migratability as our allocation, claim the whole block.
2524 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2525 page_group_by_mobility_disabled
)
2526 set_pageblock_migratetype(page
, start_type
);
2531 move_to_free_list(page
, zone
, current_order
, start_type
);
2535 * Check whether there is a suitable fallback freepage with requested order.
2536 * If only_stealable is true, this function returns fallback_mt only if
2537 * we can steal other freepages all together. This would help to reduce
2538 * fragmentation due to mixed migratetype pages in one pageblock.
2540 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2541 int migratetype
, bool only_stealable
, bool *can_steal
)
2546 if (area
->nr_free
== 0)
2551 fallback_mt
= fallbacks
[migratetype
][i
];
2552 if (fallback_mt
== MIGRATE_TYPES
)
2555 if (free_area_empty(area
, fallback_mt
))
2558 if (can_steal_fallback(order
, migratetype
))
2561 if (!only_stealable
)
2572 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2573 * there are no empty page blocks that contain a page with a suitable order
2575 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2576 unsigned int alloc_order
)
2579 unsigned long max_managed
, flags
;
2582 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2583 * Check is race-prone but harmless.
2585 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2586 if (zone
->nr_reserved_highatomic
>= max_managed
)
2589 spin_lock_irqsave(&zone
->lock
, flags
);
2591 /* Recheck the nr_reserved_highatomic limit under the lock */
2592 if (zone
->nr_reserved_highatomic
>= max_managed
)
2596 mt
= get_pageblock_migratetype(page
);
2597 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2598 && !is_migrate_cma(mt
)) {
2599 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2600 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2601 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2605 spin_unlock_irqrestore(&zone
->lock
, flags
);
2609 * Used when an allocation is about to fail under memory pressure. This
2610 * potentially hurts the reliability of high-order allocations when under
2611 * intense memory pressure but failed atomic allocations should be easier
2612 * to recover from than an OOM.
2614 * If @force is true, try to unreserve a pageblock even though highatomic
2615 * pageblock is exhausted.
2617 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2620 struct zonelist
*zonelist
= ac
->zonelist
;
2621 unsigned long flags
;
2628 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->highest_zoneidx
,
2631 * Preserve at least one pageblock unless memory pressure
2634 if (!force
&& zone
->nr_reserved_highatomic
<=
2638 spin_lock_irqsave(&zone
->lock
, flags
);
2639 for (order
= 0; order
< MAX_ORDER
; order
++) {
2640 struct free_area
*area
= &(zone
->free_area
[order
]);
2642 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2647 * In page freeing path, migratetype change is racy so
2648 * we can counter several free pages in a pageblock
2649 * in this loop althoug we changed the pageblock type
2650 * from highatomic to ac->migratetype. So we should
2651 * adjust the count once.
2653 if (is_migrate_highatomic_page(page
)) {
2655 * It should never happen but changes to
2656 * locking could inadvertently allow a per-cpu
2657 * drain to add pages to MIGRATE_HIGHATOMIC
2658 * while unreserving so be safe and watch for
2661 zone
->nr_reserved_highatomic
-= min(
2663 zone
->nr_reserved_highatomic
);
2667 * Convert to ac->migratetype and avoid the normal
2668 * pageblock stealing heuristics. Minimally, the caller
2669 * is doing the work and needs the pages. More
2670 * importantly, if the block was always converted to
2671 * MIGRATE_UNMOVABLE or another type then the number
2672 * of pageblocks that cannot be completely freed
2675 set_pageblock_migratetype(page
, ac
->migratetype
);
2676 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2679 spin_unlock_irqrestore(&zone
->lock
, flags
);
2683 spin_unlock_irqrestore(&zone
->lock
, flags
);
2690 * Try finding a free buddy page on the fallback list and put it on the free
2691 * list of requested migratetype, possibly along with other pages from the same
2692 * block, depending on fragmentation avoidance heuristics. Returns true if
2693 * fallback was found so that __rmqueue_smallest() can grab it.
2695 * The use of signed ints for order and current_order is a deliberate
2696 * deviation from the rest of this file, to make the for loop
2697 * condition simpler.
2699 static __always_inline
bool
2700 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2701 unsigned int alloc_flags
)
2703 struct free_area
*area
;
2705 int min_order
= order
;
2711 * Do not steal pages from freelists belonging to other pageblocks
2712 * i.e. orders < pageblock_order. If there are no local zones free,
2713 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2715 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2716 min_order
= pageblock_order
;
2719 * Find the largest available free page in the other list. This roughly
2720 * approximates finding the pageblock with the most free pages, which
2721 * would be too costly to do exactly.
2723 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2725 area
= &(zone
->free_area
[current_order
]);
2726 fallback_mt
= find_suitable_fallback(area
, current_order
,
2727 start_migratetype
, false, &can_steal
);
2728 if (fallback_mt
== -1)
2732 * We cannot steal all free pages from the pageblock and the
2733 * requested migratetype is movable. In that case it's better to
2734 * steal and split the smallest available page instead of the
2735 * largest available page, because even if the next movable
2736 * allocation falls back into a different pageblock than this
2737 * one, it won't cause permanent fragmentation.
2739 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2740 && current_order
> order
)
2749 for (current_order
= order
; current_order
< MAX_ORDER
;
2751 area
= &(zone
->free_area
[current_order
]);
2752 fallback_mt
= find_suitable_fallback(area
, current_order
,
2753 start_migratetype
, false, &can_steal
);
2754 if (fallback_mt
!= -1)
2759 * This should not happen - we already found a suitable fallback
2760 * when looking for the largest page.
2762 VM_BUG_ON(current_order
== MAX_ORDER
);
2765 page
= get_page_from_free_area(area
, fallback_mt
);
2767 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2770 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2771 start_migratetype
, fallback_mt
);
2778 * Do the hard work of removing an element from the buddy allocator.
2779 * Call me with the zone->lock already held.
2781 static __always_inline
struct page
*
2782 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2783 unsigned int alloc_flags
)
2789 * Balance movable allocations between regular and CMA areas by
2790 * allocating from CMA when over half of the zone's free memory
2791 * is in the CMA area.
2793 if (migratetype
== MIGRATE_MOVABLE
&&
2794 zone_page_state(zone
, NR_FREE_CMA_PAGES
) >
2795 zone_page_state(zone
, NR_FREE_PAGES
) / 2) {
2796 page
= __rmqueue_cma_fallback(zone
, order
);
2802 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2803 if (unlikely(!page
)) {
2804 if (migratetype
== MIGRATE_MOVABLE
)
2805 page
= __rmqueue_cma_fallback(zone
, order
);
2807 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2812 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2817 * Obtain a specified number of elements from the buddy allocator, all under
2818 * a single hold of the lock, for efficiency. Add them to the supplied list.
2819 * Returns the number of new pages which were placed at *list.
2821 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2822 unsigned long count
, struct list_head
*list
,
2823 int migratetype
, unsigned int alloc_flags
)
2827 spin_lock(&zone
->lock
);
2828 for (i
= 0; i
< count
; ++i
) {
2829 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2831 if (unlikely(page
== NULL
))
2834 if (unlikely(check_pcp_refill(page
)))
2838 * Split buddy pages returned by expand() are received here in
2839 * physical page order. The page is added to the tail of
2840 * caller's list. From the callers perspective, the linked list
2841 * is ordered by page number under some conditions. This is
2842 * useful for IO devices that can forward direction from the
2843 * head, thus also in the physical page order. This is useful
2844 * for IO devices that can merge IO requests if the physical
2845 * pages are ordered properly.
2847 list_add_tail(&page
->lru
, list
);
2849 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2850 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2855 * i pages were removed from the buddy list even if some leak due
2856 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2857 * on i. Do not confuse with 'alloced' which is the number of
2858 * pages added to the pcp list.
2860 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2861 spin_unlock(&zone
->lock
);
2867 * Called from the vmstat counter updater to drain pagesets of this
2868 * currently executing processor on remote nodes after they have
2871 * Note that this function must be called with the thread pinned to
2872 * a single processor.
2874 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2876 unsigned long flags
;
2877 int to_drain
, batch
;
2879 local_irq_save(flags
);
2880 batch
= READ_ONCE(pcp
->batch
);
2881 to_drain
= min(pcp
->count
, batch
);
2883 free_pcppages_bulk(zone
, to_drain
, pcp
);
2884 local_irq_restore(flags
);
2889 * Drain pcplists of the indicated processor and zone.
2891 * The processor must either be the current processor and the
2892 * thread pinned to the current processor or a processor that
2895 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2897 unsigned long flags
;
2898 struct per_cpu_pageset
*pset
;
2899 struct per_cpu_pages
*pcp
;
2901 local_irq_save(flags
);
2902 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2906 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2907 local_irq_restore(flags
);
2911 * Drain pcplists of all zones on the indicated processor.
2913 * The processor must either be the current processor and the
2914 * thread pinned to the current processor or a processor that
2917 static void drain_pages(unsigned int cpu
)
2921 for_each_populated_zone(zone
) {
2922 drain_pages_zone(cpu
, zone
);
2927 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2929 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2930 * the single zone's pages.
2932 void drain_local_pages(struct zone
*zone
)
2934 int cpu
= smp_processor_id();
2937 drain_pages_zone(cpu
, zone
);
2942 static void drain_local_pages_wq(struct work_struct
*work
)
2944 struct pcpu_drain
*drain
;
2946 drain
= container_of(work
, struct pcpu_drain
, work
);
2949 * drain_all_pages doesn't use proper cpu hotplug protection so
2950 * we can race with cpu offline when the WQ can move this from
2951 * a cpu pinned worker to an unbound one. We can operate on a different
2952 * cpu which is allright but we also have to make sure to not move to
2956 drain_local_pages(drain
->zone
);
2961 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2963 * When zone parameter is non-NULL, spill just the single zone's pages.
2965 * Note that this can be extremely slow as the draining happens in a workqueue.
2967 void drain_all_pages(struct zone
*zone
)
2972 * Allocate in the BSS so we wont require allocation in
2973 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2975 static cpumask_t cpus_with_pcps
;
2978 * Make sure nobody triggers this path before mm_percpu_wq is fully
2981 if (WARN_ON_ONCE(!mm_percpu_wq
))
2985 * Do not drain if one is already in progress unless it's specific to
2986 * a zone. Such callers are primarily CMA and memory hotplug and need
2987 * the drain to be complete when the call returns.
2989 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2992 mutex_lock(&pcpu_drain_mutex
);
2996 * We don't care about racing with CPU hotplug event
2997 * as offline notification will cause the notified
2998 * cpu to drain that CPU pcps and on_each_cpu_mask
2999 * disables preemption as part of its processing
3001 for_each_online_cpu(cpu
) {
3002 struct per_cpu_pageset
*pcp
;
3004 bool has_pcps
= false;
3007 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3011 for_each_populated_zone(z
) {
3012 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
3013 if (pcp
->pcp
.count
) {
3021 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
3023 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
3026 for_each_cpu(cpu
, &cpus_with_pcps
) {
3027 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
3030 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
3031 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
3033 for_each_cpu(cpu
, &cpus_with_pcps
)
3034 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
3036 mutex_unlock(&pcpu_drain_mutex
);
3039 #ifdef CONFIG_HIBERNATION
3042 * Touch the watchdog for every WD_PAGE_COUNT pages.
3044 #define WD_PAGE_COUNT (128*1024)
3046 void mark_free_pages(struct zone
*zone
)
3048 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
3049 unsigned long flags
;
3050 unsigned int order
, t
;
3053 if (zone_is_empty(zone
))
3056 spin_lock_irqsave(&zone
->lock
, flags
);
3058 max_zone_pfn
= zone_end_pfn(zone
);
3059 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
3060 if (pfn_valid(pfn
)) {
3061 page
= pfn_to_page(pfn
);
3063 if (!--page_count
) {
3064 touch_nmi_watchdog();
3065 page_count
= WD_PAGE_COUNT
;
3068 if (page_zone(page
) != zone
)
3071 if (!swsusp_page_is_forbidden(page
))
3072 swsusp_unset_page_free(page
);
3075 for_each_migratetype_order(order
, t
) {
3076 list_for_each_entry(page
,
3077 &zone
->free_area
[order
].free_list
[t
], lru
) {
3080 pfn
= page_to_pfn(page
);
3081 for (i
= 0; i
< (1UL << order
); i
++) {
3082 if (!--page_count
) {
3083 touch_nmi_watchdog();
3084 page_count
= WD_PAGE_COUNT
;
3086 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
3090 spin_unlock_irqrestore(&zone
->lock
, flags
);
3092 #endif /* CONFIG_PM */
3094 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
3098 if (!free_pcp_prepare(page
))
3101 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3102 set_pcppage_migratetype(page
, migratetype
);
3106 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
3108 struct zone
*zone
= page_zone(page
);
3109 struct per_cpu_pages
*pcp
;
3112 migratetype
= get_pcppage_migratetype(page
);
3113 __count_vm_event(PGFREE
);
3116 * We only track unmovable, reclaimable and movable on pcp lists.
3117 * Free ISOLATE pages back to the allocator because they are being
3118 * offlined but treat HIGHATOMIC as movable pages so we can get those
3119 * areas back if necessary. Otherwise, we may have to free
3120 * excessively into the page allocator
3122 if (migratetype
>= MIGRATE_PCPTYPES
) {
3123 if (unlikely(is_migrate_isolate(migratetype
))) {
3124 free_one_page(zone
, page
, pfn
, 0, migratetype
);
3127 migratetype
= MIGRATE_MOVABLE
;
3130 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3131 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
3133 if (pcp
->count
>= pcp
->high
) {
3134 unsigned long batch
= READ_ONCE(pcp
->batch
);
3135 free_pcppages_bulk(zone
, batch
, pcp
);
3140 * Free a 0-order page
3142 void free_unref_page(struct page
*page
)
3144 unsigned long flags
;
3145 unsigned long pfn
= page_to_pfn(page
);
3147 if (!free_unref_page_prepare(page
, pfn
))
3150 local_irq_save(flags
);
3151 free_unref_page_commit(page
, pfn
);
3152 local_irq_restore(flags
);
3156 * Free a list of 0-order pages
3158 void free_unref_page_list(struct list_head
*list
)
3160 struct page
*page
, *next
;
3161 unsigned long flags
, pfn
;
3162 int batch_count
= 0;
3164 /* Prepare pages for freeing */
3165 list_for_each_entry_safe(page
, next
, list
, lru
) {
3166 pfn
= page_to_pfn(page
);
3167 if (!free_unref_page_prepare(page
, pfn
))
3168 list_del(&page
->lru
);
3169 set_page_private(page
, pfn
);
3172 local_irq_save(flags
);
3173 list_for_each_entry_safe(page
, next
, list
, lru
) {
3174 unsigned long pfn
= page_private(page
);
3176 set_page_private(page
, 0);
3177 trace_mm_page_free_batched(page
);
3178 free_unref_page_commit(page
, pfn
);
3181 * Guard against excessive IRQ disabled times when we get
3182 * a large list of pages to free.
3184 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3185 local_irq_restore(flags
);
3187 local_irq_save(flags
);
3190 local_irq_restore(flags
);
3194 * split_page takes a non-compound higher-order page, and splits it into
3195 * n (1<<order) sub-pages: page[0..n]
3196 * Each sub-page must be freed individually.
3198 * Note: this is probably too low level an operation for use in drivers.
3199 * Please consult with lkml before using this in your driver.
3201 void split_page(struct page
*page
, unsigned int order
)
3205 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3206 VM_BUG_ON_PAGE(!page_count(page
), page
);
3208 for (i
= 1; i
< (1 << order
); i
++)
3209 set_page_refcounted(page
+ i
);
3210 split_page_owner(page
, order
);
3212 EXPORT_SYMBOL_GPL(split_page
);
3214 int __isolate_free_page(struct page
*page
, unsigned int order
)
3216 unsigned long watermark
;
3220 BUG_ON(!PageBuddy(page
));
3222 zone
= page_zone(page
);
3223 mt
= get_pageblock_migratetype(page
);
3225 if (!is_migrate_isolate(mt
)) {
3227 * Obey watermarks as if the page was being allocated. We can
3228 * emulate a high-order watermark check with a raised order-0
3229 * watermark, because we already know our high-order page
3232 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3233 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3236 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3239 /* Remove page from free list */
3241 del_page_from_free_list(page
, zone
, order
);
3244 * Set the pageblock if the isolated page is at least half of a
3247 if (order
>= pageblock_order
- 1) {
3248 struct page
*endpage
= page
+ (1 << order
) - 1;
3249 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3250 int mt
= get_pageblock_migratetype(page
);
3251 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3252 && !is_migrate_highatomic(mt
))
3253 set_pageblock_migratetype(page
,
3259 return 1UL << order
;
3263 * __putback_isolated_page - Return a now-isolated page back where we got it
3264 * @page: Page that was isolated
3265 * @order: Order of the isolated page
3266 * @mt: The page's pageblock's migratetype
3268 * This function is meant to return a page pulled from the free lists via
3269 * __isolate_free_page back to the free lists they were pulled from.
3271 void __putback_isolated_page(struct page
*page
, unsigned int order
, int mt
)
3273 struct zone
*zone
= page_zone(page
);
3275 /* zone lock should be held when this function is called */
3276 lockdep_assert_held(&zone
->lock
);
3278 /* Return isolated page to tail of freelist. */
3279 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
, false);
3283 * Update NUMA hit/miss statistics
3285 * Must be called with interrupts disabled.
3287 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
3290 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3292 /* skip numa counters update if numa stats is disabled */
3293 if (!static_branch_likely(&vm_numa_stat_key
))
3296 if (zone_to_nid(z
) != numa_node_id())
3297 local_stat
= NUMA_OTHER
;
3299 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3300 __inc_numa_state(z
, NUMA_HIT
);
3302 __inc_numa_state(z
, NUMA_MISS
);
3303 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3305 __inc_numa_state(z
, local_stat
);
3309 /* Remove page from the per-cpu list, caller must protect the list */
3310 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3311 unsigned int alloc_flags
,
3312 struct per_cpu_pages
*pcp
,
3313 struct list_head
*list
)
3318 if (list_empty(list
)) {
3319 pcp
->count
+= rmqueue_bulk(zone
, 0,
3321 migratetype
, alloc_flags
);
3322 if (unlikely(list_empty(list
)))
3326 page
= list_first_entry(list
, struct page
, lru
);
3327 list_del(&page
->lru
);
3329 } while (check_new_pcp(page
));
3334 /* Lock and remove page from the per-cpu list */
3335 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3336 struct zone
*zone
, gfp_t gfp_flags
,
3337 int migratetype
, unsigned int alloc_flags
)
3339 struct per_cpu_pages
*pcp
;
3340 struct list_head
*list
;
3342 unsigned long flags
;
3344 local_irq_save(flags
);
3345 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3346 list
= &pcp
->lists
[migratetype
];
3347 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3349 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3350 zone_statistics(preferred_zone
, zone
);
3352 local_irq_restore(flags
);
3357 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3360 struct page
*rmqueue(struct zone
*preferred_zone
,
3361 struct zone
*zone
, unsigned int order
,
3362 gfp_t gfp_flags
, unsigned int alloc_flags
,
3365 unsigned long flags
;
3368 if (likely(order
== 0)) {
3369 page
= rmqueue_pcplist(preferred_zone
, zone
, gfp_flags
,
3370 migratetype
, alloc_flags
);
3375 * We most definitely don't want callers attempting to
3376 * allocate greater than order-1 page units with __GFP_NOFAIL.
3378 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3379 spin_lock_irqsave(&zone
->lock
, flags
);
3383 if (alloc_flags
& ALLOC_HARDER
) {
3384 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3386 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3389 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3390 } while (page
&& check_new_pages(page
, order
));
3391 spin_unlock(&zone
->lock
);
3394 __mod_zone_freepage_state(zone
, -(1 << order
),
3395 get_pcppage_migratetype(page
));
3397 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3398 zone_statistics(preferred_zone
, zone
);
3399 local_irq_restore(flags
);
3402 /* Separate test+clear to avoid unnecessary atomics */
3403 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3404 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3405 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3408 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3412 local_irq_restore(flags
);
3416 #ifdef CONFIG_FAIL_PAGE_ALLOC
3419 struct fault_attr attr
;
3421 bool ignore_gfp_highmem
;
3422 bool ignore_gfp_reclaim
;
3424 } fail_page_alloc
= {
3425 .attr
= FAULT_ATTR_INITIALIZER
,
3426 .ignore_gfp_reclaim
= true,
3427 .ignore_gfp_highmem
= true,
3431 static int __init
setup_fail_page_alloc(char *str
)
3433 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3435 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3437 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3439 if (order
< fail_page_alloc
.min_order
)
3441 if (gfp_mask
& __GFP_NOFAIL
)
3443 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3445 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3446 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3449 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3452 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3454 static int __init
fail_page_alloc_debugfs(void)
3456 umode_t mode
= S_IFREG
| 0600;
3459 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3460 &fail_page_alloc
.attr
);
3462 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3463 &fail_page_alloc
.ignore_gfp_reclaim
);
3464 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3465 &fail_page_alloc
.ignore_gfp_highmem
);
3466 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3471 late_initcall(fail_page_alloc_debugfs
);
3473 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3475 #else /* CONFIG_FAIL_PAGE_ALLOC */
3477 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3482 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3484 static noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3486 return __should_fail_alloc_page(gfp_mask
, order
);
3488 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3491 * Return true if free base pages are above 'mark'. For high-order checks it
3492 * will return true of the order-0 watermark is reached and there is at least
3493 * one free page of a suitable size. Checking now avoids taking the zone lock
3494 * to check in the allocation paths if no pages are free.
3496 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3497 int highest_zoneidx
, unsigned int alloc_flags
,
3502 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3504 /* free_pages may go negative - that's OK */
3505 free_pages
-= (1 << order
) - 1;
3507 if (alloc_flags
& ALLOC_HIGH
)
3511 * If the caller does not have rights to ALLOC_HARDER then subtract
3512 * the high-atomic reserves. This will over-estimate the size of the
3513 * atomic reserve but it avoids a search.
3515 if (likely(!alloc_harder
)) {
3516 free_pages
-= z
->nr_reserved_highatomic
;
3519 * OOM victims can try even harder than normal ALLOC_HARDER
3520 * users on the grounds that it's definitely going to be in
3521 * the exit path shortly and free memory. Any allocation it
3522 * makes during the free path will be small and short-lived.
3524 if (alloc_flags
& ALLOC_OOM
)
3532 /* If allocation can't use CMA areas don't use free CMA pages */
3533 if (!(alloc_flags
& ALLOC_CMA
))
3534 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3538 * Check watermarks for an order-0 allocation request. If these
3539 * are not met, then a high-order request also cannot go ahead
3540 * even if a suitable page happened to be free.
3542 if (free_pages
<= min
+ z
->lowmem_reserve
[highest_zoneidx
])
3545 /* If this is an order-0 request then the watermark is fine */
3549 /* For a high-order request, check at least one suitable page is free */
3550 for (o
= order
; o
< MAX_ORDER
; o
++) {
3551 struct free_area
*area
= &z
->free_area
[o
];
3557 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3558 if (!free_area_empty(area
, mt
))
3563 if ((alloc_flags
& ALLOC_CMA
) &&
3564 !free_area_empty(area
, MIGRATE_CMA
)) {
3568 if (alloc_harder
&& !free_area_empty(area
, MIGRATE_HIGHATOMIC
))
3574 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3575 int highest_zoneidx
, unsigned int alloc_flags
)
3577 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3578 zone_page_state(z
, NR_FREE_PAGES
));
3581 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3582 unsigned long mark
, int highest_zoneidx
,
3583 unsigned int alloc_flags
)
3585 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3589 /* If allocation can't use CMA areas don't use free CMA pages */
3590 if (!(alloc_flags
& ALLOC_CMA
))
3591 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3595 * Fast check for order-0 only. If this fails then the reserves
3596 * need to be calculated. There is a corner case where the check
3597 * passes but only the high-order atomic reserve are free. If
3598 * the caller is !atomic then it'll uselessly search the free
3599 * list. That corner case is then slower but it is harmless.
3601 if (!order
&& (free_pages
- cma_pages
) >
3602 mark
+ z
->lowmem_reserve
[highest_zoneidx
])
3605 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3609 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3610 unsigned long mark
, int highest_zoneidx
)
3612 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3614 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3615 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3617 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, 0,
3622 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3624 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3625 node_reclaim_distance
;
3627 #else /* CONFIG_NUMA */
3628 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3632 #endif /* CONFIG_NUMA */
3635 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3636 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3637 * premature use of a lower zone may cause lowmem pressure problems that
3638 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3639 * probably too small. It only makes sense to spread allocations to avoid
3640 * fragmentation between the Normal and DMA32 zones.
3642 static inline unsigned int
3643 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3645 unsigned int alloc_flags
;
3648 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3651 alloc_flags
= (__force
int) (gfp_mask
& __GFP_KSWAPD_RECLAIM
);
3653 #ifdef CONFIG_ZONE_DMA32
3657 if (zone_idx(zone
) != ZONE_NORMAL
)
3661 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3662 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3663 * on UMA that if Normal is populated then so is DMA32.
3665 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3666 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3669 alloc_flags
|= ALLOC_NOFRAGMENT
;
3670 #endif /* CONFIG_ZONE_DMA32 */
3675 * get_page_from_freelist goes through the zonelist trying to allocate
3678 static struct page
*
3679 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3680 const struct alloc_context
*ac
)
3684 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3689 * Scan zonelist, looking for a zone with enough free.
3690 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3692 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3693 z
= ac
->preferred_zoneref
;
3694 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3695 ac
->highest_zoneidx
, ac
->nodemask
) {
3699 if (cpusets_enabled() &&
3700 (alloc_flags
& ALLOC_CPUSET
) &&
3701 !__cpuset_zone_allowed(zone
, gfp_mask
))
3704 * When allocating a page cache page for writing, we
3705 * want to get it from a node that is within its dirty
3706 * limit, such that no single node holds more than its
3707 * proportional share of globally allowed dirty pages.
3708 * The dirty limits take into account the node's
3709 * lowmem reserves and high watermark so that kswapd
3710 * should be able to balance it without having to
3711 * write pages from its LRU list.
3713 * XXX: For now, allow allocations to potentially
3714 * exceed the per-node dirty limit in the slowpath
3715 * (spread_dirty_pages unset) before going into reclaim,
3716 * which is important when on a NUMA setup the allowed
3717 * nodes are together not big enough to reach the
3718 * global limit. The proper fix for these situations
3719 * will require awareness of nodes in the
3720 * dirty-throttling and the flusher threads.
3722 if (ac
->spread_dirty_pages
) {
3723 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3726 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3727 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3732 if (no_fallback
&& nr_online_nodes
> 1 &&
3733 zone
!= ac
->preferred_zoneref
->zone
) {
3737 * If moving to a remote node, retry but allow
3738 * fragmenting fallbacks. Locality is more important
3739 * than fragmentation avoidance.
3741 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3742 if (zone_to_nid(zone
) != local_nid
) {
3743 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3748 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3749 if (!zone_watermark_fast(zone
, order
, mark
,
3750 ac
->highest_zoneidx
, alloc_flags
)) {
3753 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3755 * Watermark failed for this zone, but see if we can
3756 * grow this zone if it contains deferred pages.
3758 if (static_branch_unlikely(&deferred_pages
)) {
3759 if (_deferred_grow_zone(zone
, order
))
3763 /* Checked here to keep the fast path fast */
3764 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3765 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3768 if (node_reclaim_mode
== 0 ||
3769 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3772 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3774 case NODE_RECLAIM_NOSCAN
:
3777 case NODE_RECLAIM_FULL
:
3778 /* scanned but unreclaimable */
3781 /* did we reclaim enough */
3782 if (zone_watermark_ok(zone
, order
, mark
,
3783 ac
->highest_zoneidx
, alloc_flags
))
3791 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3792 gfp_mask
, alloc_flags
, ac
->migratetype
);
3794 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3797 * If this is a high-order atomic allocation then check
3798 * if the pageblock should be reserved for the future
3800 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3801 reserve_highatomic_pageblock(page
, zone
, order
);
3805 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3806 /* Try again if zone has deferred pages */
3807 if (static_branch_unlikely(&deferred_pages
)) {
3808 if (_deferred_grow_zone(zone
, order
))
3816 * It's possible on a UMA machine to get through all zones that are
3817 * fragmented. If avoiding fragmentation, reset and try again.
3820 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3827 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3829 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3832 * This documents exceptions given to allocations in certain
3833 * contexts that are allowed to allocate outside current's set
3836 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3837 if (tsk_is_oom_victim(current
) ||
3838 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3839 filter
&= ~SHOW_MEM_FILTER_NODES
;
3840 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3841 filter
&= ~SHOW_MEM_FILTER_NODES
;
3843 show_mem(filter
, nodemask
);
3846 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3848 struct va_format vaf
;
3850 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
3852 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3855 va_start(args
, fmt
);
3858 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3859 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3860 nodemask_pr_args(nodemask
));
3863 cpuset_print_current_mems_allowed();
3866 warn_alloc_show_mem(gfp_mask
, nodemask
);
3869 static inline struct page
*
3870 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3871 unsigned int alloc_flags
,
3872 const struct alloc_context
*ac
)
3876 page
= get_page_from_freelist(gfp_mask
, order
,
3877 alloc_flags
|ALLOC_CPUSET
, ac
);
3879 * fallback to ignore cpuset restriction if our nodes
3883 page
= get_page_from_freelist(gfp_mask
, order
,
3889 static inline struct page
*
3890 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3891 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3893 struct oom_control oc
= {
3894 .zonelist
= ac
->zonelist
,
3895 .nodemask
= ac
->nodemask
,
3897 .gfp_mask
= gfp_mask
,
3902 *did_some_progress
= 0;
3905 * Acquire the oom lock. If that fails, somebody else is
3906 * making progress for us.
3908 if (!mutex_trylock(&oom_lock
)) {
3909 *did_some_progress
= 1;
3910 schedule_timeout_uninterruptible(1);
3915 * Go through the zonelist yet one more time, keep very high watermark
3916 * here, this is only to catch a parallel oom killing, we must fail if
3917 * we're still under heavy pressure. But make sure that this reclaim
3918 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3919 * allocation which will never fail due to oom_lock already held.
3921 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3922 ~__GFP_DIRECT_RECLAIM
, order
,
3923 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3927 /* Coredumps can quickly deplete all memory reserves */
3928 if (current
->flags
& PF_DUMPCORE
)
3930 /* The OOM killer will not help higher order allocs */
3931 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3934 * We have already exhausted all our reclaim opportunities without any
3935 * success so it is time to admit defeat. We will skip the OOM killer
3936 * because it is very likely that the caller has a more reasonable
3937 * fallback than shooting a random task.
3939 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3941 /* The OOM killer does not needlessly kill tasks for lowmem */
3942 if (ac
->highest_zoneidx
< ZONE_NORMAL
)
3944 if (pm_suspended_storage())
3947 * XXX: GFP_NOFS allocations should rather fail than rely on
3948 * other request to make a forward progress.
3949 * We are in an unfortunate situation where out_of_memory cannot
3950 * do much for this context but let's try it to at least get
3951 * access to memory reserved if the current task is killed (see
3952 * out_of_memory). Once filesystems are ready to handle allocation
3953 * failures more gracefully we should just bail out here.
3956 /* The OOM killer may not free memory on a specific node */
3957 if (gfp_mask
& __GFP_THISNODE
)
3960 /* Exhausted what can be done so it's blame time */
3961 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3962 *did_some_progress
= 1;
3965 * Help non-failing allocations by giving them access to memory
3968 if (gfp_mask
& __GFP_NOFAIL
)
3969 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3970 ALLOC_NO_WATERMARKS
, ac
);
3973 mutex_unlock(&oom_lock
);
3978 * Maximum number of compaction retries wit a progress before OOM
3979 * killer is consider as the only way to move forward.
3981 #define MAX_COMPACT_RETRIES 16
3983 #ifdef CONFIG_COMPACTION
3984 /* Try memory compaction for high-order allocations before reclaim */
3985 static struct page
*
3986 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3987 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3988 enum compact_priority prio
, enum compact_result
*compact_result
)
3990 struct page
*page
= NULL
;
3991 unsigned long pflags
;
3992 unsigned int noreclaim_flag
;
3997 psi_memstall_enter(&pflags
);
3998 noreclaim_flag
= memalloc_noreclaim_save();
4000 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
4003 memalloc_noreclaim_restore(noreclaim_flag
);
4004 psi_memstall_leave(&pflags
);
4007 * At least in one zone compaction wasn't deferred or skipped, so let's
4008 * count a compaction stall
4010 count_vm_event(COMPACTSTALL
);
4012 /* Prep a captured page if available */
4014 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
4016 /* Try get a page from the freelist if available */
4018 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4021 struct zone
*zone
= page_zone(page
);
4023 zone
->compact_blockskip_flush
= false;
4024 compaction_defer_reset(zone
, order
, true);
4025 count_vm_event(COMPACTSUCCESS
);
4030 * It's bad if compaction run occurs and fails. The most likely reason
4031 * is that pages exist, but not enough to satisfy watermarks.
4033 count_vm_event(COMPACTFAIL
);
4041 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
4042 enum compact_result compact_result
,
4043 enum compact_priority
*compact_priority
,
4044 int *compaction_retries
)
4046 int max_retries
= MAX_COMPACT_RETRIES
;
4049 int retries
= *compaction_retries
;
4050 enum compact_priority priority
= *compact_priority
;
4055 if (compaction_made_progress(compact_result
))
4056 (*compaction_retries
)++;
4059 * compaction considers all the zone as desperately out of memory
4060 * so it doesn't really make much sense to retry except when the
4061 * failure could be caused by insufficient priority
4063 if (compaction_failed(compact_result
))
4064 goto check_priority
;
4067 * compaction was skipped because there are not enough order-0 pages
4068 * to work with, so we retry only if it looks like reclaim can help.
4070 if (compaction_needs_reclaim(compact_result
)) {
4071 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
4076 * make sure the compaction wasn't deferred or didn't bail out early
4077 * due to locks contention before we declare that we should give up.
4078 * But the next retry should use a higher priority if allowed, so
4079 * we don't just keep bailing out endlessly.
4081 if (compaction_withdrawn(compact_result
)) {
4082 goto check_priority
;
4086 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4087 * costly ones because they are de facto nofail and invoke OOM
4088 * killer to move on while costly can fail and users are ready
4089 * to cope with that. 1/4 retries is rather arbitrary but we
4090 * would need much more detailed feedback from compaction to
4091 * make a better decision.
4093 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4095 if (*compaction_retries
<= max_retries
) {
4101 * Make sure there are attempts at the highest priority if we exhausted
4102 * all retries or failed at the lower priorities.
4105 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
4106 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
4108 if (*compact_priority
> min_priority
) {
4109 (*compact_priority
)--;
4110 *compaction_retries
= 0;
4114 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4118 static inline struct page
*
4119 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4120 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4121 enum compact_priority prio
, enum compact_result
*compact_result
)
4123 *compact_result
= COMPACT_SKIPPED
;
4128 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4129 enum compact_result compact_result
,
4130 enum compact_priority
*compact_priority
,
4131 int *compaction_retries
)
4136 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4140 * There are setups with compaction disabled which would prefer to loop
4141 * inside the allocator rather than hit the oom killer prematurely.
4142 * Let's give them a good hope and keep retrying while the order-0
4143 * watermarks are OK.
4145 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4146 ac
->highest_zoneidx
, ac
->nodemask
) {
4147 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4148 ac
->highest_zoneidx
, alloc_flags
))
4153 #endif /* CONFIG_COMPACTION */
4155 #ifdef CONFIG_LOCKDEP
4156 static struct lockdep_map __fs_reclaim_map
=
4157 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4159 static bool __need_fs_reclaim(gfp_t gfp_mask
)
4161 gfp_mask
= current_gfp_context(gfp_mask
);
4163 /* no reclaim without waiting on it */
4164 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4167 /* this guy won't enter reclaim */
4168 if (current
->flags
& PF_MEMALLOC
)
4171 /* We're only interested __GFP_FS allocations for now */
4172 if (!(gfp_mask
& __GFP_FS
))
4175 if (gfp_mask
& __GFP_NOLOCKDEP
)
4181 void __fs_reclaim_acquire(void)
4183 lock_map_acquire(&__fs_reclaim_map
);
4186 void __fs_reclaim_release(void)
4188 lock_map_release(&__fs_reclaim_map
);
4191 void fs_reclaim_acquire(gfp_t gfp_mask
)
4193 if (__need_fs_reclaim(gfp_mask
))
4194 __fs_reclaim_acquire();
4196 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4198 void fs_reclaim_release(gfp_t gfp_mask
)
4200 if (__need_fs_reclaim(gfp_mask
))
4201 __fs_reclaim_release();
4203 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4206 /* Perform direct synchronous page reclaim */
4208 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4209 const struct alloc_context
*ac
)
4212 unsigned int noreclaim_flag
;
4213 unsigned long pflags
;
4217 /* We now go into synchronous reclaim */
4218 cpuset_memory_pressure_bump();
4219 psi_memstall_enter(&pflags
);
4220 fs_reclaim_acquire(gfp_mask
);
4221 noreclaim_flag
= memalloc_noreclaim_save();
4223 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4226 memalloc_noreclaim_restore(noreclaim_flag
);
4227 fs_reclaim_release(gfp_mask
);
4228 psi_memstall_leave(&pflags
);
4235 /* The really slow allocator path where we enter direct reclaim */
4236 static inline struct page
*
4237 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4238 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4239 unsigned long *did_some_progress
)
4241 struct page
*page
= NULL
;
4242 bool drained
= false;
4244 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4245 if (unlikely(!(*did_some_progress
)))
4249 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4252 * If an allocation failed after direct reclaim, it could be because
4253 * pages are pinned on the per-cpu lists or in high alloc reserves.
4254 * Shrink them them and try again
4256 if (!page
&& !drained
) {
4257 unreserve_highatomic_pageblock(ac
, false);
4258 drain_all_pages(NULL
);
4266 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4267 const struct alloc_context
*ac
)
4271 pg_data_t
*last_pgdat
= NULL
;
4272 enum zone_type highest_zoneidx
= ac
->highest_zoneidx
;
4274 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, highest_zoneidx
,
4276 if (last_pgdat
!= zone
->zone_pgdat
)
4277 wakeup_kswapd(zone
, gfp_mask
, order
, highest_zoneidx
);
4278 last_pgdat
= zone
->zone_pgdat
;
4282 static inline unsigned int
4283 gfp_to_alloc_flags(gfp_t gfp_mask
)
4285 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4288 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4289 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4290 * to save two branches.
4292 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4293 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM
!= (__force gfp_t
) ALLOC_KSWAPD
);
4296 * The caller may dip into page reserves a bit more if the caller
4297 * cannot run direct reclaim, or if the caller has realtime scheduling
4298 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4299 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4301 alloc_flags
|= (__force
int)
4302 (gfp_mask
& (__GFP_HIGH
| __GFP_KSWAPD_RECLAIM
));
4304 if (gfp_mask
& __GFP_ATOMIC
) {
4306 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4307 * if it can't schedule.
4309 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4310 alloc_flags
|= ALLOC_HARDER
;
4312 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4313 * comment for __cpuset_node_allowed().
4315 alloc_flags
&= ~ALLOC_CPUSET
;
4316 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4317 alloc_flags
|= ALLOC_HARDER
;
4320 if (gfp_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
4321 alloc_flags
|= ALLOC_CMA
;
4326 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4328 if (!tsk_is_oom_victim(tsk
))
4332 * !MMU doesn't have oom reaper so give access to memory reserves
4333 * only to the thread with TIF_MEMDIE set
4335 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4342 * Distinguish requests which really need access to full memory
4343 * reserves from oom victims which can live with a portion of it
4345 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4347 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4349 if (gfp_mask
& __GFP_MEMALLOC
)
4350 return ALLOC_NO_WATERMARKS
;
4351 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4352 return ALLOC_NO_WATERMARKS
;
4353 if (!in_interrupt()) {
4354 if (current
->flags
& PF_MEMALLOC
)
4355 return ALLOC_NO_WATERMARKS
;
4356 else if (oom_reserves_allowed(current
))
4363 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4365 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4369 * Checks whether it makes sense to retry the reclaim to make a forward progress
4370 * for the given allocation request.
4372 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4373 * without success, or when we couldn't even meet the watermark if we
4374 * reclaimed all remaining pages on the LRU lists.
4376 * Returns true if a retry is viable or false to enter the oom path.
4379 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4380 struct alloc_context
*ac
, int alloc_flags
,
4381 bool did_some_progress
, int *no_progress_loops
)
4388 * Costly allocations might have made a progress but this doesn't mean
4389 * their order will become available due to high fragmentation so
4390 * always increment the no progress counter for them
4392 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4393 *no_progress_loops
= 0;
4395 (*no_progress_loops
)++;
4398 * Make sure we converge to OOM if we cannot make any progress
4399 * several times in the row.
4401 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4402 /* Before OOM, exhaust highatomic_reserve */
4403 return unreserve_highatomic_pageblock(ac
, true);
4407 * Keep reclaiming pages while there is a chance this will lead
4408 * somewhere. If none of the target zones can satisfy our allocation
4409 * request even if all reclaimable pages are considered then we are
4410 * screwed and have to go OOM.
4412 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4413 ac
->highest_zoneidx
, ac
->nodemask
) {
4414 unsigned long available
;
4415 unsigned long reclaimable
;
4416 unsigned long min_wmark
= min_wmark_pages(zone
);
4419 available
= reclaimable
= zone_reclaimable_pages(zone
);
4420 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4423 * Would the allocation succeed if we reclaimed all
4424 * reclaimable pages?
4426 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4427 ac
->highest_zoneidx
, alloc_flags
, available
);
4428 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4429 available
, min_wmark
, *no_progress_loops
, wmark
);
4432 * If we didn't make any progress and have a lot of
4433 * dirty + writeback pages then we should wait for
4434 * an IO to complete to slow down the reclaim and
4435 * prevent from pre mature OOM
4437 if (!did_some_progress
) {
4438 unsigned long write_pending
;
4440 write_pending
= zone_page_state_snapshot(zone
,
4441 NR_ZONE_WRITE_PENDING
);
4443 if (2 * write_pending
> reclaimable
) {
4444 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4456 * Memory allocation/reclaim might be called from a WQ context and the
4457 * current implementation of the WQ concurrency control doesn't
4458 * recognize that a particular WQ is congested if the worker thread is
4459 * looping without ever sleeping. Therefore we have to do a short sleep
4460 * here rather than calling cond_resched().
4462 if (current
->flags
& PF_WQ_WORKER
)
4463 schedule_timeout_uninterruptible(1);
4470 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4473 * It's possible that cpuset's mems_allowed and the nodemask from
4474 * mempolicy don't intersect. This should be normally dealt with by
4475 * policy_nodemask(), but it's possible to race with cpuset update in
4476 * such a way the check therein was true, and then it became false
4477 * before we got our cpuset_mems_cookie here.
4478 * This assumes that for all allocations, ac->nodemask can come only
4479 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4480 * when it does not intersect with the cpuset restrictions) or the
4481 * caller can deal with a violated nodemask.
4483 if (cpusets_enabled() && ac
->nodemask
&&
4484 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4485 ac
->nodemask
= NULL
;
4490 * When updating a task's mems_allowed or mempolicy nodemask, it is
4491 * possible to race with parallel threads in such a way that our
4492 * allocation can fail while the mask is being updated. If we are about
4493 * to fail, check if the cpuset changed during allocation and if so,
4496 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4502 static inline struct page
*
4503 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4504 struct alloc_context
*ac
)
4506 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4507 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4508 struct page
*page
= NULL
;
4509 unsigned int alloc_flags
;
4510 unsigned long did_some_progress
;
4511 enum compact_priority compact_priority
;
4512 enum compact_result compact_result
;
4513 int compaction_retries
;
4514 int no_progress_loops
;
4515 unsigned int cpuset_mems_cookie
;
4519 * We also sanity check to catch abuse of atomic reserves being used by
4520 * callers that are not in atomic context.
4522 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4523 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4524 gfp_mask
&= ~__GFP_ATOMIC
;
4527 compaction_retries
= 0;
4528 no_progress_loops
= 0;
4529 compact_priority
= DEF_COMPACT_PRIORITY
;
4530 cpuset_mems_cookie
= read_mems_allowed_begin();
4533 * The fast path uses conservative alloc_flags to succeed only until
4534 * kswapd needs to be woken up, and to avoid the cost of setting up
4535 * alloc_flags precisely. So we do that now.
4537 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4540 * We need to recalculate the starting point for the zonelist iterator
4541 * because we might have used different nodemask in the fast path, or
4542 * there was a cpuset modification and we are retrying - otherwise we
4543 * could end up iterating over non-eligible zones endlessly.
4545 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4546 ac
->highest_zoneidx
, ac
->nodemask
);
4547 if (!ac
->preferred_zoneref
->zone
)
4550 if (alloc_flags
& ALLOC_KSWAPD
)
4551 wake_all_kswapds(order
, gfp_mask
, ac
);
4554 * The adjusted alloc_flags might result in immediate success, so try
4557 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4562 * For costly allocations, try direct compaction first, as it's likely
4563 * that we have enough base pages and don't need to reclaim. For non-
4564 * movable high-order allocations, do that as well, as compaction will
4565 * try prevent permanent fragmentation by migrating from blocks of the
4567 * Don't try this for allocations that are allowed to ignore
4568 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4570 if (can_direct_reclaim
&&
4572 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4573 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4574 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4576 INIT_COMPACT_PRIORITY
,
4582 * Checks for costly allocations with __GFP_NORETRY, which
4583 * includes some THP page fault allocations
4585 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4587 * If allocating entire pageblock(s) and compaction
4588 * failed because all zones are below low watermarks
4589 * or is prohibited because it recently failed at this
4590 * order, fail immediately unless the allocator has
4591 * requested compaction and reclaim retry.
4594 * - potentially very expensive because zones are far
4595 * below their low watermarks or this is part of very
4596 * bursty high order allocations,
4597 * - not guaranteed to help because isolate_freepages()
4598 * may not iterate over freed pages as part of its
4600 * - unlikely to make entire pageblocks free on its
4603 if (compact_result
== COMPACT_SKIPPED
||
4604 compact_result
== COMPACT_DEFERRED
)
4608 * Looks like reclaim/compaction is worth trying, but
4609 * sync compaction could be very expensive, so keep
4610 * using async compaction.
4612 compact_priority
= INIT_COMPACT_PRIORITY
;
4617 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4618 if (alloc_flags
& ALLOC_KSWAPD
)
4619 wake_all_kswapds(order
, gfp_mask
, ac
);
4621 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4623 alloc_flags
= reserve_flags
;
4626 * Reset the nodemask and zonelist iterators if memory policies can be
4627 * ignored. These allocations are high priority and system rather than
4630 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4631 ac
->nodemask
= NULL
;
4632 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4633 ac
->highest_zoneidx
, ac
->nodemask
);
4636 /* Attempt with potentially adjusted zonelist and alloc_flags */
4637 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4641 /* Caller is not willing to reclaim, we can't balance anything */
4642 if (!can_direct_reclaim
)
4645 /* Avoid recursion of direct reclaim */
4646 if (current
->flags
& PF_MEMALLOC
)
4649 /* Try direct reclaim and then allocating */
4650 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4651 &did_some_progress
);
4655 /* Try direct compaction and then allocating */
4656 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4657 compact_priority
, &compact_result
);
4661 /* Do not loop if specifically requested */
4662 if (gfp_mask
& __GFP_NORETRY
)
4666 * Do not retry costly high order allocations unless they are
4667 * __GFP_RETRY_MAYFAIL
4669 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4672 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4673 did_some_progress
> 0, &no_progress_loops
))
4677 * It doesn't make any sense to retry for the compaction if the order-0
4678 * reclaim is not able to make any progress because the current
4679 * implementation of the compaction depends on the sufficient amount
4680 * of free memory (see __compaction_suitable)
4682 if (did_some_progress
> 0 &&
4683 should_compact_retry(ac
, order
, alloc_flags
,
4684 compact_result
, &compact_priority
,
4685 &compaction_retries
))
4689 /* Deal with possible cpuset update races before we start OOM killing */
4690 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4693 /* Reclaim has failed us, start killing things */
4694 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4698 /* Avoid allocations with no watermarks from looping endlessly */
4699 if (tsk_is_oom_victim(current
) &&
4700 (alloc_flags
== ALLOC_OOM
||
4701 (gfp_mask
& __GFP_NOMEMALLOC
)))
4704 /* Retry as long as the OOM killer is making progress */
4705 if (did_some_progress
) {
4706 no_progress_loops
= 0;
4711 /* Deal with possible cpuset update races before we fail */
4712 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4716 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4719 if (gfp_mask
& __GFP_NOFAIL
) {
4721 * All existing users of the __GFP_NOFAIL are blockable, so warn
4722 * of any new users that actually require GFP_NOWAIT
4724 if (WARN_ON_ONCE(!can_direct_reclaim
))
4728 * PF_MEMALLOC request from this context is rather bizarre
4729 * because we cannot reclaim anything and only can loop waiting
4730 * for somebody to do a work for us
4732 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4735 * non failing costly orders are a hard requirement which we
4736 * are not prepared for much so let's warn about these users
4737 * so that we can identify them and convert them to something
4740 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4743 * Help non-failing allocations by giving them access to memory
4744 * reserves but do not use ALLOC_NO_WATERMARKS because this
4745 * could deplete whole memory reserves which would just make
4746 * the situation worse
4748 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4756 warn_alloc(gfp_mask
, ac
->nodemask
,
4757 "page allocation failure: order:%u", order
);
4762 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4763 int preferred_nid
, nodemask_t
*nodemask
,
4764 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4765 unsigned int *alloc_flags
)
4767 ac
->highest_zoneidx
= gfp_zone(gfp_mask
);
4768 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4769 ac
->nodemask
= nodemask
;
4770 ac
->migratetype
= gfp_migratetype(gfp_mask
);
4772 if (cpusets_enabled()) {
4773 *alloc_mask
|= __GFP_HARDWALL
;
4775 ac
->nodemask
= &cpuset_current_mems_allowed
;
4777 *alloc_flags
|= ALLOC_CPUSET
;
4780 fs_reclaim_acquire(gfp_mask
);
4781 fs_reclaim_release(gfp_mask
);
4783 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4785 if (should_fail_alloc_page(gfp_mask
, order
))
4788 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4789 *alloc_flags
|= ALLOC_CMA
;
4794 /* Determine whether to spread dirty pages and what the first usable zone */
4795 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4797 /* Dirty zone balancing only done in the fast path */
4798 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4801 * The preferred zone is used for statistics but crucially it is
4802 * also used as the starting point for the zonelist iterator. It
4803 * may get reset for allocations that ignore memory policies.
4805 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4806 ac
->highest_zoneidx
, ac
->nodemask
);
4810 * This is the 'heart' of the zoned buddy allocator.
4813 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4814 nodemask_t
*nodemask
)
4817 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4818 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4819 struct alloc_context ac
= { };
4822 * There are several places where we assume that the order value is sane
4823 * so bail out early if the request is out of bound.
4825 if (unlikely(order
>= MAX_ORDER
)) {
4826 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4830 gfp_mask
&= gfp_allowed_mask
;
4831 alloc_mask
= gfp_mask
;
4832 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4835 finalise_ac(gfp_mask
, &ac
);
4838 * Forbid the first pass from falling back to types that fragment
4839 * memory until all local zones are considered.
4841 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4843 /* First allocation attempt */
4844 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4849 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4850 * resp. GFP_NOIO which has to be inherited for all allocation requests
4851 * from a particular context which has been marked by
4852 * memalloc_no{fs,io}_{save,restore}.
4854 alloc_mask
= current_gfp_context(gfp_mask
);
4855 ac
.spread_dirty_pages
= false;
4858 * Restore the original nodemask if it was potentially replaced with
4859 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4861 ac
.nodemask
= nodemask
;
4863 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4866 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4867 unlikely(__memcg_kmem_charge_page(page
, gfp_mask
, order
) != 0)) {
4868 __free_pages(page
, order
);
4872 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4876 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4879 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4880 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4881 * you need to access high mem.
4883 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4887 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4890 return (unsigned long) page_address(page
);
4892 EXPORT_SYMBOL(__get_free_pages
);
4894 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4896 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4898 EXPORT_SYMBOL(get_zeroed_page
);
4900 static inline void free_the_page(struct page
*page
, unsigned int order
)
4902 if (order
== 0) /* Via pcp? */
4903 free_unref_page(page
);
4905 __free_pages_ok(page
, order
);
4908 void __free_pages(struct page
*page
, unsigned int order
)
4910 if (put_page_testzero(page
))
4911 free_the_page(page
, order
);
4913 EXPORT_SYMBOL(__free_pages
);
4915 void free_pages(unsigned long addr
, unsigned int order
)
4918 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4919 __free_pages(virt_to_page((void *)addr
), order
);
4923 EXPORT_SYMBOL(free_pages
);
4927 * An arbitrary-length arbitrary-offset area of memory which resides
4928 * within a 0 or higher order page. Multiple fragments within that page
4929 * are individually refcounted, in the page's reference counter.
4931 * The page_frag functions below provide a simple allocation framework for
4932 * page fragments. This is used by the network stack and network device
4933 * drivers to provide a backing region of memory for use as either an
4934 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4936 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4939 struct page
*page
= NULL
;
4940 gfp_t gfp
= gfp_mask
;
4942 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4943 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4945 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4946 PAGE_FRAG_CACHE_MAX_ORDER
);
4947 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4949 if (unlikely(!page
))
4950 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4952 nc
->va
= page
? page_address(page
) : NULL
;
4957 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4959 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4961 if (page_ref_sub_and_test(page
, count
))
4962 free_the_page(page
, compound_order(page
));
4964 EXPORT_SYMBOL(__page_frag_cache_drain
);
4966 void *page_frag_alloc(struct page_frag_cache
*nc
,
4967 unsigned int fragsz
, gfp_t gfp_mask
)
4969 unsigned int size
= PAGE_SIZE
;
4973 if (unlikely(!nc
->va
)) {
4975 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4979 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4980 /* if size can vary use size else just use PAGE_SIZE */
4983 /* Even if we own the page, we do not use atomic_set().
4984 * This would break get_page_unless_zero() users.
4986 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
4988 /* reset page count bias and offset to start of new frag */
4989 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4990 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4994 offset
= nc
->offset
- fragsz
;
4995 if (unlikely(offset
< 0)) {
4996 page
= virt_to_page(nc
->va
);
4998 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
5001 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5002 /* if size can vary use size else just use PAGE_SIZE */
5005 /* OK, page count is 0, we can safely set it */
5006 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
5008 /* reset page count bias and offset to start of new frag */
5009 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
5010 offset
= size
- fragsz
;
5014 nc
->offset
= offset
;
5016 return nc
->va
+ offset
;
5018 EXPORT_SYMBOL(page_frag_alloc
);
5021 * Frees a page fragment allocated out of either a compound or order 0 page.
5023 void page_frag_free(void *addr
)
5025 struct page
*page
= virt_to_head_page(addr
);
5027 if (unlikely(put_page_testzero(page
)))
5028 free_the_page(page
, compound_order(page
));
5030 EXPORT_SYMBOL(page_frag_free
);
5032 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
5036 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
5037 unsigned long used
= addr
+ PAGE_ALIGN(size
);
5039 split_page(virt_to_page((void *)addr
), order
);
5040 while (used
< alloc_end
) {
5045 return (void *)addr
;
5049 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5050 * @size: the number of bytes to allocate
5051 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5053 * This function is similar to alloc_pages(), except that it allocates the
5054 * minimum number of pages to satisfy the request. alloc_pages() can only
5055 * allocate memory in power-of-two pages.
5057 * This function is also limited by MAX_ORDER.
5059 * Memory allocated by this function must be released by free_pages_exact().
5061 * Return: pointer to the allocated area or %NULL in case of error.
5063 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
5065 unsigned int order
= get_order(size
);
5068 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5069 gfp_mask
&= ~__GFP_COMP
;
5071 addr
= __get_free_pages(gfp_mask
, order
);
5072 return make_alloc_exact(addr
, order
, size
);
5074 EXPORT_SYMBOL(alloc_pages_exact
);
5077 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5079 * @nid: the preferred node ID where memory should be allocated
5080 * @size: the number of bytes to allocate
5081 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5083 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5086 * Return: pointer to the allocated area or %NULL in case of error.
5088 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
5090 unsigned int order
= get_order(size
);
5093 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5094 gfp_mask
&= ~__GFP_COMP
;
5096 p
= alloc_pages_node(nid
, gfp_mask
, order
);
5099 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5103 * free_pages_exact - release memory allocated via alloc_pages_exact()
5104 * @virt: the value returned by alloc_pages_exact.
5105 * @size: size of allocation, same value as passed to alloc_pages_exact().
5107 * Release the memory allocated by a previous call to alloc_pages_exact.
5109 void free_pages_exact(void *virt
, size_t size
)
5111 unsigned long addr
= (unsigned long)virt
;
5112 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5114 while (addr
< end
) {
5119 EXPORT_SYMBOL(free_pages_exact
);
5122 * nr_free_zone_pages - count number of pages beyond high watermark
5123 * @offset: The zone index of the highest zone
5125 * nr_free_zone_pages() counts the number of pages which are beyond the
5126 * high watermark within all zones at or below a given zone index. For each
5127 * zone, the number of pages is calculated as:
5129 * nr_free_zone_pages = managed_pages - high_pages
5131 * Return: number of pages beyond high watermark.
5133 static unsigned long nr_free_zone_pages(int offset
)
5138 /* Just pick one node, since fallback list is circular */
5139 unsigned long sum
= 0;
5141 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5143 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5144 unsigned long size
= zone_managed_pages(zone
);
5145 unsigned long high
= high_wmark_pages(zone
);
5154 * nr_free_buffer_pages - count number of pages beyond high watermark
5156 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5157 * watermark within ZONE_DMA and ZONE_NORMAL.
5159 * Return: number of pages beyond high watermark within ZONE_DMA and
5162 unsigned long nr_free_buffer_pages(void)
5164 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5166 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5169 * nr_free_pagecache_pages - count number of pages beyond high watermark
5171 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5172 * high watermark within all zones.
5174 * Return: number of pages beyond high watermark within all zones.
5176 unsigned long nr_free_pagecache_pages(void)
5178 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5181 static inline void show_node(struct zone
*zone
)
5183 if (IS_ENABLED(CONFIG_NUMA
))
5184 printk("Node %d ", zone_to_nid(zone
));
5187 long si_mem_available(void)
5190 unsigned long pagecache
;
5191 unsigned long wmark_low
= 0;
5192 unsigned long pages
[NR_LRU_LISTS
];
5193 unsigned long reclaimable
;
5197 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5198 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5201 wmark_low
+= low_wmark_pages(zone
);
5204 * Estimate the amount of memory available for userspace allocations,
5205 * without causing swapping.
5207 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5210 * Not all the page cache can be freed, otherwise the system will
5211 * start swapping. Assume at least half of the page cache, or the
5212 * low watermark worth of cache, needs to stay.
5214 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5215 pagecache
-= min(pagecache
/ 2, wmark_low
);
5216 available
+= pagecache
;
5219 * Part of the reclaimable slab and other kernel memory consists of
5220 * items that are in use, and cannot be freed. Cap this estimate at the
5223 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
5224 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5225 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5231 EXPORT_SYMBOL_GPL(si_mem_available
);
5233 void si_meminfo(struct sysinfo
*val
)
5235 val
->totalram
= totalram_pages();
5236 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5237 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5238 val
->bufferram
= nr_blockdev_pages();
5239 val
->totalhigh
= totalhigh_pages();
5240 val
->freehigh
= nr_free_highpages();
5241 val
->mem_unit
= PAGE_SIZE
;
5244 EXPORT_SYMBOL(si_meminfo
);
5247 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5249 int zone_type
; /* needs to be signed */
5250 unsigned long managed_pages
= 0;
5251 unsigned long managed_highpages
= 0;
5252 unsigned long free_highpages
= 0;
5253 pg_data_t
*pgdat
= NODE_DATA(nid
);
5255 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5256 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5257 val
->totalram
= managed_pages
;
5258 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5259 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5260 #ifdef CONFIG_HIGHMEM
5261 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5262 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5264 if (is_highmem(zone
)) {
5265 managed_highpages
+= zone_managed_pages(zone
);
5266 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5269 val
->totalhigh
= managed_highpages
;
5270 val
->freehigh
= free_highpages
;
5272 val
->totalhigh
= managed_highpages
;
5273 val
->freehigh
= free_highpages
;
5275 val
->mem_unit
= PAGE_SIZE
;
5280 * Determine whether the node should be displayed or not, depending on whether
5281 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5283 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5285 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5289 * no node mask - aka implicit memory numa policy. Do not bother with
5290 * the synchronization - read_mems_allowed_begin - because we do not
5291 * have to be precise here.
5294 nodemask
= &cpuset_current_mems_allowed
;
5296 return !node_isset(nid
, *nodemask
);
5299 #define K(x) ((x) << (PAGE_SHIFT-10))
5301 static void show_migration_types(unsigned char type
)
5303 static const char types
[MIGRATE_TYPES
] = {
5304 [MIGRATE_UNMOVABLE
] = 'U',
5305 [MIGRATE_MOVABLE
] = 'M',
5306 [MIGRATE_RECLAIMABLE
] = 'E',
5307 [MIGRATE_HIGHATOMIC
] = 'H',
5309 [MIGRATE_CMA
] = 'C',
5311 #ifdef CONFIG_MEMORY_ISOLATION
5312 [MIGRATE_ISOLATE
] = 'I',
5315 char tmp
[MIGRATE_TYPES
+ 1];
5319 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5320 if (type
& (1 << i
))
5325 printk(KERN_CONT
"(%s) ", tmp
);
5329 * Show free area list (used inside shift_scroll-lock stuff)
5330 * We also calculate the percentage fragmentation. We do this by counting the
5331 * memory on each free list with the exception of the first item on the list.
5334 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5337 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5339 unsigned long free_pcp
= 0;
5344 for_each_populated_zone(zone
) {
5345 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5348 for_each_online_cpu(cpu
)
5349 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5352 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5353 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5354 " unevictable:%lu dirty:%lu writeback:%lu\n"
5355 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5356 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5357 " free:%lu free_pcp:%lu free_cma:%lu\n",
5358 global_node_page_state(NR_ACTIVE_ANON
),
5359 global_node_page_state(NR_INACTIVE_ANON
),
5360 global_node_page_state(NR_ISOLATED_ANON
),
5361 global_node_page_state(NR_ACTIVE_FILE
),
5362 global_node_page_state(NR_INACTIVE_FILE
),
5363 global_node_page_state(NR_ISOLATED_FILE
),
5364 global_node_page_state(NR_UNEVICTABLE
),
5365 global_node_page_state(NR_FILE_DIRTY
),
5366 global_node_page_state(NR_WRITEBACK
),
5367 global_node_page_state(NR_SLAB_RECLAIMABLE
),
5368 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
5369 global_node_page_state(NR_FILE_MAPPED
),
5370 global_node_page_state(NR_SHMEM
),
5371 global_zone_page_state(NR_PAGETABLE
),
5372 global_zone_page_state(NR_BOUNCE
),
5373 global_zone_page_state(NR_FREE_PAGES
),
5375 global_zone_page_state(NR_FREE_CMA_PAGES
));
5377 for_each_online_pgdat(pgdat
) {
5378 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5382 " active_anon:%lukB"
5383 " inactive_anon:%lukB"
5384 " active_file:%lukB"
5385 " inactive_file:%lukB"
5386 " unevictable:%lukB"
5387 " isolated(anon):%lukB"
5388 " isolated(file):%lukB"
5393 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5395 " shmem_pmdmapped: %lukB"
5398 " writeback_tmp:%lukB"
5399 " all_unreclaimable? %s"
5402 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5403 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5404 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5405 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5406 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5407 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5408 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5409 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5410 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5411 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5412 K(node_page_state(pgdat
, NR_SHMEM
)),
5413 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5414 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5415 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5417 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5419 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5420 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5424 for_each_populated_zone(zone
) {
5427 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5431 for_each_online_cpu(cpu
)
5432 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5441 " reserved_highatomic:%luKB"
5442 " active_anon:%lukB"
5443 " inactive_anon:%lukB"
5444 " active_file:%lukB"
5445 " inactive_file:%lukB"
5446 " unevictable:%lukB"
5447 " writepending:%lukB"
5451 " kernel_stack:%lukB"
5452 #ifdef CONFIG_SHADOW_CALL_STACK
5453 " shadow_call_stack:%lukB"
5462 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5463 K(min_wmark_pages(zone
)),
5464 K(low_wmark_pages(zone
)),
5465 K(high_wmark_pages(zone
)),
5466 K(zone
->nr_reserved_highatomic
),
5467 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5468 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5469 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5470 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5471 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5472 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5473 K(zone
->present_pages
),
5474 K(zone_managed_pages(zone
)),
5475 K(zone_page_state(zone
, NR_MLOCK
)),
5476 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5477 #ifdef CONFIG_SHADOW_CALL_STACK
5478 zone_page_state(zone
, NR_KERNEL_SCS_KB
),
5480 K(zone_page_state(zone
, NR_PAGETABLE
)),
5481 K(zone_page_state(zone
, NR_BOUNCE
)),
5483 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5484 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5485 printk("lowmem_reserve[]:");
5486 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5487 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5488 printk(KERN_CONT
"\n");
5491 for_each_populated_zone(zone
) {
5493 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5494 unsigned char types
[MAX_ORDER
];
5496 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5499 printk(KERN_CONT
"%s: ", zone
->name
);
5501 spin_lock_irqsave(&zone
->lock
, flags
);
5502 for (order
= 0; order
< MAX_ORDER
; order
++) {
5503 struct free_area
*area
= &zone
->free_area
[order
];
5506 nr
[order
] = area
->nr_free
;
5507 total
+= nr
[order
] << order
;
5510 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5511 if (!free_area_empty(area
, type
))
5512 types
[order
] |= 1 << type
;
5515 spin_unlock_irqrestore(&zone
->lock
, flags
);
5516 for (order
= 0; order
< MAX_ORDER
; order
++) {
5517 printk(KERN_CONT
"%lu*%lukB ",
5518 nr
[order
], K(1UL) << order
);
5520 show_migration_types(types
[order
]);
5522 printk(KERN_CONT
"= %lukB\n", K(total
));
5525 hugetlb_show_meminfo();
5527 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5529 show_swap_cache_info();
5532 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5534 zoneref
->zone
= zone
;
5535 zoneref
->zone_idx
= zone_idx(zone
);
5539 * Builds allocation fallback zone lists.
5541 * Add all populated zones of a node to the zonelist.
5543 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5546 enum zone_type zone_type
= MAX_NR_ZONES
;
5551 zone
= pgdat
->node_zones
+ zone_type
;
5552 if (managed_zone(zone
)) {
5553 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5554 check_highest_zone(zone_type
);
5556 } while (zone_type
);
5563 static int __parse_numa_zonelist_order(char *s
)
5566 * We used to support different zonlists modes but they turned
5567 * out to be just not useful. Let's keep the warning in place
5568 * if somebody still use the cmd line parameter so that we do
5569 * not fail it silently
5571 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5572 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5578 char numa_zonelist_order
[] = "Node";
5581 * sysctl handler for numa_zonelist_order
5583 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5584 void *buffer
, size_t *length
, loff_t
*ppos
)
5587 return __parse_numa_zonelist_order(buffer
);
5588 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5592 #define MAX_NODE_LOAD (nr_online_nodes)
5593 static int node_load
[MAX_NUMNODES
];
5596 * find_next_best_node - find the next node that should appear in a given node's fallback list
5597 * @node: node whose fallback list we're appending
5598 * @used_node_mask: nodemask_t of already used nodes
5600 * We use a number of factors to determine which is the next node that should
5601 * appear on a given node's fallback list. The node should not have appeared
5602 * already in @node's fallback list, and it should be the next closest node
5603 * according to the distance array (which contains arbitrary distance values
5604 * from each node to each node in the system), and should also prefer nodes
5605 * with no CPUs, since presumably they'll have very little allocation pressure
5606 * on them otherwise.
5608 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5610 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5613 int min_val
= INT_MAX
;
5614 int best_node
= NUMA_NO_NODE
;
5615 const struct cpumask
*tmp
= cpumask_of_node(0);
5617 /* Use the local node if we haven't already */
5618 if (!node_isset(node
, *used_node_mask
)) {
5619 node_set(node
, *used_node_mask
);
5623 for_each_node_state(n
, N_MEMORY
) {
5625 /* Don't want a node to appear more than once */
5626 if (node_isset(n
, *used_node_mask
))
5629 /* Use the distance array to find the distance */
5630 val
= node_distance(node
, n
);
5632 /* Penalize nodes under us ("prefer the next node") */
5635 /* Give preference to headless and unused nodes */
5636 tmp
= cpumask_of_node(n
);
5637 if (!cpumask_empty(tmp
))
5638 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5640 /* Slight preference for less loaded node */
5641 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5642 val
+= node_load
[n
];
5644 if (val
< min_val
) {
5651 node_set(best_node
, *used_node_mask
);
5658 * Build zonelists ordered by node and zones within node.
5659 * This results in maximum locality--normal zone overflows into local
5660 * DMA zone, if any--but risks exhausting DMA zone.
5662 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5665 struct zoneref
*zonerefs
;
5668 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5670 for (i
= 0; i
< nr_nodes
; i
++) {
5673 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5675 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5676 zonerefs
+= nr_zones
;
5678 zonerefs
->zone
= NULL
;
5679 zonerefs
->zone_idx
= 0;
5683 * Build gfp_thisnode zonelists
5685 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5687 struct zoneref
*zonerefs
;
5690 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5691 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5692 zonerefs
+= nr_zones
;
5693 zonerefs
->zone
= NULL
;
5694 zonerefs
->zone_idx
= 0;
5698 * Build zonelists ordered by zone and nodes within zones.
5699 * This results in conserving DMA zone[s] until all Normal memory is
5700 * exhausted, but results in overflowing to remote node while memory
5701 * may still exist in local DMA zone.
5704 static void build_zonelists(pg_data_t
*pgdat
)
5706 static int node_order
[MAX_NUMNODES
];
5707 int node
, load
, nr_nodes
= 0;
5708 nodemask_t used_mask
= NODE_MASK_NONE
;
5709 int local_node
, prev_node
;
5711 /* NUMA-aware ordering of nodes */
5712 local_node
= pgdat
->node_id
;
5713 load
= nr_online_nodes
;
5714 prev_node
= local_node
;
5716 memset(node_order
, 0, sizeof(node_order
));
5717 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5719 * We don't want to pressure a particular node.
5720 * So adding penalty to the first node in same
5721 * distance group to make it round-robin.
5723 if (node_distance(local_node
, node
) !=
5724 node_distance(local_node
, prev_node
))
5725 node_load
[node
] = load
;
5727 node_order
[nr_nodes
++] = node
;
5732 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5733 build_thisnode_zonelists(pgdat
);
5736 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5738 * Return node id of node used for "local" allocations.
5739 * I.e., first node id of first zone in arg node's generic zonelist.
5740 * Used for initializing percpu 'numa_mem', which is used primarily
5741 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5743 int local_memory_node(int node
)
5747 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5748 gfp_zone(GFP_KERNEL
),
5750 return zone_to_nid(z
->zone
);
5754 static void setup_min_unmapped_ratio(void);
5755 static void setup_min_slab_ratio(void);
5756 #else /* CONFIG_NUMA */
5758 static void build_zonelists(pg_data_t
*pgdat
)
5760 int node
, local_node
;
5761 struct zoneref
*zonerefs
;
5764 local_node
= pgdat
->node_id
;
5766 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5767 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5768 zonerefs
+= nr_zones
;
5771 * Now we build the zonelist so that it contains the zones
5772 * of all the other nodes.
5773 * We don't want to pressure a particular node, so when
5774 * building the zones for node N, we make sure that the
5775 * zones coming right after the local ones are those from
5776 * node N+1 (modulo N)
5778 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5779 if (!node_online(node
))
5781 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5782 zonerefs
+= nr_zones
;
5784 for (node
= 0; node
< local_node
; node
++) {
5785 if (!node_online(node
))
5787 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5788 zonerefs
+= nr_zones
;
5791 zonerefs
->zone
= NULL
;
5792 zonerefs
->zone_idx
= 0;
5795 #endif /* CONFIG_NUMA */
5798 * Boot pageset table. One per cpu which is going to be used for all
5799 * zones and all nodes. The parameters will be set in such a way
5800 * that an item put on a list will immediately be handed over to
5801 * the buddy list. This is safe since pageset manipulation is done
5802 * with interrupts disabled.
5804 * The boot_pagesets must be kept even after bootup is complete for
5805 * unused processors and/or zones. They do play a role for bootstrapping
5806 * hotplugged processors.
5808 * zoneinfo_show() and maybe other functions do
5809 * not check if the processor is online before following the pageset pointer.
5810 * Other parts of the kernel may not check if the zone is available.
5812 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5813 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5814 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5816 static void __build_all_zonelists(void *data
)
5819 int __maybe_unused cpu
;
5820 pg_data_t
*self
= data
;
5821 static DEFINE_SPINLOCK(lock
);
5826 memset(node_load
, 0, sizeof(node_load
));
5830 * This node is hotadded and no memory is yet present. So just
5831 * building zonelists is fine - no need to touch other nodes.
5833 if (self
&& !node_online(self
->node_id
)) {
5834 build_zonelists(self
);
5836 for_each_online_node(nid
) {
5837 pg_data_t
*pgdat
= NODE_DATA(nid
);
5839 build_zonelists(pgdat
);
5842 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5844 * We now know the "local memory node" for each node--
5845 * i.e., the node of the first zone in the generic zonelist.
5846 * Set up numa_mem percpu variable for on-line cpus. During
5847 * boot, only the boot cpu should be on-line; we'll init the
5848 * secondary cpus' numa_mem as they come on-line. During
5849 * node/memory hotplug, we'll fixup all on-line cpus.
5851 for_each_online_cpu(cpu
)
5852 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5859 static noinline
void __init
5860 build_all_zonelists_init(void)
5864 __build_all_zonelists(NULL
);
5867 * Initialize the boot_pagesets that are going to be used
5868 * for bootstrapping processors. The real pagesets for
5869 * each zone will be allocated later when the per cpu
5870 * allocator is available.
5872 * boot_pagesets are used also for bootstrapping offline
5873 * cpus if the system is already booted because the pagesets
5874 * are needed to initialize allocators on a specific cpu too.
5875 * F.e. the percpu allocator needs the page allocator which
5876 * needs the percpu allocator in order to allocate its pagesets
5877 * (a chicken-egg dilemma).
5879 for_each_possible_cpu(cpu
)
5880 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5882 mminit_verify_zonelist();
5883 cpuset_init_current_mems_allowed();
5887 * unless system_state == SYSTEM_BOOTING.
5889 * __ref due to call of __init annotated helper build_all_zonelists_init
5890 * [protected by SYSTEM_BOOTING].
5892 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5894 if (system_state
== SYSTEM_BOOTING
) {
5895 build_all_zonelists_init();
5897 __build_all_zonelists(pgdat
);
5898 /* cpuset refresh routine should be here */
5900 vm_total_pages
= nr_free_pagecache_pages();
5902 * Disable grouping by mobility if the number of pages in the
5903 * system is too low to allow the mechanism to work. It would be
5904 * more accurate, but expensive to check per-zone. This check is
5905 * made on memory-hotadd so a system can start with mobility
5906 * disabled and enable it later
5908 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5909 page_group_by_mobility_disabled
= 1;
5911 page_group_by_mobility_disabled
= 0;
5913 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5915 page_group_by_mobility_disabled
? "off" : "on",
5918 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5922 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5923 static bool __meminit
5924 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5926 static struct memblock_region
*r
;
5928 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5929 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5930 for_each_memblock(memory
, r
) {
5931 if (*pfn
< memblock_region_memory_end_pfn(r
))
5935 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5936 memblock_is_mirror(r
)) {
5937 *pfn
= memblock_region_memory_end_pfn(r
);
5945 * Initially all pages are reserved - free ones are freed
5946 * up by memblock_free_all() once the early boot process is
5947 * done. Non-atomic initialization, single-pass.
5949 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5950 unsigned long start_pfn
, enum memmap_context context
,
5951 struct vmem_altmap
*altmap
)
5953 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5956 if (highest_memmap_pfn
< end_pfn
- 1)
5957 highest_memmap_pfn
= end_pfn
- 1;
5959 #ifdef CONFIG_ZONE_DEVICE
5961 * Honor reservation requested by the driver for this ZONE_DEVICE
5962 * memory. We limit the total number of pages to initialize to just
5963 * those that might contain the memory mapping. We will defer the
5964 * ZONE_DEVICE page initialization until after we have released
5967 if (zone
== ZONE_DEVICE
) {
5971 if (start_pfn
== altmap
->base_pfn
)
5972 start_pfn
+= altmap
->reserve
;
5973 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5977 for (pfn
= start_pfn
; pfn
< end_pfn
; ) {
5979 * There can be holes in boot-time mem_map[]s handed to this
5980 * function. They do not exist on hotplugged memory.
5982 if (context
== MEMMAP_EARLY
) {
5983 if (overlap_memmap_init(zone
, &pfn
))
5985 if (defer_init(nid
, pfn
, end_pfn
))
5989 page
= pfn_to_page(pfn
);
5990 __init_single_page(page
, pfn
, zone
, nid
);
5991 if (context
== MEMMAP_HOTPLUG
)
5992 __SetPageReserved(page
);
5995 * Mark the block movable so that blocks are reserved for
5996 * movable at startup. This will force kernel allocations
5997 * to reserve their blocks rather than leaking throughout
5998 * the address space during boot when many long-lived
5999 * kernel allocations are made.
6001 * bitmap is created for zone's valid pfn range. but memmap
6002 * can be created for invalid pages (for alignment)
6003 * check here not to call set_pageblock_migratetype() against
6006 if (!(pfn
& (pageblock_nr_pages
- 1))) {
6007 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6014 #ifdef CONFIG_ZONE_DEVICE
6015 void __ref
memmap_init_zone_device(struct zone
*zone
,
6016 unsigned long start_pfn
,
6017 unsigned long nr_pages
,
6018 struct dev_pagemap
*pgmap
)
6020 unsigned long pfn
, end_pfn
= start_pfn
+ nr_pages
;
6021 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6022 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
6023 unsigned long zone_idx
= zone_idx(zone
);
6024 unsigned long start
= jiffies
;
6025 int nid
= pgdat
->node_id
;
6027 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
6031 * The call to memmap_init_zone should have already taken care
6032 * of the pages reserved for the memmap, so we can just jump to
6033 * the end of that region and start processing the device pages.
6036 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
6037 nr_pages
= end_pfn
- start_pfn
;
6040 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
6041 struct page
*page
= pfn_to_page(pfn
);
6043 __init_single_page(page
, pfn
, zone_idx
, nid
);
6046 * Mark page reserved as it will need to wait for onlining
6047 * phase for it to be fully associated with a zone.
6049 * We can use the non-atomic __set_bit operation for setting
6050 * the flag as we are still initializing the pages.
6052 __SetPageReserved(page
);
6055 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6056 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6057 * ever freed or placed on a driver-private list.
6059 page
->pgmap
= pgmap
;
6060 page
->zone_device_data
= NULL
;
6063 * Mark the block movable so that blocks are reserved for
6064 * movable at startup. This will force kernel allocations
6065 * to reserve their blocks rather than leaking throughout
6066 * the address space during boot when many long-lived
6067 * kernel allocations are made.
6069 * bitmap is created for zone's valid pfn range. but memmap
6070 * can be created for invalid pages (for alignment)
6071 * check here not to call set_pageblock_migratetype() against
6074 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6075 * because this is done early in section_activate()
6077 if (!(pfn
& (pageblock_nr_pages
- 1))) {
6078 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6083 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6084 nr_pages
, jiffies_to_msecs(jiffies
- start
));
6088 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6090 unsigned int order
, t
;
6091 for_each_migratetype_order(order
, t
) {
6092 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6093 zone
->free_area
[order
].nr_free
= 0;
6097 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
6099 unsigned long range_start_pfn
)
6101 unsigned long start_pfn
, end_pfn
;
6102 unsigned long range_end_pfn
= range_start_pfn
+ size
;
6105 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6106 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6107 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6109 if (end_pfn
> start_pfn
) {
6110 size
= end_pfn
- start_pfn
;
6111 memmap_init_zone(size
, nid
, zone
, start_pfn
,
6112 MEMMAP_EARLY
, NULL
);
6117 static int zone_batchsize(struct zone
*zone
)
6123 * The per-cpu-pages pools are set to around 1000th of the
6126 batch
= zone_managed_pages(zone
) / 1024;
6127 /* But no more than a meg. */
6128 if (batch
* PAGE_SIZE
> 1024 * 1024)
6129 batch
= (1024 * 1024) / PAGE_SIZE
;
6130 batch
/= 4; /* We effectively *= 4 below */
6135 * Clamp the batch to a 2^n - 1 value. Having a power
6136 * of 2 value was found to be more likely to have
6137 * suboptimal cache aliasing properties in some cases.
6139 * For example if 2 tasks are alternately allocating
6140 * batches of pages, one task can end up with a lot
6141 * of pages of one half of the possible page colors
6142 * and the other with pages of the other colors.
6144 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6149 /* The deferral and batching of frees should be suppressed under NOMMU
6152 * The problem is that NOMMU needs to be able to allocate large chunks
6153 * of contiguous memory as there's no hardware page translation to
6154 * assemble apparent contiguous memory from discontiguous pages.
6156 * Queueing large contiguous runs of pages for batching, however,
6157 * causes the pages to actually be freed in smaller chunks. As there
6158 * can be a significant delay between the individual batches being
6159 * recycled, this leads to the once large chunks of space being
6160 * fragmented and becoming unavailable for high-order allocations.
6167 * pcp->high and pcp->batch values are related and dependent on one another:
6168 * ->batch must never be higher then ->high.
6169 * The following function updates them in a safe manner without read side
6172 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6173 * those fields changing asynchronously (acording the the above rule).
6175 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6176 * outside of boot time (or some other assurance that no concurrent updaters
6179 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6180 unsigned long batch
)
6182 /* start with a fail safe value for batch */
6186 /* Update high, then batch, in order */
6193 /* a companion to pageset_set_high() */
6194 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
6196 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
6199 static void pageset_init(struct per_cpu_pageset
*p
)
6201 struct per_cpu_pages
*pcp
;
6204 memset(p
, 0, sizeof(*p
));
6207 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
6208 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
6211 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
6214 pageset_set_batch(p
, batch
);
6218 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6219 * to the value high for the pageset p.
6221 static void pageset_set_high(struct per_cpu_pageset
*p
,
6224 unsigned long batch
= max(1UL, high
/ 4);
6225 if ((high
/ 4) > (PAGE_SHIFT
* 8))
6226 batch
= PAGE_SHIFT
* 8;
6228 pageset_update(&p
->pcp
, high
, batch
);
6231 static void pageset_set_high_and_batch(struct zone
*zone
,
6232 struct per_cpu_pageset
*pcp
)
6234 if (percpu_pagelist_fraction
)
6235 pageset_set_high(pcp
,
6236 (zone_managed_pages(zone
) /
6237 percpu_pagelist_fraction
));
6239 pageset_set_batch(pcp
, zone_batchsize(zone
));
6242 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
6244 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
6247 pageset_set_high_and_batch(zone
, pcp
);
6250 void __meminit
setup_zone_pageset(struct zone
*zone
)
6253 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
6254 for_each_possible_cpu(cpu
)
6255 zone_pageset_init(zone
, cpu
);
6259 * Allocate per cpu pagesets and initialize them.
6260 * Before this call only boot pagesets were available.
6262 void __init
setup_per_cpu_pageset(void)
6264 struct pglist_data
*pgdat
;
6266 int __maybe_unused cpu
;
6268 for_each_populated_zone(zone
)
6269 setup_zone_pageset(zone
);
6273 * Unpopulated zones continue using the boot pagesets.
6274 * The numa stats for these pagesets need to be reset.
6275 * Otherwise, they will end up skewing the stats of
6276 * the nodes these zones are associated with.
6278 for_each_possible_cpu(cpu
) {
6279 struct per_cpu_pageset
*pcp
= &per_cpu(boot_pageset
, cpu
);
6280 memset(pcp
->vm_numa_stat_diff
, 0,
6281 sizeof(pcp
->vm_numa_stat_diff
));
6285 for_each_online_pgdat(pgdat
)
6286 pgdat
->per_cpu_nodestats
=
6287 alloc_percpu(struct per_cpu_nodestat
);
6290 static __meminit
void zone_pcp_init(struct zone
*zone
)
6293 * per cpu subsystem is not up at this point. The following code
6294 * relies on the ability of the linker to provide the
6295 * offset of a (static) per cpu variable into the per cpu area.
6297 zone
->pageset
= &boot_pageset
;
6299 if (populated_zone(zone
))
6300 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
6301 zone
->name
, zone
->present_pages
,
6302 zone_batchsize(zone
));
6305 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6306 unsigned long zone_start_pfn
,
6309 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6310 int zone_idx
= zone_idx(zone
) + 1;
6312 if (zone_idx
> pgdat
->nr_zones
)
6313 pgdat
->nr_zones
= zone_idx
;
6315 zone
->zone_start_pfn
= zone_start_pfn
;
6317 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6318 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6320 (unsigned long)zone_idx(zone
),
6321 zone_start_pfn
, (zone_start_pfn
+ size
));
6323 zone_init_free_lists(zone
);
6324 zone
->initialized
= 1;
6328 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6329 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6331 * If an architecture guarantees that all ranges registered contain no holes and may
6332 * be freed, this function may be used instead of calling memory_present() manually.
6334 void __init
sparse_memory_present_with_active_regions(int nid
)
6336 unsigned long start_pfn
, end_pfn
;
6339 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
6340 memory_present(this_nid
, start_pfn
, end_pfn
);
6344 * get_pfn_range_for_nid - Return the start and end page frames for a node
6345 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6346 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6347 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6349 * It returns the start and end page frame of a node based on information
6350 * provided by memblock_set_node(). If called for a node
6351 * with no available memory, a warning is printed and the start and end
6354 void __init
get_pfn_range_for_nid(unsigned int nid
,
6355 unsigned long *start_pfn
, unsigned long *end_pfn
)
6357 unsigned long this_start_pfn
, this_end_pfn
;
6363 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6364 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6365 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6368 if (*start_pfn
== -1UL)
6373 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6374 * assumption is made that zones within a node are ordered in monotonic
6375 * increasing memory addresses so that the "highest" populated zone is used
6377 static void __init
find_usable_zone_for_movable(void)
6380 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6381 if (zone_index
== ZONE_MOVABLE
)
6384 if (arch_zone_highest_possible_pfn
[zone_index
] >
6385 arch_zone_lowest_possible_pfn
[zone_index
])
6389 VM_BUG_ON(zone_index
== -1);
6390 movable_zone
= zone_index
;
6394 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6395 * because it is sized independent of architecture. Unlike the other zones,
6396 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6397 * in each node depending on the size of each node and how evenly kernelcore
6398 * is distributed. This helper function adjusts the zone ranges
6399 * provided by the architecture for a given node by using the end of the
6400 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6401 * zones within a node are in order of monotonic increases memory addresses
6403 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6404 unsigned long zone_type
,
6405 unsigned long node_start_pfn
,
6406 unsigned long node_end_pfn
,
6407 unsigned long *zone_start_pfn
,
6408 unsigned long *zone_end_pfn
)
6410 /* Only adjust if ZONE_MOVABLE is on this node */
6411 if (zone_movable_pfn
[nid
]) {
6412 /* Size ZONE_MOVABLE */
6413 if (zone_type
== ZONE_MOVABLE
) {
6414 *zone_start_pfn
= zone_movable_pfn
[nid
];
6415 *zone_end_pfn
= min(node_end_pfn
,
6416 arch_zone_highest_possible_pfn
[movable_zone
]);
6418 /* Adjust for ZONE_MOVABLE starting within this range */
6419 } else if (!mirrored_kernelcore
&&
6420 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6421 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6422 *zone_end_pfn
= zone_movable_pfn
[nid
];
6424 /* Check if this whole range is within ZONE_MOVABLE */
6425 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6426 *zone_start_pfn
= *zone_end_pfn
;
6431 * Return the number of pages a zone spans in a node, including holes
6432 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6434 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6435 unsigned long zone_type
,
6436 unsigned long node_start_pfn
,
6437 unsigned long node_end_pfn
,
6438 unsigned long *zone_start_pfn
,
6439 unsigned long *zone_end_pfn
)
6441 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6442 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6443 /* When hotadd a new node from cpu_up(), the node should be empty */
6444 if (!node_start_pfn
&& !node_end_pfn
)
6447 /* Get the start and end of the zone */
6448 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6449 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6450 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6451 node_start_pfn
, node_end_pfn
,
6452 zone_start_pfn
, zone_end_pfn
);
6454 /* Check that this node has pages within the zone's required range */
6455 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6458 /* Move the zone boundaries inside the node if necessary */
6459 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6460 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6462 /* Return the spanned pages */
6463 return *zone_end_pfn
- *zone_start_pfn
;
6467 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6468 * then all holes in the requested range will be accounted for.
6470 unsigned long __init
__absent_pages_in_range(int nid
,
6471 unsigned long range_start_pfn
,
6472 unsigned long range_end_pfn
)
6474 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6475 unsigned long start_pfn
, end_pfn
;
6478 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6479 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6480 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6481 nr_absent
-= end_pfn
- start_pfn
;
6487 * absent_pages_in_range - Return number of page frames in holes within a range
6488 * @start_pfn: The start PFN to start searching for holes
6489 * @end_pfn: The end PFN to stop searching for holes
6491 * Return: the number of pages frames in memory holes within a range.
6493 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6494 unsigned long end_pfn
)
6496 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6499 /* Return the number of page frames in holes in a zone on a node */
6500 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6501 unsigned long zone_type
,
6502 unsigned long node_start_pfn
,
6503 unsigned long node_end_pfn
)
6505 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6506 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6507 unsigned long zone_start_pfn
, zone_end_pfn
;
6508 unsigned long nr_absent
;
6510 /* When hotadd a new node from cpu_up(), the node should be empty */
6511 if (!node_start_pfn
&& !node_end_pfn
)
6514 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6515 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6517 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6518 node_start_pfn
, node_end_pfn
,
6519 &zone_start_pfn
, &zone_end_pfn
);
6520 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6523 * ZONE_MOVABLE handling.
6524 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6527 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6528 unsigned long start_pfn
, end_pfn
;
6529 struct memblock_region
*r
;
6531 for_each_memblock(memory
, r
) {
6532 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6533 zone_start_pfn
, zone_end_pfn
);
6534 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6535 zone_start_pfn
, zone_end_pfn
);
6537 if (zone_type
== ZONE_MOVABLE
&&
6538 memblock_is_mirror(r
))
6539 nr_absent
+= end_pfn
- start_pfn
;
6541 if (zone_type
== ZONE_NORMAL
&&
6542 !memblock_is_mirror(r
))
6543 nr_absent
+= end_pfn
- start_pfn
;
6550 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6551 unsigned long node_start_pfn
,
6552 unsigned long node_end_pfn
)
6554 unsigned long realtotalpages
= 0, totalpages
= 0;
6557 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6558 struct zone
*zone
= pgdat
->node_zones
+ i
;
6559 unsigned long zone_start_pfn
, zone_end_pfn
;
6560 unsigned long spanned
, absent
;
6561 unsigned long size
, real_size
;
6563 spanned
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6568 absent
= zone_absent_pages_in_node(pgdat
->node_id
, i
,
6573 real_size
= size
- absent
;
6576 zone
->zone_start_pfn
= zone_start_pfn
;
6578 zone
->zone_start_pfn
= 0;
6579 zone
->spanned_pages
= size
;
6580 zone
->present_pages
= real_size
;
6583 realtotalpages
+= real_size
;
6586 pgdat
->node_spanned_pages
= totalpages
;
6587 pgdat
->node_present_pages
= realtotalpages
;
6588 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6592 #ifndef CONFIG_SPARSEMEM
6594 * Calculate the size of the zone->blockflags rounded to an unsigned long
6595 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6596 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6597 * round what is now in bits to nearest long in bits, then return it in
6600 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6602 unsigned long usemapsize
;
6604 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6605 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6606 usemapsize
= usemapsize
>> pageblock_order
;
6607 usemapsize
*= NR_PAGEBLOCK_BITS
;
6608 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6610 return usemapsize
/ 8;
6613 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6615 unsigned long zone_start_pfn
,
6616 unsigned long zonesize
)
6618 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6619 zone
->pageblock_flags
= NULL
;
6621 zone
->pageblock_flags
=
6622 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
6624 if (!zone
->pageblock_flags
)
6625 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6626 usemapsize
, zone
->name
, pgdat
->node_id
);
6630 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6631 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6632 #endif /* CONFIG_SPARSEMEM */
6634 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6636 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6637 void __init
set_pageblock_order(void)
6641 /* Check that pageblock_nr_pages has not already been setup */
6642 if (pageblock_order
)
6645 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6646 order
= HUGETLB_PAGE_ORDER
;
6648 order
= MAX_ORDER
- 1;
6651 * Assume the largest contiguous order of interest is a huge page.
6652 * This value may be variable depending on boot parameters on IA64 and
6655 pageblock_order
= order
;
6657 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6660 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6661 * is unused as pageblock_order is set at compile-time. See
6662 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6665 void __init
set_pageblock_order(void)
6669 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6671 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6672 unsigned long present_pages
)
6674 unsigned long pages
= spanned_pages
;
6677 * Provide a more accurate estimation if there are holes within
6678 * the zone and SPARSEMEM is in use. If there are holes within the
6679 * zone, each populated memory region may cost us one or two extra
6680 * memmap pages due to alignment because memmap pages for each
6681 * populated regions may not be naturally aligned on page boundary.
6682 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6684 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6685 IS_ENABLED(CONFIG_SPARSEMEM
))
6686 pages
= present_pages
;
6688 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6691 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6692 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6694 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
6696 spin_lock_init(&ds_queue
->split_queue_lock
);
6697 INIT_LIST_HEAD(&ds_queue
->split_queue
);
6698 ds_queue
->split_queue_len
= 0;
6701 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6704 #ifdef CONFIG_COMPACTION
6705 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6707 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6710 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6713 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6715 pgdat_resize_init(pgdat
);
6717 pgdat_init_split_queue(pgdat
);
6718 pgdat_init_kcompactd(pgdat
);
6720 init_waitqueue_head(&pgdat
->kswapd_wait
);
6721 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6723 pgdat_page_ext_init(pgdat
);
6724 spin_lock_init(&pgdat
->lru_lock
);
6725 lruvec_init(&pgdat
->__lruvec
);
6728 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6729 unsigned long remaining_pages
)
6731 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6732 zone_set_nid(zone
, nid
);
6733 zone
->name
= zone_names
[idx
];
6734 zone
->zone_pgdat
= NODE_DATA(nid
);
6735 spin_lock_init(&zone
->lock
);
6736 zone_seqlock_init(zone
);
6737 zone_pcp_init(zone
);
6741 * Set up the zone data structures
6742 * - init pgdat internals
6743 * - init all zones belonging to this node
6745 * NOTE: this function is only called during memory hotplug
6747 #ifdef CONFIG_MEMORY_HOTPLUG
6748 void __ref
free_area_init_core_hotplug(int nid
)
6751 pg_data_t
*pgdat
= NODE_DATA(nid
);
6753 pgdat_init_internals(pgdat
);
6754 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6755 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6760 * Set up the zone data structures:
6761 * - mark all pages reserved
6762 * - mark all memory queues empty
6763 * - clear the memory bitmaps
6765 * NOTE: pgdat should get zeroed by caller.
6766 * NOTE: this function is only called during early init.
6768 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6771 int nid
= pgdat
->node_id
;
6773 pgdat_init_internals(pgdat
);
6774 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6776 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6777 struct zone
*zone
= pgdat
->node_zones
+ j
;
6778 unsigned long size
, freesize
, memmap_pages
;
6779 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6781 size
= zone
->spanned_pages
;
6782 freesize
= zone
->present_pages
;
6785 * Adjust freesize so that it accounts for how much memory
6786 * is used by this zone for memmap. This affects the watermark
6787 * and per-cpu initialisations
6789 memmap_pages
= calc_memmap_size(size
, freesize
);
6790 if (!is_highmem_idx(j
)) {
6791 if (freesize
>= memmap_pages
) {
6792 freesize
-= memmap_pages
;
6795 " %s zone: %lu pages used for memmap\n",
6796 zone_names
[j
], memmap_pages
);
6798 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6799 zone_names
[j
], memmap_pages
, freesize
);
6802 /* Account for reserved pages */
6803 if (j
== 0 && freesize
> dma_reserve
) {
6804 freesize
-= dma_reserve
;
6805 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6806 zone_names
[0], dma_reserve
);
6809 if (!is_highmem_idx(j
))
6810 nr_kernel_pages
+= freesize
;
6811 /* Charge for highmem memmap if there are enough kernel pages */
6812 else if (nr_kernel_pages
> memmap_pages
* 2)
6813 nr_kernel_pages
-= memmap_pages
;
6814 nr_all_pages
+= freesize
;
6817 * Set an approximate value for lowmem here, it will be adjusted
6818 * when the bootmem allocator frees pages into the buddy system.
6819 * And all highmem pages will be managed by the buddy system.
6821 zone_init_internals(zone
, j
, nid
, freesize
);
6826 set_pageblock_order();
6827 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6828 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6829 memmap_init(size
, nid
, j
, zone_start_pfn
);
6833 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6834 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6836 unsigned long __maybe_unused start
= 0;
6837 unsigned long __maybe_unused offset
= 0;
6839 /* Skip empty nodes */
6840 if (!pgdat
->node_spanned_pages
)
6843 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6844 offset
= pgdat
->node_start_pfn
- start
;
6845 /* ia64 gets its own node_mem_map, before this, without bootmem */
6846 if (!pgdat
->node_mem_map
) {
6847 unsigned long size
, end
;
6851 * The zone's endpoints aren't required to be MAX_ORDER
6852 * aligned but the node_mem_map endpoints must be in order
6853 * for the buddy allocator to function correctly.
6855 end
= pgdat_end_pfn(pgdat
);
6856 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6857 size
= (end
- start
) * sizeof(struct page
);
6858 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
6861 panic("Failed to allocate %ld bytes for node %d memory map\n",
6862 size
, pgdat
->node_id
);
6863 pgdat
->node_mem_map
= map
+ offset
;
6865 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6866 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6867 (unsigned long)pgdat
->node_mem_map
);
6868 #ifndef CONFIG_NEED_MULTIPLE_NODES
6870 * With no DISCONTIG, the global mem_map is just set as node 0's
6872 if (pgdat
== NODE_DATA(0)) {
6873 mem_map
= NODE_DATA(0)->node_mem_map
;
6874 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6880 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6881 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6883 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6884 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6886 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6889 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6892 static void __init
free_area_init_node(int nid
)
6894 pg_data_t
*pgdat
= NODE_DATA(nid
);
6895 unsigned long start_pfn
= 0;
6896 unsigned long end_pfn
= 0;
6898 /* pg_data_t should be reset to zero when it's allocated */
6899 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_highest_zoneidx
);
6901 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6903 pgdat
->node_id
= nid
;
6904 pgdat
->node_start_pfn
= start_pfn
;
6905 pgdat
->per_cpu_nodestats
= NULL
;
6907 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6908 (u64
)start_pfn
<< PAGE_SHIFT
,
6909 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6910 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
);
6912 alloc_node_mem_map(pgdat
);
6913 pgdat_set_deferred_range(pgdat
);
6915 free_area_init_core(pgdat
);
6918 void __init
free_area_init_memoryless_node(int nid
)
6920 free_area_init_node(nid
);
6923 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6925 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6926 * PageReserved(). Return the number of struct pages that were initialized.
6928 static u64 __init
init_unavailable_range(unsigned long spfn
, unsigned long epfn
)
6933 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6934 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6935 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6936 + pageblock_nr_pages
- 1;
6940 * Use a fake node/zone (0) for now. Some of these pages
6941 * (in memblock.reserved but not in memblock.memory) will
6942 * get re-initialized via reserve_bootmem_region() later.
6944 __init_single_page(pfn_to_page(pfn
), pfn
, 0, 0);
6945 __SetPageReserved(pfn_to_page(pfn
));
6953 * Only struct pages that are backed by physical memory are zeroed and
6954 * initialized by going through __init_single_page(). But, there are some
6955 * struct pages which are reserved in memblock allocator and their fields
6956 * may be accessed (for example page_to_pfn() on some configuration accesses
6957 * flags). We must explicitly initialize those struct pages.
6959 * This function also addresses a similar issue where struct pages are left
6960 * uninitialized because the physical address range is not covered by
6961 * memblock.memory or memblock.reserved. That could happen when memblock
6962 * layout is manually configured via memmap=, or when the highest physical
6963 * address (max_pfn) does not end on a section boundary.
6965 static void __init
init_unavailable_mem(void)
6967 phys_addr_t start
, end
;
6969 phys_addr_t next
= 0;
6972 * Loop through unavailable ranges not covered by memblock.memory.
6975 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6976 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6978 pgcnt
+= init_unavailable_range(PFN_DOWN(next
),
6984 * Early sections always have a fully populated memmap for the whole
6985 * section - see pfn_valid(). If the last section has holes at the
6986 * end and that section is marked "online", the memmap will be
6987 * considered initialized. Make sure that memmap has a well defined
6990 pgcnt
+= init_unavailable_range(PFN_DOWN(next
),
6991 round_up(max_pfn
, PAGES_PER_SECTION
));
6994 * Struct pages that do not have backing memory. This could be because
6995 * firmware is using some of this memory, or for some other reasons.
6998 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
7001 static inline void __init
init_unavailable_mem(void)
7004 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7006 #if MAX_NUMNODES > 1
7008 * Figure out the number of possible node ids.
7010 void __init
setup_nr_node_ids(void)
7012 unsigned int highest
;
7014 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
7015 nr_node_ids
= highest
+ 1;
7020 * node_map_pfn_alignment - determine the maximum internode alignment
7022 * This function should be called after node map is populated and sorted.
7023 * It calculates the maximum power of two alignment which can distinguish
7026 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7027 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7028 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7029 * shifted, 1GiB is enough and this function will indicate so.
7031 * This is used to test whether pfn -> nid mapping of the chosen memory
7032 * model has fine enough granularity to avoid incorrect mapping for the
7033 * populated node map.
7035 * Return: the determined alignment in pfn's. 0 if there is no alignment
7036 * requirement (single node).
7038 unsigned long __init
node_map_pfn_alignment(void)
7040 unsigned long accl_mask
= 0, last_end
= 0;
7041 unsigned long start
, end
, mask
;
7042 int last_nid
= NUMA_NO_NODE
;
7045 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
7046 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
7053 * Start with a mask granular enough to pin-point to the
7054 * start pfn and tick off bits one-by-one until it becomes
7055 * too coarse to separate the current node from the last.
7057 mask
= ~((1 << __ffs(start
)) - 1);
7058 while (mask
&& last_end
<= (start
& (mask
<< 1)))
7061 /* accumulate all internode masks */
7065 /* convert mask to number of pages */
7066 return ~accl_mask
+ 1;
7070 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7072 * Return: the minimum PFN based on information provided via
7073 * memblock_set_node().
7075 unsigned long __init
find_min_pfn_with_active_regions(void)
7077 return PHYS_PFN(memblock_start_of_DRAM());
7081 * early_calculate_totalpages()
7082 * Sum pages in active regions for movable zone.
7083 * Populate N_MEMORY for calculating usable_nodes.
7085 static unsigned long __init
early_calculate_totalpages(void)
7087 unsigned long totalpages
= 0;
7088 unsigned long start_pfn
, end_pfn
;
7091 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7092 unsigned long pages
= end_pfn
- start_pfn
;
7094 totalpages
+= pages
;
7096 node_set_state(nid
, N_MEMORY
);
7102 * Find the PFN the Movable zone begins in each node. Kernel memory
7103 * is spread evenly between nodes as long as the nodes have enough
7104 * memory. When they don't, some nodes will have more kernelcore than
7107 static void __init
find_zone_movable_pfns_for_nodes(void)
7110 unsigned long usable_startpfn
;
7111 unsigned long kernelcore_node
, kernelcore_remaining
;
7112 /* save the state before borrow the nodemask */
7113 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
7114 unsigned long totalpages
= early_calculate_totalpages();
7115 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
7116 struct memblock_region
*r
;
7118 /* Need to find movable_zone earlier when movable_node is specified. */
7119 find_usable_zone_for_movable();
7122 * If movable_node is specified, ignore kernelcore and movablecore
7125 if (movable_node_is_enabled()) {
7126 for_each_memblock(memory
, r
) {
7127 if (!memblock_is_hotpluggable(r
))
7130 nid
= memblock_get_region_node(r
);
7132 usable_startpfn
= PFN_DOWN(r
->base
);
7133 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7134 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7142 * If kernelcore=mirror is specified, ignore movablecore option
7144 if (mirrored_kernelcore
) {
7145 bool mem_below_4gb_not_mirrored
= false;
7147 for_each_memblock(memory
, r
) {
7148 if (memblock_is_mirror(r
))
7151 nid
= memblock_get_region_node(r
);
7153 usable_startpfn
= memblock_region_memory_base_pfn(r
);
7155 if (usable_startpfn
< 0x100000) {
7156 mem_below_4gb_not_mirrored
= true;
7160 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7161 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7165 if (mem_below_4gb_not_mirrored
)
7166 pr_warn("This configuration results in unmirrored kernel memory.\n");
7172 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7173 * amount of necessary memory.
7175 if (required_kernelcore_percent
)
7176 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
7178 if (required_movablecore_percent
)
7179 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
7183 * If movablecore= was specified, calculate what size of
7184 * kernelcore that corresponds so that memory usable for
7185 * any allocation type is evenly spread. If both kernelcore
7186 * and movablecore are specified, then the value of kernelcore
7187 * will be used for required_kernelcore if it's greater than
7188 * what movablecore would have allowed.
7190 if (required_movablecore
) {
7191 unsigned long corepages
;
7194 * Round-up so that ZONE_MOVABLE is at least as large as what
7195 * was requested by the user
7197 required_movablecore
=
7198 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
7199 required_movablecore
= min(totalpages
, required_movablecore
);
7200 corepages
= totalpages
- required_movablecore
;
7202 required_kernelcore
= max(required_kernelcore
, corepages
);
7206 * If kernelcore was not specified or kernelcore size is larger
7207 * than totalpages, there is no ZONE_MOVABLE.
7209 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
7212 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7213 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
7216 /* Spread kernelcore memory as evenly as possible throughout nodes */
7217 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7218 for_each_node_state(nid
, N_MEMORY
) {
7219 unsigned long start_pfn
, end_pfn
;
7222 * Recalculate kernelcore_node if the division per node
7223 * now exceeds what is necessary to satisfy the requested
7224 * amount of memory for the kernel
7226 if (required_kernelcore
< kernelcore_node
)
7227 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7230 * As the map is walked, we track how much memory is usable
7231 * by the kernel using kernelcore_remaining. When it is
7232 * 0, the rest of the node is usable by ZONE_MOVABLE
7234 kernelcore_remaining
= kernelcore_node
;
7236 /* Go through each range of PFNs within this node */
7237 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7238 unsigned long size_pages
;
7240 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
7241 if (start_pfn
>= end_pfn
)
7244 /* Account for what is only usable for kernelcore */
7245 if (start_pfn
< usable_startpfn
) {
7246 unsigned long kernel_pages
;
7247 kernel_pages
= min(end_pfn
, usable_startpfn
)
7250 kernelcore_remaining
-= min(kernel_pages
,
7251 kernelcore_remaining
);
7252 required_kernelcore
-= min(kernel_pages
,
7253 required_kernelcore
);
7255 /* Continue if range is now fully accounted */
7256 if (end_pfn
<= usable_startpfn
) {
7259 * Push zone_movable_pfn to the end so
7260 * that if we have to rebalance
7261 * kernelcore across nodes, we will
7262 * not double account here
7264 zone_movable_pfn
[nid
] = end_pfn
;
7267 start_pfn
= usable_startpfn
;
7271 * The usable PFN range for ZONE_MOVABLE is from
7272 * start_pfn->end_pfn. Calculate size_pages as the
7273 * number of pages used as kernelcore
7275 size_pages
= end_pfn
- start_pfn
;
7276 if (size_pages
> kernelcore_remaining
)
7277 size_pages
= kernelcore_remaining
;
7278 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7281 * Some kernelcore has been met, update counts and
7282 * break if the kernelcore for this node has been
7285 required_kernelcore
-= min(required_kernelcore
,
7287 kernelcore_remaining
-= size_pages
;
7288 if (!kernelcore_remaining
)
7294 * If there is still required_kernelcore, we do another pass with one
7295 * less node in the count. This will push zone_movable_pfn[nid] further
7296 * along on the nodes that still have memory until kernelcore is
7300 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7304 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7305 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7306 zone_movable_pfn
[nid
] =
7307 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7310 /* restore the node_state */
7311 node_states
[N_MEMORY
] = saved_node_state
;
7314 /* Any regular or high memory on that node ? */
7315 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7317 enum zone_type zone_type
;
7319 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7320 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7321 if (populated_zone(zone
)) {
7322 if (IS_ENABLED(CONFIG_HIGHMEM
))
7323 node_set_state(nid
, N_HIGH_MEMORY
);
7324 if (zone_type
<= ZONE_NORMAL
)
7325 node_set_state(nid
, N_NORMAL_MEMORY
);
7332 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7333 * such cases we allow max_zone_pfn sorted in the descending order
7335 bool __weak
arch_has_descending_max_zone_pfns(void)
7341 * free_area_init - Initialise all pg_data_t and zone data
7342 * @max_zone_pfn: an array of max PFNs for each zone
7344 * This will call free_area_init_node() for each active node in the system.
7345 * Using the page ranges provided by memblock_set_node(), the size of each
7346 * zone in each node and their holes is calculated. If the maximum PFN
7347 * between two adjacent zones match, it is assumed that the zone is empty.
7348 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7349 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7350 * starts where the previous one ended. For example, ZONE_DMA32 starts
7351 * at arch_max_dma_pfn.
7353 void __init
free_area_init(unsigned long *max_zone_pfn
)
7355 unsigned long start_pfn
, end_pfn
;
7359 /* Record where the zone boundaries are */
7360 memset(arch_zone_lowest_possible_pfn
, 0,
7361 sizeof(arch_zone_lowest_possible_pfn
));
7362 memset(arch_zone_highest_possible_pfn
, 0,
7363 sizeof(arch_zone_highest_possible_pfn
));
7365 start_pfn
= find_min_pfn_with_active_regions();
7366 descending
= arch_has_descending_max_zone_pfns();
7368 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7370 zone
= MAX_NR_ZONES
- i
- 1;
7374 if (zone
== ZONE_MOVABLE
)
7377 end_pfn
= max(max_zone_pfn
[zone
], start_pfn
);
7378 arch_zone_lowest_possible_pfn
[zone
] = start_pfn
;
7379 arch_zone_highest_possible_pfn
[zone
] = end_pfn
;
7381 start_pfn
= end_pfn
;
7384 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7385 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7386 find_zone_movable_pfns_for_nodes();
7388 /* Print out the zone ranges */
7389 pr_info("Zone ranges:\n");
7390 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7391 if (i
== ZONE_MOVABLE
)
7393 pr_info(" %-8s ", zone_names
[i
]);
7394 if (arch_zone_lowest_possible_pfn
[i
] ==
7395 arch_zone_highest_possible_pfn
[i
])
7398 pr_cont("[mem %#018Lx-%#018Lx]\n",
7399 (u64
)arch_zone_lowest_possible_pfn
[i
]
7401 ((u64
)arch_zone_highest_possible_pfn
[i
]
7402 << PAGE_SHIFT
) - 1);
7405 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7406 pr_info("Movable zone start for each node\n");
7407 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7408 if (zone_movable_pfn
[i
])
7409 pr_info(" Node %d: %#018Lx\n", i
,
7410 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7414 * Print out the early node map, and initialize the
7415 * subsection-map relative to active online memory ranges to
7416 * enable future "sub-section" extensions of the memory map.
7418 pr_info("Early memory node ranges\n");
7419 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7420 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7421 (u64
)start_pfn
<< PAGE_SHIFT
,
7422 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7423 subsection_map_init(start_pfn
, end_pfn
- start_pfn
);
7426 /* Initialise every node */
7427 mminit_verify_pageflags_layout();
7428 setup_nr_node_ids();
7429 init_unavailable_mem();
7430 for_each_online_node(nid
) {
7431 pg_data_t
*pgdat
= NODE_DATA(nid
);
7432 free_area_init_node(nid
);
7434 /* Any memory on that node */
7435 if (pgdat
->node_present_pages
)
7436 node_set_state(nid
, N_MEMORY
);
7437 check_for_memory(pgdat
, nid
);
7441 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7442 unsigned long *percent
)
7444 unsigned long long coremem
;
7450 /* Value may be a percentage of total memory, otherwise bytes */
7451 coremem
= simple_strtoull(p
, &endptr
, 0);
7452 if (*endptr
== '%') {
7453 /* Paranoid check for percent values greater than 100 */
7454 WARN_ON(coremem
> 100);
7458 coremem
= memparse(p
, &p
);
7459 /* Paranoid check that UL is enough for the coremem value */
7460 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7462 *core
= coremem
>> PAGE_SHIFT
;
7469 * kernelcore=size sets the amount of memory for use for allocations that
7470 * cannot be reclaimed or migrated.
7472 static int __init
cmdline_parse_kernelcore(char *p
)
7474 /* parse kernelcore=mirror */
7475 if (parse_option_str(p
, "mirror")) {
7476 mirrored_kernelcore
= true;
7480 return cmdline_parse_core(p
, &required_kernelcore
,
7481 &required_kernelcore_percent
);
7485 * movablecore=size sets the amount of memory for use for allocations that
7486 * can be reclaimed or migrated.
7488 static int __init
cmdline_parse_movablecore(char *p
)
7490 return cmdline_parse_core(p
, &required_movablecore
,
7491 &required_movablecore_percent
);
7494 early_param("kernelcore", cmdline_parse_kernelcore
);
7495 early_param("movablecore", cmdline_parse_movablecore
);
7497 void adjust_managed_page_count(struct page
*page
, long count
)
7499 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7500 totalram_pages_add(count
);
7501 #ifdef CONFIG_HIGHMEM
7502 if (PageHighMem(page
))
7503 totalhigh_pages_add(count
);
7506 EXPORT_SYMBOL(adjust_managed_page_count
);
7508 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7511 unsigned long pages
= 0;
7513 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7514 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7515 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7516 struct page
*page
= virt_to_page(pos
);
7517 void *direct_map_addr
;
7520 * 'direct_map_addr' might be different from 'pos'
7521 * because some architectures' virt_to_page()
7522 * work with aliases. Getting the direct map
7523 * address ensures that we get a _writeable_
7524 * alias for the memset().
7526 direct_map_addr
= page_address(page
);
7527 if ((unsigned int)poison
<= 0xFF)
7528 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7530 free_reserved_page(page
);
7534 pr_info("Freeing %s memory: %ldK\n",
7535 s
, pages
<< (PAGE_SHIFT
- 10));
7540 #ifdef CONFIG_HIGHMEM
7541 void free_highmem_page(struct page
*page
)
7543 __free_reserved_page(page
);
7544 totalram_pages_inc();
7545 atomic_long_inc(&page_zone(page
)->managed_pages
);
7546 totalhigh_pages_inc();
7551 void __init
mem_init_print_info(const char *str
)
7553 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7554 unsigned long init_code_size
, init_data_size
;
7556 physpages
= get_num_physpages();
7557 codesize
= _etext
- _stext
;
7558 datasize
= _edata
- _sdata
;
7559 rosize
= __end_rodata
- __start_rodata
;
7560 bss_size
= __bss_stop
- __bss_start
;
7561 init_data_size
= __init_end
- __init_begin
;
7562 init_code_size
= _einittext
- _sinittext
;
7565 * Detect special cases and adjust section sizes accordingly:
7566 * 1) .init.* may be embedded into .data sections
7567 * 2) .init.text.* may be out of [__init_begin, __init_end],
7568 * please refer to arch/tile/kernel/vmlinux.lds.S.
7569 * 3) .rodata.* may be embedded into .text or .data sections.
7571 #define adj_init_size(start, end, size, pos, adj) \
7573 if (start <= pos && pos < end && size > adj) \
7577 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7578 _sinittext
, init_code_size
);
7579 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7580 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7581 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7582 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7584 #undef adj_init_size
7586 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7587 #ifdef CONFIG_HIGHMEM
7591 nr_free_pages() << (PAGE_SHIFT
- 10),
7592 physpages
<< (PAGE_SHIFT
- 10),
7593 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7594 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7595 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7596 totalcma_pages
<< (PAGE_SHIFT
- 10),
7597 #ifdef CONFIG_HIGHMEM
7598 totalhigh_pages() << (PAGE_SHIFT
- 10),
7600 str
? ", " : "", str
? str
: "");
7604 * set_dma_reserve - set the specified number of pages reserved in the first zone
7605 * @new_dma_reserve: The number of pages to mark reserved
7607 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7608 * In the DMA zone, a significant percentage may be consumed by kernel image
7609 * and other unfreeable allocations which can skew the watermarks badly. This
7610 * function may optionally be used to account for unfreeable pages in the
7611 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7612 * smaller per-cpu batchsize.
7614 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7616 dma_reserve
= new_dma_reserve
;
7619 static int page_alloc_cpu_dead(unsigned int cpu
)
7622 lru_add_drain_cpu(cpu
);
7626 * Spill the event counters of the dead processor
7627 * into the current processors event counters.
7628 * This artificially elevates the count of the current
7631 vm_events_fold_cpu(cpu
);
7634 * Zero the differential counters of the dead processor
7635 * so that the vm statistics are consistent.
7637 * This is only okay since the processor is dead and cannot
7638 * race with what we are doing.
7640 cpu_vm_stats_fold(cpu
);
7645 int hashdist
= HASHDIST_DEFAULT
;
7647 static int __init
set_hashdist(char *str
)
7651 hashdist
= simple_strtoul(str
, &str
, 0);
7654 __setup("hashdist=", set_hashdist
);
7657 void __init
page_alloc_init(void)
7662 if (num_node_state(N_MEMORY
) == 1)
7666 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7667 "mm/page_alloc:dead", NULL
,
7668 page_alloc_cpu_dead
);
7673 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7674 * or min_free_kbytes changes.
7676 static void calculate_totalreserve_pages(void)
7678 struct pglist_data
*pgdat
;
7679 unsigned long reserve_pages
= 0;
7680 enum zone_type i
, j
;
7682 for_each_online_pgdat(pgdat
) {
7684 pgdat
->totalreserve_pages
= 0;
7686 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7687 struct zone
*zone
= pgdat
->node_zones
+ i
;
7689 unsigned long managed_pages
= zone_managed_pages(zone
);
7691 /* Find valid and maximum lowmem_reserve in the zone */
7692 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7693 if (zone
->lowmem_reserve
[j
] > max
)
7694 max
= zone
->lowmem_reserve
[j
];
7697 /* we treat the high watermark as reserved pages. */
7698 max
+= high_wmark_pages(zone
);
7700 if (max
> managed_pages
)
7701 max
= managed_pages
;
7703 pgdat
->totalreserve_pages
+= max
;
7705 reserve_pages
+= max
;
7708 totalreserve_pages
= reserve_pages
;
7712 * setup_per_zone_lowmem_reserve - called whenever
7713 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7714 * has a correct pages reserved value, so an adequate number of
7715 * pages are left in the zone after a successful __alloc_pages().
7717 static void setup_per_zone_lowmem_reserve(void)
7719 struct pglist_data
*pgdat
;
7720 enum zone_type j
, idx
;
7722 for_each_online_pgdat(pgdat
) {
7723 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7724 struct zone
*zone
= pgdat
->node_zones
+ j
;
7725 unsigned long managed_pages
= zone_managed_pages(zone
);
7727 zone
->lowmem_reserve
[j
] = 0;
7731 struct zone
*lower_zone
;
7734 lower_zone
= pgdat
->node_zones
+ idx
;
7736 if (!sysctl_lowmem_reserve_ratio
[idx
] ||
7737 !zone_managed_pages(lower_zone
)) {
7738 lower_zone
->lowmem_reserve
[j
] = 0;
7741 lower_zone
->lowmem_reserve
[j
] =
7742 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7744 managed_pages
+= zone_managed_pages(lower_zone
);
7749 /* update totalreserve_pages */
7750 calculate_totalreserve_pages();
7753 static void __setup_per_zone_wmarks(void)
7755 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7756 unsigned long lowmem_pages
= 0;
7758 unsigned long flags
;
7760 /* Calculate total number of !ZONE_HIGHMEM pages */
7761 for_each_zone(zone
) {
7762 if (!is_highmem(zone
))
7763 lowmem_pages
+= zone_managed_pages(zone
);
7766 for_each_zone(zone
) {
7769 spin_lock_irqsave(&zone
->lock
, flags
);
7770 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7771 do_div(tmp
, lowmem_pages
);
7772 if (is_highmem(zone
)) {
7774 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7775 * need highmem pages, so cap pages_min to a small
7778 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7779 * deltas control async page reclaim, and so should
7780 * not be capped for highmem.
7782 unsigned long min_pages
;
7784 min_pages
= zone_managed_pages(zone
) / 1024;
7785 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7786 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7789 * If it's a lowmem zone, reserve a number of pages
7790 * proportionate to the zone's size.
7792 zone
->_watermark
[WMARK_MIN
] = tmp
;
7796 * Set the kswapd watermarks distance according to the
7797 * scale factor in proportion to available memory, but
7798 * ensure a minimum size on small systems.
7800 tmp
= max_t(u64
, tmp
>> 2,
7801 mult_frac(zone_managed_pages(zone
),
7802 watermark_scale_factor
, 10000));
7804 zone
->watermark_boost
= 0;
7805 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7806 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7808 spin_unlock_irqrestore(&zone
->lock
, flags
);
7811 /* update totalreserve_pages */
7812 calculate_totalreserve_pages();
7816 * setup_per_zone_wmarks - called when min_free_kbytes changes
7817 * or when memory is hot-{added|removed}
7819 * Ensures that the watermark[min,low,high] values for each zone are set
7820 * correctly with respect to min_free_kbytes.
7822 void setup_per_zone_wmarks(void)
7824 static DEFINE_SPINLOCK(lock
);
7827 __setup_per_zone_wmarks();
7832 * Initialise min_free_kbytes.
7834 * For small machines we want it small (128k min). For large machines
7835 * we want it large (64MB max). But it is not linear, because network
7836 * bandwidth does not increase linearly with machine size. We use
7838 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7839 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7855 int __meminit
init_per_zone_wmark_min(void)
7857 unsigned long lowmem_kbytes
;
7858 int new_min_free_kbytes
;
7860 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7861 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7863 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7864 min_free_kbytes
= new_min_free_kbytes
;
7865 if (min_free_kbytes
< 128)
7866 min_free_kbytes
= 128;
7867 if (min_free_kbytes
> 262144)
7868 min_free_kbytes
= 262144;
7870 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7871 new_min_free_kbytes
, user_min_free_kbytes
);
7873 setup_per_zone_wmarks();
7874 refresh_zone_stat_thresholds();
7875 setup_per_zone_lowmem_reserve();
7878 setup_min_unmapped_ratio();
7879 setup_min_slab_ratio();
7884 core_initcall(init_per_zone_wmark_min
)
7887 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7888 * that we can call two helper functions whenever min_free_kbytes
7891 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7892 void *buffer
, size_t *length
, loff_t
*ppos
)
7896 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7901 user_min_free_kbytes
= min_free_kbytes
;
7902 setup_per_zone_wmarks();
7907 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7908 void *buffer
, size_t *length
, loff_t
*ppos
)
7912 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7917 setup_per_zone_wmarks();
7923 static void setup_min_unmapped_ratio(void)
7928 for_each_online_pgdat(pgdat
)
7929 pgdat
->min_unmapped_pages
= 0;
7932 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
7933 sysctl_min_unmapped_ratio
) / 100;
7937 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7938 void *buffer
, size_t *length
, loff_t
*ppos
)
7942 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7946 setup_min_unmapped_ratio();
7951 static void setup_min_slab_ratio(void)
7956 for_each_online_pgdat(pgdat
)
7957 pgdat
->min_slab_pages
= 0;
7960 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
7961 sysctl_min_slab_ratio
) / 100;
7964 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7965 void *buffer
, size_t *length
, loff_t
*ppos
)
7969 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7973 setup_min_slab_ratio();
7980 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7981 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7982 * whenever sysctl_lowmem_reserve_ratio changes.
7984 * The reserve ratio obviously has absolutely no relation with the
7985 * minimum watermarks. The lowmem reserve ratio can only make sense
7986 * if in function of the boot time zone sizes.
7988 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7989 void *buffer
, size_t *length
, loff_t
*ppos
)
7993 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7995 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7996 if (sysctl_lowmem_reserve_ratio
[i
] < 1)
7997 sysctl_lowmem_reserve_ratio
[i
] = 0;
8000 setup_per_zone_lowmem_reserve();
8004 static void __zone_pcp_update(struct zone
*zone
)
8008 for_each_possible_cpu(cpu
)
8009 pageset_set_high_and_batch(zone
,
8010 per_cpu_ptr(zone
->pageset
, cpu
));
8014 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8015 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8016 * pagelist can have before it gets flushed back to buddy allocator.
8018 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
8019 void *buffer
, size_t *length
, loff_t
*ppos
)
8022 int old_percpu_pagelist_fraction
;
8025 mutex_lock(&pcp_batch_high_lock
);
8026 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
8028 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8029 if (!write
|| ret
< 0)
8032 /* Sanity checking to avoid pcp imbalance */
8033 if (percpu_pagelist_fraction
&&
8034 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
8035 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
8041 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
8044 for_each_populated_zone(zone
)
8045 __zone_pcp_update(zone
);
8047 mutex_unlock(&pcp_batch_high_lock
);
8051 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8053 * Returns the number of pages that arch has reserved but
8054 * is not known to alloc_large_system_hash().
8056 static unsigned long __init
arch_reserved_kernel_pages(void)
8063 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8064 * machines. As memory size is increased the scale is also increased but at
8065 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8066 * quadruples the scale is increased by one, which means the size of hash table
8067 * only doubles, instead of quadrupling as well.
8068 * Because 32-bit systems cannot have large physical memory, where this scaling
8069 * makes sense, it is disabled on such platforms.
8071 #if __BITS_PER_LONG > 32
8072 #define ADAPT_SCALE_BASE (64ul << 30)
8073 #define ADAPT_SCALE_SHIFT 2
8074 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8078 * allocate a large system hash table from bootmem
8079 * - it is assumed that the hash table must contain an exact power-of-2
8080 * quantity of entries
8081 * - limit is the number of hash buckets, not the total allocation size
8083 void *__init
alloc_large_system_hash(const char *tablename
,
8084 unsigned long bucketsize
,
8085 unsigned long numentries
,
8088 unsigned int *_hash_shift
,
8089 unsigned int *_hash_mask
,
8090 unsigned long low_limit
,
8091 unsigned long high_limit
)
8093 unsigned long long max
= high_limit
;
8094 unsigned long log2qty
, size
;
8099 /* allow the kernel cmdline to have a say */
8101 /* round applicable memory size up to nearest megabyte */
8102 numentries
= nr_kernel_pages
;
8103 numentries
-= arch_reserved_kernel_pages();
8105 /* It isn't necessary when PAGE_SIZE >= 1MB */
8106 if (PAGE_SHIFT
< 20)
8107 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
8109 #if __BITS_PER_LONG > 32
8111 unsigned long adapt
;
8113 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
8114 adapt
<<= ADAPT_SCALE_SHIFT
)
8119 /* limit to 1 bucket per 2^scale bytes of low memory */
8120 if (scale
> PAGE_SHIFT
)
8121 numentries
>>= (scale
- PAGE_SHIFT
);
8123 numentries
<<= (PAGE_SHIFT
- scale
);
8125 /* Make sure we've got at least a 0-order allocation.. */
8126 if (unlikely(flags
& HASH_SMALL
)) {
8127 /* Makes no sense without HASH_EARLY */
8128 WARN_ON(!(flags
& HASH_EARLY
));
8129 if (!(numentries
>> *_hash_shift
)) {
8130 numentries
= 1UL << *_hash_shift
;
8131 BUG_ON(!numentries
);
8133 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
8134 numentries
= PAGE_SIZE
/ bucketsize
;
8136 numentries
= roundup_pow_of_two(numentries
);
8138 /* limit allocation size to 1/16 total memory by default */
8140 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
8141 do_div(max
, bucketsize
);
8143 max
= min(max
, 0x80000000ULL
);
8145 if (numentries
< low_limit
)
8146 numentries
= low_limit
;
8147 if (numentries
> max
)
8150 log2qty
= ilog2(numentries
);
8152 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
8155 size
= bucketsize
<< log2qty
;
8156 if (flags
& HASH_EARLY
) {
8157 if (flags
& HASH_ZERO
)
8158 table
= memblock_alloc(size
, SMP_CACHE_BYTES
);
8160 table
= memblock_alloc_raw(size
,
8162 } else if (get_order(size
) >= MAX_ORDER
|| hashdist
) {
8163 table
= __vmalloc(size
, gfp_flags
);
8167 * If bucketsize is not a power-of-two, we may free
8168 * some pages at the end of hash table which
8169 * alloc_pages_exact() automatically does
8171 table
= alloc_pages_exact(size
, gfp_flags
);
8172 kmemleak_alloc(table
, size
, 1, gfp_flags
);
8174 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
8177 panic("Failed to allocate %s hash table\n", tablename
);
8179 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8180 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
,
8181 virt
? "vmalloc" : "linear");
8184 *_hash_shift
= log2qty
;
8186 *_hash_mask
= (1 << log2qty
) - 1;
8192 * This function checks whether pageblock includes unmovable pages or not.
8194 * PageLRU check without isolation or lru_lock could race so that
8195 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8196 * check without lock_page also may miss some movable non-lru pages at
8197 * race condition. So you can't expect this function should be exact.
8199 * Returns a page without holding a reference. If the caller wants to
8200 * dereference that page (e.g., dumping), it has to make sure that that it
8201 * cannot get removed (e.g., via memory unplug) concurrently.
8204 struct page
*has_unmovable_pages(struct zone
*zone
, struct page
*page
,
8205 int migratetype
, int flags
)
8207 unsigned long iter
= 0;
8208 unsigned long pfn
= page_to_pfn(page
);
8211 * TODO we could make this much more efficient by not checking every
8212 * page in the range if we know all of them are in MOVABLE_ZONE and
8213 * that the movable zone guarantees that pages are migratable but
8214 * the later is not the case right now unfortunatelly. E.g. movablecore
8215 * can still lead to having bootmem allocations in zone_movable.
8218 if (is_migrate_cma_page(page
)) {
8220 * CMA allocations (alloc_contig_range) really need to mark
8221 * isolate CMA pageblocks even when they are not movable in fact
8222 * so consider them movable here.
8224 if (is_migrate_cma(migratetype
))
8230 for (; iter
< pageblock_nr_pages
; iter
++) {
8231 if (!pfn_valid_within(pfn
+ iter
))
8234 page
= pfn_to_page(pfn
+ iter
);
8236 if (PageReserved(page
))
8240 * If the zone is movable and we have ruled out all reserved
8241 * pages then it should be reasonably safe to assume the rest
8244 if (zone_idx(zone
) == ZONE_MOVABLE
)
8248 * Hugepages are not in LRU lists, but they're movable.
8249 * THPs are on the LRU, but need to be counted as #small pages.
8250 * We need not scan over tail pages because we don't
8251 * handle each tail page individually in migration.
8253 if (PageHuge(page
) || PageTransCompound(page
)) {
8254 struct page
*head
= compound_head(page
);
8255 unsigned int skip_pages
;
8257 if (PageHuge(page
)) {
8258 if (!hugepage_migration_supported(page_hstate(head
)))
8260 } else if (!PageLRU(head
) && !__PageMovable(head
)) {
8264 skip_pages
= compound_nr(head
) - (page
- head
);
8265 iter
+= skip_pages
- 1;
8270 * We can't use page_count without pin a page
8271 * because another CPU can free compound page.
8272 * This check already skips compound tails of THP
8273 * because their page->_refcount is zero at all time.
8275 if (!page_ref_count(page
)) {
8276 if (PageBuddy(page
))
8277 iter
+= (1 << page_order(page
)) - 1;
8282 * The HWPoisoned page may be not in buddy system, and
8283 * page_count() is not 0.
8285 if ((flags
& MEMORY_OFFLINE
) && PageHWPoison(page
))
8289 * We treat all PageOffline() pages as movable when offlining
8290 * to give drivers a chance to decrement their reference count
8291 * in MEM_GOING_OFFLINE in order to indicate that these pages
8292 * can be offlined as there are no direct references anymore.
8293 * For actually unmovable PageOffline() where the driver does
8294 * not support this, we will fail later when trying to actually
8295 * move these pages that still have a reference count > 0.
8296 * (false negatives in this function only)
8298 if ((flags
& MEMORY_OFFLINE
) && PageOffline(page
))
8301 if (__PageMovable(page
) || PageLRU(page
))
8305 * If there are RECLAIMABLE pages, we need to check
8306 * it. But now, memory offline itself doesn't call
8307 * shrink_node_slabs() and it still to be fixed.
8310 * If the page is not RAM, page_count()should be 0.
8311 * we don't need more check. This is an _used_ not-movable page.
8313 * The problematic thing here is PG_reserved pages. PG_reserved
8314 * is set to both of a memory hole page and a _used_ kernel
8322 #ifdef CONFIG_CONTIG_ALLOC
8323 static unsigned long pfn_max_align_down(unsigned long pfn
)
8325 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8326 pageblock_nr_pages
) - 1);
8329 static unsigned long pfn_max_align_up(unsigned long pfn
)
8331 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8332 pageblock_nr_pages
));
8335 /* [start, end) must belong to a single zone. */
8336 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8337 unsigned long start
, unsigned long end
)
8339 /* This function is based on compact_zone() from compaction.c. */
8340 unsigned int nr_reclaimed
;
8341 unsigned long pfn
= start
;
8342 unsigned int tries
= 0;
8347 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8348 if (fatal_signal_pending(current
)) {
8353 if (list_empty(&cc
->migratepages
)) {
8354 cc
->nr_migratepages
= 0;
8355 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8361 } else if (++tries
== 5) {
8362 ret
= ret
< 0 ? ret
: -EBUSY
;
8366 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8368 cc
->nr_migratepages
-= nr_reclaimed
;
8370 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
8371 NULL
, 0, cc
->mode
, MR_CONTIG_RANGE
);
8374 putback_movable_pages(&cc
->migratepages
);
8381 * alloc_contig_range() -- tries to allocate given range of pages
8382 * @start: start PFN to allocate
8383 * @end: one-past-the-last PFN to allocate
8384 * @migratetype: migratetype of the underlaying pageblocks (either
8385 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8386 * in range must have the same migratetype and it must
8387 * be either of the two.
8388 * @gfp_mask: GFP mask to use during compaction
8390 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8391 * aligned. The PFN range must belong to a single zone.
8393 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8394 * pageblocks in the range. Once isolated, the pageblocks should not
8395 * be modified by others.
8397 * Return: zero on success or negative error code. On success all
8398 * pages which PFN is in [start, end) are allocated for the caller and
8399 * need to be freed with free_contig_range().
8401 int alloc_contig_range(unsigned long start
, unsigned long end
,
8402 unsigned migratetype
, gfp_t gfp_mask
)
8404 unsigned long outer_start
, outer_end
;
8408 struct compact_control cc
= {
8409 .nr_migratepages
= 0,
8411 .zone
= page_zone(pfn_to_page(start
)),
8412 .mode
= MIGRATE_SYNC
,
8413 .ignore_skip_hint
= true,
8414 .no_set_skip_hint
= true,
8415 .gfp_mask
= current_gfp_context(gfp_mask
),
8416 .alloc_contig
= true,
8418 INIT_LIST_HEAD(&cc
.migratepages
);
8421 * What we do here is we mark all pageblocks in range as
8422 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8423 * have different sizes, and due to the way page allocator
8424 * work, we align the range to biggest of the two pages so
8425 * that page allocator won't try to merge buddies from
8426 * different pageblocks and change MIGRATE_ISOLATE to some
8427 * other migration type.
8429 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8430 * migrate the pages from an unaligned range (ie. pages that
8431 * we are interested in). This will put all the pages in
8432 * range back to page allocator as MIGRATE_ISOLATE.
8434 * When this is done, we take the pages in range from page
8435 * allocator removing them from the buddy system. This way
8436 * page allocator will never consider using them.
8438 * This lets us mark the pageblocks back as
8439 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8440 * aligned range but not in the unaligned, original range are
8441 * put back to page allocator so that buddy can use them.
8444 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8445 pfn_max_align_up(end
), migratetype
, 0);
8450 * In case of -EBUSY, we'd like to know which page causes problem.
8451 * So, just fall through. test_pages_isolated() has a tracepoint
8452 * which will report the busy page.
8454 * It is possible that busy pages could become available before
8455 * the call to test_pages_isolated, and the range will actually be
8456 * allocated. So, if we fall through be sure to clear ret so that
8457 * -EBUSY is not accidentally used or returned to caller.
8459 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8460 if (ret
&& ret
!= -EBUSY
)
8465 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8466 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8467 * more, all pages in [start, end) are free in page allocator.
8468 * What we are going to do is to allocate all pages from
8469 * [start, end) (that is remove them from page allocator).
8471 * The only problem is that pages at the beginning and at the
8472 * end of interesting range may be not aligned with pages that
8473 * page allocator holds, ie. they can be part of higher order
8474 * pages. Because of this, we reserve the bigger range and
8475 * once this is done free the pages we are not interested in.
8477 * We don't have to hold zone->lock here because the pages are
8478 * isolated thus they won't get removed from buddy.
8481 lru_add_drain_all();
8484 outer_start
= start
;
8485 while (!PageBuddy(pfn_to_page(outer_start
))) {
8486 if (++order
>= MAX_ORDER
) {
8487 outer_start
= start
;
8490 outer_start
&= ~0UL << order
;
8493 if (outer_start
!= start
) {
8494 order
= page_order(pfn_to_page(outer_start
));
8497 * outer_start page could be small order buddy page and
8498 * it doesn't include start page. Adjust outer_start
8499 * in this case to report failed page properly
8500 * on tracepoint in test_pages_isolated()
8502 if (outer_start
+ (1UL << order
) <= start
)
8503 outer_start
= start
;
8506 /* Make sure the range is really isolated. */
8507 if (test_pages_isolated(outer_start
, end
, 0)) {
8508 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8509 __func__
, outer_start
, end
);
8514 /* Grab isolated pages from freelists. */
8515 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8521 /* Free head and tail (if any) */
8522 if (start
!= outer_start
)
8523 free_contig_range(outer_start
, start
- outer_start
);
8524 if (end
!= outer_end
)
8525 free_contig_range(end
, outer_end
- end
);
8528 undo_isolate_page_range(pfn_max_align_down(start
),
8529 pfn_max_align_up(end
), migratetype
);
8532 EXPORT_SYMBOL(alloc_contig_range
);
8534 static int __alloc_contig_pages(unsigned long start_pfn
,
8535 unsigned long nr_pages
, gfp_t gfp_mask
)
8537 unsigned long end_pfn
= start_pfn
+ nr_pages
;
8539 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
8543 static bool pfn_range_valid_contig(struct zone
*z
, unsigned long start_pfn
,
8544 unsigned long nr_pages
)
8546 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
8549 for (i
= start_pfn
; i
< end_pfn
; i
++) {
8550 page
= pfn_to_online_page(i
);
8554 if (page_zone(page
) != z
)
8557 if (PageReserved(page
))
8560 if (page_count(page
) > 0)
8569 static bool zone_spans_last_pfn(const struct zone
*zone
,
8570 unsigned long start_pfn
, unsigned long nr_pages
)
8572 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
8574 return zone_spans_pfn(zone
, last_pfn
);
8578 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8579 * @nr_pages: Number of contiguous pages to allocate
8580 * @gfp_mask: GFP mask to limit search and used during compaction
8582 * @nodemask: Mask for other possible nodes
8584 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8585 * on an applicable zonelist to find a contiguous pfn range which can then be
8586 * tried for allocation with alloc_contig_range(). This routine is intended
8587 * for allocation requests which can not be fulfilled with the buddy allocator.
8589 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8590 * power of two then the alignment is guaranteed to be to the given nr_pages
8591 * (e.g. 1GB request would be aligned to 1GB).
8593 * Allocated pages can be freed with free_contig_range() or by manually calling
8594 * __free_page() on each allocated page.
8596 * Return: pointer to contiguous pages on success, or NULL if not successful.
8598 struct page
*alloc_contig_pages(unsigned long nr_pages
, gfp_t gfp_mask
,
8599 int nid
, nodemask_t
*nodemask
)
8601 unsigned long ret
, pfn
, flags
;
8602 struct zonelist
*zonelist
;
8606 zonelist
= node_zonelist(nid
, gfp_mask
);
8607 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
8608 gfp_zone(gfp_mask
), nodemask
) {
8609 spin_lock_irqsave(&zone
->lock
, flags
);
8611 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
8612 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
8613 if (pfn_range_valid_contig(zone
, pfn
, nr_pages
)) {
8615 * We release the zone lock here because
8616 * alloc_contig_range() will also lock the zone
8617 * at some point. If there's an allocation
8618 * spinning on this lock, it may win the race
8619 * and cause alloc_contig_range() to fail...
8621 spin_unlock_irqrestore(&zone
->lock
, flags
);
8622 ret
= __alloc_contig_pages(pfn
, nr_pages
,
8625 return pfn_to_page(pfn
);
8626 spin_lock_irqsave(&zone
->lock
, flags
);
8630 spin_unlock_irqrestore(&zone
->lock
, flags
);
8634 #endif /* CONFIG_CONTIG_ALLOC */
8636 void free_contig_range(unsigned long pfn
, unsigned int nr_pages
)
8638 unsigned int count
= 0;
8640 for (; nr_pages
--; pfn
++) {
8641 struct page
*page
= pfn_to_page(pfn
);
8643 count
+= page_count(page
) != 1;
8646 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8648 EXPORT_SYMBOL(free_contig_range
);
8651 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8652 * page high values need to be recalulated.
8654 void __meminit
zone_pcp_update(struct zone
*zone
)
8656 mutex_lock(&pcp_batch_high_lock
);
8657 __zone_pcp_update(zone
);
8658 mutex_unlock(&pcp_batch_high_lock
);
8661 void zone_pcp_reset(struct zone
*zone
)
8663 unsigned long flags
;
8665 struct per_cpu_pageset
*pset
;
8667 /* avoid races with drain_pages() */
8668 local_irq_save(flags
);
8669 if (zone
->pageset
!= &boot_pageset
) {
8670 for_each_online_cpu(cpu
) {
8671 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8672 drain_zonestat(zone
, pset
);
8674 free_percpu(zone
->pageset
);
8675 zone
->pageset
= &boot_pageset
;
8677 local_irq_restore(flags
);
8680 #ifdef CONFIG_MEMORY_HOTREMOVE
8682 * All pages in the range must be in a single zone and isolated
8683 * before calling this.
8686 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8692 unsigned long flags
;
8693 unsigned long offlined_pages
= 0;
8695 /* find the first valid pfn */
8696 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8700 return offlined_pages
;
8702 offline_mem_sections(pfn
, end_pfn
);
8703 zone
= page_zone(pfn_to_page(pfn
));
8704 spin_lock_irqsave(&zone
->lock
, flags
);
8706 while (pfn
< end_pfn
) {
8707 if (!pfn_valid(pfn
)) {
8711 page
= pfn_to_page(pfn
);
8713 * The HWPoisoned page may be not in buddy system, and
8714 * page_count() is not 0.
8716 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8722 * At this point all remaining PageOffline() pages have a
8723 * reference count of 0 and can simply be skipped.
8725 if (PageOffline(page
)) {
8726 BUG_ON(page_count(page
));
8727 BUG_ON(PageBuddy(page
));
8733 BUG_ON(page_count(page
));
8734 BUG_ON(!PageBuddy(page
));
8735 order
= page_order(page
);
8736 offlined_pages
+= 1 << order
;
8737 del_page_from_free_list(page
, zone
, order
);
8738 pfn
+= (1 << order
);
8740 spin_unlock_irqrestore(&zone
->lock
, flags
);
8742 return offlined_pages
;
8746 bool is_free_buddy_page(struct page
*page
)
8748 struct zone
*zone
= page_zone(page
);
8749 unsigned long pfn
= page_to_pfn(page
);
8750 unsigned long flags
;
8753 spin_lock_irqsave(&zone
->lock
, flags
);
8754 for (order
= 0; order
< MAX_ORDER
; order
++) {
8755 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8757 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8760 spin_unlock_irqrestore(&zone
->lock
, flags
);
8762 return order
< MAX_ORDER
;
8765 #ifdef CONFIG_MEMORY_FAILURE
8767 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8768 * test is performed under the zone lock to prevent a race against page
8771 bool set_hwpoison_free_buddy_page(struct page
*page
)
8773 struct zone
*zone
= page_zone(page
);
8774 unsigned long pfn
= page_to_pfn(page
);
8775 unsigned long flags
;
8777 bool hwpoisoned
= false;
8779 spin_lock_irqsave(&zone
->lock
, flags
);
8780 for (order
= 0; order
< MAX_ORDER
; order
++) {
8781 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8783 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8784 if (!TestSetPageHWPoison(page
))
8789 spin_unlock_irqrestore(&zone
->lock
, flags
);