1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
78 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
79 static DEFINE_MUTEX(pcp_batch_high_lock
);
80 #define MIN_PERCPU_PAGELIST_FRACTION (8)
82 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
83 DEFINE_PER_CPU(int, numa_node
);
84 EXPORT_PER_CPU_SYMBOL(numa_node
);
87 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
89 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
91 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
92 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
93 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
94 * defined in <linux/topology.h>.
96 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
97 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
98 int _node_numa_mem_
[MAX_NUMNODES
];
101 /* work_structs for global per-cpu drains */
104 struct work_struct work
;
106 DEFINE_MUTEX(pcpu_drain_mutex
);
107 DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
109 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
110 volatile unsigned long latent_entropy __latent_entropy
;
111 EXPORT_SYMBOL(latent_entropy
);
115 * Array of node states.
117 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
118 [N_POSSIBLE
] = NODE_MASK_ALL
,
119 [N_ONLINE
] = { { [0] = 1UL } },
121 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
122 #ifdef CONFIG_HIGHMEM
123 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
125 [N_MEMORY
] = { { [0] = 1UL } },
126 [N_CPU
] = { { [0] = 1UL } },
129 EXPORT_SYMBOL(node_states
);
131 atomic_long_t _totalram_pages __read_mostly
;
132 EXPORT_SYMBOL(_totalram_pages
);
133 unsigned long totalreserve_pages __read_mostly
;
134 unsigned long totalcma_pages __read_mostly
;
136 int percpu_pagelist_fraction
;
137 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
138 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
139 DEFINE_STATIC_KEY_TRUE(init_on_alloc
);
141 DEFINE_STATIC_KEY_FALSE(init_on_alloc
);
143 EXPORT_SYMBOL(init_on_alloc
);
145 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
146 DEFINE_STATIC_KEY_TRUE(init_on_free
);
148 DEFINE_STATIC_KEY_FALSE(init_on_free
);
150 EXPORT_SYMBOL(init_on_free
);
152 static int __init
early_init_on_alloc(char *buf
)
159 ret
= kstrtobool(buf
, &bool_result
);
160 if (bool_result
&& page_poisoning_enabled())
161 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
163 static_branch_enable(&init_on_alloc
);
165 static_branch_disable(&init_on_alloc
);
168 early_param("init_on_alloc", early_init_on_alloc
);
170 static int __init
early_init_on_free(char *buf
)
177 ret
= kstrtobool(buf
, &bool_result
);
178 if (bool_result
&& page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
181 static_branch_enable(&init_on_free
);
183 static_branch_disable(&init_on_free
);
186 early_param("init_on_free", early_init_on_free
);
189 * A cached value of the page's pageblock's migratetype, used when the page is
190 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
191 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
192 * Also the migratetype set in the page does not necessarily match the pcplist
193 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
194 * other index - this ensures that it will be put on the correct CMA freelist.
196 static inline int get_pcppage_migratetype(struct page
*page
)
201 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
203 page
->index
= migratetype
;
206 #ifdef CONFIG_PM_SLEEP
208 * The following functions are used by the suspend/hibernate code to temporarily
209 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
210 * while devices are suspended. To avoid races with the suspend/hibernate code,
211 * they should always be called with system_transition_mutex held
212 * (gfp_allowed_mask also should only be modified with system_transition_mutex
213 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
214 * with that modification).
217 static gfp_t saved_gfp_mask
;
219 void pm_restore_gfp_mask(void)
221 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
222 if (saved_gfp_mask
) {
223 gfp_allowed_mask
= saved_gfp_mask
;
228 void pm_restrict_gfp_mask(void)
230 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
231 WARN_ON(saved_gfp_mask
);
232 saved_gfp_mask
= gfp_allowed_mask
;
233 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
236 bool pm_suspended_storage(void)
238 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
242 #endif /* CONFIG_PM_SLEEP */
244 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
245 unsigned int pageblock_order __read_mostly
;
248 static void __free_pages_ok(struct page
*page
, unsigned int order
);
251 * results with 256, 32 in the lowmem_reserve sysctl:
252 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
253 * 1G machine -> (16M dma, 784M normal, 224M high)
254 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
255 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
256 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
258 * TBD: should special case ZONE_DMA32 machines here - in those we normally
259 * don't need any ZONE_NORMAL reservation
261 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
262 #ifdef CONFIG_ZONE_DMA
265 #ifdef CONFIG_ZONE_DMA32
269 #ifdef CONFIG_HIGHMEM
275 static char * const zone_names
[MAX_NR_ZONES
] = {
276 #ifdef CONFIG_ZONE_DMA
279 #ifdef CONFIG_ZONE_DMA32
283 #ifdef CONFIG_HIGHMEM
287 #ifdef CONFIG_ZONE_DEVICE
292 const char * const migratetype_names
[MIGRATE_TYPES
] = {
300 #ifdef CONFIG_MEMORY_ISOLATION
305 compound_page_dtor
* const compound_page_dtors
[] = {
308 #ifdef CONFIG_HUGETLB_PAGE
311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
316 int min_free_kbytes
= 1024;
317 int user_min_free_kbytes
= -1;
318 #ifdef CONFIG_DISCONTIGMEM
320 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
321 * are not on separate NUMA nodes. Functionally this works but with
322 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
323 * quite small. By default, do not boost watermarks on discontigmem as in
324 * many cases very high-order allocations like THP are likely to be
325 * unsupported and the premature reclaim offsets the advantage of long-term
326 * fragmentation avoidance.
328 int watermark_boost_factor __read_mostly
;
330 int watermark_boost_factor __read_mostly
= 15000;
332 int watermark_scale_factor
= 10;
334 static unsigned long nr_kernel_pages __initdata
;
335 static unsigned long nr_all_pages __initdata
;
336 static unsigned long dma_reserve __initdata
;
338 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
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
);
351 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
354 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
355 unsigned int nr_online_nodes __read_mostly
= 1;
356 EXPORT_SYMBOL(nr_node_ids
);
357 EXPORT_SYMBOL(nr_online_nodes
);
360 int page_group_by_mobility_disabled __read_mostly
;
362 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
368 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
383 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
385 if (!static_branch_unlikely(&deferred_pages
))
386 kasan_free_pages(page
, order
);
389 /* Returns true if the struct page for the pfn is uninitialised */
390 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
392 int nid
= early_pfn_to_nid(pfn
);
394 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
404 static bool __meminit
405 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
407 static unsigned long prev_end_pfn
, nr_initialised
;
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
413 if (prev_end_pfn
!= end_pfn
) {
414 prev_end_pfn
= end_pfn
;
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
427 if ((nr_initialised
> PAGES_PER_SECTION
) &&
428 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
429 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
435 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
437 static inline bool early_page_uninitialised(unsigned long pfn
)
442 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
448 /* Return a pointer to the bitmap storing bits affecting a block of pages */
449 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
452 #ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn
));
455 return page_zone(page
)->pageblock_flags
;
456 #endif /* CONFIG_SPARSEMEM */
459 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
461 #ifdef CONFIG_SPARSEMEM
462 pfn
&= (PAGES_PER_SECTION
-1);
463 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
465 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
466 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
467 #endif /* CONFIG_SPARSEMEM */
471 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
472 * @page: The page within the block of interest
473 * @pfn: The target page frame number
474 * @end_bitidx: The last bit of interest to retrieve
475 * @mask: mask of bits that the caller is interested in
477 * Return: pageblock_bits flags
479 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
481 unsigned long end_bitidx
,
484 unsigned long *bitmap
;
485 unsigned long bitidx
, word_bitidx
;
488 bitmap
= get_pageblock_bitmap(page
, pfn
);
489 bitidx
= pfn_to_bitidx(page
, pfn
);
490 word_bitidx
= bitidx
/ BITS_PER_LONG
;
491 bitidx
&= (BITS_PER_LONG
-1);
493 word
= bitmap
[word_bitidx
];
494 bitidx
+= end_bitidx
;
495 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
498 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
499 unsigned long end_bitidx
,
502 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
505 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
507 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @end_bitidx: The last bit of interest
516 * @mask: mask of bits that the caller is interested in
518 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
520 unsigned long end_bitidx
,
523 unsigned long *bitmap
;
524 unsigned long bitidx
, word_bitidx
;
525 unsigned long old_word
, word
;
527 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
528 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
530 bitmap
= get_pageblock_bitmap(page
, pfn
);
531 bitidx
= pfn_to_bitidx(page
, pfn
);
532 word_bitidx
= bitidx
/ BITS_PER_LONG
;
533 bitidx
&= (BITS_PER_LONG
-1);
535 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
537 bitidx
+= end_bitidx
;
538 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
539 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
541 word
= READ_ONCE(bitmap
[word_bitidx
]);
543 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
544 if (word
== old_word
)
550 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
552 if (unlikely(page_group_by_mobility_disabled
&&
553 migratetype
< MIGRATE_PCPTYPES
))
554 migratetype
= MIGRATE_UNMOVABLE
;
556 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
557 PB_migrate
, PB_migrate_end
);
560 #ifdef CONFIG_DEBUG_VM
561 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
565 unsigned long pfn
= page_to_pfn(page
);
566 unsigned long sp
, start_pfn
;
569 seq
= zone_span_seqbegin(zone
);
570 start_pfn
= zone
->zone_start_pfn
;
571 sp
= zone
->spanned_pages
;
572 if (!zone_spans_pfn(zone
, pfn
))
574 } while (zone_span_seqretry(zone
, seq
));
577 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
578 pfn
, zone_to_nid(zone
), zone
->name
,
579 start_pfn
, start_pfn
+ sp
);
584 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
586 if (!pfn_valid_within(page_to_pfn(page
)))
588 if (zone
!= page_zone(page
))
594 * Temporary debugging check for pages not lying within a given zone.
596 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
598 if (page_outside_zone_boundaries(zone
, page
))
600 if (!page_is_consistent(zone
, page
))
606 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
612 static void bad_page(struct page
*page
, const char *reason
,
613 unsigned long bad_flags
)
615 static unsigned long resume
;
616 static unsigned long nr_shown
;
617 static unsigned long nr_unshown
;
620 * Allow a burst of 60 reports, then keep quiet for that minute;
621 * or allow a steady drip of one report per second.
623 if (nr_shown
== 60) {
624 if (time_before(jiffies
, resume
)) {
630 "BUG: Bad page state: %lu messages suppressed\n",
637 resume
= jiffies
+ 60 * HZ
;
639 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
640 current
->comm
, page_to_pfn(page
));
641 __dump_page(page
, reason
);
642 bad_flags
&= page
->flags
;
644 pr_alert("bad because of flags: %#lx(%pGp)\n",
645 bad_flags
, &bad_flags
);
646 dump_page_owner(page
);
651 /* Leave bad fields for debug, except PageBuddy could make trouble */
652 page_mapcount_reset(page
); /* remove PageBuddy */
653 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
657 * Higher-order pages are called "compound pages". They are structured thusly:
659 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
661 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
662 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
664 * The first tail page's ->compound_dtor holds the offset in array of compound
665 * page destructors. See compound_page_dtors.
667 * The first tail page's ->compound_order holds the order of allocation.
668 * This usage means that zero-order pages may not be compound.
671 void free_compound_page(struct page
*page
)
673 mem_cgroup_uncharge(page
);
674 __free_pages_ok(page
, compound_order(page
));
677 void prep_compound_page(struct page
*page
, unsigned int order
)
680 int nr_pages
= 1 << order
;
682 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
683 set_compound_order(page
, order
);
685 for (i
= 1; i
< nr_pages
; i
++) {
686 struct page
*p
= page
+ i
;
687 set_page_count(p
, 0);
688 p
->mapping
= TAIL_MAPPING
;
689 set_compound_head(p
, page
);
691 atomic_set(compound_mapcount_ptr(page
), -1);
694 #ifdef CONFIG_DEBUG_PAGEALLOC
695 unsigned int _debug_guardpage_minorder
;
697 #ifdef CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
698 DEFINE_STATIC_KEY_TRUE(_debug_pagealloc_enabled
);
700 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
702 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
704 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
706 static int __init
early_debug_pagealloc(char *buf
)
710 if (kstrtobool(buf
, &enable
))
714 static_branch_enable(&_debug_pagealloc_enabled
);
718 early_param("debug_pagealloc", early_debug_pagealloc
);
720 static void init_debug_guardpage(void)
722 if (!debug_pagealloc_enabled())
725 if (!debug_guardpage_minorder())
728 static_branch_enable(&_debug_guardpage_enabled
);
731 static int __init
debug_guardpage_minorder_setup(char *buf
)
735 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
736 pr_err("Bad debug_guardpage_minorder value\n");
739 _debug_guardpage_minorder
= res
;
740 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
743 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
745 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
746 unsigned int order
, int migratetype
)
748 if (!debug_guardpage_enabled())
751 if (order
>= debug_guardpage_minorder())
754 __SetPageGuard(page
);
755 INIT_LIST_HEAD(&page
->lru
);
756 set_page_private(page
, order
);
757 /* Guard pages are not available for any usage */
758 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
763 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
764 unsigned int order
, int migratetype
)
766 if (!debug_guardpage_enabled())
769 __ClearPageGuard(page
);
771 set_page_private(page
, 0);
772 if (!is_migrate_isolate(migratetype
))
773 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
776 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
777 unsigned int order
, int migratetype
) { return false; }
778 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
779 unsigned int order
, int migratetype
) {}
782 static inline void set_page_order(struct page
*page
, unsigned int order
)
784 set_page_private(page
, order
);
785 __SetPageBuddy(page
);
789 * This function checks whether a page is free && is the buddy
790 * we can coalesce a page and its buddy if
791 * (a) the buddy is not in a hole (check before calling!) &&
792 * (b) the buddy is in the buddy system &&
793 * (c) a page and its buddy have the same order &&
794 * (d) a page and its buddy are in the same zone.
796 * For recording whether a page is in the buddy system, we set PageBuddy.
797 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
799 * For recording page's order, we use page_private(page).
801 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
804 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
805 if (page_zone_id(page
) != page_zone_id(buddy
))
808 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
813 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
815 * zone check is done late to avoid uselessly
816 * calculating zone/node ids for pages that could
819 if (page_zone_id(page
) != page_zone_id(buddy
))
822 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
829 #ifdef CONFIG_COMPACTION
830 static inline struct capture_control
*task_capc(struct zone
*zone
)
832 struct capture_control
*capc
= current
->capture_control
;
835 !(current
->flags
& PF_KTHREAD
) &&
837 capc
->cc
->zone
== zone
&&
838 capc
->cc
->direct_compaction
? capc
: NULL
;
842 compaction_capture(struct capture_control
*capc
, struct page
*page
,
843 int order
, int migratetype
)
845 if (!capc
|| order
!= capc
->cc
->order
)
848 /* Do not accidentally pollute CMA or isolated regions*/
849 if (is_migrate_cma(migratetype
) ||
850 is_migrate_isolate(migratetype
))
854 * Do not let lower order allocations polluate a movable pageblock.
855 * This might let an unmovable request use a reclaimable pageblock
856 * and vice-versa but no more than normal fallback logic which can
857 * have trouble finding a high-order free page.
859 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
867 static inline struct capture_control
*task_capc(struct zone
*zone
)
873 compaction_capture(struct capture_control
*capc
, struct page
*page
,
874 int order
, int migratetype
)
878 #endif /* CONFIG_COMPACTION */
881 * Freeing function for a buddy system allocator.
883 * The concept of a buddy system is to maintain direct-mapped table
884 * (containing bit values) for memory blocks of various "orders".
885 * The bottom level table contains the map for the smallest allocatable
886 * units of memory (here, pages), and each level above it describes
887 * pairs of units from the levels below, hence, "buddies".
888 * At a high level, all that happens here is marking the table entry
889 * at the bottom level available, and propagating the changes upward
890 * as necessary, plus some accounting needed to play nicely with other
891 * parts of the VM system.
892 * At each level, we keep a list of pages, which are heads of continuous
893 * free pages of length of (1 << order) and marked with PageBuddy.
894 * Page's order is recorded in page_private(page) field.
895 * So when we are allocating or freeing one, we can derive the state of the
896 * other. That is, if we allocate a small block, and both were
897 * free, the remainder of the region must be split into blocks.
898 * If a block is freed, and its buddy is also free, then this
899 * triggers coalescing into a block of larger size.
904 static inline void __free_one_page(struct page
*page
,
906 struct zone
*zone
, unsigned int order
,
909 unsigned long combined_pfn
;
910 unsigned long uninitialized_var(buddy_pfn
);
912 unsigned int max_order
;
913 struct capture_control
*capc
= task_capc(zone
);
915 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
917 VM_BUG_ON(!zone_is_initialized(zone
));
918 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
920 VM_BUG_ON(migratetype
== -1);
921 if (likely(!is_migrate_isolate(migratetype
)))
922 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
924 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
925 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
928 while (order
< max_order
- 1) {
929 if (compaction_capture(capc
, page
, order
, migratetype
)) {
930 __mod_zone_freepage_state(zone
, -(1 << order
),
934 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
935 buddy
= page
+ (buddy_pfn
- pfn
);
937 if (!pfn_valid_within(buddy_pfn
))
939 if (!page_is_buddy(page
, buddy
, order
))
942 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
943 * merge with it and move up one order.
945 if (page_is_guard(buddy
))
946 clear_page_guard(zone
, buddy
, order
, migratetype
);
948 del_page_from_free_area(buddy
, &zone
->free_area
[order
]);
949 combined_pfn
= buddy_pfn
& pfn
;
950 page
= page
+ (combined_pfn
- pfn
);
954 if (max_order
< MAX_ORDER
) {
955 /* If we are here, it means order is >= pageblock_order.
956 * We want to prevent merge between freepages on isolate
957 * pageblock and normal pageblock. Without this, pageblock
958 * isolation could cause incorrect freepage or CMA accounting.
960 * We don't want to hit this code for the more frequent
963 if (unlikely(has_isolate_pageblock(zone
))) {
966 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
967 buddy
= page
+ (buddy_pfn
- pfn
);
968 buddy_mt
= get_pageblock_migratetype(buddy
);
970 if (migratetype
!= buddy_mt
971 && (is_migrate_isolate(migratetype
) ||
972 is_migrate_isolate(buddy_mt
)))
976 goto continue_merging
;
980 set_page_order(page
, order
);
983 * If this is not the largest possible page, check if the buddy
984 * of the next-highest order is free. If it is, it's possible
985 * that pages are being freed that will coalesce soon. In case,
986 * that is happening, add the free page to the tail of the list
987 * so it's less likely to be used soon and more likely to be merged
988 * as a higher order page
990 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)
991 && !is_shuffle_order(order
)) {
992 struct page
*higher_page
, *higher_buddy
;
993 combined_pfn
= buddy_pfn
& pfn
;
994 higher_page
= page
+ (combined_pfn
- pfn
);
995 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
996 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
997 if (pfn_valid_within(buddy_pfn
) &&
998 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
999 add_to_free_area_tail(page
, &zone
->free_area
[order
],
1005 if (is_shuffle_order(order
))
1006 add_to_free_area_random(page
, &zone
->free_area
[order
],
1009 add_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
1014 * A bad page could be due to a number of fields. Instead of multiple branches,
1015 * try and check multiple fields with one check. The caller must do a detailed
1016 * check if necessary.
1018 static inline bool page_expected_state(struct page
*page
,
1019 unsigned long check_flags
)
1021 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1024 if (unlikely((unsigned long)page
->mapping
|
1025 page_ref_count(page
) |
1027 (unsigned long)page
->mem_cgroup
|
1029 (page
->flags
& check_flags
)))
1035 static void free_pages_check_bad(struct page
*page
)
1037 const char *bad_reason
;
1038 unsigned long bad_flags
;
1043 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1044 bad_reason
= "nonzero mapcount";
1045 if (unlikely(page
->mapping
!= NULL
))
1046 bad_reason
= "non-NULL mapping";
1047 if (unlikely(page_ref_count(page
) != 0))
1048 bad_reason
= "nonzero _refcount";
1049 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
1050 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1051 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
1054 if (unlikely(page
->mem_cgroup
))
1055 bad_reason
= "page still charged to cgroup";
1057 bad_page(page
, bad_reason
, bad_flags
);
1060 static inline int free_pages_check(struct page
*page
)
1062 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1065 /* Something has gone sideways, find it */
1066 free_pages_check_bad(page
);
1070 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1075 * We rely page->lru.next never has bit 0 set, unless the page
1076 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1078 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1080 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1084 switch (page
- head_page
) {
1086 /* the first tail page: ->mapping may be compound_mapcount() */
1087 if (unlikely(compound_mapcount(page
))) {
1088 bad_page(page
, "nonzero compound_mapcount", 0);
1094 * the second tail page: ->mapping is
1095 * deferred_list.next -- ignore value.
1099 if (page
->mapping
!= TAIL_MAPPING
) {
1100 bad_page(page
, "corrupted mapping in tail page", 0);
1105 if (unlikely(!PageTail(page
))) {
1106 bad_page(page
, "PageTail not set", 0);
1109 if (unlikely(compound_head(page
) != head_page
)) {
1110 bad_page(page
, "compound_head not consistent", 0);
1115 page
->mapping
= NULL
;
1116 clear_compound_head(page
);
1120 static void kernel_init_free_pages(struct page
*page
, int numpages
)
1124 for (i
= 0; i
< numpages
; i
++)
1125 clear_highpage(page
+ i
);
1128 static __always_inline
bool free_pages_prepare(struct page
*page
,
1129 unsigned int order
, bool check_free
)
1133 VM_BUG_ON_PAGE(PageTail(page
), page
);
1135 trace_mm_page_free(page
, order
);
1138 * Check tail pages before head page information is cleared to
1139 * avoid checking PageCompound for order-0 pages.
1141 if (unlikely(order
)) {
1142 bool compound
= PageCompound(page
);
1145 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1148 ClearPageDoubleMap(page
);
1149 for (i
= 1; i
< (1 << order
); i
++) {
1151 bad
+= free_tail_pages_check(page
, page
+ i
);
1152 if (unlikely(free_pages_check(page
+ i
))) {
1156 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1159 if (PageMappingFlags(page
))
1160 page
->mapping
= NULL
;
1161 if (memcg_kmem_enabled() && PageKmemcg(page
))
1162 __memcg_kmem_uncharge(page
, order
);
1164 bad
+= free_pages_check(page
);
1168 page_cpupid_reset_last(page
);
1169 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1170 reset_page_owner(page
, order
);
1172 if (!PageHighMem(page
)) {
1173 debug_check_no_locks_freed(page_address(page
),
1174 PAGE_SIZE
<< order
);
1175 debug_check_no_obj_freed(page_address(page
),
1176 PAGE_SIZE
<< order
);
1178 if (want_init_on_free())
1179 kernel_init_free_pages(page
, 1 << order
);
1181 kernel_poison_pages(page
, 1 << order
, 0);
1183 * arch_free_page() can make the page's contents inaccessible. s390
1184 * does this. So nothing which can access the page's contents should
1185 * happen after this.
1187 arch_free_page(page
, order
);
1189 if (debug_pagealloc_enabled())
1190 kernel_map_pages(page
, 1 << order
, 0);
1192 kasan_free_nondeferred_pages(page
, order
);
1197 #ifdef CONFIG_DEBUG_VM
1199 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1200 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1201 * moved from pcp lists to free lists.
1203 static bool free_pcp_prepare(struct page
*page
)
1205 return free_pages_prepare(page
, 0, true);
1208 static bool bulkfree_pcp_prepare(struct page
*page
)
1210 if (debug_pagealloc_enabled())
1211 return free_pages_check(page
);
1217 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1218 * moving from pcp lists to free list in order to reduce overhead. With
1219 * debug_pagealloc enabled, they are checked also immediately when being freed
1222 static bool free_pcp_prepare(struct page
*page
)
1224 if (debug_pagealloc_enabled())
1225 return free_pages_prepare(page
, 0, true);
1227 return free_pages_prepare(page
, 0, false);
1230 static bool bulkfree_pcp_prepare(struct page
*page
)
1232 return free_pages_check(page
);
1234 #endif /* CONFIG_DEBUG_VM */
1236 static inline void prefetch_buddy(struct page
*page
)
1238 unsigned long pfn
= page_to_pfn(page
);
1239 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1240 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1246 * Frees a number of pages from the PCP lists
1247 * Assumes all pages on list are in same zone, and of same order.
1248 * count is the number of pages to free.
1250 * If the zone was previously in an "all pages pinned" state then look to
1251 * see if this freeing clears that state.
1253 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1254 * pinned" detection logic.
1256 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1257 struct per_cpu_pages
*pcp
)
1259 int migratetype
= 0;
1261 int prefetch_nr
= 0;
1262 bool isolated_pageblocks
;
1263 struct page
*page
, *tmp
;
1267 struct list_head
*list
;
1270 * Remove pages from lists in a round-robin fashion. A
1271 * batch_free count is maintained that is incremented when an
1272 * empty list is encountered. This is so more pages are freed
1273 * off fuller lists instead of spinning excessively around empty
1278 if (++migratetype
== MIGRATE_PCPTYPES
)
1280 list
= &pcp
->lists
[migratetype
];
1281 } while (list_empty(list
));
1283 /* This is the only non-empty list. Free them all. */
1284 if (batch_free
== MIGRATE_PCPTYPES
)
1288 page
= list_last_entry(list
, struct page
, lru
);
1289 /* must delete to avoid corrupting pcp list */
1290 list_del(&page
->lru
);
1293 if (bulkfree_pcp_prepare(page
))
1296 list_add_tail(&page
->lru
, &head
);
1299 * We are going to put the page back to the global
1300 * pool, prefetch its buddy to speed up later access
1301 * under zone->lock. It is believed the overhead of
1302 * an additional test and calculating buddy_pfn here
1303 * can be offset by reduced memory latency later. To
1304 * avoid excessive prefetching due to large count, only
1305 * prefetch buddy for the first pcp->batch nr of pages.
1307 if (prefetch_nr
++ < pcp
->batch
)
1308 prefetch_buddy(page
);
1309 } while (--count
&& --batch_free
&& !list_empty(list
));
1312 spin_lock(&zone
->lock
);
1313 isolated_pageblocks
= has_isolate_pageblock(zone
);
1316 * Use safe version since after __free_one_page(),
1317 * page->lru.next will not point to original list.
1319 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1320 int mt
= get_pcppage_migratetype(page
);
1321 /* MIGRATE_ISOLATE page should not go to pcplists */
1322 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1323 /* Pageblock could have been isolated meanwhile */
1324 if (unlikely(isolated_pageblocks
))
1325 mt
= get_pageblock_migratetype(page
);
1327 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1328 trace_mm_page_pcpu_drain(page
, 0, mt
);
1330 spin_unlock(&zone
->lock
);
1333 static void free_one_page(struct zone
*zone
,
1334 struct page
*page
, unsigned long pfn
,
1338 spin_lock(&zone
->lock
);
1339 if (unlikely(has_isolate_pageblock(zone
) ||
1340 is_migrate_isolate(migratetype
))) {
1341 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1343 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1344 spin_unlock(&zone
->lock
);
1347 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1348 unsigned long zone
, int nid
)
1350 mm_zero_struct_page(page
);
1351 set_page_links(page
, zone
, nid
, pfn
);
1352 init_page_count(page
);
1353 page_mapcount_reset(page
);
1354 page_cpupid_reset_last(page
);
1355 page_kasan_tag_reset(page
);
1357 INIT_LIST_HEAD(&page
->lru
);
1358 #ifdef WANT_PAGE_VIRTUAL
1359 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1360 if (!is_highmem_idx(zone
))
1361 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1365 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1366 static void __meminit
init_reserved_page(unsigned long pfn
)
1371 if (!early_page_uninitialised(pfn
))
1374 nid
= early_pfn_to_nid(pfn
);
1375 pgdat
= NODE_DATA(nid
);
1377 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1378 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1380 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1383 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1386 static inline void init_reserved_page(unsigned long pfn
)
1389 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1392 * Initialised pages do not have PageReserved set. This function is
1393 * called for each range allocated by the bootmem allocator and
1394 * marks the pages PageReserved. The remaining valid pages are later
1395 * sent to the buddy page allocator.
1397 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1399 unsigned long start_pfn
= PFN_DOWN(start
);
1400 unsigned long end_pfn
= PFN_UP(end
);
1402 for (; start_pfn
< end_pfn
; start_pfn
++) {
1403 if (pfn_valid(start_pfn
)) {
1404 struct page
*page
= pfn_to_page(start_pfn
);
1406 init_reserved_page(start_pfn
);
1408 /* Avoid false-positive PageTail() */
1409 INIT_LIST_HEAD(&page
->lru
);
1412 * no need for atomic set_bit because the struct
1413 * page is not visible yet so nobody should
1416 __SetPageReserved(page
);
1421 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1423 unsigned long flags
;
1425 unsigned long pfn
= page_to_pfn(page
);
1427 if (!free_pages_prepare(page
, order
, true))
1430 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1431 local_irq_save(flags
);
1432 __count_vm_events(PGFREE
, 1 << order
);
1433 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1434 local_irq_restore(flags
);
1437 void __free_pages_core(struct page
*page
, unsigned int order
)
1439 unsigned int nr_pages
= 1 << order
;
1440 struct page
*p
= page
;
1444 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1446 __ClearPageReserved(p
);
1447 set_page_count(p
, 0);
1449 __ClearPageReserved(p
);
1450 set_page_count(p
, 0);
1452 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1453 set_page_refcounted(page
);
1454 __free_pages(page
, order
);
1457 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1458 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1460 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1462 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1464 static DEFINE_SPINLOCK(early_pfn_lock
);
1467 spin_lock(&early_pfn_lock
);
1468 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1470 nid
= first_online_node
;
1471 spin_unlock(&early_pfn_lock
);
1477 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1478 /* Only safe to use early in boot when initialisation is single-threaded */
1479 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1483 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1484 if (nid
>= 0 && nid
!= node
)
1490 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1497 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1500 if (early_page_uninitialised(pfn
))
1502 __free_pages_core(page
, order
);
1506 * Check that the whole (or subset of) a pageblock given by the interval of
1507 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1508 * with the migration of free compaction scanner. The scanners then need to
1509 * use only pfn_valid_within() check for arches that allow holes within
1512 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1514 * It's possible on some configurations to have a setup like node0 node1 node0
1515 * i.e. it's possible that all pages within a zones range of pages do not
1516 * belong to a single zone. We assume that a border between node0 and node1
1517 * can occur within a single pageblock, but not a node0 node1 node0
1518 * interleaving within a single pageblock. It is therefore sufficient to check
1519 * the first and last page of a pageblock and avoid checking each individual
1520 * page in a pageblock.
1522 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1523 unsigned long end_pfn
, struct zone
*zone
)
1525 struct page
*start_page
;
1526 struct page
*end_page
;
1528 /* end_pfn is one past the range we are checking */
1531 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1534 start_page
= pfn_to_online_page(start_pfn
);
1538 if (page_zone(start_page
) != zone
)
1541 end_page
= pfn_to_page(end_pfn
);
1543 /* This gives a shorter code than deriving page_zone(end_page) */
1544 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1550 void set_zone_contiguous(struct zone
*zone
)
1552 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1553 unsigned long block_end_pfn
;
1555 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1556 for (; block_start_pfn
< zone_end_pfn(zone
);
1557 block_start_pfn
= block_end_pfn
,
1558 block_end_pfn
+= pageblock_nr_pages
) {
1560 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1562 if (!__pageblock_pfn_to_page(block_start_pfn
,
1563 block_end_pfn
, zone
))
1567 /* We confirm that there is no hole */
1568 zone
->contiguous
= true;
1571 void clear_zone_contiguous(struct zone
*zone
)
1573 zone
->contiguous
= false;
1576 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1577 static void __init
deferred_free_range(unsigned long pfn
,
1578 unsigned long nr_pages
)
1586 page
= pfn_to_page(pfn
);
1588 /* Free a large naturally-aligned chunk if possible */
1589 if (nr_pages
== pageblock_nr_pages
&&
1590 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1591 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1592 __free_pages_core(page
, pageblock_order
);
1596 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1597 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1598 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1599 __free_pages_core(page
, 0);
1603 /* Completion tracking for deferred_init_memmap() threads */
1604 static atomic_t pgdat_init_n_undone __initdata
;
1605 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1607 static inline void __init
pgdat_init_report_one_done(void)
1609 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1610 complete(&pgdat_init_all_done_comp
);
1614 * Returns true if page needs to be initialized or freed to buddy allocator.
1616 * First we check if pfn is valid on architectures where it is possible to have
1617 * holes within pageblock_nr_pages. On systems where it is not possible, this
1618 * function is optimized out.
1620 * Then, we check if a current large page is valid by only checking the validity
1623 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1625 if (!pfn_valid_within(pfn
))
1627 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1633 * Free pages to buddy allocator. Try to free aligned pages in
1634 * pageblock_nr_pages sizes.
1636 static void __init
deferred_free_pages(unsigned long pfn
,
1637 unsigned long end_pfn
)
1639 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1640 unsigned long nr_free
= 0;
1642 for (; pfn
< end_pfn
; pfn
++) {
1643 if (!deferred_pfn_valid(pfn
)) {
1644 deferred_free_range(pfn
- nr_free
, nr_free
);
1646 } else if (!(pfn
& nr_pgmask
)) {
1647 deferred_free_range(pfn
- nr_free
, nr_free
);
1649 touch_nmi_watchdog();
1654 /* Free the last block of pages to allocator */
1655 deferred_free_range(pfn
- nr_free
, nr_free
);
1659 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1660 * by performing it only once every pageblock_nr_pages.
1661 * Return number of pages initialized.
1663 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1665 unsigned long end_pfn
)
1667 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1668 int nid
= zone_to_nid(zone
);
1669 unsigned long nr_pages
= 0;
1670 int zid
= zone_idx(zone
);
1671 struct page
*page
= NULL
;
1673 for (; pfn
< end_pfn
; pfn
++) {
1674 if (!deferred_pfn_valid(pfn
)) {
1677 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1678 page
= pfn_to_page(pfn
);
1679 touch_nmi_watchdog();
1683 __init_single_page(page
, pfn
, zid
, nid
);
1690 * This function is meant to pre-load the iterator for the zone init.
1691 * Specifically it walks through the ranges until we are caught up to the
1692 * first_init_pfn value and exits there. If we never encounter the value we
1693 * return false indicating there are no valid ranges left.
1696 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1697 unsigned long *spfn
, unsigned long *epfn
,
1698 unsigned long first_init_pfn
)
1703 * Start out by walking through the ranges in this zone that have
1704 * already been initialized. We don't need to do anything with them
1705 * so we just need to flush them out of the system.
1707 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1708 if (*epfn
<= first_init_pfn
)
1710 if (*spfn
< first_init_pfn
)
1711 *spfn
= first_init_pfn
;
1720 * Initialize and free pages. We do it in two loops: first we initialize
1721 * struct page, then free to buddy allocator, because while we are
1722 * freeing pages we can access pages that are ahead (computing buddy
1723 * page in __free_one_page()).
1725 * In order to try and keep some memory in the cache we have the loop
1726 * broken along max page order boundaries. This way we will not cause
1727 * any issues with the buddy page computation.
1729 static unsigned long __init
1730 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1731 unsigned long *end_pfn
)
1733 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1734 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1735 unsigned long nr_pages
= 0;
1738 /* First we loop through and initialize the page values */
1739 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1742 if (mo_pfn
<= *start_pfn
)
1745 t
= min(mo_pfn
, *end_pfn
);
1746 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1748 if (mo_pfn
< *end_pfn
) {
1749 *start_pfn
= mo_pfn
;
1754 /* Reset values and now loop through freeing pages as needed */
1757 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1763 t
= min(mo_pfn
, epfn
);
1764 deferred_free_pages(spfn
, t
);
1773 /* Initialise remaining memory on a node */
1774 static int __init
deferred_init_memmap(void *data
)
1776 pg_data_t
*pgdat
= data
;
1777 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1778 unsigned long spfn
= 0, epfn
= 0, nr_pages
= 0;
1779 unsigned long first_init_pfn
, flags
;
1780 unsigned long start
= jiffies
;
1785 /* Bind memory initialisation thread to a local node if possible */
1786 if (!cpumask_empty(cpumask
))
1787 set_cpus_allowed_ptr(current
, cpumask
);
1789 pgdat_resize_lock(pgdat
, &flags
);
1790 first_init_pfn
= pgdat
->first_deferred_pfn
;
1791 if (first_init_pfn
== ULONG_MAX
) {
1792 pgdat_resize_unlock(pgdat
, &flags
);
1793 pgdat_init_report_one_done();
1797 /* Sanity check boundaries */
1798 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1799 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1800 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1802 /* Only the highest zone is deferred so find it */
1803 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1804 zone
= pgdat
->node_zones
+ zid
;
1805 if (first_init_pfn
< zone_end_pfn(zone
))
1809 /* If the zone is empty somebody else may have cleared out the zone */
1810 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1815 * Initialize and free pages in MAX_ORDER sized increments so
1816 * that we can avoid introducing any issues with the buddy
1820 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1822 pgdat_resize_unlock(pgdat
, &flags
);
1824 /* Sanity check that the next zone really is unpopulated */
1825 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1827 pr_info("node %d initialised, %lu pages in %ums\n",
1828 pgdat
->node_id
, nr_pages
, jiffies_to_msecs(jiffies
- start
));
1830 pgdat_init_report_one_done();
1835 * If this zone has deferred pages, try to grow it by initializing enough
1836 * deferred pages to satisfy the allocation specified by order, rounded up to
1837 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1838 * of SECTION_SIZE bytes by initializing struct pages in increments of
1839 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1841 * Return true when zone was grown, otherwise return false. We return true even
1842 * when we grow less than requested, to let the caller decide if there are
1843 * enough pages to satisfy the allocation.
1845 * Note: We use noinline because this function is needed only during boot, and
1846 * it is called from a __ref function _deferred_grow_zone. This way we are
1847 * making sure that it is not inlined into permanent text section.
1849 static noinline
bool __init
1850 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1852 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1853 pg_data_t
*pgdat
= zone
->zone_pgdat
;
1854 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1855 unsigned long spfn
, epfn
, flags
;
1856 unsigned long nr_pages
= 0;
1859 /* Only the last zone may have deferred pages */
1860 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1863 pgdat_resize_lock(pgdat
, &flags
);
1866 * If deferred pages have been initialized while we were waiting for
1867 * the lock, return true, as the zone was grown. The caller will retry
1868 * this zone. We won't return to this function since the caller also
1869 * has this static branch.
1871 if (!static_branch_unlikely(&deferred_pages
)) {
1872 pgdat_resize_unlock(pgdat
, &flags
);
1877 * If someone grew this zone while we were waiting for spinlock, return
1878 * true, as there might be enough pages already.
1880 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1881 pgdat_resize_unlock(pgdat
, &flags
);
1885 /* If the zone is empty somebody else may have cleared out the zone */
1886 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1887 first_deferred_pfn
)) {
1888 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1889 pgdat_resize_unlock(pgdat
, &flags
);
1890 /* Retry only once. */
1891 return first_deferred_pfn
!= ULONG_MAX
;
1895 * Initialize and free pages in MAX_ORDER sized increments so
1896 * that we can avoid introducing any issues with the buddy
1899 while (spfn
< epfn
) {
1900 /* update our first deferred PFN for this section */
1901 first_deferred_pfn
= spfn
;
1903 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1905 /* We should only stop along section boundaries */
1906 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
1909 /* If our quota has been met we can stop here */
1910 if (nr_pages
>= nr_pages_needed
)
1914 pgdat
->first_deferred_pfn
= spfn
;
1915 pgdat_resize_unlock(pgdat
, &flags
);
1917 return nr_pages
> 0;
1921 * deferred_grow_zone() is __init, but it is called from
1922 * get_page_from_freelist() during early boot until deferred_pages permanently
1923 * disables this call. This is why we have refdata wrapper to avoid warning,
1924 * and to ensure that the function body gets unloaded.
1927 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1929 return deferred_grow_zone(zone
, order
);
1932 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1934 void __init
page_alloc_init_late(void)
1939 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1941 /* There will be num_node_state(N_MEMORY) threads */
1942 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1943 for_each_node_state(nid
, N_MEMORY
) {
1944 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1947 /* Block until all are initialised */
1948 wait_for_completion(&pgdat_init_all_done_comp
);
1951 * We initialized the rest of the deferred pages. Permanently disable
1952 * on-demand struct page initialization.
1954 static_branch_disable(&deferred_pages
);
1956 /* Reinit limits that are based on free pages after the kernel is up */
1957 files_maxfiles_init();
1960 /* Discard memblock private memory */
1963 for_each_node_state(nid
, N_MEMORY
)
1964 shuffle_free_memory(NODE_DATA(nid
));
1966 for_each_populated_zone(zone
)
1967 set_zone_contiguous(zone
);
1969 #ifdef CONFIG_DEBUG_PAGEALLOC
1970 init_debug_guardpage();
1975 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1976 void __init
init_cma_reserved_pageblock(struct page
*page
)
1978 unsigned i
= pageblock_nr_pages
;
1979 struct page
*p
= page
;
1982 __ClearPageReserved(p
);
1983 set_page_count(p
, 0);
1986 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1988 if (pageblock_order
>= MAX_ORDER
) {
1989 i
= pageblock_nr_pages
;
1992 set_page_refcounted(p
);
1993 __free_pages(p
, MAX_ORDER
- 1);
1994 p
+= MAX_ORDER_NR_PAGES
;
1995 } while (i
-= MAX_ORDER_NR_PAGES
);
1997 set_page_refcounted(page
);
1998 __free_pages(page
, pageblock_order
);
2001 adjust_managed_page_count(page
, pageblock_nr_pages
);
2006 * The order of subdivision here is critical for the IO subsystem.
2007 * Please do not alter this order without good reasons and regression
2008 * testing. Specifically, as large blocks of memory are subdivided,
2009 * the order in which smaller blocks are delivered depends on the order
2010 * they're subdivided in this function. This is the primary factor
2011 * influencing the order in which pages are delivered to the IO
2012 * subsystem according to empirical testing, and this is also justified
2013 * by considering the behavior of a buddy system containing a single
2014 * large block of memory acted on by a series of small allocations.
2015 * This behavior is a critical factor in sglist merging's success.
2019 static inline void expand(struct zone
*zone
, struct page
*page
,
2020 int low
, int high
, struct free_area
*area
,
2023 unsigned long size
= 1 << high
;
2025 while (high
> low
) {
2029 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2032 * Mark as guard pages (or page), that will allow to
2033 * merge back to allocator when buddy will be freed.
2034 * Corresponding page table entries will not be touched,
2035 * pages will stay not present in virtual address space
2037 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2040 add_to_free_area(&page
[size
], area
, migratetype
);
2041 set_page_order(&page
[size
], high
);
2045 static void check_new_page_bad(struct page
*page
)
2047 const char *bad_reason
= NULL
;
2048 unsigned long bad_flags
= 0;
2050 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
2051 bad_reason
= "nonzero mapcount";
2052 if (unlikely(page
->mapping
!= NULL
))
2053 bad_reason
= "non-NULL mapping";
2054 if (unlikely(page_ref_count(page
) != 0))
2055 bad_reason
= "nonzero _refcount";
2056 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2057 bad_reason
= "HWPoisoned (hardware-corrupted)";
2058 bad_flags
= __PG_HWPOISON
;
2059 /* Don't complain about hwpoisoned pages */
2060 page_mapcount_reset(page
); /* remove PageBuddy */
2063 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
2064 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
2065 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
2068 if (unlikely(page
->mem_cgroup
))
2069 bad_reason
= "page still charged to cgroup";
2071 bad_page(page
, bad_reason
, bad_flags
);
2075 * This page is about to be returned from the page allocator
2077 static inline int check_new_page(struct page
*page
)
2079 if (likely(page_expected_state(page
,
2080 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2083 check_new_page_bad(page
);
2087 static inline bool free_pages_prezeroed(void)
2089 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
2090 page_poisoning_enabled()) || want_init_on_free();
2093 #ifdef CONFIG_DEBUG_VM
2095 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2096 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2097 * also checked when pcp lists are refilled from the free lists.
2099 static inline bool check_pcp_refill(struct page
*page
)
2101 if (debug_pagealloc_enabled())
2102 return check_new_page(page
);
2107 static inline bool check_new_pcp(struct page
*page
)
2109 return check_new_page(page
);
2113 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2114 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2115 * enabled, they are also checked when being allocated from the pcp lists.
2117 static inline bool check_pcp_refill(struct page
*page
)
2119 return check_new_page(page
);
2121 static inline bool check_new_pcp(struct page
*page
)
2123 if (debug_pagealloc_enabled())
2124 return check_new_page(page
);
2128 #endif /* CONFIG_DEBUG_VM */
2130 static bool check_new_pages(struct page
*page
, unsigned int order
)
2133 for (i
= 0; i
< (1 << order
); i
++) {
2134 struct page
*p
= page
+ i
;
2136 if (unlikely(check_new_page(p
)))
2143 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2146 set_page_private(page
, 0);
2147 set_page_refcounted(page
);
2149 arch_alloc_page(page
, order
);
2150 if (debug_pagealloc_enabled())
2151 kernel_map_pages(page
, 1 << order
, 1);
2152 kasan_alloc_pages(page
, order
);
2153 kernel_poison_pages(page
, 1 << order
, 1);
2154 set_page_owner(page
, order
, gfp_flags
);
2157 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2158 unsigned int alloc_flags
)
2160 post_alloc_hook(page
, order
, gfp_flags
);
2162 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags
))
2163 kernel_init_free_pages(page
, 1 << order
);
2165 if (order
&& (gfp_flags
& __GFP_COMP
))
2166 prep_compound_page(page
, order
);
2169 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2170 * allocate the page. The expectation is that the caller is taking
2171 * steps that will free more memory. The caller should avoid the page
2172 * being used for !PFMEMALLOC purposes.
2174 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2175 set_page_pfmemalloc(page
);
2177 clear_page_pfmemalloc(page
);
2181 * Go through the free lists for the given migratetype and remove
2182 * the smallest available page from the freelists
2184 static __always_inline
2185 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2188 unsigned int current_order
;
2189 struct free_area
*area
;
2192 /* Find a page of the appropriate size in the preferred list */
2193 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2194 area
= &(zone
->free_area
[current_order
]);
2195 page
= get_page_from_free_area(area
, migratetype
);
2198 del_page_from_free_area(page
, area
);
2199 expand(zone
, page
, order
, current_order
, area
, migratetype
);
2200 set_pcppage_migratetype(page
, migratetype
);
2209 * This array describes the order lists are fallen back to when
2210 * the free lists for the desirable migrate type are depleted
2212 static int fallbacks
[MIGRATE_TYPES
][4] = {
2213 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2214 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2215 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2217 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2219 #ifdef CONFIG_MEMORY_ISOLATION
2220 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2225 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2228 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2231 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2232 unsigned int order
) { return NULL
; }
2236 * Move the free pages in a range to the free lists of the requested type.
2237 * Note that start_page and end_pages are not aligned on a pageblock
2238 * boundary. If alignment is required, use move_freepages_block()
2240 static int move_freepages(struct zone
*zone
,
2241 struct page
*start_page
, struct page
*end_page
,
2242 int migratetype
, int *num_movable
)
2246 int pages_moved
= 0;
2248 for (page
= start_page
; page
<= end_page
;) {
2249 if (!pfn_valid_within(page_to_pfn(page
))) {
2254 if (!PageBuddy(page
)) {
2256 * We assume that pages that could be isolated for
2257 * migration are movable. But we don't actually try
2258 * isolating, as that would be expensive.
2261 (PageLRU(page
) || __PageMovable(page
)))
2268 /* Make sure we are not inadvertently changing nodes */
2269 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2270 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2272 order
= page_order(page
);
2273 move_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
2275 pages_moved
+= 1 << order
;
2281 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2282 int migratetype
, int *num_movable
)
2284 unsigned long start_pfn
, end_pfn
;
2285 struct page
*start_page
, *end_page
;
2290 start_pfn
= page_to_pfn(page
);
2291 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2292 start_page
= pfn_to_page(start_pfn
);
2293 end_page
= start_page
+ pageblock_nr_pages
- 1;
2294 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2296 /* Do not cross zone boundaries */
2297 if (!zone_spans_pfn(zone
, start_pfn
))
2299 if (!zone_spans_pfn(zone
, end_pfn
))
2302 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2306 static void change_pageblock_range(struct page
*pageblock_page
,
2307 int start_order
, int migratetype
)
2309 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2311 while (nr_pageblocks
--) {
2312 set_pageblock_migratetype(pageblock_page
, migratetype
);
2313 pageblock_page
+= pageblock_nr_pages
;
2318 * When we are falling back to another migratetype during allocation, try to
2319 * steal extra free pages from the same pageblocks to satisfy further
2320 * allocations, instead of polluting multiple pageblocks.
2322 * If we are stealing a relatively large buddy page, it is likely there will
2323 * be more free pages in the pageblock, so try to steal them all. For
2324 * reclaimable and unmovable allocations, we steal regardless of page size,
2325 * as fragmentation caused by those allocations polluting movable pageblocks
2326 * is worse than movable allocations stealing from unmovable and reclaimable
2329 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2332 * Leaving this order check is intended, although there is
2333 * relaxed order check in next check. The reason is that
2334 * we can actually steal whole pageblock if this condition met,
2335 * but, below check doesn't guarantee it and that is just heuristic
2336 * so could be changed anytime.
2338 if (order
>= pageblock_order
)
2341 if (order
>= pageblock_order
/ 2 ||
2342 start_mt
== MIGRATE_RECLAIMABLE
||
2343 start_mt
== MIGRATE_UNMOVABLE
||
2344 page_group_by_mobility_disabled
)
2350 static inline void boost_watermark(struct zone
*zone
)
2352 unsigned long max_boost
;
2354 if (!watermark_boost_factor
)
2357 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2358 watermark_boost_factor
, 10000);
2361 * high watermark may be uninitialised if fragmentation occurs
2362 * very early in boot so do not boost. We do not fall
2363 * through and boost by pageblock_nr_pages as failing
2364 * allocations that early means that reclaim is not going
2365 * to help and it may even be impossible to reclaim the
2366 * boosted watermark resulting in a hang.
2371 max_boost
= max(pageblock_nr_pages
, max_boost
);
2373 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2378 * This function implements actual steal behaviour. If order is large enough,
2379 * we can steal whole pageblock. If not, we first move freepages in this
2380 * pageblock to our migratetype and determine how many already-allocated pages
2381 * are there in the pageblock with a compatible migratetype. If at least half
2382 * of pages are free or compatible, we can change migratetype of the pageblock
2383 * itself, so pages freed in the future will be put on the correct free list.
2385 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2386 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2388 unsigned int current_order
= page_order(page
);
2389 struct free_area
*area
;
2390 int free_pages
, movable_pages
, alike_pages
;
2393 old_block_type
= get_pageblock_migratetype(page
);
2396 * This can happen due to races and we want to prevent broken
2397 * highatomic accounting.
2399 if (is_migrate_highatomic(old_block_type
))
2402 /* Take ownership for orders >= pageblock_order */
2403 if (current_order
>= pageblock_order
) {
2404 change_pageblock_range(page
, current_order
, start_type
);
2409 * Boost watermarks to increase reclaim pressure to reduce the
2410 * likelihood of future fallbacks. Wake kswapd now as the node
2411 * may be balanced overall and kswapd will not wake naturally.
2413 boost_watermark(zone
);
2414 if (alloc_flags
& ALLOC_KSWAPD
)
2415 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2417 /* We are not allowed to try stealing from the whole block */
2421 free_pages
= move_freepages_block(zone
, page
, start_type
,
2424 * Determine how many pages are compatible with our allocation.
2425 * For movable allocation, it's the number of movable pages which
2426 * we just obtained. For other types it's a bit more tricky.
2428 if (start_type
== MIGRATE_MOVABLE
) {
2429 alike_pages
= movable_pages
;
2432 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2433 * to MOVABLE pageblock, consider all non-movable pages as
2434 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2435 * vice versa, be conservative since we can't distinguish the
2436 * exact migratetype of non-movable pages.
2438 if (old_block_type
== MIGRATE_MOVABLE
)
2439 alike_pages
= pageblock_nr_pages
2440 - (free_pages
+ movable_pages
);
2445 /* moving whole block can fail due to zone boundary conditions */
2450 * If a sufficient number of pages in the block are either free or of
2451 * comparable migratability as our allocation, claim the whole block.
2453 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2454 page_group_by_mobility_disabled
)
2455 set_pageblock_migratetype(page
, start_type
);
2460 area
= &zone
->free_area
[current_order
];
2461 move_to_free_area(page
, area
, start_type
);
2465 * Check whether there is a suitable fallback freepage with requested order.
2466 * If only_stealable is true, this function returns fallback_mt only if
2467 * we can steal other freepages all together. This would help to reduce
2468 * fragmentation due to mixed migratetype pages in one pageblock.
2470 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2471 int migratetype
, bool only_stealable
, bool *can_steal
)
2476 if (area
->nr_free
== 0)
2481 fallback_mt
= fallbacks
[migratetype
][i
];
2482 if (fallback_mt
== MIGRATE_TYPES
)
2485 if (free_area_empty(area
, fallback_mt
))
2488 if (can_steal_fallback(order
, migratetype
))
2491 if (!only_stealable
)
2502 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2503 * there are no empty page blocks that contain a page with a suitable order
2505 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2506 unsigned int alloc_order
)
2509 unsigned long max_managed
, flags
;
2512 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2513 * Check is race-prone but harmless.
2515 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2516 if (zone
->nr_reserved_highatomic
>= max_managed
)
2519 spin_lock_irqsave(&zone
->lock
, flags
);
2521 /* Recheck the nr_reserved_highatomic limit under the lock */
2522 if (zone
->nr_reserved_highatomic
>= max_managed
)
2526 mt
= get_pageblock_migratetype(page
);
2527 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2528 && !is_migrate_cma(mt
)) {
2529 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2530 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2531 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2535 spin_unlock_irqrestore(&zone
->lock
, flags
);
2539 * Used when an allocation is about to fail under memory pressure. This
2540 * potentially hurts the reliability of high-order allocations when under
2541 * intense memory pressure but failed atomic allocations should be easier
2542 * to recover from than an OOM.
2544 * If @force is true, try to unreserve a pageblock even though highatomic
2545 * pageblock is exhausted.
2547 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2550 struct zonelist
*zonelist
= ac
->zonelist
;
2551 unsigned long flags
;
2558 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2561 * Preserve at least one pageblock unless memory pressure
2564 if (!force
&& zone
->nr_reserved_highatomic
<=
2568 spin_lock_irqsave(&zone
->lock
, flags
);
2569 for (order
= 0; order
< MAX_ORDER
; order
++) {
2570 struct free_area
*area
= &(zone
->free_area
[order
]);
2572 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2577 * In page freeing path, migratetype change is racy so
2578 * we can counter several free pages in a pageblock
2579 * in this loop althoug we changed the pageblock type
2580 * from highatomic to ac->migratetype. So we should
2581 * adjust the count once.
2583 if (is_migrate_highatomic_page(page
)) {
2585 * It should never happen but changes to
2586 * locking could inadvertently allow a per-cpu
2587 * drain to add pages to MIGRATE_HIGHATOMIC
2588 * while unreserving so be safe and watch for
2591 zone
->nr_reserved_highatomic
-= min(
2593 zone
->nr_reserved_highatomic
);
2597 * Convert to ac->migratetype and avoid the normal
2598 * pageblock stealing heuristics. Minimally, the caller
2599 * is doing the work and needs the pages. More
2600 * importantly, if the block was always converted to
2601 * MIGRATE_UNMOVABLE or another type then the number
2602 * of pageblocks that cannot be completely freed
2605 set_pageblock_migratetype(page
, ac
->migratetype
);
2606 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2609 spin_unlock_irqrestore(&zone
->lock
, flags
);
2613 spin_unlock_irqrestore(&zone
->lock
, flags
);
2620 * Try finding a free buddy page on the fallback list and put it on the free
2621 * list of requested migratetype, possibly along with other pages from the same
2622 * block, depending on fragmentation avoidance heuristics. Returns true if
2623 * fallback was found so that __rmqueue_smallest() can grab it.
2625 * The use of signed ints for order and current_order is a deliberate
2626 * deviation from the rest of this file, to make the for loop
2627 * condition simpler.
2629 static __always_inline
bool
2630 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2631 unsigned int alloc_flags
)
2633 struct free_area
*area
;
2635 int min_order
= order
;
2641 * Do not steal pages from freelists belonging to other pageblocks
2642 * i.e. orders < pageblock_order. If there are no local zones free,
2643 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2645 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2646 min_order
= pageblock_order
;
2649 * Find the largest available free page in the other list. This roughly
2650 * approximates finding the pageblock with the most free pages, which
2651 * would be too costly to do exactly.
2653 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2655 area
= &(zone
->free_area
[current_order
]);
2656 fallback_mt
= find_suitable_fallback(area
, current_order
,
2657 start_migratetype
, false, &can_steal
);
2658 if (fallback_mt
== -1)
2662 * We cannot steal all free pages from the pageblock and the
2663 * requested migratetype is movable. In that case it's better to
2664 * steal and split the smallest available page instead of the
2665 * largest available page, because even if the next movable
2666 * allocation falls back into a different pageblock than this
2667 * one, it won't cause permanent fragmentation.
2669 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2670 && current_order
> order
)
2679 for (current_order
= order
; current_order
< MAX_ORDER
;
2681 area
= &(zone
->free_area
[current_order
]);
2682 fallback_mt
= find_suitable_fallback(area
, current_order
,
2683 start_migratetype
, false, &can_steal
);
2684 if (fallback_mt
!= -1)
2689 * This should not happen - we already found a suitable fallback
2690 * when looking for the largest page.
2692 VM_BUG_ON(current_order
== MAX_ORDER
);
2695 page
= get_page_from_free_area(area
, fallback_mt
);
2697 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2700 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2701 start_migratetype
, fallback_mt
);
2708 * Do the hard work of removing an element from the buddy allocator.
2709 * Call me with the zone->lock already held.
2711 static __always_inline
struct page
*
2712 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2713 unsigned int alloc_flags
)
2718 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2719 if (unlikely(!page
)) {
2720 if (migratetype
== MIGRATE_MOVABLE
)
2721 page
= __rmqueue_cma_fallback(zone
, order
);
2723 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2728 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2733 * Obtain a specified number of elements from the buddy allocator, all under
2734 * a single hold of the lock, for efficiency. Add them to the supplied list.
2735 * Returns the number of new pages which were placed at *list.
2737 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2738 unsigned long count
, struct list_head
*list
,
2739 int migratetype
, unsigned int alloc_flags
)
2743 spin_lock(&zone
->lock
);
2744 for (i
= 0; i
< count
; ++i
) {
2745 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2747 if (unlikely(page
== NULL
))
2750 if (unlikely(check_pcp_refill(page
)))
2754 * Split buddy pages returned by expand() are received here in
2755 * physical page order. The page is added to the tail of
2756 * caller's list. From the callers perspective, the linked list
2757 * is ordered by page number under some conditions. This is
2758 * useful for IO devices that can forward direction from the
2759 * head, thus also in the physical page order. This is useful
2760 * for IO devices that can merge IO requests if the physical
2761 * pages are ordered properly.
2763 list_add_tail(&page
->lru
, list
);
2765 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2766 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2771 * i pages were removed from the buddy list even if some leak due
2772 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2773 * on i. Do not confuse with 'alloced' which is the number of
2774 * pages added to the pcp list.
2776 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2777 spin_unlock(&zone
->lock
);
2783 * Called from the vmstat counter updater to drain pagesets of this
2784 * currently executing processor on remote nodes after they have
2787 * Note that this function must be called with the thread pinned to
2788 * a single processor.
2790 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2792 unsigned long flags
;
2793 int to_drain
, batch
;
2795 local_irq_save(flags
);
2796 batch
= READ_ONCE(pcp
->batch
);
2797 to_drain
= min(pcp
->count
, batch
);
2799 free_pcppages_bulk(zone
, to_drain
, pcp
);
2800 local_irq_restore(flags
);
2805 * Drain pcplists of the indicated processor and zone.
2807 * The processor must either be the current processor and the
2808 * thread pinned to the current processor or a processor that
2811 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2813 unsigned long flags
;
2814 struct per_cpu_pageset
*pset
;
2815 struct per_cpu_pages
*pcp
;
2817 local_irq_save(flags
);
2818 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2822 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2823 local_irq_restore(flags
);
2827 * Drain pcplists of all zones on the indicated processor.
2829 * The processor must either be the current processor and the
2830 * thread pinned to the current processor or a processor that
2833 static void drain_pages(unsigned int cpu
)
2837 for_each_populated_zone(zone
) {
2838 drain_pages_zone(cpu
, zone
);
2843 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2845 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2846 * the single zone's pages.
2848 void drain_local_pages(struct zone
*zone
)
2850 int cpu
= smp_processor_id();
2853 drain_pages_zone(cpu
, zone
);
2858 static void drain_local_pages_wq(struct work_struct
*work
)
2860 struct pcpu_drain
*drain
;
2862 drain
= container_of(work
, struct pcpu_drain
, work
);
2865 * drain_all_pages doesn't use proper cpu hotplug protection so
2866 * we can race with cpu offline when the WQ can move this from
2867 * a cpu pinned worker to an unbound one. We can operate on a different
2868 * cpu which is allright but we also have to make sure to not move to
2872 drain_local_pages(drain
->zone
);
2877 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2879 * When zone parameter is non-NULL, spill just the single zone's pages.
2881 * Note that this can be extremely slow as the draining happens in a workqueue.
2883 void drain_all_pages(struct zone
*zone
)
2888 * Allocate in the BSS so we wont require allocation in
2889 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2891 static cpumask_t cpus_with_pcps
;
2894 * Make sure nobody triggers this path before mm_percpu_wq is fully
2897 if (WARN_ON_ONCE(!mm_percpu_wq
))
2901 * Do not drain if one is already in progress unless it's specific to
2902 * a zone. Such callers are primarily CMA and memory hotplug and need
2903 * the drain to be complete when the call returns.
2905 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2908 mutex_lock(&pcpu_drain_mutex
);
2912 * We don't care about racing with CPU hotplug event
2913 * as offline notification will cause the notified
2914 * cpu to drain that CPU pcps and on_each_cpu_mask
2915 * disables preemption as part of its processing
2917 for_each_online_cpu(cpu
) {
2918 struct per_cpu_pageset
*pcp
;
2920 bool has_pcps
= false;
2923 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2927 for_each_populated_zone(z
) {
2928 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2929 if (pcp
->pcp
.count
) {
2937 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2939 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2942 for_each_cpu(cpu
, &cpus_with_pcps
) {
2943 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
2946 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
2947 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
2949 for_each_cpu(cpu
, &cpus_with_pcps
)
2950 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
2952 mutex_unlock(&pcpu_drain_mutex
);
2955 #ifdef CONFIG_HIBERNATION
2958 * Touch the watchdog for every WD_PAGE_COUNT pages.
2960 #define WD_PAGE_COUNT (128*1024)
2962 void mark_free_pages(struct zone
*zone
)
2964 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2965 unsigned long flags
;
2966 unsigned int order
, t
;
2969 if (zone_is_empty(zone
))
2972 spin_lock_irqsave(&zone
->lock
, flags
);
2974 max_zone_pfn
= zone_end_pfn(zone
);
2975 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2976 if (pfn_valid(pfn
)) {
2977 page
= pfn_to_page(pfn
);
2979 if (!--page_count
) {
2980 touch_nmi_watchdog();
2981 page_count
= WD_PAGE_COUNT
;
2984 if (page_zone(page
) != zone
)
2987 if (!swsusp_page_is_forbidden(page
))
2988 swsusp_unset_page_free(page
);
2991 for_each_migratetype_order(order
, t
) {
2992 list_for_each_entry(page
,
2993 &zone
->free_area
[order
].free_list
[t
], lru
) {
2996 pfn
= page_to_pfn(page
);
2997 for (i
= 0; i
< (1UL << order
); i
++) {
2998 if (!--page_count
) {
2999 touch_nmi_watchdog();
3000 page_count
= WD_PAGE_COUNT
;
3002 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
3006 spin_unlock_irqrestore(&zone
->lock
, flags
);
3008 #endif /* CONFIG_PM */
3010 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
3014 if (!free_pcp_prepare(page
))
3017 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3018 set_pcppage_migratetype(page
, migratetype
);
3022 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
3024 struct zone
*zone
= page_zone(page
);
3025 struct per_cpu_pages
*pcp
;
3028 migratetype
= get_pcppage_migratetype(page
);
3029 __count_vm_event(PGFREE
);
3032 * We only track unmovable, reclaimable and movable on pcp lists.
3033 * Free ISOLATE pages back to the allocator because they are being
3034 * offlined but treat HIGHATOMIC as movable pages so we can get those
3035 * areas back if necessary. Otherwise, we may have to free
3036 * excessively into the page allocator
3038 if (migratetype
>= MIGRATE_PCPTYPES
) {
3039 if (unlikely(is_migrate_isolate(migratetype
))) {
3040 free_one_page(zone
, page
, pfn
, 0, migratetype
);
3043 migratetype
= MIGRATE_MOVABLE
;
3046 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3047 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
3049 if (pcp
->count
>= pcp
->high
) {
3050 unsigned long batch
= READ_ONCE(pcp
->batch
);
3051 free_pcppages_bulk(zone
, batch
, pcp
);
3056 * Free a 0-order page
3058 void free_unref_page(struct page
*page
)
3060 unsigned long flags
;
3061 unsigned long pfn
= page_to_pfn(page
);
3063 if (!free_unref_page_prepare(page
, pfn
))
3066 local_irq_save(flags
);
3067 free_unref_page_commit(page
, pfn
);
3068 local_irq_restore(flags
);
3072 * Free a list of 0-order pages
3074 void free_unref_page_list(struct list_head
*list
)
3076 struct page
*page
, *next
;
3077 unsigned long flags
, pfn
;
3078 int batch_count
= 0;
3080 /* Prepare pages for freeing */
3081 list_for_each_entry_safe(page
, next
, list
, lru
) {
3082 pfn
= page_to_pfn(page
);
3083 if (!free_unref_page_prepare(page
, pfn
))
3084 list_del(&page
->lru
);
3085 set_page_private(page
, pfn
);
3088 local_irq_save(flags
);
3089 list_for_each_entry_safe(page
, next
, list
, lru
) {
3090 unsigned long pfn
= page_private(page
);
3092 set_page_private(page
, 0);
3093 trace_mm_page_free_batched(page
);
3094 free_unref_page_commit(page
, pfn
);
3097 * Guard against excessive IRQ disabled times when we get
3098 * a large list of pages to free.
3100 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3101 local_irq_restore(flags
);
3103 local_irq_save(flags
);
3106 local_irq_restore(flags
);
3110 * split_page takes a non-compound higher-order page, and splits it into
3111 * n (1<<order) sub-pages: page[0..n]
3112 * Each sub-page must be freed individually.
3114 * Note: this is probably too low level an operation for use in drivers.
3115 * Please consult with lkml before using this in your driver.
3117 void split_page(struct page
*page
, unsigned int order
)
3121 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3122 VM_BUG_ON_PAGE(!page_count(page
), page
);
3124 for (i
= 1; i
< (1 << order
); i
++)
3125 set_page_refcounted(page
+ i
);
3126 split_page_owner(page
, order
);
3128 EXPORT_SYMBOL_GPL(split_page
);
3130 int __isolate_free_page(struct page
*page
, unsigned int order
)
3132 struct free_area
*area
= &page_zone(page
)->free_area
[order
];
3133 unsigned long watermark
;
3137 BUG_ON(!PageBuddy(page
));
3139 zone
= page_zone(page
);
3140 mt
= get_pageblock_migratetype(page
);
3142 if (!is_migrate_isolate(mt
)) {
3144 * Obey watermarks as if the page was being allocated. We can
3145 * emulate a high-order watermark check with a raised order-0
3146 * watermark, because we already know our high-order page
3149 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3150 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3153 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3156 /* Remove page from free list */
3158 del_page_from_free_area(page
, area
);
3161 * Set the pageblock if the isolated page is at least half of a
3164 if (order
>= pageblock_order
- 1) {
3165 struct page
*endpage
= page
+ (1 << order
) - 1;
3166 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3167 int mt
= get_pageblock_migratetype(page
);
3168 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3169 && !is_migrate_highatomic(mt
))
3170 set_pageblock_migratetype(page
,
3176 return 1UL << order
;
3180 * Update NUMA hit/miss statistics
3182 * Must be called with interrupts disabled.
3184 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
3187 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3189 /* skip numa counters update if numa stats is disabled */
3190 if (!static_branch_likely(&vm_numa_stat_key
))
3193 if (zone_to_nid(z
) != numa_node_id())
3194 local_stat
= NUMA_OTHER
;
3196 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3197 __inc_numa_state(z
, NUMA_HIT
);
3199 __inc_numa_state(z
, NUMA_MISS
);
3200 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3202 __inc_numa_state(z
, local_stat
);
3206 /* Remove page from the per-cpu list, caller must protect the list */
3207 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3208 unsigned int alloc_flags
,
3209 struct per_cpu_pages
*pcp
,
3210 struct list_head
*list
)
3215 if (list_empty(list
)) {
3216 pcp
->count
+= rmqueue_bulk(zone
, 0,
3218 migratetype
, alloc_flags
);
3219 if (unlikely(list_empty(list
)))
3223 page
= list_first_entry(list
, struct page
, lru
);
3224 list_del(&page
->lru
);
3226 } while (check_new_pcp(page
));
3231 /* Lock and remove page from the per-cpu list */
3232 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3233 struct zone
*zone
, gfp_t gfp_flags
,
3234 int migratetype
, unsigned int alloc_flags
)
3236 struct per_cpu_pages
*pcp
;
3237 struct list_head
*list
;
3239 unsigned long flags
;
3241 local_irq_save(flags
);
3242 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3243 list
= &pcp
->lists
[migratetype
];
3244 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3246 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3247 zone_statistics(preferred_zone
, zone
);
3249 local_irq_restore(flags
);
3254 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3257 struct page
*rmqueue(struct zone
*preferred_zone
,
3258 struct zone
*zone
, unsigned int order
,
3259 gfp_t gfp_flags
, unsigned int alloc_flags
,
3262 unsigned long flags
;
3265 if (likely(order
== 0)) {
3266 page
= rmqueue_pcplist(preferred_zone
, zone
, gfp_flags
,
3267 migratetype
, alloc_flags
);
3272 * We most definitely don't want callers attempting to
3273 * allocate greater than order-1 page units with __GFP_NOFAIL.
3275 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3276 spin_lock_irqsave(&zone
->lock
, flags
);
3280 if (alloc_flags
& ALLOC_HARDER
) {
3281 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3283 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3286 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3287 } while (page
&& check_new_pages(page
, order
));
3288 spin_unlock(&zone
->lock
);
3291 __mod_zone_freepage_state(zone
, -(1 << order
),
3292 get_pcppage_migratetype(page
));
3294 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3295 zone_statistics(preferred_zone
, zone
);
3296 local_irq_restore(flags
);
3299 /* Separate test+clear to avoid unnecessary atomics */
3300 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3301 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3302 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3305 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3309 local_irq_restore(flags
);
3313 #ifdef CONFIG_FAIL_PAGE_ALLOC
3316 struct fault_attr attr
;
3318 bool ignore_gfp_highmem
;
3319 bool ignore_gfp_reclaim
;
3321 } fail_page_alloc
= {
3322 .attr
= FAULT_ATTR_INITIALIZER
,
3323 .ignore_gfp_reclaim
= true,
3324 .ignore_gfp_highmem
= true,
3328 static int __init
setup_fail_page_alloc(char *str
)
3330 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3332 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3334 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3336 if (order
< fail_page_alloc
.min_order
)
3338 if (gfp_mask
& __GFP_NOFAIL
)
3340 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3342 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3343 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3346 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3349 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3351 static int __init
fail_page_alloc_debugfs(void)
3353 umode_t mode
= S_IFREG
| 0600;
3356 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3357 &fail_page_alloc
.attr
);
3359 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3360 &fail_page_alloc
.ignore_gfp_reclaim
);
3361 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3362 &fail_page_alloc
.ignore_gfp_highmem
);
3363 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3368 late_initcall(fail_page_alloc_debugfs
);
3370 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3372 #else /* CONFIG_FAIL_PAGE_ALLOC */
3374 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3379 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3381 static noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3383 return __should_fail_alloc_page(gfp_mask
, order
);
3385 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3388 * Return true if free base pages are above 'mark'. For high-order checks it
3389 * will return true of the order-0 watermark is reached and there is at least
3390 * one free page of a suitable size. Checking now avoids taking the zone lock
3391 * to check in the allocation paths if no pages are free.
3393 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3394 int classzone_idx
, unsigned int alloc_flags
,
3399 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3401 /* free_pages may go negative - that's OK */
3402 free_pages
-= (1 << order
) - 1;
3404 if (alloc_flags
& ALLOC_HIGH
)
3408 * If the caller does not have rights to ALLOC_HARDER then subtract
3409 * the high-atomic reserves. This will over-estimate the size of the
3410 * atomic reserve but it avoids a search.
3412 if (likely(!alloc_harder
)) {
3413 free_pages
-= z
->nr_reserved_highatomic
;
3416 * OOM victims can try even harder than normal ALLOC_HARDER
3417 * users on the grounds that it's definitely going to be in
3418 * the exit path shortly and free memory. Any allocation it
3419 * makes during the free path will be small and short-lived.
3421 if (alloc_flags
& ALLOC_OOM
)
3429 /* If allocation can't use CMA areas don't use free CMA pages */
3430 if (!(alloc_flags
& ALLOC_CMA
))
3431 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3435 * Check watermarks for an order-0 allocation request. If these
3436 * are not met, then a high-order request also cannot go ahead
3437 * even if a suitable page happened to be free.
3439 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3442 /* If this is an order-0 request then the watermark is fine */
3446 /* For a high-order request, check at least one suitable page is free */
3447 for (o
= order
; o
< MAX_ORDER
; o
++) {
3448 struct free_area
*area
= &z
->free_area
[o
];
3454 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3455 if (!free_area_empty(area
, mt
))
3460 if ((alloc_flags
& ALLOC_CMA
) &&
3461 !free_area_empty(area
, MIGRATE_CMA
)) {
3466 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3472 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3473 int classzone_idx
, unsigned int alloc_flags
)
3475 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3476 zone_page_state(z
, NR_FREE_PAGES
));
3479 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3480 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3482 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3486 /* If allocation can't use CMA areas don't use free CMA pages */
3487 if (!(alloc_flags
& ALLOC_CMA
))
3488 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3492 * Fast check for order-0 only. If this fails then the reserves
3493 * need to be calculated. There is a corner case where the check
3494 * passes but only the high-order atomic reserve are free. If
3495 * the caller is !atomic then it'll uselessly search the free
3496 * list. That corner case is then slower but it is harmless.
3498 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3501 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3505 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3506 unsigned long mark
, int classzone_idx
)
3508 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3510 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3511 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3513 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3518 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3520 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3521 node_reclaim_distance
;
3523 #else /* CONFIG_NUMA */
3524 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3528 #endif /* CONFIG_NUMA */
3531 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3532 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3533 * premature use of a lower zone may cause lowmem pressure problems that
3534 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3535 * probably too small. It only makes sense to spread allocations to avoid
3536 * fragmentation between the Normal and DMA32 zones.
3538 static inline unsigned int
3539 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3541 unsigned int alloc_flags
= 0;
3543 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3544 alloc_flags
|= ALLOC_KSWAPD
;
3546 #ifdef CONFIG_ZONE_DMA32
3550 if (zone_idx(zone
) != ZONE_NORMAL
)
3554 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3555 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3556 * on UMA that if Normal is populated then so is DMA32.
3558 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3559 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3562 alloc_flags
|= ALLOC_NOFRAGMENT
;
3563 #endif /* CONFIG_ZONE_DMA32 */
3568 * get_page_from_freelist goes through the zonelist trying to allocate
3571 static struct page
*
3572 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3573 const struct alloc_context
*ac
)
3577 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3582 * Scan zonelist, looking for a zone with enough free.
3583 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3585 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3586 z
= ac
->preferred_zoneref
;
3587 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3592 if (cpusets_enabled() &&
3593 (alloc_flags
& ALLOC_CPUSET
) &&
3594 !__cpuset_zone_allowed(zone
, gfp_mask
))
3597 * When allocating a page cache page for writing, we
3598 * want to get it from a node that is within its dirty
3599 * limit, such that no single node holds more than its
3600 * proportional share of globally allowed dirty pages.
3601 * The dirty limits take into account the node's
3602 * lowmem reserves and high watermark so that kswapd
3603 * should be able to balance it without having to
3604 * write pages from its LRU list.
3606 * XXX: For now, allow allocations to potentially
3607 * exceed the per-node dirty limit in the slowpath
3608 * (spread_dirty_pages unset) before going into reclaim,
3609 * which is important when on a NUMA setup the allowed
3610 * nodes are together not big enough to reach the
3611 * global limit. The proper fix for these situations
3612 * will require awareness of nodes in the
3613 * dirty-throttling and the flusher threads.
3615 if (ac
->spread_dirty_pages
) {
3616 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3619 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3620 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3625 if (no_fallback
&& nr_online_nodes
> 1 &&
3626 zone
!= ac
->preferred_zoneref
->zone
) {
3630 * If moving to a remote node, retry but allow
3631 * fragmenting fallbacks. Locality is more important
3632 * than fragmentation avoidance.
3634 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3635 if (zone_to_nid(zone
) != local_nid
) {
3636 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3641 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3642 if (!zone_watermark_fast(zone
, order
, mark
,
3643 ac_classzone_idx(ac
), alloc_flags
)) {
3646 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3648 * Watermark failed for this zone, but see if we can
3649 * grow this zone if it contains deferred pages.
3651 if (static_branch_unlikely(&deferred_pages
)) {
3652 if (_deferred_grow_zone(zone
, order
))
3656 /* Checked here to keep the fast path fast */
3657 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3658 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3661 if (node_reclaim_mode
== 0 ||
3662 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3665 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3667 case NODE_RECLAIM_NOSCAN
:
3670 case NODE_RECLAIM_FULL
:
3671 /* scanned but unreclaimable */
3674 /* did we reclaim enough */
3675 if (zone_watermark_ok(zone
, order
, mark
,
3676 ac_classzone_idx(ac
), alloc_flags
))
3684 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3685 gfp_mask
, alloc_flags
, ac
->migratetype
);
3687 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3690 * If this is a high-order atomic allocation then check
3691 * if the pageblock should be reserved for the future
3693 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3694 reserve_highatomic_pageblock(page
, zone
, order
);
3698 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3699 /* Try again if zone has deferred pages */
3700 if (static_branch_unlikely(&deferred_pages
)) {
3701 if (_deferred_grow_zone(zone
, order
))
3709 * It's possible on a UMA machine to get through all zones that are
3710 * fragmented. If avoiding fragmentation, reset and try again.
3713 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3720 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3722 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3723 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3725 if (!__ratelimit(&show_mem_rs
))
3729 * This documents exceptions given to allocations in certain
3730 * contexts that are allowed to allocate outside current's set
3733 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3734 if (tsk_is_oom_victim(current
) ||
3735 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3736 filter
&= ~SHOW_MEM_FILTER_NODES
;
3737 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3738 filter
&= ~SHOW_MEM_FILTER_NODES
;
3740 show_mem(filter
, nodemask
);
3743 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3745 struct va_format vaf
;
3747 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3748 DEFAULT_RATELIMIT_BURST
);
3750 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3753 va_start(args
, fmt
);
3756 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3757 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3758 nodemask_pr_args(nodemask
));
3761 cpuset_print_current_mems_allowed();
3764 warn_alloc_show_mem(gfp_mask
, nodemask
);
3767 static inline struct page
*
3768 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3769 unsigned int alloc_flags
,
3770 const struct alloc_context
*ac
)
3774 page
= get_page_from_freelist(gfp_mask
, order
,
3775 alloc_flags
|ALLOC_CPUSET
, ac
);
3777 * fallback to ignore cpuset restriction if our nodes
3781 page
= get_page_from_freelist(gfp_mask
, order
,
3787 static inline struct page
*
3788 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3789 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3791 struct oom_control oc
= {
3792 .zonelist
= ac
->zonelist
,
3793 .nodemask
= ac
->nodemask
,
3795 .gfp_mask
= gfp_mask
,
3800 *did_some_progress
= 0;
3803 * Acquire the oom lock. If that fails, somebody else is
3804 * making progress for us.
3806 if (!mutex_trylock(&oom_lock
)) {
3807 *did_some_progress
= 1;
3808 schedule_timeout_uninterruptible(1);
3813 * Go through the zonelist yet one more time, keep very high watermark
3814 * here, this is only to catch a parallel oom killing, we must fail if
3815 * we're still under heavy pressure. But make sure that this reclaim
3816 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3817 * allocation which will never fail due to oom_lock already held.
3819 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3820 ~__GFP_DIRECT_RECLAIM
, order
,
3821 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3825 /* Coredumps can quickly deplete all memory reserves */
3826 if (current
->flags
& PF_DUMPCORE
)
3828 /* The OOM killer will not help higher order allocs */
3829 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3832 * We have already exhausted all our reclaim opportunities without any
3833 * success so it is time to admit defeat. We will skip the OOM killer
3834 * because it is very likely that the caller has a more reasonable
3835 * fallback than shooting a random task.
3837 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3839 /* The OOM killer does not needlessly kill tasks for lowmem */
3840 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3842 if (pm_suspended_storage())
3845 * XXX: GFP_NOFS allocations should rather fail than rely on
3846 * other request to make a forward progress.
3847 * We are in an unfortunate situation where out_of_memory cannot
3848 * do much for this context but let's try it to at least get
3849 * access to memory reserved if the current task is killed (see
3850 * out_of_memory). Once filesystems are ready to handle allocation
3851 * failures more gracefully we should just bail out here.
3854 /* The OOM killer may not free memory on a specific node */
3855 if (gfp_mask
& __GFP_THISNODE
)
3858 /* Exhausted what can be done so it's blame time */
3859 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3860 *did_some_progress
= 1;
3863 * Help non-failing allocations by giving them access to memory
3866 if (gfp_mask
& __GFP_NOFAIL
)
3867 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3868 ALLOC_NO_WATERMARKS
, ac
);
3871 mutex_unlock(&oom_lock
);
3876 * Maximum number of compaction retries wit a progress before OOM
3877 * killer is consider as the only way to move forward.
3879 #define MAX_COMPACT_RETRIES 16
3881 #ifdef CONFIG_COMPACTION
3882 /* Try memory compaction for high-order allocations before reclaim */
3883 static struct page
*
3884 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3885 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3886 enum compact_priority prio
, enum compact_result
*compact_result
)
3888 struct page
*page
= NULL
;
3889 unsigned long pflags
;
3890 unsigned int noreclaim_flag
;
3895 psi_memstall_enter(&pflags
);
3896 noreclaim_flag
= memalloc_noreclaim_save();
3898 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3901 memalloc_noreclaim_restore(noreclaim_flag
);
3902 psi_memstall_leave(&pflags
);
3905 * At least in one zone compaction wasn't deferred or skipped, so let's
3906 * count a compaction stall
3908 count_vm_event(COMPACTSTALL
);
3910 /* Prep a captured page if available */
3912 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3914 /* Try get a page from the freelist if available */
3916 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3919 struct zone
*zone
= page_zone(page
);
3921 zone
->compact_blockskip_flush
= false;
3922 compaction_defer_reset(zone
, order
, true);
3923 count_vm_event(COMPACTSUCCESS
);
3928 * It's bad if compaction run occurs and fails. The most likely reason
3929 * is that pages exist, but not enough to satisfy watermarks.
3931 count_vm_event(COMPACTFAIL
);
3939 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3940 enum compact_result compact_result
,
3941 enum compact_priority
*compact_priority
,
3942 int *compaction_retries
)
3944 int max_retries
= MAX_COMPACT_RETRIES
;
3947 int retries
= *compaction_retries
;
3948 enum compact_priority priority
= *compact_priority
;
3953 if (compaction_made_progress(compact_result
))
3954 (*compaction_retries
)++;
3957 * compaction considers all the zone as desperately out of memory
3958 * so it doesn't really make much sense to retry except when the
3959 * failure could be caused by insufficient priority
3961 if (compaction_failed(compact_result
))
3962 goto check_priority
;
3965 * compaction was skipped because there are not enough order-0 pages
3966 * to work with, so we retry only if it looks like reclaim can help.
3968 if (compaction_needs_reclaim(compact_result
)) {
3969 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3974 * make sure the compaction wasn't deferred or didn't bail out early
3975 * due to locks contention before we declare that we should give up.
3976 * But the next retry should use a higher priority if allowed, so
3977 * we don't just keep bailing out endlessly.
3979 if (compaction_withdrawn(compact_result
)) {
3980 goto check_priority
;
3984 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3985 * costly ones because they are de facto nofail and invoke OOM
3986 * killer to move on while costly can fail and users are ready
3987 * to cope with that. 1/4 retries is rather arbitrary but we
3988 * would need much more detailed feedback from compaction to
3989 * make a better decision.
3991 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3993 if (*compaction_retries
<= max_retries
) {
3999 * Make sure there are attempts at the highest priority if we exhausted
4000 * all retries or failed at the lower priorities.
4003 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
4004 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
4006 if (*compact_priority
> min_priority
) {
4007 (*compact_priority
)--;
4008 *compaction_retries
= 0;
4012 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4016 static inline struct page
*
4017 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4018 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4019 enum compact_priority prio
, enum compact_result
*compact_result
)
4021 *compact_result
= COMPACT_SKIPPED
;
4026 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4027 enum compact_result compact_result
,
4028 enum compact_priority
*compact_priority
,
4029 int *compaction_retries
)
4034 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4038 * There are setups with compaction disabled which would prefer to loop
4039 * inside the allocator rather than hit the oom killer prematurely.
4040 * Let's give them a good hope and keep retrying while the order-0
4041 * watermarks are OK.
4043 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4045 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4046 ac_classzone_idx(ac
), alloc_flags
))
4051 #endif /* CONFIG_COMPACTION */
4053 #ifdef CONFIG_LOCKDEP
4054 static struct lockdep_map __fs_reclaim_map
=
4055 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4057 static bool __need_fs_reclaim(gfp_t gfp_mask
)
4059 gfp_mask
= current_gfp_context(gfp_mask
);
4061 /* no reclaim without waiting on it */
4062 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4065 /* this guy won't enter reclaim */
4066 if (current
->flags
& PF_MEMALLOC
)
4069 /* We're only interested __GFP_FS allocations for now */
4070 if (!(gfp_mask
& __GFP_FS
))
4073 if (gfp_mask
& __GFP_NOLOCKDEP
)
4079 void __fs_reclaim_acquire(void)
4081 lock_map_acquire(&__fs_reclaim_map
);
4084 void __fs_reclaim_release(void)
4086 lock_map_release(&__fs_reclaim_map
);
4089 void fs_reclaim_acquire(gfp_t gfp_mask
)
4091 if (__need_fs_reclaim(gfp_mask
))
4092 __fs_reclaim_acquire();
4094 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4096 void fs_reclaim_release(gfp_t gfp_mask
)
4098 if (__need_fs_reclaim(gfp_mask
))
4099 __fs_reclaim_release();
4101 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4104 /* Perform direct synchronous page reclaim */
4106 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4107 const struct alloc_context
*ac
)
4110 unsigned int noreclaim_flag
;
4111 unsigned long pflags
;
4115 /* We now go into synchronous reclaim */
4116 cpuset_memory_pressure_bump();
4117 psi_memstall_enter(&pflags
);
4118 fs_reclaim_acquire(gfp_mask
);
4119 noreclaim_flag
= memalloc_noreclaim_save();
4121 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4124 memalloc_noreclaim_restore(noreclaim_flag
);
4125 fs_reclaim_release(gfp_mask
);
4126 psi_memstall_leave(&pflags
);
4133 /* The really slow allocator path where we enter direct reclaim */
4134 static inline struct page
*
4135 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4136 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4137 unsigned long *did_some_progress
)
4139 struct page
*page
= NULL
;
4140 bool drained
= false;
4142 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4143 if (unlikely(!(*did_some_progress
)))
4147 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4150 * If an allocation failed after direct reclaim, it could be because
4151 * pages are pinned on the per-cpu lists or in high alloc reserves.
4152 * Shrink them them and try again
4154 if (!page
&& !drained
) {
4155 unreserve_highatomic_pageblock(ac
, false);
4156 drain_all_pages(NULL
);
4164 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4165 const struct alloc_context
*ac
)
4169 pg_data_t
*last_pgdat
= NULL
;
4170 enum zone_type high_zoneidx
= ac
->high_zoneidx
;
4172 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, high_zoneidx
,
4174 if (last_pgdat
!= zone
->zone_pgdat
)
4175 wakeup_kswapd(zone
, gfp_mask
, order
, high_zoneidx
);
4176 last_pgdat
= zone
->zone_pgdat
;
4180 static inline unsigned int
4181 gfp_to_alloc_flags(gfp_t gfp_mask
)
4183 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4185 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4186 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4189 * The caller may dip into page reserves a bit more if the caller
4190 * cannot run direct reclaim, or if the caller has realtime scheduling
4191 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4192 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4194 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
4196 if (gfp_mask
& __GFP_ATOMIC
) {
4198 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4199 * if it can't schedule.
4201 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4202 alloc_flags
|= ALLOC_HARDER
;
4204 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4205 * comment for __cpuset_node_allowed().
4207 alloc_flags
&= ~ALLOC_CPUSET
;
4208 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4209 alloc_flags
|= ALLOC_HARDER
;
4211 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4212 alloc_flags
|= ALLOC_KSWAPD
;
4215 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
4216 alloc_flags
|= ALLOC_CMA
;
4221 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4223 if (!tsk_is_oom_victim(tsk
))
4227 * !MMU doesn't have oom reaper so give access to memory reserves
4228 * only to the thread with TIF_MEMDIE set
4230 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4237 * Distinguish requests which really need access to full memory
4238 * reserves from oom victims which can live with a portion of it
4240 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4242 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4244 if (gfp_mask
& __GFP_MEMALLOC
)
4245 return ALLOC_NO_WATERMARKS
;
4246 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4247 return ALLOC_NO_WATERMARKS
;
4248 if (!in_interrupt()) {
4249 if (current
->flags
& PF_MEMALLOC
)
4250 return ALLOC_NO_WATERMARKS
;
4251 else if (oom_reserves_allowed(current
))
4258 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4260 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4264 * Checks whether it makes sense to retry the reclaim to make a forward progress
4265 * for the given allocation request.
4267 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4268 * without success, or when we couldn't even meet the watermark if we
4269 * reclaimed all remaining pages on the LRU lists.
4271 * Returns true if a retry is viable or false to enter the oom path.
4274 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4275 struct alloc_context
*ac
, int alloc_flags
,
4276 bool did_some_progress
, int *no_progress_loops
)
4283 * Costly allocations might have made a progress but this doesn't mean
4284 * their order will become available due to high fragmentation so
4285 * always increment the no progress counter for them
4287 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4288 *no_progress_loops
= 0;
4290 (*no_progress_loops
)++;
4293 * Make sure we converge to OOM if we cannot make any progress
4294 * several times in the row.
4296 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4297 /* Before OOM, exhaust highatomic_reserve */
4298 return unreserve_highatomic_pageblock(ac
, true);
4302 * Keep reclaiming pages while there is a chance this will lead
4303 * somewhere. If none of the target zones can satisfy our allocation
4304 * request even if all reclaimable pages are considered then we are
4305 * screwed and have to go OOM.
4307 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4309 unsigned long available
;
4310 unsigned long reclaimable
;
4311 unsigned long min_wmark
= min_wmark_pages(zone
);
4314 available
= reclaimable
= zone_reclaimable_pages(zone
);
4315 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4318 * Would the allocation succeed if we reclaimed all
4319 * reclaimable pages?
4321 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4322 ac_classzone_idx(ac
), alloc_flags
, available
);
4323 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4324 available
, min_wmark
, *no_progress_loops
, wmark
);
4327 * If we didn't make any progress and have a lot of
4328 * dirty + writeback pages then we should wait for
4329 * an IO to complete to slow down the reclaim and
4330 * prevent from pre mature OOM
4332 if (!did_some_progress
) {
4333 unsigned long write_pending
;
4335 write_pending
= zone_page_state_snapshot(zone
,
4336 NR_ZONE_WRITE_PENDING
);
4338 if (2 * write_pending
> reclaimable
) {
4339 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4351 * Memory allocation/reclaim might be called from a WQ context and the
4352 * current implementation of the WQ concurrency control doesn't
4353 * recognize that a particular WQ is congested if the worker thread is
4354 * looping without ever sleeping. Therefore we have to do a short sleep
4355 * here rather than calling cond_resched().
4357 if (current
->flags
& PF_WQ_WORKER
)
4358 schedule_timeout_uninterruptible(1);
4365 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4368 * It's possible that cpuset's mems_allowed and the nodemask from
4369 * mempolicy don't intersect. This should be normally dealt with by
4370 * policy_nodemask(), but it's possible to race with cpuset update in
4371 * such a way the check therein was true, and then it became false
4372 * before we got our cpuset_mems_cookie here.
4373 * This assumes that for all allocations, ac->nodemask can come only
4374 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4375 * when it does not intersect with the cpuset restrictions) or the
4376 * caller can deal with a violated nodemask.
4378 if (cpusets_enabled() && ac
->nodemask
&&
4379 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4380 ac
->nodemask
= NULL
;
4385 * When updating a task's mems_allowed or mempolicy nodemask, it is
4386 * possible to race with parallel threads in such a way that our
4387 * allocation can fail while the mask is being updated. If we are about
4388 * to fail, check if the cpuset changed during allocation and if so,
4391 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4397 static inline struct page
*
4398 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4399 struct alloc_context
*ac
)
4401 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4402 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4403 struct page
*page
= NULL
;
4404 unsigned int alloc_flags
;
4405 unsigned long did_some_progress
;
4406 enum compact_priority compact_priority
;
4407 enum compact_result compact_result
;
4408 int compaction_retries
;
4409 int no_progress_loops
;
4410 unsigned int cpuset_mems_cookie
;
4414 * We also sanity check to catch abuse of atomic reserves being used by
4415 * callers that are not in atomic context.
4417 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4418 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4419 gfp_mask
&= ~__GFP_ATOMIC
;
4422 compaction_retries
= 0;
4423 no_progress_loops
= 0;
4424 compact_priority
= DEF_COMPACT_PRIORITY
;
4425 cpuset_mems_cookie
= read_mems_allowed_begin();
4428 * The fast path uses conservative alloc_flags to succeed only until
4429 * kswapd needs to be woken up, and to avoid the cost of setting up
4430 * alloc_flags precisely. So we do that now.
4432 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4435 * We need to recalculate the starting point for the zonelist iterator
4436 * because we might have used different nodemask in the fast path, or
4437 * there was a cpuset modification and we are retrying - otherwise we
4438 * could end up iterating over non-eligible zones endlessly.
4440 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4441 ac
->high_zoneidx
, ac
->nodemask
);
4442 if (!ac
->preferred_zoneref
->zone
)
4445 if (alloc_flags
& ALLOC_KSWAPD
)
4446 wake_all_kswapds(order
, gfp_mask
, ac
);
4449 * The adjusted alloc_flags might result in immediate success, so try
4452 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4457 * For costly allocations, try direct compaction first, as it's likely
4458 * that we have enough base pages and don't need to reclaim. For non-
4459 * movable high-order allocations, do that as well, as compaction will
4460 * try prevent permanent fragmentation by migrating from blocks of the
4462 * Don't try this for allocations that are allowed to ignore
4463 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4465 if (can_direct_reclaim
&&
4467 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4468 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4469 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4471 INIT_COMPACT_PRIORITY
,
4476 if (order
>= pageblock_order
&& (gfp_mask
& __GFP_IO
)) {
4478 * If allocating entire pageblock(s) and compaction
4479 * failed because all zones are below low watermarks
4480 * or is prohibited because it recently failed at this
4481 * order, fail immediately.
4484 * - potentially very expensive because zones are far
4485 * below their low watermarks or this is part of very
4486 * bursty high order allocations,
4487 * - not guaranteed to help because isolate_freepages()
4488 * may not iterate over freed pages as part of its
4490 * - unlikely to make entire pageblocks free on its
4493 if (compact_result
== COMPACT_SKIPPED
||
4494 compact_result
== COMPACT_DEFERRED
)
4499 * Checks for costly allocations with __GFP_NORETRY, which
4500 * includes THP page fault allocations
4502 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4504 * If compaction is deferred for high-order allocations,
4505 * it is because sync compaction recently failed. If
4506 * this is the case and the caller requested a THP
4507 * allocation, we do not want to heavily disrupt the
4508 * system, so we fail the allocation instead of entering
4511 if (compact_result
== COMPACT_DEFERRED
)
4515 * Looks like reclaim/compaction is worth trying, but
4516 * sync compaction could be very expensive, so keep
4517 * using async compaction.
4519 compact_priority
= INIT_COMPACT_PRIORITY
;
4524 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4525 if (alloc_flags
& ALLOC_KSWAPD
)
4526 wake_all_kswapds(order
, gfp_mask
, ac
);
4528 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4530 alloc_flags
= reserve_flags
;
4533 * Reset the nodemask and zonelist iterators if memory policies can be
4534 * ignored. These allocations are high priority and system rather than
4537 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4538 ac
->nodemask
= NULL
;
4539 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4540 ac
->high_zoneidx
, ac
->nodemask
);
4543 /* Attempt with potentially adjusted zonelist and alloc_flags */
4544 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4548 /* Caller is not willing to reclaim, we can't balance anything */
4549 if (!can_direct_reclaim
)
4552 /* Avoid recursion of direct reclaim */
4553 if (current
->flags
& PF_MEMALLOC
)
4556 /* Try direct reclaim and then allocating */
4557 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4558 &did_some_progress
);
4562 /* Try direct compaction and then allocating */
4563 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4564 compact_priority
, &compact_result
);
4568 /* Do not loop if specifically requested */
4569 if (gfp_mask
& __GFP_NORETRY
)
4573 * Do not retry costly high order allocations unless they are
4574 * __GFP_RETRY_MAYFAIL
4576 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4579 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4580 did_some_progress
> 0, &no_progress_loops
))
4584 * It doesn't make any sense to retry for the compaction if the order-0
4585 * reclaim is not able to make any progress because the current
4586 * implementation of the compaction depends on the sufficient amount
4587 * of free memory (see __compaction_suitable)
4589 if (did_some_progress
> 0 &&
4590 should_compact_retry(ac
, order
, alloc_flags
,
4591 compact_result
, &compact_priority
,
4592 &compaction_retries
))
4596 /* Deal with possible cpuset update races before we start OOM killing */
4597 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4600 /* Reclaim has failed us, start killing things */
4601 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4605 /* Avoid allocations with no watermarks from looping endlessly */
4606 if (tsk_is_oom_victim(current
) &&
4607 (alloc_flags
== ALLOC_OOM
||
4608 (gfp_mask
& __GFP_NOMEMALLOC
)))
4611 /* Retry as long as the OOM killer is making progress */
4612 if (did_some_progress
) {
4613 no_progress_loops
= 0;
4618 /* Deal with possible cpuset update races before we fail */
4619 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4623 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4626 if (gfp_mask
& __GFP_NOFAIL
) {
4628 * All existing users of the __GFP_NOFAIL are blockable, so warn
4629 * of any new users that actually require GFP_NOWAIT
4631 if (WARN_ON_ONCE(!can_direct_reclaim
))
4635 * PF_MEMALLOC request from this context is rather bizarre
4636 * because we cannot reclaim anything and only can loop waiting
4637 * for somebody to do a work for us
4639 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4642 * non failing costly orders are a hard requirement which we
4643 * are not prepared for much so let's warn about these users
4644 * so that we can identify them and convert them to something
4647 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4650 * Help non-failing allocations by giving them access to memory
4651 * reserves but do not use ALLOC_NO_WATERMARKS because this
4652 * could deplete whole memory reserves which would just make
4653 * the situation worse
4655 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4663 warn_alloc(gfp_mask
, ac
->nodemask
,
4664 "page allocation failure: order:%u", order
);
4669 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4670 int preferred_nid
, nodemask_t
*nodemask
,
4671 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4672 unsigned int *alloc_flags
)
4674 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4675 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4676 ac
->nodemask
= nodemask
;
4677 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4679 if (cpusets_enabled()) {
4680 *alloc_mask
|= __GFP_HARDWALL
;
4682 ac
->nodemask
= &cpuset_current_mems_allowed
;
4684 *alloc_flags
|= ALLOC_CPUSET
;
4687 fs_reclaim_acquire(gfp_mask
);
4688 fs_reclaim_release(gfp_mask
);
4690 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4692 if (should_fail_alloc_page(gfp_mask
, order
))
4695 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4696 *alloc_flags
|= ALLOC_CMA
;
4701 /* Determine whether to spread dirty pages and what the first usable zone */
4702 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4704 /* Dirty zone balancing only done in the fast path */
4705 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4708 * The preferred zone is used for statistics but crucially it is
4709 * also used as the starting point for the zonelist iterator. It
4710 * may get reset for allocations that ignore memory policies.
4712 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4713 ac
->high_zoneidx
, ac
->nodemask
);
4717 * This is the 'heart' of the zoned buddy allocator.
4720 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4721 nodemask_t
*nodemask
)
4724 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4725 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4726 struct alloc_context ac
= { };
4729 * There are several places where we assume that the order value is sane
4730 * so bail out early if the request is out of bound.
4732 if (unlikely(order
>= MAX_ORDER
)) {
4733 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4737 gfp_mask
&= gfp_allowed_mask
;
4738 alloc_mask
= gfp_mask
;
4739 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4742 finalise_ac(gfp_mask
, &ac
);
4745 * Forbid the first pass from falling back to types that fragment
4746 * memory until all local zones are considered.
4748 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4750 /* First allocation attempt */
4751 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4756 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4757 * resp. GFP_NOIO which has to be inherited for all allocation requests
4758 * from a particular context which has been marked by
4759 * memalloc_no{fs,io}_{save,restore}.
4761 alloc_mask
= current_gfp_context(gfp_mask
);
4762 ac
.spread_dirty_pages
= false;
4765 * Restore the original nodemask if it was potentially replaced with
4766 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4768 if (unlikely(ac
.nodemask
!= nodemask
))
4769 ac
.nodemask
= nodemask
;
4771 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4774 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4775 unlikely(__memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4776 __free_pages(page
, order
);
4780 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4784 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4787 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4788 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4789 * you need to access high mem.
4791 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4795 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4798 return (unsigned long) page_address(page
);
4800 EXPORT_SYMBOL(__get_free_pages
);
4802 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4804 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4806 EXPORT_SYMBOL(get_zeroed_page
);
4808 static inline void free_the_page(struct page
*page
, unsigned int order
)
4810 if (order
== 0) /* Via pcp? */
4811 free_unref_page(page
);
4813 __free_pages_ok(page
, order
);
4816 void __free_pages(struct page
*page
, unsigned int order
)
4818 if (put_page_testzero(page
))
4819 free_the_page(page
, order
);
4821 EXPORT_SYMBOL(__free_pages
);
4823 void free_pages(unsigned long addr
, unsigned int order
)
4826 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4827 __free_pages(virt_to_page((void *)addr
), order
);
4831 EXPORT_SYMBOL(free_pages
);
4835 * An arbitrary-length arbitrary-offset area of memory which resides
4836 * within a 0 or higher order page. Multiple fragments within that page
4837 * are individually refcounted, in the page's reference counter.
4839 * The page_frag functions below provide a simple allocation framework for
4840 * page fragments. This is used by the network stack and network device
4841 * drivers to provide a backing region of memory for use as either an
4842 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4844 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4847 struct page
*page
= NULL
;
4848 gfp_t gfp
= gfp_mask
;
4850 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4851 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4853 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4854 PAGE_FRAG_CACHE_MAX_ORDER
);
4855 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4857 if (unlikely(!page
))
4858 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4860 nc
->va
= page
? page_address(page
) : NULL
;
4865 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4867 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4869 if (page_ref_sub_and_test(page
, count
))
4870 free_the_page(page
, compound_order(page
));
4872 EXPORT_SYMBOL(__page_frag_cache_drain
);
4874 void *page_frag_alloc(struct page_frag_cache
*nc
,
4875 unsigned int fragsz
, gfp_t gfp_mask
)
4877 unsigned int size
= PAGE_SIZE
;
4881 if (unlikely(!nc
->va
)) {
4883 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4887 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4888 /* if size can vary use size else just use PAGE_SIZE */
4891 /* Even if we own the page, we do not use atomic_set().
4892 * This would break get_page_unless_zero() users.
4894 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
4896 /* reset page count bias and offset to start of new frag */
4897 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4898 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4902 offset
= nc
->offset
- fragsz
;
4903 if (unlikely(offset
< 0)) {
4904 page
= virt_to_page(nc
->va
);
4906 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4909 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4910 /* if size can vary use size else just use PAGE_SIZE */
4913 /* OK, page count is 0, we can safely set it */
4914 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
4916 /* reset page count bias and offset to start of new frag */
4917 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4918 offset
= size
- fragsz
;
4922 nc
->offset
= offset
;
4924 return nc
->va
+ offset
;
4926 EXPORT_SYMBOL(page_frag_alloc
);
4929 * Frees a page fragment allocated out of either a compound or order 0 page.
4931 void page_frag_free(void *addr
)
4933 struct page
*page
= virt_to_head_page(addr
);
4935 if (unlikely(put_page_testzero(page
)))
4936 free_the_page(page
, compound_order(page
));
4938 EXPORT_SYMBOL(page_frag_free
);
4940 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4944 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4945 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4947 split_page(virt_to_page((void *)addr
), order
);
4948 while (used
< alloc_end
) {
4953 return (void *)addr
;
4957 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4958 * @size: the number of bytes to allocate
4959 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4961 * This function is similar to alloc_pages(), except that it allocates the
4962 * minimum number of pages to satisfy the request. alloc_pages() can only
4963 * allocate memory in power-of-two pages.
4965 * This function is also limited by MAX_ORDER.
4967 * Memory allocated by this function must be released by free_pages_exact().
4969 * Return: pointer to the allocated area or %NULL in case of error.
4971 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4973 unsigned int order
= get_order(size
);
4976 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
4977 gfp_mask
&= ~__GFP_COMP
;
4979 addr
= __get_free_pages(gfp_mask
, order
);
4980 return make_alloc_exact(addr
, order
, size
);
4982 EXPORT_SYMBOL(alloc_pages_exact
);
4985 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4987 * @nid: the preferred node ID where memory should be allocated
4988 * @size: the number of bytes to allocate
4989 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4991 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4994 * Return: pointer to the allocated area or %NULL in case of error.
4996 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4998 unsigned int order
= get_order(size
);
5001 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5002 gfp_mask
&= ~__GFP_COMP
;
5004 p
= alloc_pages_node(nid
, gfp_mask
, order
);
5007 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5011 * free_pages_exact - release memory allocated via alloc_pages_exact()
5012 * @virt: the value returned by alloc_pages_exact.
5013 * @size: size of allocation, same value as passed to alloc_pages_exact().
5015 * Release the memory allocated by a previous call to alloc_pages_exact.
5017 void free_pages_exact(void *virt
, size_t size
)
5019 unsigned long addr
= (unsigned long)virt
;
5020 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5022 while (addr
< end
) {
5027 EXPORT_SYMBOL(free_pages_exact
);
5030 * nr_free_zone_pages - count number of pages beyond high watermark
5031 * @offset: The zone index of the highest zone
5033 * nr_free_zone_pages() counts the number of pages which are beyond the
5034 * high watermark within all zones at or below a given zone index. For each
5035 * zone, the number of pages is calculated as:
5037 * nr_free_zone_pages = managed_pages - high_pages
5039 * Return: number of pages beyond high watermark.
5041 static unsigned long nr_free_zone_pages(int offset
)
5046 /* Just pick one node, since fallback list is circular */
5047 unsigned long sum
= 0;
5049 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5051 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5052 unsigned long size
= zone_managed_pages(zone
);
5053 unsigned long high
= high_wmark_pages(zone
);
5062 * nr_free_buffer_pages - count number of pages beyond high watermark
5064 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5065 * watermark within ZONE_DMA and ZONE_NORMAL.
5067 * Return: number of pages beyond high watermark within ZONE_DMA and
5070 unsigned long nr_free_buffer_pages(void)
5072 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5074 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5077 * nr_free_pagecache_pages - count number of pages beyond high watermark
5079 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5080 * high watermark within all zones.
5082 * Return: number of pages beyond high watermark within all zones.
5084 unsigned long nr_free_pagecache_pages(void)
5086 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5089 static inline void show_node(struct zone
*zone
)
5091 if (IS_ENABLED(CONFIG_NUMA
))
5092 printk("Node %d ", zone_to_nid(zone
));
5095 long si_mem_available(void)
5098 unsigned long pagecache
;
5099 unsigned long wmark_low
= 0;
5100 unsigned long pages
[NR_LRU_LISTS
];
5101 unsigned long reclaimable
;
5105 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5106 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5109 wmark_low
+= low_wmark_pages(zone
);
5112 * Estimate the amount of memory available for userspace allocations,
5113 * without causing swapping.
5115 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5118 * Not all the page cache can be freed, otherwise the system will
5119 * start swapping. Assume at least half of the page cache, or the
5120 * low watermark worth of cache, needs to stay.
5122 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5123 pagecache
-= min(pagecache
/ 2, wmark_low
);
5124 available
+= pagecache
;
5127 * Part of the reclaimable slab and other kernel memory consists of
5128 * items that are in use, and cannot be freed. Cap this estimate at the
5131 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
5132 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5133 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5139 EXPORT_SYMBOL_GPL(si_mem_available
);
5141 void si_meminfo(struct sysinfo
*val
)
5143 val
->totalram
= totalram_pages();
5144 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5145 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5146 val
->bufferram
= nr_blockdev_pages();
5147 val
->totalhigh
= totalhigh_pages();
5148 val
->freehigh
= nr_free_highpages();
5149 val
->mem_unit
= PAGE_SIZE
;
5152 EXPORT_SYMBOL(si_meminfo
);
5155 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5157 int zone_type
; /* needs to be signed */
5158 unsigned long managed_pages
= 0;
5159 unsigned long managed_highpages
= 0;
5160 unsigned long free_highpages
= 0;
5161 pg_data_t
*pgdat
= NODE_DATA(nid
);
5163 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5164 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5165 val
->totalram
= managed_pages
;
5166 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5167 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5168 #ifdef CONFIG_HIGHMEM
5169 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5170 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5172 if (is_highmem(zone
)) {
5173 managed_highpages
+= zone_managed_pages(zone
);
5174 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5177 val
->totalhigh
= managed_highpages
;
5178 val
->freehigh
= free_highpages
;
5180 val
->totalhigh
= managed_highpages
;
5181 val
->freehigh
= free_highpages
;
5183 val
->mem_unit
= PAGE_SIZE
;
5188 * Determine whether the node should be displayed or not, depending on whether
5189 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5191 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5193 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5197 * no node mask - aka implicit memory numa policy. Do not bother with
5198 * the synchronization - read_mems_allowed_begin - because we do not
5199 * have to be precise here.
5202 nodemask
= &cpuset_current_mems_allowed
;
5204 return !node_isset(nid
, *nodemask
);
5207 #define K(x) ((x) << (PAGE_SHIFT-10))
5209 static void show_migration_types(unsigned char type
)
5211 static const char types
[MIGRATE_TYPES
] = {
5212 [MIGRATE_UNMOVABLE
] = 'U',
5213 [MIGRATE_MOVABLE
] = 'M',
5214 [MIGRATE_RECLAIMABLE
] = 'E',
5215 [MIGRATE_HIGHATOMIC
] = 'H',
5217 [MIGRATE_CMA
] = 'C',
5219 #ifdef CONFIG_MEMORY_ISOLATION
5220 [MIGRATE_ISOLATE
] = 'I',
5223 char tmp
[MIGRATE_TYPES
+ 1];
5227 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5228 if (type
& (1 << i
))
5233 printk(KERN_CONT
"(%s) ", tmp
);
5237 * Show free area list (used inside shift_scroll-lock stuff)
5238 * We also calculate the percentage fragmentation. We do this by counting the
5239 * memory on each free list with the exception of the first item on the list.
5242 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5245 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5247 unsigned long free_pcp
= 0;
5252 for_each_populated_zone(zone
) {
5253 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5256 for_each_online_cpu(cpu
)
5257 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5260 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5261 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5262 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5263 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5264 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5265 " free:%lu free_pcp:%lu free_cma:%lu\n",
5266 global_node_page_state(NR_ACTIVE_ANON
),
5267 global_node_page_state(NR_INACTIVE_ANON
),
5268 global_node_page_state(NR_ISOLATED_ANON
),
5269 global_node_page_state(NR_ACTIVE_FILE
),
5270 global_node_page_state(NR_INACTIVE_FILE
),
5271 global_node_page_state(NR_ISOLATED_FILE
),
5272 global_node_page_state(NR_UNEVICTABLE
),
5273 global_node_page_state(NR_FILE_DIRTY
),
5274 global_node_page_state(NR_WRITEBACK
),
5275 global_node_page_state(NR_UNSTABLE_NFS
),
5276 global_node_page_state(NR_SLAB_RECLAIMABLE
),
5277 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
5278 global_node_page_state(NR_FILE_MAPPED
),
5279 global_node_page_state(NR_SHMEM
),
5280 global_zone_page_state(NR_PAGETABLE
),
5281 global_zone_page_state(NR_BOUNCE
),
5282 global_zone_page_state(NR_FREE_PAGES
),
5284 global_zone_page_state(NR_FREE_CMA_PAGES
));
5286 for_each_online_pgdat(pgdat
) {
5287 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5291 " active_anon:%lukB"
5292 " inactive_anon:%lukB"
5293 " active_file:%lukB"
5294 " inactive_file:%lukB"
5295 " unevictable:%lukB"
5296 " isolated(anon):%lukB"
5297 " isolated(file):%lukB"
5302 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5304 " shmem_pmdmapped: %lukB"
5307 " writeback_tmp:%lukB"
5309 " all_unreclaimable? %s"
5312 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5313 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5314 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5315 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5316 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5317 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5318 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5319 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5320 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5321 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5322 K(node_page_state(pgdat
, NR_SHMEM
)),
5323 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5324 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5325 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5327 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5329 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5330 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
5331 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5335 for_each_populated_zone(zone
) {
5338 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5342 for_each_online_cpu(cpu
)
5343 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5352 " active_anon:%lukB"
5353 " inactive_anon:%lukB"
5354 " active_file:%lukB"
5355 " inactive_file:%lukB"
5356 " unevictable:%lukB"
5357 " writepending:%lukB"
5361 " kernel_stack:%lukB"
5369 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5370 K(min_wmark_pages(zone
)),
5371 K(low_wmark_pages(zone
)),
5372 K(high_wmark_pages(zone
)),
5373 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5374 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5375 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5376 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5377 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5378 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5379 K(zone
->present_pages
),
5380 K(zone_managed_pages(zone
)),
5381 K(zone_page_state(zone
, NR_MLOCK
)),
5382 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5383 K(zone_page_state(zone
, NR_PAGETABLE
)),
5384 K(zone_page_state(zone
, NR_BOUNCE
)),
5386 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5387 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5388 printk("lowmem_reserve[]:");
5389 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5390 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5391 printk(KERN_CONT
"\n");
5394 for_each_populated_zone(zone
) {
5396 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5397 unsigned char types
[MAX_ORDER
];
5399 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5402 printk(KERN_CONT
"%s: ", zone
->name
);
5404 spin_lock_irqsave(&zone
->lock
, flags
);
5405 for (order
= 0; order
< MAX_ORDER
; order
++) {
5406 struct free_area
*area
= &zone
->free_area
[order
];
5409 nr
[order
] = area
->nr_free
;
5410 total
+= nr
[order
] << order
;
5413 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5414 if (!free_area_empty(area
, type
))
5415 types
[order
] |= 1 << type
;
5418 spin_unlock_irqrestore(&zone
->lock
, flags
);
5419 for (order
= 0; order
< MAX_ORDER
; order
++) {
5420 printk(KERN_CONT
"%lu*%lukB ",
5421 nr
[order
], K(1UL) << order
);
5423 show_migration_types(types
[order
]);
5425 printk(KERN_CONT
"= %lukB\n", K(total
));
5428 hugetlb_show_meminfo();
5430 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5432 show_swap_cache_info();
5435 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5437 zoneref
->zone
= zone
;
5438 zoneref
->zone_idx
= zone_idx(zone
);
5442 * Builds allocation fallback zone lists.
5444 * Add all populated zones of a node to the zonelist.
5446 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5449 enum zone_type zone_type
= MAX_NR_ZONES
;
5454 zone
= pgdat
->node_zones
+ zone_type
;
5455 if (managed_zone(zone
)) {
5456 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5457 check_highest_zone(zone_type
);
5459 } while (zone_type
);
5466 static int __parse_numa_zonelist_order(char *s
)
5469 * We used to support different zonlists modes but they turned
5470 * out to be just not useful. Let's keep the warning in place
5471 * if somebody still use the cmd line parameter so that we do
5472 * not fail it silently
5474 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5475 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5481 static __init
int setup_numa_zonelist_order(char *s
)
5486 return __parse_numa_zonelist_order(s
);
5488 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
5490 char numa_zonelist_order
[] = "Node";
5493 * sysctl handler for numa_zonelist_order
5495 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5496 void __user
*buffer
, size_t *length
,
5503 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5504 str
= memdup_user_nul(buffer
, 16);
5506 return PTR_ERR(str
);
5508 ret
= __parse_numa_zonelist_order(str
);
5514 #define MAX_NODE_LOAD (nr_online_nodes)
5515 static int node_load
[MAX_NUMNODES
];
5518 * find_next_best_node - find the next node that should appear in a given node's fallback list
5519 * @node: node whose fallback list we're appending
5520 * @used_node_mask: nodemask_t of already used nodes
5522 * We use a number of factors to determine which is the next node that should
5523 * appear on a given node's fallback list. The node should not have appeared
5524 * already in @node's fallback list, and it should be the next closest node
5525 * according to the distance array (which contains arbitrary distance values
5526 * from each node to each node in the system), and should also prefer nodes
5527 * with no CPUs, since presumably they'll have very little allocation pressure
5528 * on them otherwise.
5530 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5532 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5535 int min_val
= INT_MAX
;
5536 int best_node
= NUMA_NO_NODE
;
5537 const struct cpumask
*tmp
= cpumask_of_node(0);
5539 /* Use the local node if we haven't already */
5540 if (!node_isset(node
, *used_node_mask
)) {
5541 node_set(node
, *used_node_mask
);
5545 for_each_node_state(n
, N_MEMORY
) {
5547 /* Don't want a node to appear more than once */
5548 if (node_isset(n
, *used_node_mask
))
5551 /* Use the distance array to find the distance */
5552 val
= node_distance(node
, n
);
5554 /* Penalize nodes under us ("prefer the next node") */
5557 /* Give preference to headless and unused nodes */
5558 tmp
= cpumask_of_node(n
);
5559 if (!cpumask_empty(tmp
))
5560 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5562 /* Slight preference for less loaded node */
5563 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5564 val
+= node_load
[n
];
5566 if (val
< min_val
) {
5573 node_set(best_node
, *used_node_mask
);
5580 * Build zonelists ordered by node and zones within node.
5581 * This results in maximum locality--normal zone overflows into local
5582 * DMA zone, if any--but risks exhausting DMA zone.
5584 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5587 struct zoneref
*zonerefs
;
5590 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5592 for (i
= 0; i
< nr_nodes
; i
++) {
5595 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5597 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5598 zonerefs
+= nr_zones
;
5600 zonerefs
->zone
= NULL
;
5601 zonerefs
->zone_idx
= 0;
5605 * Build gfp_thisnode zonelists
5607 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5609 struct zoneref
*zonerefs
;
5612 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5613 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5614 zonerefs
+= nr_zones
;
5615 zonerefs
->zone
= NULL
;
5616 zonerefs
->zone_idx
= 0;
5620 * Build zonelists ordered by zone and nodes within zones.
5621 * This results in conserving DMA zone[s] until all Normal memory is
5622 * exhausted, but results in overflowing to remote node while memory
5623 * may still exist in local DMA zone.
5626 static void build_zonelists(pg_data_t
*pgdat
)
5628 static int node_order
[MAX_NUMNODES
];
5629 int node
, load
, nr_nodes
= 0;
5630 nodemask_t used_mask
;
5631 int local_node
, prev_node
;
5633 /* NUMA-aware ordering of nodes */
5634 local_node
= pgdat
->node_id
;
5635 load
= nr_online_nodes
;
5636 prev_node
= local_node
;
5637 nodes_clear(used_mask
);
5639 memset(node_order
, 0, sizeof(node_order
));
5640 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5642 * We don't want to pressure a particular node.
5643 * So adding penalty to the first node in same
5644 * distance group to make it round-robin.
5646 if (node_distance(local_node
, node
) !=
5647 node_distance(local_node
, prev_node
))
5648 node_load
[node
] = load
;
5650 node_order
[nr_nodes
++] = node
;
5655 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5656 build_thisnode_zonelists(pgdat
);
5659 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5661 * Return node id of node used for "local" allocations.
5662 * I.e., first node id of first zone in arg node's generic zonelist.
5663 * Used for initializing percpu 'numa_mem', which is used primarily
5664 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5666 int local_memory_node(int node
)
5670 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5671 gfp_zone(GFP_KERNEL
),
5673 return zone_to_nid(z
->zone
);
5677 static void setup_min_unmapped_ratio(void);
5678 static void setup_min_slab_ratio(void);
5679 #else /* CONFIG_NUMA */
5681 static void build_zonelists(pg_data_t
*pgdat
)
5683 int node
, local_node
;
5684 struct zoneref
*zonerefs
;
5687 local_node
= pgdat
->node_id
;
5689 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5690 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5691 zonerefs
+= nr_zones
;
5694 * Now we build the zonelist so that it contains the zones
5695 * of all the other nodes.
5696 * We don't want to pressure a particular node, so when
5697 * building the zones for node N, we make sure that the
5698 * zones coming right after the local ones are those from
5699 * node N+1 (modulo N)
5701 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5702 if (!node_online(node
))
5704 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5705 zonerefs
+= nr_zones
;
5707 for (node
= 0; node
< local_node
; node
++) {
5708 if (!node_online(node
))
5710 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5711 zonerefs
+= nr_zones
;
5714 zonerefs
->zone
= NULL
;
5715 zonerefs
->zone_idx
= 0;
5718 #endif /* CONFIG_NUMA */
5721 * Boot pageset table. One per cpu which is going to be used for all
5722 * zones and all nodes. The parameters will be set in such a way
5723 * that an item put on a list will immediately be handed over to
5724 * the buddy list. This is safe since pageset manipulation is done
5725 * with interrupts disabled.
5727 * The boot_pagesets must be kept even after bootup is complete for
5728 * unused processors and/or zones. They do play a role for bootstrapping
5729 * hotplugged processors.
5731 * zoneinfo_show() and maybe other functions do
5732 * not check if the processor is online before following the pageset pointer.
5733 * Other parts of the kernel may not check if the zone is available.
5735 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5736 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5737 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5739 static void __build_all_zonelists(void *data
)
5742 int __maybe_unused cpu
;
5743 pg_data_t
*self
= data
;
5744 static DEFINE_SPINLOCK(lock
);
5749 memset(node_load
, 0, sizeof(node_load
));
5753 * This node is hotadded and no memory is yet present. So just
5754 * building zonelists is fine - no need to touch other nodes.
5756 if (self
&& !node_online(self
->node_id
)) {
5757 build_zonelists(self
);
5759 for_each_online_node(nid
) {
5760 pg_data_t
*pgdat
= NODE_DATA(nid
);
5762 build_zonelists(pgdat
);
5765 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5767 * We now know the "local memory node" for each node--
5768 * i.e., the node of the first zone in the generic zonelist.
5769 * Set up numa_mem percpu variable for on-line cpus. During
5770 * boot, only the boot cpu should be on-line; we'll init the
5771 * secondary cpus' numa_mem as they come on-line. During
5772 * node/memory hotplug, we'll fixup all on-line cpus.
5774 for_each_online_cpu(cpu
)
5775 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5782 static noinline
void __init
5783 build_all_zonelists_init(void)
5787 __build_all_zonelists(NULL
);
5790 * Initialize the boot_pagesets that are going to be used
5791 * for bootstrapping processors. The real pagesets for
5792 * each zone will be allocated later when the per cpu
5793 * allocator is available.
5795 * boot_pagesets are used also for bootstrapping offline
5796 * cpus if the system is already booted because the pagesets
5797 * are needed to initialize allocators on a specific cpu too.
5798 * F.e. the percpu allocator needs the page allocator which
5799 * needs the percpu allocator in order to allocate its pagesets
5800 * (a chicken-egg dilemma).
5802 for_each_possible_cpu(cpu
)
5803 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5805 mminit_verify_zonelist();
5806 cpuset_init_current_mems_allowed();
5810 * unless system_state == SYSTEM_BOOTING.
5812 * __ref due to call of __init annotated helper build_all_zonelists_init
5813 * [protected by SYSTEM_BOOTING].
5815 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5817 if (system_state
== SYSTEM_BOOTING
) {
5818 build_all_zonelists_init();
5820 __build_all_zonelists(pgdat
);
5821 /* cpuset refresh routine should be here */
5823 vm_total_pages
= nr_free_pagecache_pages();
5825 * Disable grouping by mobility if the number of pages in the
5826 * system is too low to allow the mechanism to work. It would be
5827 * more accurate, but expensive to check per-zone. This check is
5828 * made on memory-hotadd so a system can start with mobility
5829 * disabled and enable it later
5831 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5832 page_group_by_mobility_disabled
= 1;
5834 page_group_by_mobility_disabled
= 0;
5836 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5838 page_group_by_mobility_disabled
? "off" : "on",
5841 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5845 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5846 static bool __meminit
5847 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5849 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5850 static struct memblock_region
*r
;
5852 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5853 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5854 for_each_memblock(memory
, r
) {
5855 if (*pfn
< memblock_region_memory_end_pfn(r
))
5859 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5860 memblock_is_mirror(r
)) {
5861 *pfn
= memblock_region_memory_end_pfn(r
);
5870 * Initially all pages are reserved - free ones are freed
5871 * up by memblock_free_all() once the early boot process is
5872 * done. Non-atomic initialization, single-pass.
5874 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5875 unsigned long start_pfn
, enum memmap_context context
,
5876 struct vmem_altmap
*altmap
)
5878 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5881 if (highest_memmap_pfn
< end_pfn
- 1)
5882 highest_memmap_pfn
= end_pfn
- 1;
5884 #ifdef CONFIG_ZONE_DEVICE
5886 * Honor reservation requested by the driver for this ZONE_DEVICE
5887 * memory. We limit the total number of pages to initialize to just
5888 * those that might contain the memory mapping. We will defer the
5889 * ZONE_DEVICE page initialization until after we have released
5892 if (zone
== ZONE_DEVICE
) {
5896 if (start_pfn
== altmap
->base_pfn
)
5897 start_pfn
+= altmap
->reserve
;
5898 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5902 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5904 * There can be holes in boot-time mem_map[]s handed to this
5905 * function. They do not exist on hotplugged memory.
5907 if (context
== MEMMAP_EARLY
) {
5908 if (!early_pfn_valid(pfn
))
5910 if (!early_pfn_in_nid(pfn
, nid
))
5912 if (overlap_memmap_init(zone
, &pfn
))
5914 if (defer_init(nid
, pfn
, end_pfn
))
5918 page
= pfn_to_page(pfn
);
5919 __init_single_page(page
, pfn
, zone
, nid
);
5920 if (context
== MEMMAP_HOTPLUG
)
5921 __SetPageReserved(page
);
5924 * Mark the block movable so that blocks are reserved for
5925 * movable at startup. This will force kernel allocations
5926 * to reserve their blocks rather than leaking throughout
5927 * the address space during boot when many long-lived
5928 * kernel allocations are made.
5930 * bitmap is created for zone's valid pfn range. but memmap
5931 * can be created for invalid pages (for alignment)
5932 * check here not to call set_pageblock_migratetype() against
5935 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5936 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5942 #ifdef CONFIG_ZONE_DEVICE
5943 void __ref
memmap_init_zone_device(struct zone
*zone
,
5944 unsigned long start_pfn
,
5946 struct dev_pagemap
*pgmap
)
5948 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5949 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5950 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
5951 unsigned long zone_idx
= zone_idx(zone
);
5952 unsigned long start
= jiffies
;
5953 int nid
= pgdat
->node_id
;
5955 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
5959 * The call to memmap_init_zone should have already taken care
5960 * of the pages reserved for the memmap, so we can just jump to
5961 * the end of that region and start processing the device pages.
5964 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5965 size
= end_pfn
- start_pfn
;
5968 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5969 struct page
*page
= pfn_to_page(pfn
);
5971 __init_single_page(page
, pfn
, zone_idx
, nid
);
5974 * Mark page reserved as it will need to wait for onlining
5975 * phase for it to be fully associated with a zone.
5977 * We can use the non-atomic __set_bit operation for setting
5978 * the flag as we are still initializing the pages.
5980 __SetPageReserved(page
);
5983 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5984 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5985 * ever freed or placed on a driver-private list.
5987 page
->pgmap
= pgmap
;
5988 page
->zone_device_data
= NULL
;
5991 * Mark the block movable so that blocks are reserved for
5992 * movable at startup. This will force kernel allocations
5993 * to reserve their blocks rather than leaking throughout
5994 * the address space during boot when many long-lived
5995 * kernel allocations are made.
5997 * bitmap is created for zone's valid pfn range. but memmap
5998 * can be created for invalid pages (for alignment)
5999 * check here not to call set_pageblock_migratetype() against
6002 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6003 * because this is done early in section_activate()
6005 if (!(pfn
& (pageblock_nr_pages
- 1))) {
6006 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6011 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6012 size
, jiffies_to_msecs(jiffies
- start
));
6016 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6018 unsigned int order
, t
;
6019 for_each_migratetype_order(order
, t
) {
6020 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6021 zone
->free_area
[order
].nr_free
= 0;
6025 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
6026 unsigned long zone
, unsigned long start_pfn
)
6028 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
, NULL
);
6031 static int zone_batchsize(struct zone
*zone
)
6037 * The per-cpu-pages pools are set to around 1000th of the
6040 batch
= zone_managed_pages(zone
) / 1024;
6041 /* But no more than a meg. */
6042 if (batch
* PAGE_SIZE
> 1024 * 1024)
6043 batch
= (1024 * 1024) / PAGE_SIZE
;
6044 batch
/= 4; /* We effectively *= 4 below */
6049 * Clamp the batch to a 2^n - 1 value. Having a power
6050 * of 2 value was found to be more likely to have
6051 * suboptimal cache aliasing properties in some cases.
6053 * For example if 2 tasks are alternately allocating
6054 * batches of pages, one task can end up with a lot
6055 * of pages of one half of the possible page colors
6056 * and the other with pages of the other colors.
6058 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6063 /* The deferral and batching of frees should be suppressed under NOMMU
6066 * The problem is that NOMMU needs to be able to allocate large chunks
6067 * of contiguous memory as there's no hardware page translation to
6068 * assemble apparent contiguous memory from discontiguous pages.
6070 * Queueing large contiguous runs of pages for batching, however,
6071 * causes the pages to actually be freed in smaller chunks. As there
6072 * can be a significant delay between the individual batches being
6073 * recycled, this leads to the once large chunks of space being
6074 * fragmented and becoming unavailable for high-order allocations.
6081 * pcp->high and pcp->batch values are related and dependent on one another:
6082 * ->batch must never be higher then ->high.
6083 * The following function updates them in a safe manner without read side
6086 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6087 * those fields changing asynchronously (acording the the above rule).
6089 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6090 * outside of boot time (or some other assurance that no concurrent updaters
6093 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6094 unsigned long batch
)
6096 /* start with a fail safe value for batch */
6100 /* Update high, then batch, in order */
6107 /* a companion to pageset_set_high() */
6108 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
6110 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
6113 static void pageset_init(struct per_cpu_pageset
*p
)
6115 struct per_cpu_pages
*pcp
;
6118 memset(p
, 0, sizeof(*p
));
6121 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
6122 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
6125 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
6128 pageset_set_batch(p
, batch
);
6132 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6133 * to the value high for the pageset p.
6135 static void pageset_set_high(struct per_cpu_pageset
*p
,
6138 unsigned long batch
= max(1UL, high
/ 4);
6139 if ((high
/ 4) > (PAGE_SHIFT
* 8))
6140 batch
= PAGE_SHIFT
* 8;
6142 pageset_update(&p
->pcp
, high
, batch
);
6145 static void pageset_set_high_and_batch(struct zone
*zone
,
6146 struct per_cpu_pageset
*pcp
)
6148 if (percpu_pagelist_fraction
)
6149 pageset_set_high(pcp
,
6150 (zone_managed_pages(zone
) /
6151 percpu_pagelist_fraction
));
6153 pageset_set_batch(pcp
, zone_batchsize(zone
));
6156 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
6158 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
6161 pageset_set_high_and_batch(zone
, pcp
);
6164 void __meminit
setup_zone_pageset(struct zone
*zone
)
6167 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
6168 for_each_possible_cpu(cpu
)
6169 zone_pageset_init(zone
, cpu
);
6173 * Allocate per cpu pagesets and initialize them.
6174 * Before this call only boot pagesets were available.
6176 void __init
setup_per_cpu_pageset(void)
6178 struct pglist_data
*pgdat
;
6181 for_each_populated_zone(zone
)
6182 setup_zone_pageset(zone
);
6184 for_each_online_pgdat(pgdat
)
6185 pgdat
->per_cpu_nodestats
=
6186 alloc_percpu(struct per_cpu_nodestat
);
6189 static __meminit
void zone_pcp_init(struct zone
*zone
)
6192 * per cpu subsystem is not up at this point. The following code
6193 * relies on the ability of the linker to provide the
6194 * offset of a (static) per cpu variable into the per cpu area.
6196 zone
->pageset
= &boot_pageset
;
6198 if (populated_zone(zone
))
6199 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
6200 zone
->name
, zone
->present_pages
,
6201 zone_batchsize(zone
));
6204 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6205 unsigned long zone_start_pfn
,
6208 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6209 int zone_idx
= zone_idx(zone
) + 1;
6211 if (zone_idx
> pgdat
->nr_zones
)
6212 pgdat
->nr_zones
= zone_idx
;
6214 zone
->zone_start_pfn
= zone_start_pfn
;
6216 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6217 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6219 (unsigned long)zone_idx(zone
),
6220 zone_start_pfn
, (zone_start_pfn
+ size
));
6222 zone_init_free_lists(zone
);
6223 zone
->initialized
= 1;
6226 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6227 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6230 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6232 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
6233 struct mminit_pfnnid_cache
*state
)
6235 unsigned long start_pfn
, end_pfn
;
6238 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
6239 return state
->last_nid
;
6241 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
6242 if (nid
!= NUMA_NO_NODE
) {
6243 state
->last_start
= start_pfn
;
6244 state
->last_end
= end_pfn
;
6245 state
->last_nid
= nid
;
6250 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6253 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6254 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6255 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6257 * If an architecture guarantees that all ranges registered contain no holes
6258 * and may be freed, this this function may be used instead of calling
6259 * memblock_free_early_nid() manually.
6261 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
6263 unsigned long start_pfn
, end_pfn
;
6266 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
6267 start_pfn
= min(start_pfn
, max_low_pfn
);
6268 end_pfn
= min(end_pfn
, max_low_pfn
);
6270 if (start_pfn
< end_pfn
)
6271 memblock_free_early_nid(PFN_PHYS(start_pfn
),
6272 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
6278 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6279 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6281 * If an architecture guarantees that all ranges registered contain no holes and may
6282 * be freed, this function may be used instead of calling memory_present() manually.
6284 void __init
sparse_memory_present_with_active_regions(int nid
)
6286 unsigned long start_pfn
, end_pfn
;
6289 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
6290 memory_present(this_nid
, start_pfn
, end_pfn
);
6294 * get_pfn_range_for_nid - Return the start and end page frames for a node
6295 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6296 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6297 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6299 * It returns the start and end page frame of a node based on information
6300 * provided by memblock_set_node(). If called for a node
6301 * with no available memory, a warning is printed and the start and end
6304 void __init
get_pfn_range_for_nid(unsigned int nid
,
6305 unsigned long *start_pfn
, unsigned long *end_pfn
)
6307 unsigned long this_start_pfn
, this_end_pfn
;
6313 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6314 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6315 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6318 if (*start_pfn
== -1UL)
6323 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6324 * assumption is made that zones within a node are ordered in monotonic
6325 * increasing memory addresses so that the "highest" populated zone is used
6327 static void __init
find_usable_zone_for_movable(void)
6330 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6331 if (zone_index
== ZONE_MOVABLE
)
6334 if (arch_zone_highest_possible_pfn
[zone_index
] >
6335 arch_zone_lowest_possible_pfn
[zone_index
])
6339 VM_BUG_ON(zone_index
== -1);
6340 movable_zone
= zone_index
;
6344 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6345 * because it is sized independent of architecture. Unlike the other zones,
6346 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6347 * in each node depending on the size of each node and how evenly kernelcore
6348 * is distributed. This helper function adjusts the zone ranges
6349 * provided by the architecture for a given node by using the end of the
6350 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6351 * zones within a node are in order of monotonic increases memory addresses
6353 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6354 unsigned long zone_type
,
6355 unsigned long node_start_pfn
,
6356 unsigned long node_end_pfn
,
6357 unsigned long *zone_start_pfn
,
6358 unsigned long *zone_end_pfn
)
6360 /* Only adjust if ZONE_MOVABLE is on this node */
6361 if (zone_movable_pfn
[nid
]) {
6362 /* Size ZONE_MOVABLE */
6363 if (zone_type
== ZONE_MOVABLE
) {
6364 *zone_start_pfn
= zone_movable_pfn
[nid
];
6365 *zone_end_pfn
= min(node_end_pfn
,
6366 arch_zone_highest_possible_pfn
[movable_zone
]);
6368 /* Adjust for ZONE_MOVABLE starting within this range */
6369 } else if (!mirrored_kernelcore
&&
6370 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6371 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6372 *zone_end_pfn
= zone_movable_pfn
[nid
];
6374 /* Check if this whole range is within ZONE_MOVABLE */
6375 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6376 *zone_start_pfn
= *zone_end_pfn
;
6381 * Return the number of pages a zone spans in a node, including holes
6382 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6384 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6385 unsigned long zone_type
,
6386 unsigned long node_start_pfn
,
6387 unsigned long node_end_pfn
,
6388 unsigned long *zone_start_pfn
,
6389 unsigned long *zone_end_pfn
,
6390 unsigned long *ignored
)
6392 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6393 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6394 /* When hotadd a new node from cpu_up(), the node should be empty */
6395 if (!node_start_pfn
&& !node_end_pfn
)
6398 /* Get the start and end of the zone */
6399 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6400 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6401 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6402 node_start_pfn
, node_end_pfn
,
6403 zone_start_pfn
, zone_end_pfn
);
6405 /* Check that this node has pages within the zone's required range */
6406 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6409 /* Move the zone boundaries inside the node if necessary */
6410 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6411 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6413 /* Return the spanned pages */
6414 return *zone_end_pfn
- *zone_start_pfn
;
6418 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6419 * then all holes in the requested range will be accounted for.
6421 unsigned long __init
__absent_pages_in_range(int nid
,
6422 unsigned long range_start_pfn
,
6423 unsigned long range_end_pfn
)
6425 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6426 unsigned long start_pfn
, end_pfn
;
6429 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6430 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6431 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6432 nr_absent
-= end_pfn
- start_pfn
;
6438 * absent_pages_in_range - Return number of page frames in holes within a range
6439 * @start_pfn: The start PFN to start searching for holes
6440 * @end_pfn: The end PFN to stop searching for holes
6442 * Return: the number of pages frames in memory holes within a range.
6444 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6445 unsigned long end_pfn
)
6447 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6450 /* Return the number of page frames in holes in a zone on a node */
6451 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6452 unsigned long zone_type
,
6453 unsigned long node_start_pfn
,
6454 unsigned long node_end_pfn
,
6455 unsigned long *ignored
)
6457 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6458 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6459 unsigned long zone_start_pfn
, zone_end_pfn
;
6460 unsigned long nr_absent
;
6462 /* When hotadd a new node from cpu_up(), the node should be empty */
6463 if (!node_start_pfn
&& !node_end_pfn
)
6466 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6467 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6469 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6470 node_start_pfn
, node_end_pfn
,
6471 &zone_start_pfn
, &zone_end_pfn
);
6472 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6475 * ZONE_MOVABLE handling.
6476 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6479 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6480 unsigned long start_pfn
, end_pfn
;
6481 struct memblock_region
*r
;
6483 for_each_memblock(memory
, r
) {
6484 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6485 zone_start_pfn
, zone_end_pfn
);
6486 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6487 zone_start_pfn
, zone_end_pfn
);
6489 if (zone_type
== ZONE_MOVABLE
&&
6490 memblock_is_mirror(r
))
6491 nr_absent
+= end_pfn
- start_pfn
;
6493 if (zone_type
== ZONE_NORMAL
&&
6494 !memblock_is_mirror(r
))
6495 nr_absent
+= end_pfn
- start_pfn
;
6502 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6503 static inline unsigned long __init
zone_spanned_pages_in_node(int nid
,
6504 unsigned long zone_type
,
6505 unsigned long node_start_pfn
,
6506 unsigned long node_end_pfn
,
6507 unsigned long *zone_start_pfn
,
6508 unsigned long *zone_end_pfn
,
6509 unsigned long *zones_size
)
6513 *zone_start_pfn
= node_start_pfn
;
6514 for (zone
= 0; zone
< zone_type
; zone
++)
6515 *zone_start_pfn
+= zones_size
[zone
];
6517 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
6519 return zones_size
[zone_type
];
6522 static inline unsigned long __init
zone_absent_pages_in_node(int nid
,
6523 unsigned long zone_type
,
6524 unsigned long node_start_pfn
,
6525 unsigned long node_end_pfn
,
6526 unsigned long *zholes_size
)
6531 return zholes_size
[zone_type
];
6534 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6536 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6537 unsigned long node_start_pfn
,
6538 unsigned long node_end_pfn
,
6539 unsigned long *zones_size
,
6540 unsigned long *zholes_size
)
6542 unsigned long realtotalpages
= 0, totalpages
= 0;
6545 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6546 struct zone
*zone
= pgdat
->node_zones
+ i
;
6547 unsigned long zone_start_pfn
, zone_end_pfn
;
6548 unsigned long size
, real_size
;
6550 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6556 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
6557 node_start_pfn
, node_end_pfn
,
6560 zone
->zone_start_pfn
= zone_start_pfn
;
6562 zone
->zone_start_pfn
= 0;
6563 zone
->spanned_pages
= size
;
6564 zone
->present_pages
= real_size
;
6567 realtotalpages
+= real_size
;
6570 pgdat
->node_spanned_pages
= totalpages
;
6571 pgdat
->node_present_pages
= realtotalpages
;
6572 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6576 #ifndef CONFIG_SPARSEMEM
6578 * Calculate the size of the zone->blockflags rounded to an unsigned long
6579 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6580 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6581 * round what is now in bits to nearest long in bits, then return it in
6584 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6586 unsigned long usemapsize
;
6588 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6589 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6590 usemapsize
= usemapsize
>> pageblock_order
;
6591 usemapsize
*= NR_PAGEBLOCK_BITS
;
6592 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6594 return usemapsize
/ 8;
6597 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6599 unsigned long zone_start_pfn
,
6600 unsigned long zonesize
)
6602 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6603 zone
->pageblock_flags
= NULL
;
6605 zone
->pageblock_flags
=
6606 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
6608 if (!zone
->pageblock_flags
)
6609 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6610 usemapsize
, zone
->name
, pgdat
->node_id
);
6614 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6615 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6616 #endif /* CONFIG_SPARSEMEM */
6618 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6620 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6621 void __init
set_pageblock_order(void)
6625 /* Check that pageblock_nr_pages has not already been setup */
6626 if (pageblock_order
)
6629 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6630 order
= HUGETLB_PAGE_ORDER
;
6632 order
= MAX_ORDER
- 1;
6635 * Assume the largest contiguous order of interest is a huge page.
6636 * This value may be variable depending on boot parameters on IA64 and
6639 pageblock_order
= order
;
6641 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6644 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6645 * is unused as pageblock_order is set at compile-time. See
6646 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6649 void __init
set_pageblock_order(void)
6653 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6655 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6656 unsigned long present_pages
)
6658 unsigned long pages
= spanned_pages
;
6661 * Provide a more accurate estimation if there are holes within
6662 * the zone and SPARSEMEM is in use. If there are holes within the
6663 * zone, each populated memory region may cost us one or two extra
6664 * memmap pages due to alignment because memmap pages for each
6665 * populated regions may not be naturally aligned on page boundary.
6666 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6668 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6669 IS_ENABLED(CONFIG_SPARSEMEM
))
6670 pages
= present_pages
;
6672 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6675 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6676 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6678 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
6680 spin_lock_init(&ds_queue
->split_queue_lock
);
6681 INIT_LIST_HEAD(&ds_queue
->split_queue
);
6682 ds_queue
->split_queue_len
= 0;
6685 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6688 #ifdef CONFIG_COMPACTION
6689 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6691 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6694 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6697 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6699 pgdat_resize_init(pgdat
);
6701 pgdat_init_split_queue(pgdat
);
6702 pgdat_init_kcompactd(pgdat
);
6704 init_waitqueue_head(&pgdat
->kswapd_wait
);
6705 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6707 pgdat_page_ext_init(pgdat
);
6708 spin_lock_init(&pgdat
->lru_lock
);
6709 lruvec_init(node_lruvec(pgdat
));
6712 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6713 unsigned long remaining_pages
)
6715 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6716 zone_set_nid(zone
, nid
);
6717 zone
->name
= zone_names
[idx
];
6718 zone
->zone_pgdat
= NODE_DATA(nid
);
6719 spin_lock_init(&zone
->lock
);
6720 zone_seqlock_init(zone
);
6721 zone_pcp_init(zone
);
6725 * Set up the zone data structures
6726 * - init pgdat internals
6727 * - init all zones belonging to this node
6729 * NOTE: this function is only called during memory hotplug
6731 #ifdef CONFIG_MEMORY_HOTPLUG
6732 void __ref
free_area_init_core_hotplug(int nid
)
6735 pg_data_t
*pgdat
= NODE_DATA(nid
);
6737 pgdat_init_internals(pgdat
);
6738 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6739 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6744 * Set up the zone data structures:
6745 * - mark all pages reserved
6746 * - mark all memory queues empty
6747 * - clear the memory bitmaps
6749 * NOTE: pgdat should get zeroed by caller.
6750 * NOTE: this function is only called during early init.
6752 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6755 int nid
= pgdat
->node_id
;
6757 pgdat_init_internals(pgdat
);
6758 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6760 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6761 struct zone
*zone
= pgdat
->node_zones
+ j
;
6762 unsigned long size
, freesize
, memmap_pages
;
6763 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6765 size
= zone
->spanned_pages
;
6766 freesize
= zone
->present_pages
;
6769 * Adjust freesize so that it accounts for how much memory
6770 * is used by this zone for memmap. This affects the watermark
6771 * and per-cpu initialisations
6773 memmap_pages
= calc_memmap_size(size
, freesize
);
6774 if (!is_highmem_idx(j
)) {
6775 if (freesize
>= memmap_pages
) {
6776 freesize
-= memmap_pages
;
6779 " %s zone: %lu pages used for memmap\n",
6780 zone_names
[j
], memmap_pages
);
6782 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6783 zone_names
[j
], memmap_pages
, freesize
);
6786 /* Account for reserved pages */
6787 if (j
== 0 && freesize
> dma_reserve
) {
6788 freesize
-= dma_reserve
;
6789 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6790 zone_names
[0], dma_reserve
);
6793 if (!is_highmem_idx(j
))
6794 nr_kernel_pages
+= freesize
;
6795 /* Charge for highmem memmap if there are enough kernel pages */
6796 else if (nr_kernel_pages
> memmap_pages
* 2)
6797 nr_kernel_pages
-= memmap_pages
;
6798 nr_all_pages
+= freesize
;
6801 * Set an approximate value for lowmem here, it will be adjusted
6802 * when the bootmem allocator frees pages into the buddy system.
6803 * And all highmem pages will be managed by the buddy system.
6805 zone_init_internals(zone
, j
, nid
, freesize
);
6810 set_pageblock_order();
6811 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6812 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6813 memmap_init(size
, nid
, j
, zone_start_pfn
);
6817 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6818 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6820 unsigned long __maybe_unused start
= 0;
6821 unsigned long __maybe_unused offset
= 0;
6823 /* Skip empty nodes */
6824 if (!pgdat
->node_spanned_pages
)
6827 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6828 offset
= pgdat
->node_start_pfn
- start
;
6829 /* ia64 gets its own node_mem_map, before this, without bootmem */
6830 if (!pgdat
->node_mem_map
) {
6831 unsigned long size
, end
;
6835 * The zone's endpoints aren't required to be MAX_ORDER
6836 * aligned but the node_mem_map endpoints must be in order
6837 * for the buddy allocator to function correctly.
6839 end
= pgdat_end_pfn(pgdat
);
6840 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6841 size
= (end
- start
) * sizeof(struct page
);
6842 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
6845 panic("Failed to allocate %ld bytes for node %d memory map\n",
6846 size
, pgdat
->node_id
);
6847 pgdat
->node_mem_map
= map
+ offset
;
6849 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6850 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6851 (unsigned long)pgdat
->node_mem_map
);
6852 #ifndef CONFIG_NEED_MULTIPLE_NODES
6854 * With no DISCONTIG, the global mem_map is just set as node 0's
6856 if (pgdat
== NODE_DATA(0)) {
6857 mem_map
= NODE_DATA(0)->node_mem_map
;
6858 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6859 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6861 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6866 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6867 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6869 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6870 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6872 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6875 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6878 void __init
free_area_init_node(int nid
, unsigned long *zones_size
,
6879 unsigned long node_start_pfn
,
6880 unsigned long *zholes_size
)
6882 pg_data_t
*pgdat
= NODE_DATA(nid
);
6883 unsigned long start_pfn
= 0;
6884 unsigned long end_pfn
= 0;
6886 /* pg_data_t should be reset to zero when it's allocated */
6887 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6889 pgdat
->node_id
= nid
;
6890 pgdat
->node_start_pfn
= node_start_pfn
;
6891 pgdat
->per_cpu_nodestats
= NULL
;
6892 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6893 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6894 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6895 (u64
)start_pfn
<< PAGE_SHIFT
,
6896 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6898 start_pfn
= node_start_pfn
;
6900 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6901 zones_size
, zholes_size
);
6903 alloc_node_mem_map(pgdat
);
6904 pgdat_set_deferred_range(pgdat
);
6906 free_area_init_core(pgdat
);
6909 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6911 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6914 static u64
zero_pfn_range(unsigned long spfn
, unsigned long epfn
)
6919 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6920 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6921 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6922 + pageblock_nr_pages
- 1;
6925 mm_zero_struct_page(pfn_to_page(pfn
));
6933 * Only struct pages that are backed by physical memory are zeroed and
6934 * initialized by going through __init_single_page(). But, there are some
6935 * struct pages which are reserved in memblock allocator and their fields
6936 * may be accessed (for example page_to_pfn() on some configuration accesses
6937 * flags). We must explicitly zero those struct pages.
6939 * This function also addresses a similar issue where struct pages are left
6940 * uninitialized because the physical address range is not covered by
6941 * memblock.memory or memblock.reserved. That could happen when memblock
6942 * layout is manually configured via memmap=.
6944 void __init
zero_resv_unavail(void)
6946 phys_addr_t start
, end
;
6948 phys_addr_t next
= 0;
6951 * Loop through unavailable ranges not covered by memblock.memory.
6954 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6955 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6957 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), PFN_UP(start
));
6960 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), max_pfn
);
6963 * Struct pages that do not have backing memory. This could be because
6964 * firmware is using some of this memory, or for some other reasons.
6967 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
6969 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6971 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6973 #if MAX_NUMNODES > 1
6975 * Figure out the number of possible node ids.
6977 void __init
setup_nr_node_ids(void)
6979 unsigned int highest
;
6981 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6982 nr_node_ids
= highest
+ 1;
6987 * node_map_pfn_alignment - determine the maximum internode alignment
6989 * This function should be called after node map is populated and sorted.
6990 * It calculates the maximum power of two alignment which can distinguish
6993 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6994 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6995 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6996 * shifted, 1GiB is enough and this function will indicate so.
6998 * This is used to test whether pfn -> nid mapping of the chosen memory
6999 * model has fine enough granularity to avoid incorrect mapping for the
7000 * populated node map.
7002 * Return: the determined alignment in pfn's. 0 if there is no alignment
7003 * requirement (single node).
7005 unsigned long __init
node_map_pfn_alignment(void)
7007 unsigned long accl_mask
= 0, last_end
= 0;
7008 unsigned long start
, end
, mask
;
7009 int last_nid
= NUMA_NO_NODE
;
7012 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
7013 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
7020 * Start with a mask granular enough to pin-point to the
7021 * start pfn and tick off bits one-by-one until it becomes
7022 * too coarse to separate the current node from the last.
7024 mask
= ~((1 << __ffs(start
)) - 1);
7025 while (mask
&& last_end
<= (start
& (mask
<< 1)))
7028 /* accumulate all internode masks */
7032 /* convert mask to number of pages */
7033 return ~accl_mask
+ 1;
7036 /* Find the lowest pfn for a node */
7037 static unsigned long __init
find_min_pfn_for_node(int nid
)
7039 unsigned long min_pfn
= ULONG_MAX
;
7040 unsigned long start_pfn
;
7043 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
7044 min_pfn
= min(min_pfn
, start_pfn
);
7046 if (min_pfn
== ULONG_MAX
) {
7047 pr_warn("Could not find start_pfn for node %d\n", nid
);
7055 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7057 * Return: the minimum PFN based on information provided via
7058 * memblock_set_node().
7060 unsigned long __init
find_min_pfn_with_active_regions(void)
7062 return find_min_pfn_for_node(MAX_NUMNODES
);
7066 * early_calculate_totalpages()
7067 * Sum pages in active regions for movable zone.
7068 * Populate N_MEMORY for calculating usable_nodes.
7070 static unsigned long __init
early_calculate_totalpages(void)
7072 unsigned long totalpages
= 0;
7073 unsigned long start_pfn
, end_pfn
;
7076 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7077 unsigned long pages
= end_pfn
- start_pfn
;
7079 totalpages
+= pages
;
7081 node_set_state(nid
, N_MEMORY
);
7087 * Find the PFN the Movable zone begins in each node. Kernel memory
7088 * is spread evenly between nodes as long as the nodes have enough
7089 * memory. When they don't, some nodes will have more kernelcore than
7092 static void __init
find_zone_movable_pfns_for_nodes(void)
7095 unsigned long usable_startpfn
;
7096 unsigned long kernelcore_node
, kernelcore_remaining
;
7097 /* save the state before borrow the nodemask */
7098 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
7099 unsigned long totalpages
= early_calculate_totalpages();
7100 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
7101 struct memblock_region
*r
;
7103 /* Need to find movable_zone earlier when movable_node is specified. */
7104 find_usable_zone_for_movable();
7107 * If movable_node is specified, ignore kernelcore and movablecore
7110 if (movable_node_is_enabled()) {
7111 for_each_memblock(memory
, r
) {
7112 if (!memblock_is_hotpluggable(r
))
7117 usable_startpfn
= PFN_DOWN(r
->base
);
7118 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7119 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7127 * If kernelcore=mirror is specified, ignore movablecore option
7129 if (mirrored_kernelcore
) {
7130 bool mem_below_4gb_not_mirrored
= false;
7132 for_each_memblock(memory
, r
) {
7133 if (memblock_is_mirror(r
))
7138 usable_startpfn
= memblock_region_memory_base_pfn(r
);
7140 if (usable_startpfn
< 0x100000) {
7141 mem_below_4gb_not_mirrored
= true;
7145 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7146 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7150 if (mem_below_4gb_not_mirrored
)
7151 pr_warn("This configuration results in unmirrored kernel memory.");
7157 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7158 * amount of necessary memory.
7160 if (required_kernelcore_percent
)
7161 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
7163 if (required_movablecore_percent
)
7164 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
7168 * If movablecore= was specified, calculate what size of
7169 * kernelcore that corresponds so that memory usable for
7170 * any allocation type is evenly spread. If both kernelcore
7171 * and movablecore are specified, then the value of kernelcore
7172 * will be used for required_kernelcore if it's greater than
7173 * what movablecore would have allowed.
7175 if (required_movablecore
) {
7176 unsigned long corepages
;
7179 * Round-up so that ZONE_MOVABLE is at least as large as what
7180 * was requested by the user
7182 required_movablecore
=
7183 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
7184 required_movablecore
= min(totalpages
, required_movablecore
);
7185 corepages
= totalpages
- required_movablecore
;
7187 required_kernelcore
= max(required_kernelcore
, corepages
);
7191 * If kernelcore was not specified or kernelcore size is larger
7192 * than totalpages, there is no ZONE_MOVABLE.
7194 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
7197 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7198 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
7201 /* Spread kernelcore memory as evenly as possible throughout nodes */
7202 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7203 for_each_node_state(nid
, N_MEMORY
) {
7204 unsigned long start_pfn
, end_pfn
;
7207 * Recalculate kernelcore_node if the division per node
7208 * now exceeds what is necessary to satisfy the requested
7209 * amount of memory for the kernel
7211 if (required_kernelcore
< kernelcore_node
)
7212 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7215 * As the map is walked, we track how much memory is usable
7216 * by the kernel using kernelcore_remaining. When it is
7217 * 0, the rest of the node is usable by ZONE_MOVABLE
7219 kernelcore_remaining
= kernelcore_node
;
7221 /* Go through each range of PFNs within this node */
7222 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7223 unsigned long size_pages
;
7225 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
7226 if (start_pfn
>= end_pfn
)
7229 /* Account for what is only usable for kernelcore */
7230 if (start_pfn
< usable_startpfn
) {
7231 unsigned long kernel_pages
;
7232 kernel_pages
= min(end_pfn
, usable_startpfn
)
7235 kernelcore_remaining
-= min(kernel_pages
,
7236 kernelcore_remaining
);
7237 required_kernelcore
-= min(kernel_pages
,
7238 required_kernelcore
);
7240 /* Continue if range is now fully accounted */
7241 if (end_pfn
<= usable_startpfn
) {
7244 * Push zone_movable_pfn to the end so
7245 * that if we have to rebalance
7246 * kernelcore across nodes, we will
7247 * not double account here
7249 zone_movable_pfn
[nid
] = end_pfn
;
7252 start_pfn
= usable_startpfn
;
7256 * The usable PFN range for ZONE_MOVABLE is from
7257 * start_pfn->end_pfn. Calculate size_pages as the
7258 * number of pages used as kernelcore
7260 size_pages
= end_pfn
- start_pfn
;
7261 if (size_pages
> kernelcore_remaining
)
7262 size_pages
= kernelcore_remaining
;
7263 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7266 * Some kernelcore has been met, update counts and
7267 * break if the kernelcore for this node has been
7270 required_kernelcore
-= min(required_kernelcore
,
7272 kernelcore_remaining
-= size_pages
;
7273 if (!kernelcore_remaining
)
7279 * If there is still required_kernelcore, we do another pass with one
7280 * less node in the count. This will push zone_movable_pfn[nid] further
7281 * along on the nodes that still have memory until kernelcore is
7285 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7289 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7290 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7291 zone_movable_pfn
[nid
] =
7292 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7295 /* restore the node_state */
7296 node_states
[N_MEMORY
] = saved_node_state
;
7299 /* Any regular or high memory on that node ? */
7300 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7302 enum zone_type zone_type
;
7304 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7305 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7306 if (populated_zone(zone
)) {
7307 if (IS_ENABLED(CONFIG_HIGHMEM
))
7308 node_set_state(nid
, N_HIGH_MEMORY
);
7309 if (zone_type
<= ZONE_NORMAL
)
7310 node_set_state(nid
, N_NORMAL_MEMORY
);
7317 * free_area_init_nodes - Initialise all pg_data_t and zone data
7318 * @max_zone_pfn: an array of max PFNs for each zone
7320 * This will call free_area_init_node() for each active node in the system.
7321 * Using the page ranges provided by memblock_set_node(), the size of each
7322 * zone in each node and their holes is calculated. If the maximum PFN
7323 * between two adjacent zones match, it is assumed that the zone is empty.
7324 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7325 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7326 * starts where the previous one ended. For example, ZONE_DMA32 starts
7327 * at arch_max_dma_pfn.
7329 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
7331 unsigned long start_pfn
, end_pfn
;
7334 /* Record where the zone boundaries are */
7335 memset(arch_zone_lowest_possible_pfn
, 0,
7336 sizeof(arch_zone_lowest_possible_pfn
));
7337 memset(arch_zone_highest_possible_pfn
, 0,
7338 sizeof(arch_zone_highest_possible_pfn
));
7340 start_pfn
= find_min_pfn_with_active_regions();
7342 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7343 if (i
== ZONE_MOVABLE
)
7346 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
7347 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
7348 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
7350 start_pfn
= end_pfn
;
7353 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7354 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7355 find_zone_movable_pfns_for_nodes();
7357 /* Print out the zone ranges */
7358 pr_info("Zone ranges:\n");
7359 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7360 if (i
== ZONE_MOVABLE
)
7362 pr_info(" %-8s ", zone_names
[i
]);
7363 if (arch_zone_lowest_possible_pfn
[i
] ==
7364 arch_zone_highest_possible_pfn
[i
])
7367 pr_cont("[mem %#018Lx-%#018Lx]\n",
7368 (u64
)arch_zone_lowest_possible_pfn
[i
]
7370 ((u64
)arch_zone_highest_possible_pfn
[i
]
7371 << PAGE_SHIFT
) - 1);
7374 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7375 pr_info("Movable zone start for each node\n");
7376 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7377 if (zone_movable_pfn
[i
])
7378 pr_info(" Node %d: %#018Lx\n", i
,
7379 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7383 * Print out the early node map, and initialize the
7384 * subsection-map relative to active online memory ranges to
7385 * enable future "sub-section" extensions of the memory map.
7387 pr_info("Early memory node ranges\n");
7388 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7389 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7390 (u64
)start_pfn
<< PAGE_SHIFT
,
7391 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7392 subsection_map_init(start_pfn
, end_pfn
- start_pfn
);
7395 /* Initialise every node */
7396 mminit_verify_pageflags_layout();
7397 setup_nr_node_ids();
7398 zero_resv_unavail();
7399 for_each_online_node(nid
) {
7400 pg_data_t
*pgdat
= NODE_DATA(nid
);
7401 free_area_init_node(nid
, NULL
,
7402 find_min_pfn_for_node(nid
), NULL
);
7404 /* Any memory on that node */
7405 if (pgdat
->node_present_pages
)
7406 node_set_state(nid
, N_MEMORY
);
7407 check_for_memory(pgdat
, nid
);
7411 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7412 unsigned long *percent
)
7414 unsigned long long coremem
;
7420 /* Value may be a percentage of total memory, otherwise bytes */
7421 coremem
= simple_strtoull(p
, &endptr
, 0);
7422 if (*endptr
== '%') {
7423 /* Paranoid check for percent values greater than 100 */
7424 WARN_ON(coremem
> 100);
7428 coremem
= memparse(p
, &p
);
7429 /* Paranoid check that UL is enough for the coremem value */
7430 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7432 *core
= coremem
>> PAGE_SHIFT
;
7439 * kernelcore=size sets the amount of memory for use for allocations that
7440 * cannot be reclaimed or migrated.
7442 static int __init
cmdline_parse_kernelcore(char *p
)
7444 /* parse kernelcore=mirror */
7445 if (parse_option_str(p
, "mirror")) {
7446 mirrored_kernelcore
= true;
7450 return cmdline_parse_core(p
, &required_kernelcore
,
7451 &required_kernelcore_percent
);
7455 * movablecore=size sets the amount of memory for use for allocations that
7456 * can be reclaimed or migrated.
7458 static int __init
cmdline_parse_movablecore(char *p
)
7460 return cmdline_parse_core(p
, &required_movablecore
,
7461 &required_movablecore_percent
);
7464 early_param("kernelcore", cmdline_parse_kernelcore
);
7465 early_param("movablecore", cmdline_parse_movablecore
);
7467 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7469 void adjust_managed_page_count(struct page
*page
, long count
)
7471 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7472 totalram_pages_add(count
);
7473 #ifdef CONFIG_HIGHMEM
7474 if (PageHighMem(page
))
7475 totalhigh_pages_add(count
);
7478 EXPORT_SYMBOL(adjust_managed_page_count
);
7480 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7483 unsigned long pages
= 0;
7485 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7486 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7487 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7488 struct page
*page
= virt_to_page(pos
);
7489 void *direct_map_addr
;
7492 * 'direct_map_addr' might be different from 'pos'
7493 * because some architectures' virt_to_page()
7494 * work with aliases. Getting the direct map
7495 * address ensures that we get a _writeable_
7496 * alias for the memset().
7498 direct_map_addr
= page_address(page
);
7499 if ((unsigned int)poison
<= 0xFF)
7500 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7502 free_reserved_page(page
);
7506 pr_info("Freeing %s memory: %ldK\n",
7507 s
, pages
<< (PAGE_SHIFT
- 10));
7512 #ifdef CONFIG_HIGHMEM
7513 void free_highmem_page(struct page
*page
)
7515 __free_reserved_page(page
);
7516 totalram_pages_inc();
7517 atomic_long_inc(&page_zone(page
)->managed_pages
);
7518 totalhigh_pages_inc();
7523 void __init
mem_init_print_info(const char *str
)
7525 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7526 unsigned long init_code_size
, init_data_size
;
7528 physpages
= get_num_physpages();
7529 codesize
= _etext
- _stext
;
7530 datasize
= _edata
- _sdata
;
7531 rosize
= __end_rodata
- __start_rodata
;
7532 bss_size
= __bss_stop
- __bss_start
;
7533 init_data_size
= __init_end
- __init_begin
;
7534 init_code_size
= _einittext
- _sinittext
;
7537 * Detect special cases and adjust section sizes accordingly:
7538 * 1) .init.* may be embedded into .data sections
7539 * 2) .init.text.* may be out of [__init_begin, __init_end],
7540 * please refer to arch/tile/kernel/vmlinux.lds.S.
7541 * 3) .rodata.* may be embedded into .text or .data sections.
7543 #define adj_init_size(start, end, size, pos, adj) \
7545 if (start <= pos && pos < end && size > adj) \
7549 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7550 _sinittext
, init_code_size
);
7551 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7552 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7553 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7554 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7556 #undef adj_init_size
7558 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7559 #ifdef CONFIG_HIGHMEM
7563 nr_free_pages() << (PAGE_SHIFT
- 10),
7564 physpages
<< (PAGE_SHIFT
- 10),
7565 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7566 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7567 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7568 totalcma_pages
<< (PAGE_SHIFT
- 10),
7569 #ifdef CONFIG_HIGHMEM
7570 totalhigh_pages() << (PAGE_SHIFT
- 10),
7572 str
? ", " : "", str
? str
: "");
7576 * set_dma_reserve - set the specified number of pages reserved in the first zone
7577 * @new_dma_reserve: The number of pages to mark reserved
7579 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7580 * In the DMA zone, a significant percentage may be consumed by kernel image
7581 * and other unfreeable allocations which can skew the watermarks badly. This
7582 * function may optionally be used to account for unfreeable pages in the
7583 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7584 * smaller per-cpu batchsize.
7586 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7588 dma_reserve
= new_dma_reserve
;
7591 void __init
free_area_init(unsigned long *zones_size
)
7593 zero_resv_unavail();
7594 free_area_init_node(0, zones_size
,
7595 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
7598 static int page_alloc_cpu_dead(unsigned int cpu
)
7601 lru_add_drain_cpu(cpu
);
7605 * Spill the event counters of the dead processor
7606 * into the current processors event counters.
7607 * This artificially elevates the count of the current
7610 vm_events_fold_cpu(cpu
);
7613 * Zero the differential counters of the dead processor
7614 * so that the vm statistics are consistent.
7616 * This is only okay since the processor is dead and cannot
7617 * race with what we are doing.
7619 cpu_vm_stats_fold(cpu
);
7624 int hashdist
= HASHDIST_DEFAULT
;
7626 static int __init
set_hashdist(char *str
)
7630 hashdist
= simple_strtoul(str
, &str
, 0);
7633 __setup("hashdist=", set_hashdist
);
7636 void __init
page_alloc_init(void)
7641 if (num_node_state(N_MEMORY
) == 1)
7645 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7646 "mm/page_alloc:dead", NULL
,
7647 page_alloc_cpu_dead
);
7652 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7653 * or min_free_kbytes changes.
7655 static void calculate_totalreserve_pages(void)
7657 struct pglist_data
*pgdat
;
7658 unsigned long reserve_pages
= 0;
7659 enum zone_type i
, j
;
7661 for_each_online_pgdat(pgdat
) {
7663 pgdat
->totalreserve_pages
= 0;
7665 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7666 struct zone
*zone
= pgdat
->node_zones
+ i
;
7668 unsigned long managed_pages
= zone_managed_pages(zone
);
7670 /* Find valid and maximum lowmem_reserve in the zone */
7671 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7672 if (zone
->lowmem_reserve
[j
] > max
)
7673 max
= zone
->lowmem_reserve
[j
];
7676 /* we treat the high watermark as reserved pages. */
7677 max
+= high_wmark_pages(zone
);
7679 if (max
> managed_pages
)
7680 max
= managed_pages
;
7682 pgdat
->totalreserve_pages
+= max
;
7684 reserve_pages
+= max
;
7687 totalreserve_pages
= reserve_pages
;
7691 * setup_per_zone_lowmem_reserve - called whenever
7692 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7693 * has a correct pages reserved value, so an adequate number of
7694 * pages are left in the zone after a successful __alloc_pages().
7696 static void setup_per_zone_lowmem_reserve(void)
7698 struct pglist_data
*pgdat
;
7699 enum zone_type j
, idx
;
7701 for_each_online_pgdat(pgdat
) {
7702 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7703 struct zone
*zone
= pgdat
->node_zones
+ j
;
7704 unsigned long managed_pages
= zone_managed_pages(zone
);
7706 zone
->lowmem_reserve
[j
] = 0;
7710 struct zone
*lower_zone
;
7713 lower_zone
= pgdat
->node_zones
+ idx
;
7715 if (sysctl_lowmem_reserve_ratio
[idx
] < 1) {
7716 sysctl_lowmem_reserve_ratio
[idx
] = 0;
7717 lower_zone
->lowmem_reserve
[j
] = 0;
7719 lower_zone
->lowmem_reserve
[j
] =
7720 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7722 managed_pages
+= zone_managed_pages(lower_zone
);
7727 /* update totalreserve_pages */
7728 calculate_totalreserve_pages();
7731 static void __setup_per_zone_wmarks(void)
7733 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7734 unsigned long lowmem_pages
= 0;
7736 unsigned long flags
;
7738 /* Calculate total number of !ZONE_HIGHMEM pages */
7739 for_each_zone(zone
) {
7740 if (!is_highmem(zone
))
7741 lowmem_pages
+= zone_managed_pages(zone
);
7744 for_each_zone(zone
) {
7747 spin_lock_irqsave(&zone
->lock
, flags
);
7748 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7749 do_div(tmp
, lowmem_pages
);
7750 if (is_highmem(zone
)) {
7752 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7753 * need highmem pages, so cap pages_min to a small
7756 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7757 * deltas control async page reclaim, and so should
7758 * not be capped for highmem.
7760 unsigned long min_pages
;
7762 min_pages
= zone_managed_pages(zone
) / 1024;
7763 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7764 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7767 * If it's a lowmem zone, reserve a number of pages
7768 * proportionate to the zone's size.
7770 zone
->_watermark
[WMARK_MIN
] = tmp
;
7774 * Set the kswapd watermarks distance according to the
7775 * scale factor in proportion to available memory, but
7776 * ensure a minimum size on small systems.
7778 tmp
= max_t(u64
, tmp
>> 2,
7779 mult_frac(zone_managed_pages(zone
),
7780 watermark_scale_factor
, 10000));
7782 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7783 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7784 zone
->watermark_boost
= 0;
7786 spin_unlock_irqrestore(&zone
->lock
, flags
);
7789 /* update totalreserve_pages */
7790 calculate_totalreserve_pages();
7794 * setup_per_zone_wmarks - called when min_free_kbytes changes
7795 * or when memory is hot-{added|removed}
7797 * Ensures that the watermark[min,low,high] values for each zone are set
7798 * correctly with respect to min_free_kbytes.
7800 void setup_per_zone_wmarks(void)
7802 static DEFINE_SPINLOCK(lock
);
7805 __setup_per_zone_wmarks();
7810 * Initialise min_free_kbytes.
7812 * For small machines we want it small (128k min). For large machines
7813 * we want it large (64MB max). But it is not linear, because network
7814 * bandwidth does not increase linearly with machine size. We use
7816 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7817 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7833 int __meminit
init_per_zone_wmark_min(void)
7835 unsigned long lowmem_kbytes
;
7836 int new_min_free_kbytes
;
7838 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7839 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7841 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7842 min_free_kbytes
= new_min_free_kbytes
;
7843 if (min_free_kbytes
< 128)
7844 min_free_kbytes
= 128;
7845 if (min_free_kbytes
> 65536)
7846 min_free_kbytes
= 65536;
7848 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7849 new_min_free_kbytes
, user_min_free_kbytes
);
7851 setup_per_zone_wmarks();
7852 refresh_zone_stat_thresholds();
7853 setup_per_zone_lowmem_reserve();
7856 setup_min_unmapped_ratio();
7857 setup_min_slab_ratio();
7862 core_initcall(init_per_zone_wmark_min
)
7865 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7866 * that we can call two helper functions whenever min_free_kbytes
7869 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7870 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7874 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7879 user_min_free_kbytes
= min_free_kbytes
;
7880 setup_per_zone_wmarks();
7885 int watermark_boost_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7886 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7890 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7897 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7898 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7902 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7907 setup_per_zone_wmarks();
7913 static void setup_min_unmapped_ratio(void)
7918 for_each_online_pgdat(pgdat
)
7919 pgdat
->min_unmapped_pages
= 0;
7922 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
7923 sysctl_min_unmapped_ratio
) / 100;
7927 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7928 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7932 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7936 setup_min_unmapped_ratio();
7941 static void setup_min_slab_ratio(void)
7946 for_each_online_pgdat(pgdat
)
7947 pgdat
->min_slab_pages
= 0;
7950 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
7951 sysctl_min_slab_ratio
) / 100;
7954 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7955 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7959 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7963 setup_min_slab_ratio();
7970 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7971 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7972 * whenever sysctl_lowmem_reserve_ratio changes.
7974 * The reserve ratio obviously has absolutely no relation with the
7975 * minimum watermarks. The lowmem reserve ratio can only make sense
7976 * if in function of the boot time zone sizes.
7978 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7979 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7981 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7982 setup_per_zone_lowmem_reserve();
7987 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7988 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7989 * pagelist can have before it gets flushed back to buddy allocator.
7991 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7992 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7995 int old_percpu_pagelist_fraction
;
7998 mutex_lock(&pcp_batch_high_lock
);
7999 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
8001 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8002 if (!write
|| ret
< 0)
8005 /* Sanity checking to avoid pcp imbalance */
8006 if (percpu_pagelist_fraction
&&
8007 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
8008 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
8014 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
8017 for_each_populated_zone(zone
) {
8020 for_each_possible_cpu(cpu
)
8021 pageset_set_high_and_batch(zone
,
8022 per_cpu_ptr(zone
->pageset
, cpu
));
8025 mutex_unlock(&pcp_batch_high_lock
);
8029 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8031 * Returns the number of pages that arch has reserved but
8032 * is not known to alloc_large_system_hash().
8034 static unsigned long __init
arch_reserved_kernel_pages(void)
8041 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8042 * machines. As memory size is increased the scale is also increased but at
8043 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8044 * quadruples the scale is increased by one, which means the size of hash table
8045 * only doubles, instead of quadrupling as well.
8046 * Because 32-bit systems cannot have large physical memory, where this scaling
8047 * makes sense, it is disabled on such platforms.
8049 #if __BITS_PER_LONG > 32
8050 #define ADAPT_SCALE_BASE (64ul << 30)
8051 #define ADAPT_SCALE_SHIFT 2
8052 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8056 * allocate a large system hash table from bootmem
8057 * - it is assumed that the hash table must contain an exact power-of-2
8058 * quantity of entries
8059 * - limit is the number of hash buckets, not the total allocation size
8061 void *__init
alloc_large_system_hash(const char *tablename
,
8062 unsigned long bucketsize
,
8063 unsigned long numentries
,
8066 unsigned int *_hash_shift
,
8067 unsigned int *_hash_mask
,
8068 unsigned long low_limit
,
8069 unsigned long high_limit
)
8071 unsigned long long max
= high_limit
;
8072 unsigned long log2qty
, size
;
8077 /* allow the kernel cmdline to have a say */
8079 /* round applicable memory size up to nearest megabyte */
8080 numentries
= nr_kernel_pages
;
8081 numentries
-= arch_reserved_kernel_pages();
8083 /* It isn't necessary when PAGE_SIZE >= 1MB */
8084 if (PAGE_SHIFT
< 20)
8085 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
8087 #if __BITS_PER_LONG > 32
8089 unsigned long adapt
;
8091 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
8092 adapt
<<= ADAPT_SCALE_SHIFT
)
8097 /* limit to 1 bucket per 2^scale bytes of low memory */
8098 if (scale
> PAGE_SHIFT
)
8099 numentries
>>= (scale
- PAGE_SHIFT
);
8101 numentries
<<= (PAGE_SHIFT
- scale
);
8103 /* Make sure we've got at least a 0-order allocation.. */
8104 if (unlikely(flags
& HASH_SMALL
)) {
8105 /* Makes no sense without HASH_EARLY */
8106 WARN_ON(!(flags
& HASH_EARLY
));
8107 if (!(numentries
>> *_hash_shift
)) {
8108 numentries
= 1UL << *_hash_shift
;
8109 BUG_ON(!numentries
);
8111 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
8112 numentries
= PAGE_SIZE
/ bucketsize
;
8114 numentries
= roundup_pow_of_two(numentries
);
8116 /* limit allocation size to 1/16 total memory by default */
8118 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
8119 do_div(max
, bucketsize
);
8121 max
= min(max
, 0x80000000ULL
);
8123 if (numentries
< low_limit
)
8124 numentries
= low_limit
;
8125 if (numentries
> max
)
8128 log2qty
= ilog2(numentries
);
8130 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
8133 size
= bucketsize
<< log2qty
;
8134 if (flags
& HASH_EARLY
) {
8135 if (flags
& HASH_ZERO
)
8136 table
= memblock_alloc(size
, SMP_CACHE_BYTES
);
8138 table
= memblock_alloc_raw(size
,
8140 } else if (get_order(size
) >= MAX_ORDER
|| hashdist
) {
8141 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
8145 * If bucketsize is not a power-of-two, we may free
8146 * some pages at the end of hash table which
8147 * alloc_pages_exact() automatically does
8149 table
= alloc_pages_exact(size
, gfp_flags
);
8150 kmemleak_alloc(table
, size
, 1, gfp_flags
);
8152 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
8155 panic("Failed to allocate %s hash table\n", tablename
);
8157 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8158 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
,
8159 virt
? "vmalloc" : "linear");
8162 *_hash_shift
= log2qty
;
8164 *_hash_mask
= (1 << log2qty
) - 1;
8170 * This function checks whether pageblock includes unmovable pages or not.
8171 * If @count is not zero, it is okay to include less @count unmovable pages
8173 * PageLRU check without isolation or lru_lock could race so that
8174 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8175 * check without lock_page also may miss some movable non-lru pages at
8176 * race condition. So you can't expect this function should be exact.
8178 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
8179 int migratetype
, int flags
)
8181 unsigned long found
;
8182 unsigned long iter
= 0;
8183 unsigned long pfn
= page_to_pfn(page
);
8184 const char *reason
= "unmovable page";
8187 * TODO we could make this much more efficient by not checking every
8188 * page in the range if we know all of them are in MOVABLE_ZONE and
8189 * that the movable zone guarantees that pages are migratable but
8190 * the later is not the case right now unfortunatelly. E.g. movablecore
8191 * can still lead to having bootmem allocations in zone_movable.
8194 if (is_migrate_cma_page(page
)) {
8196 * CMA allocations (alloc_contig_range) really need to mark
8197 * isolate CMA pageblocks even when they are not movable in fact
8198 * so consider them movable here.
8200 if (is_migrate_cma(migratetype
))
8203 reason
= "CMA page";
8207 for (found
= 0; iter
< pageblock_nr_pages
; iter
++) {
8208 unsigned long check
= pfn
+ iter
;
8210 if (!pfn_valid_within(check
))
8213 page
= pfn_to_page(check
);
8215 if (PageReserved(page
))
8219 * If the zone is movable and we have ruled out all reserved
8220 * pages then it should be reasonably safe to assume the rest
8223 if (zone_idx(zone
) == ZONE_MOVABLE
)
8227 * Hugepages are not in LRU lists, but they're movable.
8228 * We need not scan over tail pages because we don't
8229 * handle each tail page individually in migration.
8231 if (PageHuge(page
)) {
8232 struct page
*head
= compound_head(page
);
8233 unsigned int skip_pages
;
8235 if (!hugepage_migration_supported(page_hstate(head
)))
8238 skip_pages
= compound_nr(head
) - (page
- head
);
8239 iter
+= skip_pages
- 1;
8244 * We can't use page_count without pin a page
8245 * because another CPU can free compound page.
8246 * This check already skips compound tails of THP
8247 * because their page->_refcount is zero at all time.
8249 if (!page_ref_count(page
)) {
8250 if (PageBuddy(page
))
8251 iter
+= (1 << page_order(page
)) - 1;
8256 * The HWPoisoned page may be not in buddy system, and
8257 * page_count() is not 0.
8259 if ((flags
& SKIP_HWPOISON
) && PageHWPoison(page
))
8262 if (__PageMovable(page
))
8268 * If there are RECLAIMABLE pages, we need to check
8269 * it. But now, memory offline itself doesn't call
8270 * shrink_node_slabs() and it still to be fixed.
8273 * If the page is not RAM, page_count()should be 0.
8274 * we don't need more check. This is an _used_ not-movable page.
8276 * The problematic thing here is PG_reserved pages. PG_reserved
8277 * is set to both of a memory hole page and a _used_ kernel
8285 WARN_ON_ONCE(zone_idx(zone
) == ZONE_MOVABLE
);
8286 if (flags
& REPORT_FAILURE
)
8287 dump_page(pfn_to_page(pfn
+ iter
), reason
);
8291 #ifdef CONFIG_CONTIG_ALLOC
8292 static unsigned long pfn_max_align_down(unsigned long pfn
)
8294 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8295 pageblock_nr_pages
) - 1);
8298 static unsigned long pfn_max_align_up(unsigned long pfn
)
8300 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8301 pageblock_nr_pages
));
8304 /* [start, end) must belong to a single zone. */
8305 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8306 unsigned long start
, unsigned long end
)
8308 /* This function is based on compact_zone() from compaction.c. */
8309 unsigned long nr_reclaimed
;
8310 unsigned long pfn
= start
;
8311 unsigned int tries
= 0;
8316 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8317 if (fatal_signal_pending(current
)) {
8322 if (list_empty(&cc
->migratepages
)) {
8323 cc
->nr_migratepages
= 0;
8324 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8330 } else if (++tries
== 5) {
8331 ret
= ret
< 0 ? ret
: -EBUSY
;
8335 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8337 cc
->nr_migratepages
-= nr_reclaimed
;
8339 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
8340 NULL
, 0, cc
->mode
, MR_CONTIG_RANGE
);
8343 putback_movable_pages(&cc
->migratepages
);
8350 * alloc_contig_range() -- tries to allocate given range of pages
8351 * @start: start PFN to allocate
8352 * @end: one-past-the-last PFN to allocate
8353 * @migratetype: migratetype of the underlaying pageblocks (either
8354 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8355 * in range must have the same migratetype and it must
8356 * be either of the two.
8357 * @gfp_mask: GFP mask to use during compaction
8359 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8360 * aligned. The PFN range must belong to a single zone.
8362 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8363 * pageblocks in the range. Once isolated, the pageblocks should not
8364 * be modified by others.
8366 * Return: zero on success or negative error code. On success all
8367 * pages which PFN is in [start, end) are allocated for the caller and
8368 * need to be freed with free_contig_range().
8370 int alloc_contig_range(unsigned long start
, unsigned long end
,
8371 unsigned migratetype
, gfp_t gfp_mask
)
8373 unsigned long outer_start
, outer_end
;
8377 struct compact_control cc
= {
8378 .nr_migratepages
= 0,
8380 .zone
= page_zone(pfn_to_page(start
)),
8381 .mode
= MIGRATE_SYNC
,
8382 .ignore_skip_hint
= true,
8383 .no_set_skip_hint
= true,
8384 .gfp_mask
= current_gfp_context(gfp_mask
),
8386 INIT_LIST_HEAD(&cc
.migratepages
);
8389 * What we do here is we mark all pageblocks in range as
8390 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8391 * have different sizes, and due to the way page allocator
8392 * work, we align the range to biggest of the two pages so
8393 * that page allocator won't try to merge buddies from
8394 * different pageblocks and change MIGRATE_ISOLATE to some
8395 * other migration type.
8397 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8398 * migrate the pages from an unaligned range (ie. pages that
8399 * we are interested in). This will put all the pages in
8400 * range back to page allocator as MIGRATE_ISOLATE.
8402 * When this is done, we take the pages in range from page
8403 * allocator removing them from the buddy system. This way
8404 * page allocator will never consider using them.
8406 * This lets us mark the pageblocks back as
8407 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8408 * aligned range but not in the unaligned, original range are
8409 * put back to page allocator so that buddy can use them.
8412 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8413 pfn_max_align_up(end
), migratetype
, 0);
8418 * In case of -EBUSY, we'd like to know which page causes problem.
8419 * So, just fall through. test_pages_isolated() has a tracepoint
8420 * which will report the busy page.
8422 * It is possible that busy pages could become available before
8423 * the call to test_pages_isolated, and the range will actually be
8424 * allocated. So, if we fall through be sure to clear ret so that
8425 * -EBUSY is not accidentally used or returned to caller.
8427 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8428 if (ret
&& ret
!= -EBUSY
)
8433 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8434 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8435 * more, all pages in [start, end) are free in page allocator.
8436 * What we are going to do is to allocate all pages from
8437 * [start, end) (that is remove them from page allocator).
8439 * The only problem is that pages at the beginning and at the
8440 * end of interesting range may be not aligned with pages that
8441 * page allocator holds, ie. they can be part of higher order
8442 * pages. Because of this, we reserve the bigger range and
8443 * once this is done free the pages we are not interested in.
8445 * We don't have to hold zone->lock here because the pages are
8446 * isolated thus they won't get removed from buddy.
8449 lru_add_drain_all();
8452 outer_start
= start
;
8453 while (!PageBuddy(pfn_to_page(outer_start
))) {
8454 if (++order
>= MAX_ORDER
) {
8455 outer_start
= start
;
8458 outer_start
&= ~0UL << order
;
8461 if (outer_start
!= start
) {
8462 order
= page_order(pfn_to_page(outer_start
));
8465 * outer_start page could be small order buddy page and
8466 * it doesn't include start page. Adjust outer_start
8467 * in this case to report failed page properly
8468 * on tracepoint in test_pages_isolated()
8470 if (outer_start
+ (1UL << order
) <= start
)
8471 outer_start
= start
;
8474 /* Make sure the range is really isolated. */
8475 if (test_pages_isolated(outer_start
, end
, false)) {
8476 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8477 __func__
, outer_start
, end
);
8482 /* Grab isolated pages from freelists. */
8483 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8489 /* Free head and tail (if any) */
8490 if (start
!= outer_start
)
8491 free_contig_range(outer_start
, start
- outer_start
);
8492 if (end
!= outer_end
)
8493 free_contig_range(end
, outer_end
- end
);
8496 undo_isolate_page_range(pfn_max_align_down(start
),
8497 pfn_max_align_up(end
), migratetype
);
8500 #endif /* CONFIG_CONTIG_ALLOC */
8502 void free_contig_range(unsigned long pfn
, unsigned int nr_pages
)
8504 unsigned int count
= 0;
8506 for (; nr_pages
--; pfn
++) {
8507 struct page
*page
= pfn_to_page(pfn
);
8509 count
+= page_count(page
) != 1;
8512 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8515 #ifdef CONFIG_MEMORY_HOTPLUG
8517 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8518 * page high values need to be recalulated.
8520 void __meminit
zone_pcp_update(struct zone
*zone
)
8523 mutex_lock(&pcp_batch_high_lock
);
8524 for_each_possible_cpu(cpu
)
8525 pageset_set_high_and_batch(zone
,
8526 per_cpu_ptr(zone
->pageset
, cpu
));
8527 mutex_unlock(&pcp_batch_high_lock
);
8531 void zone_pcp_reset(struct zone
*zone
)
8533 unsigned long flags
;
8535 struct per_cpu_pageset
*pset
;
8537 /* avoid races with drain_pages() */
8538 local_irq_save(flags
);
8539 if (zone
->pageset
!= &boot_pageset
) {
8540 for_each_online_cpu(cpu
) {
8541 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8542 drain_zonestat(zone
, pset
);
8544 free_percpu(zone
->pageset
);
8545 zone
->pageset
= &boot_pageset
;
8547 local_irq_restore(flags
);
8550 #ifdef CONFIG_MEMORY_HOTREMOVE
8552 * All pages in the range must be in a single zone and isolated
8553 * before calling this.
8556 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8560 unsigned int order
, i
;
8562 unsigned long flags
;
8563 unsigned long offlined_pages
= 0;
8565 /* find the first valid pfn */
8566 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8570 return offlined_pages
;
8572 offline_mem_sections(pfn
, end_pfn
);
8573 zone
= page_zone(pfn_to_page(pfn
));
8574 spin_lock_irqsave(&zone
->lock
, flags
);
8576 while (pfn
< end_pfn
) {
8577 if (!pfn_valid(pfn
)) {
8581 page
= pfn_to_page(pfn
);
8583 * The HWPoisoned page may be not in buddy system, and
8584 * page_count() is not 0.
8586 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8588 SetPageReserved(page
);
8593 BUG_ON(page_count(page
));
8594 BUG_ON(!PageBuddy(page
));
8595 order
= page_order(page
);
8596 offlined_pages
+= 1 << order
;
8597 #ifdef CONFIG_DEBUG_VM
8598 pr_info("remove from free list %lx %d %lx\n",
8599 pfn
, 1 << order
, end_pfn
);
8601 del_page_from_free_area(page
, &zone
->free_area
[order
]);
8602 for (i
= 0; i
< (1 << order
); i
++)
8603 SetPageReserved((page
+i
));
8604 pfn
+= (1 << order
);
8606 spin_unlock_irqrestore(&zone
->lock
, flags
);
8608 return offlined_pages
;
8612 bool is_free_buddy_page(struct page
*page
)
8614 struct zone
*zone
= page_zone(page
);
8615 unsigned long pfn
= page_to_pfn(page
);
8616 unsigned long flags
;
8619 spin_lock_irqsave(&zone
->lock
, flags
);
8620 for (order
= 0; order
< MAX_ORDER
; order
++) {
8621 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8623 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8626 spin_unlock_irqrestore(&zone
->lock
, flags
);
8628 return order
< MAX_ORDER
;
8631 #ifdef CONFIG_MEMORY_FAILURE
8633 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8634 * test is performed under the zone lock to prevent a race against page
8637 bool set_hwpoison_free_buddy_page(struct page
*page
)
8639 struct zone
*zone
= page_zone(page
);
8640 unsigned long pfn
= page_to_pfn(page
);
8641 unsigned long flags
;
8643 bool hwpoisoned
= false;
8645 spin_lock_irqsave(&zone
->lock
, flags
);
8646 for (order
= 0; order
< MAX_ORDER
; order
++) {
8647 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8649 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8650 if (!TestSetPageHWPoison(page
))
8655 spin_unlock_irqrestore(&zone
->lock
, flags
);