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 int watermark_boost_factor __read_mostly
;
319 int watermark_scale_factor
= 10;
321 static unsigned long nr_kernel_pages __initdata
;
322 static unsigned long nr_all_pages __initdata
;
323 static unsigned long dma_reserve __initdata
;
325 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
326 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
327 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
328 static unsigned long required_kernelcore __initdata
;
329 static unsigned long required_kernelcore_percent __initdata
;
330 static unsigned long required_movablecore __initdata
;
331 static unsigned long required_movablecore_percent __initdata
;
332 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
333 static bool mirrored_kernelcore __meminitdata
;
335 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
337 EXPORT_SYMBOL(movable_zone
);
338 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
341 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
342 unsigned int nr_online_nodes __read_mostly
= 1;
343 EXPORT_SYMBOL(nr_node_ids
);
344 EXPORT_SYMBOL(nr_online_nodes
);
347 int page_group_by_mobility_disabled __read_mostly
;
349 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
351 * During boot we initialize deferred pages on-demand, as needed, but once
352 * page_alloc_init_late() has finished, the deferred pages are all initialized,
353 * and we can permanently disable that path.
355 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
358 * Calling kasan_free_pages() only after deferred memory initialization
359 * has completed. Poisoning pages during deferred memory init will greatly
360 * lengthen the process and cause problem in large memory systems as the
361 * deferred pages initialization is done with interrupt disabled.
363 * Assuming that there will be no reference to those newly initialized
364 * pages before they are ever allocated, this should have no effect on
365 * KASAN memory tracking as the poison will be properly inserted at page
366 * allocation time. The only corner case is when pages are allocated by
367 * on-demand allocation and then freed again before the deferred pages
368 * initialization is done, but this is not likely to happen.
370 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
372 if (!static_branch_unlikely(&deferred_pages
))
373 kasan_free_pages(page
, order
);
376 /* Returns true if the struct page for the pfn is uninitialised */
377 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
379 int nid
= early_pfn_to_nid(pfn
);
381 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
388 * Returns true when the remaining initialisation should be deferred until
389 * later in the boot cycle when it can be parallelised.
391 static bool __meminit
392 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
394 static unsigned long prev_end_pfn
, nr_initialised
;
397 * prev_end_pfn static that contains the end of previous zone
398 * No need to protect because called very early in boot before smp_init.
400 if (prev_end_pfn
!= end_pfn
) {
401 prev_end_pfn
= end_pfn
;
405 /* Always populate low zones for address-constrained allocations */
406 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
410 * We start only with one section of pages, more pages are added as
411 * needed until the rest of deferred pages are initialized.
414 if ((nr_initialised
> PAGES_PER_SECTION
) &&
415 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
416 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
422 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
424 static inline bool early_page_uninitialised(unsigned long pfn
)
429 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
435 /* Return a pointer to the bitmap storing bits affecting a block of pages */
436 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
439 #ifdef CONFIG_SPARSEMEM
440 return section_to_usemap(__pfn_to_section(pfn
));
442 return page_zone(page
)->pageblock_flags
;
443 #endif /* CONFIG_SPARSEMEM */
446 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
448 #ifdef CONFIG_SPARSEMEM
449 pfn
&= (PAGES_PER_SECTION
-1);
450 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
452 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
453 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
454 #endif /* CONFIG_SPARSEMEM */
458 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
459 * @page: The page within the block of interest
460 * @pfn: The target page frame number
461 * @end_bitidx: The last bit of interest to retrieve
462 * @mask: mask of bits that the caller is interested in
464 * Return: pageblock_bits flags
466 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
468 unsigned long end_bitidx
,
471 unsigned long *bitmap
;
472 unsigned long bitidx
, word_bitidx
;
475 bitmap
= get_pageblock_bitmap(page
, pfn
);
476 bitidx
= pfn_to_bitidx(page
, pfn
);
477 word_bitidx
= bitidx
/ BITS_PER_LONG
;
478 bitidx
&= (BITS_PER_LONG
-1);
480 word
= bitmap
[word_bitidx
];
481 bitidx
+= end_bitidx
;
482 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
485 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
486 unsigned long end_bitidx
,
489 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
492 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
494 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
498 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
499 * @page: The page within the block of interest
500 * @flags: The flags to set
501 * @pfn: The target page frame number
502 * @end_bitidx: The last bit of interest
503 * @mask: mask of bits that the caller is interested in
505 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
507 unsigned long end_bitidx
,
510 unsigned long *bitmap
;
511 unsigned long bitidx
, word_bitidx
;
512 unsigned long old_word
, word
;
514 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
515 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
517 bitmap
= get_pageblock_bitmap(page
, pfn
);
518 bitidx
= pfn_to_bitidx(page
, pfn
);
519 word_bitidx
= bitidx
/ BITS_PER_LONG
;
520 bitidx
&= (BITS_PER_LONG
-1);
522 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
524 bitidx
+= end_bitidx
;
525 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
526 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
528 word
= READ_ONCE(bitmap
[word_bitidx
]);
530 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
531 if (word
== old_word
)
537 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
539 if (unlikely(page_group_by_mobility_disabled
&&
540 migratetype
< MIGRATE_PCPTYPES
))
541 migratetype
= MIGRATE_UNMOVABLE
;
543 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
544 PB_migrate
, PB_migrate_end
);
547 #ifdef CONFIG_DEBUG_VM
548 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
552 unsigned long pfn
= page_to_pfn(page
);
553 unsigned long sp
, start_pfn
;
556 seq
= zone_span_seqbegin(zone
);
557 start_pfn
= zone
->zone_start_pfn
;
558 sp
= zone
->spanned_pages
;
559 if (!zone_spans_pfn(zone
, pfn
))
561 } while (zone_span_seqretry(zone
, seq
));
564 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
565 pfn
, zone_to_nid(zone
), zone
->name
,
566 start_pfn
, start_pfn
+ sp
);
571 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
573 if (!pfn_valid_within(page_to_pfn(page
)))
575 if (zone
!= page_zone(page
))
581 * Temporary debugging check for pages not lying within a given zone.
583 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
585 if (page_outside_zone_boundaries(zone
, page
))
587 if (!page_is_consistent(zone
, page
))
593 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
599 static void bad_page(struct page
*page
, const char *reason
,
600 unsigned long bad_flags
)
602 static unsigned long resume
;
603 static unsigned long nr_shown
;
604 static unsigned long nr_unshown
;
607 * Allow a burst of 60 reports, then keep quiet for that minute;
608 * or allow a steady drip of one report per second.
610 if (nr_shown
== 60) {
611 if (time_before(jiffies
, resume
)) {
617 "BUG: Bad page state: %lu messages suppressed\n",
624 resume
= jiffies
+ 60 * HZ
;
626 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
627 current
->comm
, page_to_pfn(page
));
628 __dump_page(page
, reason
);
629 bad_flags
&= page
->flags
;
631 pr_alert("bad because of flags: %#lx(%pGp)\n",
632 bad_flags
, &bad_flags
);
633 dump_page_owner(page
);
638 /* Leave bad fields for debug, except PageBuddy could make trouble */
639 page_mapcount_reset(page
); /* remove PageBuddy */
640 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
644 * Higher-order pages are called "compound pages". They are structured thusly:
646 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
648 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
649 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
651 * The first tail page's ->compound_dtor holds the offset in array of compound
652 * page destructors. See compound_page_dtors.
654 * The first tail page's ->compound_order holds the order of allocation.
655 * This usage means that zero-order pages may not be compound.
658 void free_compound_page(struct page
*page
)
660 mem_cgroup_uncharge(page
);
661 __free_pages_ok(page
, compound_order(page
));
664 void prep_compound_page(struct page
*page
, unsigned int order
)
667 int nr_pages
= 1 << order
;
669 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
670 set_compound_order(page
, order
);
672 for (i
= 1; i
< nr_pages
; i
++) {
673 struct page
*p
= page
+ i
;
674 set_page_count(p
, 0);
675 p
->mapping
= TAIL_MAPPING
;
676 set_compound_head(p
, page
);
678 atomic_set(compound_mapcount_ptr(page
), -1);
681 #ifdef CONFIG_DEBUG_PAGEALLOC
682 unsigned int _debug_guardpage_minorder
;
684 bool _debug_pagealloc_enabled_early __read_mostly
685 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
686 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
687 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
688 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
690 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
692 static int __init
early_debug_pagealloc(char *buf
)
694 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
696 early_param("debug_pagealloc", early_debug_pagealloc
);
698 void init_debug_pagealloc(void)
700 if (!debug_pagealloc_enabled())
703 static_branch_enable(&_debug_pagealloc_enabled
);
705 if (!debug_guardpage_minorder())
708 static_branch_enable(&_debug_guardpage_enabled
);
711 static int __init
debug_guardpage_minorder_setup(char *buf
)
715 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
716 pr_err("Bad debug_guardpage_minorder value\n");
719 _debug_guardpage_minorder
= res
;
720 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
723 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
725 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
726 unsigned int order
, int migratetype
)
728 if (!debug_guardpage_enabled())
731 if (order
>= debug_guardpage_minorder())
734 __SetPageGuard(page
);
735 INIT_LIST_HEAD(&page
->lru
);
736 set_page_private(page
, order
);
737 /* Guard pages are not available for any usage */
738 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
743 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
744 unsigned int order
, int migratetype
)
746 if (!debug_guardpage_enabled())
749 __ClearPageGuard(page
);
751 set_page_private(page
, 0);
752 if (!is_migrate_isolate(migratetype
))
753 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
756 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
757 unsigned int order
, int migratetype
) { return false; }
758 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
759 unsigned int order
, int migratetype
) {}
762 static inline void set_page_order(struct page
*page
, unsigned int order
)
764 set_page_private(page
, order
);
765 __SetPageBuddy(page
);
769 * This function checks whether a page is free && is the buddy
770 * we can coalesce a page and its buddy if
771 * (a) the buddy is not in a hole (check before calling!) &&
772 * (b) the buddy is in the buddy system &&
773 * (c) a page and its buddy have the same order &&
774 * (d) a page and its buddy are in the same zone.
776 * For recording whether a page is in the buddy system, we set PageBuddy.
777 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
779 * For recording page's order, we use page_private(page).
781 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
784 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
785 if (page_zone_id(page
) != page_zone_id(buddy
))
788 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
793 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
795 * zone check is done late to avoid uselessly
796 * calculating zone/node ids for pages that could
799 if (page_zone_id(page
) != page_zone_id(buddy
))
802 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
809 #ifdef CONFIG_COMPACTION
810 static inline struct capture_control
*task_capc(struct zone
*zone
)
812 struct capture_control
*capc
= current
->capture_control
;
815 !(current
->flags
& PF_KTHREAD
) &&
817 capc
->cc
->zone
== zone
&&
818 capc
->cc
->direct_compaction
? capc
: NULL
;
822 compaction_capture(struct capture_control
*capc
, struct page
*page
,
823 int order
, int migratetype
)
825 if (!capc
|| order
!= capc
->cc
->order
)
828 /* Do not accidentally pollute CMA or isolated regions*/
829 if (is_migrate_cma(migratetype
) ||
830 is_migrate_isolate(migratetype
))
834 * Do not let lower order allocations polluate a movable pageblock.
835 * This might let an unmovable request use a reclaimable pageblock
836 * and vice-versa but no more than normal fallback logic which can
837 * have trouble finding a high-order free page.
839 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
847 static inline struct capture_control
*task_capc(struct zone
*zone
)
853 compaction_capture(struct capture_control
*capc
, struct page
*page
,
854 int order
, int migratetype
)
858 #endif /* CONFIG_COMPACTION */
861 * Freeing function for a buddy system allocator.
863 * The concept of a buddy system is to maintain direct-mapped table
864 * (containing bit values) for memory blocks of various "orders".
865 * The bottom level table contains the map for the smallest allocatable
866 * units of memory (here, pages), and each level above it describes
867 * pairs of units from the levels below, hence, "buddies".
868 * At a high level, all that happens here is marking the table entry
869 * at the bottom level available, and propagating the changes upward
870 * as necessary, plus some accounting needed to play nicely with other
871 * parts of the VM system.
872 * At each level, we keep a list of pages, which are heads of continuous
873 * free pages of length of (1 << order) and marked with PageBuddy.
874 * Page's order is recorded in page_private(page) field.
875 * So when we are allocating or freeing one, we can derive the state of the
876 * other. That is, if we allocate a small block, and both were
877 * free, the remainder of the region must be split into blocks.
878 * If a block is freed, and its buddy is also free, then this
879 * triggers coalescing into a block of larger size.
884 static inline void __free_one_page(struct page
*page
,
886 struct zone
*zone
, unsigned int order
,
889 unsigned long combined_pfn
;
890 unsigned long uninitialized_var(buddy_pfn
);
892 unsigned int max_order
;
893 struct capture_control
*capc
= task_capc(zone
);
895 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
897 VM_BUG_ON(!zone_is_initialized(zone
));
898 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
900 VM_BUG_ON(migratetype
== -1);
901 if (likely(!is_migrate_isolate(migratetype
)))
902 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
904 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
905 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
908 while (order
< max_order
- 1) {
909 if (compaction_capture(capc
, page
, order
, migratetype
)) {
910 __mod_zone_freepage_state(zone
, -(1 << order
),
914 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
915 buddy
= page
+ (buddy_pfn
- pfn
);
917 if (!pfn_valid_within(buddy_pfn
))
919 if (!page_is_buddy(page
, buddy
, order
))
922 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
923 * merge with it and move up one order.
925 if (page_is_guard(buddy
))
926 clear_page_guard(zone
, buddy
, order
, migratetype
);
928 del_page_from_free_area(buddy
, &zone
->free_area
[order
]);
929 combined_pfn
= buddy_pfn
& pfn
;
930 page
= page
+ (combined_pfn
- pfn
);
934 if (max_order
< MAX_ORDER
) {
935 /* If we are here, it means order is >= pageblock_order.
936 * We want to prevent merge between freepages on isolate
937 * pageblock and normal pageblock. Without this, pageblock
938 * isolation could cause incorrect freepage or CMA accounting.
940 * We don't want to hit this code for the more frequent
943 if (unlikely(has_isolate_pageblock(zone
))) {
946 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
947 buddy
= page
+ (buddy_pfn
- pfn
);
948 buddy_mt
= get_pageblock_migratetype(buddy
);
950 if (migratetype
!= buddy_mt
951 && (is_migrate_isolate(migratetype
) ||
952 is_migrate_isolate(buddy_mt
)))
956 goto continue_merging
;
960 set_page_order(page
, order
);
963 * If this is not the largest possible page, check if the buddy
964 * of the next-highest order is free. If it is, it's possible
965 * that pages are being freed that will coalesce soon. In case,
966 * that is happening, add the free page to the tail of the list
967 * so it's less likely to be used soon and more likely to be merged
968 * as a higher order page
970 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)
971 && !is_shuffle_order(order
)) {
972 struct page
*higher_page
, *higher_buddy
;
973 combined_pfn
= buddy_pfn
& pfn
;
974 higher_page
= page
+ (combined_pfn
- pfn
);
975 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
976 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
977 if (pfn_valid_within(buddy_pfn
) &&
978 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
979 add_to_free_area_tail(page
, &zone
->free_area
[order
],
985 if (is_shuffle_order(order
))
986 add_to_free_area_random(page
, &zone
->free_area
[order
],
989 add_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
994 * A bad page could be due to a number of fields. Instead of multiple branches,
995 * try and check multiple fields with one check. The caller must do a detailed
996 * check if necessary.
998 static inline bool page_expected_state(struct page
*page
,
999 unsigned long check_flags
)
1001 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1004 if (unlikely((unsigned long)page
->mapping
|
1005 page_ref_count(page
) |
1007 (unsigned long)page
->mem_cgroup
|
1009 (page
->flags
& check_flags
)))
1015 static void free_pages_check_bad(struct page
*page
)
1017 const char *bad_reason
;
1018 unsigned long bad_flags
;
1023 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1024 bad_reason
= "nonzero mapcount";
1025 if (unlikely(page
->mapping
!= NULL
))
1026 bad_reason
= "non-NULL mapping";
1027 if (unlikely(page_ref_count(page
) != 0))
1028 bad_reason
= "nonzero _refcount";
1029 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
1030 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1031 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
1034 if (unlikely(page
->mem_cgroup
))
1035 bad_reason
= "page still charged to cgroup";
1037 bad_page(page
, bad_reason
, bad_flags
);
1040 static inline int free_pages_check(struct page
*page
)
1042 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1045 /* Something has gone sideways, find it */
1046 free_pages_check_bad(page
);
1050 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1055 * We rely page->lru.next never has bit 0 set, unless the page
1056 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1058 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1060 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1064 switch (page
- head_page
) {
1066 /* the first tail page: ->mapping may be compound_mapcount() */
1067 if (unlikely(compound_mapcount(page
))) {
1068 bad_page(page
, "nonzero compound_mapcount", 0);
1074 * the second tail page: ->mapping is
1075 * deferred_list.next -- ignore value.
1079 if (page
->mapping
!= TAIL_MAPPING
) {
1080 bad_page(page
, "corrupted mapping in tail page", 0);
1085 if (unlikely(!PageTail(page
))) {
1086 bad_page(page
, "PageTail not set", 0);
1089 if (unlikely(compound_head(page
) != head_page
)) {
1090 bad_page(page
, "compound_head not consistent", 0);
1095 page
->mapping
= NULL
;
1096 clear_compound_head(page
);
1100 static void kernel_init_free_pages(struct page
*page
, int numpages
)
1104 for (i
= 0; i
< numpages
; i
++)
1105 clear_highpage(page
+ i
);
1108 static __always_inline
bool free_pages_prepare(struct page
*page
,
1109 unsigned int order
, bool check_free
)
1113 VM_BUG_ON_PAGE(PageTail(page
), page
);
1115 trace_mm_page_free(page
, order
);
1118 * Check tail pages before head page information is cleared to
1119 * avoid checking PageCompound for order-0 pages.
1121 if (unlikely(order
)) {
1122 bool compound
= PageCompound(page
);
1125 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1128 ClearPageDoubleMap(page
);
1129 for (i
= 1; i
< (1 << order
); i
++) {
1131 bad
+= free_tail_pages_check(page
, page
+ i
);
1132 if (unlikely(free_pages_check(page
+ i
))) {
1136 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1139 if (PageMappingFlags(page
))
1140 page
->mapping
= NULL
;
1141 if (memcg_kmem_enabled() && PageKmemcg(page
))
1142 __memcg_kmem_uncharge(page
, order
);
1144 bad
+= free_pages_check(page
);
1148 page_cpupid_reset_last(page
);
1149 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1150 reset_page_owner(page
, order
);
1152 if (!PageHighMem(page
)) {
1153 debug_check_no_locks_freed(page_address(page
),
1154 PAGE_SIZE
<< order
);
1155 debug_check_no_obj_freed(page_address(page
),
1156 PAGE_SIZE
<< order
);
1158 if (want_init_on_free())
1159 kernel_init_free_pages(page
, 1 << order
);
1161 kernel_poison_pages(page
, 1 << order
, 0);
1163 * arch_free_page() can make the page's contents inaccessible. s390
1164 * does this. So nothing which can access the page's contents should
1165 * happen after this.
1167 arch_free_page(page
, order
);
1169 if (debug_pagealloc_enabled_static())
1170 kernel_map_pages(page
, 1 << order
, 0);
1172 kasan_free_nondeferred_pages(page
, order
);
1177 #ifdef CONFIG_DEBUG_VM
1179 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1180 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1181 * moved from pcp lists to free lists.
1183 static bool free_pcp_prepare(struct page
*page
)
1185 return free_pages_prepare(page
, 0, true);
1188 static bool bulkfree_pcp_prepare(struct page
*page
)
1190 if (debug_pagealloc_enabled_static())
1191 return free_pages_check(page
);
1197 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1198 * moving from pcp lists to free list in order to reduce overhead. With
1199 * debug_pagealloc enabled, they are checked also immediately when being freed
1202 static bool free_pcp_prepare(struct page
*page
)
1204 if (debug_pagealloc_enabled_static())
1205 return free_pages_prepare(page
, 0, true);
1207 return free_pages_prepare(page
, 0, false);
1210 static bool bulkfree_pcp_prepare(struct page
*page
)
1212 return free_pages_check(page
);
1214 #endif /* CONFIG_DEBUG_VM */
1216 static inline void prefetch_buddy(struct page
*page
)
1218 unsigned long pfn
= page_to_pfn(page
);
1219 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1220 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1226 * Frees a number of pages from the PCP lists
1227 * Assumes all pages on list are in same zone, and of same order.
1228 * count is the number of pages to free.
1230 * If the zone was previously in an "all pages pinned" state then look to
1231 * see if this freeing clears that state.
1233 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1234 * pinned" detection logic.
1236 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1237 struct per_cpu_pages
*pcp
)
1239 int migratetype
= 0;
1241 int prefetch_nr
= 0;
1242 bool isolated_pageblocks
;
1243 struct page
*page
, *tmp
;
1247 * Ensure proper count is passed which otherwise would stuck in the
1248 * below while (list_empty(list)) loop.
1250 count
= min(pcp
->count
, count
);
1252 struct list_head
*list
;
1255 * Remove pages from lists in a round-robin fashion. A
1256 * batch_free count is maintained that is incremented when an
1257 * empty list is encountered. This is so more pages are freed
1258 * off fuller lists instead of spinning excessively around empty
1263 if (++migratetype
== MIGRATE_PCPTYPES
)
1265 list
= &pcp
->lists
[migratetype
];
1266 } while (list_empty(list
));
1268 /* This is the only non-empty list. Free them all. */
1269 if (batch_free
== MIGRATE_PCPTYPES
)
1273 page
= list_last_entry(list
, struct page
, lru
);
1274 /* must delete to avoid corrupting pcp list */
1275 list_del(&page
->lru
);
1278 if (bulkfree_pcp_prepare(page
))
1281 list_add_tail(&page
->lru
, &head
);
1284 * We are going to put the page back to the global
1285 * pool, prefetch its buddy to speed up later access
1286 * under zone->lock. It is believed the overhead of
1287 * an additional test and calculating buddy_pfn here
1288 * can be offset by reduced memory latency later. To
1289 * avoid excessive prefetching due to large count, only
1290 * prefetch buddy for the first pcp->batch nr of pages.
1292 if (prefetch_nr
++ < pcp
->batch
)
1293 prefetch_buddy(page
);
1294 } while (--count
&& --batch_free
&& !list_empty(list
));
1297 spin_lock(&zone
->lock
);
1298 isolated_pageblocks
= has_isolate_pageblock(zone
);
1301 * Use safe version since after __free_one_page(),
1302 * page->lru.next will not point to original list.
1304 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1305 int mt
= get_pcppage_migratetype(page
);
1306 /* MIGRATE_ISOLATE page should not go to pcplists */
1307 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1308 /* Pageblock could have been isolated meanwhile */
1309 if (unlikely(isolated_pageblocks
))
1310 mt
= get_pageblock_migratetype(page
);
1312 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1313 trace_mm_page_pcpu_drain(page
, 0, mt
);
1315 spin_unlock(&zone
->lock
);
1318 static void free_one_page(struct zone
*zone
,
1319 struct page
*page
, unsigned long pfn
,
1323 spin_lock(&zone
->lock
);
1324 if (unlikely(has_isolate_pageblock(zone
) ||
1325 is_migrate_isolate(migratetype
))) {
1326 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1328 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1329 spin_unlock(&zone
->lock
);
1332 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1333 unsigned long zone
, int nid
)
1335 mm_zero_struct_page(page
);
1336 set_page_links(page
, zone
, nid
, pfn
);
1337 init_page_count(page
);
1338 page_mapcount_reset(page
);
1339 page_cpupid_reset_last(page
);
1340 page_kasan_tag_reset(page
);
1342 INIT_LIST_HEAD(&page
->lru
);
1343 #ifdef WANT_PAGE_VIRTUAL
1344 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1345 if (!is_highmem_idx(zone
))
1346 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1350 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1351 static void __meminit
init_reserved_page(unsigned long pfn
)
1356 if (!early_page_uninitialised(pfn
))
1359 nid
= early_pfn_to_nid(pfn
);
1360 pgdat
= NODE_DATA(nid
);
1362 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1363 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1365 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1368 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1371 static inline void init_reserved_page(unsigned long pfn
)
1374 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1377 * Initialised pages do not have PageReserved set. This function is
1378 * called for each range allocated by the bootmem allocator and
1379 * marks the pages PageReserved. The remaining valid pages are later
1380 * sent to the buddy page allocator.
1382 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1384 unsigned long start_pfn
= PFN_DOWN(start
);
1385 unsigned long end_pfn
= PFN_UP(end
);
1387 for (; start_pfn
< end_pfn
; start_pfn
++) {
1388 if (pfn_valid(start_pfn
)) {
1389 struct page
*page
= pfn_to_page(start_pfn
);
1391 init_reserved_page(start_pfn
);
1393 /* Avoid false-positive PageTail() */
1394 INIT_LIST_HEAD(&page
->lru
);
1397 * no need for atomic set_bit because the struct
1398 * page is not visible yet so nobody should
1401 __SetPageReserved(page
);
1406 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1408 unsigned long flags
;
1410 unsigned long pfn
= page_to_pfn(page
);
1412 if (!free_pages_prepare(page
, order
, true))
1415 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1416 local_irq_save(flags
);
1417 __count_vm_events(PGFREE
, 1 << order
);
1418 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1419 local_irq_restore(flags
);
1422 void __free_pages_core(struct page
*page
, unsigned int order
)
1424 unsigned int nr_pages
= 1 << order
;
1425 struct page
*p
= page
;
1429 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1431 __ClearPageReserved(p
);
1432 set_page_count(p
, 0);
1434 __ClearPageReserved(p
);
1435 set_page_count(p
, 0);
1437 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1438 set_page_refcounted(page
);
1439 __free_pages(page
, order
);
1442 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1443 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1445 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1447 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1449 static DEFINE_SPINLOCK(early_pfn_lock
);
1452 spin_lock(&early_pfn_lock
);
1453 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1455 nid
= first_online_node
;
1456 spin_unlock(&early_pfn_lock
);
1462 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1463 /* Only safe to use early in boot when initialisation is single-threaded */
1464 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1468 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1469 if (nid
>= 0 && nid
!= node
)
1475 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1482 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1485 if (early_page_uninitialised(pfn
))
1487 __free_pages_core(page
, order
);
1491 * Check that the whole (or subset of) a pageblock given by the interval of
1492 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1493 * with the migration of free compaction scanner. The scanners then need to
1494 * use only pfn_valid_within() check for arches that allow holes within
1497 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1499 * It's possible on some configurations to have a setup like node0 node1 node0
1500 * i.e. it's possible that all pages within a zones range of pages do not
1501 * belong to a single zone. We assume that a border between node0 and node1
1502 * can occur within a single pageblock, but not a node0 node1 node0
1503 * interleaving within a single pageblock. It is therefore sufficient to check
1504 * the first and last page of a pageblock and avoid checking each individual
1505 * page in a pageblock.
1507 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1508 unsigned long end_pfn
, struct zone
*zone
)
1510 struct page
*start_page
;
1511 struct page
*end_page
;
1513 /* end_pfn is one past the range we are checking */
1516 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1519 start_page
= pfn_to_online_page(start_pfn
);
1523 if (page_zone(start_page
) != zone
)
1526 end_page
= pfn_to_page(end_pfn
);
1528 /* This gives a shorter code than deriving page_zone(end_page) */
1529 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1535 void set_zone_contiguous(struct zone
*zone
)
1537 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1538 unsigned long block_end_pfn
;
1540 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1541 for (; block_start_pfn
< zone_end_pfn(zone
);
1542 block_start_pfn
= block_end_pfn
,
1543 block_end_pfn
+= pageblock_nr_pages
) {
1545 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1547 if (!__pageblock_pfn_to_page(block_start_pfn
,
1548 block_end_pfn
, zone
))
1553 /* We confirm that there is no hole */
1554 zone
->contiguous
= true;
1557 void clear_zone_contiguous(struct zone
*zone
)
1559 zone
->contiguous
= false;
1562 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1563 static void __init
deferred_free_range(unsigned long pfn
,
1564 unsigned long nr_pages
)
1572 page
= pfn_to_page(pfn
);
1574 /* Free a large naturally-aligned chunk if possible */
1575 if (nr_pages
== pageblock_nr_pages
&&
1576 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1577 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1578 __free_pages_core(page
, pageblock_order
);
1582 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1583 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1584 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1585 __free_pages_core(page
, 0);
1589 /* Completion tracking for deferred_init_memmap() threads */
1590 static atomic_t pgdat_init_n_undone __initdata
;
1591 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1593 static inline void __init
pgdat_init_report_one_done(void)
1595 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1596 complete(&pgdat_init_all_done_comp
);
1600 * Returns true if page needs to be initialized or freed to buddy allocator.
1602 * First we check if pfn is valid on architectures where it is possible to have
1603 * holes within pageblock_nr_pages. On systems where it is not possible, this
1604 * function is optimized out.
1606 * Then, we check if a current large page is valid by only checking the validity
1609 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1611 if (!pfn_valid_within(pfn
))
1613 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1619 * Free pages to buddy allocator. Try to free aligned pages in
1620 * pageblock_nr_pages sizes.
1622 static void __init
deferred_free_pages(unsigned long pfn
,
1623 unsigned long end_pfn
)
1625 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1626 unsigned long nr_free
= 0;
1628 for (; pfn
< end_pfn
; pfn
++) {
1629 if (!deferred_pfn_valid(pfn
)) {
1630 deferred_free_range(pfn
- nr_free
, nr_free
);
1632 } else if (!(pfn
& nr_pgmask
)) {
1633 deferred_free_range(pfn
- nr_free
, nr_free
);
1639 /* Free the last block of pages to allocator */
1640 deferred_free_range(pfn
- nr_free
, nr_free
);
1644 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1645 * by performing it only once every pageblock_nr_pages.
1646 * Return number of pages initialized.
1648 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1650 unsigned long end_pfn
)
1652 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1653 int nid
= zone_to_nid(zone
);
1654 unsigned long nr_pages
= 0;
1655 int zid
= zone_idx(zone
);
1656 struct page
*page
= NULL
;
1658 for (; pfn
< end_pfn
; pfn
++) {
1659 if (!deferred_pfn_valid(pfn
)) {
1662 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1663 page
= pfn_to_page(pfn
);
1667 __init_single_page(page
, pfn
, zid
, nid
);
1674 * This function is meant to pre-load the iterator for the zone init.
1675 * Specifically it walks through the ranges until we are caught up to the
1676 * first_init_pfn value and exits there. If we never encounter the value we
1677 * return false indicating there are no valid ranges left.
1680 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1681 unsigned long *spfn
, unsigned long *epfn
,
1682 unsigned long first_init_pfn
)
1687 * Start out by walking through the ranges in this zone that have
1688 * already been initialized. We don't need to do anything with them
1689 * so we just need to flush them out of the system.
1691 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1692 if (*epfn
<= first_init_pfn
)
1694 if (*spfn
< first_init_pfn
)
1695 *spfn
= first_init_pfn
;
1704 * Initialize and free pages. We do it in two loops: first we initialize
1705 * struct page, then free to buddy allocator, because while we are
1706 * freeing pages we can access pages that are ahead (computing buddy
1707 * page in __free_one_page()).
1709 * In order to try and keep some memory in the cache we have the loop
1710 * broken along max page order boundaries. This way we will not cause
1711 * any issues with the buddy page computation.
1713 static unsigned long __init
1714 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1715 unsigned long *end_pfn
)
1717 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1718 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1719 unsigned long nr_pages
= 0;
1722 /* First we loop through and initialize the page values */
1723 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1726 if (mo_pfn
<= *start_pfn
)
1729 t
= min(mo_pfn
, *end_pfn
);
1730 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1732 if (mo_pfn
< *end_pfn
) {
1733 *start_pfn
= mo_pfn
;
1738 /* Reset values and now loop through freeing pages as needed */
1741 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1747 t
= min(mo_pfn
, epfn
);
1748 deferred_free_pages(spfn
, t
);
1757 /* Initialise remaining memory on a node */
1758 static int __init
deferred_init_memmap(void *data
)
1760 pg_data_t
*pgdat
= data
;
1761 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1762 unsigned long spfn
= 0, epfn
= 0, nr_pages
= 0;
1763 unsigned long first_init_pfn
, flags
;
1764 unsigned long start
= jiffies
;
1769 /* Bind memory initialisation thread to a local node if possible */
1770 if (!cpumask_empty(cpumask
))
1771 set_cpus_allowed_ptr(current
, cpumask
);
1773 pgdat_resize_lock(pgdat
, &flags
);
1774 first_init_pfn
= pgdat
->first_deferred_pfn
;
1775 if (first_init_pfn
== ULONG_MAX
) {
1776 pgdat_resize_unlock(pgdat
, &flags
);
1777 pgdat_init_report_one_done();
1781 /* Sanity check boundaries */
1782 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1783 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1784 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1787 * Once we unlock here, the zone cannot be grown anymore, thus if an
1788 * interrupt thread must allocate this early in boot, zone must be
1789 * pre-grown prior to start of deferred page initialization.
1791 pgdat_resize_unlock(pgdat
, &flags
);
1793 /* Only the highest zone is deferred so find it */
1794 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1795 zone
= pgdat
->node_zones
+ zid
;
1796 if (first_init_pfn
< zone_end_pfn(zone
))
1800 /* If the zone is empty somebody else may have cleared out the zone */
1801 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1806 * Initialize and free pages in MAX_ORDER sized increments so
1807 * that we can avoid introducing any issues with the buddy
1810 while (spfn
< epfn
) {
1811 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1815 /* Sanity check that the next zone really is unpopulated */
1816 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1818 pr_info("node %d initialised, %lu pages in %ums\n",
1819 pgdat
->node_id
, nr_pages
, jiffies_to_msecs(jiffies
- start
));
1821 pgdat_init_report_one_done();
1826 * If this zone has deferred pages, try to grow it by initializing enough
1827 * deferred pages to satisfy the allocation specified by order, rounded up to
1828 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1829 * of SECTION_SIZE bytes by initializing struct pages in increments of
1830 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1832 * Return true when zone was grown, otherwise return false. We return true even
1833 * when we grow less than requested, to let the caller decide if there are
1834 * enough pages to satisfy the allocation.
1836 * Note: We use noinline because this function is needed only during boot, and
1837 * it is called from a __ref function _deferred_grow_zone. This way we are
1838 * making sure that it is not inlined into permanent text section.
1840 static noinline
bool __init
1841 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1843 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1844 pg_data_t
*pgdat
= zone
->zone_pgdat
;
1845 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1846 unsigned long spfn
, epfn
, flags
;
1847 unsigned long nr_pages
= 0;
1850 /* Only the last zone may have deferred pages */
1851 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1854 pgdat_resize_lock(pgdat
, &flags
);
1857 * If someone grew this zone while we were waiting for spinlock, return
1858 * true, as there might be enough pages already.
1860 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1861 pgdat_resize_unlock(pgdat
, &flags
);
1865 /* If the zone is empty somebody else may have cleared out the zone */
1866 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1867 first_deferred_pfn
)) {
1868 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1869 pgdat_resize_unlock(pgdat
, &flags
);
1870 /* Retry only once. */
1871 return first_deferred_pfn
!= ULONG_MAX
;
1875 * Initialize and free pages in MAX_ORDER sized increments so
1876 * that we can avoid introducing any issues with the buddy
1879 while (spfn
< epfn
) {
1880 /* update our first deferred PFN for this section */
1881 first_deferred_pfn
= spfn
;
1883 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1884 touch_nmi_watchdog();
1886 /* We should only stop along section boundaries */
1887 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
1890 /* If our quota has been met we can stop here */
1891 if (nr_pages
>= nr_pages_needed
)
1895 pgdat
->first_deferred_pfn
= spfn
;
1896 pgdat_resize_unlock(pgdat
, &flags
);
1898 return nr_pages
> 0;
1902 * deferred_grow_zone() is __init, but it is called from
1903 * get_page_from_freelist() during early boot until deferred_pages permanently
1904 * disables this call. This is why we have refdata wrapper to avoid warning,
1905 * and to ensure that the function body gets unloaded.
1908 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1910 return deferred_grow_zone(zone
, order
);
1913 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1915 void __init
page_alloc_init_late(void)
1920 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1922 /* There will be num_node_state(N_MEMORY) threads */
1923 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1924 for_each_node_state(nid
, N_MEMORY
) {
1925 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1928 /* Block until all are initialised */
1929 wait_for_completion(&pgdat_init_all_done_comp
);
1932 * The number of managed pages has changed due to the initialisation
1933 * so the pcpu batch and high limits needs to be updated or the limits
1934 * will be artificially small.
1936 for_each_populated_zone(zone
)
1937 zone_pcp_update(zone
);
1940 * We initialized the rest of the deferred pages. Permanently disable
1941 * on-demand struct page initialization.
1943 static_branch_disable(&deferred_pages
);
1945 /* Reinit limits that are based on free pages after the kernel is up */
1946 files_maxfiles_init();
1949 /* Discard memblock private memory */
1952 for_each_node_state(nid
, N_MEMORY
)
1953 shuffle_free_memory(NODE_DATA(nid
));
1955 for_each_populated_zone(zone
)
1956 set_zone_contiguous(zone
);
1960 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1961 void __init
init_cma_reserved_pageblock(struct page
*page
)
1963 unsigned i
= pageblock_nr_pages
;
1964 struct page
*p
= page
;
1967 __ClearPageReserved(p
);
1968 set_page_count(p
, 0);
1971 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1973 if (pageblock_order
>= MAX_ORDER
) {
1974 i
= pageblock_nr_pages
;
1977 set_page_refcounted(p
);
1978 __free_pages(p
, MAX_ORDER
- 1);
1979 p
+= MAX_ORDER_NR_PAGES
;
1980 } while (i
-= MAX_ORDER_NR_PAGES
);
1982 set_page_refcounted(page
);
1983 __free_pages(page
, pageblock_order
);
1986 adjust_managed_page_count(page
, pageblock_nr_pages
);
1991 * The order of subdivision here is critical for the IO subsystem.
1992 * Please do not alter this order without good reasons and regression
1993 * testing. Specifically, as large blocks of memory are subdivided,
1994 * the order in which smaller blocks are delivered depends on the order
1995 * they're subdivided in this function. This is the primary factor
1996 * influencing the order in which pages are delivered to the IO
1997 * subsystem according to empirical testing, and this is also justified
1998 * by considering the behavior of a buddy system containing a single
1999 * large block of memory acted on by a series of small allocations.
2000 * This behavior is a critical factor in sglist merging's success.
2004 static inline void expand(struct zone
*zone
, struct page
*page
,
2005 int low
, int high
, struct free_area
*area
,
2008 unsigned long size
= 1 << high
;
2010 while (high
> low
) {
2014 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2017 * Mark as guard pages (or page), that will allow to
2018 * merge back to allocator when buddy will be freed.
2019 * Corresponding page table entries will not be touched,
2020 * pages will stay not present in virtual address space
2022 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2025 add_to_free_area(&page
[size
], area
, migratetype
);
2026 set_page_order(&page
[size
], high
);
2030 static void check_new_page_bad(struct page
*page
)
2032 const char *bad_reason
= NULL
;
2033 unsigned long bad_flags
= 0;
2035 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
2036 bad_reason
= "nonzero mapcount";
2037 if (unlikely(page
->mapping
!= NULL
))
2038 bad_reason
= "non-NULL mapping";
2039 if (unlikely(page_ref_count(page
) != 0))
2040 bad_reason
= "nonzero _refcount";
2041 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2042 bad_reason
= "HWPoisoned (hardware-corrupted)";
2043 bad_flags
= __PG_HWPOISON
;
2044 /* Don't complain about hwpoisoned pages */
2045 page_mapcount_reset(page
); /* remove PageBuddy */
2048 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
2049 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
2050 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
2053 if (unlikely(page
->mem_cgroup
))
2054 bad_reason
= "page still charged to cgroup";
2056 bad_page(page
, bad_reason
, bad_flags
);
2060 * This page is about to be returned from the page allocator
2062 static inline int check_new_page(struct page
*page
)
2064 if (likely(page_expected_state(page
,
2065 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2068 check_new_page_bad(page
);
2072 static inline bool free_pages_prezeroed(void)
2074 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
2075 page_poisoning_enabled()) || want_init_on_free();
2078 #ifdef CONFIG_DEBUG_VM
2080 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2081 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2082 * also checked when pcp lists are refilled from the free lists.
2084 static inline bool check_pcp_refill(struct page
*page
)
2086 if (debug_pagealloc_enabled_static())
2087 return check_new_page(page
);
2092 static inline bool check_new_pcp(struct page
*page
)
2094 return check_new_page(page
);
2098 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2099 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2100 * enabled, they are also checked when being allocated from the pcp lists.
2102 static inline bool check_pcp_refill(struct page
*page
)
2104 return check_new_page(page
);
2106 static inline bool check_new_pcp(struct page
*page
)
2108 if (debug_pagealloc_enabled_static())
2109 return check_new_page(page
);
2113 #endif /* CONFIG_DEBUG_VM */
2115 static bool check_new_pages(struct page
*page
, unsigned int order
)
2118 for (i
= 0; i
< (1 << order
); i
++) {
2119 struct page
*p
= page
+ i
;
2121 if (unlikely(check_new_page(p
)))
2128 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2131 set_page_private(page
, 0);
2132 set_page_refcounted(page
);
2134 arch_alloc_page(page
, order
);
2135 if (debug_pagealloc_enabled_static())
2136 kernel_map_pages(page
, 1 << order
, 1);
2137 kasan_alloc_pages(page
, order
);
2138 kernel_poison_pages(page
, 1 << order
, 1);
2139 set_page_owner(page
, order
, gfp_flags
);
2142 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2143 unsigned int alloc_flags
)
2145 post_alloc_hook(page
, order
, gfp_flags
);
2147 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags
))
2148 kernel_init_free_pages(page
, 1 << order
);
2150 if (order
&& (gfp_flags
& __GFP_COMP
))
2151 prep_compound_page(page
, order
);
2154 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2155 * allocate the page. The expectation is that the caller is taking
2156 * steps that will free more memory. The caller should avoid the page
2157 * being used for !PFMEMALLOC purposes.
2159 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2160 set_page_pfmemalloc(page
);
2162 clear_page_pfmemalloc(page
);
2166 * Go through the free lists for the given migratetype and remove
2167 * the smallest available page from the freelists
2169 static __always_inline
2170 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2173 unsigned int current_order
;
2174 struct free_area
*area
;
2177 /* Find a page of the appropriate size in the preferred list */
2178 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2179 area
= &(zone
->free_area
[current_order
]);
2180 page
= get_page_from_free_area(area
, migratetype
);
2183 del_page_from_free_area(page
, area
);
2184 expand(zone
, page
, order
, current_order
, area
, migratetype
);
2185 set_pcppage_migratetype(page
, migratetype
);
2194 * This array describes the order lists are fallen back to when
2195 * the free lists for the desirable migrate type are depleted
2197 static int fallbacks
[MIGRATE_TYPES
][4] = {
2198 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2199 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2200 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2202 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2204 #ifdef CONFIG_MEMORY_ISOLATION
2205 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2210 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2213 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2216 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2217 unsigned int order
) { return NULL
; }
2221 * Move the free pages in a range to the free lists of the requested type.
2222 * Note that start_page and end_pages are not aligned on a pageblock
2223 * boundary. If alignment is required, use move_freepages_block()
2225 static int move_freepages(struct zone
*zone
,
2226 struct page
*start_page
, struct page
*end_page
,
2227 int migratetype
, int *num_movable
)
2231 int pages_moved
= 0;
2233 for (page
= start_page
; page
<= end_page
;) {
2234 if (!pfn_valid_within(page_to_pfn(page
))) {
2239 if (!PageBuddy(page
)) {
2241 * We assume that pages that could be isolated for
2242 * migration are movable. But we don't actually try
2243 * isolating, as that would be expensive.
2246 (PageLRU(page
) || __PageMovable(page
)))
2253 /* Make sure we are not inadvertently changing nodes */
2254 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2255 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2257 order
= page_order(page
);
2258 move_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
2260 pages_moved
+= 1 << order
;
2266 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2267 int migratetype
, int *num_movable
)
2269 unsigned long start_pfn
, end_pfn
;
2270 struct page
*start_page
, *end_page
;
2275 start_pfn
= page_to_pfn(page
);
2276 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2277 start_page
= pfn_to_page(start_pfn
);
2278 end_page
= start_page
+ pageblock_nr_pages
- 1;
2279 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2281 /* Do not cross zone boundaries */
2282 if (!zone_spans_pfn(zone
, start_pfn
))
2284 if (!zone_spans_pfn(zone
, end_pfn
))
2287 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2291 static void change_pageblock_range(struct page
*pageblock_page
,
2292 int start_order
, int migratetype
)
2294 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2296 while (nr_pageblocks
--) {
2297 set_pageblock_migratetype(pageblock_page
, migratetype
);
2298 pageblock_page
+= pageblock_nr_pages
;
2303 * When we are falling back to another migratetype during allocation, try to
2304 * steal extra free pages from the same pageblocks to satisfy further
2305 * allocations, instead of polluting multiple pageblocks.
2307 * If we are stealing a relatively large buddy page, it is likely there will
2308 * be more free pages in the pageblock, so try to steal them all. For
2309 * reclaimable and unmovable allocations, we steal regardless of page size,
2310 * as fragmentation caused by those allocations polluting movable pageblocks
2311 * is worse than movable allocations stealing from unmovable and reclaimable
2314 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2317 * Leaving this order check is intended, although there is
2318 * relaxed order check in next check. The reason is that
2319 * we can actually steal whole pageblock if this condition met,
2320 * but, below check doesn't guarantee it and that is just heuristic
2321 * so could be changed anytime.
2323 if (order
>= pageblock_order
)
2326 if (order
>= pageblock_order
/ 2 ||
2327 start_mt
== MIGRATE_RECLAIMABLE
||
2328 start_mt
== MIGRATE_UNMOVABLE
||
2329 page_group_by_mobility_disabled
)
2335 static inline void boost_watermark(struct zone
*zone
)
2337 unsigned long max_boost
;
2339 if (!watermark_boost_factor
)
2342 * Don't bother in zones that are unlikely to produce results.
2343 * On small machines, including kdump capture kernels running
2344 * in a small area, boosting the watermark can cause an out of
2345 * memory situation immediately.
2347 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
2350 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2351 watermark_boost_factor
, 10000);
2354 * high watermark may be uninitialised if fragmentation occurs
2355 * very early in boot so do not boost. We do not fall
2356 * through and boost by pageblock_nr_pages as failing
2357 * allocations that early means that reclaim is not going
2358 * to help and it may even be impossible to reclaim the
2359 * boosted watermark resulting in a hang.
2364 max_boost
= max(pageblock_nr_pages
, max_boost
);
2366 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2371 * This function implements actual steal behaviour. If order is large enough,
2372 * we can steal whole pageblock. If not, we first move freepages in this
2373 * pageblock to our migratetype and determine how many already-allocated pages
2374 * are there in the pageblock with a compatible migratetype. If at least half
2375 * of pages are free or compatible, we can change migratetype of the pageblock
2376 * itself, so pages freed in the future will be put on the correct free list.
2378 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2379 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2381 unsigned int current_order
= page_order(page
);
2382 struct free_area
*area
;
2383 int free_pages
, movable_pages
, alike_pages
;
2386 old_block_type
= get_pageblock_migratetype(page
);
2389 * This can happen due to races and we want to prevent broken
2390 * highatomic accounting.
2392 if (is_migrate_highatomic(old_block_type
))
2395 /* Take ownership for orders >= pageblock_order */
2396 if (current_order
>= pageblock_order
) {
2397 change_pageblock_range(page
, current_order
, start_type
);
2402 * Boost watermarks to increase reclaim pressure to reduce the
2403 * likelihood of future fallbacks. Wake kswapd now as the node
2404 * may be balanced overall and kswapd will not wake naturally.
2406 boost_watermark(zone
);
2407 if (alloc_flags
& ALLOC_KSWAPD
)
2408 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2410 /* We are not allowed to try stealing from the whole block */
2414 free_pages
= move_freepages_block(zone
, page
, start_type
,
2417 * Determine how many pages are compatible with our allocation.
2418 * For movable allocation, it's the number of movable pages which
2419 * we just obtained. For other types it's a bit more tricky.
2421 if (start_type
== MIGRATE_MOVABLE
) {
2422 alike_pages
= movable_pages
;
2425 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2426 * to MOVABLE pageblock, consider all non-movable pages as
2427 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2428 * vice versa, be conservative since we can't distinguish the
2429 * exact migratetype of non-movable pages.
2431 if (old_block_type
== MIGRATE_MOVABLE
)
2432 alike_pages
= pageblock_nr_pages
2433 - (free_pages
+ movable_pages
);
2438 /* moving whole block can fail due to zone boundary conditions */
2443 * If a sufficient number of pages in the block are either free or of
2444 * comparable migratability as our allocation, claim the whole block.
2446 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2447 page_group_by_mobility_disabled
)
2448 set_pageblock_migratetype(page
, start_type
);
2453 area
= &zone
->free_area
[current_order
];
2454 move_to_free_area(page
, area
, start_type
);
2458 * Check whether there is a suitable fallback freepage with requested order.
2459 * If only_stealable is true, this function returns fallback_mt only if
2460 * we can steal other freepages all together. This would help to reduce
2461 * fragmentation due to mixed migratetype pages in one pageblock.
2463 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2464 int migratetype
, bool only_stealable
, bool *can_steal
)
2469 if (area
->nr_free
== 0)
2474 fallback_mt
= fallbacks
[migratetype
][i
];
2475 if (fallback_mt
== MIGRATE_TYPES
)
2478 if (free_area_empty(area
, fallback_mt
))
2481 if (can_steal_fallback(order
, migratetype
))
2484 if (!only_stealable
)
2495 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2496 * there are no empty page blocks that contain a page with a suitable order
2498 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2499 unsigned int alloc_order
)
2502 unsigned long max_managed
, flags
;
2505 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2506 * Check is race-prone but harmless.
2508 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2509 if (zone
->nr_reserved_highatomic
>= max_managed
)
2512 spin_lock_irqsave(&zone
->lock
, flags
);
2514 /* Recheck the nr_reserved_highatomic limit under the lock */
2515 if (zone
->nr_reserved_highatomic
>= max_managed
)
2519 mt
= get_pageblock_migratetype(page
);
2520 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2521 && !is_migrate_cma(mt
)) {
2522 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2523 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2524 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2528 spin_unlock_irqrestore(&zone
->lock
, flags
);
2532 * Used when an allocation is about to fail under memory pressure. This
2533 * potentially hurts the reliability of high-order allocations when under
2534 * intense memory pressure but failed atomic allocations should be easier
2535 * to recover from than an OOM.
2537 * If @force is true, try to unreserve a pageblock even though highatomic
2538 * pageblock is exhausted.
2540 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2543 struct zonelist
*zonelist
= ac
->zonelist
;
2544 unsigned long flags
;
2551 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2554 * Preserve at least one pageblock unless memory pressure
2557 if (!force
&& zone
->nr_reserved_highatomic
<=
2561 spin_lock_irqsave(&zone
->lock
, flags
);
2562 for (order
= 0; order
< MAX_ORDER
; order
++) {
2563 struct free_area
*area
= &(zone
->free_area
[order
]);
2565 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2570 * In page freeing path, migratetype change is racy so
2571 * we can counter several free pages in a pageblock
2572 * in this loop althoug we changed the pageblock type
2573 * from highatomic to ac->migratetype. So we should
2574 * adjust the count once.
2576 if (is_migrate_highatomic_page(page
)) {
2578 * It should never happen but changes to
2579 * locking could inadvertently allow a per-cpu
2580 * drain to add pages to MIGRATE_HIGHATOMIC
2581 * while unreserving so be safe and watch for
2584 zone
->nr_reserved_highatomic
-= min(
2586 zone
->nr_reserved_highatomic
);
2590 * Convert to ac->migratetype and avoid the normal
2591 * pageblock stealing heuristics. Minimally, the caller
2592 * is doing the work and needs the pages. More
2593 * importantly, if the block was always converted to
2594 * MIGRATE_UNMOVABLE or another type then the number
2595 * of pageblocks that cannot be completely freed
2598 set_pageblock_migratetype(page
, ac
->migratetype
);
2599 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2602 spin_unlock_irqrestore(&zone
->lock
, flags
);
2606 spin_unlock_irqrestore(&zone
->lock
, flags
);
2613 * Try finding a free buddy page on the fallback list and put it on the free
2614 * list of requested migratetype, possibly along with other pages from the same
2615 * block, depending on fragmentation avoidance heuristics. Returns true if
2616 * fallback was found so that __rmqueue_smallest() can grab it.
2618 * The use of signed ints for order and current_order is a deliberate
2619 * deviation from the rest of this file, to make the for loop
2620 * condition simpler.
2622 static __always_inline
bool
2623 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2624 unsigned int alloc_flags
)
2626 struct free_area
*area
;
2628 int min_order
= order
;
2634 * Do not steal pages from freelists belonging to other pageblocks
2635 * i.e. orders < pageblock_order. If there are no local zones free,
2636 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2638 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2639 min_order
= pageblock_order
;
2642 * Find the largest available free page in the other list. This roughly
2643 * approximates finding the pageblock with the most free pages, which
2644 * would be too costly to do exactly.
2646 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2648 area
= &(zone
->free_area
[current_order
]);
2649 fallback_mt
= find_suitable_fallback(area
, current_order
,
2650 start_migratetype
, false, &can_steal
);
2651 if (fallback_mt
== -1)
2655 * We cannot steal all free pages from the pageblock and the
2656 * requested migratetype is movable. In that case it's better to
2657 * steal and split the smallest available page instead of the
2658 * largest available page, because even if the next movable
2659 * allocation falls back into a different pageblock than this
2660 * one, it won't cause permanent fragmentation.
2662 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2663 && current_order
> order
)
2672 for (current_order
= order
; current_order
< MAX_ORDER
;
2674 area
= &(zone
->free_area
[current_order
]);
2675 fallback_mt
= find_suitable_fallback(area
, current_order
,
2676 start_migratetype
, false, &can_steal
);
2677 if (fallback_mt
!= -1)
2682 * This should not happen - we already found a suitable fallback
2683 * when looking for the largest page.
2685 VM_BUG_ON(current_order
== MAX_ORDER
);
2688 page
= get_page_from_free_area(area
, fallback_mt
);
2690 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2693 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2694 start_migratetype
, fallback_mt
);
2701 * Do the hard work of removing an element from the buddy allocator.
2702 * Call me with the zone->lock already held.
2704 static __always_inline
struct page
*
2705 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2706 unsigned int alloc_flags
)
2711 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2712 if (unlikely(!page
)) {
2713 if (migratetype
== MIGRATE_MOVABLE
)
2714 page
= __rmqueue_cma_fallback(zone
, order
);
2716 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2721 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2726 * Obtain a specified number of elements from the buddy allocator, all under
2727 * a single hold of the lock, for efficiency. Add them to the supplied list.
2728 * Returns the number of new pages which were placed at *list.
2730 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2731 unsigned long count
, struct list_head
*list
,
2732 int migratetype
, unsigned int alloc_flags
)
2736 spin_lock(&zone
->lock
);
2737 for (i
= 0; i
< count
; ++i
) {
2738 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2740 if (unlikely(page
== NULL
))
2743 if (unlikely(check_pcp_refill(page
)))
2747 * Split buddy pages returned by expand() are received here in
2748 * physical page order. The page is added to the tail of
2749 * caller's list. From the callers perspective, the linked list
2750 * is ordered by page number under some conditions. This is
2751 * useful for IO devices that can forward direction from the
2752 * head, thus also in the physical page order. This is useful
2753 * for IO devices that can merge IO requests if the physical
2754 * pages are ordered properly.
2756 list_add_tail(&page
->lru
, list
);
2758 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2759 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2764 * i pages were removed from the buddy list even if some leak due
2765 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2766 * on i. Do not confuse with 'alloced' which is the number of
2767 * pages added to the pcp list.
2769 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2770 spin_unlock(&zone
->lock
);
2776 * Called from the vmstat counter updater to drain pagesets of this
2777 * currently executing processor on remote nodes after they have
2780 * Note that this function must be called with the thread pinned to
2781 * a single processor.
2783 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2785 unsigned long flags
;
2786 int to_drain
, batch
;
2788 local_irq_save(flags
);
2789 batch
= READ_ONCE(pcp
->batch
);
2790 to_drain
= min(pcp
->count
, batch
);
2792 free_pcppages_bulk(zone
, to_drain
, pcp
);
2793 local_irq_restore(flags
);
2798 * Drain pcplists of the indicated processor and zone.
2800 * The processor must either be the current processor and the
2801 * thread pinned to the current processor or a processor that
2804 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2806 unsigned long flags
;
2807 struct per_cpu_pageset
*pset
;
2808 struct per_cpu_pages
*pcp
;
2810 local_irq_save(flags
);
2811 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2815 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2816 local_irq_restore(flags
);
2820 * Drain pcplists of all zones on the indicated processor.
2822 * The processor must either be the current processor and the
2823 * thread pinned to the current processor or a processor that
2826 static void drain_pages(unsigned int cpu
)
2830 for_each_populated_zone(zone
) {
2831 drain_pages_zone(cpu
, zone
);
2836 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2838 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2839 * the single zone's pages.
2841 void drain_local_pages(struct zone
*zone
)
2843 int cpu
= smp_processor_id();
2846 drain_pages_zone(cpu
, zone
);
2851 static void drain_local_pages_wq(struct work_struct
*work
)
2853 struct pcpu_drain
*drain
;
2855 drain
= container_of(work
, struct pcpu_drain
, work
);
2858 * drain_all_pages doesn't use proper cpu hotplug protection so
2859 * we can race with cpu offline when the WQ can move this from
2860 * a cpu pinned worker to an unbound one. We can operate on a different
2861 * cpu which is allright but we also have to make sure to not move to
2865 drain_local_pages(drain
->zone
);
2870 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2872 * When zone parameter is non-NULL, spill just the single zone's pages.
2874 * Note that this can be extremely slow as the draining happens in a workqueue.
2876 void drain_all_pages(struct zone
*zone
)
2881 * Allocate in the BSS so we wont require allocation in
2882 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2884 static cpumask_t cpus_with_pcps
;
2887 * Make sure nobody triggers this path before mm_percpu_wq is fully
2890 if (WARN_ON_ONCE(!mm_percpu_wq
))
2894 * Do not drain if one is already in progress unless it's specific to
2895 * a zone. Such callers are primarily CMA and memory hotplug and need
2896 * the drain to be complete when the call returns.
2898 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2901 mutex_lock(&pcpu_drain_mutex
);
2905 * We don't care about racing with CPU hotplug event
2906 * as offline notification will cause the notified
2907 * cpu to drain that CPU pcps and on_each_cpu_mask
2908 * disables preemption as part of its processing
2910 for_each_online_cpu(cpu
) {
2911 struct per_cpu_pageset
*pcp
;
2913 bool has_pcps
= false;
2916 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2920 for_each_populated_zone(z
) {
2921 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2922 if (pcp
->pcp
.count
) {
2930 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2932 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2935 for_each_cpu(cpu
, &cpus_with_pcps
) {
2936 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
2939 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
2940 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
2942 for_each_cpu(cpu
, &cpus_with_pcps
)
2943 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
2945 mutex_unlock(&pcpu_drain_mutex
);
2948 #ifdef CONFIG_HIBERNATION
2951 * Touch the watchdog for every WD_PAGE_COUNT pages.
2953 #define WD_PAGE_COUNT (128*1024)
2955 void mark_free_pages(struct zone
*zone
)
2957 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2958 unsigned long flags
;
2959 unsigned int order
, t
;
2962 if (zone_is_empty(zone
))
2965 spin_lock_irqsave(&zone
->lock
, flags
);
2967 max_zone_pfn
= zone_end_pfn(zone
);
2968 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2969 if (pfn_valid(pfn
)) {
2970 page
= pfn_to_page(pfn
);
2972 if (!--page_count
) {
2973 touch_nmi_watchdog();
2974 page_count
= WD_PAGE_COUNT
;
2977 if (page_zone(page
) != zone
)
2980 if (!swsusp_page_is_forbidden(page
))
2981 swsusp_unset_page_free(page
);
2984 for_each_migratetype_order(order
, t
) {
2985 list_for_each_entry(page
,
2986 &zone
->free_area
[order
].free_list
[t
], lru
) {
2989 pfn
= page_to_pfn(page
);
2990 for (i
= 0; i
< (1UL << order
); i
++) {
2991 if (!--page_count
) {
2992 touch_nmi_watchdog();
2993 page_count
= WD_PAGE_COUNT
;
2995 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2999 spin_unlock_irqrestore(&zone
->lock
, flags
);
3001 #endif /* CONFIG_PM */
3003 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
3007 if (!free_pcp_prepare(page
))
3010 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3011 set_pcppage_migratetype(page
, migratetype
);
3015 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
3017 struct zone
*zone
= page_zone(page
);
3018 struct per_cpu_pages
*pcp
;
3021 migratetype
= get_pcppage_migratetype(page
);
3022 __count_vm_event(PGFREE
);
3025 * We only track unmovable, reclaimable and movable on pcp lists.
3026 * Free ISOLATE pages back to the allocator because they are being
3027 * offlined but treat HIGHATOMIC as movable pages so we can get those
3028 * areas back if necessary. Otherwise, we may have to free
3029 * excessively into the page allocator
3031 if (migratetype
>= MIGRATE_PCPTYPES
) {
3032 if (unlikely(is_migrate_isolate(migratetype
))) {
3033 free_one_page(zone
, page
, pfn
, 0, migratetype
);
3036 migratetype
= MIGRATE_MOVABLE
;
3039 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3040 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
3042 if (pcp
->count
>= pcp
->high
) {
3043 unsigned long batch
= READ_ONCE(pcp
->batch
);
3044 free_pcppages_bulk(zone
, batch
, pcp
);
3049 * Free a 0-order page
3051 void free_unref_page(struct page
*page
)
3053 unsigned long flags
;
3054 unsigned long pfn
= page_to_pfn(page
);
3056 if (!free_unref_page_prepare(page
, pfn
))
3059 local_irq_save(flags
);
3060 free_unref_page_commit(page
, pfn
);
3061 local_irq_restore(flags
);
3065 * Free a list of 0-order pages
3067 void free_unref_page_list(struct list_head
*list
)
3069 struct page
*page
, *next
;
3070 unsigned long flags
, pfn
;
3071 int batch_count
= 0;
3073 /* Prepare pages for freeing */
3074 list_for_each_entry_safe(page
, next
, list
, lru
) {
3075 pfn
= page_to_pfn(page
);
3076 if (!free_unref_page_prepare(page
, pfn
))
3077 list_del(&page
->lru
);
3078 set_page_private(page
, pfn
);
3081 local_irq_save(flags
);
3082 list_for_each_entry_safe(page
, next
, list
, lru
) {
3083 unsigned long pfn
= page_private(page
);
3085 set_page_private(page
, 0);
3086 trace_mm_page_free_batched(page
);
3087 free_unref_page_commit(page
, pfn
);
3090 * Guard against excessive IRQ disabled times when we get
3091 * a large list of pages to free.
3093 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3094 local_irq_restore(flags
);
3096 local_irq_save(flags
);
3099 local_irq_restore(flags
);
3103 * split_page takes a non-compound higher-order page, and splits it into
3104 * n (1<<order) sub-pages: page[0..n]
3105 * Each sub-page must be freed individually.
3107 * Note: this is probably too low level an operation for use in drivers.
3108 * Please consult with lkml before using this in your driver.
3110 void split_page(struct page
*page
, unsigned int order
)
3114 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3115 VM_BUG_ON_PAGE(!page_count(page
), page
);
3117 for (i
= 1; i
< (1 << order
); i
++)
3118 set_page_refcounted(page
+ i
);
3119 split_page_owner(page
, order
);
3121 EXPORT_SYMBOL_GPL(split_page
);
3123 int __isolate_free_page(struct page
*page
, unsigned int order
)
3125 struct free_area
*area
= &page_zone(page
)->free_area
[order
];
3126 unsigned long watermark
;
3130 BUG_ON(!PageBuddy(page
));
3132 zone
= page_zone(page
);
3133 mt
= get_pageblock_migratetype(page
);
3135 if (!is_migrate_isolate(mt
)) {
3137 * Obey watermarks as if the page was being allocated. We can
3138 * emulate a high-order watermark check with a raised order-0
3139 * watermark, because we already know our high-order page
3142 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3143 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3146 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3149 /* Remove page from free list */
3151 del_page_from_free_area(page
, area
);
3154 * Set the pageblock if the isolated page is at least half of a
3157 if (order
>= pageblock_order
- 1) {
3158 struct page
*endpage
= page
+ (1 << order
) - 1;
3159 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3160 int mt
= get_pageblock_migratetype(page
);
3161 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3162 && !is_migrate_highatomic(mt
))
3163 set_pageblock_migratetype(page
,
3169 return 1UL << order
;
3173 * Update NUMA hit/miss statistics
3175 * Must be called with interrupts disabled.
3177 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
3180 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3182 /* skip numa counters update if numa stats is disabled */
3183 if (!static_branch_likely(&vm_numa_stat_key
))
3186 if (zone_to_nid(z
) != numa_node_id())
3187 local_stat
= NUMA_OTHER
;
3189 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3190 __inc_numa_state(z
, NUMA_HIT
);
3192 __inc_numa_state(z
, NUMA_MISS
);
3193 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3195 __inc_numa_state(z
, local_stat
);
3199 /* Remove page from the per-cpu list, caller must protect the list */
3200 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3201 unsigned int alloc_flags
,
3202 struct per_cpu_pages
*pcp
,
3203 struct list_head
*list
)
3208 if (list_empty(list
)) {
3209 pcp
->count
+= rmqueue_bulk(zone
, 0,
3211 migratetype
, alloc_flags
);
3212 if (unlikely(list_empty(list
)))
3216 page
= list_first_entry(list
, struct page
, lru
);
3217 list_del(&page
->lru
);
3219 } while (check_new_pcp(page
));
3224 /* Lock and remove page from the per-cpu list */
3225 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3226 struct zone
*zone
, gfp_t gfp_flags
,
3227 int migratetype
, unsigned int alloc_flags
)
3229 struct per_cpu_pages
*pcp
;
3230 struct list_head
*list
;
3232 unsigned long flags
;
3234 local_irq_save(flags
);
3235 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3236 list
= &pcp
->lists
[migratetype
];
3237 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3239 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3240 zone_statistics(preferred_zone
, zone
);
3242 local_irq_restore(flags
);
3247 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3250 struct page
*rmqueue(struct zone
*preferred_zone
,
3251 struct zone
*zone
, unsigned int order
,
3252 gfp_t gfp_flags
, unsigned int alloc_flags
,
3255 unsigned long flags
;
3258 if (likely(order
== 0)) {
3259 page
= rmqueue_pcplist(preferred_zone
, zone
, gfp_flags
,
3260 migratetype
, alloc_flags
);
3265 * We most definitely don't want callers attempting to
3266 * allocate greater than order-1 page units with __GFP_NOFAIL.
3268 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3269 spin_lock_irqsave(&zone
->lock
, flags
);
3273 if (alloc_flags
& ALLOC_HARDER
) {
3274 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3276 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3279 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3280 } while (page
&& check_new_pages(page
, order
));
3281 spin_unlock(&zone
->lock
);
3284 __mod_zone_freepage_state(zone
, -(1 << order
),
3285 get_pcppage_migratetype(page
));
3287 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3288 zone_statistics(preferred_zone
, zone
);
3289 local_irq_restore(flags
);
3292 /* Separate test+clear to avoid unnecessary atomics */
3293 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3294 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3295 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3298 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3302 local_irq_restore(flags
);
3306 #ifdef CONFIG_FAIL_PAGE_ALLOC
3309 struct fault_attr attr
;
3311 bool ignore_gfp_highmem
;
3312 bool ignore_gfp_reclaim
;
3314 } fail_page_alloc
= {
3315 .attr
= FAULT_ATTR_INITIALIZER
,
3316 .ignore_gfp_reclaim
= true,
3317 .ignore_gfp_highmem
= true,
3321 static int __init
setup_fail_page_alloc(char *str
)
3323 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3325 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3327 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3329 if (order
< fail_page_alloc
.min_order
)
3331 if (gfp_mask
& __GFP_NOFAIL
)
3333 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3335 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3336 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3339 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3342 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3344 static int __init
fail_page_alloc_debugfs(void)
3346 umode_t mode
= S_IFREG
| 0600;
3349 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3350 &fail_page_alloc
.attr
);
3352 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3353 &fail_page_alloc
.ignore_gfp_reclaim
);
3354 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3355 &fail_page_alloc
.ignore_gfp_highmem
);
3356 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3361 late_initcall(fail_page_alloc_debugfs
);
3363 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3365 #else /* CONFIG_FAIL_PAGE_ALLOC */
3367 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3372 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3374 static noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3376 return __should_fail_alloc_page(gfp_mask
, order
);
3378 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3381 * Return true if free base pages are above 'mark'. For high-order checks it
3382 * will return true of the order-0 watermark is reached and there is at least
3383 * one free page of a suitable size. Checking now avoids taking the zone lock
3384 * to check in the allocation paths if no pages are free.
3386 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3387 int classzone_idx
, unsigned int alloc_flags
,
3392 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3394 /* free_pages may go negative - that's OK */
3395 free_pages
-= (1 << order
) - 1;
3397 if (alloc_flags
& ALLOC_HIGH
)
3401 * If the caller does not have rights to ALLOC_HARDER then subtract
3402 * the high-atomic reserves. This will over-estimate the size of the
3403 * atomic reserve but it avoids a search.
3405 if (likely(!alloc_harder
)) {
3406 free_pages
-= z
->nr_reserved_highatomic
;
3409 * OOM victims can try even harder than normal ALLOC_HARDER
3410 * users on the grounds that it's definitely going to be in
3411 * the exit path shortly and free memory. Any allocation it
3412 * makes during the free path will be small and short-lived.
3414 if (alloc_flags
& ALLOC_OOM
)
3422 /* If allocation can't use CMA areas don't use free CMA pages */
3423 if (!(alloc_flags
& ALLOC_CMA
))
3424 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3428 * Check watermarks for an order-0 allocation request. If these
3429 * are not met, then a high-order request also cannot go ahead
3430 * even if a suitable page happened to be free.
3432 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3435 /* If this is an order-0 request then the watermark is fine */
3439 /* For a high-order request, check at least one suitable page is free */
3440 for (o
= order
; o
< MAX_ORDER
; o
++) {
3441 struct free_area
*area
= &z
->free_area
[o
];
3447 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3448 if (!free_area_empty(area
, mt
))
3453 if ((alloc_flags
& ALLOC_CMA
) &&
3454 !free_area_empty(area
, MIGRATE_CMA
)) {
3459 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3465 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3466 int classzone_idx
, unsigned int alloc_flags
)
3468 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3469 zone_page_state(z
, NR_FREE_PAGES
));
3472 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3473 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3475 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3479 /* If allocation can't use CMA areas don't use free CMA pages */
3480 if (!(alloc_flags
& ALLOC_CMA
))
3481 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3485 * Fast check for order-0 only. If this fails then the reserves
3486 * need to be calculated. There is a corner case where the check
3487 * passes but only the high-order atomic reserve are free. If
3488 * the caller is !atomic then it'll uselessly search the free
3489 * list. That corner case is then slower but it is harmless.
3491 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3494 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3498 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3499 unsigned long mark
, int classzone_idx
)
3501 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3503 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3504 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3506 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3511 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3513 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3514 node_reclaim_distance
;
3516 #else /* CONFIG_NUMA */
3517 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3521 #endif /* CONFIG_NUMA */
3524 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3525 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3526 * premature use of a lower zone may cause lowmem pressure problems that
3527 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3528 * probably too small. It only makes sense to spread allocations to avoid
3529 * fragmentation between the Normal and DMA32 zones.
3531 static inline unsigned int
3532 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3534 unsigned int alloc_flags
= 0;
3536 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3537 alloc_flags
|= ALLOC_KSWAPD
;
3539 #ifdef CONFIG_ZONE_DMA32
3543 if (zone_idx(zone
) != ZONE_NORMAL
)
3547 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3548 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3549 * on UMA that if Normal is populated then so is DMA32.
3551 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3552 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3555 alloc_flags
|= ALLOC_NOFRAGMENT
;
3556 #endif /* CONFIG_ZONE_DMA32 */
3561 * get_page_from_freelist goes through the zonelist trying to allocate
3564 static struct page
*
3565 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3566 const struct alloc_context
*ac
)
3570 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3575 * Scan zonelist, looking for a zone with enough free.
3576 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3578 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3579 z
= ac
->preferred_zoneref
;
3580 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3585 if (cpusets_enabled() &&
3586 (alloc_flags
& ALLOC_CPUSET
) &&
3587 !__cpuset_zone_allowed(zone
, gfp_mask
))
3590 * When allocating a page cache page for writing, we
3591 * want to get it from a node that is within its dirty
3592 * limit, such that no single node holds more than its
3593 * proportional share of globally allowed dirty pages.
3594 * The dirty limits take into account the node's
3595 * lowmem reserves and high watermark so that kswapd
3596 * should be able to balance it without having to
3597 * write pages from its LRU list.
3599 * XXX: For now, allow allocations to potentially
3600 * exceed the per-node dirty limit in the slowpath
3601 * (spread_dirty_pages unset) before going into reclaim,
3602 * which is important when on a NUMA setup the allowed
3603 * nodes are together not big enough to reach the
3604 * global limit. The proper fix for these situations
3605 * will require awareness of nodes in the
3606 * dirty-throttling and the flusher threads.
3608 if (ac
->spread_dirty_pages
) {
3609 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3612 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3613 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3618 if (no_fallback
&& nr_online_nodes
> 1 &&
3619 zone
!= ac
->preferred_zoneref
->zone
) {
3623 * If moving to a remote node, retry but allow
3624 * fragmenting fallbacks. Locality is more important
3625 * than fragmentation avoidance.
3627 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3628 if (zone_to_nid(zone
) != local_nid
) {
3629 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3634 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3635 if (!zone_watermark_fast(zone
, order
, mark
,
3636 ac_classzone_idx(ac
), alloc_flags
)) {
3639 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3641 * Watermark failed for this zone, but see if we can
3642 * grow this zone if it contains deferred pages.
3644 if (static_branch_unlikely(&deferred_pages
)) {
3645 if (_deferred_grow_zone(zone
, order
))
3649 /* Checked here to keep the fast path fast */
3650 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3651 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3654 if (node_reclaim_mode
== 0 ||
3655 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3658 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3660 case NODE_RECLAIM_NOSCAN
:
3663 case NODE_RECLAIM_FULL
:
3664 /* scanned but unreclaimable */
3667 /* did we reclaim enough */
3668 if (zone_watermark_ok(zone
, order
, mark
,
3669 ac_classzone_idx(ac
), alloc_flags
))
3677 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3678 gfp_mask
, alloc_flags
, ac
->migratetype
);
3680 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3683 * If this is a high-order atomic allocation then check
3684 * if the pageblock should be reserved for the future
3686 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3687 reserve_highatomic_pageblock(page
, zone
, order
);
3691 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3692 /* Try again if zone has deferred pages */
3693 if (static_branch_unlikely(&deferred_pages
)) {
3694 if (_deferred_grow_zone(zone
, order
))
3702 * It's possible on a UMA machine to get through all zones that are
3703 * fragmented. If avoiding fragmentation, reset and try again.
3706 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3713 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3715 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3718 * This documents exceptions given to allocations in certain
3719 * contexts that are allowed to allocate outside current's set
3722 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3723 if (tsk_is_oom_victim(current
) ||
3724 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3725 filter
&= ~SHOW_MEM_FILTER_NODES
;
3726 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3727 filter
&= ~SHOW_MEM_FILTER_NODES
;
3729 show_mem(filter
, nodemask
);
3732 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3734 struct va_format vaf
;
3736 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
3738 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3741 va_start(args
, fmt
);
3744 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3745 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3746 nodemask_pr_args(nodemask
));
3749 cpuset_print_current_mems_allowed();
3752 warn_alloc_show_mem(gfp_mask
, nodemask
);
3755 static inline struct page
*
3756 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3757 unsigned int alloc_flags
,
3758 const struct alloc_context
*ac
)
3762 page
= get_page_from_freelist(gfp_mask
, order
,
3763 alloc_flags
|ALLOC_CPUSET
, ac
);
3765 * fallback to ignore cpuset restriction if our nodes
3769 page
= get_page_from_freelist(gfp_mask
, order
,
3775 static inline struct page
*
3776 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3777 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3779 struct oom_control oc
= {
3780 .zonelist
= ac
->zonelist
,
3781 .nodemask
= ac
->nodemask
,
3783 .gfp_mask
= gfp_mask
,
3788 *did_some_progress
= 0;
3791 * Acquire the oom lock. If that fails, somebody else is
3792 * making progress for us.
3794 if (!mutex_trylock(&oom_lock
)) {
3795 *did_some_progress
= 1;
3796 schedule_timeout_uninterruptible(1);
3801 * Go through the zonelist yet one more time, keep very high watermark
3802 * here, this is only to catch a parallel oom killing, we must fail if
3803 * we're still under heavy pressure. But make sure that this reclaim
3804 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3805 * allocation which will never fail due to oom_lock already held.
3807 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3808 ~__GFP_DIRECT_RECLAIM
, order
,
3809 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3813 /* Coredumps can quickly deplete all memory reserves */
3814 if (current
->flags
& PF_DUMPCORE
)
3816 /* The OOM killer will not help higher order allocs */
3817 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3820 * We have already exhausted all our reclaim opportunities without any
3821 * success so it is time to admit defeat. We will skip the OOM killer
3822 * because it is very likely that the caller has a more reasonable
3823 * fallback than shooting a random task.
3825 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3827 /* The OOM killer does not needlessly kill tasks for lowmem */
3828 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3830 if (pm_suspended_storage())
3833 * XXX: GFP_NOFS allocations should rather fail than rely on
3834 * other request to make a forward progress.
3835 * We are in an unfortunate situation where out_of_memory cannot
3836 * do much for this context but let's try it to at least get
3837 * access to memory reserved if the current task is killed (see
3838 * out_of_memory). Once filesystems are ready to handle allocation
3839 * failures more gracefully we should just bail out here.
3842 /* The OOM killer may not free memory on a specific node */
3843 if (gfp_mask
& __GFP_THISNODE
)
3846 /* Exhausted what can be done so it's blame time */
3847 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3848 *did_some_progress
= 1;
3851 * Help non-failing allocations by giving them access to memory
3854 if (gfp_mask
& __GFP_NOFAIL
)
3855 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3856 ALLOC_NO_WATERMARKS
, ac
);
3859 mutex_unlock(&oom_lock
);
3864 * Maximum number of compaction retries wit a progress before OOM
3865 * killer is consider as the only way to move forward.
3867 #define MAX_COMPACT_RETRIES 16
3869 #ifdef CONFIG_COMPACTION
3870 /* Try memory compaction for high-order allocations before reclaim */
3871 static struct page
*
3872 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3873 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3874 enum compact_priority prio
, enum compact_result
*compact_result
)
3876 struct page
*page
= NULL
;
3877 unsigned long pflags
;
3878 unsigned int noreclaim_flag
;
3883 psi_memstall_enter(&pflags
);
3884 noreclaim_flag
= memalloc_noreclaim_save();
3886 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3889 memalloc_noreclaim_restore(noreclaim_flag
);
3890 psi_memstall_leave(&pflags
);
3893 * At least in one zone compaction wasn't deferred or skipped, so let's
3894 * count a compaction stall
3896 count_vm_event(COMPACTSTALL
);
3898 /* Prep a captured page if available */
3900 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3902 /* Try get a page from the freelist if available */
3904 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3907 struct zone
*zone
= page_zone(page
);
3909 zone
->compact_blockskip_flush
= false;
3910 compaction_defer_reset(zone
, order
, true);
3911 count_vm_event(COMPACTSUCCESS
);
3916 * It's bad if compaction run occurs and fails. The most likely reason
3917 * is that pages exist, but not enough to satisfy watermarks.
3919 count_vm_event(COMPACTFAIL
);
3927 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3928 enum compact_result compact_result
,
3929 enum compact_priority
*compact_priority
,
3930 int *compaction_retries
)
3932 int max_retries
= MAX_COMPACT_RETRIES
;
3935 int retries
= *compaction_retries
;
3936 enum compact_priority priority
= *compact_priority
;
3941 if (compaction_made_progress(compact_result
))
3942 (*compaction_retries
)++;
3945 * compaction considers all the zone as desperately out of memory
3946 * so it doesn't really make much sense to retry except when the
3947 * failure could be caused by insufficient priority
3949 if (compaction_failed(compact_result
))
3950 goto check_priority
;
3953 * compaction was skipped because there are not enough order-0 pages
3954 * to work with, so we retry only if it looks like reclaim can help.
3956 if (compaction_needs_reclaim(compact_result
)) {
3957 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3962 * make sure the compaction wasn't deferred or didn't bail out early
3963 * due to locks contention before we declare that we should give up.
3964 * But the next retry should use a higher priority if allowed, so
3965 * we don't just keep bailing out endlessly.
3967 if (compaction_withdrawn(compact_result
)) {
3968 goto check_priority
;
3972 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3973 * costly ones because they are de facto nofail and invoke OOM
3974 * killer to move on while costly can fail and users are ready
3975 * to cope with that. 1/4 retries is rather arbitrary but we
3976 * would need much more detailed feedback from compaction to
3977 * make a better decision.
3979 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3981 if (*compaction_retries
<= max_retries
) {
3987 * Make sure there are attempts at the highest priority if we exhausted
3988 * all retries or failed at the lower priorities.
3991 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3992 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3994 if (*compact_priority
> min_priority
) {
3995 (*compact_priority
)--;
3996 *compaction_retries
= 0;
4000 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4004 static inline struct page
*
4005 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4006 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4007 enum compact_priority prio
, enum compact_result
*compact_result
)
4009 *compact_result
= COMPACT_SKIPPED
;
4014 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4015 enum compact_result compact_result
,
4016 enum compact_priority
*compact_priority
,
4017 int *compaction_retries
)
4022 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4026 * There are setups with compaction disabled which would prefer to loop
4027 * inside the allocator rather than hit the oom killer prematurely.
4028 * Let's give them a good hope and keep retrying while the order-0
4029 * watermarks are OK.
4031 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4033 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4034 ac_classzone_idx(ac
), alloc_flags
))
4039 #endif /* CONFIG_COMPACTION */
4041 #ifdef CONFIG_LOCKDEP
4042 static struct lockdep_map __fs_reclaim_map
=
4043 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4045 static bool __need_fs_reclaim(gfp_t gfp_mask
)
4047 gfp_mask
= current_gfp_context(gfp_mask
);
4049 /* no reclaim without waiting on it */
4050 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4053 /* this guy won't enter reclaim */
4054 if (current
->flags
& PF_MEMALLOC
)
4057 /* We're only interested __GFP_FS allocations for now */
4058 if (!(gfp_mask
& __GFP_FS
))
4061 if (gfp_mask
& __GFP_NOLOCKDEP
)
4067 void __fs_reclaim_acquire(void)
4069 lock_map_acquire(&__fs_reclaim_map
);
4072 void __fs_reclaim_release(void)
4074 lock_map_release(&__fs_reclaim_map
);
4077 void fs_reclaim_acquire(gfp_t gfp_mask
)
4079 if (__need_fs_reclaim(gfp_mask
))
4080 __fs_reclaim_acquire();
4082 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4084 void fs_reclaim_release(gfp_t gfp_mask
)
4086 if (__need_fs_reclaim(gfp_mask
))
4087 __fs_reclaim_release();
4089 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4092 /* Perform direct synchronous page reclaim */
4094 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4095 const struct alloc_context
*ac
)
4098 unsigned int noreclaim_flag
;
4099 unsigned long pflags
;
4103 /* We now go into synchronous reclaim */
4104 cpuset_memory_pressure_bump();
4105 psi_memstall_enter(&pflags
);
4106 fs_reclaim_acquire(gfp_mask
);
4107 noreclaim_flag
= memalloc_noreclaim_save();
4109 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4112 memalloc_noreclaim_restore(noreclaim_flag
);
4113 fs_reclaim_release(gfp_mask
);
4114 psi_memstall_leave(&pflags
);
4121 /* The really slow allocator path where we enter direct reclaim */
4122 static inline struct page
*
4123 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4124 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4125 unsigned long *did_some_progress
)
4127 struct page
*page
= NULL
;
4128 bool drained
= false;
4130 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4131 if (unlikely(!(*did_some_progress
)))
4135 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4138 * If an allocation failed after direct reclaim, it could be because
4139 * pages are pinned on the per-cpu lists or in high alloc reserves.
4140 * Shrink them them and try again
4142 if (!page
&& !drained
) {
4143 unreserve_highatomic_pageblock(ac
, false);
4144 drain_all_pages(NULL
);
4152 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4153 const struct alloc_context
*ac
)
4157 pg_data_t
*last_pgdat
= NULL
;
4158 enum zone_type high_zoneidx
= ac
->high_zoneidx
;
4160 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, high_zoneidx
,
4162 if (last_pgdat
!= zone
->zone_pgdat
)
4163 wakeup_kswapd(zone
, gfp_mask
, order
, high_zoneidx
);
4164 last_pgdat
= zone
->zone_pgdat
;
4168 static inline unsigned int
4169 gfp_to_alloc_flags(gfp_t gfp_mask
)
4171 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4173 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4174 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4177 * The caller may dip into page reserves a bit more if the caller
4178 * cannot run direct reclaim, or if the caller has realtime scheduling
4179 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4180 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4182 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
4184 if (gfp_mask
& __GFP_ATOMIC
) {
4186 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4187 * if it can't schedule.
4189 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4190 alloc_flags
|= ALLOC_HARDER
;
4192 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4193 * comment for __cpuset_node_allowed().
4195 alloc_flags
&= ~ALLOC_CPUSET
;
4196 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4197 alloc_flags
|= ALLOC_HARDER
;
4199 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4200 alloc_flags
|= ALLOC_KSWAPD
;
4203 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
4204 alloc_flags
|= ALLOC_CMA
;
4209 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4211 if (!tsk_is_oom_victim(tsk
))
4215 * !MMU doesn't have oom reaper so give access to memory reserves
4216 * only to the thread with TIF_MEMDIE set
4218 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4225 * Distinguish requests which really need access to full memory
4226 * reserves from oom victims which can live with a portion of it
4228 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4230 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4232 if (gfp_mask
& __GFP_MEMALLOC
)
4233 return ALLOC_NO_WATERMARKS
;
4234 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4235 return ALLOC_NO_WATERMARKS
;
4236 if (!in_interrupt()) {
4237 if (current
->flags
& PF_MEMALLOC
)
4238 return ALLOC_NO_WATERMARKS
;
4239 else if (oom_reserves_allowed(current
))
4246 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4248 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4252 * Checks whether it makes sense to retry the reclaim to make a forward progress
4253 * for the given allocation request.
4255 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4256 * without success, or when we couldn't even meet the watermark if we
4257 * reclaimed all remaining pages on the LRU lists.
4259 * Returns true if a retry is viable or false to enter the oom path.
4262 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4263 struct alloc_context
*ac
, int alloc_flags
,
4264 bool did_some_progress
, int *no_progress_loops
)
4271 * Costly allocations might have made a progress but this doesn't mean
4272 * their order will become available due to high fragmentation so
4273 * always increment the no progress counter for them
4275 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4276 *no_progress_loops
= 0;
4278 (*no_progress_loops
)++;
4281 * Make sure we converge to OOM if we cannot make any progress
4282 * several times in the row.
4284 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4285 /* Before OOM, exhaust highatomic_reserve */
4286 return unreserve_highatomic_pageblock(ac
, true);
4290 * Keep reclaiming pages while there is a chance this will lead
4291 * somewhere. If none of the target zones can satisfy our allocation
4292 * request even if all reclaimable pages are considered then we are
4293 * screwed and have to go OOM.
4295 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4297 unsigned long available
;
4298 unsigned long reclaimable
;
4299 unsigned long min_wmark
= min_wmark_pages(zone
);
4302 available
= reclaimable
= zone_reclaimable_pages(zone
);
4303 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4306 * Would the allocation succeed if we reclaimed all
4307 * reclaimable pages?
4309 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4310 ac_classzone_idx(ac
), alloc_flags
, available
);
4311 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4312 available
, min_wmark
, *no_progress_loops
, wmark
);
4315 * If we didn't make any progress and have a lot of
4316 * dirty + writeback pages then we should wait for
4317 * an IO to complete to slow down the reclaim and
4318 * prevent from pre mature OOM
4320 if (!did_some_progress
) {
4321 unsigned long write_pending
;
4323 write_pending
= zone_page_state_snapshot(zone
,
4324 NR_ZONE_WRITE_PENDING
);
4326 if (2 * write_pending
> reclaimable
) {
4327 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4339 * Memory allocation/reclaim might be called from a WQ context and the
4340 * current implementation of the WQ concurrency control doesn't
4341 * recognize that a particular WQ is congested if the worker thread is
4342 * looping without ever sleeping. Therefore we have to do a short sleep
4343 * here rather than calling cond_resched().
4345 if (current
->flags
& PF_WQ_WORKER
)
4346 schedule_timeout_uninterruptible(1);
4353 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4356 * It's possible that cpuset's mems_allowed and the nodemask from
4357 * mempolicy don't intersect. This should be normally dealt with by
4358 * policy_nodemask(), but it's possible to race with cpuset update in
4359 * such a way the check therein was true, and then it became false
4360 * before we got our cpuset_mems_cookie here.
4361 * This assumes that for all allocations, ac->nodemask can come only
4362 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4363 * when it does not intersect with the cpuset restrictions) or the
4364 * caller can deal with a violated nodemask.
4366 if (cpusets_enabled() && ac
->nodemask
&&
4367 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4368 ac
->nodemask
= NULL
;
4373 * When updating a task's mems_allowed or mempolicy nodemask, it is
4374 * possible to race with parallel threads in such a way that our
4375 * allocation can fail while the mask is being updated. If we are about
4376 * to fail, check if the cpuset changed during allocation and if so,
4379 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4385 static inline struct page
*
4386 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4387 struct alloc_context
*ac
)
4389 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4390 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4391 struct page
*page
= NULL
;
4392 unsigned int alloc_flags
;
4393 unsigned long did_some_progress
;
4394 enum compact_priority compact_priority
;
4395 enum compact_result compact_result
;
4396 int compaction_retries
;
4397 int no_progress_loops
;
4398 unsigned int cpuset_mems_cookie
;
4402 * We also sanity check to catch abuse of atomic reserves being used by
4403 * callers that are not in atomic context.
4405 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4406 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4407 gfp_mask
&= ~__GFP_ATOMIC
;
4410 compaction_retries
= 0;
4411 no_progress_loops
= 0;
4412 compact_priority
= DEF_COMPACT_PRIORITY
;
4413 cpuset_mems_cookie
= read_mems_allowed_begin();
4416 * The fast path uses conservative alloc_flags to succeed only until
4417 * kswapd needs to be woken up, and to avoid the cost of setting up
4418 * alloc_flags precisely. So we do that now.
4420 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4423 * We need to recalculate the starting point for the zonelist iterator
4424 * because we might have used different nodemask in the fast path, or
4425 * there was a cpuset modification and we are retrying - otherwise we
4426 * could end up iterating over non-eligible zones endlessly.
4428 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4429 ac
->high_zoneidx
, ac
->nodemask
);
4430 if (!ac
->preferred_zoneref
->zone
)
4433 if (alloc_flags
& ALLOC_KSWAPD
)
4434 wake_all_kswapds(order
, gfp_mask
, ac
);
4437 * The adjusted alloc_flags might result in immediate success, so try
4440 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4445 * For costly allocations, try direct compaction first, as it's likely
4446 * that we have enough base pages and don't need to reclaim. For non-
4447 * movable high-order allocations, do that as well, as compaction will
4448 * try prevent permanent fragmentation by migrating from blocks of the
4450 * Don't try this for allocations that are allowed to ignore
4451 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4453 if (can_direct_reclaim
&&
4455 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4456 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4457 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4459 INIT_COMPACT_PRIORITY
,
4464 if (order
>= pageblock_order
&& (gfp_mask
& __GFP_IO
) &&
4465 !(gfp_mask
& __GFP_RETRY_MAYFAIL
)) {
4467 * If allocating entire pageblock(s) and compaction
4468 * failed because all zones are below low watermarks
4469 * or is prohibited because it recently failed at this
4470 * order, fail immediately unless the allocator has
4471 * requested compaction and reclaim retry.
4474 * - potentially very expensive because zones are far
4475 * below their low watermarks or this is part of very
4476 * bursty high order allocations,
4477 * - not guaranteed to help because isolate_freepages()
4478 * may not iterate over freed pages as part of its
4480 * - unlikely to make entire pageblocks free on its
4483 if (compact_result
== COMPACT_SKIPPED
||
4484 compact_result
== COMPACT_DEFERRED
)
4489 * Checks for costly allocations with __GFP_NORETRY, which
4490 * includes THP page fault allocations
4492 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4494 * If compaction is deferred for high-order allocations,
4495 * it is because sync compaction recently failed. If
4496 * this is the case and the caller requested a THP
4497 * allocation, we do not want to heavily disrupt the
4498 * system, so we fail the allocation instead of entering
4501 if (compact_result
== COMPACT_DEFERRED
)
4505 * Looks like reclaim/compaction is worth trying, but
4506 * sync compaction could be very expensive, so keep
4507 * using async compaction.
4509 compact_priority
= INIT_COMPACT_PRIORITY
;
4514 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4515 if (alloc_flags
& ALLOC_KSWAPD
)
4516 wake_all_kswapds(order
, gfp_mask
, ac
);
4518 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4520 alloc_flags
= reserve_flags
;
4523 * Reset the nodemask and zonelist iterators if memory policies can be
4524 * ignored. These allocations are high priority and system rather than
4527 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4528 ac
->nodemask
= NULL
;
4529 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4530 ac
->high_zoneidx
, ac
->nodemask
);
4533 /* Attempt with potentially adjusted zonelist and alloc_flags */
4534 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4538 /* Caller is not willing to reclaim, we can't balance anything */
4539 if (!can_direct_reclaim
)
4542 /* Avoid recursion of direct reclaim */
4543 if (current
->flags
& PF_MEMALLOC
)
4546 /* Try direct reclaim and then allocating */
4547 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4548 &did_some_progress
);
4552 /* Try direct compaction and then allocating */
4553 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4554 compact_priority
, &compact_result
);
4558 /* Do not loop if specifically requested */
4559 if (gfp_mask
& __GFP_NORETRY
)
4563 * Do not retry costly high order allocations unless they are
4564 * __GFP_RETRY_MAYFAIL
4566 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4569 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4570 did_some_progress
> 0, &no_progress_loops
))
4574 * It doesn't make any sense to retry for the compaction if the order-0
4575 * reclaim is not able to make any progress because the current
4576 * implementation of the compaction depends on the sufficient amount
4577 * of free memory (see __compaction_suitable)
4579 if (did_some_progress
> 0 &&
4580 should_compact_retry(ac
, order
, alloc_flags
,
4581 compact_result
, &compact_priority
,
4582 &compaction_retries
))
4586 /* Deal with possible cpuset update races before we start OOM killing */
4587 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4590 /* Reclaim has failed us, start killing things */
4591 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4595 /* Avoid allocations with no watermarks from looping endlessly */
4596 if (tsk_is_oom_victim(current
) &&
4597 (alloc_flags
== ALLOC_OOM
||
4598 (gfp_mask
& __GFP_NOMEMALLOC
)))
4601 /* Retry as long as the OOM killer is making progress */
4602 if (did_some_progress
) {
4603 no_progress_loops
= 0;
4608 /* Deal with possible cpuset update races before we fail */
4609 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4613 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4616 if (gfp_mask
& __GFP_NOFAIL
) {
4618 * All existing users of the __GFP_NOFAIL are blockable, so warn
4619 * of any new users that actually require GFP_NOWAIT
4621 if (WARN_ON_ONCE(!can_direct_reclaim
))
4625 * PF_MEMALLOC request from this context is rather bizarre
4626 * because we cannot reclaim anything and only can loop waiting
4627 * for somebody to do a work for us
4629 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4632 * non failing costly orders are a hard requirement which we
4633 * are not prepared for much so let's warn about these users
4634 * so that we can identify them and convert them to something
4637 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4640 * Help non-failing allocations by giving them access to memory
4641 * reserves but do not use ALLOC_NO_WATERMARKS because this
4642 * could deplete whole memory reserves which would just make
4643 * the situation worse
4645 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4653 warn_alloc(gfp_mask
, ac
->nodemask
,
4654 "page allocation failure: order:%u", order
);
4659 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4660 int preferred_nid
, nodemask_t
*nodemask
,
4661 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4662 unsigned int *alloc_flags
)
4664 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4665 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4666 ac
->nodemask
= nodemask
;
4667 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4669 if (cpusets_enabled()) {
4670 *alloc_mask
|= __GFP_HARDWALL
;
4672 ac
->nodemask
= &cpuset_current_mems_allowed
;
4674 *alloc_flags
|= ALLOC_CPUSET
;
4677 fs_reclaim_acquire(gfp_mask
);
4678 fs_reclaim_release(gfp_mask
);
4680 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4682 if (should_fail_alloc_page(gfp_mask
, order
))
4685 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4686 *alloc_flags
|= ALLOC_CMA
;
4691 /* Determine whether to spread dirty pages and what the first usable zone */
4692 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4694 /* Dirty zone balancing only done in the fast path */
4695 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4698 * The preferred zone is used for statistics but crucially it is
4699 * also used as the starting point for the zonelist iterator. It
4700 * may get reset for allocations that ignore memory policies.
4702 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4703 ac
->high_zoneidx
, ac
->nodemask
);
4707 * This is the 'heart' of the zoned buddy allocator.
4710 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4711 nodemask_t
*nodemask
)
4714 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4715 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4716 struct alloc_context ac
= { };
4719 * There are several places where we assume that the order value is sane
4720 * so bail out early if the request is out of bound.
4722 if (unlikely(order
>= MAX_ORDER
)) {
4723 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4727 gfp_mask
&= gfp_allowed_mask
;
4728 alloc_mask
= gfp_mask
;
4729 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4732 finalise_ac(gfp_mask
, &ac
);
4735 * Forbid the first pass from falling back to types that fragment
4736 * memory until all local zones are considered.
4738 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4740 /* First allocation attempt */
4741 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4746 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4747 * resp. GFP_NOIO which has to be inherited for all allocation requests
4748 * from a particular context which has been marked by
4749 * memalloc_no{fs,io}_{save,restore}.
4751 alloc_mask
= current_gfp_context(gfp_mask
);
4752 ac
.spread_dirty_pages
= false;
4755 * Restore the original nodemask if it was potentially replaced with
4756 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4758 if (unlikely(ac
.nodemask
!= nodemask
))
4759 ac
.nodemask
= nodemask
;
4761 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4764 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4765 unlikely(__memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4766 __free_pages(page
, order
);
4770 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4774 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4777 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4778 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4779 * you need to access high mem.
4781 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4785 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4788 return (unsigned long) page_address(page
);
4790 EXPORT_SYMBOL(__get_free_pages
);
4792 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4794 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4796 EXPORT_SYMBOL(get_zeroed_page
);
4798 static inline void free_the_page(struct page
*page
, unsigned int order
)
4800 if (order
== 0) /* Via pcp? */
4801 free_unref_page(page
);
4803 __free_pages_ok(page
, order
);
4806 void __free_pages(struct page
*page
, unsigned int order
)
4808 if (put_page_testzero(page
))
4809 free_the_page(page
, order
);
4811 EXPORT_SYMBOL(__free_pages
);
4813 void free_pages(unsigned long addr
, unsigned int order
)
4816 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4817 __free_pages(virt_to_page((void *)addr
), order
);
4821 EXPORT_SYMBOL(free_pages
);
4825 * An arbitrary-length arbitrary-offset area of memory which resides
4826 * within a 0 or higher order page. Multiple fragments within that page
4827 * are individually refcounted, in the page's reference counter.
4829 * The page_frag functions below provide a simple allocation framework for
4830 * page fragments. This is used by the network stack and network device
4831 * drivers to provide a backing region of memory for use as either an
4832 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4834 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4837 struct page
*page
= NULL
;
4838 gfp_t gfp
= gfp_mask
;
4840 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4841 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4843 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4844 PAGE_FRAG_CACHE_MAX_ORDER
);
4845 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4847 if (unlikely(!page
))
4848 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4850 nc
->va
= page
? page_address(page
) : NULL
;
4855 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4857 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4859 if (page_ref_sub_and_test(page
, count
))
4860 free_the_page(page
, compound_order(page
));
4862 EXPORT_SYMBOL(__page_frag_cache_drain
);
4864 void *page_frag_alloc(struct page_frag_cache
*nc
,
4865 unsigned int fragsz
, gfp_t gfp_mask
)
4867 unsigned int size
= PAGE_SIZE
;
4871 if (unlikely(!nc
->va
)) {
4873 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4877 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4878 /* if size can vary use size else just use PAGE_SIZE */
4881 /* Even if we own the page, we do not use atomic_set().
4882 * This would break get_page_unless_zero() users.
4884 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
4886 /* reset page count bias and offset to start of new frag */
4887 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4888 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4892 offset
= nc
->offset
- fragsz
;
4893 if (unlikely(offset
< 0)) {
4894 page
= virt_to_page(nc
->va
);
4896 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4899 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4900 /* if size can vary use size else just use PAGE_SIZE */
4903 /* OK, page count is 0, we can safely set it */
4904 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
4906 /* reset page count bias and offset to start of new frag */
4907 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4908 offset
= size
- fragsz
;
4912 nc
->offset
= offset
;
4914 return nc
->va
+ offset
;
4916 EXPORT_SYMBOL(page_frag_alloc
);
4919 * Frees a page fragment allocated out of either a compound or order 0 page.
4921 void page_frag_free(void *addr
)
4923 struct page
*page
= virt_to_head_page(addr
);
4925 if (unlikely(put_page_testzero(page
)))
4926 free_the_page(page
, compound_order(page
));
4928 EXPORT_SYMBOL(page_frag_free
);
4930 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4934 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4935 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4937 split_page(virt_to_page((void *)addr
), order
);
4938 while (used
< alloc_end
) {
4943 return (void *)addr
;
4947 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4948 * @size: the number of bytes to allocate
4949 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4951 * This function is similar to alloc_pages(), except that it allocates the
4952 * minimum number of pages to satisfy the request. alloc_pages() can only
4953 * allocate memory in power-of-two pages.
4955 * This function is also limited by MAX_ORDER.
4957 * Memory allocated by this function must be released by free_pages_exact().
4959 * Return: pointer to the allocated area or %NULL in case of error.
4961 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4963 unsigned int order
= get_order(size
);
4966 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
4967 gfp_mask
&= ~__GFP_COMP
;
4969 addr
= __get_free_pages(gfp_mask
, order
);
4970 return make_alloc_exact(addr
, order
, size
);
4972 EXPORT_SYMBOL(alloc_pages_exact
);
4975 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4977 * @nid: the preferred node ID where memory should be allocated
4978 * @size: the number of bytes to allocate
4979 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4981 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4984 * Return: pointer to the allocated area or %NULL in case of error.
4986 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4988 unsigned int order
= get_order(size
);
4991 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
4992 gfp_mask
&= ~__GFP_COMP
;
4994 p
= alloc_pages_node(nid
, gfp_mask
, order
);
4997 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5001 * free_pages_exact - release memory allocated via alloc_pages_exact()
5002 * @virt: the value returned by alloc_pages_exact.
5003 * @size: size of allocation, same value as passed to alloc_pages_exact().
5005 * Release the memory allocated by a previous call to alloc_pages_exact.
5007 void free_pages_exact(void *virt
, size_t size
)
5009 unsigned long addr
= (unsigned long)virt
;
5010 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5012 while (addr
< end
) {
5017 EXPORT_SYMBOL(free_pages_exact
);
5020 * nr_free_zone_pages - count number of pages beyond high watermark
5021 * @offset: The zone index of the highest zone
5023 * nr_free_zone_pages() counts the number of pages which are beyond the
5024 * high watermark within all zones at or below a given zone index. For each
5025 * zone, the number of pages is calculated as:
5027 * nr_free_zone_pages = managed_pages - high_pages
5029 * Return: number of pages beyond high watermark.
5031 static unsigned long nr_free_zone_pages(int offset
)
5036 /* Just pick one node, since fallback list is circular */
5037 unsigned long sum
= 0;
5039 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5041 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5042 unsigned long size
= zone_managed_pages(zone
);
5043 unsigned long high
= high_wmark_pages(zone
);
5052 * nr_free_buffer_pages - count number of pages beyond high watermark
5054 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5055 * watermark within ZONE_DMA and ZONE_NORMAL.
5057 * Return: number of pages beyond high watermark within ZONE_DMA and
5060 unsigned long nr_free_buffer_pages(void)
5062 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5064 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5067 * nr_free_pagecache_pages - count number of pages beyond high watermark
5069 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5070 * high watermark within all zones.
5072 * Return: number of pages beyond high watermark within all zones.
5074 unsigned long nr_free_pagecache_pages(void)
5076 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5079 static inline void show_node(struct zone
*zone
)
5081 if (IS_ENABLED(CONFIG_NUMA
))
5082 printk("Node %d ", zone_to_nid(zone
));
5085 long si_mem_available(void)
5088 unsigned long pagecache
;
5089 unsigned long wmark_low
= 0;
5090 unsigned long pages
[NR_LRU_LISTS
];
5091 unsigned long reclaimable
;
5095 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5096 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5099 wmark_low
+= low_wmark_pages(zone
);
5102 * Estimate the amount of memory available for userspace allocations,
5103 * without causing swapping.
5105 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5108 * Not all the page cache can be freed, otherwise the system will
5109 * start swapping. Assume at least half of the page cache, or the
5110 * low watermark worth of cache, needs to stay.
5112 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5113 pagecache
-= min(pagecache
/ 2, wmark_low
);
5114 available
+= pagecache
;
5117 * Part of the reclaimable slab and other kernel memory consists of
5118 * items that are in use, and cannot be freed. Cap this estimate at the
5121 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
5122 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5123 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5129 EXPORT_SYMBOL_GPL(si_mem_available
);
5131 void si_meminfo(struct sysinfo
*val
)
5133 val
->totalram
= totalram_pages();
5134 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5135 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5136 val
->bufferram
= nr_blockdev_pages();
5137 val
->totalhigh
= totalhigh_pages();
5138 val
->freehigh
= nr_free_highpages();
5139 val
->mem_unit
= PAGE_SIZE
;
5142 EXPORT_SYMBOL(si_meminfo
);
5145 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5147 int zone_type
; /* needs to be signed */
5148 unsigned long managed_pages
= 0;
5149 unsigned long managed_highpages
= 0;
5150 unsigned long free_highpages
= 0;
5151 pg_data_t
*pgdat
= NODE_DATA(nid
);
5153 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5154 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5155 val
->totalram
= managed_pages
;
5156 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5157 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5158 #ifdef CONFIG_HIGHMEM
5159 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5160 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5162 if (is_highmem(zone
)) {
5163 managed_highpages
+= zone_managed_pages(zone
);
5164 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5167 val
->totalhigh
= managed_highpages
;
5168 val
->freehigh
= free_highpages
;
5170 val
->totalhigh
= managed_highpages
;
5171 val
->freehigh
= free_highpages
;
5173 val
->mem_unit
= PAGE_SIZE
;
5178 * Determine whether the node should be displayed or not, depending on whether
5179 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5181 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5183 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5187 * no node mask - aka implicit memory numa policy. Do not bother with
5188 * the synchronization - read_mems_allowed_begin - because we do not
5189 * have to be precise here.
5192 nodemask
= &cpuset_current_mems_allowed
;
5194 return !node_isset(nid
, *nodemask
);
5197 #define K(x) ((x) << (PAGE_SHIFT-10))
5199 static void show_migration_types(unsigned char type
)
5201 static const char types
[MIGRATE_TYPES
] = {
5202 [MIGRATE_UNMOVABLE
] = 'U',
5203 [MIGRATE_MOVABLE
] = 'M',
5204 [MIGRATE_RECLAIMABLE
] = 'E',
5205 [MIGRATE_HIGHATOMIC
] = 'H',
5207 [MIGRATE_CMA
] = 'C',
5209 #ifdef CONFIG_MEMORY_ISOLATION
5210 [MIGRATE_ISOLATE
] = 'I',
5213 char tmp
[MIGRATE_TYPES
+ 1];
5217 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5218 if (type
& (1 << i
))
5223 printk(KERN_CONT
"(%s) ", tmp
);
5227 * Show free area list (used inside shift_scroll-lock stuff)
5228 * We also calculate the percentage fragmentation. We do this by counting the
5229 * memory on each free list with the exception of the first item on the list.
5232 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5235 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5237 unsigned long free_pcp
= 0;
5242 for_each_populated_zone(zone
) {
5243 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5246 for_each_online_cpu(cpu
)
5247 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5250 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5251 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5252 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5253 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5254 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5255 " free:%lu free_pcp:%lu free_cma:%lu\n",
5256 global_node_page_state(NR_ACTIVE_ANON
),
5257 global_node_page_state(NR_INACTIVE_ANON
),
5258 global_node_page_state(NR_ISOLATED_ANON
),
5259 global_node_page_state(NR_ACTIVE_FILE
),
5260 global_node_page_state(NR_INACTIVE_FILE
),
5261 global_node_page_state(NR_ISOLATED_FILE
),
5262 global_node_page_state(NR_UNEVICTABLE
),
5263 global_node_page_state(NR_FILE_DIRTY
),
5264 global_node_page_state(NR_WRITEBACK
),
5265 global_node_page_state(NR_UNSTABLE_NFS
),
5266 global_node_page_state(NR_SLAB_RECLAIMABLE
),
5267 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
5268 global_node_page_state(NR_FILE_MAPPED
),
5269 global_node_page_state(NR_SHMEM
),
5270 global_zone_page_state(NR_PAGETABLE
),
5271 global_zone_page_state(NR_BOUNCE
),
5272 global_zone_page_state(NR_FREE_PAGES
),
5274 global_zone_page_state(NR_FREE_CMA_PAGES
));
5276 for_each_online_pgdat(pgdat
) {
5277 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5281 " active_anon:%lukB"
5282 " inactive_anon:%lukB"
5283 " active_file:%lukB"
5284 " inactive_file:%lukB"
5285 " unevictable:%lukB"
5286 " isolated(anon):%lukB"
5287 " isolated(file):%lukB"
5292 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5294 " shmem_pmdmapped: %lukB"
5297 " writeback_tmp:%lukB"
5299 " all_unreclaimable? %s"
5302 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5303 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5304 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5305 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5306 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5307 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5308 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5309 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5310 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5311 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5312 K(node_page_state(pgdat
, NR_SHMEM
)),
5313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5314 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5315 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5317 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5319 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5320 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
5321 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5325 for_each_populated_zone(zone
) {
5328 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5332 for_each_online_cpu(cpu
)
5333 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5342 " active_anon:%lukB"
5343 " inactive_anon:%lukB"
5344 " active_file:%lukB"
5345 " inactive_file:%lukB"
5346 " unevictable:%lukB"
5347 " writepending:%lukB"
5351 " kernel_stack:%lukB"
5359 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5360 K(min_wmark_pages(zone
)),
5361 K(low_wmark_pages(zone
)),
5362 K(high_wmark_pages(zone
)),
5363 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5364 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5365 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5366 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5367 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5368 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5369 K(zone
->present_pages
),
5370 K(zone_managed_pages(zone
)),
5371 K(zone_page_state(zone
, NR_MLOCK
)),
5372 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5373 K(zone_page_state(zone
, NR_PAGETABLE
)),
5374 K(zone_page_state(zone
, NR_BOUNCE
)),
5376 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5377 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5378 printk("lowmem_reserve[]:");
5379 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5380 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5381 printk(KERN_CONT
"\n");
5384 for_each_populated_zone(zone
) {
5386 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5387 unsigned char types
[MAX_ORDER
];
5389 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5392 printk(KERN_CONT
"%s: ", zone
->name
);
5394 spin_lock_irqsave(&zone
->lock
, flags
);
5395 for (order
= 0; order
< MAX_ORDER
; order
++) {
5396 struct free_area
*area
= &zone
->free_area
[order
];
5399 nr
[order
] = area
->nr_free
;
5400 total
+= nr
[order
] << order
;
5403 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5404 if (!free_area_empty(area
, type
))
5405 types
[order
] |= 1 << type
;
5408 spin_unlock_irqrestore(&zone
->lock
, flags
);
5409 for (order
= 0; order
< MAX_ORDER
; order
++) {
5410 printk(KERN_CONT
"%lu*%lukB ",
5411 nr
[order
], K(1UL) << order
);
5413 show_migration_types(types
[order
]);
5415 printk(KERN_CONT
"= %lukB\n", K(total
));
5418 hugetlb_show_meminfo();
5420 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5422 show_swap_cache_info();
5425 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5427 zoneref
->zone
= zone
;
5428 zoneref
->zone_idx
= zone_idx(zone
);
5432 * Builds allocation fallback zone lists.
5434 * Add all populated zones of a node to the zonelist.
5436 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5439 enum zone_type zone_type
= MAX_NR_ZONES
;
5444 zone
= pgdat
->node_zones
+ zone_type
;
5445 if (managed_zone(zone
)) {
5446 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5447 check_highest_zone(zone_type
);
5449 } while (zone_type
);
5456 static int __parse_numa_zonelist_order(char *s
)
5459 * We used to support different zonlists modes but they turned
5460 * out to be just not useful. Let's keep the warning in place
5461 * if somebody still use the cmd line parameter so that we do
5462 * not fail it silently
5464 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5465 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5471 static __init
int setup_numa_zonelist_order(char *s
)
5476 return __parse_numa_zonelist_order(s
);
5478 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
5480 char numa_zonelist_order
[] = "Node";
5483 * sysctl handler for numa_zonelist_order
5485 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5486 void __user
*buffer
, size_t *length
,
5493 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5494 str
= memdup_user_nul(buffer
, 16);
5496 return PTR_ERR(str
);
5498 ret
= __parse_numa_zonelist_order(str
);
5504 #define MAX_NODE_LOAD (nr_online_nodes)
5505 static int node_load
[MAX_NUMNODES
];
5508 * find_next_best_node - find the next node that should appear in a given node's fallback list
5509 * @node: node whose fallback list we're appending
5510 * @used_node_mask: nodemask_t of already used nodes
5512 * We use a number of factors to determine which is the next node that should
5513 * appear on a given node's fallback list. The node should not have appeared
5514 * already in @node's fallback list, and it should be the next closest node
5515 * according to the distance array (which contains arbitrary distance values
5516 * from each node to each node in the system), and should also prefer nodes
5517 * with no CPUs, since presumably they'll have very little allocation pressure
5518 * on them otherwise.
5520 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5522 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5525 int min_val
= INT_MAX
;
5526 int best_node
= NUMA_NO_NODE
;
5527 const struct cpumask
*tmp
= cpumask_of_node(0);
5529 /* Use the local node if we haven't already */
5530 if (!node_isset(node
, *used_node_mask
)) {
5531 node_set(node
, *used_node_mask
);
5535 for_each_node_state(n
, N_MEMORY
) {
5537 /* Don't want a node to appear more than once */
5538 if (node_isset(n
, *used_node_mask
))
5541 /* Use the distance array to find the distance */
5542 val
= node_distance(node
, n
);
5544 /* Penalize nodes under us ("prefer the next node") */
5547 /* Give preference to headless and unused nodes */
5548 tmp
= cpumask_of_node(n
);
5549 if (!cpumask_empty(tmp
))
5550 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5552 /* Slight preference for less loaded node */
5553 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5554 val
+= node_load
[n
];
5556 if (val
< min_val
) {
5563 node_set(best_node
, *used_node_mask
);
5570 * Build zonelists ordered by node and zones within node.
5571 * This results in maximum locality--normal zone overflows into local
5572 * DMA zone, if any--but risks exhausting DMA zone.
5574 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5577 struct zoneref
*zonerefs
;
5580 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5582 for (i
= 0; i
< nr_nodes
; i
++) {
5585 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5587 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5588 zonerefs
+= nr_zones
;
5590 zonerefs
->zone
= NULL
;
5591 zonerefs
->zone_idx
= 0;
5595 * Build gfp_thisnode zonelists
5597 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5599 struct zoneref
*zonerefs
;
5602 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5603 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5604 zonerefs
+= nr_zones
;
5605 zonerefs
->zone
= NULL
;
5606 zonerefs
->zone_idx
= 0;
5610 * Build zonelists ordered by zone and nodes within zones.
5611 * This results in conserving DMA zone[s] until all Normal memory is
5612 * exhausted, but results in overflowing to remote node while memory
5613 * may still exist in local DMA zone.
5616 static void build_zonelists(pg_data_t
*pgdat
)
5618 static int node_order
[MAX_NUMNODES
];
5619 int node
, load
, nr_nodes
= 0;
5620 nodemask_t used_mask
;
5621 int local_node
, prev_node
;
5623 /* NUMA-aware ordering of nodes */
5624 local_node
= pgdat
->node_id
;
5625 load
= nr_online_nodes
;
5626 prev_node
= local_node
;
5627 nodes_clear(used_mask
);
5629 memset(node_order
, 0, sizeof(node_order
));
5630 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5632 * We don't want to pressure a particular node.
5633 * So adding penalty to the first node in same
5634 * distance group to make it round-robin.
5636 if (node_distance(local_node
, node
) !=
5637 node_distance(local_node
, prev_node
))
5638 node_load
[node
] = load
;
5640 node_order
[nr_nodes
++] = node
;
5645 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5646 build_thisnode_zonelists(pgdat
);
5649 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5651 * Return node id of node used for "local" allocations.
5652 * I.e., first node id of first zone in arg node's generic zonelist.
5653 * Used for initializing percpu 'numa_mem', which is used primarily
5654 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5656 int local_memory_node(int node
)
5660 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5661 gfp_zone(GFP_KERNEL
),
5663 return zone_to_nid(z
->zone
);
5667 static void setup_min_unmapped_ratio(void);
5668 static void setup_min_slab_ratio(void);
5669 #else /* CONFIG_NUMA */
5671 static void build_zonelists(pg_data_t
*pgdat
)
5673 int node
, local_node
;
5674 struct zoneref
*zonerefs
;
5677 local_node
= pgdat
->node_id
;
5679 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5680 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5681 zonerefs
+= nr_zones
;
5684 * Now we build the zonelist so that it contains the zones
5685 * of all the other nodes.
5686 * We don't want to pressure a particular node, so when
5687 * building the zones for node N, we make sure that the
5688 * zones coming right after the local ones are those from
5689 * node N+1 (modulo N)
5691 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5692 if (!node_online(node
))
5694 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5695 zonerefs
+= nr_zones
;
5697 for (node
= 0; node
< local_node
; node
++) {
5698 if (!node_online(node
))
5700 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5701 zonerefs
+= nr_zones
;
5704 zonerefs
->zone
= NULL
;
5705 zonerefs
->zone_idx
= 0;
5708 #endif /* CONFIG_NUMA */
5711 * Boot pageset table. One per cpu which is going to be used for all
5712 * zones and all nodes. The parameters will be set in such a way
5713 * that an item put on a list will immediately be handed over to
5714 * the buddy list. This is safe since pageset manipulation is done
5715 * with interrupts disabled.
5717 * The boot_pagesets must be kept even after bootup is complete for
5718 * unused processors and/or zones. They do play a role for bootstrapping
5719 * hotplugged processors.
5721 * zoneinfo_show() and maybe other functions do
5722 * not check if the processor is online before following the pageset pointer.
5723 * Other parts of the kernel may not check if the zone is available.
5725 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5726 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5727 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5729 static void __build_all_zonelists(void *data
)
5732 int __maybe_unused cpu
;
5733 pg_data_t
*self
= data
;
5734 static DEFINE_SPINLOCK(lock
);
5739 memset(node_load
, 0, sizeof(node_load
));
5743 * This node is hotadded and no memory is yet present. So just
5744 * building zonelists is fine - no need to touch other nodes.
5746 if (self
&& !node_online(self
->node_id
)) {
5747 build_zonelists(self
);
5749 for_each_online_node(nid
) {
5750 pg_data_t
*pgdat
= NODE_DATA(nid
);
5752 build_zonelists(pgdat
);
5755 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5757 * We now know the "local memory node" for each node--
5758 * i.e., the node of the first zone in the generic zonelist.
5759 * Set up numa_mem percpu variable for on-line cpus. During
5760 * boot, only the boot cpu should be on-line; we'll init the
5761 * secondary cpus' numa_mem as they come on-line. During
5762 * node/memory hotplug, we'll fixup all on-line cpus.
5764 for_each_online_cpu(cpu
)
5765 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5772 static noinline
void __init
5773 build_all_zonelists_init(void)
5777 __build_all_zonelists(NULL
);
5780 * Initialize the boot_pagesets that are going to be used
5781 * for bootstrapping processors. The real pagesets for
5782 * each zone will be allocated later when the per cpu
5783 * allocator is available.
5785 * boot_pagesets are used also for bootstrapping offline
5786 * cpus if the system is already booted because the pagesets
5787 * are needed to initialize allocators on a specific cpu too.
5788 * F.e. the percpu allocator needs the page allocator which
5789 * needs the percpu allocator in order to allocate its pagesets
5790 * (a chicken-egg dilemma).
5792 for_each_possible_cpu(cpu
)
5793 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5795 mminit_verify_zonelist();
5796 cpuset_init_current_mems_allowed();
5800 * unless system_state == SYSTEM_BOOTING.
5802 * __ref due to call of __init annotated helper build_all_zonelists_init
5803 * [protected by SYSTEM_BOOTING].
5805 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5807 if (system_state
== SYSTEM_BOOTING
) {
5808 build_all_zonelists_init();
5810 __build_all_zonelists(pgdat
);
5811 /* cpuset refresh routine should be here */
5813 vm_total_pages
= nr_free_pagecache_pages();
5815 * Disable grouping by mobility if the number of pages in the
5816 * system is too low to allow the mechanism to work. It would be
5817 * more accurate, but expensive to check per-zone. This check is
5818 * made on memory-hotadd so a system can start with mobility
5819 * disabled and enable it later
5821 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5822 page_group_by_mobility_disabled
= 1;
5824 page_group_by_mobility_disabled
= 0;
5826 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5828 page_group_by_mobility_disabled
? "off" : "on",
5831 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5835 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5836 static bool __meminit
5837 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5839 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5840 static struct memblock_region
*r
;
5842 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5843 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5844 for_each_memblock(memory
, r
) {
5845 if (*pfn
< memblock_region_memory_end_pfn(r
))
5849 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5850 memblock_is_mirror(r
)) {
5851 *pfn
= memblock_region_memory_end_pfn(r
);
5860 * Initially all pages are reserved - free ones are freed
5861 * up by memblock_free_all() once the early boot process is
5862 * done. Non-atomic initialization, single-pass.
5864 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5865 unsigned long start_pfn
, enum memmap_context context
,
5866 struct vmem_altmap
*altmap
)
5868 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5871 if (highest_memmap_pfn
< end_pfn
- 1)
5872 highest_memmap_pfn
= end_pfn
- 1;
5874 #ifdef CONFIG_ZONE_DEVICE
5876 * Honor reservation requested by the driver for this ZONE_DEVICE
5877 * memory. We limit the total number of pages to initialize to just
5878 * those that might contain the memory mapping. We will defer the
5879 * ZONE_DEVICE page initialization until after we have released
5882 if (zone
== ZONE_DEVICE
) {
5886 if (start_pfn
== altmap
->base_pfn
)
5887 start_pfn
+= altmap
->reserve
;
5888 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5892 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5894 * There can be holes in boot-time mem_map[]s handed to this
5895 * function. They do not exist on hotplugged memory.
5897 if (context
== MEMMAP_EARLY
) {
5898 if (!early_pfn_valid(pfn
))
5900 if (!early_pfn_in_nid(pfn
, nid
))
5902 if (overlap_memmap_init(zone
, &pfn
))
5904 if (defer_init(nid
, pfn
, end_pfn
))
5908 page
= pfn_to_page(pfn
);
5909 __init_single_page(page
, pfn
, zone
, nid
);
5910 if (context
== MEMMAP_HOTPLUG
)
5911 __SetPageReserved(page
);
5914 * Mark the block movable so that blocks are reserved for
5915 * movable at startup. This will force kernel allocations
5916 * to reserve their blocks rather than leaking throughout
5917 * the address space during boot when many long-lived
5918 * kernel allocations are made.
5920 * bitmap is created for zone's valid pfn range. but memmap
5921 * can be created for invalid pages (for alignment)
5922 * check here not to call set_pageblock_migratetype() against
5925 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5926 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5932 #ifdef CONFIG_ZONE_DEVICE
5933 void __ref
memmap_init_zone_device(struct zone
*zone
,
5934 unsigned long start_pfn
,
5936 struct dev_pagemap
*pgmap
)
5938 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5939 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5940 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
5941 unsigned long zone_idx
= zone_idx(zone
);
5942 unsigned long start
= jiffies
;
5943 int nid
= pgdat
->node_id
;
5945 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
5949 * The call to memmap_init_zone should have already taken care
5950 * of the pages reserved for the memmap, so we can just jump to
5951 * the end of that region and start processing the device pages.
5954 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5955 size
= end_pfn
- start_pfn
;
5958 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5959 struct page
*page
= pfn_to_page(pfn
);
5961 __init_single_page(page
, pfn
, zone_idx
, nid
);
5964 * Mark page reserved as it will need to wait for onlining
5965 * phase for it to be fully associated with a zone.
5967 * We can use the non-atomic __set_bit operation for setting
5968 * the flag as we are still initializing the pages.
5970 __SetPageReserved(page
);
5973 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5974 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5975 * ever freed or placed on a driver-private list.
5977 page
->pgmap
= pgmap
;
5978 page
->zone_device_data
= NULL
;
5981 * Mark the block movable so that blocks are reserved for
5982 * movable at startup. This will force kernel allocations
5983 * to reserve their blocks rather than leaking throughout
5984 * the address space during boot when many long-lived
5985 * kernel allocations are made.
5987 * bitmap is created for zone's valid pfn range. but memmap
5988 * can be created for invalid pages (for alignment)
5989 * check here not to call set_pageblock_migratetype() against
5992 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5993 * because this is done early in section_activate()
5995 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5996 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6001 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6002 size
, jiffies_to_msecs(jiffies
- start
));
6006 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6008 unsigned int order
, t
;
6009 for_each_migratetype_order(order
, t
) {
6010 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6011 zone
->free_area
[order
].nr_free
= 0;
6015 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
6016 unsigned long zone
, unsigned long start_pfn
)
6018 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
, NULL
);
6021 static int zone_batchsize(struct zone
*zone
)
6027 * The per-cpu-pages pools are set to around 1000th of the
6030 batch
= zone_managed_pages(zone
) / 1024;
6031 /* But no more than a meg. */
6032 if (batch
* PAGE_SIZE
> 1024 * 1024)
6033 batch
= (1024 * 1024) / PAGE_SIZE
;
6034 batch
/= 4; /* We effectively *= 4 below */
6039 * Clamp the batch to a 2^n - 1 value. Having a power
6040 * of 2 value was found to be more likely to have
6041 * suboptimal cache aliasing properties in some cases.
6043 * For example if 2 tasks are alternately allocating
6044 * batches of pages, one task can end up with a lot
6045 * of pages of one half of the possible page colors
6046 * and the other with pages of the other colors.
6048 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6053 /* The deferral and batching of frees should be suppressed under NOMMU
6056 * The problem is that NOMMU needs to be able to allocate large chunks
6057 * of contiguous memory as there's no hardware page translation to
6058 * assemble apparent contiguous memory from discontiguous pages.
6060 * Queueing large contiguous runs of pages for batching, however,
6061 * causes the pages to actually be freed in smaller chunks. As there
6062 * can be a significant delay between the individual batches being
6063 * recycled, this leads to the once large chunks of space being
6064 * fragmented and becoming unavailable for high-order allocations.
6071 * pcp->high and pcp->batch values are related and dependent on one another:
6072 * ->batch must never be higher then ->high.
6073 * The following function updates them in a safe manner without read side
6076 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6077 * those fields changing asynchronously (acording the the above rule).
6079 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6080 * outside of boot time (or some other assurance that no concurrent updaters
6083 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6084 unsigned long batch
)
6086 /* start with a fail safe value for batch */
6090 /* Update high, then batch, in order */
6097 /* a companion to pageset_set_high() */
6098 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
6100 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
6103 static void pageset_init(struct per_cpu_pageset
*p
)
6105 struct per_cpu_pages
*pcp
;
6108 memset(p
, 0, sizeof(*p
));
6111 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
6112 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
6115 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
6118 pageset_set_batch(p
, batch
);
6122 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6123 * to the value high for the pageset p.
6125 static void pageset_set_high(struct per_cpu_pageset
*p
,
6128 unsigned long batch
= max(1UL, high
/ 4);
6129 if ((high
/ 4) > (PAGE_SHIFT
* 8))
6130 batch
= PAGE_SHIFT
* 8;
6132 pageset_update(&p
->pcp
, high
, batch
);
6135 static void pageset_set_high_and_batch(struct zone
*zone
,
6136 struct per_cpu_pageset
*pcp
)
6138 if (percpu_pagelist_fraction
)
6139 pageset_set_high(pcp
,
6140 (zone_managed_pages(zone
) /
6141 percpu_pagelist_fraction
));
6143 pageset_set_batch(pcp
, zone_batchsize(zone
));
6146 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
6148 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
6151 pageset_set_high_and_batch(zone
, pcp
);
6154 void __meminit
setup_zone_pageset(struct zone
*zone
)
6157 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
6158 for_each_possible_cpu(cpu
)
6159 zone_pageset_init(zone
, cpu
);
6163 * Allocate per cpu pagesets and initialize them.
6164 * Before this call only boot pagesets were available.
6166 void __init
setup_per_cpu_pageset(void)
6168 struct pglist_data
*pgdat
;
6171 for_each_populated_zone(zone
)
6172 setup_zone_pageset(zone
);
6174 for_each_online_pgdat(pgdat
)
6175 pgdat
->per_cpu_nodestats
=
6176 alloc_percpu(struct per_cpu_nodestat
);
6179 static __meminit
void zone_pcp_init(struct zone
*zone
)
6182 * per cpu subsystem is not up at this point. The following code
6183 * relies on the ability of the linker to provide the
6184 * offset of a (static) per cpu variable into the per cpu area.
6186 zone
->pageset
= &boot_pageset
;
6188 if (populated_zone(zone
))
6189 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
6190 zone
->name
, zone
->present_pages
,
6191 zone_batchsize(zone
));
6194 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6195 unsigned long zone_start_pfn
,
6198 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6199 int zone_idx
= zone_idx(zone
) + 1;
6201 if (zone_idx
> pgdat
->nr_zones
)
6202 pgdat
->nr_zones
= zone_idx
;
6204 zone
->zone_start_pfn
= zone_start_pfn
;
6206 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6207 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6209 (unsigned long)zone_idx(zone
),
6210 zone_start_pfn
, (zone_start_pfn
+ size
));
6212 zone_init_free_lists(zone
);
6213 zone
->initialized
= 1;
6216 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6217 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6220 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6222 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
6223 struct mminit_pfnnid_cache
*state
)
6225 unsigned long start_pfn
, end_pfn
;
6228 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
6229 return state
->last_nid
;
6231 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
6232 if (nid
!= NUMA_NO_NODE
) {
6233 state
->last_start
= start_pfn
;
6234 state
->last_end
= end_pfn
;
6235 state
->last_nid
= nid
;
6240 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6243 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6244 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6245 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6247 * If an architecture guarantees that all ranges registered contain no holes
6248 * and may be freed, this this function may be used instead of calling
6249 * memblock_free_early_nid() manually.
6251 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
6253 unsigned long start_pfn
, end_pfn
;
6256 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
6257 start_pfn
= min(start_pfn
, max_low_pfn
);
6258 end_pfn
= min(end_pfn
, max_low_pfn
);
6260 if (start_pfn
< end_pfn
)
6261 memblock_free_early_nid(PFN_PHYS(start_pfn
),
6262 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
6268 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6269 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6271 * If an architecture guarantees that all ranges registered contain no holes and may
6272 * be freed, this function may be used instead of calling memory_present() manually.
6274 void __init
sparse_memory_present_with_active_regions(int nid
)
6276 unsigned long start_pfn
, end_pfn
;
6279 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
6280 memory_present(this_nid
, start_pfn
, end_pfn
);
6284 * get_pfn_range_for_nid - Return the start and end page frames for a node
6285 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6286 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6287 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6289 * It returns the start and end page frame of a node based on information
6290 * provided by memblock_set_node(). If called for a node
6291 * with no available memory, a warning is printed and the start and end
6294 void __init
get_pfn_range_for_nid(unsigned int nid
,
6295 unsigned long *start_pfn
, unsigned long *end_pfn
)
6297 unsigned long this_start_pfn
, this_end_pfn
;
6303 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6304 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6305 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6308 if (*start_pfn
== -1UL)
6313 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6314 * assumption is made that zones within a node are ordered in monotonic
6315 * increasing memory addresses so that the "highest" populated zone is used
6317 static void __init
find_usable_zone_for_movable(void)
6320 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6321 if (zone_index
== ZONE_MOVABLE
)
6324 if (arch_zone_highest_possible_pfn
[zone_index
] >
6325 arch_zone_lowest_possible_pfn
[zone_index
])
6329 VM_BUG_ON(zone_index
== -1);
6330 movable_zone
= zone_index
;
6334 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6335 * because it is sized independent of architecture. Unlike the other zones,
6336 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6337 * in each node depending on the size of each node and how evenly kernelcore
6338 * is distributed. This helper function adjusts the zone ranges
6339 * provided by the architecture for a given node by using the end of the
6340 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6341 * zones within a node are in order of monotonic increases memory addresses
6343 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6344 unsigned long zone_type
,
6345 unsigned long node_start_pfn
,
6346 unsigned long node_end_pfn
,
6347 unsigned long *zone_start_pfn
,
6348 unsigned long *zone_end_pfn
)
6350 /* Only adjust if ZONE_MOVABLE is on this node */
6351 if (zone_movable_pfn
[nid
]) {
6352 /* Size ZONE_MOVABLE */
6353 if (zone_type
== ZONE_MOVABLE
) {
6354 *zone_start_pfn
= zone_movable_pfn
[nid
];
6355 *zone_end_pfn
= min(node_end_pfn
,
6356 arch_zone_highest_possible_pfn
[movable_zone
]);
6358 /* Adjust for ZONE_MOVABLE starting within this range */
6359 } else if (!mirrored_kernelcore
&&
6360 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6361 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6362 *zone_end_pfn
= zone_movable_pfn
[nid
];
6364 /* Check if this whole range is within ZONE_MOVABLE */
6365 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6366 *zone_start_pfn
= *zone_end_pfn
;
6371 * Return the number of pages a zone spans in a node, including holes
6372 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6374 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6375 unsigned long zone_type
,
6376 unsigned long node_start_pfn
,
6377 unsigned long node_end_pfn
,
6378 unsigned long *zone_start_pfn
,
6379 unsigned long *zone_end_pfn
,
6380 unsigned long *ignored
)
6382 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6383 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6384 /* When hotadd a new node from cpu_up(), the node should be empty */
6385 if (!node_start_pfn
&& !node_end_pfn
)
6388 /* Get the start and end of the zone */
6389 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6390 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6391 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6392 node_start_pfn
, node_end_pfn
,
6393 zone_start_pfn
, zone_end_pfn
);
6395 /* Check that this node has pages within the zone's required range */
6396 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6399 /* Move the zone boundaries inside the node if necessary */
6400 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6401 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6403 /* Return the spanned pages */
6404 return *zone_end_pfn
- *zone_start_pfn
;
6408 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6409 * then all holes in the requested range will be accounted for.
6411 unsigned long __init
__absent_pages_in_range(int nid
,
6412 unsigned long range_start_pfn
,
6413 unsigned long range_end_pfn
)
6415 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6416 unsigned long start_pfn
, end_pfn
;
6419 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6420 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6421 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6422 nr_absent
-= end_pfn
- start_pfn
;
6428 * absent_pages_in_range - Return number of page frames in holes within a range
6429 * @start_pfn: The start PFN to start searching for holes
6430 * @end_pfn: The end PFN to stop searching for holes
6432 * Return: the number of pages frames in memory holes within a range.
6434 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6435 unsigned long end_pfn
)
6437 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6440 /* Return the number of page frames in holes in a zone on a node */
6441 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6442 unsigned long zone_type
,
6443 unsigned long node_start_pfn
,
6444 unsigned long node_end_pfn
,
6445 unsigned long *ignored
)
6447 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6448 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6449 unsigned long zone_start_pfn
, zone_end_pfn
;
6450 unsigned long nr_absent
;
6452 /* When hotadd a new node from cpu_up(), the node should be empty */
6453 if (!node_start_pfn
&& !node_end_pfn
)
6456 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6457 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6459 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6460 node_start_pfn
, node_end_pfn
,
6461 &zone_start_pfn
, &zone_end_pfn
);
6462 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6465 * ZONE_MOVABLE handling.
6466 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6469 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6470 unsigned long start_pfn
, end_pfn
;
6471 struct memblock_region
*r
;
6473 for_each_memblock(memory
, r
) {
6474 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6475 zone_start_pfn
, zone_end_pfn
);
6476 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6477 zone_start_pfn
, zone_end_pfn
);
6479 if (zone_type
== ZONE_MOVABLE
&&
6480 memblock_is_mirror(r
))
6481 nr_absent
+= end_pfn
- start_pfn
;
6483 if (zone_type
== ZONE_NORMAL
&&
6484 !memblock_is_mirror(r
))
6485 nr_absent
+= end_pfn
- start_pfn
;
6492 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6493 static inline unsigned long __init
zone_spanned_pages_in_node(int nid
,
6494 unsigned long zone_type
,
6495 unsigned long node_start_pfn
,
6496 unsigned long node_end_pfn
,
6497 unsigned long *zone_start_pfn
,
6498 unsigned long *zone_end_pfn
,
6499 unsigned long *zones_size
)
6503 *zone_start_pfn
= node_start_pfn
;
6504 for (zone
= 0; zone
< zone_type
; zone
++)
6505 *zone_start_pfn
+= zones_size
[zone
];
6507 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
6509 return zones_size
[zone_type
];
6512 static inline unsigned long __init
zone_absent_pages_in_node(int nid
,
6513 unsigned long zone_type
,
6514 unsigned long node_start_pfn
,
6515 unsigned long node_end_pfn
,
6516 unsigned long *zholes_size
)
6521 return zholes_size
[zone_type
];
6524 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6526 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6527 unsigned long node_start_pfn
,
6528 unsigned long node_end_pfn
,
6529 unsigned long *zones_size
,
6530 unsigned long *zholes_size
)
6532 unsigned long realtotalpages
= 0, totalpages
= 0;
6535 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6536 struct zone
*zone
= pgdat
->node_zones
+ i
;
6537 unsigned long zone_start_pfn
, zone_end_pfn
;
6538 unsigned long size
, real_size
;
6540 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6546 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
6547 node_start_pfn
, node_end_pfn
,
6550 zone
->zone_start_pfn
= zone_start_pfn
;
6552 zone
->zone_start_pfn
= 0;
6553 zone
->spanned_pages
= size
;
6554 zone
->present_pages
= real_size
;
6557 realtotalpages
+= real_size
;
6560 pgdat
->node_spanned_pages
= totalpages
;
6561 pgdat
->node_present_pages
= realtotalpages
;
6562 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6566 #ifndef CONFIG_SPARSEMEM
6568 * Calculate the size of the zone->blockflags rounded to an unsigned long
6569 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6570 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6571 * round what is now in bits to nearest long in bits, then return it in
6574 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6576 unsigned long usemapsize
;
6578 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6579 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6580 usemapsize
= usemapsize
>> pageblock_order
;
6581 usemapsize
*= NR_PAGEBLOCK_BITS
;
6582 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6584 return usemapsize
/ 8;
6587 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6589 unsigned long zone_start_pfn
,
6590 unsigned long zonesize
)
6592 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6593 zone
->pageblock_flags
= NULL
;
6595 zone
->pageblock_flags
=
6596 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
6598 if (!zone
->pageblock_flags
)
6599 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6600 usemapsize
, zone
->name
, pgdat
->node_id
);
6604 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6605 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6606 #endif /* CONFIG_SPARSEMEM */
6608 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6610 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6611 void __init
set_pageblock_order(void)
6615 /* Check that pageblock_nr_pages has not already been setup */
6616 if (pageblock_order
)
6619 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6620 order
= HUGETLB_PAGE_ORDER
;
6622 order
= MAX_ORDER
- 1;
6625 * Assume the largest contiguous order of interest is a huge page.
6626 * This value may be variable depending on boot parameters on IA64 and
6629 pageblock_order
= order
;
6631 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6634 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6635 * is unused as pageblock_order is set at compile-time. See
6636 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6639 void __init
set_pageblock_order(void)
6643 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6645 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6646 unsigned long present_pages
)
6648 unsigned long pages
= spanned_pages
;
6651 * Provide a more accurate estimation if there are holes within
6652 * the zone and SPARSEMEM is in use. If there are holes within the
6653 * zone, each populated memory region may cost us one or two extra
6654 * memmap pages due to alignment because memmap pages for each
6655 * populated regions may not be naturally aligned on page boundary.
6656 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6658 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6659 IS_ENABLED(CONFIG_SPARSEMEM
))
6660 pages
= present_pages
;
6662 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6665 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6666 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6668 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
6670 spin_lock_init(&ds_queue
->split_queue_lock
);
6671 INIT_LIST_HEAD(&ds_queue
->split_queue
);
6672 ds_queue
->split_queue_len
= 0;
6675 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6678 #ifdef CONFIG_COMPACTION
6679 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6681 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6684 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6687 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6689 pgdat_resize_init(pgdat
);
6691 pgdat_init_split_queue(pgdat
);
6692 pgdat_init_kcompactd(pgdat
);
6694 init_waitqueue_head(&pgdat
->kswapd_wait
);
6695 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6697 pgdat_page_ext_init(pgdat
);
6698 spin_lock_init(&pgdat
->lru_lock
);
6699 lruvec_init(node_lruvec(pgdat
));
6702 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6703 unsigned long remaining_pages
)
6705 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6706 zone_set_nid(zone
, nid
);
6707 zone
->name
= zone_names
[idx
];
6708 zone
->zone_pgdat
= NODE_DATA(nid
);
6709 spin_lock_init(&zone
->lock
);
6710 zone_seqlock_init(zone
);
6711 zone_pcp_init(zone
);
6715 * Set up the zone data structures
6716 * - init pgdat internals
6717 * - init all zones belonging to this node
6719 * NOTE: this function is only called during memory hotplug
6721 #ifdef CONFIG_MEMORY_HOTPLUG
6722 void __ref
free_area_init_core_hotplug(int nid
)
6725 pg_data_t
*pgdat
= NODE_DATA(nid
);
6727 pgdat_init_internals(pgdat
);
6728 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6729 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6734 * Set up the zone data structures:
6735 * - mark all pages reserved
6736 * - mark all memory queues empty
6737 * - clear the memory bitmaps
6739 * NOTE: pgdat should get zeroed by caller.
6740 * NOTE: this function is only called during early init.
6742 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6745 int nid
= pgdat
->node_id
;
6747 pgdat_init_internals(pgdat
);
6748 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6750 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6751 struct zone
*zone
= pgdat
->node_zones
+ j
;
6752 unsigned long size
, freesize
, memmap_pages
;
6753 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6755 size
= zone
->spanned_pages
;
6756 freesize
= zone
->present_pages
;
6759 * Adjust freesize so that it accounts for how much memory
6760 * is used by this zone for memmap. This affects the watermark
6761 * and per-cpu initialisations
6763 memmap_pages
= calc_memmap_size(size
, freesize
);
6764 if (!is_highmem_idx(j
)) {
6765 if (freesize
>= memmap_pages
) {
6766 freesize
-= memmap_pages
;
6769 " %s zone: %lu pages used for memmap\n",
6770 zone_names
[j
], memmap_pages
);
6772 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6773 zone_names
[j
], memmap_pages
, freesize
);
6776 /* Account for reserved pages */
6777 if (j
== 0 && freesize
> dma_reserve
) {
6778 freesize
-= dma_reserve
;
6779 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6780 zone_names
[0], dma_reserve
);
6783 if (!is_highmem_idx(j
))
6784 nr_kernel_pages
+= freesize
;
6785 /* Charge for highmem memmap if there are enough kernel pages */
6786 else if (nr_kernel_pages
> memmap_pages
* 2)
6787 nr_kernel_pages
-= memmap_pages
;
6788 nr_all_pages
+= freesize
;
6791 * Set an approximate value for lowmem here, it will be adjusted
6792 * when the bootmem allocator frees pages into the buddy system.
6793 * And all highmem pages will be managed by the buddy system.
6795 zone_init_internals(zone
, j
, nid
, freesize
);
6800 set_pageblock_order();
6801 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6802 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6803 memmap_init(size
, nid
, j
, zone_start_pfn
);
6807 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6808 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6810 unsigned long __maybe_unused start
= 0;
6811 unsigned long __maybe_unused offset
= 0;
6813 /* Skip empty nodes */
6814 if (!pgdat
->node_spanned_pages
)
6817 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6818 offset
= pgdat
->node_start_pfn
- start
;
6819 /* ia64 gets its own node_mem_map, before this, without bootmem */
6820 if (!pgdat
->node_mem_map
) {
6821 unsigned long size
, end
;
6825 * The zone's endpoints aren't required to be MAX_ORDER
6826 * aligned but the node_mem_map endpoints must be in order
6827 * for the buddy allocator to function correctly.
6829 end
= pgdat_end_pfn(pgdat
);
6830 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6831 size
= (end
- start
) * sizeof(struct page
);
6832 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
6835 panic("Failed to allocate %ld bytes for node %d memory map\n",
6836 size
, pgdat
->node_id
);
6837 pgdat
->node_mem_map
= map
+ offset
;
6839 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6840 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6841 (unsigned long)pgdat
->node_mem_map
);
6842 #ifndef CONFIG_NEED_MULTIPLE_NODES
6844 * With no DISCONTIG, the global mem_map is just set as node 0's
6846 if (pgdat
== NODE_DATA(0)) {
6847 mem_map
= NODE_DATA(0)->node_mem_map
;
6848 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6849 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6851 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6856 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6857 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6859 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6860 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6862 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6865 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6868 void __init
free_area_init_node(int nid
, unsigned long *zones_size
,
6869 unsigned long node_start_pfn
,
6870 unsigned long *zholes_size
)
6872 pg_data_t
*pgdat
= NODE_DATA(nid
);
6873 unsigned long start_pfn
= 0;
6874 unsigned long end_pfn
= 0;
6876 /* pg_data_t should be reset to zero when it's allocated */
6877 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6879 pgdat
->node_id
= nid
;
6880 pgdat
->node_start_pfn
= node_start_pfn
;
6881 pgdat
->per_cpu_nodestats
= NULL
;
6882 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6883 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6884 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6885 (u64
)start_pfn
<< PAGE_SHIFT
,
6886 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6888 start_pfn
= node_start_pfn
;
6890 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6891 zones_size
, zholes_size
);
6893 alloc_node_mem_map(pgdat
);
6894 pgdat_set_deferred_range(pgdat
);
6896 free_area_init_core(pgdat
);
6899 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6901 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6904 static u64
zero_pfn_range(unsigned long spfn
, unsigned long epfn
)
6909 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6910 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6911 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6912 + pageblock_nr_pages
- 1;
6915 mm_zero_struct_page(pfn_to_page(pfn
));
6923 * Only struct pages that are backed by physical memory are zeroed and
6924 * initialized by going through __init_single_page(). But, there are some
6925 * struct pages which are reserved in memblock allocator and their fields
6926 * may be accessed (for example page_to_pfn() on some configuration accesses
6927 * flags). We must explicitly zero those struct pages.
6929 * This function also addresses a similar issue where struct pages are left
6930 * uninitialized because the physical address range is not covered by
6931 * memblock.memory or memblock.reserved. That could happen when memblock
6932 * layout is manually configured via memmap=, or when the highest physical
6933 * address (max_pfn) does not end on a section boundary.
6935 void __init
zero_resv_unavail(void)
6937 phys_addr_t start
, end
;
6939 phys_addr_t next
= 0;
6942 * Loop through unavailable ranges not covered by memblock.memory.
6945 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6946 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6948 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), PFN_UP(start
));
6953 * Early sections always have a fully populated memmap for the whole
6954 * section - see pfn_valid(). If the last section has holes at the
6955 * end and that section is marked "online", the memmap will be
6956 * considered initialized. Make sure that memmap has a well defined
6959 pgcnt
+= zero_pfn_range(PFN_DOWN(next
),
6960 round_up(max_pfn
, PAGES_PER_SECTION
));
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 postcore_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
);
8516 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8517 * page high values need to be recalulated.
8519 void __meminit
zone_pcp_update(struct zone
*zone
)
8522 mutex_lock(&pcp_batch_high_lock
);
8523 for_each_possible_cpu(cpu
)
8524 pageset_set_high_and_batch(zone
,
8525 per_cpu_ptr(zone
->pageset
, cpu
));
8526 mutex_unlock(&pcp_batch_high_lock
);
8529 void zone_pcp_reset(struct zone
*zone
)
8531 unsigned long flags
;
8533 struct per_cpu_pageset
*pset
;
8535 /* avoid races with drain_pages() */
8536 local_irq_save(flags
);
8537 if (zone
->pageset
!= &boot_pageset
) {
8538 for_each_online_cpu(cpu
) {
8539 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8540 drain_zonestat(zone
, pset
);
8542 free_percpu(zone
->pageset
);
8543 zone
->pageset
= &boot_pageset
;
8545 local_irq_restore(flags
);
8548 #ifdef CONFIG_MEMORY_HOTREMOVE
8550 * All pages in the range must be in a single zone and isolated
8551 * before calling this.
8554 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8558 unsigned int order
, i
;
8560 unsigned long flags
;
8561 unsigned long offlined_pages
= 0;
8563 /* find the first valid pfn */
8564 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8568 return offlined_pages
;
8570 offline_mem_sections(pfn
, end_pfn
);
8571 zone
= page_zone(pfn_to_page(pfn
));
8572 spin_lock_irqsave(&zone
->lock
, flags
);
8574 while (pfn
< end_pfn
) {
8575 if (!pfn_valid(pfn
)) {
8579 page
= pfn_to_page(pfn
);
8581 * The HWPoisoned page may be not in buddy system, and
8582 * page_count() is not 0.
8584 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8586 SetPageReserved(page
);
8591 BUG_ON(page_count(page
));
8592 BUG_ON(!PageBuddy(page
));
8593 order
= page_order(page
);
8594 offlined_pages
+= 1 << order
;
8595 #ifdef CONFIG_DEBUG_VM
8596 pr_info("remove from free list %lx %d %lx\n",
8597 pfn
, 1 << order
, end_pfn
);
8599 del_page_from_free_area(page
, &zone
->free_area
[order
]);
8600 for (i
= 0; i
< (1 << order
); i
++)
8601 SetPageReserved((page
+i
));
8602 pfn
+= (1 << order
);
8604 spin_unlock_irqrestore(&zone
->lock
, flags
);
8606 return offlined_pages
;
8610 bool is_free_buddy_page(struct page
*page
)
8612 struct zone
*zone
= page_zone(page
);
8613 unsigned long pfn
= page_to_pfn(page
);
8614 unsigned long flags
;
8617 spin_lock_irqsave(&zone
->lock
, flags
);
8618 for (order
= 0; order
< MAX_ORDER
; order
++) {
8619 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8621 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8624 spin_unlock_irqrestore(&zone
->lock
, flags
);
8626 return order
< MAX_ORDER
;
8629 #ifdef CONFIG_MEMORY_FAILURE
8631 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8632 * test is performed under the zone lock to prevent a race against page
8635 bool set_hwpoison_free_buddy_page(struct page
*page
)
8637 struct zone
*zone
= page_zone(page
);
8638 unsigned long pfn
= page_to_pfn(page
);
8639 unsigned long flags
;
8641 bool hwpoisoned
= false;
8643 spin_lock_irqsave(&zone
->lock
, flags
);
8644 for (order
= 0; order
< MAX_ORDER
; order
++) {
8645 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8647 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8648 if (!TestSetPageHWPoison(page
))
8653 spin_unlock_irqrestore(&zone
->lock
, flags
);