1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
71 #include <linux/khugepaged.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
75 #include <asm/div64.h>
79 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
80 static DEFINE_MUTEX(pcp_batch_high_lock
);
81 #define MIN_PERCPU_PAGELIST_FRACTION (8)
83 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
84 DEFINE_PER_CPU(int, numa_node
);
85 EXPORT_PER_CPU_SYMBOL(numa_node
);
88 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
90 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
92 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
93 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
94 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
95 * defined in <linux/topology.h>.
97 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
98 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
99 int _node_numa_mem_
[MAX_NUMNODES
];
102 /* work_structs for global per-cpu drains */
105 struct work_struct work
;
107 DEFINE_MUTEX(pcpu_drain_mutex
);
108 DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
110 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
111 volatile unsigned long latent_entropy __latent_entropy
;
112 EXPORT_SYMBOL(latent_entropy
);
116 * Array of node states.
118 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
119 [N_POSSIBLE
] = NODE_MASK_ALL
,
120 [N_ONLINE
] = { { [0] = 1UL } },
122 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
123 #ifdef CONFIG_HIGHMEM
124 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
126 [N_MEMORY
] = { { [0] = 1UL } },
127 [N_CPU
] = { { [0] = 1UL } },
130 EXPORT_SYMBOL(node_states
);
132 atomic_long_t _totalram_pages __read_mostly
;
133 EXPORT_SYMBOL(_totalram_pages
);
134 unsigned long totalreserve_pages __read_mostly
;
135 unsigned long totalcma_pages __read_mostly
;
137 int percpu_pagelist_fraction
;
138 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
139 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
140 DEFINE_STATIC_KEY_TRUE(init_on_alloc
);
142 DEFINE_STATIC_KEY_FALSE(init_on_alloc
);
144 EXPORT_SYMBOL(init_on_alloc
);
146 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
147 DEFINE_STATIC_KEY_TRUE(init_on_free
);
149 DEFINE_STATIC_KEY_FALSE(init_on_free
);
151 EXPORT_SYMBOL(init_on_free
);
153 static int __init
early_init_on_alloc(char *buf
)
160 ret
= kstrtobool(buf
, &bool_result
);
161 if (bool_result
&& page_poisoning_enabled())
162 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
164 static_branch_enable(&init_on_alloc
);
166 static_branch_disable(&init_on_alloc
);
169 early_param("init_on_alloc", early_init_on_alloc
);
171 static int __init
early_init_on_free(char *buf
)
178 ret
= kstrtobool(buf
, &bool_result
);
179 if (bool_result
&& page_poisoning_enabled())
180 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
182 static_branch_enable(&init_on_free
);
184 static_branch_disable(&init_on_free
);
187 early_param("init_on_free", early_init_on_free
);
190 * A cached value of the page's pageblock's migratetype, used when the page is
191 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
192 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
193 * Also the migratetype set in the page does not necessarily match the pcplist
194 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
195 * other index - this ensures that it will be put on the correct CMA freelist.
197 static inline int get_pcppage_migratetype(struct page
*page
)
202 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
204 page
->index
= migratetype
;
207 #ifdef CONFIG_PM_SLEEP
209 * The following functions are used by the suspend/hibernate code to temporarily
210 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
211 * while devices are suspended. To avoid races with the suspend/hibernate code,
212 * they should always be called with system_transition_mutex held
213 * (gfp_allowed_mask also should only be modified with system_transition_mutex
214 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
215 * with that modification).
218 static gfp_t saved_gfp_mask
;
220 void pm_restore_gfp_mask(void)
222 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
223 if (saved_gfp_mask
) {
224 gfp_allowed_mask
= saved_gfp_mask
;
229 void pm_restrict_gfp_mask(void)
231 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
232 WARN_ON(saved_gfp_mask
);
233 saved_gfp_mask
= gfp_allowed_mask
;
234 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
237 bool pm_suspended_storage(void)
239 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
243 #endif /* CONFIG_PM_SLEEP */
245 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
246 unsigned int pageblock_order __read_mostly
;
249 static void __free_pages_ok(struct page
*page
, unsigned int order
);
252 * results with 256, 32 in the lowmem_reserve sysctl:
253 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
254 * 1G machine -> (16M dma, 784M normal, 224M high)
255 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
256 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
257 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
259 * TBD: should special case ZONE_DMA32 machines here - in those we normally
260 * don't need any ZONE_NORMAL reservation
262 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
263 #ifdef CONFIG_ZONE_DMA
266 #ifdef CONFIG_ZONE_DMA32
270 #ifdef CONFIG_HIGHMEM
276 static char * const zone_names
[MAX_NR_ZONES
] = {
277 #ifdef CONFIG_ZONE_DMA
280 #ifdef CONFIG_ZONE_DMA32
284 #ifdef CONFIG_HIGHMEM
288 #ifdef CONFIG_ZONE_DEVICE
293 const char * const migratetype_names
[MIGRATE_TYPES
] = {
301 #ifdef CONFIG_MEMORY_ISOLATION
306 compound_page_dtor
* const compound_page_dtors
[] = {
309 #ifdef CONFIG_HUGETLB_PAGE
312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
317 int min_free_kbytes
= 1024;
318 int user_min_free_kbytes
= -1;
319 int watermark_boost_factor __read_mostly
;
320 int watermark_scale_factor
= 10;
322 static unsigned long nr_kernel_pages __initdata
;
323 static unsigned long nr_all_pages __initdata
;
324 static unsigned long dma_reserve __initdata
;
326 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
327 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
328 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
329 static unsigned long required_kernelcore __initdata
;
330 static unsigned long required_kernelcore_percent __initdata
;
331 static unsigned long required_movablecore __initdata
;
332 static unsigned long required_movablecore_percent __initdata
;
333 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
334 static bool mirrored_kernelcore __meminitdata
;
336 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
338 EXPORT_SYMBOL(movable_zone
);
339 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
342 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
343 unsigned int nr_online_nodes __read_mostly
= 1;
344 EXPORT_SYMBOL(nr_node_ids
);
345 EXPORT_SYMBOL(nr_online_nodes
);
348 int page_group_by_mobility_disabled __read_mostly
;
350 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
352 * During boot we initialize deferred pages on-demand, as needed, but once
353 * page_alloc_init_late() has finished, the deferred pages are all initialized,
354 * and we can permanently disable that path.
356 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
359 * Calling kasan_free_pages() only after deferred memory initialization
360 * has completed. Poisoning pages during deferred memory init will greatly
361 * lengthen the process and cause problem in large memory systems as the
362 * deferred pages initialization is done with interrupt disabled.
364 * Assuming that there will be no reference to those newly initialized
365 * pages before they are ever allocated, this should have no effect on
366 * KASAN memory tracking as the poison will be properly inserted at page
367 * allocation time. The only corner case is when pages are allocated by
368 * on-demand allocation and then freed again before the deferred pages
369 * initialization is done, but this is not likely to happen.
371 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
373 if (!static_branch_unlikely(&deferred_pages
))
374 kasan_free_pages(page
, order
);
377 /* Returns true if the struct page for the pfn is uninitialised */
378 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
380 int nid
= early_pfn_to_nid(pfn
);
382 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
389 * Returns true when the remaining initialisation should be deferred until
390 * later in the boot cycle when it can be parallelised.
392 static bool __meminit
393 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
395 static unsigned long prev_end_pfn
, nr_initialised
;
398 * prev_end_pfn static that contains the end of previous zone
399 * No need to protect because called very early in boot before smp_init.
401 if (prev_end_pfn
!= end_pfn
) {
402 prev_end_pfn
= end_pfn
;
406 /* Always populate low zones for address-constrained allocations */
407 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
411 * We start only with one section of pages, more pages are added as
412 * needed until the rest of deferred pages are initialized.
415 if ((nr_initialised
> PAGES_PER_SECTION
) &&
416 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
417 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
423 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
425 static inline bool early_page_uninitialised(unsigned long pfn
)
430 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
436 /* Return a pointer to the bitmap storing bits affecting a block of pages */
437 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
440 #ifdef CONFIG_SPARSEMEM
441 return section_to_usemap(__pfn_to_section(pfn
));
443 return page_zone(page
)->pageblock_flags
;
444 #endif /* CONFIG_SPARSEMEM */
447 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
449 #ifdef CONFIG_SPARSEMEM
450 pfn
&= (PAGES_PER_SECTION
-1);
451 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
453 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
454 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
455 #endif /* CONFIG_SPARSEMEM */
459 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
460 * @page: The page within the block of interest
461 * @pfn: The target page frame number
462 * @end_bitidx: The last bit of interest to retrieve
463 * @mask: mask of bits that the caller is interested in
465 * Return: pageblock_bits flags
467 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
469 unsigned long end_bitidx
,
472 unsigned long *bitmap
;
473 unsigned long bitidx
, word_bitidx
;
476 bitmap
= get_pageblock_bitmap(page
, pfn
);
477 bitidx
= pfn_to_bitidx(page
, pfn
);
478 word_bitidx
= bitidx
/ BITS_PER_LONG
;
479 bitidx
&= (BITS_PER_LONG
-1);
481 word
= bitmap
[word_bitidx
];
482 bitidx
+= end_bitidx
;
483 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
486 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
487 unsigned long end_bitidx
,
490 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
493 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
495 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
499 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
500 * @page: The page within the block of interest
501 * @flags: The flags to set
502 * @pfn: The target page frame number
503 * @end_bitidx: The last bit of interest
504 * @mask: mask of bits that the caller is interested in
506 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
508 unsigned long end_bitidx
,
511 unsigned long *bitmap
;
512 unsigned long bitidx
, word_bitidx
;
513 unsigned long old_word
, word
;
515 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
516 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
518 bitmap
= get_pageblock_bitmap(page
, pfn
);
519 bitidx
= pfn_to_bitidx(page
, pfn
);
520 word_bitidx
= bitidx
/ BITS_PER_LONG
;
521 bitidx
&= (BITS_PER_LONG
-1);
523 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
525 bitidx
+= end_bitidx
;
526 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
527 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
529 word
= READ_ONCE(bitmap
[word_bitidx
]);
531 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
532 if (word
== old_word
)
538 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
540 if (unlikely(page_group_by_mobility_disabled
&&
541 migratetype
< MIGRATE_PCPTYPES
))
542 migratetype
= MIGRATE_UNMOVABLE
;
544 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
545 PB_migrate
, PB_migrate_end
);
548 #ifdef CONFIG_DEBUG_VM
549 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
553 unsigned long pfn
= page_to_pfn(page
);
554 unsigned long sp
, start_pfn
;
557 seq
= zone_span_seqbegin(zone
);
558 start_pfn
= zone
->zone_start_pfn
;
559 sp
= zone
->spanned_pages
;
560 if (!zone_spans_pfn(zone
, pfn
))
562 } while (zone_span_seqretry(zone
, seq
));
565 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
566 pfn
, zone_to_nid(zone
), zone
->name
,
567 start_pfn
, start_pfn
+ sp
);
572 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
574 if (!pfn_valid_within(page_to_pfn(page
)))
576 if (zone
!= page_zone(page
))
582 * Temporary debugging check for pages not lying within a given zone.
584 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
586 if (page_outside_zone_boundaries(zone
, page
))
588 if (!page_is_consistent(zone
, page
))
594 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
600 static void bad_page(struct page
*page
, const char *reason
,
601 unsigned long bad_flags
)
603 static unsigned long resume
;
604 static unsigned long nr_shown
;
605 static unsigned long nr_unshown
;
608 * Allow a burst of 60 reports, then keep quiet for that minute;
609 * or allow a steady drip of one report per second.
611 if (nr_shown
== 60) {
612 if (time_before(jiffies
, resume
)) {
618 "BUG: Bad page state: %lu messages suppressed\n",
625 resume
= jiffies
+ 60 * HZ
;
627 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
628 current
->comm
, page_to_pfn(page
));
629 __dump_page(page
, reason
);
630 bad_flags
&= page
->flags
;
632 pr_alert("bad because of flags: %#lx(%pGp)\n",
633 bad_flags
, &bad_flags
);
634 dump_page_owner(page
);
639 /* Leave bad fields for debug, except PageBuddy could make trouble */
640 page_mapcount_reset(page
); /* remove PageBuddy */
641 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
645 * Higher-order pages are called "compound pages". They are structured thusly:
647 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
649 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
650 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
652 * The first tail page's ->compound_dtor holds the offset in array of compound
653 * page destructors. See compound_page_dtors.
655 * The first tail page's ->compound_order holds the order of allocation.
656 * This usage means that zero-order pages may not be compound.
659 void free_compound_page(struct page
*page
)
661 mem_cgroup_uncharge(page
);
662 __free_pages_ok(page
, compound_order(page
));
665 void prep_compound_page(struct page
*page
, unsigned int order
)
668 int nr_pages
= 1 << order
;
670 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
671 set_compound_order(page
, order
);
673 for (i
= 1; i
< nr_pages
; i
++) {
674 struct page
*p
= page
+ i
;
675 set_page_count(p
, 0);
676 p
->mapping
= TAIL_MAPPING
;
677 set_compound_head(p
, page
);
679 atomic_set(compound_mapcount_ptr(page
), -1);
682 #ifdef CONFIG_DEBUG_PAGEALLOC
683 unsigned int _debug_guardpage_minorder
;
685 bool _debug_pagealloc_enabled_early __read_mostly
686 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
687 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
688 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
689 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
691 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
693 static int __init
early_debug_pagealloc(char *buf
)
695 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
697 early_param("debug_pagealloc", early_debug_pagealloc
);
699 void init_debug_pagealloc(void)
701 if (!debug_pagealloc_enabled())
704 static_branch_enable(&_debug_pagealloc_enabled
);
706 if (!debug_guardpage_minorder())
709 static_branch_enable(&_debug_guardpage_enabled
);
712 static int __init
debug_guardpage_minorder_setup(char *buf
)
716 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
717 pr_err("Bad debug_guardpage_minorder value\n");
720 _debug_guardpage_minorder
= res
;
721 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
724 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
726 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
727 unsigned int order
, int migratetype
)
729 if (!debug_guardpage_enabled())
732 if (order
>= debug_guardpage_minorder())
735 __SetPageGuard(page
);
736 INIT_LIST_HEAD(&page
->lru
);
737 set_page_private(page
, order
);
738 /* Guard pages are not available for any usage */
739 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
744 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
745 unsigned int order
, int migratetype
)
747 if (!debug_guardpage_enabled())
750 __ClearPageGuard(page
);
752 set_page_private(page
, 0);
753 if (!is_migrate_isolate(migratetype
))
754 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
757 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
758 unsigned int order
, int migratetype
) { return false; }
759 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
760 unsigned int order
, int migratetype
) {}
763 static inline void set_page_order(struct page
*page
, unsigned int order
)
765 set_page_private(page
, order
);
766 __SetPageBuddy(page
);
770 * This function checks whether a page is free && is the buddy
771 * we can coalesce a page and its buddy if
772 * (a) the buddy is not in a hole (check before calling!) &&
773 * (b) the buddy is in the buddy system &&
774 * (c) a page and its buddy have the same order &&
775 * (d) a page and its buddy are in the same zone.
777 * For recording whether a page is in the buddy system, we set PageBuddy.
778 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
780 * For recording page's order, we use page_private(page).
782 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
785 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
786 if (page_zone_id(page
) != page_zone_id(buddy
))
789 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
794 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
796 * zone check is done late to avoid uselessly
797 * calculating zone/node ids for pages that could
800 if (page_zone_id(page
) != page_zone_id(buddy
))
803 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
810 #ifdef CONFIG_COMPACTION
811 static inline struct capture_control
*task_capc(struct zone
*zone
)
813 struct capture_control
*capc
= current
->capture_control
;
816 !(current
->flags
& PF_KTHREAD
) &&
818 capc
->cc
->zone
== zone
&&
819 capc
->cc
->direct_compaction
? capc
: NULL
;
823 compaction_capture(struct capture_control
*capc
, struct page
*page
,
824 int order
, int migratetype
)
826 if (!capc
|| order
!= capc
->cc
->order
)
829 /* Do not accidentally pollute CMA or isolated regions*/
830 if (is_migrate_cma(migratetype
) ||
831 is_migrate_isolate(migratetype
))
835 * Do not let lower order allocations polluate a movable pageblock.
836 * This might let an unmovable request use a reclaimable pageblock
837 * and vice-versa but no more than normal fallback logic which can
838 * have trouble finding a high-order free page.
840 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
848 static inline struct capture_control
*task_capc(struct zone
*zone
)
854 compaction_capture(struct capture_control
*capc
, struct page
*page
,
855 int order
, int migratetype
)
859 #endif /* CONFIG_COMPACTION */
862 * Freeing function for a buddy system allocator.
864 * The concept of a buddy system is to maintain direct-mapped table
865 * (containing bit values) for memory blocks of various "orders".
866 * The bottom level table contains the map for the smallest allocatable
867 * units of memory (here, pages), and each level above it describes
868 * pairs of units from the levels below, hence, "buddies".
869 * At a high level, all that happens here is marking the table entry
870 * at the bottom level available, and propagating the changes upward
871 * as necessary, plus some accounting needed to play nicely with other
872 * parts of the VM system.
873 * At each level, we keep a list of pages, which are heads of continuous
874 * free pages of length of (1 << order) and marked with PageBuddy.
875 * Page's order is recorded in page_private(page) field.
876 * So when we are allocating or freeing one, we can derive the state of the
877 * other. That is, if we allocate a small block, and both were
878 * free, the remainder of the region must be split into blocks.
879 * If a block is freed, and its buddy is also free, then this
880 * triggers coalescing into a block of larger size.
885 static inline void __free_one_page(struct page
*page
,
887 struct zone
*zone
, unsigned int order
,
890 unsigned long combined_pfn
;
891 unsigned long uninitialized_var(buddy_pfn
);
893 unsigned int max_order
;
894 struct capture_control
*capc
= task_capc(zone
);
896 max_order
= min_t(unsigned int, MAX_ORDER
- 1, pageblock_order
);
898 VM_BUG_ON(!zone_is_initialized(zone
));
899 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
901 VM_BUG_ON(migratetype
== -1);
902 if (likely(!is_migrate_isolate(migratetype
)))
903 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
905 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
906 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
909 while (order
< max_order
) {
910 if (compaction_capture(capc
, page
, order
, migratetype
)) {
911 __mod_zone_freepage_state(zone
, -(1 << order
),
915 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
916 buddy
= page
+ (buddy_pfn
- pfn
);
918 if (!pfn_valid_within(buddy_pfn
))
920 if (!page_is_buddy(page
, buddy
, order
))
923 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
924 * merge with it and move up one order.
926 if (page_is_guard(buddy
))
927 clear_page_guard(zone
, buddy
, order
, migratetype
);
929 del_page_from_free_area(buddy
, &zone
->free_area
[order
]);
930 combined_pfn
= buddy_pfn
& pfn
;
931 page
= page
+ (combined_pfn
- pfn
);
935 if (order
< MAX_ORDER
- 1) {
936 /* If we are here, it means order is >= pageblock_order.
937 * We want to prevent merge between freepages on isolate
938 * pageblock and normal pageblock. Without this, pageblock
939 * isolation could cause incorrect freepage or CMA accounting.
941 * We don't want to hit this code for the more frequent
944 if (unlikely(has_isolate_pageblock(zone
))) {
947 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
948 buddy
= page
+ (buddy_pfn
- pfn
);
949 buddy_mt
= get_pageblock_migratetype(buddy
);
951 if (migratetype
!= buddy_mt
952 && (is_migrate_isolate(migratetype
) ||
953 is_migrate_isolate(buddy_mt
)))
956 max_order
= order
+ 1;
957 goto continue_merging
;
961 set_page_order(page
, order
);
964 * If this is not the largest possible page, check if the buddy
965 * of the next-highest order is free. If it is, it's possible
966 * that pages are being freed that will coalesce soon. In case,
967 * that is happening, add the free page to the tail of the list
968 * so it's less likely to be used soon and more likely to be merged
969 * as a higher order page
971 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)
972 && !is_shuffle_order(order
)) {
973 struct page
*higher_page
, *higher_buddy
;
974 combined_pfn
= buddy_pfn
& pfn
;
975 higher_page
= page
+ (combined_pfn
- pfn
);
976 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
977 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
978 if (pfn_valid_within(buddy_pfn
) &&
979 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
980 add_to_free_area_tail(page
, &zone
->free_area
[order
],
986 if (is_shuffle_order(order
))
987 add_to_free_area_random(page
, &zone
->free_area
[order
],
990 add_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
995 * A bad page could be due to a number of fields. Instead of multiple branches,
996 * try and check multiple fields with one check. The caller must do a detailed
997 * check if necessary.
999 static inline bool page_expected_state(struct page
*page
,
1000 unsigned long check_flags
)
1002 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1005 if (unlikely((unsigned long)page
->mapping
|
1006 page_ref_count(page
) |
1008 (unsigned long)page
->mem_cgroup
|
1010 (page
->flags
& check_flags
)))
1016 static void free_pages_check_bad(struct page
*page
)
1018 const char *bad_reason
;
1019 unsigned long bad_flags
;
1024 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1025 bad_reason
= "nonzero mapcount";
1026 if (unlikely(page
->mapping
!= NULL
))
1027 bad_reason
= "non-NULL mapping";
1028 if (unlikely(page_ref_count(page
) != 0))
1029 bad_reason
= "nonzero _refcount";
1030 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
1031 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1032 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
1035 if (unlikely(page
->mem_cgroup
))
1036 bad_reason
= "page still charged to cgroup";
1038 bad_page(page
, bad_reason
, bad_flags
);
1041 static inline int free_pages_check(struct page
*page
)
1043 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1046 /* Something has gone sideways, find it */
1047 free_pages_check_bad(page
);
1051 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1056 * We rely page->lru.next never has bit 0 set, unless the page
1057 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1059 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1061 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1065 switch (page
- head_page
) {
1067 /* the first tail page: ->mapping may be compound_mapcount() */
1068 if (unlikely(compound_mapcount(page
))) {
1069 bad_page(page
, "nonzero compound_mapcount", 0);
1075 * the second tail page: ->mapping is
1076 * deferred_list.next -- ignore value.
1080 if (page
->mapping
!= TAIL_MAPPING
) {
1081 bad_page(page
, "corrupted mapping in tail page", 0);
1086 if (unlikely(!PageTail(page
))) {
1087 bad_page(page
, "PageTail not set", 0);
1090 if (unlikely(compound_head(page
) != head_page
)) {
1091 bad_page(page
, "compound_head not consistent", 0);
1096 page
->mapping
= NULL
;
1097 clear_compound_head(page
);
1101 static void kernel_init_free_pages(struct page
*page
, int numpages
)
1105 for (i
= 0; i
< numpages
; i
++)
1106 clear_highpage(page
+ i
);
1109 static __always_inline
bool free_pages_prepare(struct page
*page
,
1110 unsigned int order
, bool check_free
)
1114 VM_BUG_ON_PAGE(PageTail(page
), page
);
1116 trace_mm_page_free(page
, order
);
1119 * Check tail pages before head page information is cleared to
1120 * avoid checking PageCompound for order-0 pages.
1122 if (unlikely(order
)) {
1123 bool compound
= PageCompound(page
);
1126 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1129 ClearPageDoubleMap(page
);
1130 for (i
= 1; i
< (1 << order
); i
++) {
1132 bad
+= free_tail_pages_check(page
, page
+ i
);
1133 if (unlikely(free_pages_check(page
+ i
))) {
1137 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1140 if (PageMappingFlags(page
))
1141 page
->mapping
= NULL
;
1142 if (memcg_kmem_enabled() && PageKmemcg(page
))
1143 __memcg_kmem_uncharge(page
, order
);
1145 bad
+= free_pages_check(page
);
1149 page_cpupid_reset_last(page
);
1150 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1151 reset_page_owner(page
, order
);
1153 if (!PageHighMem(page
)) {
1154 debug_check_no_locks_freed(page_address(page
),
1155 PAGE_SIZE
<< order
);
1156 debug_check_no_obj_freed(page_address(page
),
1157 PAGE_SIZE
<< order
);
1159 if (want_init_on_free())
1160 kernel_init_free_pages(page
, 1 << order
);
1162 kernel_poison_pages(page
, 1 << order
, 0);
1164 * arch_free_page() can make the page's contents inaccessible. s390
1165 * does this. So nothing which can access the page's contents should
1166 * happen after this.
1168 arch_free_page(page
, order
);
1170 if (debug_pagealloc_enabled_static())
1171 kernel_map_pages(page
, 1 << order
, 0);
1173 kasan_free_nondeferred_pages(page
, order
);
1178 #ifdef CONFIG_DEBUG_VM
1180 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1181 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1182 * moved from pcp lists to free lists.
1184 static bool free_pcp_prepare(struct page
*page
)
1186 return free_pages_prepare(page
, 0, true);
1189 static bool bulkfree_pcp_prepare(struct page
*page
)
1191 if (debug_pagealloc_enabled_static())
1192 return free_pages_check(page
);
1198 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1199 * moving from pcp lists to free list in order to reduce overhead. With
1200 * debug_pagealloc enabled, they are checked also immediately when being freed
1203 static bool free_pcp_prepare(struct page
*page
)
1205 if (debug_pagealloc_enabled_static())
1206 return free_pages_prepare(page
, 0, true);
1208 return free_pages_prepare(page
, 0, false);
1211 static bool bulkfree_pcp_prepare(struct page
*page
)
1213 return free_pages_check(page
);
1215 #endif /* CONFIG_DEBUG_VM */
1217 static inline void prefetch_buddy(struct page
*page
)
1219 unsigned long pfn
= page_to_pfn(page
);
1220 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1221 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1227 * Frees a number of pages from the PCP lists
1228 * Assumes all pages on list are in same zone, and of same order.
1229 * count is the number of pages to free.
1231 * If the zone was previously in an "all pages pinned" state then look to
1232 * see if this freeing clears that state.
1234 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1235 * pinned" detection logic.
1237 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1238 struct per_cpu_pages
*pcp
)
1240 int migratetype
= 0;
1242 int prefetch_nr
= 0;
1243 bool isolated_pageblocks
;
1244 struct page
*page
, *tmp
;
1248 * Ensure proper count is passed which otherwise would stuck in the
1249 * below while (list_empty(list)) loop.
1251 count
= min(pcp
->count
, count
);
1253 struct list_head
*list
;
1256 * Remove pages from lists in a round-robin fashion. A
1257 * batch_free count is maintained that is incremented when an
1258 * empty list is encountered. This is so more pages are freed
1259 * off fuller lists instead of spinning excessively around empty
1264 if (++migratetype
== MIGRATE_PCPTYPES
)
1266 list
= &pcp
->lists
[migratetype
];
1267 } while (list_empty(list
));
1269 /* This is the only non-empty list. Free them all. */
1270 if (batch_free
== MIGRATE_PCPTYPES
)
1274 page
= list_last_entry(list
, struct page
, lru
);
1275 /* must delete to avoid corrupting pcp list */
1276 list_del(&page
->lru
);
1279 if (bulkfree_pcp_prepare(page
))
1282 list_add_tail(&page
->lru
, &head
);
1285 * We are going to put the page back to the global
1286 * pool, prefetch its buddy to speed up later access
1287 * under zone->lock. It is believed the overhead of
1288 * an additional test and calculating buddy_pfn here
1289 * can be offset by reduced memory latency later. To
1290 * avoid excessive prefetching due to large count, only
1291 * prefetch buddy for the first pcp->batch nr of pages.
1293 if (prefetch_nr
++ < pcp
->batch
)
1294 prefetch_buddy(page
);
1295 } while (--count
&& --batch_free
&& !list_empty(list
));
1298 spin_lock(&zone
->lock
);
1299 isolated_pageblocks
= has_isolate_pageblock(zone
);
1302 * Use safe version since after __free_one_page(),
1303 * page->lru.next will not point to original list.
1305 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1306 int mt
= get_pcppage_migratetype(page
);
1307 /* MIGRATE_ISOLATE page should not go to pcplists */
1308 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1309 /* Pageblock could have been isolated meanwhile */
1310 if (unlikely(isolated_pageblocks
))
1311 mt
= get_pageblock_migratetype(page
);
1313 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1314 trace_mm_page_pcpu_drain(page
, 0, mt
);
1316 spin_unlock(&zone
->lock
);
1319 static void free_one_page(struct zone
*zone
,
1320 struct page
*page
, unsigned long pfn
,
1324 spin_lock(&zone
->lock
);
1325 if (unlikely(has_isolate_pageblock(zone
) ||
1326 is_migrate_isolate(migratetype
))) {
1327 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1329 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1330 spin_unlock(&zone
->lock
);
1333 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1334 unsigned long zone
, int nid
)
1336 mm_zero_struct_page(page
);
1337 set_page_links(page
, zone
, nid
, pfn
);
1338 init_page_count(page
);
1339 page_mapcount_reset(page
);
1340 page_cpupid_reset_last(page
);
1341 page_kasan_tag_reset(page
);
1343 INIT_LIST_HEAD(&page
->lru
);
1344 #ifdef WANT_PAGE_VIRTUAL
1345 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1346 if (!is_highmem_idx(zone
))
1347 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1351 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1352 static void __meminit
init_reserved_page(unsigned long pfn
)
1357 if (!early_page_uninitialised(pfn
))
1360 nid
= early_pfn_to_nid(pfn
);
1361 pgdat
= NODE_DATA(nid
);
1363 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1364 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1366 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1369 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1372 static inline void init_reserved_page(unsigned long pfn
)
1375 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1378 * Initialised pages do not have PageReserved set. This function is
1379 * called for each range allocated by the bootmem allocator and
1380 * marks the pages PageReserved. The remaining valid pages are later
1381 * sent to the buddy page allocator.
1383 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1385 unsigned long start_pfn
= PFN_DOWN(start
);
1386 unsigned long end_pfn
= PFN_UP(end
);
1388 for (; start_pfn
< end_pfn
; start_pfn
++) {
1389 if (pfn_valid(start_pfn
)) {
1390 struct page
*page
= pfn_to_page(start_pfn
);
1392 init_reserved_page(start_pfn
);
1394 /* Avoid false-positive PageTail() */
1395 INIT_LIST_HEAD(&page
->lru
);
1398 * no need for atomic set_bit because the struct
1399 * page is not visible yet so nobody should
1402 __SetPageReserved(page
);
1407 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1409 unsigned long flags
;
1411 unsigned long pfn
= page_to_pfn(page
);
1413 if (!free_pages_prepare(page
, order
, true))
1416 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1417 local_irq_save(flags
);
1418 __count_vm_events(PGFREE
, 1 << order
);
1419 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1420 local_irq_restore(flags
);
1423 void __free_pages_core(struct page
*page
, unsigned int order
)
1425 unsigned int nr_pages
= 1 << order
;
1426 struct page
*p
= page
;
1430 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1432 __ClearPageReserved(p
);
1433 set_page_count(p
, 0);
1435 __ClearPageReserved(p
);
1436 set_page_count(p
, 0);
1438 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1439 set_page_refcounted(page
);
1440 __free_pages(page
, order
);
1443 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1444 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1446 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1448 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1450 static DEFINE_SPINLOCK(early_pfn_lock
);
1453 spin_lock(&early_pfn_lock
);
1454 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1456 nid
= first_online_node
;
1457 spin_unlock(&early_pfn_lock
);
1463 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1464 /* Only safe to use early in boot when initialisation is single-threaded */
1465 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1469 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1470 if (nid
>= 0 && nid
!= node
)
1476 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1483 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1486 if (early_page_uninitialised(pfn
))
1488 __free_pages_core(page
, order
);
1492 * Check that the whole (or subset of) a pageblock given by the interval of
1493 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1494 * with the migration of free compaction scanner. The scanners then need to
1495 * use only pfn_valid_within() check for arches that allow holes within
1498 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1500 * It's possible on some configurations to have a setup like node0 node1 node0
1501 * i.e. it's possible that all pages within a zones range of pages do not
1502 * belong to a single zone. We assume that a border between node0 and node1
1503 * can occur within a single pageblock, but not a node0 node1 node0
1504 * interleaving within a single pageblock. It is therefore sufficient to check
1505 * the first and last page of a pageblock and avoid checking each individual
1506 * page in a pageblock.
1508 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1509 unsigned long end_pfn
, struct zone
*zone
)
1511 struct page
*start_page
;
1512 struct page
*end_page
;
1514 /* end_pfn is one past the range we are checking */
1517 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1520 start_page
= pfn_to_online_page(start_pfn
);
1524 if (page_zone(start_page
) != zone
)
1527 end_page
= pfn_to_page(end_pfn
);
1529 /* This gives a shorter code than deriving page_zone(end_page) */
1530 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1536 void set_zone_contiguous(struct zone
*zone
)
1538 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1539 unsigned long block_end_pfn
;
1541 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1542 for (; block_start_pfn
< zone_end_pfn(zone
);
1543 block_start_pfn
= block_end_pfn
,
1544 block_end_pfn
+= pageblock_nr_pages
) {
1546 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1548 if (!__pageblock_pfn_to_page(block_start_pfn
,
1549 block_end_pfn
, zone
))
1554 /* We confirm that there is no hole */
1555 zone
->contiguous
= true;
1558 void clear_zone_contiguous(struct zone
*zone
)
1560 zone
->contiguous
= false;
1563 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1564 static void __init
deferred_free_range(unsigned long pfn
,
1565 unsigned long nr_pages
)
1573 page
= pfn_to_page(pfn
);
1575 /* Free a large naturally-aligned chunk if possible */
1576 if (nr_pages
== pageblock_nr_pages
&&
1577 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1578 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1579 __free_pages_core(page
, pageblock_order
);
1583 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1584 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1585 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1586 __free_pages_core(page
, 0);
1590 /* Completion tracking for deferred_init_memmap() threads */
1591 static atomic_t pgdat_init_n_undone __initdata
;
1592 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1594 static inline void __init
pgdat_init_report_one_done(void)
1596 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1597 complete(&pgdat_init_all_done_comp
);
1601 * Returns true if page needs to be initialized or freed to buddy allocator.
1603 * First we check if pfn is valid on architectures where it is possible to have
1604 * holes within pageblock_nr_pages. On systems where it is not possible, this
1605 * function is optimized out.
1607 * Then, we check if a current large page is valid by only checking the validity
1610 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1612 if (!pfn_valid_within(pfn
))
1614 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1620 * Free pages to buddy allocator. Try to free aligned pages in
1621 * pageblock_nr_pages sizes.
1623 static void __init
deferred_free_pages(unsigned long pfn
,
1624 unsigned long end_pfn
)
1626 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1627 unsigned long nr_free
= 0;
1629 for (; pfn
< end_pfn
; pfn
++) {
1630 if (!deferred_pfn_valid(pfn
)) {
1631 deferred_free_range(pfn
- nr_free
, nr_free
);
1633 } else if (!(pfn
& nr_pgmask
)) {
1634 deferred_free_range(pfn
- nr_free
, nr_free
);
1640 /* Free the last block of pages to allocator */
1641 deferred_free_range(pfn
- nr_free
, nr_free
);
1645 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1646 * by performing it only once every pageblock_nr_pages.
1647 * Return number of pages initialized.
1649 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1651 unsigned long end_pfn
)
1653 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1654 int nid
= zone_to_nid(zone
);
1655 unsigned long nr_pages
= 0;
1656 int zid
= zone_idx(zone
);
1657 struct page
*page
= NULL
;
1659 for (; pfn
< end_pfn
; pfn
++) {
1660 if (!deferred_pfn_valid(pfn
)) {
1663 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1664 page
= pfn_to_page(pfn
);
1668 __init_single_page(page
, pfn
, zid
, nid
);
1675 * This function is meant to pre-load the iterator for the zone init.
1676 * Specifically it walks through the ranges until we are caught up to the
1677 * first_init_pfn value and exits there. If we never encounter the value we
1678 * return false indicating there are no valid ranges left.
1681 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1682 unsigned long *spfn
, unsigned long *epfn
,
1683 unsigned long first_init_pfn
)
1688 * Start out by walking through the ranges in this zone that have
1689 * already been initialized. We don't need to do anything with them
1690 * so we just need to flush them out of the system.
1692 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1693 if (*epfn
<= first_init_pfn
)
1695 if (*spfn
< first_init_pfn
)
1696 *spfn
= first_init_pfn
;
1705 * Initialize and free pages. We do it in two loops: first we initialize
1706 * struct page, then free to buddy allocator, because while we are
1707 * freeing pages we can access pages that are ahead (computing buddy
1708 * page in __free_one_page()).
1710 * In order to try and keep some memory in the cache we have the loop
1711 * broken along max page order boundaries. This way we will not cause
1712 * any issues with the buddy page computation.
1714 static unsigned long __init
1715 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1716 unsigned long *end_pfn
)
1718 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1719 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1720 unsigned long nr_pages
= 0;
1723 /* First we loop through and initialize the page values */
1724 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1727 if (mo_pfn
<= *start_pfn
)
1730 t
= min(mo_pfn
, *end_pfn
);
1731 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1733 if (mo_pfn
< *end_pfn
) {
1734 *start_pfn
= mo_pfn
;
1739 /* Reset values and now loop through freeing pages as needed */
1742 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1748 t
= min(mo_pfn
, epfn
);
1749 deferred_free_pages(spfn
, t
);
1758 /* Initialise remaining memory on a node */
1759 static int __init
deferred_init_memmap(void *data
)
1761 pg_data_t
*pgdat
= data
;
1762 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1763 unsigned long spfn
= 0, epfn
= 0, nr_pages
= 0;
1764 unsigned long first_init_pfn
, flags
;
1765 unsigned long start
= jiffies
;
1770 /* Bind memory initialisation thread to a local node if possible */
1771 if (!cpumask_empty(cpumask
))
1772 set_cpus_allowed_ptr(current
, cpumask
);
1774 pgdat_resize_lock(pgdat
, &flags
);
1775 first_init_pfn
= pgdat
->first_deferred_pfn
;
1776 if (first_init_pfn
== ULONG_MAX
) {
1777 pgdat_resize_unlock(pgdat
, &flags
);
1778 pgdat_init_report_one_done();
1782 /* Sanity check boundaries */
1783 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1784 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1785 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1788 * Once we unlock here, the zone cannot be grown anymore, thus if an
1789 * interrupt thread must allocate this early in boot, zone must be
1790 * pre-grown prior to start of deferred page initialization.
1792 pgdat_resize_unlock(pgdat
, &flags
);
1794 /* Only the highest zone is deferred so find it */
1795 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1796 zone
= pgdat
->node_zones
+ zid
;
1797 if (first_init_pfn
< zone_end_pfn(zone
))
1801 /* If the zone is empty somebody else may have cleared out the zone */
1802 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1807 * Initialize and free pages in MAX_ORDER sized increments so
1808 * that we can avoid introducing any issues with the buddy
1811 while (spfn
< epfn
) {
1812 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1816 /* Sanity check that the next zone really is unpopulated */
1817 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1819 pr_info("node %d initialised, %lu pages in %ums\n",
1820 pgdat
->node_id
, nr_pages
, jiffies_to_msecs(jiffies
- start
));
1822 pgdat_init_report_one_done();
1827 * If this zone has deferred pages, try to grow it by initializing enough
1828 * deferred pages to satisfy the allocation specified by order, rounded up to
1829 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1830 * of SECTION_SIZE bytes by initializing struct pages in increments of
1831 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1833 * Return true when zone was grown, otherwise return false. We return true even
1834 * when we grow less than requested, to let the caller decide if there are
1835 * enough pages to satisfy the allocation.
1837 * Note: We use noinline because this function is needed only during boot, and
1838 * it is called from a __ref function _deferred_grow_zone. This way we are
1839 * making sure that it is not inlined into permanent text section.
1841 static noinline
bool __init
1842 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1844 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1845 pg_data_t
*pgdat
= zone
->zone_pgdat
;
1846 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1847 unsigned long spfn
, epfn
, flags
;
1848 unsigned long nr_pages
= 0;
1851 /* Only the last zone may have deferred pages */
1852 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1855 pgdat_resize_lock(pgdat
, &flags
);
1858 * If someone grew this zone while we were waiting for spinlock, return
1859 * true, as there might be enough pages already.
1861 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1862 pgdat_resize_unlock(pgdat
, &flags
);
1866 /* If the zone is empty somebody else may have cleared out the zone */
1867 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1868 first_deferred_pfn
)) {
1869 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1870 pgdat_resize_unlock(pgdat
, &flags
);
1871 /* Retry only once. */
1872 return first_deferred_pfn
!= ULONG_MAX
;
1876 * Initialize and free pages in MAX_ORDER sized increments so
1877 * that we can avoid introducing any issues with the buddy
1880 while (spfn
< epfn
) {
1881 /* update our first deferred PFN for this section */
1882 first_deferred_pfn
= spfn
;
1884 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1885 touch_nmi_watchdog();
1887 /* We should only stop along section boundaries */
1888 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
1891 /* If our quota has been met we can stop here */
1892 if (nr_pages
>= nr_pages_needed
)
1896 pgdat
->first_deferred_pfn
= spfn
;
1897 pgdat_resize_unlock(pgdat
, &flags
);
1899 return nr_pages
> 0;
1903 * deferred_grow_zone() is __init, but it is called from
1904 * get_page_from_freelist() during early boot until deferred_pages permanently
1905 * disables this call. This is why we have refdata wrapper to avoid warning,
1906 * and to ensure that the function body gets unloaded.
1909 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1911 return deferred_grow_zone(zone
, order
);
1914 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1916 void __init
page_alloc_init_late(void)
1921 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1923 /* There will be num_node_state(N_MEMORY) threads */
1924 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1925 for_each_node_state(nid
, N_MEMORY
) {
1926 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1929 /* Block until all are initialised */
1930 wait_for_completion(&pgdat_init_all_done_comp
);
1933 * The number of managed pages has changed due to the initialisation
1934 * so the pcpu batch and high limits needs to be updated or the limits
1935 * will be artificially small.
1937 for_each_populated_zone(zone
)
1938 zone_pcp_update(zone
);
1941 * We initialized the rest of the deferred pages. Permanently disable
1942 * on-demand struct page initialization.
1944 static_branch_disable(&deferred_pages
);
1946 /* Reinit limits that are based on free pages after the kernel is up */
1947 files_maxfiles_init();
1950 /* Discard memblock private memory */
1953 for_each_node_state(nid
, N_MEMORY
)
1954 shuffle_free_memory(NODE_DATA(nid
));
1956 for_each_populated_zone(zone
)
1957 set_zone_contiguous(zone
);
1961 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1962 void __init
init_cma_reserved_pageblock(struct page
*page
)
1964 unsigned i
= pageblock_nr_pages
;
1965 struct page
*p
= page
;
1968 __ClearPageReserved(p
);
1969 set_page_count(p
, 0);
1972 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1974 if (pageblock_order
>= MAX_ORDER
) {
1975 i
= pageblock_nr_pages
;
1978 set_page_refcounted(p
);
1979 __free_pages(p
, MAX_ORDER
- 1);
1980 p
+= MAX_ORDER_NR_PAGES
;
1981 } while (i
-= MAX_ORDER_NR_PAGES
);
1983 set_page_refcounted(page
);
1984 __free_pages(page
, pageblock_order
);
1987 adjust_managed_page_count(page
, pageblock_nr_pages
);
1992 * The order of subdivision here is critical for the IO subsystem.
1993 * Please do not alter this order without good reasons and regression
1994 * testing. Specifically, as large blocks of memory are subdivided,
1995 * the order in which smaller blocks are delivered depends on the order
1996 * they're subdivided in this function. This is the primary factor
1997 * influencing the order in which pages are delivered to the IO
1998 * subsystem according to empirical testing, and this is also justified
1999 * by considering the behavior of a buddy system containing a single
2000 * large block of memory acted on by a series of small allocations.
2001 * This behavior is a critical factor in sglist merging's success.
2005 static inline void expand(struct zone
*zone
, struct page
*page
,
2006 int low
, int high
, struct free_area
*area
,
2009 unsigned long size
= 1 << high
;
2011 while (high
> low
) {
2015 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2018 * Mark as guard pages (or page), that will allow to
2019 * merge back to allocator when buddy will be freed.
2020 * Corresponding page table entries will not be touched,
2021 * pages will stay not present in virtual address space
2023 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2026 add_to_free_area(&page
[size
], area
, migratetype
);
2027 set_page_order(&page
[size
], high
);
2031 static void check_new_page_bad(struct page
*page
)
2033 const char *bad_reason
= NULL
;
2034 unsigned long bad_flags
= 0;
2036 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
2037 bad_reason
= "nonzero mapcount";
2038 if (unlikely(page
->mapping
!= NULL
))
2039 bad_reason
= "non-NULL mapping";
2040 if (unlikely(page_ref_count(page
) != 0))
2041 bad_reason
= "nonzero _refcount";
2042 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2043 bad_reason
= "HWPoisoned (hardware-corrupted)";
2044 bad_flags
= __PG_HWPOISON
;
2045 /* Don't complain about hwpoisoned pages */
2046 page_mapcount_reset(page
); /* remove PageBuddy */
2049 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
2050 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
2051 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
2054 if (unlikely(page
->mem_cgroup
))
2055 bad_reason
= "page still charged to cgroup";
2057 bad_page(page
, bad_reason
, bad_flags
);
2061 * This page is about to be returned from the page allocator
2063 static inline int check_new_page(struct page
*page
)
2065 if (likely(page_expected_state(page
,
2066 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2069 check_new_page_bad(page
);
2073 static inline bool free_pages_prezeroed(void)
2075 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
2076 page_poisoning_enabled()) || want_init_on_free();
2079 #ifdef CONFIG_DEBUG_VM
2081 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2082 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2083 * also checked when pcp lists are refilled from the free lists.
2085 static inline bool check_pcp_refill(struct page
*page
)
2087 if (debug_pagealloc_enabled_static())
2088 return check_new_page(page
);
2093 static inline bool check_new_pcp(struct page
*page
)
2095 return check_new_page(page
);
2099 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2100 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2101 * enabled, they are also checked when being allocated from the pcp lists.
2103 static inline bool check_pcp_refill(struct page
*page
)
2105 return check_new_page(page
);
2107 static inline bool check_new_pcp(struct page
*page
)
2109 if (debug_pagealloc_enabled_static())
2110 return check_new_page(page
);
2114 #endif /* CONFIG_DEBUG_VM */
2116 static bool check_new_pages(struct page
*page
, unsigned int order
)
2119 for (i
= 0; i
< (1 << order
); i
++) {
2120 struct page
*p
= page
+ i
;
2122 if (unlikely(check_new_page(p
)))
2129 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2132 set_page_private(page
, 0);
2133 set_page_refcounted(page
);
2135 arch_alloc_page(page
, order
);
2136 if (debug_pagealloc_enabled_static())
2137 kernel_map_pages(page
, 1 << order
, 1);
2138 kasan_alloc_pages(page
, order
);
2139 kernel_poison_pages(page
, 1 << order
, 1);
2140 set_page_owner(page
, order
, gfp_flags
);
2143 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2144 unsigned int alloc_flags
)
2146 post_alloc_hook(page
, order
, gfp_flags
);
2148 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags
))
2149 kernel_init_free_pages(page
, 1 << order
);
2151 if (order
&& (gfp_flags
& __GFP_COMP
))
2152 prep_compound_page(page
, order
);
2155 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2156 * allocate the page. The expectation is that the caller is taking
2157 * steps that will free more memory. The caller should avoid the page
2158 * being used for !PFMEMALLOC purposes.
2160 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2161 set_page_pfmemalloc(page
);
2163 clear_page_pfmemalloc(page
);
2167 * Go through the free lists for the given migratetype and remove
2168 * the smallest available page from the freelists
2170 static __always_inline
2171 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2174 unsigned int current_order
;
2175 struct free_area
*area
;
2178 /* Find a page of the appropriate size in the preferred list */
2179 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2180 area
= &(zone
->free_area
[current_order
]);
2181 page
= get_page_from_free_area(area
, migratetype
);
2184 del_page_from_free_area(page
, area
);
2185 expand(zone
, page
, order
, current_order
, area
, migratetype
);
2186 set_pcppage_migratetype(page
, migratetype
);
2195 * This array describes the order lists are fallen back to when
2196 * the free lists for the desirable migrate type are depleted
2198 static int fallbacks
[MIGRATE_TYPES
][4] = {
2199 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2200 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2201 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2203 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2205 #ifdef CONFIG_MEMORY_ISOLATION
2206 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2211 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2214 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2217 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2218 unsigned int order
) { return NULL
; }
2222 * Move the free pages in a range to the free lists of the requested type.
2223 * Note that start_page and end_pages are not aligned on a pageblock
2224 * boundary. If alignment is required, use move_freepages_block()
2226 static int move_freepages(struct zone
*zone
,
2227 struct page
*start_page
, struct page
*end_page
,
2228 int migratetype
, int *num_movable
)
2232 int pages_moved
= 0;
2234 for (page
= start_page
; page
<= end_page
;) {
2235 if (!pfn_valid_within(page_to_pfn(page
))) {
2240 if (!PageBuddy(page
)) {
2242 * We assume that pages that could be isolated for
2243 * migration are movable. But we don't actually try
2244 * isolating, as that would be expensive.
2247 (PageLRU(page
) || __PageMovable(page
)))
2254 /* Make sure we are not inadvertently changing nodes */
2255 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2256 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2258 order
= page_order(page
);
2259 move_to_free_area(page
, &zone
->free_area
[order
], migratetype
);
2261 pages_moved
+= 1 << order
;
2267 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2268 int migratetype
, int *num_movable
)
2270 unsigned long start_pfn
, end_pfn
;
2271 struct page
*start_page
, *end_page
;
2276 start_pfn
= page_to_pfn(page
);
2277 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2278 start_page
= pfn_to_page(start_pfn
);
2279 end_page
= start_page
+ pageblock_nr_pages
- 1;
2280 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2282 /* Do not cross zone boundaries */
2283 if (!zone_spans_pfn(zone
, start_pfn
))
2285 if (!zone_spans_pfn(zone
, end_pfn
))
2288 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2292 static void change_pageblock_range(struct page
*pageblock_page
,
2293 int start_order
, int migratetype
)
2295 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2297 while (nr_pageblocks
--) {
2298 set_pageblock_migratetype(pageblock_page
, migratetype
);
2299 pageblock_page
+= pageblock_nr_pages
;
2304 * When we are falling back to another migratetype during allocation, try to
2305 * steal extra free pages from the same pageblocks to satisfy further
2306 * allocations, instead of polluting multiple pageblocks.
2308 * If we are stealing a relatively large buddy page, it is likely there will
2309 * be more free pages in the pageblock, so try to steal them all. For
2310 * reclaimable and unmovable allocations, we steal regardless of page size,
2311 * as fragmentation caused by those allocations polluting movable pageblocks
2312 * is worse than movable allocations stealing from unmovable and reclaimable
2315 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2318 * Leaving this order check is intended, although there is
2319 * relaxed order check in next check. The reason is that
2320 * we can actually steal whole pageblock if this condition met,
2321 * but, below check doesn't guarantee it and that is just heuristic
2322 * so could be changed anytime.
2324 if (order
>= pageblock_order
)
2327 if (order
>= pageblock_order
/ 2 ||
2328 start_mt
== MIGRATE_RECLAIMABLE
||
2329 start_mt
== MIGRATE_UNMOVABLE
||
2330 page_group_by_mobility_disabled
)
2336 static inline bool boost_watermark(struct zone
*zone
)
2338 unsigned long max_boost
;
2340 if (!watermark_boost_factor
)
2343 * Don't bother in zones that are unlikely to produce results.
2344 * On small machines, including kdump capture kernels running
2345 * in a small area, boosting the watermark can cause an out of
2346 * memory situation immediately.
2348 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
2351 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2352 watermark_boost_factor
, 10000);
2355 * high watermark may be uninitialised if fragmentation occurs
2356 * very early in boot so do not boost. We do not fall
2357 * through and boost by pageblock_nr_pages as failing
2358 * allocations that early means that reclaim is not going
2359 * to help and it may even be impossible to reclaim the
2360 * boosted watermark resulting in a hang.
2365 max_boost
= max(pageblock_nr_pages
, max_boost
);
2367 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2374 * This function implements actual steal behaviour. If order is large enough,
2375 * we can steal whole pageblock. If not, we first move freepages in this
2376 * pageblock to our migratetype and determine how many already-allocated pages
2377 * are there in the pageblock with a compatible migratetype. If at least half
2378 * of pages are free or compatible, we can change migratetype of the pageblock
2379 * itself, so pages freed in the future will be put on the correct free list.
2381 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2382 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2384 unsigned int current_order
= page_order(page
);
2385 struct free_area
*area
;
2386 int free_pages
, movable_pages
, alike_pages
;
2389 old_block_type
= get_pageblock_migratetype(page
);
2392 * This can happen due to races and we want to prevent broken
2393 * highatomic accounting.
2395 if (is_migrate_highatomic(old_block_type
))
2398 /* Take ownership for orders >= pageblock_order */
2399 if (current_order
>= pageblock_order
) {
2400 change_pageblock_range(page
, current_order
, start_type
);
2405 * Boost watermarks to increase reclaim pressure to reduce the
2406 * likelihood of future fallbacks. Wake kswapd now as the node
2407 * may be balanced overall and kswapd will not wake naturally.
2409 if (boost_watermark(zone
) && (alloc_flags
& ALLOC_KSWAPD
))
2410 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2412 /* We are not allowed to try stealing from the whole block */
2416 free_pages
= move_freepages_block(zone
, page
, start_type
,
2419 * Determine how many pages are compatible with our allocation.
2420 * For movable allocation, it's the number of movable pages which
2421 * we just obtained. For other types it's a bit more tricky.
2423 if (start_type
== MIGRATE_MOVABLE
) {
2424 alike_pages
= movable_pages
;
2427 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2428 * to MOVABLE pageblock, consider all non-movable pages as
2429 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2430 * vice versa, be conservative since we can't distinguish the
2431 * exact migratetype of non-movable pages.
2433 if (old_block_type
== MIGRATE_MOVABLE
)
2434 alike_pages
= pageblock_nr_pages
2435 - (free_pages
+ movable_pages
);
2440 /* moving whole block can fail due to zone boundary conditions */
2445 * If a sufficient number of pages in the block are either free or of
2446 * comparable migratability as our allocation, claim the whole block.
2448 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2449 page_group_by_mobility_disabled
)
2450 set_pageblock_migratetype(page
, start_type
);
2455 area
= &zone
->free_area
[current_order
];
2456 move_to_free_area(page
, area
, start_type
);
2460 * Check whether there is a suitable fallback freepage with requested order.
2461 * If only_stealable is true, this function returns fallback_mt only if
2462 * we can steal other freepages all together. This would help to reduce
2463 * fragmentation due to mixed migratetype pages in one pageblock.
2465 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2466 int migratetype
, bool only_stealable
, bool *can_steal
)
2471 if (area
->nr_free
== 0)
2476 fallback_mt
= fallbacks
[migratetype
][i
];
2477 if (fallback_mt
== MIGRATE_TYPES
)
2480 if (free_area_empty(area
, fallback_mt
))
2483 if (can_steal_fallback(order
, migratetype
))
2486 if (!only_stealable
)
2497 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2498 * there are no empty page blocks that contain a page with a suitable order
2500 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2501 unsigned int alloc_order
)
2504 unsigned long max_managed
, flags
;
2507 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2508 * Check is race-prone but harmless.
2510 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2511 if (zone
->nr_reserved_highatomic
>= max_managed
)
2514 spin_lock_irqsave(&zone
->lock
, flags
);
2516 /* Recheck the nr_reserved_highatomic limit under the lock */
2517 if (zone
->nr_reserved_highatomic
>= max_managed
)
2521 mt
= get_pageblock_migratetype(page
);
2522 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2523 && !is_migrate_cma(mt
)) {
2524 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2525 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2526 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2530 spin_unlock_irqrestore(&zone
->lock
, flags
);
2534 * Used when an allocation is about to fail under memory pressure. This
2535 * potentially hurts the reliability of high-order allocations when under
2536 * intense memory pressure but failed atomic allocations should be easier
2537 * to recover from than an OOM.
2539 * If @force is true, try to unreserve a pageblock even though highatomic
2540 * pageblock is exhausted.
2542 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2545 struct zonelist
*zonelist
= ac
->zonelist
;
2546 unsigned long flags
;
2553 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2556 * Preserve at least one pageblock unless memory pressure
2559 if (!force
&& zone
->nr_reserved_highatomic
<=
2563 spin_lock_irqsave(&zone
->lock
, flags
);
2564 for (order
= 0; order
< MAX_ORDER
; order
++) {
2565 struct free_area
*area
= &(zone
->free_area
[order
]);
2567 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2572 * In page freeing path, migratetype change is racy so
2573 * we can counter several free pages in a pageblock
2574 * in this loop althoug we changed the pageblock type
2575 * from highatomic to ac->migratetype. So we should
2576 * adjust the count once.
2578 if (is_migrate_highatomic_page(page
)) {
2580 * It should never happen but changes to
2581 * locking could inadvertently allow a per-cpu
2582 * drain to add pages to MIGRATE_HIGHATOMIC
2583 * while unreserving so be safe and watch for
2586 zone
->nr_reserved_highatomic
-= min(
2588 zone
->nr_reserved_highatomic
);
2592 * Convert to ac->migratetype and avoid the normal
2593 * pageblock stealing heuristics. Minimally, the caller
2594 * is doing the work and needs the pages. More
2595 * importantly, if the block was always converted to
2596 * MIGRATE_UNMOVABLE or another type then the number
2597 * of pageblocks that cannot be completely freed
2600 set_pageblock_migratetype(page
, ac
->migratetype
);
2601 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2604 spin_unlock_irqrestore(&zone
->lock
, flags
);
2608 spin_unlock_irqrestore(&zone
->lock
, flags
);
2615 * Try finding a free buddy page on the fallback list and put it on the free
2616 * list of requested migratetype, possibly along with other pages from the same
2617 * block, depending on fragmentation avoidance heuristics. Returns true if
2618 * fallback was found so that __rmqueue_smallest() can grab it.
2620 * The use of signed ints for order and current_order is a deliberate
2621 * deviation from the rest of this file, to make the for loop
2622 * condition simpler.
2624 static __always_inline
bool
2625 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2626 unsigned int alloc_flags
)
2628 struct free_area
*area
;
2630 int min_order
= order
;
2636 * Do not steal pages from freelists belonging to other pageblocks
2637 * i.e. orders < pageblock_order. If there are no local zones free,
2638 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2640 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2641 min_order
= pageblock_order
;
2644 * Find the largest available free page in the other list. This roughly
2645 * approximates finding the pageblock with the most free pages, which
2646 * would be too costly to do exactly.
2648 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2650 area
= &(zone
->free_area
[current_order
]);
2651 fallback_mt
= find_suitable_fallback(area
, current_order
,
2652 start_migratetype
, false, &can_steal
);
2653 if (fallback_mt
== -1)
2657 * We cannot steal all free pages from the pageblock and the
2658 * requested migratetype is movable. In that case it's better to
2659 * steal and split the smallest available page instead of the
2660 * largest available page, because even if the next movable
2661 * allocation falls back into a different pageblock than this
2662 * one, it won't cause permanent fragmentation.
2664 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2665 && current_order
> order
)
2674 for (current_order
= order
; current_order
< MAX_ORDER
;
2676 area
= &(zone
->free_area
[current_order
]);
2677 fallback_mt
= find_suitable_fallback(area
, current_order
,
2678 start_migratetype
, false, &can_steal
);
2679 if (fallback_mt
!= -1)
2684 * This should not happen - we already found a suitable fallback
2685 * when looking for the largest page.
2687 VM_BUG_ON(current_order
== MAX_ORDER
);
2690 page
= get_page_from_free_area(area
, fallback_mt
);
2692 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2695 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2696 start_migratetype
, fallback_mt
);
2703 * Do the hard work of removing an element from the buddy allocator.
2704 * Call me with the zone->lock already held.
2706 static __always_inline
struct page
*
2707 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2708 unsigned int alloc_flags
)
2713 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2714 if (unlikely(!page
)) {
2715 if (migratetype
== MIGRATE_MOVABLE
)
2716 page
= __rmqueue_cma_fallback(zone
, order
);
2718 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2723 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2728 * Obtain a specified number of elements from the buddy allocator, all under
2729 * a single hold of the lock, for efficiency. Add them to the supplied list.
2730 * Returns the number of new pages which were placed at *list.
2732 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2733 unsigned long count
, struct list_head
*list
,
2734 int migratetype
, unsigned int alloc_flags
)
2738 spin_lock(&zone
->lock
);
2739 for (i
= 0; i
< count
; ++i
) {
2740 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2742 if (unlikely(page
== NULL
))
2745 if (unlikely(check_pcp_refill(page
)))
2749 * Split buddy pages returned by expand() are received here in
2750 * physical page order. The page is added to the tail of
2751 * caller's list. From the callers perspective, the linked list
2752 * is ordered by page number under some conditions. This is
2753 * useful for IO devices that can forward direction from the
2754 * head, thus also in the physical page order. This is useful
2755 * for IO devices that can merge IO requests if the physical
2756 * pages are ordered properly.
2758 list_add_tail(&page
->lru
, list
);
2760 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2761 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2766 * i pages were removed from the buddy list even if some leak due
2767 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2768 * on i. Do not confuse with 'alloced' which is the number of
2769 * pages added to the pcp list.
2771 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2772 spin_unlock(&zone
->lock
);
2778 * Called from the vmstat counter updater to drain pagesets of this
2779 * currently executing processor on remote nodes after they have
2782 * Note that this function must be called with the thread pinned to
2783 * a single processor.
2785 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2787 unsigned long flags
;
2788 int to_drain
, batch
;
2790 local_irq_save(flags
);
2791 batch
= READ_ONCE(pcp
->batch
);
2792 to_drain
= min(pcp
->count
, batch
);
2794 free_pcppages_bulk(zone
, to_drain
, pcp
);
2795 local_irq_restore(flags
);
2800 * Drain pcplists of the indicated processor and zone.
2802 * The processor must either be the current processor and the
2803 * thread pinned to the current processor or a processor that
2806 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2808 unsigned long flags
;
2809 struct per_cpu_pageset
*pset
;
2810 struct per_cpu_pages
*pcp
;
2812 local_irq_save(flags
);
2813 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2817 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2818 local_irq_restore(flags
);
2822 * Drain pcplists of all zones on the indicated processor.
2824 * The processor must either be the current processor and the
2825 * thread pinned to the current processor or a processor that
2828 static void drain_pages(unsigned int cpu
)
2832 for_each_populated_zone(zone
) {
2833 drain_pages_zone(cpu
, zone
);
2838 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2840 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2841 * the single zone's pages.
2843 void drain_local_pages(struct zone
*zone
)
2845 int cpu
= smp_processor_id();
2848 drain_pages_zone(cpu
, zone
);
2853 static void drain_local_pages_wq(struct work_struct
*work
)
2855 struct pcpu_drain
*drain
;
2857 drain
= container_of(work
, struct pcpu_drain
, work
);
2860 * drain_all_pages doesn't use proper cpu hotplug protection so
2861 * we can race with cpu offline when the WQ can move this from
2862 * a cpu pinned worker to an unbound one. We can operate on a different
2863 * cpu which is allright but we also have to make sure to not move to
2867 drain_local_pages(drain
->zone
);
2872 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2874 * When zone parameter is non-NULL, spill just the single zone's pages.
2876 * Note that this can be extremely slow as the draining happens in a workqueue.
2878 void drain_all_pages(struct zone
*zone
)
2883 * Allocate in the BSS so we wont require allocation in
2884 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2886 static cpumask_t cpus_with_pcps
;
2889 * Make sure nobody triggers this path before mm_percpu_wq is fully
2892 if (WARN_ON_ONCE(!mm_percpu_wq
))
2896 * Do not drain if one is already in progress unless it's specific to
2897 * a zone. Such callers are primarily CMA and memory hotplug and need
2898 * the drain to be complete when the call returns.
2900 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2903 mutex_lock(&pcpu_drain_mutex
);
2907 * We don't care about racing with CPU hotplug event
2908 * as offline notification will cause the notified
2909 * cpu to drain that CPU pcps and on_each_cpu_mask
2910 * disables preemption as part of its processing
2912 for_each_online_cpu(cpu
) {
2913 struct per_cpu_pageset
*pcp
;
2915 bool has_pcps
= false;
2918 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2922 for_each_populated_zone(z
) {
2923 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2924 if (pcp
->pcp
.count
) {
2932 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2934 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2937 for_each_cpu(cpu
, &cpus_with_pcps
) {
2938 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
2941 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
2942 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
2944 for_each_cpu(cpu
, &cpus_with_pcps
)
2945 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
2947 mutex_unlock(&pcpu_drain_mutex
);
2950 #ifdef CONFIG_HIBERNATION
2953 * Touch the watchdog for every WD_PAGE_COUNT pages.
2955 #define WD_PAGE_COUNT (128*1024)
2957 void mark_free_pages(struct zone
*zone
)
2959 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2960 unsigned long flags
;
2961 unsigned int order
, t
;
2964 if (zone_is_empty(zone
))
2967 spin_lock_irqsave(&zone
->lock
, flags
);
2969 max_zone_pfn
= zone_end_pfn(zone
);
2970 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2971 if (pfn_valid(pfn
)) {
2972 page
= pfn_to_page(pfn
);
2974 if (!--page_count
) {
2975 touch_nmi_watchdog();
2976 page_count
= WD_PAGE_COUNT
;
2979 if (page_zone(page
) != zone
)
2982 if (!swsusp_page_is_forbidden(page
))
2983 swsusp_unset_page_free(page
);
2986 for_each_migratetype_order(order
, t
) {
2987 list_for_each_entry(page
,
2988 &zone
->free_area
[order
].free_list
[t
], lru
) {
2991 pfn
= page_to_pfn(page
);
2992 for (i
= 0; i
< (1UL << order
); i
++) {
2993 if (!--page_count
) {
2994 touch_nmi_watchdog();
2995 page_count
= WD_PAGE_COUNT
;
2997 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
3001 spin_unlock_irqrestore(&zone
->lock
, flags
);
3003 #endif /* CONFIG_PM */
3005 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
3009 if (!free_pcp_prepare(page
))
3012 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3013 set_pcppage_migratetype(page
, migratetype
);
3017 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
3019 struct zone
*zone
= page_zone(page
);
3020 struct per_cpu_pages
*pcp
;
3023 migratetype
= get_pcppage_migratetype(page
);
3024 __count_vm_event(PGFREE
);
3027 * We only track unmovable, reclaimable and movable on pcp lists.
3028 * Free ISOLATE pages back to the allocator because they are being
3029 * offlined but treat HIGHATOMIC as movable pages so we can get those
3030 * areas back if necessary. Otherwise, we may have to free
3031 * excessively into the page allocator
3033 if (migratetype
>= MIGRATE_PCPTYPES
) {
3034 if (unlikely(is_migrate_isolate(migratetype
))) {
3035 free_one_page(zone
, page
, pfn
, 0, migratetype
);
3038 migratetype
= MIGRATE_MOVABLE
;
3041 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3042 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
3044 if (pcp
->count
>= pcp
->high
) {
3045 unsigned long batch
= READ_ONCE(pcp
->batch
);
3046 free_pcppages_bulk(zone
, batch
, pcp
);
3051 * Free a 0-order page
3053 void free_unref_page(struct page
*page
)
3055 unsigned long flags
;
3056 unsigned long pfn
= page_to_pfn(page
);
3058 if (!free_unref_page_prepare(page
, pfn
))
3061 local_irq_save(flags
);
3062 free_unref_page_commit(page
, pfn
);
3063 local_irq_restore(flags
);
3067 * Free a list of 0-order pages
3069 void free_unref_page_list(struct list_head
*list
)
3071 struct page
*page
, *next
;
3072 unsigned long flags
, pfn
;
3073 int batch_count
= 0;
3075 /* Prepare pages for freeing */
3076 list_for_each_entry_safe(page
, next
, list
, lru
) {
3077 pfn
= page_to_pfn(page
);
3078 if (!free_unref_page_prepare(page
, pfn
))
3079 list_del(&page
->lru
);
3080 set_page_private(page
, pfn
);
3083 local_irq_save(flags
);
3084 list_for_each_entry_safe(page
, next
, list
, lru
) {
3085 unsigned long pfn
= page_private(page
);
3087 set_page_private(page
, 0);
3088 trace_mm_page_free_batched(page
);
3089 free_unref_page_commit(page
, pfn
);
3092 * Guard against excessive IRQ disabled times when we get
3093 * a large list of pages to free.
3095 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3096 local_irq_restore(flags
);
3098 local_irq_save(flags
);
3101 local_irq_restore(flags
);
3105 * split_page takes a non-compound higher-order page, and splits it into
3106 * n (1<<order) sub-pages: page[0..n]
3107 * Each sub-page must be freed individually.
3109 * Note: this is probably too low level an operation for use in drivers.
3110 * Please consult with lkml before using this in your driver.
3112 void split_page(struct page
*page
, unsigned int order
)
3116 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3117 VM_BUG_ON_PAGE(!page_count(page
), page
);
3119 for (i
= 1; i
< (1 << order
); i
++)
3120 set_page_refcounted(page
+ i
);
3121 split_page_owner(page
, 1 << order
);
3123 EXPORT_SYMBOL_GPL(split_page
);
3125 int __isolate_free_page(struct page
*page
, unsigned int order
)
3127 struct free_area
*area
= &page_zone(page
)->free_area
[order
];
3128 unsigned long watermark
;
3132 BUG_ON(!PageBuddy(page
));
3134 zone
= page_zone(page
);
3135 mt
= get_pageblock_migratetype(page
);
3137 if (!is_migrate_isolate(mt
)) {
3139 * Obey watermarks as if the page was being allocated. We can
3140 * emulate a high-order watermark check with a raised order-0
3141 * watermark, because we already know our high-order page
3144 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3145 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3148 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3151 /* Remove page from free list */
3153 del_page_from_free_area(page
, area
);
3156 * Set the pageblock if the isolated page is at least half of a
3159 if (order
>= pageblock_order
- 1) {
3160 struct page
*endpage
= page
+ (1 << order
) - 1;
3161 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3162 int mt
= get_pageblock_migratetype(page
);
3163 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3164 && !is_migrate_highatomic(mt
))
3165 set_pageblock_migratetype(page
,
3171 return 1UL << order
;
3175 * Update NUMA hit/miss statistics
3177 * Must be called with interrupts disabled.
3179 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
3182 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3184 /* skip numa counters update if numa stats is disabled */
3185 if (!static_branch_likely(&vm_numa_stat_key
))
3188 if (zone_to_nid(z
) != numa_node_id())
3189 local_stat
= NUMA_OTHER
;
3191 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3192 __inc_numa_state(z
, NUMA_HIT
);
3194 __inc_numa_state(z
, NUMA_MISS
);
3195 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3197 __inc_numa_state(z
, local_stat
);
3201 /* Remove page from the per-cpu list, caller must protect the list */
3202 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3203 unsigned int alloc_flags
,
3204 struct per_cpu_pages
*pcp
,
3205 struct list_head
*list
)
3210 if (list_empty(list
)) {
3211 pcp
->count
+= rmqueue_bulk(zone
, 0,
3213 migratetype
, alloc_flags
);
3214 if (unlikely(list_empty(list
)))
3218 page
= list_first_entry(list
, struct page
, lru
);
3219 list_del(&page
->lru
);
3221 } while (check_new_pcp(page
));
3226 /* Lock and remove page from the per-cpu list */
3227 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3228 struct zone
*zone
, gfp_t gfp_flags
,
3229 int migratetype
, unsigned int alloc_flags
)
3231 struct per_cpu_pages
*pcp
;
3232 struct list_head
*list
;
3234 unsigned long flags
;
3236 local_irq_save(flags
);
3237 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3238 list
= &pcp
->lists
[migratetype
];
3239 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3241 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3242 zone_statistics(preferred_zone
, zone
);
3244 local_irq_restore(flags
);
3249 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3252 struct page
*rmqueue(struct zone
*preferred_zone
,
3253 struct zone
*zone
, unsigned int order
,
3254 gfp_t gfp_flags
, unsigned int alloc_flags
,
3257 unsigned long flags
;
3260 if (likely(order
== 0)) {
3261 page
= rmqueue_pcplist(preferred_zone
, zone
, gfp_flags
,
3262 migratetype
, alloc_flags
);
3267 * We most definitely don't want callers attempting to
3268 * allocate greater than order-1 page units with __GFP_NOFAIL.
3270 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3271 spin_lock_irqsave(&zone
->lock
, flags
);
3275 if (alloc_flags
& ALLOC_HARDER
) {
3276 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3278 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3281 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3282 } while (page
&& check_new_pages(page
, order
));
3283 spin_unlock(&zone
->lock
);
3286 __mod_zone_freepage_state(zone
, -(1 << order
),
3287 get_pcppage_migratetype(page
));
3289 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3290 zone_statistics(preferred_zone
, zone
);
3291 local_irq_restore(flags
);
3294 /* Separate test+clear to avoid unnecessary atomics */
3295 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3296 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3297 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3300 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3304 local_irq_restore(flags
);
3308 #ifdef CONFIG_FAIL_PAGE_ALLOC
3311 struct fault_attr attr
;
3313 bool ignore_gfp_highmem
;
3314 bool ignore_gfp_reclaim
;
3316 } fail_page_alloc
= {
3317 .attr
= FAULT_ATTR_INITIALIZER
,
3318 .ignore_gfp_reclaim
= true,
3319 .ignore_gfp_highmem
= true,
3323 static int __init
setup_fail_page_alloc(char *str
)
3325 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3327 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3329 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3331 if (order
< fail_page_alloc
.min_order
)
3333 if (gfp_mask
& __GFP_NOFAIL
)
3335 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3337 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3338 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3341 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3344 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3346 static int __init
fail_page_alloc_debugfs(void)
3348 umode_t mode
= S_IFREG
| 0600;
3351 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3352 &fail_page_alloc
.attr
);
3354 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3355 &fail_page_alloc
.ignore_gfp_reclaim
);
3356 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3357 &fail_page_alloc
.ignore_gfp_highmem
);
3358 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3363 late_initcall(fail_page_alloc_debugfs
);
3365 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3367 #else /* CONFIG_FAIL_PAGE_ALLOC */
3369 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3374 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3376 noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3378 return __should_fail_alloc_page(gfp_mask
, order
);
3380 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3383 * Return true if free base pages are above 'mark'. For high-order checks it
3384 * will return true of the order-0 watermark is reached and there is at least
3385 * one free page of a suitable size. Checking now avoids taking the zone lock
3386 * to check in the allocation paths if no pages are free.
3388 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3389 int classzone_idx
, unsigned int alloc_flags
,
3394 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3396 /* free_pages may go negative - that's OK */
3397 free_pages
-= (1 << order
) - 1;
3399 if (alloc_flags
& ALLOC_HIGH
)
3403 * If the caller does not have rights to ALLOC_HARDER then subtract
3404 * the high-atomic reserves. This will over-estimate the size of the
3405 * atomic reserve but it avoids a search.
3407 if (likely(!alloc_harder
)) {
3408 free_pages
-= z
->nr_reserved_highatomic
;
3411 * OOM victims can try even harder than normal ALLOC_HARDER
3412 * users on the grounds that it's definitely going to be in
3413 * the exit path shortly and free memory. Any allocation it
3414 * makes during the free path will be small and short-lived.
3416 if (alloc_flags
& ALLOC_OOM
)
3424 /* If allocation can't use CMA areas don't use free CMA pages */
3425 if (!(alloc_flags
& ALLOC_CMA
))
3426 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3430 * Check watermarks for an order-0 allocation request. If these
3431 * are not met, then a high-order request also cannot go ahead
3432 * even if a suitable page happened to be free.
3434 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3437 /* If this is an order-0 request then the watermark is fine */
3441 /* For a high-order request, check at least one suitable page is free */
3442 for (o
= order
; o
< MAX_ORDER
; o
++) {
3443 struct free_area
*area
= &z
->free_area
[o
];
3449 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3450 if (!free_area_empty(area
, mt
))
3455 if ((alloc_flags
& ALLOC_CMA
) &&
3456 !free_area_empty(area
, MIGRATE_CMA
)) {
3461 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3467 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3468 int classzone_idx
, unsigned int alloc_flags
)
3470 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3471 zone_page_state(z
, NR_FREE_PAGES
));
3474 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3475 unsigned long mark
, int classzone_idx
,
3476 unsigned int alloc_flags
, gfp_t gfp_mask
)
3478 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3482 /* If allocation can't use CMA areas don't use free CMA pages */
3483 if (!(alloc_flags
& ALLOC_CMA
))
3484 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3488 * Fast check for order-0 only. If this fails then the reserves
3489 * need to be calculated. There is a corner case where the check
3490 * passes but only the high-order atomic reserve are free. If
3491 * the caller is !atomic then it'll uselessly search the free
3492 * list. That corner case is then slower but it is harmless.
3494 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3497 if (__zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3501 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3502 * when checking the min watermark. The min watermark is the
3503 * point where boosting is ignored so that kswapd is woken up
3504 * when below the low watermark.
3506 if (unlikely(!order
&& (gfp_mask
& __GFP_ATOMIC
) && z
->watermark_boost
3507 && ((alloc_flags
& ALLOC_WMARK_MASK
) == WMARK_MIN
))) {
3508 mark
= z
->_watermark
[WMARK_MIN
];
3509 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
,
3510 alloc_flags
, free_pages
);
3516 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3517 unsigned long mark
, int classzone_idx
)
3519 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3521 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3522 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3524 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3529 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3531 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3532 node_reclaim_distance
;
3534 #else /* CONFIG_NUMA */
3535 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3539 #endif /* CONFIG_NUMA */
3542 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3543 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3544 * premature use of a lower zone may cause lowmem pressure problems that
3545 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3546 * probably too small. It only makes sense to spread allocations to avoid
3547 * fragmentation between the Normal and DMA32 zones.
3549 static inline unsigned int
3550 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3552 unsigned int alloc_flags
= 0;
3554 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3555 alloc_flags
|= ALLOC_KSWAPD
;
3557 #ifdef CONFIG_ZONE_DMA32
3561 if (zone_idx(zone
) != ZONE_NORMAL
)
3565 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3566 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3567 * on UMA that if Normal is populated then so is DMA32.
3569 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3570 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3573 alloc_flags
|= ALLOC_NOFRAGMENT
;
3574 #endif /* CONFIG_ZONE_DMA32 */
3579 * get_page_from_freelist goes through the zonelist trying to allocate
3582 static struct page
*
3583 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3584 const struct alloc_context
*ac
)
3588 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3593 * Scan zonelist, looking for a zone with enough free.
3594 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3596 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3597 z
= ac
->preferred_zoneref
;
3598 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3603 if (cpusets_enabled() &&
3604 (alloc_flags
& ALLOC_CPUSET
) &&
3605 !__cpuset_zone_allowed(zone
, gfp_mask
))
3608 * When allocating a page cache page for writing, we
3609 * want to get it from a node that is within its dirty
3610 * limit, such that no single node holds more than its
3611 * proportional share of globally allowed dirty pages.
3612 * The dirty limits take into account the node's
3613 * lowmem reserves and high watermark so that kswapd
3614 * should be able to balance it without having to
3615 * write pages from its LRU list.
3617 * XXX: For now, allow allocations to potentially
3618 * exceed the per-node dirty limit in the slowpath
3619 * (spread_dirty_pages unset) before going into reclaim,
3620 * which is important when on a NUMA setup the allowed
3621 * nodes are together not big enough to reach the
3622 * global limit. The proper fix for these situations
3623 * will require awareness of nodes in the
3624 * dirty-throttling and the flusher threads.
3626 if (ac
->spread_dirty_pages
) {
3627 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3630 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3631 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3636 if (no_fallback
&& nr_online_nodes
> 1 &&
3637 zone
!= ac
->preferred_zoneref
->zone
) {
3641 * If moving to a remote node, retry but allow
3642 * fragmenting fallbacks. Locality is more important
3643 * than fragmentation avoidance.
3645 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3646 if (zone_to_nid(zone
) != local_nid
) {
3647 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3652 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3653 if (!zone_watermark_fast(zone
, order
, mark
,
3654 ac_classzone_idx(ac
), alloc_flags
,
3658 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3660 * Watermark failed for this zone, but see if we can
3661 * grow this zone if it contains deferred pages.
3663 if (static_branch_unlikely(&deferred_pages
)) {
3664 if (_deferred_grow_zone(zone
, order
))
3668 /* Checked here to keep the fast path fast */
3669 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3670 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3673 if (node_reclaim_mode
== 0 ||
3674 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3677 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3679 case NODE_RECLAIM_NOSCAN
:
3682 case NODE_RECLAIM_FULL
:
3683 /* scanned but unreclaimable */
3686 /* did we reclaim enough */
3687 if (zone_watermark_ok(zone
, order
, mark
,
3688 ac_classzone_idx(ac
), alloc_flags
))
3696 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3697 gfp_mask
, alloc_flags
, ac
->migratetype
);
3699 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3702 * If this is a high-order atomic allocation then check
3703 * if the pageblock should be reserved for the future
3705 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3706 reserve_highatomic_pageblock(page
, zone
, order
);
3710 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3711 /* Try again if zone has deferred pages */
3712 if (static_branch_unlikely(&deferred_pages
)) {
3713 if (_deferred_grow_zone(zone
, order
))
3721 * It's possible on a UMA machine to get through all zones that are
3722 * fragmented. If avoiding fragmentation, reset and try again.
3725 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3732 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3734 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3737 * This documents exceptions given to allocations in certain
3738 * contexts that are allowed to allocate outside current's set
3741 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3742 if (tsk_is_oom_victim(current
) ||
3743 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3744 filter
&= ~SHOW_MEM_FILTER_NODES
;
3745 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3746 filter
&= ~SHOW_MEM_FILTER_NODES
;
3748 show_mem(filter
, nodemask
);
3751 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3753 struct va_format vaf
;
3755 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
3757 if ((gfp_mask
& __GFP_NOWARN
) ||
3758 !__ratelimit(&nopage_rs
) ||
3759 ((gfp_mask
& __GFP_DMA
) && !has_managed_dma()))
3762 va_start(args
, fmt
);
3765 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3766 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3767 nodemask_pr_args(nodemask
));
3770 cpuset_print_current_mems_allowed();
3773 warn_alloc_show_mem(gfp_mask
, nodemask
);
3776 static inline struct page
*
3777 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3778 unsigned int alloc_flags
,
3779 const struct alloc_context
*ac
)
3783 page
= get_page_from_freelist(gfp_mask
, order
,
3784 alloc_flags
|ALLOC_CPUSET
, ac
);
3786 * fallback to ignore cpuset restriction if our nodes
3790 page
= get_page_from_freelist(gfp_mask
, order
,
3796 static inline struct page
*
3797 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3798 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3800 struct oom_control oc
= {
3801 .zonelist
= ac
->zonelist
,
3802 .nodemask
= ac
->nodemask
,
3804 .gfp_mask
= gfp_mask
,
3809 *did_some_progress
= 0;
3812 * Acquire the oom lock. If that fails, somebody else is
3813 * making progress for us.
3815 if (!mutex_trylock(&oom_lock
)) {
3816 *did_some_progress
= 1;
3817 schedule_timeout_uninterruptible(1);
3822 * Go through the zonelist yet one more time, keep very high watermark
3823 * here, this is only to catch a parallel oom killing, we must fail if
3824 * we're still under heavy pressure. But make sure that this reclaim
3825 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3826 * allocation which will never fail due to oom_lock already held.
3828 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3829 ~__GFP_DIRECT_RECLAIM
, order
,
3830 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3834 /* Coredumps can quickly deplete all memory reserves */
3835 if (current
->flags
& PF_DUMPCORE
)
3837 /* The OOM killer will not help higher order allocs */
3838 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3841 * We have already exhausted all our reclaim opportunities without any
3842 * success so it is time to admit defeat. We will skip the OOM killer
3843 * because it is very likely that the caller has a more reasonable
3844 * fallback than shooting a random task.
3846 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3848 /* The OOM killer does not needlessly kill tasks for lowmem */
3849 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3851 if (pm_suspended_storage())
3854 * XXX: GFP_NOFS allocations should rather fail than rely on
3855 * other request to make a forward progress.
3856 * We are in an unfortunate situation where out_of_memory cannot
3857 * do much for this context but let's try it to at least get
3858 * access to memory reserved if the current task is killed (see
3859 * out_of_memory). Once filesystems are ready to handle allocation
3860 * failures more gracefully we should just bail out here.
3863 /* The OOM killer may not free memory on a specific node */
3864 if (gfp_mask
& __GFP_THISNODE
)
3867 /* Exhausted what can be done so it's blame time */
3868 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3869 *did_some_progress
= 1;
3872 * Help non-failing allocations by giving them access to memory
3875 if (gfp_mask
& __GFP_NOFAIL
)
3876 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3877 ALLOC_NO_WATERMARKS
, ac
);
3880 mutex_unlock(&oom_lock
);
3885 * Maximum number of compaction retries wit a progress before OOM
3886 * killer is consider as the only way to move forward.
3888 #define MAX_COMPACT_RETRIES 16
3890 #ifdef CONFIG_COMPACTION
3891 /* Try memory compaction for high-order allocations before reclaim */
3892 static struct page
*
3893 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3894 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3895 enum compact_priority prio
, enum compact_result
*compact_result
)
3897 struct page
*page
= NULL
;
3898 unsigned long pflags
;
3899 unsigned int noreclaim_flag
;
3904 psi_memstall_enter(&pflags
);
3905 noreclaim_flag
= memalloc_noreclaim_save();
3907 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3910 memalloc_noreclaim_restore(noreclaim_flag
);
3911 psi_memstall_leave(&pflags
);
3914 * At least in one zone compaction wasn't deferred or skipped, so let's
3915 * count a compaction stall
3917 count_vm_event(COMPACTSTALL
);
3919 /* Prep a captured page if available */
3921 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3923 /* Try get a page from the freelist if available */
3925 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3928 struct zone
*zone
= page_zone(page
);
3930 zone
->compact_blockskip_flush
= false;
3931 compaction_defer_reset(zone
, order
, true);
3932 count_vm_event(COMPACTSUCCESS
);
3937 * It's bad if compaction run occurs and fails. The most likely reason
3938 * is that pages exist, but not enough to satisfy watermarks.
3940 count_vm_event(COMPACTFAIL
);
3948 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3949 enum compact_result compact_result
,
3950 enum compact_priority
*compact_priority
,
3951 int *compaction_retries
)
3953 int max_retries
= MAX_COMPACT_RETRIES
;
3956 int retries
= *compaction_retries
;
3957 enum compact_priority priority
= *compact_priority
;
3962 if (compaction_made_progress(compact_result
))
3963 (*compaction_retries
)++;
3966 * compaction considers all the zone as desperately out of memory
3967 * so it doesn't really make much sense to retry except when the
3968 * failure could be caused by insufficient priority
3970 if (compaction_failed(compact_result
))
3971 goto check_priority
;
3974 * compaction was skipped because there are not enough order-0 pages
3975 * to work with, so we retry only if it looks like reclaim can help.
3977 if (compaction_needs_reclaim(compact_result
)) {
3978 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3983 * make sure the compaction wasn't deferred or didn't bail out early
3984 * due to locks contention before we declare that we should give up.
3985 * But the next retry should use a higher priority if allowed, so
3986 * we don't just keep bailing out endlessly.
3988 if (compaction_withdrawn(compact_result
)) {
3989 goto check_priority
;
3993 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3994 * costly ones because they are de facto nofail and invoke OOM
3995 * killer to move on while costly can fail and users are ready
3996 * to cope with that. 1/4 retries is rather arbitrary but we
3997 * would need much more detailed feedback from compaction to
3998 * make a better decision.
4000 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4002 if (*compaction_retries
<= max_retries
) {
4008 * Make sure there are attempts at the highest priority if we exhausted
4009 * all retries or failed at the lower priorities.
4012 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
4013 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
4015 if (*compact_priority
> min_priority
) {
4016 (*compact_priority
)--;
4017 *compaction_retries
= 0;
4021 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4025 static inline struct page
*
4026 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4027 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4028 enum compact_priority prio
, enum compact_result
*compact_result
)
4030 *compact_result
= COMPACT_SKIPPED
;
4035 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4036 enum compact_result compact_result
,
4037 enum compact_priority
*compact_priority
,
4038 int *compaction_retries
)
4043 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4047 * There are setups with compaction disabled which would prefer to loop
4048 * inside the allocator rather than hit the oom killer prematurely.
4049 * Let's give them a good hope and keep retrying while the order-0
4050 * watermarks are OK.
4052 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4054 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4055 ac_classzone_idx(ac
), alloc_flags
))
4060 #endif /* CONFIG_COMPACTION */
4062 #ifdef CONFIG_LOCKDEP
4063 static struct lockdep_map __fs_reclaim_map
=
4064 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4066 static bool __need_fs_reclaim(gfp_t gfp_mask
)
4068 gfp_mask
= current_gfp_context(gfp_mask
);
4070 /* no reclaim without waiting on it */
4071 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4074 /* this guy won't enter reclaim */
4075 if (current
->flags
& PF_MEMALLOC
)
4078 /* We're only interested __GFP_FS allocations for now */
4079 if (!(gfp_mask
& __GFP_FS
))
4082 if (gfp_mask
& __GFP_NOLOCKDEP
)
4088 void __fs_reclaim_acquire(void)
4090 lock_map_acquire(&__fs_reclaim_map
);
4093 void __fs_reclaim_release(void)
4095 lock_map_release(&__fs_reclaim_map
);
4098 void fs_reclaim_acquire(gfp_t gfp_mask
)
4100 if (__need_fs_reclaim(gfp_mask
))
4101 __fs_reclaim_acquire();
4103 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4105 void fs_reclaim_release(gfp_t gfp_mask
)
4107 if (__need_fs_reclaim(gfp_mask
))
4108 __fs_reclaim_release();
4110 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4113 /* Perform direct synchronous page reclaim */
4115 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4116 const struct alloc_context
*ac
)
4119 unsigned int noreclaim_flag
;
4120 unsigned long pflags
;
4124 /* We now go into synchronous reclaim */
4125 cpuset_memory_pressure_bump();
4126 psi_memstall_enter(&pflags
);
4127 fs_reclaim_acquire(gfp_mask
);
4128 noreclaim_flag
= memalloc_noreclaim_save();
4130 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4133 memalloc_noreclaim_restore(noreclaim_flag
);
4134 fs_reclaim_release(gfp_mask
);
4135 psi_memstall_leave(&pflags
);
4142 /* The really slow allocator path where we enter direct reclaim */
4143 static inline struct page
*
4144 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4145 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4146 unsigned long *did_some_progress
)
4148 struct page
*page
= NULL
;
4149 bool drained
= false;
4151 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4152 if (unlikely(!(*did_some_progress
)))
4156 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4159 * If an allocation failed after direct reclaim, it could be because
4160 * pages are pinned on the per-cpu lists or in high alloc reserves.
4161 * Shrink them them and try again
4163 if (!page
&& !drained
) {
4164 unreserve_highatomic_pageblock(ac
, false);
4165 drain_all_pages(NULL
);
4173 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4174 const struct alloc_context
*ac
)
4178 pg_data_t
*last_pgdat
= NULL
;
4179 enum zone_type high_zoneidx
= ac
->high_zoneidx
;
4181 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, high_zoneidx
,
4183 if (last_pgdat
!= zone
->zone_pgdat
)
4184 wakeup_kswapd(zone
, gfp_mask
, order
, high_zoneidx
);
4185 last_pgdat
= zone
->zone_pgdat
;
4189 static inline unsigned int
4190 gfp_to_alloc_flags(gfp_t gfp_mask
)
4192 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4194 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4195 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4198 * The caller may dip into page reserves a bit more if the caller
4199 * cannot run direct reclaim, or if the caller has realtime scheduling
4200 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4201 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4203 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
4205 if (gfp_mask
& __GFP_ATOMIC
) {
4207 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4208 * if it can't schedule.
4210 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4211 alloc_flags
|= ALLOC_HARDER
;
4213 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4214 * comment for __cpuset_node_allowed().
4216 alloc_flags
&= ~ALLOC_CPUSET
;
4217 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4218 alloc_flags
|= ALLOC_HARDER
;
4220 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4221 alloc_flags
|= ALLOC_KSWAPD
;
4224 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
4225 alloc_flags
|= ALLOC_CMA
;
4230 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4232 if (!tsk_is_oom_victim(tsk
))
4236 * !MMU doesn't have oom reaper so give access to memory reserves
4237 * only to the thread with TIF_MEMDIE set
4239 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4246 * Distinguish requests which really need access to full memory
4247 * reserves from oom victims which can live with a portion of it
4249 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4251 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4253 if (gfp_mask
& __GFP_MEMALLOC
)
4254 return ALLOC_NO_WATERMARKS
;
4255 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4256 return ALLOC_NO_WATERMARKS
;
4257 if (!in_interrupt()) {
4258 if (current
->flags
& PF_MEMALLOC
)
4259 return ALLOC_NO_WATERMARKS
;
4260 else if (oom_reserves_allowed(current
))
4267 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4269 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4273 * Checks whether it makes sense to retry the reclaim to make a forward progress
4274 * for the given allocation request.
4276 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4277 * without success, or when we couldn't even meet the watermark if we
4278 * reclaimed all remaining pages on the LRU lists.
4280 * Returns true if a retry is viable or false to enter the oom path.
4283 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4284 struct alloc_context
*ac
, int alloc_flags
,
4285 bool did_some_progress
, int *no_progress_loops
)
4292 * Costly allocations might have made a progress but this doesn't mean
4293 * their order will become available due to high fragmentation so
4294 * always increment the no progress counter for them
4296 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4297 *no_progress_loops
= 0;
4299 (*no_progress_loops
)++;
4302 * Make sure we converge to OOM if we cannot make any progress
4303 * several times in the row.
4305 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4306 /* Before OOM, exhaust highatomic_reserve */
4307 return unreserve_highatomic_pageblock(ac
, true);
4311 * Keep reclaiming pages while there is a chance this will lead
4312 * somewhere. If none of the target zones can satisfy our allocation
4313 * request even if all reclaimable pages are considered then we are
4314 * screwed and have to go OOM.
4316 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
4318 unsigned long available
;
4319 unsigned long reclaimable
;
4320 unsigned long min_wmark
= min_wmark_pages(zone
);
4323 available
= reclaimable
= zone_reclaimable_pages(zone
);
4324 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4327 * Would the allocation succeed if we reclaimed all
4328 * reclaimable pages?
4330 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4331 ac_classzone_idx(ac
), alloc_flags
, available
);
4332 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4333 available
, min_wmark
, *no_progress_loops
, wmark
);
4336 * If we didn't make any progress and have a lot of
4337 * dirty + writeback pages then we should wait for
4338 * an IO to complete to slow down the reclaim and
4339 * prevent from pre mature OOM
4341 if (!did_some_progress
) {
4342 unsigned long write_pending
;
4344 write_pending
= zone_page_state_snapshot(zone
,
4345 NR_ZONE_WRITE_PENDING
);
4347 if (2 * write_pending
> reclaimable
) {
4348 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4360 * Memory allocation/reclaim might be called from a WQ context and the
4361 * current implementation of the WQ concurrency control doesn't
4362 * recognize that a particular WQ is congested if the worker thread is
4363 * looping without ever sleeping. Therefore we have to do a short sleep
4364 * here rather than calling cond_resched().
4366 if (current
->flags
& PF_WQ_WORKER
)
4367 schedule_timeout_uninterruptible(1);
4374 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4377 * It's possible that cpuset's mems_allowed and the nodemask from
4378 * mempolicy don't intersect. This should be normally dealt with by
4379 * policy_nodemask(), but it's possible to race with cpuset update in
4380 * such a way the check therein was true, and then it became false
4381 * before we got our cpuset_mems_cookie here.
4382 * This assumes that for all allocations, ac->nodemask can come only
4383 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4384 * when it does not intersect with the cpuset restrictions) or the
4385 * caller can deal with a violated nodemask.
4387 if (cpusets_enabled() && ac
->nodemask
&&
4388 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4389 ac
->nodemask
= NULL
;
4394 * When updating a task's mems_allowed or mempolicy nodemask, it is
4395 * possible to race with parallel threads in such a way that our
4396 * allocation can fail while the mask is being updated. If we are about
4397 * to fail, check if the cpuset changed during allocation and if so,
4400 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4406 static inline struct page
*
4407 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4408 struct alloc_context
*ac
)
4410 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4411 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4412 struct page
*page
= NULL
;
4413 unsigned int alloc_flags
;
4414 unsigned long did_some_progress
;
4415 enum compact_priority compact_priority
;
4416 enum compact_result compact_result
;
4417 int compaction_retries
;
4418 int no_progress_loops
;
4419 unsigned int cpuset_mems_cookie
;
4423 * We also sanity check to catch abuse of atomic reserves being used by
4424 * callers that are not in atomic context.
4426 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4427 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4428 gfp_mask
&= ~__GFP_ATOMIC
;
4431 compaction_retries
= 0;
4432 no_progress_loops
= 0;
4433 compact_priority
= DEF_COMPACT_PRIORITY
;
4434 cpuset_mems_cookie
= read_mems_allowed_begin();
4437 * The fast path uses conservative alloc_flags to succeed only until
4438 * kswapd needs to be woken up, and to avoid the cost of setting up
4439 * alloc_flags precisely. So we do that now.
4441 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4444 * We need to recalculate the starting point for the zonelist iterator
4445 * because we might have used different nodemask in the fast path, or
4446 * there was a cpuset modification and we are retrying - otherwise we
4447 * could end up iterating over non-eligible zones endlessly.
4449 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4450 ac
->high_zoneidx
, ac
->nodemask
);
4451 if (!ac
->preferred_zoneref
->zone
)
4454 if (alloc_flags
& ALLOC_KSWAPD
)
4455 wake_all_kswapds(order
, gfp_mask
, ac
);
4458 * The adjusted alloc_flags might result in immediate success, so try
4461 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4466 * For costly allocations, try direct compaction first, as it's likely
4467 * that we have enough base pages and don't need to reclaim. For non-
4468 * movable high-order allocations, do that as well, as compaction will
4469 * try prevent permanent fragmentation by migrating from blocks of the
4471 * Don't try this for allocations that are allowed to ignore
4472 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4474 if (can_direct_reclaim
&&
4476 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4477 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4478 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4480 INIT_COMPACT_PRIORITY
,
4485 if (order
>= pageblock_order
&& (gfp_mask
& __GFP_IO
) &&
4486 !(gfp_mask
& __GFP_RETRY_MAYFAIL
)) {
4488 * If allocating entire pageblock(s) and compaction
4489 * failed because all zones are below low watermarks
4490 * or is prohibited because it recently failed at this
4491 * order, fail immediately unless the allocator has
4492 * requested compaction and reclaim retry.
4495 * - potentially very expensive because zones are far
4496 * below their low watermarks or this is part of very
4497 * bursty high order allocations,
4498 * - not guaranteed to help because isolate_freepages()
4499 * may not iterate over freed pages as part of its
4501 * - unlikely to make entire pageblocks free on its
4504 if (compact_result
== COMPACT_SKIPPED
||
4505 compact_result
== COMPACT_DEFERRED
)
4510 * Checks for costly allocations with __GFP_NORETRY, which
4511 * includes THP page fault allocations
4513 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4515 * If compaction is deferred for high-order allocations,
4516 * it is because sync compaction recently failed. If
4517 * this is the case and the caller requested a THP
4518 * allocation, we do not want to heavily disrupt the
4519 * system, so we fail the allocation instead of entering
4522 if (compact_result
== COMPACT_DEFERRED
)
4526 * Looks like reclaim/compaction is worth trying, but
4527 * sync compaction could be very expensive, so keep
4528 * using async compaction.
4530 compact_priority
= INIT_COMPACT_PRIORITY
;
4535 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4536 if (alloc_flags
& ALLOC_KSWAPD
)
4537 wake_all_kswapds(order
, gfp_mask
, ac
);
4539 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4541 alloc_flags
= reserve_flags
;
4544 * Reset the nodemask and zonelist iterators if memory policies can be
4545 * ignored. These allocations are high priority and system rather than
4548 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4549 ac
->nodemask
= NULL
;
4550 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4551 ac
->high_zoneidx
, ac
->nodemask
);
4554 /* Attempt with potentially adjusted zonelist and alloc_flags */
4555 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4559 /* Caller is not willing to reclaim, we can't balance anything */
4560 if (!can_direct_reclaim
)
4563 /* Avoid recursion of direct reclaim */
4564 if (current
->flags
& PF_MEMALLOC
)
4567 /* Try direct reclaim and then allocating */
4568 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4569 &did_some_progress
);
4573 /* Try direct compaction and then allocating */
4574 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4575 compact_priority
, &compact_result
);
4579 /* Do not loop if specifically requested */
4580 if (gfp_mask
& __GFP_NORETRY
)
4584 * Do not retry costly high order allocations unless they are
4585 * __GFP_RETRY_MAYFAIL
4587 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4590 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4591 did_some_progress
> 0, &no_progress_loops
))
4595 * It doesn't make any sense to retry for the compaction if the order-0
4596 * reclaim is not able to make any progress because the current
4597 * implementation of the compaction depends on the sufficient amount
4598 * of free memory (see __compaction_suitable)
4600 if (did_some_progress
> 0 &&
4601 should_compact_retry(ac
, order
, alloc_flags
,
4602 compact_result
, &compact_priority
,
4603 &compaction_retries
))
4607 /* Deal with possible cpuset update races before we start OOM killing */
4608 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4611 /* Reclaim has failed us, start killing things */
4612 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4616 /* Avoid allocations with no watermarks from looping endlessly */
4617 if (tsk_is_oom_victim(current
) &&
4618 (alloc_flags
== ALLOC_OOM
||
4619 (gfp_mask
& __GFP_NOMEMALLOC
)))
4622 /* Retry as long as the OOM killer is making progress */
4623 if (did_some_progress
) {
4624 no_progress_loops
= 0;
4629 /* Deal with possible cpuset update races before we fail */
4630 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4634 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4637 if (gfp_mask
& __GFP_NOFAIL
) {
4639 * All existing users of the __GFP_NOFAIL are blockable, so warn
4640 * of any new users that actually require GFP_NOWAIT
4642 if (WARN_ON_ONCE(!can_direct_reclaim
))
4646 * PF_MEMALLOC request from this context is rather bizarre
4647 * because we cannot reclaim anything and only can loop waiting
4648 * for somebody to do a work for us
4650 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4653 * non failing costly orders are a hard requirement which we
4654 * are not prepared for much so let's warn about these users
4655 * so that we can identify them and convert them to something
4658 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4661 * Help non-failing allocations by giving them access to memory
4662 * reserves but do not use ALLOC_NO_WATERMARKS because this
4663 * could deplete whole memory reserves which would just make
4664 * the situation worse
4666 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4674 warn_alloc(gfp_mask
, ac
->nodemask
,
4675 "page allocation failure: order:%u", order
);
4680 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4681 int preferred_nid
, nodemask_t
*nodemask
,
4682 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4683 unsigned int *alloc_flags
)
4685 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4686 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4687 ac
->nodemask
= nodemask
;
4688 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4690 if (cpusets_enabled()) {
4691 *alloc_mask
|= __GFP_HARDWALL
;
4693 ac
->nodemask
= &cpuset_current_mems_allowed
;
4695 *alloc_flags
|= ALLOC_CPUSET
;
4698 fs_reclaim_acquire(gfp_mask
);
4699 fs_reclaim_release(gfp_mask
);
4701 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4703 if (should_fail_alloc_page(gfp_mask
, order
))
4706 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4707 *alloc_flags
|= ALLOC_CMA
;
4712 /* Determine whether to spread dirty pages and what the first usable zone */
4713 static inline void finalise_ac(gfp_t gfp_mask
, struct alloc_context
*ac
)
4715 /* Dirty zone balancing only done in the fast path */
4716 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4719 * The preferred zone is used for statistics but crucially it is
4720 * also used as the starting point for the zonelist iterator. It
4721 * may get reset for allocations that ignore memory policies.
4723 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4724 ac
->high_zoneidx
, ac
->nodemask
);
4728 * This is the 'heart' of the zoned buddy allocator.
4731 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4732 nodemask_t
*nodemask
)
4735 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4736 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4737 struct alloc_context ac
= { };
4740 * There are several places where we assume that the order value is sane
4741 * so bail out early if the request is out of bound.
4743 if (unlikely(order
>= MAX_ORDER
)) {
4744 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4748 gfp_mask
&= gfp_allowed_mask
;
4749 alloc_mask
= gfp_mask
;
4750 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4753 finalise_ac(gfp_mask
, &ac
);
4756 * Forbid the first pass from falling back to types that fragment
4757 * memory until all local zones are considered.
4759 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4761 /* First allocation attempt */
4762 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4767 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4768 * resp. GFP_NOIO which has to be inherited for all allocation requests
4769 * from a particular context which has been marked by
4770 * memalloc_no{fs,io}_{save,restore}.
4772 alloc_mask
= current_gfp_context(gfp_mask
);
4773 ac
.spread_dirty_pages
= false;
4776 * Restore the original nodemask if it was potentially replaced with
4777 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4779 if (unlikely(ac
.nodemask
!= nodemask
))
4780 ac
.nodemask
= nodemask
;
4782 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4785 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4786 unlikely(__memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4787 __free_pages(page
, order
);
4791 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4795 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4798 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4799 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4800 * you need to access high mem.
4802 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4806 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4809 return (unsigned long) page_address(page
);
4811 EXPORT_SYMBOL(__get_free_pages
);
4813 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4815 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4817 EXPORT_SYMBOL(get_zeroed_page
);
4819 static inline void free_the_page(struct page
*page
, unsigned int order
)
4821 if (order
== 0) /* Via pcp? */
4822 free_unref_page(page
);
4824 __free_pages_ok(page
, order
);
4827 void __free_pages(struct page
*page
, unsigned int order
)
4829 if (put_page_testzero(page
))
4830 free_the_page(page
, order
);
4832 EXPORT_SYMBOL(__free_pages
);
4834 void free_pages(unsigned long addr
, unsigned int order
)
4837 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4838 __free_pages(virt_to_page((void *)addr
), order
);
4842 EXPORT_SYMBOL(free_pages
);
4846 * An arbitrary-length arbitrary-offset area of memory which resides
4847 * within a 0 or higher order page. Multiple fragments within that page
4848 * are individually refcounted, in the page's reference counter.
4850 * The page_frag functions below provide a simple allocation framework for
4851 * page fragments. This is used by the network stack and network device
4852 * drivers to provide a backing region of memory for use as either an
4853 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4855 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4858 struct page
*page
= NULL
;
4859 gfp_t gfp
= gfp_mask
;
4861 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4862 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4864 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4865 PAGE_FRAG_CACHE_MAX_ORDER
);
4866 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4868 if (unlikely(!page
))
4869 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4871 nc
->va
= page
? page_address(page
) : NULL
;
4876 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4878 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4880 if (page_ref_sub_and_test(page
, count
))
4881 free_the_page(page
, compound_order(page
));
4883 EXPORT_SYMBOL(__page_frag_cache_drain
);
4885 void *page_frag_alloc(struct page_frag_cache
*nc
,
4886 unsigned int fragsz
, gfp_t gfp_mask
)
4888 unsigned int size
= PAGE_SIZE
;
4892 if (unlikely(!nc
->va
)) {
4894 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4898 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4899 /* if size can vary use size else just use PAGE_SIZE */
4902 /* Even if we own the page, we do not use atomic_set().
4903 * This would break get_page_unless_zero() users.
4905 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
4907 /* reset page count bias and offset to start of new frag */
4908 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4909 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4913 offset
= nc
->offset
- fragsz
;
4914 if (unlikely(offset
< 0)) {
4915 page
= virt_to_page(nc
->va
);
4917 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4920 if (unlikely(nc
->pfmemalloc
)) {
4921 free_the_page(page
, compound_order(page
));
4925 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4926 /* if size can vary use size else just use PAGE_SIZE */
4929 /* OK, page count is 0, we can safely set it */
4930 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
4932 /* reset page count bias and offset to start of new frag */
4933 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4934 offset
= size
- fragsz
;
4938 nc
->offset
= offset
;
4940 return nc
->va
+ offset
;
4942 EXPORT_SYMBOL(page_frag_alloc
);
4945 * Frees a page fragment allocated out of either a compound or order 0 page.
4947 void page_frag_free(void *addr
)
4949 struct page
*page
= virt_to_head_page(addr
);
4951 if (unlikely(put_page_testzero(page
)))
4952 free_the_page(page
, compound_order(page
));
4954 EXPORT_SYMBOL(page_frag_free
);
4956 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4960 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4961 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4963 split_page(virt_to_page((void *)addr
), order
);
4964 while (used
< alloc_end
) {
4969 return (void *)addr
;
4973 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4974 * @size: the number of bytes to allocate
4975 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4977 * This function is similar to alloc_pages(), except that it allocates the
4978 * minimum number of pages to satisfy the request. alloc_pages() can only
4979 * allocate memory in power-of-two pages.
4981 * This function is also limited by MAX_ORDER.
4983 * Memory allocated by this function must be released by free_pages_exact().
4985 * Return: pointer to the allocated area or %NULL in case of error.
4987 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4989 unsigned int order
= get_order(size
);
4992 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
4993 gfp_mask
&= ~__GFP_COMP
;
4995 addr
= __get_free_pages(gfp_mask
, order
);
4996 return make_alloc_exact(addr
, order
, size
);
4998 EXPORT_SYMBOL(alloc_pages_exact
);
5001 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5003 * @nid: the preferred node ID where memory should be allocated
5004 * @size: the number of bytes to allocate
5005 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5007 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5010 * Return: pointer to the allocated area or %NULL in case of error.
5012 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
5014 unsigned int order
= get_order(size
);
5017 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5018 gfp_mask
&= ~__GFP_COMP
;
5020 p
= alloc_pages_node(nid
, gfp_mask
, order
);
5023 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5027 * free_pages_exact - release memory allocated via alloc_pages_exact()
5028 * @virt: the value returned by alloc_pages_exact.
5029 * @size: size of allocation, same value as passed to alloc_pages_exact().
5031 * Release the memory allocated by a previous call to alloc_pages_exact.
5033 void free_pages_exact(void *virt
, size_t size
)
5035 unsigned long addr
= (unsigned long)virt
;
5036 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5038 while (addr
< end
) {
5043 EXPORT_SYMBOL(free_pages_exact
);
5046 * nr_free_zone_pages - count number of pages beyond high watermark
5047 * @offset: The zone index of the highest zone
5049 * nr_free_zone_pages() counts the number of pages which are beyond the
5050 * high watermark within all zones at or below a given zone index. For each
5051 * zone, the number of pages is calculated as:
5053 * nr_free_zone_pages = managed_pages - high_pages
5055 * Return: number of pages beyond high watermark.
5057 static unsigned long nr_free_zone_pages(int offset
)
5062 /* Just pick one node, since fallback list is circular */
5063 unsigned long sum
= 0;
5065 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5067 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5068 unsigned long size
= zone_managed_pages(zone
);
5069 unsigned long high
= high_wmark_pages(zone
);
5078 * nr_free_buffer_pages - count number of pages beyond high watermark
5080 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5081 * watermark within ZONE_DMA and ZONE_NORMAL.
5083 * Return: number of pages beyond high watermark within ZONE_DMA and
5086 unsigned long nr_free_buffer_pages(void)
5088 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5090 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5093 * nr_free_pagecache_pages - count number of pages beyond high watermark
5095 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5096 * high watermark within all zones.
5098 * Return: number of pages beyond high watermark within all zones.
5100 unsigned long nr_free_pagecache_pages(void)
5102 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5105 static inline void show_node(struct zone
*zone
)
5107 if (IS_ENABLED(CONFIG_NUMA
))
5108 printk("Node %d ", zone_to_nid(zone
));
5111 long si_mem_available(void)
5114 unsigned long pagecache
;
5115 unsigned long wmark_low
= 0;
5116 unsigned long pages
[NR_LRU_LISTS
];
5117 unsigned long reclaimable
;
5121 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5122 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5125 wmark_low
+= low_wmark_pages(zone
);
5128 * Estimate the amount of memory available for userspace allocations,
5129 * without causing swapping.
5131 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5134 * Not all the page cache can be freed, otherwise the system will
5135 * start swapping. Assume at least half of the page cache, or the
5136 * low watermark worth of cache, needs to stay.
5138 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5139 pagecache
-= min(pagecache
/ 2, wmark_low
);
5140 available
+= pagecache
;
5143 * Part of the reclaimable slab and other kernel memory consists of
5144 * items that are in use, and cannot be freed. Cap this estimate at the
5147 reclaimable
= global_node_page_state(NR_SLAB_RECLAIMABLE
) +
5148 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5149 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5155 EXPORT_SYMBOL_GPL(si_mem_available
);
5157 void si_meminfo(struct sysinfo
*val
)
5159 val
->totalram
= totalram_pages();
5160 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5161 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5162 val
->bufferram
= nr_blockdev_pages();
5163 val
->totalhigh
= totalhigh_pages();
5164 val
->freehigh
= nr_free_highpages();
5165 val
->mem_unit
= PAGE_SIZE
;
5168 EXPORT_SYMBOL(si_meminfo
);
5171 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5173 int zone_type
; /* needs to be signed */
5174 unsigned long managed_pages
= 0;
5175 unsigned long managed_highpages
= 0;
5176 unsigned long free_highpages
= 0;
5177 pg_data_t
*pgdat
= NODE_DATA(nid
);
5179 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5180 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5181 val
->totalram
= managed_pages
;
5182 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5183 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5184 #ifdef CONFIG_HIGHMEM
5185 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5186 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5188 if (is_highmem(zone
)) {
5189 managed_highpages
+= zone_managed_pages(zone
);
5190 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5193 val
->totalhigh
= managed_highpages
;
5194 val
->freehigh
= free_highpages
;
5196 val
->totalhigh
= managed_highpages
;
5197 val
->freehigh
= free_highpages
;
5199 val
->mem_unit
= PAGE_SIZE
;
5204 * Determine whether the node should be displayed or not, depending on whether
5205 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5207 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5209 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5213 * no node mask - aka implicit memory numa policy. Do not bother with
5214 * the synchronization - read_mems_allowed_begin - because we do not
5215 * have to be precise here.
5218 nodemask
= &cpuset_current_mems_allowed
;
5220 return !node_isset(nid
, *nodemask
);
5223 #define K(x) ((x) << (PAGE_SHIFT-10))
5225 static void show_migration_types(unsigned char type
)
5227 static const char types
[MIGRATE_TYPES
] = {
5228 [MIGRATE_UNMOVABLE
] = 'U',
5229 [MIGRATE_MOVABLE
] = 'M',
5230 [MIGRATE_RECLAIMABLE
] = 'E',
5231 [MIGRATE_HIGHATOMIC
] = 'H',
5233 [MIGRATE_CMA
] = 'C',
5235 #ifdef CONFIG_MEMORY_ISOLATION
5236 [MIGRATE_ISOLATE
] = 'I',
5239 char tmp
[MIGRATE_TYPES
+ 1];
5243 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5244 if (type
& (1 << i
))
5249 printk(KERN_CONT
"(%s) ", tmp
);
5253 * Show free area list (used inside shift_scroll-lock stuff)
5254 * We also calculate the percentage fragmentation. We do this by counting the
5255 * memory on each free list with the exception of the first item on the list.
5258 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5261 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5263 unsigned long free_pcp
= 0;
5268 for_each_populated_zone(zone
) {
5269 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5272 for_each_online_cpu(cpu
)
5273 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5276 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5277 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5278 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5279 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5280 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5281 " free:%lu free_pcp:%lu free_cma:%lu\n",
5282 global_node_page_state(NR_ACTIVE_ANON
),
5283 global_node_page_state(NR_INACTIVE_ANON
),
5284 global_node_page_state(NR_ISOLATED_ANON
),
5285 global_node_page_state(NR_ACTIVE_FILE
),
5286 global_node_page_state(NR_INACTIVE_FILE
),
5287 global_node_page_state(NR_ISOLATED_FILE
),
5288 global_node_page_state(NR_UNEVICTABLE
),
5289 global_node_page_state(NR_FILE_DIRTY
),
5290 global_node_page_state(NR_WRITEBACK
),
5291 global_node_page_state(NR_UNSTABLE_NFS
),
5292 global_node_page_state(NR_SLAB_RECLAIMABLE
),
5293 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
5294 global_node_page_state(NR_FILE_MAPPED
),
5295 global_node_page_state(NR_SHMEM
),
5296 global_zone_page_state(NR_PAGETABLE
),
5297 global_zone_page_state(NR_BOUNCE
),
5298 global_zone_page_state(NR_FREE_PAGES
),
5300 global_zone_page_state(NR_FREE_CMA_PAGES
));
5302 for_each_online_pgdat(pgdat
) {
5303 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5307 " active_anon:%lukB"
5308 " inactive_anon:%lukB"
5309 " active_file:%lukB"
5310 " inactive_file:%lukB"
5311 " unevictable:%lukB"
5312 " isolated(anon):%lukB"
5313 " isolated(file):%lukB"
5318 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5320 " shmem_pmdmapped: %lukB"
5323 " writeback_tmp:%lukB"
5325 " all_unreclaimable? %s"
5328 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5329 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5330 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5331 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5332 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5333 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5334 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5335 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5336 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5337 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5338 K(node_page_state(pgdat
, NR_SHMEM
)),
5339 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5340 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5341 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5343 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5345 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5346 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
5347 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5351 for_each_populated_zone(zone
) {
5354 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5358 for_each_online_cpu(cpu
)
5359 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5368 " active_anon:%lukB"
5369 " inactive_anon:%lukB"
5370 " active_file:%lukB"
5371 " inactive_file:%lukB"
5372 " unevictable:%lukB"
5373 " writepending:%lukB"
5377 " kernel_stack:%lukB"
5385 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5386 K(min_wmark_pages(zone
)),
5387 K(low_wmark_pages(zone
)),
5388 K(high_wmark_pages(zone
)),
5389 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5390 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5391 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5392 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5393 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5394 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5395 K(zone
->present_pages
),
5396 K(zone_managed_pages(zone
)),
5397 K(zone_page_state(zone
, NR_MLOCK
)),
5398 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
5399 K(zone_page_state(zone
, NR_PAGETABLE
)),
5400 K(zone_page_state(zone
, NR_BOUNCE
)),
5402 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5403 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5404 printk("lowmem_reserve[]:");
5405 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5406 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5407 printk(KERN_CONT
"\n");
5410 for_each_populated_zone(zone
) {
5412 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5413 unsigned char types
[MAX_ORDER
];
5415 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5418 printk(KERN_CONT
"%s: ", zone
->name
);
5420 spin_lock_irqsave(&zone
->lock
, flags
);
5421 for (order
= 0; order
< MAX_ORDER
; order
++) {
5422 struct free_area
*area
= &zone
->free_area
[order
];
5425 nr
[order
] = area
->nr_free
;
5426 total
+= nr
[order
] << order
;
5429 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5430 if (!free_area_empty(area
, type
))
5431 types
[order
] |= 1 << type
;
5434 spin_unlock_irqrestore(&zone
->lock
, flags
);
5435 for (order
= 0; order
< MAX_ORDER
; order
++) {
5436 printk(KERN_CONT
"%lu*%lukB ",
5437 nr
[order
], K(1UL) << order
);
5439 show_migration_types(types
[order
]);
5441 printk(KERN_CONT
"= %lukB\n", K(total
));
5444 hugetlb_show_meminfo();
5446 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5448 show_swap_cache_info();
5451 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5453 zoneref
->zone
= zone
;
5454 zoneref
->zone_idx
= zone_idx(zone
);
5458 * Builds allocation fallback zone lists.
5460 * Add all populated zones of a node to the zonelist.
5462 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5465 enum zone_type zone_type
= MAX_NR_ZONES
;
5470 zone
= pgdat
->node_zones
+ zone_type
;
5471 if (managed_zone(zone
)) {
5472 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5473 check_highest_zone(zone_type
);
5475 } while (zone_type
);
5482 static int __parse_numa_zonelist_order(char *s
)
5485 * We used to support different zonlists modes but they turned
5486 * out to be just not useful. Let's keep the warning in place
5487 * if somebody still use the cmd line parameter so that we do
5488 * not fail it silently
5490 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5491 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5497 static __init
int setup_numa_zonelist_order(char *s
)
5502 return __parse_numa_zonelist_order(s
);
5504 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
5506 char numa_zonelist_order
[] = "Node";
5509 * sysctl handler for numa_zonelist_order
5511 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5512 void __user
*buffer
, size_t *length
,
5519 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5520 str
= memdup_user_nul(buffer
, 16);
5522 return PTR_ERR(str
);
5524 ret
= __parse_numa_zonelist_order(str
);
5530 #define MAX_NODE_LOAD (nr_online_nodes)
5531 static int node_load
[MAX_NUMNODES
];
5534 * find_next_best_node - find the next node that should appear in a given node's fallback list
5535 * @node: node whose fallback list we're appending
5536 * @used_node_mask: nodemask_t of already used nodes
5538 * We use a number of factors to determine which is the next node that should
5539 * appear on a given node's fallback list. The node should not have appeared
5540 * already in @node's fallback list, and it should be the next closest node
5541 * according to the distance array (which contains arbitrary distance values
5542 * from each node to each node in the system), and should also prefer nodes
5543 * with no CPUs, since presumably they'll have very little allocation pressure
5544 * on them otherwise.
5546 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5548 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5551 int min_val
= INT_MAX
;
5552 int best_node
= NUMA_NO_NODE
;
5553 const struct cpumask
*tmp
= cpumask_of_node(0);
5555 /* Use the local node if we haven't already */
5556 if (!node_isset(node
, *used_node_mask
)) {
5557 node_set(node
, *used_node_mask
);
5561 for_each_node_state(n
, N_MEMORY
) {
5563 /* Don't want a node to appear more than once */
5564 if (node_isset(n
, *used_node_mask
))
5567 /* Use the distance array to find the distance */
5568 val
= node_distance(node
, n
);
5570 /* Penalize nodes under us ("prefer the next node") */
5573 /* Give preference to headless and unused nodes */
5574 tmp
= cpumask_of_node(n
);
5575 if (!cpumask_empty(tmp
))
5576 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5578 /* Slight preference for less loaded node */
5579 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5580 val
+= node_load
[n
];
5582 if (val
< min_val
) {
5589 node_set(best_node
, *used_node_mask
);
5596 * Build zonelists ordered by node and zones within node.
5597 * This results in maximum locality--normal zone overflows into local
5598 * DMA zone, if any--but risks exhausting DMA zone.
5600 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5603 struct zoneref
*zonerefs
;
5606 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5608 for (i
= 0; i
< nr_nodes
; i
++) {
5611 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5613 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5614 zonerefs
+= nr_zones
;
5616 zonerefs
->zone
= NULL
;
5617 zonerefs
->zone_idx
= 0;
5621 * Build gfp_thisnode zonelists
5623 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5625 struct zoneref
*zonerefs
;
5628 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5629 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5630 zonerefs
+= nr_zones
;
5631 zonerefs
->zone
= NULL
;
5632 zonerefs
->zone_idx
= 0;
5636 * Build zonelists ordered by zone and nodes within zones.
5637 * This results in conserving DMA zone[s] until all Normal memory is
5638 * exhausted, but results in overflowing to remote node while memory
5639 * may still exist in local DMA zone.
5642 static void build_zonelists(pg_data_t
*pgdat
)
5644 static int node_order
[MAX_NUMNODES
];
5645 int node
, load
, nr_nodes
= 0;
5646 nodemask_t used_mask
;
5647 int local_node
, prev_node
;
5649 /* NUMA-aware ordering of nodes */
5650 local_node
= pgdat
->node_id
;
5651 load
= nr_online_nodes
;
5652 prev_node
= local_node
;
5653 nodes_clear(used_mask
);
5655 memset(node_order
, 0, sizeof(node_order
));
5656 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5658 * We don't want to pressure a particular node.
5659 * So adding penalty to the first node in same
5660 * distance group to make it round-robin.
5662 if (node_distance(local_node
, node
) !=
5663 node_distance(local_node
, prev_node
))
5664 node_load
[node
] = load
;
5666 node_order
[nr_nodes
++] = node
;
5671 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5672 build_thisnode_zonelists(pgdat
);
5675 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5677 * Return node id of node used for "local" allocations.
5678 * I.e., first node id of first zone in arg node's generic zonelist.
5679 * Used for initializing percpu 'numa_mem', which is used primarily
5680 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5682 int local_memory_node(int node
)
5686 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5687 gfp_zone(GFP_KERNEL
),
5689 return zone_to_nid(z
->zone
);
5693 static void setup_min_unmapped_ratio(void);
5694 static void setup_min_slab_ratio(void);
5695 #else /* CONFIG_NUMA */
5697 static void build_zonelists(pg_data_t
*pgdat
)
5699 int node
, local_node
;
5700 struct zoneref
*zonerefs
;
5703 local_node
= pgdat
->node_id
;
5705 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5706 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5707 zonerefs
+= nr_zones
;
5710 * Now we build the zonelist so that it contains the zones
5711 * of all the other nodes.
5712 * We don't want to pressure a particular node, so when
5713 * building the zones for node N, we make sure that the
5714 * zones coming right after the local ones are those from
5715 * node N+1 (modulo N)
5717 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5718 if (!node_online(node
))
5720 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5721 zonerefs
+= nr_zones
;
5723 for (node
= 0; node
< local_node
; node
++) {
5724 if (!node_online(node
))
5726 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5727 zonerefs
+= nr_zones
;
5730 zonerefs
->zone
= NULL
;
5731 zonerefs
->zone_idx
= 0;
5734 #endif /* CONFIG_NUMA */
5737 * Boot pageset table. One per cpu which is going to be used for all
5738 * zones and all nodes. The parameters will be set in such a way
5739 * that an item put on a list will immediately be handed over to
5740 * the buddy list. This is safe since pageset manipulation is done
5741 * with interrupts disabled.
5743 * The boot_pagesets must be kept even after bootup is complete for
5744 * unused processors and/or zones. They do play a role for bootstrapping
5745 * hotplugged processors.
5747 * zoneinfo_show() and maybe other functions do
5748 * not check if the processor is online before following the pageset pointer.
5749 * Other parts of the kernel may not check if the zone is available.
5751 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5752 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5753 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5755 static void __build_all_zonelists(void *data
)
5758 int __maybe_unused cpu
;
5759 pg_data_t
*self
= data
;
5760 static DEFINE_SPINLOCK(lock
);
5765 memset(node_load
, 0, sizeof(node_load
));
5769 * This node is hotadded and no memory is yet present. So just
5770 * building zonelists is fine - no need to touch other nodes.
5772 if (self
&& !node_online(self
->node_id
)) {
5773 build_zonelists(self
);
5775 for_each_online_node(nid
) {
5776 pg_data_t
*pgdat
= NODE_DATA(nid
);
5778 build_zonelists(pgdat
);
5781 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5783 * We now know the "local memory node" for each node--
5784 * i.e., the node of the first zone in the generic zonelist.
5785 * Set up numa_mem percpu variable for on-line cpus. During
5786 * boot, only the boot cpu should be on-line; we'll init the
5787 * secondary cpus' numa_mem as they come on-line. During
5788 * node/memory hotplug, we'll fixup all on-line cpus.
5790 for_each_online_cpu(cpu
)
5791 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5798 static noinline
void __init
5799 build_all_zonelists_init(void)
5803 __build_all_zonelists(NULL
);
5806 * Initialize the boot_pagesets that are going to be used
5807 * for bootstrapping processors. The real pagesets for
5808 * each zone will be allocated later when the per cpu
5809 * allocator is available.
5811 * boot_pagesets are used also for bootstrapping offline
5812 * cpus if the system is already booted because the pagesets
5813 * are needed to initialize allocators on a specific cpu too.
5814 * F.e. the percpu allocator needs the page allocator which
5815 * needs the percpu allocator in order to allocate its pagesets
5816 * (a chicken-egg dilemma).
5818 for_each_possible_cpu(cpu
)
5819 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5821 mminit_verify_zonelist();
5822 cpuset_init_current_mems_allowed();
5826 * unless system_state == SYSTEM_BOOTING.
5828 * __ref due to call of __init annotated helper build_all_zonelists_init
5829 * [protected by SYSTEM_BOOTING].
5831 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5833 if (system_state
== SYSTEM_BOOTING
) {
5834 build_all_zonelists_init();
5836 __build_all_zonelists(pgdat
);
5837 /* cpuset refresh routine should be here */
5839 vm_total_pages
= nr_free_pagecache_pages();
5841 * Disable grouping by mobility if the number of pages in the
5842 * system is too low to allow the mechanism to work. It would be
5843 * more accurate, but expensive to check per-zone. This check is
5844 * made on memory-hotadd so a system can start with mobility
5845 * disabled and enable it later
5847 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5848 page_group_by_mobility_disabled
= 1;
5850 page_group_by_mobility_disabled
= 0;
5852 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5854 page_group_by_mobility_disabled
? "off" : "on",
5857 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5861 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5862 static bool __meminit
5863 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5865 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5866 static struct memblock_region
*r
;
5868 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5869 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5870 for_each_memblock(memory
, r
) {
5871 if (*pfn
< memblock_region_memory_end_pfn(r
))
5875 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5876 memblock_is_mirror(r
)) {
5877 *pfn
= memblock_region_memory_end_pfn(r
);
5886 * Initially all pages are reserved - free ones are freed
5887 * up by memblock_free_all() once the early boot process is
5888 * done. Non-atomic initialization, single-pass.
5890 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5891 unsigned long start_pfn
, enum meminit_context context
,
5892 struct vmem_altmap
*altmap
)
5894 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5897 if (highest_memmap_pfn
< end_pfn
- 1)
5898 highest_memmap_pfn
= end_pfn
- 1;
5900 #ifdef CONFIG_ZONE_DEVICE
5902 * Honor reservation requested by the driver for this ZONE_DEVICE
5903 * memory. We limit the total number of pages to initialize to just
5904 * those that might contain the memory mapping. We will defer the
5905 * ZONE_DEVICE page initialization until after we have released
5908 if (zone
== ZONE_DEVICE
) {
5912 if (start_pfn
== altmap
->base_pfn
)
5913 start_pfn
+= altmap
->reserve
;
5914 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5918 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5920 * There can be holes in boot-time mem_map[]s handed to this
5921 * function. They do not exist on hotplugged memory.
5923 if (context
== MEMINIT_EARLY
) {
5924 if (!early_pfn_valid(pfn
))
5926 if (!early_pfn_in_nid(pfn
, nid
))
5928 if (overlap_memmap_init(zone
, &pfn
))
5930 if (defer_init(nid
, pfn
, end_pfn
))
5934 page
= pfn_to_page(pfn
);
5935 __init_single_page(page
, pfn
, zone
, nid
);
5936 if (context
== MEMINIT_HOTPLUG
)
5937 __SetPageReserved(page
);
5940 * Mark the block movable so that blocks are reserved for
5941 * movable at startup. This will force kernel allocations
5942 * to reserve their blocks rather than leaking throughout
5943 * the address space during boot when many long-lived
5944 * kernel allocations are made.
5946 * bitmap is created for zone's valid pfn range. but memmap
5947 * can be created for invalid pages (for alignment)
5948 * check here not to call set_pageblock_migratetype() against
5951 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5952 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5958 #ifdef CONFIG_ZONE_DEVICE
5959 void __ref
memmap_init_zone_device(struct zone
*zone
,
5960 unsigned long start_pfn
,
5962 struct dev_pagemap
*pgmap
)
5964 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
5965 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5966 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
5967 unsigned long zone_idx
= zone_idx(zone
);
5968 unsigned long start
= jiffies
;
5969 int nid
= pgdat
->node_id
;
5971 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
5975 * The call to memmap_init_zone should have already taken care
5976 * of the pages reserved for the memmap, so we can just jump to
5977 * the end of that region and start processing the device pages.
5980 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
5981 size
= end_pfn
- start_pfn
;
5984 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5985 struct page
*page
= pfn_to_page(pfn
);
5987 __init_single_page(page
, pfn
, zone_idx
, nid
);
5990 * Mark page reserved as it will need to wait for onlining
5991 * phase for it to be fully associated with a zone.
5993 * We can use the non-atomic __set_bit operation for setting
5994 * the flag as we are still initializing the pages.
5996 __SetPageReserved(page
);
5999 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6000 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6001 * ever freed or placed on a driver-private list.
6003 page
->pgmap
= pgmap
;
6004 page
->zone_device_data
= NULL
;
6007 * Mark the block movable so that blocks are reserved for
6008 * movable at startup. This will force kernel allocations
6009 * to reserve their blocks rather than leaking throughout
6010 * the address space during boot when many long-lived
6011 * kernel allocations are made.
6013 * bitmap is created for zone's valid pfn range. but memmap
6014 * can be created for invalid pages (for alignment)
6015 * check here not to call set_pageblock_migratetype() against
6018 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6019 * because this is done early in section_activate()
6021 if (!(pfn
& (pageblock_nr_pages
- 1))) {
6022 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6027 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6028 size
, jiffies_to_msecs(jiffies
- start
));
6032 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6034 unsigned int order
, t
;
6035 for_each_migratetype_order(order
, t
) {
6036 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6037 zone
->free_area
[order
].nr_free
= 0;
6041 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
6042 unsigned long zone
, unsigned long start_pfn
)
6044 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMINIT_EARLY
, NULL
);
6047 static int zone_batchsize(struct zone
*zone
)
6053 * The per-cpu-pages pools are set to around 1000th of the
6056 batch
= zone_managed_pages(zone
) / 1024;
6057 /* But no more than a meg. */
6058 if (batch
* PAGE_SIZE
> 1024 * 1024)
6059 batch
= (1024 * 1024) / PAGE_SIZE
;
6060 batch
/= 4; /* We effectively *= 4 below */
6065 * Clamp the batch to a 2^n - 1 value. Having a power
6066 * of 2 value was found to be more likely to have
6067 * suboptimal cache aliasing properties in some cases.
6069 * For example if 2 tasks are alternately allocating
6070 * batches of pages, one task can end up with a lot
6071 * of pages of one half of the possible page colors
6072 * and the other with pages of the other colors.
6074 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6079 /* The deferral and batching of frees should be suppressed under NOMMU
6082 * The problem is that NOMMU needs to be able to allocate large chunks
6083 * of contiguous memory as there's no hardware page translation to
6084 * assemble apparent contiguous memory from discontiguous pages.
6086 * Queueing large contiguous runs of pages for batching, however,
6087 * causes the pages to actually be freed in smaller chunks. As there
6088 * can be a significant delay between the individual batches being
6089 * recycled, this leads to the once large chunks of space being
6090 * fragmented and becoming unavailable for high-order allocations.
6097 * pcp->high and pcp->batch values are related and dependent on one another:
6098 * ->batch must never be higher then ->high.
6099 * The following function updates them in a safe manner without read side
6102 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6103 * those fields changing asynchronously (acording the the above rule).
6105 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6106 * outside of boot time (or some other assurance that no concurrent updaters
6109 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6110 unsigned long batch
)
6112 /* start with a fail safe value for batch */
6116 /* Update high, then batch, in order */
6123 /* a companion to pageset_set_high() */
6124 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
6126 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
6129 static void pageset_init(struct per_cpu_pageset
*p
)
6131 struct per_cpu_pages
*pcp
;
6134 memset(p
, 0, sizeof(*p
));
6137 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
6138 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
6141 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
6144 pageset_set_batch(p
, batch
);
6148 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6149 * to the value high for the pageset p.
6151 static void pageset_set_high(struct per_cpu_pageset
*p
,
6154 unsigned long batch
= max(1UL, high
/ 4);
6155 if ((high
/ 4) > (PAGE_SHIFT
* 8))
6156 batch
= PAGE_SHIFT
* 8;
6158 pageset_update(&p
->pcp
, high
, batch
);
6161 static void pageset_set_high_and_batch(struct zone
*zone
,
6162 struct per_cpu_pageset
*pcp
)
6164 if (percpu_pagelist_fraction
)
6165 pageset_set_high(pcp
,
6166 (zone_managed_pages(zone
) /
6167 percpu_pagelist_fraction
));
6169 pageset_set_batch(pcp
, zone_batchsize(zone
));
6172 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
6174 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
6177 pageset_set_high_and_batch(zone
, pcp
);
6180 void __meminit
setup_zone_pageset(struct zone
*zone
)
6183 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
6184 for_each_possible_cpu(cpu
)
6185 zone_pageset_init(zone
, cpu
);
6189 * Allocate per cpu pagesets and initialize them.
6190 * Before this call only boot pagesets were available.
6192 void __init
setup_per_cpu_pageset(void)
6194 struct pglist_data
*pgdat
;
6197 for_each_populated_zone(zone
)
6198 setup_zone_pageset(zone
);
6200 for_each_online_pgdat(pgdat
)
6201 pgdat
->per_cpu_nodestats
=
6202 alloc_percpu(struct per_cpu_nodestat
);
6205 static __meminit
void zone_pcp_init(struct zone
*zone
)
6208 * per cpu subsystem is not up at this point. The following code
6209 * relies on the ability of the linker to provide the
6210 * offset of a (static) per cpu variable into the per cpu area.
6212 zone
->pageset
= &boot_pageset
;
6214 if (populated_zone(zone
))
6215 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
6216 zone
->name
, zone
->present_pages
,
6217 zone_batchsize(zone
));
6220 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6221 unsigned long zone_start_pfn
,
6224 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6225 int zone_idx
= zone_idx(zone
) + 1;
6227 if (zone_idx
> pgdat
->nr_zones
)
6228 pgdat
->nr_zones
= zone_idx
;
6230 zone
->zone_start_pfn
= zone_start_pfn
;
6232 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6233 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6235 (unsigned long)zone_idx(zone
),
6236 zone_start_pfn
, (zone_start_pfn
+ size
));
6238 zone_init_free_lists(zone
);
6239 zone
->initialized
= 1;
6242 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6243 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6246 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6248 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
6249 struct mminit_pfnnid_cache
*state
)
6251 unsigned long start_pfn
, end_pfn
;
6254 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
6255 return state
->last_nid
;
6257 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
6258 if (nid
!= NUMA_NO_NODE
) {
6259 state
->last_start
= start_pfn
;
6260 state
->last_end
= end_pfn
;
6261 state
->last_nid
= nid
;
6266 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6269 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6270 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6271 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6273 * If an architecture guarantees that all ranges registered contain no holes
6274 * and may be freed, this this function may be used instead of calling
6275 * memblock_free_early_nid() manually.
6277 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
6279 unsigned long start_pfn
, end_pfn
;
6282 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
6283 start_pfn
= min(start_pfn
, max_low_pfn
);
6284 end_pfn
= min(end_pfn
, max_low_pfn
);
6286 if (start_pfn
< end_pfn
)
6287 memblock_free_early_nid(PFN_PHYS(start_pfn
),
6288 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
6294 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6295 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6297 * If an architecture guarantees that all ranges registered contain no holes and may
6298 * be freed, this function may be used instead of calling memory_present() manually.
6300 void __init
sparse_memory_present_with_active_regions(int nid
)
6302 unsigned long start_pfn
, end_pfn
;
6305 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
6306 memory_present(this_nid
, start_pfn
, end_pfn
);
6310 * get_pfn_range_for_nid - Return the start and end page frames for a node
6311 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6312 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6313 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6315 * It returns the start and end page frame of a node based on information
6316 * provided by memblock_set_node(). If called for a node
6317 * with no available memory, a warning is printed and the start and end
6320 void __init
get_pfn_range_for_nid(unsigned int nid
,
6321 unsigned long *start_pfn
, unsigned long *end_pfn
)
6323 unsigned long this_start_pfn
, this_end_pfn
;
6329 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6330 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6331 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6334 if (*start_pfn
== -1UL)
6339 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6340 * assumption is made that zones within a node are ordered in monotonic
6341 * increasing memory addresses so that the "highest" populated zone is used
6343 static void __init
find_usable_zone_for_movable(void)
6346 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6347 if (zone_index
== ZONE_MOVABLE
)
6350 if (arch_zone_highest_possible_pfn
[zone_index
] >
6351 arch_zone_lowest_possible_pfn
[zone_index
])
6355 VM_BUG_ON(zone_index
== -1);
6356 movable_zone
= zone_index
;
6360 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6361 * because it is sized independent of architecture. Unlike the other zones,
6362 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6363 * in each node depending on the size of each node and how evenly kernelcore
6364 * is distributed. This helper function adjusts the zone ranges
6365 * provided by the architecture for a given node by using the end of the
6366 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6367 * zones within a node are in order of monotonic increases memory addresses
6369 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6370 unsigned long zone_type
,
6371 unsigned long node_start_pfn
,
6372 unsigned long node_end_pfn
,
6373 unsigned long *zone_start_pfn
,
6374 unsigned long *zone_end_pfn
)
6376 /* Only adjust if ZONE_MOVABLE is on this node */
6377 if (zone_movable_pfn
[nid
]) {
6378 /* Size ZONE_MOVABLE */
6379 if (zone_type
== ZONE_MOVABLE
) {
6380 *zone_start_pfn
= zone_movable_pfn
[nid
];
6381 *zone_end_pfn
= min(node_end_pfn
,
6382 arch_zone_highest_possible_pfn
[movable_zone
]);
6384 /* Adjust for ZONE_MOVABLE starting within this range */
6385 } else if (!mirrored_kernelcore
&&
6386 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6387 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6388 *zone_end_pfn
= zone_movable_pfn
[nid
];
6390 /* Check if this whole range is within ZONE_MOVABLE */
6391 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6392 *zone_start_pfn
= *zone_end_pfn
;
6397 * Return the number of pages a zone spans in a node, including holes
6398 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6400 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6401 unsigned long zone_type
,
6402 unsigned long node_start_pfn
,
6403 unsigned long node_end_pfn
,
6404 unsigned long *zone_start_pfn
,
6405 unsigned long *zone_end_pfn
,
6406 unsigned long *ignored
)
6408 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6409 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6410 /* When hotadd a new node from cpu_up(), the node should be empty */
6411 if (!node_start_pfn
&& !node_end_pfn
)
6414 /* Get the start and end of the zone */
6415 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6416 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6417 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6418 node_start_pfn
, node_end_pfn
,
6419 zone_start_pfn
, zone_end_pfn
);
6421 /* Check that this node has pages within the zone's required range */
6422 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6425 /* Move the zone boundaries inside the node if necessary */
6426 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6427 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6429 /* Return the spanned pages */
6430 return *zone_end_pfn
- *zone_start_pfn
;
6434 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6435 * then all holes in the requested range will be accounted for.
6437 unsigned long __init
__absent_pages_in_range(int nid
,
6438 unsigned long range_start_pfn
,
6439 unsigned long range_end_pfn
)
6441 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6442 unsigned long start_pfn
, end_pfn
;
6445 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6446 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6447 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6448 nr_absent
-= end_pfn
- start_pfn
;
6454 * absent_pages_in_range - Return number of page frames in holes within a range
6455 * @start_pfn: The start PFN to start searching for holes
6456 * @end_pfn: The end PFN to stop searching for holes
6458 * Return: the number of pages frames in memory holes within a range.
6460 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6461 unsigned long end_pfn
)
6463 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6466 /* Return the number of page frames in holes in a zone on a node */
6467 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6468 unsigned long zone_type
,
6469 unsigned long node_start_pfn
,
6470 unsigned long node_end_pfn
,
6471 unsigned long *ignored
)
6473 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6474 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6475 unsigned long zone_start_pfn
, zone_end_pfn
;
6476 unsigned long nr_absent
;
6478 /* When hotadd a new node from cpu_up(), the node should be empty */
6479 if (!node_start_pfn
&& !node_end_pfn
)
6482 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6483 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6485 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6486 node_start_pfn
, node_end_pfn
,
6487 &zone_start_pfn
, &zone_end_pfn
);
6488 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6491 * ZONE_MOVABLE handling.
6492 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6495 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6496 unsigned long start_pfn
, end_pfn
;
6497 struct memblock_region
*r
;
6499 for_each_memblock(memory
, r
) {
6500 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6501 zone_start_pfn
, zone_end_pfn
);
6502 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6503 zone_start_pfn
, zone_end_pfn
);
6505 if (zone_type
== ZONE_MOVABLE
&&
6506 memblock_is_mirror(r
))
6507 nr_absent
+= end_pfn
- start_pfn
;
6509 if (zone_type
== ZONE_NORMAL
&&
6510 !memblock_is_mirror(r
))
6511 nr_absent
+= end_pfn
- start_pfn
;
6518 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6519 static inline unsigned long __init
zone_spanned_pages_in_node(int nid
,
6520 unsigned long zone_type
,
6521 unsigned long node_start_pfn
,
6522 unsigned long node_end_pfn
,
6523 unsigned long *zone_start_pfn
,
6524 unsigned long *zone_end_pfn
,
6525 unsigned long *zones_size
)
6529 *zone_start_pfn
= node_start_pfn
;
6530 for (zone
= 0; zone
< zone_type
; zone
++)
6531 *zone_start_pfn
+= zones_size
[zone
];
6533 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
6535 return zones_size
[zone_type
];
6538 static inline unsigned long __init
zone_absent_pages_in_node(int nid
,
6539 unsigned long zone_type
,
6540 unsigned long node_start_pfn
,
6541 unsigned long node_end_pfn
,
6542 unsigned long *zholes_size
)
6547 return zholes_size
[zone_type
];
6550 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6552 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6553 unsigned long node_start_pfn
,
6554 unsigned long node_end_pfn
,
6555 unsigned long *zones_size
,
6556 unsigned long *zholes_size
)
6558 unsigned long realtotalpages
= 0, totalpages
= 0;
6561 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6562 struct zone
*zone
= pgdat
->node_zones
+ i
;
6563 unsigned long zone_start_pfn
, zone_end_pfn
;
6564 unsigned long size
, real_size
;
6566 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6572 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
6573 node_start_pfn
, node_end_pfn
,
6576 zone
->zone_start_pfn
= zone_start_pfn
;
6578 zone
->zone_start_pfn
= 0;
6579 zone
->spanned_pages
= size
;
6580 zone
->present_pages
= real_size
;
6583 realtotalpages
+= real_size
;
6586 pgdat
->node_spanned_pages
= totalpages
;
6587 pgdat
->node_present_pages
= realtotalpages
;
6588 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6592 #ifndef CONFIG_SPARSEMEM
6594 * Calculate the size of the zone->blockflags rounded to an unsigned long
6595 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6596 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6597 * round what is now in bits to nearest long in bits, then return it in
6600 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6602 unsigned long usemapsize
;
6604 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6605 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6606 usemapsize
= usemapsize
>> pageblock_order
;
6607 usemapsize
*= NR_PAGEBLOCK_BITS
;
6608 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6610 return usemapsize
/ 8;
6613 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6615 unsigned long zone_start_pfn
,
6616 unsigned long zonesize
)
6618 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6619 zone
->pageblock_flags
= NULL
;
6621 zone
->pageblock_flags
=
6622 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
6624 if (!zone
->pageblock_flags
)
6625 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6626 usemapsize
, zone
->name
, pgdat
->node_id
);
6630 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6631 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6632 #endif /* CONFIG_SPARSEMEM */
6634 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6636 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6637 void __init
set_pageblock_order(void)
6641 /* Check that pageblock_nr_pages has not already been setup */
6642 if (pageblock_order
)
6645 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6646 order
= HUGETLB_PAGE_ORDER
;
6648 order
= MAX_ORDER
- 1;
6651 * Assume the largest contiguous order of interest is a huge page.
6652 * This value may be variable depending on boot parameters on IA64 and
6655 pageblock_order
= order
;
6657 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6660 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6661 * is unused as pageblock_order is set at compile-time. See
6662 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6665 void __init
set_pageblock_order(void)
6669 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6671 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6672 unsigned long present_pages
)
6674 unsigned long pages
= spanned_pages
;
6677 * Provide a more accurate estimation if there are holes within
6678 * the zone and SPARSEMEM is in use. If there are holes within the
6679 * zone, each populated memory region may cost us one or two extra
6680 * memmap pages due to alignment because memmap pages for each
6681 * populated regions may not be naturally aligned on page boundary.
6682 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6684 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6685 IS_ENABLED(CONFIG_SPARSEMEM
))
6686 pages
= present_pages
;
6688 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6691 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6692 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6694 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
6696 spin_lock_init(&ds_queue
->split_queue_lock
);
6697 INIT_LIST_HEAD(&ds_queue
->split_queue
);
6698 ds_queue
->split_queue_len
= 0;
6701 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6704 #ifdef CONFIG_COMPACTION
6705 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6707 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6710 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6713 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6715 pgdat_resize_init(pgdat
);
6717 pgdat_init_split_queue(pgdat
);
6718 pgdat_init_kcompactd(pgdat
);
6720 init_waitqueue_head(&pgdat
->kswapd_wait
);
6721 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6723 pgdat_page_ext_init(pgdat
);
6724 spin_lock_init(&pgdat
->lru_lock
);
6725 lruvec_init(node_lruvec(pgdat
));
6728 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6729 unsigned long remaining_pages
)
6731 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6732 zone_set_nid(zone
, nid
);
6733 zone
->name
= zone_names
[idx
];
6734 zone
->zone_pgdat
= NODE_DATA(nid
);
6735 spin_lock_init(&zone
->lock
);
6736 zone_seqlock_init(zone
);
6737 zone_pcp_init(zone
);
6741 * Set up the zone data structures
6742 * - init pgdat internals
6743 * - init all zones belonging to this node
6745 * NOTE: this function is only called during memory hotplug
6747 #ifdef CONFIG_MEMORY_HOTPLUG
6748 void __ref
free_area_init_core_hotplug(int nid
)
6751 pg_data_t
*pgdat
= NODE_DATA(nid
);
6753 pgdat_init_internals(pgdat
);
6754 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6755 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6760 * Set up the zone data structures:
6761 * - mark all pages reserved
6762 * - mark all memory queues empty
6763 * - clear the memory bitmaps
6765 * NOTE: pgdat should get zeroed by caller.
6766 * NOTE: this function is only called during early init.
6768 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6771 int nid
= pgdat
->node_id
;
6773 pgdat_init_internals(pgdat
);
6774 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6776 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6777 struct zone
*zone
= pgdat
->node_zones
+ j
;
6778 unsigned long size
, freesize
, memmap_pages
;
6779 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6781 size
= zone
->spanned_pages
;
6782 freesize
= zone
->present_pages
;
6785 * Adjust freesize so that it accounts for how much memory
6786 * is used by this zone for memmap. This affects the watermark
6787 * and per-cpu initialisations
6789 memmap_pages
= calc_memmap_size(size
, freesize
);
6790 if (!is_highmem_idx(j
)) {
6791 if (freesize
>= memmap_pages
) {
6792 freesize
-= memmap_pages
;
6795 " %s zone: %lu pages used for memmap\n",
6796 zone_names
[j
], memmap_pages
);
6798 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6799 zone_names
[j
], memmap_pages
, freesize
);
6802 /* Account for reserved pages */
6803 if (j
== 0 && freesize
> dma_reserve
) {
6804 freesize
-= dma_reserve
;
6805 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6806 zone_names
[0], dma_reserve
);
6809 if (!is_highmem_idx(j
))
6810 nr_kernel_pages
+= freesize
;
6811 /* Charge for highmem memmap if there are enough kernel pages */
6812 else if (nr_kernel_pages
> memmap_pages
* 2)
6813 nr_kernel_pages
-= memmap_pages
;
6814 nr_all_pages
+= freesize
;
6817 * Set an approximate value for lowmem here, it will be adjusted
6818 * when the bootmem allocator frees pages into the buddy system.
6819 * And all highmem pages will be managed by the buddy system.
6821 zone_init_internals(zone
, j
, nid
, freesize
);
6826 set_pageblock_order();
6827 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6828 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6829 memmap_init(size
, nid
, j
, zone_start_pfn
);
6833 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6834 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6836 unsigned long __maybe_unused start
= 0;
6837 unsigned long __maybe_unused offset
= 0;
6839 /* Skip empty nodes */
6840 if (!pgdat
->node_spanned_pages
)
6843 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6844 offset
= pgdat
->node_start_pfn
- start
;
6845 /* ia64 gets its own node_mem_map, before this, without bootmem */
6846 if (!pgdat
->node_mem_map
) {
6847 unsigned long size
, end
;
6851 * The zone's endpoints aren't required to be MAX_ORDER
6852 * aligned but the node_mem_map endpoints must be in order
6853 * for the buddy allocator to function correctly.
6855 end
= pgdat_end_pfn(pgdat
);
6856 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6857 size
= (end
- start
) * sizeof(struct page
);
6858 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
6861 panic("Failed to allocate %ld bytes for node %d memory map\n",
6862 size
, pgdat
->node_id
);
6863 pgdat
->node_mem_map
= map
+ offset
;
6865 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6866 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6867 (unsigned long)pgdat
->node_mem_map
);
6868 #ifndef CONFIG_NEED_MULTIPLE_NODES
6870 * With no DISCONTIG, the global mem_map is just set as node 0's
6872 if (pgdat
== NODE_DATA(0)) {
6873 mem_map
= NODE_DATA(0)->node_mem_map
;
6874 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6875 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6877 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6882 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6883 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6885 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6886 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6888 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6891 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6894 void __init
free_area_init_node(int nid
, unsigned long *zones_size
,
6895 unsigned long node_start_pfn
,
6896 unsigned long *zholes_size
)
6898 pg_data_t
*pgdat
= NODE_DATA(nid
);
6899 unsigned long start_pfn
= 0;
6900 unsigned long end_pfn
= 0;
6902 /* pg_data_t should be reset to zero when it's allocated */
6903 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6905 pgdat
->node_id
= nid
;
6906 pgdat
->node_start_pfn
= node_start_pfn
;
6907 pgdat
->per_cpu_nodestats
= NULL
;
6908 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6909 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6910 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6911 (u64
)start_pfn
<< PAGE_SHIFT
,
6912 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6914 start_pfn
= node_start_pfn
;
6916 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6917 zones_size
, zholes_size
);
6919 alloc_node_mem_map(pgdat
);
6920 pgdat_set_deferred_range(pgdat
);
6922 free_area_init_core(pgdat
);
6925 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6927 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6930 static u64
zero_pfn_range(unsigned long spfn
, unsigned long epfn
)
6935 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6936 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6937 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6938 + pageblock_nr_pages
- 1;
6941 mm_zero_struct_page(pfn_to_page(pfn
));
6949 * Only struct pages that are backed by physical memory are zeroed and
6950 * initialized by going through __init_single_page(). But, there are some
6951 * struct pages which are reserved in memblock allocator and their fields
6952 * may be accessed (for example page_to_pfn() on some configuration accesses
6953 * flags). We must explicitly zero those struct pages.
6955 * This function also addresses a similar issue where struct pages are left
6956 * uninitialized because the physical address range is not covered by
6957 * memblock.memory or memblock.reserved. That could happen when memblock
6958 * layout is manually configured via memmap=, or when the highest physical
6959 * address (max_pfn) does not end on a section boundary.
6961 void __init
zero_resv_unavail(void)
6963 phys_addr_t start
, end
;
6965 phys_addr_t next
= 0;
6968 * Loop through unavailable ranges not covered by memblock.memory.
6971 for_each_mem_range(i
, &memblock
.memory
, NULL
,
6972 NUMA_NO_NODE
, MEMBLOCK_NONE
, &start
, &end
, NULL
) {
6974 pgcnt
+= zero_pfn_range(PFN_DOWN(next
), PFN_UP(start
));
6979 * Early sections always have a fully populated memmap for the whole
6980 * section - see pfn_valid(). If the last section has holes at the
6981 * end and that section is marked "online", the memmap will be
6982 * considered initialized. Make sure that memmap has a well defined
6985 pgcnt
+= zero_pfn_range(PFN_DOWN(next
),
6986 round_up(max_pfn
, PAGES_PER_SECTION
));
6989 * Struct pages that do not have backing memory. This could be because
6990 * firmware is using some of this memory, or for some other reasons.
6993 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
6995 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6997 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6999 #if MAX_NUMNODES > 1
7001 * Figure out the number of possible node ids.
7003 void __init
setup_nr_node_ids(void)
7005 unsigned int highest
;
7007 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
7008 nr_node_ids
= highest
+ 1;
7013 * node_map_pfn_alignment - determine the maximum internode alignment
7015 * This function should be called after node map is populated and sorted.
7016 * It calculates the maximum power of two alignment which can distinguish
7019 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7020 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7021 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7022 * shifted, 1GiB is enough and this function will indicate so.
7024 * This is used to test whether pfn -> nid mapping of the chosen memory
7025 * model has fine enough granularity to avoid incorrect mapping for the
7026 * populated node map.
7028 * Return: the determined alignment in pfn's. 0 if there is no alignment
7029 * requirement (single node).
7031 unsigned long __init
node_map_pfn_alignment(void)
7033 unsigned long accl_mask
= 0, last_end
= 0;
7034 unsigned long start
, end
, mask
;
7035 int last_nid
= NUMA_NO_NODE
;
7038 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
7039 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
7046 * Start with a mask granular enough to pin-point to the
7047 * start pfn and tick off bits one-by-one until it becomes
7048 * too coarse to separate the current node from the last.
7050 mask
= ~((1 << __ffs(start
)) - 1);
7051 while (mask
&& last_end
<= (start
& (mask
<< 1)))
7054 /* accumulate all internode masks */
7058 /* convert mask to number of pages */
7059 return ~accl_mask
+ 1;
7062 /* Find the lowest pfn for a node */
7063 static unsigned long __init
find_min_pfn_for_node(int nid
)
7065 unsigned long min_pfn
= ULONG_MAX
;
7066 unsigned long start_pfn
;
7069 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
7070 min_pfn
= min(min_pfn
, start_pfn
);
7072 if (min_pfn
== ULONG_MAX
) {
7073 pr_warn("Could not find start_pfn for node %d\n", nid
);
7081 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7083 * Return: the minimum PFN based on information provided via
7084 * memblock_set_node().
7086 unsigned long __init
find_min_pfn_with_active_regions(void)
7088 return find_min_pfn_for_node(MAX_NUMNODES
);
7092 * early_calculate_totalpages()
7093 * Sum pages in active regions for movable zone.
7094 * Populate N_MEMORY for calculating usable_nodes.
7096 static unsigned long __init
early_calculate_totalpages(void)
7098 unsigned long totalpages
= 0;
7099 unsigned long start_pfn
, end_pfn
;
7102 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7103 unsigned long pages
= end_pfn
- start_pfn
;
7105 totalpages
+= pages
;
7107 node_set_state(nid
, N_MEMORY
);
7113 * Find the PFN the Movable zone begins in each node. Kernel memory
7114 * is spread evenly between nodes as long as the nodes have enough
7115 * memory. When they don't, some nodes will have more kernelcore than
7118 static void __init
find_zone_movable_pfns_for_nodes(void)
7121 unsigned long usable_startpfn
;
7122 unsigned long kernelcore_node
, kernelcore_remaining
;
7123 /* save the state before borrow the nodemask */
7124 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
7125 unsigned long totalpages
= early_calculate_totalpages();
7126 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
7127 struct memblock_region
*r
;
7129 /* Need to find movable_zone earlier when movable_node is specified. */
7130 find_usable_zone_for_movable();
7133 * If movable_node is specified, ignore kernelcore and movablecore
7136 if (movable_node_is_enabled()) {
7137 for_each_memblock(memory
, r
) {
7138 if (!memblock_is_hotpluggable(r
))
7143 usable_startpfn
= PFN_DOWN(r
->base
);
7144 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7145 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7153 * If kernelcore=mirror is specified, ignore movablecore option
7155 if (mirrored_kernelcore
) {
7156 bool mem_below_4gb_not_mirrored
= false;
7158 for_each_memblock(memory
, r
) {
7159 if (memblock_is_mirror(r
))
7164 usable_startpfn
= memblock_region_memory_base_pfn(r
);
7166 if (usable_startpfn
< 0x100000) {
7167 mem_below_4gb_not_mirrored
= true;
7171 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7172 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7176 if (mem_below_4gb_not_mirrored
)
7177 pr_warn("This configuration results in unmirrored kernel memory.");
7183 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7184 * amount of necessary memory.
7186 if (required_kernelcore_percent
)
7187 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
7189 if (required_movablecore_percent
)
7190 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
7194 * If movablecore= was specified, calculate what size of
7195 * kernelcore that corresponds so that memory usable for
7196 * any allocation type is evenly spread. If both kernelcore
7197 * and movablecore are specified, then the value of kernelcore
7198 * will be used for required_kernelcore if it's greater than
7199 * what movablecore would have allowed.
7201 if (required_movablecore
) {
7202 unsigned long corepages
;
7205 * Round-up so that ZONE_MOVABLE is at least as large as what
7206 * was requested by the user
7208 required_movablecore
=
7209 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
7210 required_movablecore
= min(totalpages
, required_movablecore
);
7211 corepages
= totalpages
- required_movablecore
;
7213 required_kernelcore
= max(required_kernelcore
, corepages
);
7217 * If kernelcore was not specified or kernelcore size is larger
7218 * than totalpages, there is no ZONE_MOVABLE.
7220 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
7223 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7224 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
7227 /* Spread kernelcore memory as evenly as possible throughout nodes */
7228 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7229 for_each_node_state(nid
, N_MEMORY
) {
7230 unsigned long start_pfn
, end_pfn
;
7233 * Recalculate kernelcore_node if the division per node
7234 * now exceeds what is necessary to satisfy the requested
7235 * amount of memory for the kernel
7237 if (required_kernelcore
< kernelcore_node
)
7238 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7241 * As the map is walked, we track how much memory is usable
7242 * by the kernel using kernelcore_remaining. When it is
7243 * 0, the rest of the node is usable by ZONE_MOVABLE
7245 kernelcore_remaining
= kernelcore_node
;
7247 /* Go through each range of PFNs within this node */
7248 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7249 unsigned long size_pages
;
7251 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
7252 if (start_pfn
>= end_pfn
)
7255 /* Account for what is only usable for kernelcore */
7256 if (start_pfn
< usable_startpfn
) {
7257 unsigned long kernel_pages
;
7258 kernel_pages
= min(end_pfn
, usable_startpfn
)
7261 kernelcore_remaining
-= min(kernel_pages
,
7262 kernelcore_remaining
);
7263 required_kernelcore
-= min(kernel_pages
,
7264 required_kernelcore
);
7266 /* Continue if range is now fully accounted */
7267 if (end_pfn
<= usable_startpfn
) {
7270 * Push zone_movable_pfn to the end so
7271 * that if we have to rebalance
7272 * kernelcore across nodes, we will
7273 * not double account here
7275 zone_movable_pfn
[nid
] = end_pfn
;
7278 start_pfn
= usable_startpfn
;
7282 * The usable PFN range for ZONE_MOVABLE is from
7283 * start_pfn->end_pfn. Calculate size_pages as the
7284 * number of pages used as kernelcore
7286 size_pages
= end_pfn
- start_pfn
;
7287 if (size_pages
> kernelcore_remaining
)
7288 size_pages
= kernelcore_remaining
;
7289 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7292 * Some kernelcore has been met, update counts and
7293 * break if the kernelcore for this node has been
7296 required_kernelcore
-= min(required_kernelcore
,
7298 kernelcore_remaining
-= size_pages
;
7299 if (!kernelcore_remaining
)
7305 * If there is still required_kernelcore, we do another pass with one
7306 * less node in the count. This will push zone_movable_pfn[nid] further
7307 * along on the nodes that still have memory until kernelcore is
7311 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7315 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7316 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++) {
7317 unsigned long start_pfn
, end_pfn
;
7319 zone_movable_pfn
[nid
] =
7320 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7322 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
7323 if (zone_movable_pfn
[nid
] >= end_pfn
)
7324 zone_movable_pfn
[nid
] = 0;
7328 /* restore the node_state */
7329 node_states
[N_MEMORY
] = saved_node_state
;
7332 /* Any regular or high memory on that node ? */
7333 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7335 enum zone_type zone_type
;
7337 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7338 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7339 if (populated_zone(zone
)) {
7340 if (IS_ENABLED(CONFIG_HIGHMEM
))
7341 node_set_state(nid
, N_HIGH_MEMORY
);
7342 if (zone_type
<= ZONE_NORMAL
)
7343 node_set_state(nid
, N_NORMAL_MEMORY
);
7350 * free_area_init_nodes - Initialise all pg_data_t and zone data
7351 * @max_zone_pfn: an array of max PFNs for each zone
7353 * This will call free_area_init_node() for each active node in the system.
7354 * Using the page ranges provided by memblock_set_node(), the size of each
7355 * zone in each node and their holes is calculated. If the maximum PFN
7356 * between two adjacent zones match, it is assumed that the zone is empty.
7357 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7358 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7359 * starts where the previous one ended. For example, ZONE_DMA32 starts
7360 * at arch_max_dma_pfn.
7362 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
7364 unsigned long start_pfn
, end_pfn
;
7367 /* Record where the zone boundaries are */
7368 memset(arch_zone_lowest_possible_pfn
, 0,
7369 sizeof(arch_zone_lowest_possible_pfn
));
7370 memset(arch_zone_highest_possible_pfn
, 0,
7371 sizeof(arch_zone_highest_possible_pfn
));
7373 start_pfn
= find_min_pfn_with_active_regions();
7375 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7376 if (i
== ZONE_MOVABLE
)
7379 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
7380 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
7381 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
7383 start_pfn
= end_pfn
;
7386 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7387 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7388 find_zone_movable_pfns_for_nodes();
7390 /* Print out the zone ranges */
7391 pr_info("Zone ranges:\n");
7392 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7393 if (i
== ZONE_MOVABLE
)
7395 pr_info(" %-8s ", zone_names
[i
]);
7396 if (arch_zone_lowest_possible_pfn
[i
] ==
7397 arch_zone_highest_possible_pfn
[i
])
7400 pr_cont("[mem %#018Lx-%#018Lx]\n",
7401 (u64
)arch_zone_lowest_possible_pfn
[i
]
7403 ((u64
)arch_zone_highest_possible_pfn
[i
]
7404 << PAGE_SHIFT
) - 1);
7407 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7408 pr_info("Movable zone start for each node\n");
7409 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7410 if (zone_movable_pfn
[i
])
7411 pr_info(" Node %d: %#018Lx\n", i
,
7412 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7416 * Print out the early node map, and initialize the
7417 * subsection-map relative to active online memory ranges to
7418 * enable future "sub-section" extensions of the memory map.
7420 pr_info("Early memory node ranges\n");
7421 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7422 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7423 (u64
)start_pfn
<< PAGE_SHIFT
,
7424 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7425 subsection_map_init(start_pfn
, end_pfn
- start_pfn
);
7428 /* Initialise every node */
7429 mminit_verify_pageflags_layout();
7430 setup_nr_node_ids();
7431 zero_resv_unavail();
7432 for_each_online_node(nid
) {
7433 pg_data_t
*pgdat
= NODE_DATA(nid
);
7434 free_area_init_node(nid
, NULL
,
7435 find_min_pfn_for_node(nid
), NULL
);
7437 /* Any memory on that node */
7438 if (pgdat
->node_present_pages
)
7439 node_set_state(nid
, N_MEMORY
);
7440 check_for_memory(pgdat
, nid
);
7444 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7445 unsigned long *percent
)
7447 unsigned long long coremem
;
7453 /* Value may be a percentage of total memory, otherwise bytes */
7454 coremem
= simple_strtoull(p
, &endptr
, 0);
7455 if (*endptr
== '%') {
7456 /* Paranoid check for percent values greater than 100 */
7457 WARN_ON(coremem
> 100);
7461 coremem
= memparse(p
, &p
);
7462 /* Paranoid check that UL is enough for the coremem value */
7463 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7465 *core
= coremem
>> PAGE_SHIFT
;
7472 * kernelcore=size sets the amount of memory for use for allocations that
7473 * cannot be reclaimed or migrated.
7475 static int __init
cmdline_parse_kernelcore(char *p
)
7477 /* parse kernelcore=mirror */
7478 if (parse_option_str(p
, "mirror")) {
7479 mirrored_kernelcore
= true;
7483 return cmdline_parse_core(p
, &required_kernelcore
,
7484 &required_kernelcore_percent
);
7488 * movablecore=size sets the amount of memory for use for allocations that
7489 * can be reclaimed or migrated.
7491 static int __init
cmdline_parse_movablecore(char *p
)
7493 return cmdline_parse_core(p
, &required_movablecore
,
7494 &required_movablecore_percent
);
7497 early_param("kernelcore", cmdline_parse_kernelcore
);
7498 early_param("movablecore", cmdline_parse_movablecore
);
7500 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7502 void adjust_managed_page_count(struct page
*page
, long count
)
7504 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7505 totalram_pages_add(count
);
7506 #ifdef CONFIG_HIGHMEM
7507 if (PageHighMem(page
))
7508 totalhigh_pages_add(count
);
7511 EXPORT_SYMBOL(adjust_managed_page_count
);
7513 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7516 unsigned long pages
= 0;
7518 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7519 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7520 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7521 struct page
*page
= virt_to_page(pos
);
7522 void *direct_map_addr
;
7525 * 'direct_map_addr' might be different from 'pos'
7526 * because some architectures' virt_to_page()
7527 * work with aliases. Getting the direct map
7528 * address ensures that we get a _writeable_
7529 * alias for the memset().
7531 direct_map_addr
= page_address(page
);
7532 if ((unsigned int)poison
<= 0xFF)
7533 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7535 free_reserved_page(page
);
7539 pr_info("Freeing %s memory: %ldK\n",
7540 s
, pages
<< (PAGE_SHIFT
- 10));
7545 #ifdef CONFIG_HIGHMEM
7546 void free_highmem_page(struct page
*page
)
7548 __free_reserved_page(page
);
7549 totalram_pages_inc();
7550 atomic_long_inc(&page_zone(page
)->managed_pages
);
7551 totalhigh_pages_inc();
7556 void __init
mem_init_print_info(const char *str
)
7558 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7559 unsigned long init_code_size
, init_data_size
;
7561 physpages
= get_num_physpages();
7562 codesize
= _etext
- _stext
;
7563 datasize
= _edata
- _sdata
;
7564 rosize
= __end_rodata
- __start_rodata
;
7565 bss_size
= __bss_stop
- __bss_start
;
7566 init_data_size
= __init_end
- __init_begin
;
7567 init_code_size
= _einittext
- _sinittext
;
7570 * Detect special cases and adjust section sizes accordingly:
7571 * 1) .init.* may be embedded into .data sections
7572 * 2) .init.text.* may be out of [__init_begin, __init_end],
7573 * please refer to arch/tile/kernel/vmlinux.lds.S.
7574 * 3) .rodata.* may be embedded into .text or .data sections.
7576 #define adj_init_size(start, end, size, pos, adj) \
7578 if (start <= pos && pos < end && size > adj) \
7582 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7583 _sinittext
, init_code_size
);
7584 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7585 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7586 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7587 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7589 #undef adj_init_size
7591 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7592 #ifdef CONFIG_HIGHMEM
7596 nr_free_pages() << (PAGE_SHIFT
- 10),
7597 physpages
<< (PAGE_SHIFT
- 10),
7598 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7599 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7600 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7601 totalcma_pages
<< (PAGE_SHIFT
- 10),
7602 #ifdef CONFIG_HIGHMEM
7603 totalhigh_pages() << (PAGE_SHIFT
- 10),
7605 str
? ", " : "", str
? str
: "");
7609 * set_dma_reserve - set the specified number of pages reserved in the first zone
7610 * @new_dma_reserve: The number of pages to mark reserved
7612 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7613 * In the DMA zone, a significant percentage may be consumed by kernel image
7614 * and other unfreeable allocations which can skew the watermarks badly. This
7615 * function may optionally be used to account for unfreeable pages in the
7616 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7617 * smaller per-cpu batchsize.
7619 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7621 dma_reserve
= new_dma_reserve
;
7624 void __init
free_area_init(unsigned long *zones_size
)
7626 zero_resv_unavail();
7627 free_area_init_node(0, zones_size
,
7628 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
7631 static int page_alloc_cpu_dead(unsigned int cpu
)
7634 lru_add_drain_cpu(cpu
);
7638 * Spill the event counters of the dead processor
7639 * into the current processors event counters.
7640 * This artificially elevates the count of the current
7643 vm_events_fold_cpu(cpu
);
7646 * Zero the differential counters of the dead processor
7647 * so that the vm statistics are consistent.
7649 * This is only okay since the processor is dead and cannot
7650 * race with what we are doing.
7652 cpu_vm_stats_fold(cpu
);
7657 int hashdist
= HASHDIST_DEFAULT
;
7659 static int __init
set_hashdist(char *str
)
7663 hashdist
= simple_strtoul(str
, &str
, 0);
7666 __setup("hashdist=", set_hashdist
);
7669 void __init
page_alloc_init(void)
7674 if (num_node_state(N_MEMORY
) == 1)
7678 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7679 "mm/page_alloc:dead", NULL
,
7680 page_alloc_cpu_dead
);
7685 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7686 * or min_free_kbytes changes.
7688 static void calculate_totalreserve_pages(void)
7690 struct pglist_data
*pgdat
;
7691 unsigned long reserve_pages
= 0;
7692 enum zone_type i
, j
;
7694 for_each_online_pgdat(pgdat
) {
7696 pgdat
->totalreserve_pages
= 0;
7698 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7699 struct zone
*zone
= pgdat
->node_zones
+ i
;
7701 unsigned long managed_pages
= zone_managed_pages(zone
);
7703 /* Find valid and maximum lowmem_reserve in the zone */
7704 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7705 if (zone
->lowmem_reserve
[j
] > max
)
7706 max
= zone
->lowmem_reserve
[j
];
7709 /* we treat the high watermark as reserved pages. */
7710 max
+= high_wmark_pages(zone
);
7712 if (max
> managed_pages
)
7713 max
= managed_pages
;
7715 pgdat
->totalreserve_pages
+= max
;
7717 reserve_pages
+= max
;
7720 totalreserve_pages
= reserve_pages
;
7724 * setup_per_zone_lowmem_reserve - called whenever
7725 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7726 * has a correct pages reserved value, so an adequate number of
7727 * pages are left in the zone after a successful __alloc_pages().
7729 static void setup_per_zone_lowmem_reserve(void)
7731 struct pglist_data
*pgdat
;
7732 enum zone_type j
, idx
;
7734 for_each_online_pgdat(pgdat
) {
7735 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7736 struct zone
*zone
= pgdat
->node_zones
+ j
;
7737 unsigned long managed_pages
= zone_managed_pages(zone
);
7739 zone
->lowmem_reserve
[j
] = 0;
7743 struct zone
*lower_zone
;
7746 lower_zone
= pgdat
->node_zones
+ idx
;
7748 if (sysctl_lowmem_reserve_ratio
[idx
] < 1) {
7749 sysctl_lowmem_reserve_ratio
[idx
] = 0;
7750 lower_zone
->lowmem_reserve
[j
] = 0;
7752 lower_zone
->lowmem_reserve
[j
] =
7753 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7755 managed_pages
+= zone_managed_pages(lower_zone
);
7760 /* update totalreserve_pages */
7761 calculate_totalreserve_pages();
7764 static void __setup_per_zone_wmarks(void)
7766 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7767 unsigned long lowmem_pages
= 0;
7769 unsigned long flags
;
7771 /* Calculate total number of !ZONE_HIGHMEM pages */
7772 for_each_zone(zone
) {
7773 if (!is_highmem(zone
))
7774 lowmem_pages
+= zone_managed_pages(zone
);
7777 for_each_zone(zone
) {
7780 spin_lock_irqsave(&zone
->lock
, flags
);
7781 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7782 do_div(tmp
, lowmem_pages
);
7783 if (is_highmem(zone
)) {
7785 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7786 * need highmem pages, so cap pages_min to a small
7789 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7790 * deltas control async page reclaim, and so should
7791 * not be capped for highmem.
7793 unsigned long min_pages
;
7795 min_pages
= zone_managed_pages(zone
) / 1024;
7796 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7797 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7800 * If it's a lowmem zone, reserve a number of pages
7801 * proportionate to the zone's size.
7803 zone
->_watermark
[WMARK_MIN
] = tmp
;
7807 * Set the kswapd watermarks distance according to the
7808 * scale factor in proportion to available memory, but
7809 * ensure a minimum size on small systems.
7811 tmp
= max_t(u64
, tmp
>> 2,
7812 mult_frac(zone_managed_pages(zone
),
7813 watermark_scale_factor
, 10000));
7815 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7816 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7817 zone
->watermark_boost
= 0;
7819 spin_unlock_irqrestore(&zone
->lock
, flags
);
7822 /* update totalreserve_pages */
7823 calculate_totalreserve_pages();
7827 * setup_per_zone_wmarks - called when min_free_kbytes changes
7828 * or when memory is hot-{added|removed}
7830 * Ensures that the watermark[min,low,high] values for each zone are set
7831 * correctly with respect to min_free_kbytes.
7833 void setup_per_zone_wmarks(void)
7835 static DEFINE_SPINLOCK(lock
);
7838 __setup_per_zone_wmarks();
7843 * Initialise min_free_kbytes.
7845 * For small machines we want it small (128k min). For large machines
7846 * we want it large (64MB max). But it is not linear, because network
7847 * bandwidth does not increase linearly with machine size. We use
7849 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7850 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7866 int __meminit
init_per_zone_wmark_min(void)
7868 unsigned long lowmem_kbytes
;
7869 int new_min_free_kbytes
;
7871 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7872 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7874 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7875 min_free_kbytes
= new_min_free_kbytes
;
7876 if (min_free_kbytes
< 128)
7877 min_free_kbytes
= 128;
7878 if (min_free_kbytes
> 65536)
7879 min_free_kbytes
= 65536;
7881 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7882 new_min_free_kbytes
, user_min_free_kbytes
);
7884 setup_per_zone_wmarks();
7885 refresh_zone_stat_thresholds();
7886 setup_per_zone_lowmem_reserve();
7889 setup_min_unmapped_ratio();
7890 setup_min_slab_ratio();
7893 khugepaged_min_free_kbytes_update();
7897 postcore_initcall(init_per_zone_wmark_min
)
7900 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7901 * that we can call two helper functions whenever min_free_kbytes
7904 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7905 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7909 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7914 user_min_free_kbytes
= min_free_kbytes
;
7915 setup_per_zone_wmarks();
7920 int watermark_boost_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7921 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7925 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7932 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7933 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7937 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7942 setup_per_zone_wmarks();
7948 static void setup_min_unmapped_ratio(void)
7953 for_each_online_pgdat(pgdat
)
7954 pgdat
->min_unmapped_pages
= 0;
7957 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
7958 sysctl_min_unmapped_ratio
) / 100;
7962 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7963 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7967 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7971 setup_min_unmapped_ratio();
7976 static void setup_min_slab_ratio(void)
7981 for_each_online_pgdat(pgdat
)
7982 pgdat
->min_slab_pages
= 0;
7985 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
7986 sysctl_min_slab_ratio
) / 100;
7989 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7990 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7994 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7998 setup_min_slab_ratio();
8005 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8006 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8007 * whenever sysctl_lowmem_reserve_ratio changes.
8009 * The reserve ratio obviously has absolutely no relation with the
8010 * minimum watermarks. The lowmem reserve ratio can only make sense
8011 * if in function of the boot time zone sizes.
8013 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8014 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
8016 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8017 setup_per_zone_lowmem_reserve();
8022 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8023 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8024 * pagelist can have before it gets flushed back to buddy allocator.
8026 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
8027 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
8030 int old_percpu_pagelist_fraction
;
8033 mutex_lock(&pcp_batch_high_lock
);
8034 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
8036 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8037 if (!write
|| ret
< 0)
8040 /* Sanity checking to avoid pcp imbalance */
8041 if (percpu_pagelist_fraction
&&
8042 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
8043 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
8049 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
8052 for_each_populated_zone(zone
) {
8055 for_each_possible_cpu(cpu
)
8056 pageset_set_high_and_batch(zone
,
8057 per_cpu_ptr(zone
->pageset
, cpu
));
8060 mutex_unlock(&pcp_batch_high_lock
);
8064 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8066 * Returns the number of pages that arch has reserved but
8067 * is not known to alloc_large_system_hash().
8069 static unsigned long __init
arch_reserved_kernel_pages(void)
8076 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8077 * machines. As memory size is increased the scale is also increased but at
8078 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8079 * quadruples the scale is increased by one, which means the size of hash table
8080 * only doubles, instead of quadrupling as well.
8081 * Because 32-bit systems cannot have large physical memory, where this scaling
8082 * makes sense, it is disabled on such platforms.
8084 #if __BITS_PER_LONG > 32
8085 #define ADAPT_SCALE_BASE (64ul << 30)
8086 #define ADAPT_SCALE_SHIFT 2
8087 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8091 * allocate a large system hash table from bootmem
8092 * - it is assumed that the hash table must contain an exact power-of-2
8093 * quantity of entries
8094 * - limit is the number of hash buckets, not the total allocation size
8096 void *__init
alloc_large_system_hash(const char *tablename
,
8097 unsigned long bucketsize
,
8098 unsigned long numentries
,
8101 unsigned int *_hash_shift
,
8102 unsigned int *_hash_mask
,
8103 unsigned long low_limit
,
8104 unsigned long high_limit
)
8106 unsigned long long max
= high_limit
;
8107 unsigned long log2qty
, size
;
8112 /* allow the kernel cmdline to have a say */
8114 /* round applicable memory size up to nearest megabyte */
8115 numentries
= nr_kernel_pages
;
8116 numentries
-= arch_reserved_kernel_pages();
8118 /* It isn't necessary when PAGE_SIZE >= 1MB */
8119 if (PAGE_SHIFT
< 20)
8120 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
8122 #if __BITS_PER_LONG > 32
8124 unsigned long adapt
;
8126 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
8127 adapt
<<= ADAPT_SCALE_SHIFT
)
8132 /* limit to 1 bucket per 2^scale bytes of low memory */
8133 if (scale
> PAGE_SHIFT
)
8134 numentries
>>= (scale
- PAGE_SHIFT
);
8136 numentries
<<= (PAGE_SHIFT
- scale
);
8138 /* Make sure we've got at least a 0-order allocation.. */
8139 if (unlikely(flags
& HASH_SMALL
)) {
8140 /* Makes no sense without HASH_EARLY */
8141 WARN_ON(!(flags
& HASH_EARLY
));
8142 if (!(numentries
>> *_hash_shift
)) {
8143 numentries
= 1UL << *_hash_shift
;
8144 BUG_ON(!numentries
);
8146 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
8147 numentries
= PAGE_SIZE
/ bucketsize
;
8149 numentries
= roundup_pow_of_two(numentries
);
8151 /* limit allocation size to 1/16 total memory by default */
8153 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
8154 do_div(max
, bucketsize
);
8156 max
= min(max
, 0x80000000ULL
);
8158 if (numentries
< low_limit
)
8159 numentries
= low_limit
;
8160 if (numentries
> max
)
8163 log2qty
= ilog2(numentries
);
8165 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
8168 size
= bucketsize
<< log2qty
;
8169 if (flags
& HASH_EARLY
) {
8170 if (flags
& HASH_ZERO
)
8171 table
= memblock_alloc(size
, SMP_CACHE_BYTES
);
8173 table
= memblock_alloc_raw(size
,
8175 } else if (get_order(size
) >= MAX_ORDER
|| hashdist
) {
8176 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
8180 * If bucketsize is not a power-of-two, we may free
8181 * some pages at the end of hash table which
8182 * alloc_pages_exact() automatically does
8184 table
= alloc_pages_exact(size
, gfp_flags
);
8185 kmemleak_alloc(table
, size
, 1, gfp_flags
);
8187 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
8190 panic("Failed to allocate %s hash table\n", tablename
);
8192 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8193 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
,
8194 virt
? "vmalloc" : "linear");
8197 *_hash_shift
= log2qty
;
8199 *_hash_mask
= (1 << log2qty
) - 1;
8205 * This function checks whether pageblock includes unmovable pages or not.
8206 * If @count is not zero, it is okay to include less @count unmovable pages
8208 * PageLRU check without isolation or lru_lock could race so that
8209 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8210 * check without lock_page also may miss some movable non-lru pages at
8211 * race condition. So you can't expect this function should be exact.
8213 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
8214 int migratetype
, int flags
)
8216 unsigned long found
;
8217 unsigned long iter
= 0;
8218 unsigned long pfn
= page_to_pfn(page
);
8219 const char *reason
= "unmovable page";
8222 * TODO we could make this much more efficient by not checking every
8223 * page in the range if we know all of them are in MOVABLE_ZONE and
8224 * that the movable zone guarantees that pages are migratable but
8225 * the later is not the case right now unfortunatelly. E.g. movablecore
8226 * can still lead to having bootmem allocations in zone_movable.
8229 if (is_migrate_cma_page(page
)) {
8231 * CMA allocations (alloc_contig_range) really need to mark
8232 * isolate CMA pageblocks even when they are not movable in fact
8233 * so consider them movable here.
8235 if (is_migrate_cma(migratetype
))
8238 reason
= "CMA page";
8242 for (found
= 0; iter
< pageblock_nr_pages
; iter
++) {
8243 unsigned long check
= pfn
+ iter
;
8245 if (!pfn_valid_within(check
))
8248 page
= pfn_to_page(check
);
8250 if (PageReserved(page
))
8254 * If the zone is movable and we have ruled out all reserved
8255 * pages then it should be reasonably safe to assume the rest
8258 if (zone_idx(zone
) == ZONE_MOVABLE
)
8262 * Hugepages are not in LRU lists, but they're movable.
8263 * We need not scan over tail pages because we don't
8264 * handle each tail page individually in migration.
8266 if (PageHuge(page
)) {
8267 struct page
*head
= compound_head(page
);
8268 unsigned int skip_pages
;
8270 if (!hugepage_migration_supported(page_hstate(head
)))
8273 skip_pages
= compound_nr(head
) - (page
- head
);
8274 iter
+= skip_pages
- 1;
8279 * We can't use page_count without pin a page
8280 * because another CPU can free compound page.
8281 * This check already skips compound tails of THP
8282 * because their page->_refcount is zero at all time.
8284 if (!page_ref_count(page
)) {
8285 if (PageBuddy(page
))
8286 iter
+= (1 << page_order(page
)) - 1;
8291 * The HWPoisoned page may be not in buddy system, and
8292 * page_count() is not 0.
8294 if ((flags
& SKIP_HWPOISON
) && PageHWPoison(page
))
8297 if (__PageMovable(page
))
8303 * If there are RECLAIMABLE pages, we need to check
8304 * it. But now, memory offline itself doesn't call
8305 * shrink_node_slabs() and it still to be fixed.
8308 * If the page is not RAM, page_count()should be 0.
8309 * we don't need more check. This is an _used_ not-movable page.
8311 * The problematic thing here is PG_reserved pages. PG_reserved
8312 * is set to both of a memory hole page and a _used_ kernel
8320 WARN_ON_ONCE(zone_idx(zone
) == ZONE_MOVABLE
);
8321 if (flags
& REPORT_FAILURE
)
8322 dump_page(pfn_to_page(pfn
+ iter
), reason
);
8326 #ifdef CONFIG_CONTIG_ALLOC
8327 static unsigned long pfn_max_align_down(unsigned long pfn
)
8329 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8330 pageblock_nr_pages
) - 1);
8333 static unsigned long pfn_max_align_up(unsigned long pfn
)
8335 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8336 pageblock_nr_pages
));
8339 /* [start, end) must belong to a single zone. */
8340 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8341 unsigned long start
, unsigned long end
)
8343 /* This function is based on compact_zone() from compaction.c. */
8344 unsigned long nr_reclaimed
;
8345 unsigned long pfn
= start
;
8346 unsigned int tries
= 0;
8351 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8352 if (fatal_signal_pending(current
)) {
8357 if (list_empty(&cc
->migratepages
)) {
8358 cc
->nr_migratepages
= 0;
8359 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8365 } else if (++tries
== 5) {
8366 ret
= ret
< 0 ? ret
: -EBUSY
;
8370 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8372 cc
->nr_migratepages
-= nr_reclaimed
;
8374 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
8375 NULL
, 0, cc
->mode
, MR_CONTIG_RANGE
);
8378 putback_movable_pages(&cc
->migratepages
);
8385 * alloc_contig_range() -- tries to allocate given range of pages
8386 * @start: start PFN to allocate
8387 * @end: one-past-the-last PFN to allocate
8388 * @migratetype: migratetype of the underlaying pageblocks (either
8389 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8390 * in range must have the same migratetype and it must
8391 * be either of the two.
8392 * @gfp_mask: GFP mask to use during compaction
8394 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8395 * aligned. The PFN range must belong to a single zone.
8397 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8398 * pageblocks in the range. Once isolated, the pageblocks should not
8399 * be modified by others.
8401 * Return: zero on success or negative error code. On success all
8402 * pages which PFN is in [start, end) are allocated for the caller and
8403 * need to be freed with free_contig_range().
8405 int alloc_contig_range(unsigned long start
, unsigned long end
,
8406 unsigned migratetype
, gfp_t gfp_mask
)
8408 unsigned long outer_start
, outer_end
;
8412 struct compact_control cc
= {
8413 .nr_migratepages
= 0,
8415 .zone
= page_zone(pfn_to_page(start
)),
8416 .mode
= MIGRATE_SYNC
,
8417 .ignore_skip_hint
= true,
8418 .no_set_skip_hint
= true,
8419 .gfp_mask
= current_gfp_context(gfp_mask
),
8421 INIT_LIST_HEAD(&cc
.migratepages
);
8424 * What we do here is we mark all pageblocks in range as
8425 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8426 * have different sizes, and due to the way page allocator
8427 * work, we align the range to biggest of the two pages so
8428 * that page allocator won't try to merge buddies from
8429 * different pageblocks and change MIGRATE_ISOLATE to some
8430 * other migration type.
8432 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8433 * migrate the pages from an unaligned range (ie. pages that
8434 * we are interested in). This will put all the pages in
8435 * range back to page allocator as MIGRATE_ISOLATE.
8437 * When this is done, we take the pages in range from page
8438 * allocator removing them from the buddy system. This way
8439 * page allocator will never consider using them.
8441 * This lets us mark the pageblocks back as
8442 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8443 * aligned range but not in the unaligned, original range are
8444 * put back to page allocator so that buddy can use them.
8447 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8448 pfn_max_align_up(end
), migratetype
, 0);
8453 * In case of -EBUSY, we'd like to know which page causes problem.
8454 * So, just fall through. test_pages_isolated() has a tracepoint
8455 * which will report the busy page.
8457 * It is possible that busy pages could become available before
8458 * the call to test_pages_isolated, and the range will actually be
8459 * allocated. So, if we fall through be sure to clear ret so that
8460 * -EBUSY is not accidentally used or returned to caller.
8462 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8463 if (ret
&& ret
!= -EBUSY
)
8468 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8469 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8470 * more, all pages in [start, end) are free in page allocator.
8471 * What we are going to do is to allocate all pages from
8472 * [start, end) (that is remove them from page allocator).
8474 * The only problem is that pages at the beginning and at the
8475 * end of interesting range may be not aligned with pages that
8476 * page allocator holds, ie. they can be part of higher order
8477 * pages. Because of this, we reserve the bigger range and
8478 * once this is done free the pages we are not interested in.
8480 * We don't have to hold zone->lock here because the pages are
8481 * isolated thus they won't get removed from buddy.
8484 lru_add_drain_all();
8487 outer_start
= start
;
8488 while (!PageBuddy(pfn_to_page(outer_start
))) {
8489 if (++order
>= MAX_ORDER
) {
8490 outer_start
= start
;
8493 outer_start
&= ~0UL << order
;
8496 if (outer_start
!= start
) {
8497 order
= page_order(pfn_to_page(outer_start
));
8500 * outer_start page could be small order buddy page and
8501 * it doesn't include start page. Adjust outer_start
8502 * in this case to report failed page properly
8503 * on tracepoint in test_pages_isolated()
8505 if (outer_start
+ (1UL << order
) <= start
)
8506 outer_start
= start
;
8509 /* Make sure the range is really isolated. */
8510 if (test_pages_isolated(outer_start
, end
, false)) {
8511 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8512 __func__
, outer_start
, end
);
8517 /* Grab isolated pages from freelists. */
8518 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8524 /* Free head and tail (if any) */
8525 if (start
!= outer_start
)
8526 free_contig_range(outer_start
, start
- outer_start
);
8527 if (end
!= outer_end
)
8528 free_contig_range(end
, outer_end
- end
);
8531 undo_isolate_page_range(pfn_max_align_down(start
),
8532 pfn_max_align_up(end
), migratetype
);
8535 #endif /* CONFIG_CONTIG_ALLOC */
8537 void free_contig_range(unsigned long pfn
, unsigned int nr_pages
)
8539 unsigned int count
= 0;
8541 for (; nr_pages
--; pfn
++) {
8542 struct page
*page
= pfn_to_page(pfn
);
8544 count
+= page_count(page
) != 1;
8547 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8551 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8552 * page high values need to be recalulated.
8554 void __meminit
zone_pcp_update(struct zone
*zone
)
8557 mutex_lock(&pcp_batch_high_lock
);
8558 for_each_possible_cpu(cpu
)
8559 pageset_set_high_and_batch(zone
,
8560 per_cpu_ptr(zone
->pageset
, cpu
));
8561 mutex_unlock(&pcp_batch_high_lock
);
8564 void zone_pcp_reset(struct zone
*zone
)
8566 unsigned long flags
;
8568 struct per_cpu_pageset
*pset
;
8570 /* avoid races with drain_pages() */
8571 local_irq_save(flags
);
8572 if (zone
->pageset
!= &boot_pageset
) {
8573 for_each_online_cpu(cpu
) {
8574 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8575 drain_zonestat(zone
, pset
);
8577 free_percpu(zone
->pageset
);
8578 zone
->pageset
= &boot_pageset
;
8580 local_irq_restore(flags
);
8583 #ifdef CONFIG_MEMORY_HOTREMOVE
8585 * All pages in the range must be in a single zone and isolated
8586 * before calling this.
8589 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8593 unsigned int order
, i
;
8595 unsigned long flags
;
8596 unsigned long offlined_pages
= 0;
8598 /* find the first valid pfn */
8599 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
8603 return offlined_pages
;
8605 offline_mem_sections(pfn
, end_pfn
);
8606 zone
= page_zone(pfn_to_page(pfn
));
8607 spin_lock_irqsave(&zone
->lock
, flags
);
8609 while (pfn
< end_pfn
) {
8610 if (!pfn_valid(pfn
)) {
8614 page
= pfn_to_page(pfn
);
8616 * The HWPoisoned page may be not in buddy system, and
8617 * page_count() is not 0.
8619 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8621 SetPageReserved(page
);
8626 BUG_ON(page_count(page
));
8627 BUG_ON(!PageBuddy(page
));
8628 order
= page_order(page
);
8629 offlined_pages
+= 1 << order
;
8630 #ifdef CONFIG_DEBUG_VM
8631 pr_info("remove from free list %lx %d %lx\n",
8632 pfn
, 1 << order
, end_pfn
);
8634 del_page_from_free_area(page
, &zone
->free_area
[order
]);
8635 for (i
= 0; i
< (1 << order
); i
++)
8636 SetPageReserved((page
+i
));
8637 pfn
+= (1 << order
);
8639 spin_unlock_irqrestore(&zone
->lock
, flags
);
8641 return offlined_pages
;
8645 bool is_free_buddy_page(struct page
*page
)
8647 struct zone
*zone
= page_zone(page
);
8648 unsigned long pfn
= page_to_pfn(page
);
8649 unsigned long flags
;
8652 spin_lock_irqsave(&zone
->lock
, flags
);
8653 for (order
= 0; order
< MAX_ORDER
; order
++) {
8654 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8656 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8659 spin_unlock_irqrestore(&zone
->lock
, flags
);
8661 return order
< MAX_ORDER
;
8664 #ifdef CONFIG_MEMORY_FAILURE
8666 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8667 * test is performed under the zone lock to prevent a race against page
8670 bool set_hwpoison_free_buddy_page(struct page
*page
)
8672 struct zone
*zone
= page_zone(page
);
8673 unsigned long pfn
= page_to_pfn(page
);
8674 unsigned long flags
;
8676 bool hwpoisoned
= false;
8678 spin_lock_irqsave(&zone
->lock
, flags
);
8679 for (order
= 0; order
< MAX_ORDER
; order
++) {
8680 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8682 if (PageBuddy(page_head
) && page_order(page_head
) >= order
) {
8683 if (!TestSetPageHWPoison(page
))
8688 spin_unlock_irqrestore(&zone
->lock
, flags
);
8694 #ifdef CONFIG_ZONE_DMA
8695 bool has_managed_dma(void)
8697 struct pglist_data
*pgdat
;
8699 for_each_online_pgdat(pgdat
) {
8700 struct zone
*zone
= &pgdat
->node_zones
[ZONE_DMA
];
8702 if (managed_zone(zone
))
8707 #endif /* CONFIG_ZONE_DMA */