1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
71 #include <linux/padata.h>
72 #include <linux/khugepaged.h>
74 #include <asm/sections.h>
75 #include <asm/tlbflush.h>
76 #include <asm/div64.h>
79 #include "page_reporting.h"
81 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
82 typedef int __bitwise fpi_t
;
84 /* No special request */
85 #define FPI_NONE ((__force fpi_t)0)
88 * Skip free page reporting notification for the (possibly merged) page.
89 * This does not hinder free page reporting from grabbing the page,
90 * reporting it and marking it "reported" - it only skips notifying
91 * the free page reporting infrastructure about a newly freed page. For
92 * example, used when temporarily pulling a page from a freelist and
93 * putting it back unmodified.
95 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
97 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
98 static DEFINE_MUTEX(pcp_batch_high_lock
);
99 #define MIN_PERCPU_PAGELIST_FRACTION (8)
101 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
102 DEFINE_PER_CPU(int, numa_node
);
103 EXPORT_PER_CPU_SYMBOL(numa_node
);
106 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
108 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
110 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
111 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
112 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
113 * defined in <linux/topology.h>.
115 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
116 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
119 /* work_structs for global per-cpu drains */
122 struct work_struct work
;
124 static DEFINE_MUTEX(pcpu_drain_mutex
);
125 static DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
127 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
128 volatile unsigned long latent_entropy __latent_entropy
;
129 EXPORT_SYMBOL(latent_entropy
);
133 * Array of node states.
135 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
136 [N_POSSIBLE
] = NODE_MASK_ALL
,
137 [N_ONLINE
] = { { [0] = 1UL } },
139 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
140 #ifdef CONFIG_HIGHMEM
141 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
143 [N_MEMORY
] = { { [0] = 1UL } },
144 [N_CPU
] = { { [0] = 1UL } },
147 EXPORT_SYMBOL(node_states
);
149 atomic_long_t _totalram_pages __read_mostly
;
150 EXPORT_SYMBOL(_totalram_pages
);
151 unsigned long totalreserve_pages __read_mostly
;
152 unsigned long totalcma_pages __read_mostly
;
154 int percpu_pagelist_fraction
;
155 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
156 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
157 DEFINE_STATIC_KEY_TRUE(init_on_alloc
);
159 DEFINE_STATIC_KEY_FALSE(init_on_alloc
);
161 EXPORT_SYMBOL(init_on_alloc
);
163 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
164 DEFINE_STATIC_KEY_TRUE(init_on_free
);
166 DEFINE_STATIC_KEY_FALSE(init_on_free
);
168 EXPORT_SYMBOL(init_on_free
);
170 static int __init
early_init_on_alloc(char *buf
)
175 ret
= kstrtobool(buf
, &bool_result
);
178 if (bool_result
&& page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
181 static_branch_enable(&init_on_alloc
);
183 static_branch_disable(&init_on_alloc
);
186 early_param("init_on_alloc", early_init_on_alloc
);
188 static int __init
early_init_on_free(char *buf
)
193 ret
= kstrtobool(buf
, &bool_result
);
196 if (bool_result
&& page_poisoning_enabled())
197 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
199 static_branch_enable(&init_on_free
);
201 static_branch_disable(&init_on_free
);
204 early_param("init_on_free", early_init_on_free
);
207 * A cached value of the page's pageblock's migratetype, used when the page is
208 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
209 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
210 * Also the migratetype set in the page does not necessarily match the pcplist
211 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
212 * other index - this ensures that it will be put on the correct CMA freelist.
214 static inline int get_pcppage_migratetype(struct page
*page
)
219 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
221 page
->index
= migratetype
;
224 #ifdef CONFIG_PM_SLEEP
226 * The following functions are used by the suspend/hibernate code to temporarily
227 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
228 * while devices are suspended. To avoid races with the suspend/hibernate code,
229 * they should always be called with system_transition_mutex held
230 * (gfp_allowed_mask also should only be modified with system_transition_mutex
231 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
232 * with that modification).
235 static gfp_t saved_gfp_mask
;
237 void pm_restore_gfp_mask(void)
239 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
240 if (saved_gfp_mask
) {
241 gfp_allowed_mask
= saved_gfp_mask
;
246 void pm_restrict_gfp_mask(void)
248 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
249 WARN_ON(saved_gfp_mask
);
250 saved_gfp_mask
= gfp_allowed_mask
;
251 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
254 bool pm_suspended_storage(void)
256 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
260 #endif /* CONFIG_PM_SLEEP */
262 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
263 unsigned int pageblock_order __read_mostly
;
266 static void __free_pages_ok(struct page
*page
, unsigned int order
);
269 * results with 256, 32 in the lowmem_reserve sysctl:
270 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
271 * 1G machine -> (16M dma, 784M normal, 224M high)
272 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
273 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
274 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
276 * TBD: should special case ZONE_DMA32 machines here - in those we normally
277 * don't need any ZONE_NORMAL reservation
279 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
280 #ifdef CONFIG_ZONE_DMA
283 #ifdef CONFIG_ZONE_DMA32
287 #ifdef CONFIG_HIGHMEM
293 static char * const zone_names
[MAX_NR_ZONES
] = {
294 #ifdef CONFIG_ZONE_DMA
297 #ifdef CONFIG_ZONE_DMA32
301 #ifdef CONFIG_HIGHMEM
305 #ifdef CONFIG_ZONE_DEVICE
310 const char * const migratetype_names
[MIGRATE_TYPES
] = {
318 #ifdef CONFIG_MEMORY_ISOLATION
323 compound_page_dtor
* const compound_page_dtors
[NR_COMPOUND_DTORS
] = {
324 [NULL_COMPOUND_DTOR
] = NULL
,
325 [COMPOUND_PAGE_DTOR
] = free_compound_page
,
326 #ifdef CONFIG_HUGETLB_PAGE
327 [HUGETLB_PAGE_DTOR
] = free_huge_page
,
329 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
330 [TRANSHUGE_PAGE_DTOR
] = free_transhuge_page
,
334 int min_free_kbytes
= 1024;
335 int user_min_free_kbytes
= -1;
336 #ifdef CONFIG_DISCONTIGMEM
338 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
339 * are not on separate NUMA nodes. Functionally this works but with
340 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
341 * quite small. By default, do not boost watermarks on discontigmem as in
342 * many cases very high-order allocations like THP are likely to be
343 * unsupported and the premature reclaim offsets the advantage of long-term
344 * fragmentation avoidance.
346 int watermark_boost_factor __read_mostly
;
348 int watermark_boost_factor __read_mostly
= 15000;
350 int watermark_scale_factor
= 10;
352 static unsigned long nr_kernel_pages __initdata
;
353 static unsigned long nr_all_pages __initdata
;
354 static unsigned long dma_reserve __initdata
;
356 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
357 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
358 static unsigned long required_kernelcore __initdata
;
359 static unsigned long required_kernelcore_percent __initdata
;
360 static unsigned long required_movablecore __initdata
;
361 static unsigned long required_movablecore_percent __initdata
;
362 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
363 static bool mirrored_kernelcore __meminitdata
;
365 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
367 EXPORT_SYMBOL(movable_zone
);
370 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
371 unsigned int nr_online_nodes __read_mostly
= 1;
372 EXPORT_SYMBOL(nr_node_ids
);
373 EXPORT_SYMBOL(nr_online_nodes
);
376 int page_group_by_mobility_disabled __read_mostly
;
378 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
380 * During boot we initialize deferred pages on-demand, as needed, but once
381 * page_alloc_init_late() has finished, the deferred pages are all initialized,
382 * and we can permanently disable that path.
384 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
387 * Calling kasan_free_pages() only after deferred memory initialization
388 * has completed. Poisoning pages during deferred memory init will greatly
389 * lengthen the process and cause problem in large memory systems as the
390 * deferred pages initialization is done with interrupt disabled.
392 * Assuming that there will be no reference to those newly initialized
393 * pages before they are ever allocated, this should have no effect on
394 * KASAN memory tracking as the poison will be properly inserted at page
395 * allocation time. The only corner case is when pages are allocated by
396 * on-demand allocation and then freed again before the deferred pages
397 * initialization is done, but this is not likely to happen.
399 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
401 if (!static_branch_unlikely(&deferred_pages
))
402 kasan_free_pages(page
, order
);
405 /* Returns true if the struct page for the pfn is uninitialised */
406 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
408 int nid
= early_pfn_to_nid(pfn
);
410 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
417 * Returns true when the remaining initialisation should be deferred until
418 * later in the boot cycle when it can be parallelised.
420 static bool __meminit
421 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
423 static unsigned long prev_end_pfn
, nr_initialised
;
426 * prev_end_pfn static that contains the end of previous zone
427 * No need to protect because called very early in boot before smp_init.
429 if (prev_end_pfn
!= end_pfn
) {
430 prev_end_pfn
= end_pfn
;
434 /* Always populate low zones for address-constrained allocations */
435 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
439 * We start only with one section of pages, more pages are added as
440 * needed until the rest of deferred pages are initialized.
443 if ((nr_initialised
> PAGES_PER_SECTION
) &&
444 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
445 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
451 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
453 static inline bool early_page_uninitialised(unsigned long pfn
)
458 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
464 /* Return a pointer to the bitmap storing bits affecting a block of pages */
465 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
468 #ifdef CONFIG_SPARSEMEM
469 return section_to_usemap(__pfn_to_section(pfn
));
471 return page_zone(page
)->pageblock_flags
;
472 #endif /* CONFIG_SPARSEMEM */
475 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
477 #ifdef CONFIG_SPARSEMEM
478 pfn
&= (PAGES_PER_SECTION
-1);
480 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
481 #endif /* CONFIG_SPARSEMEM */
482 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
486 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
487 * @page: The page within the block of interest
488 * @pfn: The target page frame number
489 * @mask: mask of bits that the caller is interested in
491 * Return: pageblock_bits flags
493 static __always_inline
494 unsigned long __get_pfnblock_flags_mask(struct page
*page
,
498 unsigned long *bitmap
;
499 unsigned long bitidx
, word_bitidx
;
502 bitmap
= get_pageblock_bitmap(page
, pfn
);
503 bitidx
= pfn_to_bitidx(page
, pfn
);
504 word_bitidx
= bitidx
/ BITS_PER_LONG
;
505 bitidx
&= (BITS_PER_LONG
-1);
507 word
= bitmap
[word_bitidx
];
508 return (word
>> bitidx
) & mask
;
511 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
514 return __get_pfnblock_flags_mask(page
, pfn
, mask
);
517 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
519 return __get_pfnblock_flags_mask(page
, pfn
, MIGRATETYPE_MASK
);
523 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
524 * @page: The page within the block of interest
525 * @flags: The flags to set
526 * @pfn: The target page frame number
527 * @mask: mask of bits that the caller is interested in
529 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
533 unsigned long *bitmap
;
534 unsigned long bitidx
, word_bitidx
;
535 unsigned long old_word
, word
;
537 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
538 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
540 bitmap
= get_pageblock_bitmap(page
, pfn
);
541 bitidx
= pfn_to_bitidx(page
, pfn
);
542 word_bitidx
= bitidx
/ BITS_PER_LONG
;
543 bitidx
&= (BITS_PER_LONG
-1);
545 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
550 word
= READ_ONCE(bitmap
[word_bitidx
]);
552 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
553 if (word
== old_word
)
559 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
561 if (unlikely(page_group_by_mobility_disabled
&&
562 migratetype
< MIGRATE_PCPTYPES
))
563 migratetype
= MIGRATE_UNMOVABLE
;
565 set_pfnblock_flags_mask(page
, (unsigned long)migratetype
,
566 page_to_pfn(page
), MIGRATETYPE_MASK
);
569 #ifdef CONFIG_DEBUG_VM
570 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
574 unsigned long pfn
= page_to_pfn(page
);
575 unsigned long sp
, start_pfn
;
578 seq
= zone_span_seqbegin(zone
);
579 start_pfn
= zone
->zone_start_pfn
;
580 sp
= zone
->spanned_pages
;
581 if (!zone_spans_pfn(zone
, pfn
))
583 } while (zone_span_seqretry(zone
, seq
));
586 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
587 pfn
, zone_to_nid(zone
), zone
->name
,
588 start_pfn
, start_pfn
+ sp
);
593 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
595 if (!pfn_valid_within(page_to_pfn(page
)))
597 if (zone
!= page_zone(page
))
603 * Temporary debugging check for pages not lying within a given zone.
605 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
607 if (page_outside_zone_boundaries(zone
, page
))
609 if (!page_is_consistent(zone
, page
))
615 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
621 static void bad_page(struct page
*page
, const char *reason
)
623 static unsigned long resume
;
624 static unsigned long nr_shown
;
625 static unsigned long nr_unshown
;
628 * Allow a burst of 60 reports, then keep quiet for that minute;
629 * or allow a steady drip of one report per second.
631 if (nr_shown
== 60) {
632 if (time_before(jiffies
, resume
)) {
638 "BUG: Bad page state: %lu messages suppressed\n",
645 resume
= jiffies
+ 60 * HZ
;
647 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
648 current
->comm
, page_to_pfn(page
));
649 __dump_page(page
, reason
);
650 dump_page_owner(page
);
655 /* Leave bad fields for debug, except PageBuddy could make trouble */
656 page_mapcount_reset(page
); /* remove PageBuddy */
657 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
661 * Higher-order pages are called "compound pages". They are structured thusly:
663 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
665 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
666 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
668 * The first tail page's ->compound_dtor holds the offset in array of compound
669 * page destructors. See compound_page_dtors.
671 * The first tail page's ->compound_order holds the order of allocation.
672 * This usage means that zero-order pages may not be compound.
675 void free_compound_page(struct page
*page
)
677 mem_cgroup_uncharge(page
);
678 __free_pages_ok(page
, compound_order(page
));
681 void prep_compound_page(struct page
*page
, unsigned int order
)
684 int nr_pages
= 1 << order
;
687 for (i
= 1; i
< nr_pages
; i
++) {
688 struct page
*p
= page
+ i
;
689 set_page_count(p
, 0);
690 p
->mapping
= TAIL_MAPPING
;
691 set_compound_head(p
, page
);
694 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
695 set_compound_order(page
, order
);
696 atomic_set(compound_mapcount_ptr(page
), -1);
697 if (hpage_pincount_available(page
))
698 atomic_set(compound_pincount_ptr(page
), 0);
701 #ifdef CONFIG_DEBUG_PAGEALLOC
702 unsigned int _debug_guardpage_minorder
;
704 bool _debug_pagealloc_enabled_early __read_mostly
705 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
706 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
707 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
708 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
710 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
712 static int __init
early_debug_pagealloc(char *buf
)
714 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
716 early_param("debug_pagealloc", early_debug_pagealloc
);
718 void init_debug_pagealloc(void)
720 if (!debug_pagealloc_enabled())
723 static_branch_enable(&_debug_pagealloc_enabled
);
725 if (!debug_guardpage_minorder())
728 static_branch_enable(&_debug_guardpage_enabled
);
731 static int __init
debug_guardpage_minorder_setup(char *buf
)
735 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
736 pr_err("Bad debug_guardpage_minorder value\n");
739 _debug_guardpage_minorder
= res
;
740 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
743 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
745 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
746 unsigned int order
, int migratetype
)
748 if (!debug_guardpage_enabled())
751 if (order
>= debug_guardpage_minorder())
754 __SetPageGuard(page
);
755 INIT_LIST_HEAD(&page
->lru
);
756 set_page_private(page
, order
);
757 /* Guard pages are not available for any usage */
758 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
763 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
764 unsigned int order
, int migratetype
)
766 if (!debug_guardpage_enabled())
769 __ClearPageGuard(page
);
771 set_page_private(page
, 0);
772 if (!is_migrate_isolate(migratetype
))
773 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
776 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
777 unsigned int order
, int migratetype
) { return false; }
778 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
779 unsigned int order
, int migratetype
) {}
782 static inline void set_page_order(struct page
*page
, unsigned int order
)
784 set_page_private(page
, order
);
785 __SetPageBuddy(page
);
789 * This function checks whether a page is free && is the buddy
790 * we can coalesce a page and its buddy if
791 * (a) the buddy is not in a hole (check before calling!) &&
792 * (b) the buddy is in the buddy system &&
793 * (c) a page and its buddy have the same order &&
794 * (d) a page and its buddy are in the same zone.
796 * For recording whether a page is in the buddy system, we set PageBuddy.
797 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
799 * For recording page's order, we use page_private(page).
801 static inline bool page_is_buddy(struct page
*page
, struct page
*buddy
,
804 if (!page_is_guard(buddy
) && !PageBuddy(buddy
))
807 if (page_order(buddy
) != order
)
811 * zone check is done late to avoid uselessly calculating
812 * zone/node ids for pages that could never merge.
814 if (page_zone_id(page
) != page_zone_id(buddy
))
817 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
822 #ifdef CONFIG_COMPACTION
823 static inline struct capture_control
*task_capc(struct zone
*zone
)
825 struct capture_control
*capc
= current
->capture_control
;
827 return unlikely(capc
) &&
828 !(current
->flags
& PF_KTHREAD
) &&
830 capc
->cc
->zone
== zone
? capc
: NULL
;
834 compaction_capture(struct capture_control
*capc
, struct page
*page
,
835 int order
, int migratetype
)
837 if (!capc
|| order
!= capc
->cc
->order
)
840 /* Do not accidentally pollute CMA or isolated regions*/
841 if (is_migrate_cma(migratetype
) ||
842 is_migrate_isolate(migratetype
))
846 * Do not let lower order allocations polluate a movable pageblock.
847 * This might let an unmovable request use a reclaimable pageblock
848 * and vice-versa but no more than normal fallback logic which can
849 * have trouble finding a high-order free page.
851 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
859 static inline struct capture_control
*task_capc(struct zone
*zone
)
865 compaction_capture(struct capture_control
*capc
, struct page
*page
,
866 int order
, int migratetype
)
870 #endif /* CONFIG_COMPACTION */
872 /* Used for pages not on another list */
873 static inline void add_to_free_list(struct page
*page
, struct zone
*zone
,
874 unsigned int order
, int migratetype
)
876 struct free_area
*area
= &zone
->free_area
[order
];
878 list_add(&page
->lru
, &area
->free_list
[migratetype
]);
882 /* Used for pages not on another list */
883 static inline void add_to_free_list_tail(struct page
*page
, struct zone
*zone
,
884 unsigned int order
, int migratetype
)
886 struct free_area
*area
= &zone
->free_area
[order
];
888 list_add_tail(&page
->lru
, &area
->free_list
[migratetype
]);
892 /* Used for pages which are on another list */
893 static inline void move_to_free_list(struct page
*page
, struct zone
*zone
,
894 unsigned int order
, int migratetype
)
896 struct free_area
*area
= &zone
->free_area
[order
];
898 list_move(&page
->lru
, &area
->free_list
[migratetype
]);
901 static inline void del_page_from_free_list(struct page
*page
, struct zone
*zone
,
904 /* clear reported state and update reported page count */
905 if (page_reported(page
))
906 __ClearPageReported(page
);
908 list_del(&page
->lru
);
909 __ClearPageBuddy(page
);
910 set_page_private(page
, 0);
911 zone
->free_area
[order
].nr_free
--;
915 * If this is not the largest possible page, check if the buddy
916 * of the next-highest order is free. If it is, it's possible
917 * that pages are being freed that will coalesce soon. In case,
918 * that is happening, add the free page to the tail of the list
919 * so it's less likely to be used soon and more likely to be merged
920 * as a higher order page
923 buddy_merge_likely(unsigned long pfn
, unsigned long buddy_pfn
,
924 struct page
*page
, unsigned int order
)
926 struct page
*higher_page
, *higher_buddy
;
927 unsigned long combined_pfn
;
929 if (order
>= MAX_ORDER
- 2)
932 if (!pfn_valid_within(buddy_pfn
))
935 combined_pfn
= buddy_pfn
& pfn
;
936 higher_page
= page
+ (combined_pfn
- pfn
);
937 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
938 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
940 return pfn_valid_within(buddy_pfn
) &&
941 page_is_buddy(higher_page
, higher_buddy
, order
+ 1);
945 * Freeing function for a buddy system allocator.
947 * The concept of a buddy system is to maintain direct-mapped table
948 * (containing bit values) for memory blocks of various "orders".
949 * The bottom level table contains the map for the smallest allocatable
950 * units of memory (here, pages), and each level above it describes
951 * pairs of units from the levels below, hence, "buddies".
952 * At a high level, all that happens here is marking the table entry
953 * at the bottom level available, and propagating the changes upward
954 * as necessary, plus some accounting needed to play nicely with other
955 * parts of the VM system.
956 * At each level, we keep a list of pages, which are heads of continuous
957 * free pages of length of (1 << order) and marked with PageBuddy.
958 * Page's order is recorded in page_private(page) field.
959 * So when we are allocating or freeing one, we can derive the state of the
960 * other. That is, if we allocate a small block, and both were
961 * free, the remainder of the region must be split into blocks.
962 * If a block is freed, and its buddy is also free, then this
963 * triggers coalescing into a block of larger size.
968 static inline void __free_one_page(struct page
*page
,
970 struct zone
*zone
, unsigned int order
,
971 int migratetype
, fpi_t fpi_flags
)
973 struct capture_control
*capc
= task_capc(zone
);
974 unsigned long buddy_pfn
;
975 unsigned long combined_pfn
;
976 unsigned int max_order
;
980 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
982 VM_BUG_ON(!zone_is_initialized(zone
));
983 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
985 VM_BUG_ON(migratetype
== -1);
986 if (likely(!is_migrate_isolate(migratetype
)))
987 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
989 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
990 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
993 while (order
< max_order
- 1) {
994 if (compaction_capture(capc
, page
, order
, migratetype
)) {
995 __mod_zone_freepage_state(zone
, -(1 << order
),
999 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
1000 buddy
= page
+ (buddy_pfn
- pfn
);
1002 if (!pfn_valid_within(buddy_pfn
))
1004 if (!page_is_buddy(page
, buddy
, order
))
1007 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1008 * merge with it and move up one order.
1010 if (page_is_guard(buddy
))
1011 clear_page_guard(zone
, buddy
, order
, migratetype
);
1013 del_page_from_free_list(buddy
, zone
, order
);
1014 combined_pfn
= buddy_pfn
& pfn
;
1015 page
= page
+ (combined_pfn
- pfn
);
1019 if (max_order
< MAX_ORDER
) {
1020 /* If we are here, it means order is >= pageblock_order.
1021 * We want to prevent merge between freepages on isolate
1022 * pageblock and normal pageblock. Without this, pageblock
1023 * isolation could cause incorrect freepage or CMA accounting.
1025 * We don't want to hit this code for the more frequent
1026 * low-order merging.
1028 if (unlikely(has_isolate_pageblock(zone
))) {
1031 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
1032 buddy
= page
+ (buddy_pfn
- pfn
);
1033 buddy_mt
= get_pageblock_migratetype(buddy
);
1035 if (migratetype
!= buddy_mt
1036 && (is_migrate_isolate(migratetype
) ||
1037 is_migrate_isolate(buddy_mt
)))
1041 goto continue_merging
;
1045 set_page_order(page
, order
);
1047 if (is_shuffle_order(order
))
1048 to_tail
= shuffle_pick_tail();
1050 to_tail
= buddy_merge_likely(pfn
, buddy_pfn
, page
, order
);
1053 add_to_free_list_tail(page
, zone
, order
, migratetype
);
1055 add_to_free_list(page
, zone
, order
, migratetype
);
1057 /* Notify page reporting subsystem of freed page */
1058 if (!(fpi_flags
& FPI_SKIP_REPORT_NOTIFY
))
1059 page_reporting_notify_free(order
);
1063 * A bad page could be due to a number of fields. Instead of multiple branches,
1064 * try and check multiple fields with one check. The caller must do a detailed
1065 * check if necessary.
1067 static inline bool page_expected_state(struct page
*page
,
1068 unsigned long check_flags
)
1070 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1073 if (unlikely((unsigned long)page
->mapping
|
1074 page_ref_count(page
) |
1076 (unsigned long)page
->mem_cgroup
|
1078 (page
->flags
& check_flags
)))
1084 static const char *page_bad_reason(struct page
*page
, unsigned long flags
)
1086 const char *bad_reason
= NULL
;
1088 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1089 bad_reason
= "nonzero mapcount";
1090 if (unlikely(page
->mapping
!= NULL
))
1091 bad_reason
= "non-NULL mapping";
1092 if (unlikely(page_ref_count(page
) != 0))
1093 bad_reason
= "nonzero _refcount";
1094 if (unlikely(page
->flags
& flags
)) {
1095 if (flags
== PAGE_FLAGS_CHECK_AT_PREP
)
1096 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1098 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1101 if (unlikely(page
->mem_cgroup
))
1102 bad_reason
= "page still charged to cgroup";
1107 static void check_free_page_bad(struct page
*page
)
1110 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_FREE
));
1113 static inline int check_free_page(struct page
*page
)
1115 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1118 /* Something has gone sideways, find it */
1119 check_free_page_bad(page
);
1123 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1128 * We rely page->lru.next never has bit 0 set, unless the page
1129 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1131 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1133 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1137 switch (page
- head_page
) {
1139 /* the first tail page: ->mapping may be compound_mapcount() */
1140 if (unlikely(compound_mapcount(page
))) {
1141 bad_page(page
, "nonzero compound_mapcount");
1147 * the second tail page: ->mapping is
1148 * deferred_list.next -- ignore value.
1152 if (page
->mapping
!= TAIL_MAPPING
) {
1153 bad_page(page
, "corrupted mapping in tail page");
1158 if (unlikely(!PageTail(page
))) {
1159 bad_page(page
, "PageTail not set");
1162 if (unlikely(compound_head(page
) != head_page
)) {
1163 bad_page(page
, "compound_head not consistent");
1168 page
->mapping
= NULL
;
1169 clear_compound_head(page
);
1173 static void kernel_init_free_pages(struct page
*page
, int numpages
)
1177 /* s390's use of memset() could override KASAN redzones. */
1178 kasan_disable_current();
1179 for (i
= 0; i
< numpages
; i
++)
1180 clear_highpage(page
+ i
);
1181 kasan_enable_current();
1184 static __always_inline
bool free_pages_prepare(struct page
*page
,
1185 unsigned int order
, bool check_free
)
1189 VM_BUG_ON_PAGE(PageTail(page
), page
);
1191 trace_mm_page_free(page
, order
);
1193 if (unlikely(PageHWPoison(page
)) && !order
) {
1195 * Do not let hwpoison pages hit pcplists/buddy
1196 * Untie memcg state and reset page's owner
1198 if (memcg_kmem_enabled() && PageKmemcg(page
))
1199 __memcg_kmem_uncharge_page(page
, order
);
1200 reset_page_owner(page
, order
);
1205 * Check tail pages before head page information is cleared to
1206 * avoid checking PageCompound for order-0 pages.
1208 if (unlikely(order
)) {
1209 bool compound
= PageCompound(page
);
1212 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1215 ClearPageDoubleMap(page
);
1216 for (i
= 1; i
< (1 << order
); i
++) {
1218 bad
+= free_tail_pages_check(page
, page
+ i
);
1219 if (unlikely(check_free_page(page
+ i
))) {
1223 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1226 if (PageMappingFlags(page
))
1227 page
->mapping
= NULL
;
1228 if (memcg_kmem_enabled() && PageKmemcg(page
))
1229 __memcg_kmem_uncharge_page(page
, order
);
1231 bad
+= check_free_page(page
);
1235 page_cpupid_reset_last(page
);
1236 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1237 reset_page_owner(page
, order
);
1239 if (!PageHighMem(page
)) {
1240 debug_check_no_locks_freed(page_address(page
),
1241 PAGE_SIZE
<< order
);
1242 debug_check_no_obj_freed(page_address(page
),
1243 PAGE_SIZE
<< order
);
1245 if (want_init_on_free())
1246 kernel_init_free_pages(page
, 1 << order
);
1248 kernel_poison_pages(page
, 1 << order
, 0);
1250 * arch_free_page() can make the page's contents inaccessible. s390
1251 * does this. So nothing which can access the page's contents should
1252 * happen after this.
1254 arch_free_page(page
, order
);
1256 if (debug_pagealloc_enabled_static())
1257 kernel_map_pages(page
, 1 << order
, 0);
1259 kasan_free_nondeferred_pages(page
, order
);
1264 #ifdef CONFIG_DEBUG_VM
1266 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1267 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1268 * moved from pcp lists to free lists.
1270 static bool free_pcp_prepare(struct page
*page
)
1272 return free_pages_prepare(page
, 0, true);
1275 static bool bulkfree_pcp_prepare(struct page
*page
)
1277 if (debug_pagealloc_enabled_static())
1278 return check_free_page(page
);
1284 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1285 * moving from pcp lists to free list in order to reduce overhead. With
1286 * debug_pagealloc enabled, they are checked also immediately when being freed
1289 static bool free_pcp_prepare(struct page
*page
)
1291 if (debug_pagealloc_enabled_static())
1292 return free_pages_prepare(page
, 0, true);
1294 return free_pages_prepare(page
, 0, false);
1297 static bool bulkfree_pcp_prepare(struct page
*page
)
1299 return check_free_page(page
);
1301 #endif /* CONFIG_DEBUG_VM */
1303 static inline void prefetch_buddy(struct page
*page
)
1305 unsigned long pfn
= page_to_pfn(page
);
1306 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1307 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1313 * Frees a number of pages from the PCP lists
1314 * Assumes all pages on list are in same zone, and of same order.
1315 * count is the number of pages to free.
1317 * If the zone was previously in an "all pages pinned" state then look to
1318 * see if this freeing clears that state.
1320 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1321 * pinned" detection logic.
1323 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1324 struct per_cpu_pages
*pcp
)
1326 int migratetype
= 0;
1328 int prefetch_nr
= 0;
1329 bool isolated_pageblocks
;
1330 struct page
*page
, *tmp
;
1334 * Ensure proper count is passed which otherwise would stuck in the
1335 * below while (list_empty(list)) loop.
1337 count
= min(pcp
->count
, count
);
1339 struct list_head
*list
;
1342 * Remove pages from lists in a round-robin fashion. A
1343 * batch_free count is maintained that is incremented when an
1344 * empty list is encountered. This is so more pages are freed
1345 * off fuller lists instead of spinning excessively around empty
1350 if (++migratetype
== MIGRATE_PCPTYPES
)
1352 list
= &pcp
->lists
[migratetype
];
1353 } while (list_empty(list
));
1355 /* This is the only non-empty list. Free them all. */
1356 if (batch_free
== MIGRATE_PCPTYPES
)
1360 page
= list_last_entry(list
, struct page
, lru
);
1361 /* must delete to avoid corrupting pcp list */
1362 list_del(&page
->lru
);
1365 if (bulkfree_pcp_prepare(page
))
1368 list_add_tail(&page
->lru
, &head
);
1371 * We are going to put the page back to the global
1372 * pool, prefetch its buddy to speed up later access
1373 * under zone->lock. It is believed the overhead of
1374 * an additional test and calculating buddy_pfn here
1375 * can be offset by reduced memory latency later. To
1376 * avoid excessive prefetching due to large count, only
1377 * prefetch buddy for the first pcp->batch nr of pages.
1379 if (prefetch_nr
++ < pcp
->batch
)
1380 prefetch_buddy(page
);
1381 } while (--count
&& --batch_free
&& !list_empty(list
));
1384 spin_lock(&zone
->lock
);
1385 isolated_pageblocks
= has_isolate_pageblock(zone
);
1388 * Use safe version since after __free_one_page(),
1389 * page->lru.next will not point to original list.
1391 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1392 int mt
= get_pcppage_migratetype(page
);
1393 /* MIGRATE_ISOLATE page should not go to pcplists */
1394 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1395 /* Pageblock could have been isolated meanwhile */
1396 if (unlikely(isolated_pageblocks
))
1397 mt
= get_pageblock_migratetype(page
);
1399 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
, FPI_NONE
);
1400 trace_mm_page_pcpu_drain(page
, 0, mt
);
1402 spin_unlock(&zone
->lock
);
1405 static void free_one_page(struct zone
*zone
,
1406 struct page
*page
, unsigned long pfn
,
1410 spin_lock(&zone
->lock
);
1411 if (unlikely(has_isolate_pageblock(zone
) ||
1412 is_migrate_isolate(migratetype
))) {
1413 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1415 __free_one_page(page
, pfn
, zone
, order
, migratetype
, FPI_NONE
);
1416 spin_unlock(&zone
->lock
);
1419 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1420 unsigned long zone
, int nid
)
1422 mm_zero_struct_page(page
);
1423 set_page_links(page
, zone
, nid
, pfn
);
1424 init_page_count(page
);
1425 page_mapcount_reset(page
);
1426 page_cpupid_reset_last(page
);
1427 page_kasan_tag_reset(page
);
1429 INIT_LIST_HEAD(&page
->lru
);
1430 #ifdef WANT_PAGE_VIRTUAL
1431 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1432 if (!is_highmem_idx(zone
))
1433 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1437 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1438 static void __meminit
init_reserved_page(unsigned long pfn
)
1443 if (!early_page_uninitialised(pfn
))
1446 nid
= early_pfn_to_nid(pfn
);
1447 pgdat
= NODE_DATA(nid
);
1449 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1450 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1452 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1455 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1458 static inline void init_reserved_page(unsigned long pfn
)
1461 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1464 * Initialised pages do not have PageReserved set. This function is
1465 * called for each range allocated by the bootmem allocator and
1466 * marks the pages PageReserved. The remaining valid pages are later
1467 * sent to the buddy page allocator.
1469 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1471 unsigned long start_pfn
= PFN_DOWN(start
);
1472 unsigned long end_pfn
= PFN_UP(end
);
1474 for (; start_pfn
< end_pfn
; start_pfn
++) {
1475 if (pfn_valid(start_pfn
)) {
1476 struct page
*page
= pfn_to_page(start_pfn
);
1478 init_reserved_page(start_pfn
);
1480 /* Avoid false-positive PageTail() */
1481 INIT_LIST_HEAD(&page
->lru
);
1484 * no need for atomic set_bit because the struct
1485 * page is not visible yet so nobody should
1488 __SetPageReserved(page
);
1493 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1495 unsigned long flags
;
1497 unsigned long pfn
= page_to_pfn(page
);
1499 if (!free_pages_prepare(page
, order
, true))
1502 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1503 local_irq_save(flags
);
1504 __count_vm_events(PGFREE
, 1 << order
);
1505 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1506 local_irq_restore(flags
);
1509 void __free_pages_core(struct page
*page
, unsigned int order
)
1511 unsigned int nr_pages
= 1 << order
;
1512 struct page
*p
= page
;
1516 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1518 __ClearPageReserved(p
);
1519 set_page_count(p
, 0);
1521 __ClearPageReserved(p
);
1522 set_page_count(p
, 0);
1524 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1525 set_page_refcounted(page
);
1526 __free_pages(page
, order
);
1529 #ifdef CONFIG_NEED_MULTIPLE_NODES
1531 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1533 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1536 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1538 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
1539 struct mminit_pfnnid_cache
*state
)
1541 unsigned long start_pfn
, end_pfn
;
1544 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
1545 return state
->last_nid
;
1547 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
1548 if (nid
!= NUMA_NO_NODE
) {
1549 state
->last_start
= start_pfn
;
1550 state
->last_end
= end_pfn
;
1551 state
->last_nid
= nid
;
1556 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1558 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1560 static DEFINE_SPINLOCK(early_pfn_lock
);
1563 spin_lock(&early_pfn_lock
);
1564 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1566 nid
= first_online_node
;
1567 spin_unlock(&early_pfn_lock
);
1571 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1573 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1576 if (early_page_uninitialised(pfn
))
1578 __free_pages_core(page
, order
);
1582 * Check that the whole (or subset of) a pageblock given by the interval of
1583 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1584 * with the migration of free compaction scanner. The scanners then need to
1585 * use only pfn_valid_within() check for arches that allow holes within
1588 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1590 * It's possible on some configurations to have a setup like node0 node1 node0
1591 * i.e. it's possible that all pages within a zones range of pages do not
1592 * belong to a single zone. We assume that a border between node0 and node1
1593 * can occur within a single pageblock, but not a node0 node1 node0
1594 * interleaving within a single pageblock. It is therefore sufficient to check
1595 * the first and last page of a pageblock and avoid checking each individual
1596 * page in a pageblock.
1598 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1599 unsigned long end_pfn
, struct zone
*zone
)
1601 struct page
*start_page
;
1602 struct page
*end_page
;
1604 /* end_pfn is one past the range we are checking */
1607 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1610 start_page
= pfn_to_online_page(start_pfn
);
1614 if (page_zone(start_page
) != zone
)
1617 end_page
= pfn_to_page(end_pfn
);
1619 /* This gives a shorter code than deriving page_zone(end_page) */
1620 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1626 void set_zone_contiguous(struct zone
*zone
)
1628 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1629 unsigned long block_end_pfn
;
1631 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1632 for (; block_start_pfn
< zone_end_pfn(zone
);
1633 block_start_pfn
= block_end_pfn
,
1634 block_end_pfn
+= pageblock_nr_pages
) {
1636 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1638 if (!__pageblock_pfn_to_page(block_start_pfn
,
1639 block_end_pfn
, zone
))
1644 /* We confirm that there is no hole */
1645 zone
->contiguous
= true;
1648 void clear_zone_contiguous(struct zone
*zone
)
1650 zone
->contiguous
= false;
1653 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1654 static void __init
deferred_free_range(unsigned long pfn
,
1655 unsigned long nr_pages
)
1663 page
= pfn_to_page(pfn
);
1665 /* Free a large naturally-aligned chunk if possible */
1666 if (nr_pages
== pageblock_nr_pages
&&
1667 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1668 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1669 __free_pages_core(page
, pageblock_order
);
1673 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1674 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1675 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1676 __free_pages_core(page
, 0);
1680 /* Completion tracking for deferred_init_memmap() threads */
1681 static atomic_t pgdat_init_n_undone __initdata
;
1682 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1684 static inline void __init
pgdat_init_report_one_done(void)
1686 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1687 complete(&pgdat_init_all_done_comp
);
1691 * Returns true if page needs to be initialized or freed to buddy allocator.
1693 * First we check if pfn is valid on architectures where it is possible to have
1694 * holes within pageblock_nr_pages. On systems where it is not possible, this
1695 * function is optimized out.
1697 * Then, we check if a current large page is valid by only checking the validity
1700 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1702 if (!pfn_valid_within(pfn
))
1704 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1710 * Free pages to buddy allocator. Try to free aligned pages in
1711 * pageblock_nr_pages sizes.
1713 static void __init
deferred_free_pages(unsigned long pfn
,
1714 unsigned long end_pfn
)
1716 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1717 unsigned long nr_free
= 0;
1719 for (; pfn
< end_pfn
; pfn
++) {
1720 if (!deferred_pfn_valid(pfn
)) {
1721 deferred_free_range(pfn
- nr_free
, nr_free
);
1723 } else if (!(pfn
& nr_pgmask
)) {
1724 deferred_free_range(pfn
- nr_free
, nr_free
);
1730 /* Free the last block of pages to allocator */
1731 deferred_free_range(pfn
- nr_free
, nr_free
);
1735 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1736 * by performing it only once every pageblock_nr_pages.
1737 * Return number of pages initialized.
1739 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1741 unsigned long end_pfn
)
1743 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1744 int nid
= zone_to_nid(zone
);
1745 unsigned long nr_pages
= 0;
1746 int zid
= zone_idx(zone
);
1747 struct page
*page
= NULL
;
1749 for (; pfn
< end_pfn
; pfn
++) {
1750 if (!deferred_pfn_valid(pfn
)) {
1753 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1754 page
= pfn_to_page(pfn
);
1758 __init_single_page(page
, pfn
, zid
, nid
);
1765 * This function is meant to pre-load the iterator for the zone init.
1766 * Specifically it walks through the ranges until we are caught up to the
1767 * first_init_pfn value and exits there. If we never encounter the value we
1768 * return false indicating there are no valid ranges left.
1771 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1772 unsigned long *spfn
, unsigned long *epfn
,
1773 unsigned long first_init_pfn
)
1778 * Start out by walking through the ranges in this zone that have
1779 * already been initialized. We don't need to do anything with them
1780 * so we just need to flush them out of the system.
1782 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1783 if (*epfn
<= first_init_pfn
)
1785 if (*spfn
< first_init_pfn
)
1786 *spfn
= first_init_pfn
;
1795 * Initialize and free pages. We do it in two loops: first we initialize
1796 * struct page, then free to buddy allocator, because while we are
1797 * freeing pages we can access pages that are ahead (computing buddy
1798 * page in __free_one_page()).
1800 * In order to try and keep some memory in the cache we have the loop
1801 * broken along max page order boundaries. This way we will not cause
1802 * any issues with the buddy page computation.
1804 static unsigned long __init
1805 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1806 unsigned long *end_pfn
)
1808 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1809 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1810 unsigned long nr_pages
= 0;
1813 /* First we loop through and initialize the page values */
1814 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1817 if (mo_pfn
<= *start_pfn
)
1820 t
= min(mo_pfn
, *end_pfn
);
1821 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1823 if (mo_pfn
< *end_pfn
) {
1824 *start_pfn
= mo_pfn
;
1829 /* Reset values and now loop through freeing pages as needed */
1832 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1838 t
= min(mo_pfn
, epfn
);
1839 deferred_free_pages(spfn
, t
);
1849 deferred_init_memmap_chunk(unsigned long start_pfn
, unsigned long end_pfn
,
1852 unsigned long spfn
, epfn
;
1853 struct zone
*zone
= arg
;
1856 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
, start_pfn
);
1859 * Initialize and free pages in MAX_ORDER sized increments so that we
1860 * can avoid introducing any issues with the buddy allocator.
1862 while (spfn
< end_pfn
) {
1863 deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1868 /* An arch may override for more concurrency. */
1870 deferred_page_init_max_threads(const struct cpumask
*node_cpumask
)
1875 /* Initialise remaining memory on a node */
1876 static int __init
deferred_init_memmap(void *data
)
1878 pg_data_t
*pgdat
= data
;
1879 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1880 unsigned long spfn
= 0, epfn
= 0;
1881 unsigned long first_init_pfn
, flags
;
1882 unsigned long start
= jiffies
;
1884 int zid
, max_threads
;
1887 /* Bind memory initialisation thread to a local node if possible */
1888 if (!cpumask_empty(cpumask
))
1889 set_cpus_allowed_ptr(current
, cpumask
);
1891 pgdat_resize_lock(pgdat
, &flags
);
1892 first_init_pfn
= pgdat
->first_deferred_pfn
;
1893 if (first_init_pfn
== ULONG_MAX
) {
1894 pgdat_resize_unlock(pgdat
, &flags
);
1895 pgdat_init_report_one_done();
1899 /* Sanity check boundaries */
1900 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1901 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1902 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1905 * Once we unlock here, the zone cannot be grown anymore, thus if an
1906 * interrupt thread must allocate this early in boot, zone must be
1907 * pre-grown prior to start of deferred page initialization.
1909 pgdat_resize_unlock(pgdat
, &flags
);
1911 /* Only the highest zone is deferred so find it */
1912 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1913 zone
= pgdat
->node_zones
+ zid
;
1914 if (first_init_pfn
< zone_end_pfn(zone
))
1918 /* If the zone is empty somebody else may have cleared out the zone */
1919 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1923 max_threads
= deferred_page_init_max_threads(cpumask
);
1925 while (spfn
< epfn
) {
1926 unsigned long epfn_align
= ALIGN(epfn
, PAGES_PER_SECTION
);
1927 struct padata_mt_job job
= {
1928 .thread_fn
= deferred_init_memmap_chunk
,
1931 .size
= epfn_align
- spfn
,
1932 .align
= PAGES_PER_SECTION
,
1933 .min_chunk
= PAGES_PER_SECTION
,
1934 .max_threads
= max_threads
,
1937 padata_do_multithreaded(&job
);
1938 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1942 /* Sanity check that the next zone really is unpopulated */
1943 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1945 pr_info("node %d deferred pages initialised in %ums\n",
1946 pgdat
->node_id
, jiffies_to_msecs(jiffies
- start
));
1948 pgdat_init_report_one_done();
1953 * If this zone has deferred pages, try to grow it by initializing enough
1954 * deferred pages to satisfy the allocation specified by order, rounded up to
1955 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1956 * of SECTION_SIZE bytes by initializing struct pages in increments of
1957 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1959 * Return true when zone was grown, otherwise return false. We return true even
1960 * when we grow less than requested, to let the caller decide if there are
1961 * enough pages to satisfy the allocation.
1963 * Note: We use noinline because this function is needed only during boot, and
1964 * it is called from a __ref function _deferred_grow_zone. This way we are
1965 * making sure that it is not inlined into permanent text section.
1967 static noinline
bool __init
1968 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
1970 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
1971 pg_data_t
*pgdat
= zone
->zone_pgdat
;
1972 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
1973 unsigned long spfn
, epfn
, flags
;
1974 unsigned long nr_pages
= 0;
1977 /* Only the last zone may have deferred pages */
1978 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
1981 pgdat_resize_lock(pgdat
, &flags
);
1984 * If someone grew this zone while we were waiting for spinlock, return
1985 * true, as there might be enough pages already.
1987 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
1988 pgdat_resize_unlock(pgdat
, &flags
);
1992 /* If the zone is empty somebody else may have cleared out the zone */
1993 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1994 first_deferred_pfn
)) {
1995 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1996 pgdat_resize_unlock(pgdat
, &flags
);
1997 /* Retry only once. */
1998 return first_deferred_pfn
!= ULONG_MAX
;
2002 * Initialize and free pages in MAX_ORDER sized increments so
2003 * that we can avoid introducing any issues with the buddy
2006 while (spfn
< epfn
) {
2007 /* update our first deferred PFN for this section */
2008 first_deferred_pfn
= spfn
;
2010 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
2011 touch_nmi_watchdog();
2013 /* We should only stop along section boundaries */
2014 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
2017 /* If our quota has been met we can stop here */
2018 if (nr_pages
>= nr_pages_needed
)
2022 pgdat
->first_deferred_pfn
= spfn
;
2023 pgdat_resize_unlock(pgdat
, &flags
);
2025 return nr_pages
> 0;
2029 * deferred_grow_zone() is __init, but it is called from
2030 * get_page_from_freelist() during early boot until deferred_pages permanently
2031 * disables this call. This is why we have refdata wrapper to avoid warning,
2032 * and to ensure that the function body gets unloaded.
2035 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
2037 return deferred_grow_zone(zone
, order
);
2040 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2042 void __init
page_alloc_init_late(void)
2047 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2049 /* There will be num_node_state(N_MEMORY) threads */
2050 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
2051 for_each_node_state(nid
, N_MEMORY
) {
2052 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
2055 /* Block until all are initialised */
2056 wait_for_completion(&pgdat_init_all_done_comp
);
2059 * The number of managed pages has changed due to the initialisation
2060 * so the pcpu batch and high limits needs to be updated or the limits
2061 * will be artificially small.
2063 for_each_populated_zone(zone
)
2064 zone_pcp_update(zone
);
2067 * We initialized the rest of the deferred pages. Permanently disable
2068 * on-demand struct page initialization.
2070 static_branch_disable(&deferred_pages
);
2072 /* Reinit limits that are based on free pages after the kernel is up */
2073 files_maxfiles_init();
2076 /* Discard memblock private memory */
2079 for_each_node_state(nid
, N_MEMORY
)
2080 shuffle_free_memory(NODE_DATA(nid
));
2082 for_each_populated_zone(zone
)
2083 set_zone_contiguous(zone
);
2087 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2088 void __init
init_cma_reserved_pageblock(struct page
*page
)
2090 unsigned i
= pageblock_nr_pages
;
2091 struct page
*p
= page
;
2094 __ClearPageReserved(p
);
2095 set_page_count(p
, 0);
2098 set_pageblock_migratetype(page
, MIGRATE_CMA
);
2100 if (pageblock_order
>= MAX_ORDER
) {
2101 i
= pageblock_nr_pages
;
2104 set_page_refcounted(p
);
2105 __free_pages(p
, MAX_ORDER
- 1);
2106 p
+= MAX_ORDER_NR_PAGES
;
2107 } while (i
-= MAX_ORDER_NR_PAGES
);
2109 set_page_refcounted(page
);
2110 __free_pages(page
, pageblock_order
);
2113 adjust_managed_page_count(page
, pageblock_nr_pages
);
2118 * The order of subdivision here is critical for the IO subsystem.
2119 * Please do not alter this order without good reasons and regression
2120 * testing. Specifically, as large blocks of memory are subdivided,
2121 * the order in which smaller blocks are delivered depends on the order
2122 * they're subdivided in this function. This is the primary factor
2123 * influencing the order in which pages are delivered to the IO
2124 * subsystem according to empirical testing, and this is also justified
2125 * by considering the behavior of a buddy system containing a single
2126 * large block of memory acted on by a series of small allocations.
2127 * This behavior is a critical factor in sglist merging's success.
2131 static inline void expand(struct zone
*zone
, struct page
*page
,
2132 int low
, int high
, int migratetype
)
2134 unsigned long size
= 1 << high
;
2136 while (high
> low
) {
2139 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2142 * Mark as guard pages (or page), that will allow to
2143 * merge back to allocator when buddy will be freed.
2144 * Corresponding page table entries will not be touched,
2145 * pages will stay not present in virtual address space
2147 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2150 add_to_free_list(&page
[size
], zone
, high
, migratetype
);
2151 set_page_order(&page
[size
], high
);
2155 static void check_new_page_bad(struct page
*page
)
2157 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2158 /* Don't complain about hwpoisoned pages */
2159 page_mapcount_reset(page
); /* remove PageBuddy */
2164 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_PREP
));
2168 * This page is about to be returned from the page allocator
2170 static inline int check_new_page(struct page
*page
)
2172 if (likely(page_expected_state(page
,
2173 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2176 check_new_page_bad(page
);
2180 static inline bool free_pages_prezeroed(void)
2182 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
2183 page_poisoning_enabled()) || want_init_on_free();
2186 #ifdef CONFIG_DEBUG_VM
2188 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2189 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2190 * also checked when pcp lists are refilled from the free lists.
2192 static inline bool check_pcp_refill(struct page
*page
)
2194 if (debug_pagealloc_enabled_static())
2195 return check_new_page(page
);
2200 static inline bool check_new_pcp(struct page
*page
)
2202 return check_new_page(page
);
2206 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2207 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2208 * enabled, they are also checked when being allocated from the pcp lists.
2210 static inline bool check_pcp_refill(struct page
*page
)
2212 return check_new_page(page
);
2214 static inline bool check_new_pcp(struct page
*page
)
2216 if (debug_pagealloc_enabled_static())
2217 return check_new_page(page
);
2221 #endif /* CONFIG_DEBUG_VM */
2223 static bool check_new_pages(struct page
*page
, unsigned int order
)
2226 for (i
= 0; i
< (1 << order
); i
++) {
2227 struct page
*p
= page
+ i
;
2229 if (unlikely(check_new_page(p
)))
2236 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2239 set_page_private(page
, 0);
2240 set_page_refcounted(page
);
2242 arch_alloc_page(page
, order
);
2243 if (debug_pagealloc_enabled_static())
2244 kernel_map_pages(page
, 1 << order
, 1);
2245 kasan_alloc_pages(page
, order
);
2246 kernel_poison_pages(page
, 1 << order
, 1);
2247 set_page_owner(page
, order
, gfp_flags
);
2250 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2251 unsigned int alloc_flags
)
2253 post_alloc_hook(page
, order
, gfp_flags
);
2255 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags
))
2256 kernel_init_free_pages(page
, 1 << order
);
2258 if (order
&& (gfp_flags
& __GFP_COMP
))
2259 prep_compound_page(page
, order
);
2262 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2263 * allocate the page. The expectation is that the caller is taking
2264 * steps that will free more memory. The caller should avoid the page
2265 * being used for !PFMEMALLOC purposes.
2267 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2268 set_page_pfmemalloc(page
);
2270 clear_page_pfmemalloc(page
);
2274 * Go through the free lists for the given migratetype and remove
2275 * the smallest available page from the freelists
2277 static __always_inline
2278 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2281 unsigned int current_order
;
2282 struct free_area
*area
;
2285 /* Find a page of the appropriate size in the preferred list */
2286 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2287 area
= &(zone
->free_area
[current_order
]);
2288 page
= get_page_from_free_area(area
, migratetype
);
2291 del_page_from_free_list(page
, zone
, current_order
);
2292 expand(zone
, page
, order
, current_order
, migratetype
);
2293 set_pcppage_migratetype(page
, migratetype
);
2302 * This array describes the order lists are fallen back to when
2303 * the free lists for the desirable migrate type are depleted
2305 static int fallbacks
[MIGRATE_TYPES
][3] = {
2306 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2307 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2308 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2310 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2312 #ifdef CONFIG_MEMORY_ISOLATION
2313 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2318 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2321 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2324 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2325 unsigned int order
) { return NULL
; }
2329 * Move the free pages in a range to the free lists of the requested type.
2330 * Note that start_page and end_pages are not aligned on a pageblock
2331 * boundary. If alignment is required, use move_freepages_block()
2333 static int move_freepages(struct zone
*zone
,
2334 struct page
*start_page
, struct page
*end_page
,
2335 int migratetype
, int *num_movable
)
2339 int pages_moved
= 0;
2341 for (page
= start_page
; page
<= end_page
;) {
2342 if (!pfn_valid_within(page_to_pfn(page
))) {
2347 if (!PageBuddy(page
)) {
2349 * We assume that pages that could be isolated for
2350 * migration are movable. But we don't actually try
2351 * isolating, as that would be expensive.
2354 (PageLRU(page
) || __PageMovable(page
)))
2361 /* Make sure we are not inadvertently changing nodes */
2362 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2363 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2365 order
= page_order(page
);
2366 move_to_free_list(page
, zone
, order
, migratetype
);
2368 pages_moved
+= 1 << order
;
2374 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2375 int migratetype
, int *num_movable
)
2377 unsigned long start_pfn
, end_pfn
;
2378 struct page
*start_page
, *end_page
;
2383 start_pfn
= page_to_pfn(page
);
2384 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2385 start_page
= pfn_to_page(start_pfn
);
2386 end_page
= start_page
+ pageblock_nr_pages
- 1;
2387 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2389 /* Do not cross zone boundaries */
2390 if (!zone_spans_pfn(zone
, start_pfn
))
2392 if (!zone_spans_pfn(zone
, end_pfn
))
2395 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2399 static void change_pageblock_range(struct page
*pageblock_page
,
2400 int start_order
, int migratetype
)
2402 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2404 while (nr_pageblocks
--) {
2405 set_pageblock_migratetype(pageblock_page
, migratetype
);
2406 pageblock_page
+= pageblock_nr_pages
;
2411 * When we are falling back to another migratetype during allocation, try to
2412 * steal extra free pages from the same pageblocks to satisfy further
2413 * allocations, instead of polluting multiple pageblocks.
2415 * If we are stealing a relatively large buddy page, it is likely there will
2416 * be more free pages in the pageblock, so try to steal them all. For
2417 * reclaimable and unmovable allocations, we steal regardless of page size,
2418 * as fragmentation caused by those allocations polluting movable pageblocks
2419 * is worse than movable allocations stealing from unmovable and reclaimable
2422 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2425 * Leaving this order check is intended, although there is
2426 * relaxed order check in next check. The reason is that
2427 * we can actually steal whole pageblock if this condition met,
2428 * but, below check doesn't guarantee it and that is just heuristic
2429 * so could be changed anytime.
2431 if (order
>= pageblock_order
)
2434 if (order
>= pageblock_order
/ 2 ||
2435 start_mt
== MIGRATE_RECLAIMABLE
||
2436 start_mt
== MIGRATE_UNMOVABLE
||
2437 page_group_by_mobility_disabled
)
2443 static inline void boost_watermark(struct zone
*zone
)
2445 unsigned long max_boost
;
2447 if (!watermark_boost_factor
)
2450 * Don't bother in zones that are unlikely to produce results.
2451 * On small machines, including kdump capture kernels running
2452 * in a small area, boosting the watermark can cause an out of
2453 * memory situation immediately.
2455 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
2458 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2459 watermark_boost_factor
, 10000);
2462 * high watermark may be uninitialised if fragmentation occurs
2463 * very early in boot so do not boost. We do not fall
2464 * through and boost by pageblock_nr_pages as failing
2465 * allocations that early means that reclaim is not going
2466 * to help and it may even be impossible to reclaim the
2467 * boosted watermark resulting in a hang.
2472 max_boost
= max(pageblock_nr_pages
, max_boost
);
2474 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2479 * This function implements actual steal behaviour. If order is large enough,
2480 * we can steal whole pageblock. If not, we first move freepages in this
2481 * pageblock to our migratetype and determine how many already-allocated pages
2482 * are there in the pageblock with a compatible migratetype. If at least half
2483 * of pages are free or compatible, we can change migratetype of the pageblock
2484 * itself, so pages freed in the future will be put on the correct free list.
2486 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2487 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2489 unsigned int current_order
= page_order(page
);
2490 int free_pages
, movable_pages
, alike_pages
;
2493 old_block_type
= get_pageblock_migratetype(page
);
2496 * This can happen due to races and we want to prevent broken
2497 * highatomic accounting.
2499 if (is_migrate_highatomic(old_block_type
))
2502 /* Take ownership for orders >= pageblock_order */
2503 if (current_order
>= pageblock_order
) {
2504 change_pageblock_range(page
, current_order
, start_type
);
2509 * Boost watermarks to increase reclaim pressure to reduce the
2510 * likelihood of future fallbacks. Wake kswapd now as the node
2511 * may be balanced overall and kswapd will not wake naturally.
2513 boost_watermark(zone
);
2514 if (alloc_flags
& ALLOC_KSWAPD
)
2515 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2517 /* We are not allowed to try stealing from the whole block */
2521 free_pages
= move_freepages_block(zone
, page
, start_type
,
2524 * Determine how many pages are compatible with our allocation.
2525 * For movable allocation, it's the number of movable pages which
2526 * we just obtained. For other types it's a bit more tricky.
2528 if (start_type
== MIGRATE_MOVABLE
) {
2529 alike_pages
= movable_pages
;
2532 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2533 * to MOVABLE pageblock, consider all non-movable pages as
2534 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2535 * vice versa, be conservative since we can't distinguish the
2536 * exact migratetype of non-movable pages.
2538 if (old_block_type
== MIGRATE_MOVABLE
)
2539 alike_pages
= pageblock_nr_pages
2540 - (free_pages
+ movable_pages
);
2545 /* moving whole block can fail due to zone boundary conditions */
2550 * If a sufficient number of pages in the block are either free or of
2551 * comparable migratability as our allocation, claim the whole block.
2553 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2554 page_group_by_mobility_disabled
)
2555 set_pageblock_migratetype(page
, start_type
);
2560 move_to_free_list(page
, zone
, current_order
, start_type
);
2564 * Check whether there is a suitable fallback freepage with requested order.
2565 * If only_stealable is true, this function returns fallback_mt only if
2566 * we can steal other freepages all together. This would help to reduce
2567 * fragmentation due to mixed migratetype pages in one pageblock.
2569 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2570 int migratetype
, bool only_stealable
, bool *can_steal
)
2575 if (area
->nr_free
== 0)
2580 fallback_mt
= fallbacks
[migratetype
][i
];
2581 if (fallback_mt
== MIGRATE_TYPES
)
2584 if (free_area_empty(area
, fallback_mt
))
2587 if (can_steal_fallback(order
, migratetype
))
2590 if (!only_stealable
)
2601 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2602 * there are no empty page blocks that contain a page with a suitable order
2604 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2605 unsigned int alloc_order
)
2608 unsigned long max_managed
, flags
;
2611 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2612 * Check is race-prone but harmless.
2614 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2615 if (zone
->nr_reserved_highatomic
>= max_managed
)
2618 spin_lock_irqsave(&zone
->lock
, flags
);
2620 /* Recheck the nr_reserved_highatomic limit under the lock */
2621 if (zone
->nr_reserved_highatomic
>= max_managed
)
2625 mt
= get_pageblock_migratetype(page
);
2626 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2627 && !is_migrate_cma(mt
)) {
2628 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2629 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2630 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2634 spin_unlock_irqrestore(&zone
->lock
, flags
);
2638 * Used when an allocation is about to fail under memory pressure. This
2639 * potentially hurts the reliability of high-order allocations when under
2640 * intense memory pressure but failed atomic allocations should be easier
2641 * to recover from than an OOM.
2643 * If @force is true, try to unreserve a pageblock even though highatomic
2644 * pageblock is exhausted.
2646 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2649 struct zonelist
*zonelist
= ac
->zonelist
;
2650 unsigned long flags
;
2657 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->highest_zoneidx
,
2660 * Preserve at least one pageblock unless memory pressure
2663 if (!force
&& zone
->nr_reserved_highatomic
<=
2667 spin_lock_irqsave(&zone
->lock
, flags
);
2668 for (order
= 0; order
< MAX_ORDER
; order
++) {
2669 struct free_area
*area
= &(zone
->free_area
[order
]);
2671 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2676 * In page freeing path, migratetype change is racy so
2677 * we can counter several free pages in a pageblock
2678 * in this loop althoug we changed the pageblock type
2679 * from highatomic to ac->migratetype. So we should
2680 * adjust the count once.
2682 if (is_migrate_highatomic_page(page
)) {
2684 * It should never happen but changes to
2685 * locking could inadvertently allow a per-cpu
2686 * drain to add pages to MIGRATE_HIGHATOMIC
2687 * while unreserving so be safe and watch for
2690 zone
->nr_reserved_highatomic
-= min(
2692 zone
->nr_reserved_highatomic
);
2696 * Convert to ac->migratetype and avoid the normal
2697 * pageblock stealing heuristics. Minimally, the caller
2698 * is doing the work and needs the pages. More
2699 * importantly, if the block was always converted to
2700 * MIGRATE_UNMOVABLE or another type then the number
2701 * of pageblocks that cannot be completely freed
2704 set_pageblock_migratetype(page
, ac
->migratetype
);
2705 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2708 spin_unlock_irqrestore(&zone
->lock
, flags
);
2712 spin_unlock_irqrestore(&zone
->lock
, flags
);
2719 * Try finding a free buddy page on the fallback list and put it on the free
2720 * list of requested migratetype, possibly along with other pages from the same
2721 * block, depending on fragmentation avoidance heuristics. Returns true if
2722 * fallback was found so that __rmqueue_smallest() can grab it.
2724 * The use of signed ints for order and current_order is a deliberate
2725 * deviation from the rest of this file, to make the for loop
2726 * condition simpler.
2728 static __always_inline
bool
2729 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2730 unsigned int alloc_flags
)
2732 struct free_area
*area
;
2734 int min_order
= order
;
2740 * Do not steal pages from freelists belonging to other pageblocks
2741 * i.e. orders < pageblock_order. If there are no local zones free,
2742 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2744 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2745 min_order
= pageblock_order
;
2748 * Find the largest available free page in the other list. This roughly
2749 * approximates finding the pageblock with the most free pages, which
2750 * would be too costly to do exactly.
2752 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2754 area
= &(zone
->free_area
[current_order
]);
2755 fallback_mt
= find_suitable_fallback(area
, current_order
,
2756 start_migratetype
, false, &can_steal
);
2757 if (fallback_mt
== -1)
2761 * We cannot steal all free pages from the pageblock and the
2762 * requested migratetype is movable. In that case it's better to
2763 * steal and split the smallest available page instead of the
2764 * largest available page, because even if the next movable
2765 * allocation falls back into a different pageblock than this
2766 * one, it won't cause permanent fragmentation.
2768 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2769 && current_order
> order
)
2778 for (current_order
= order
; current_order
< MAX_ORDER
;
2780 area
= &(zone
->free_area
[current_order
]);
2781 fallback_mt
= find_suitable_fallback(area
, current_order
,
2782 start_migratetype
, false, &can_steal
);
2783 if (fallback_mt
!= -1)
2788 * This should not happen - we already found a suitable fallback
2789 * when looking for the largest page.
2791 VM_BUG_ON(current_order
== MAX_ORDER
);
2794 page
= get_page_from_free_area(area
, fallback_mt
);
2796 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2799 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2800 start_migratetype
, fallback_mt
);
2807 * Do the hard work of removing an element from the buddy allocator.
2808 * Call me with the zone->lock already held.
2810 static __always_inline
struct page
*
2811 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2812 unsigned int alloc_flags
)
2818 * Balance movable allocations between regular and CMA areas by
2819 * allocating from CMA when over half of the zone's free memory
2820 * is in the CMA area.
2822 if (alloc_flags
& ALLOC_CMA
&&
2823 zone_page_state(zone
, NR_FREE_CMA_PAGES
) >
2824 zone_page_state(zone
, NR_FREE_PAGES
) / 2) {
2825 page
= __rmqueue_cma_fallback(zone
, order
);
2831 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2832 if (unlikely(!page
)) {
2833 if (alloc_flags
& ALLOC_CMA
)
2834 page
= __rmqueue_cma_fallback(zone
, order
);
2836 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2841 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2846 * Obtain a specified number of elements from the buddy allocator, all under
2847 * a single hold of the lock, for efficiency. Add them to the supplied list.
2848 * Returns the number of new pages which were placed at *list.
2850 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2851 unsigned long count
, struct list_head
*list
,
2852 int migratetype
, unsigned int alloc_flags
)
2856 spin_lock(&zone
->lock
);
2857 for (i
= 0; i
< count
; ++i
) {
2858 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2860 if (unlikely(page
== NULL
))
2863 if (unlikely(check_pcp_refill(page
)))
2867 * Split buddy pages returned by expand() are received here in
2868 * physical page order. The page is added to the tail of
2869 * caller's list. From the callers perspective, the linked list
2870 * is ordered by page number under some conditions. This is
2871 * useful for IO devices that can forward direction from the
2872 * head, thus also in the physical page order. This is useful
2873 * for IO devices that can merge IO requests if the physical
2874 * pages are ordered properly.
2876 list_add_tail(&page
->lru
, list
);
2878 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2879 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2884 * i pages were removed from the buddy list even if some leak due
2885 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2886 * on i. Do not confuse with 'alloced' which is the number of
2887 * pages added to the pcp list.
2889 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2890 spin_unlock(&zone
->lock
);
2896 * Called from the vmstat counter updater to drain pagesets of this
2897 * currently executing processor on remote nodes after they have
2900 * Note that this function must be called with the thread pinned to
2901 * a single processor.
2903 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2905 unsigned long flags
;
2906 int to_drain
, batch
;
2908 local_irq_save(flags
);
2909 batch
= READ_ONCE(pcp
->batch
);
2910 to_drain
= min(pcp
->count
, batch
);
2912 free_pcppages_bulk(zone
, to_drain
, pcp
);
2913 local_irq_restore(flags
);
2918 * Drain pcplists of the indicated processor and zone.
2920 * The processor must either be the current processor and the
2921 * thread pinned to the current processor or a processor that
2924 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2926 unsigned long flags
;
2927 struct per_cpu_pageset
*pset
;
2928 struct per_cpu_pages
*pcp
;
2930 local_irq_save(flags
);
2931 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2935 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2936 local_irq_restore(flags
);
2940 * Drain pcplists of all zones on the indicated processor.
2942 * The processor must either be the current processor and the
2943 * thread pinned to the current processor or a processor that
2946 static void drain_pages(unsigned int cpu
)
2950 for_each_populated_zone(zone
) {
2951 drain_pages_zone(cpu
, zone
);
2956 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2958 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2959 * the single zone's pages.
2961 void drain_local_pages(struct zone
*zone
)
2963 int cpu
= smp_processor_id();
2966 drain_pages_zone(cpu
, zone
);
2971 static void drain_local_pages_wq(struct work_struct
*work
)
2973 struct pcpu_drain
*drain
;
2975 drain
= container_of(work
, struct pcpu_drain
, work
);
2978 * drain_all_pages doesn't use proper cpu hotplug protection so
2979 * we can race with cpu offline when the WQ can move this from
2980 * a cpu pinned worker to an unbound one. We can operate on a different
2981 * cpu which is allright but we also have to make sure to not move to
2985 drain_local_pages(drain
->zone
);
2990 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2992 * When zone parameter is non-NULL, spill just the single zone's pages.
2994 * Note that this can be extremely slow as the draining happens in a workqueue.
2996 void drain_all_pages(struct zone
*zone
)
3001 * Allocate in the BSS so we wont require allocation in
3002 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3004 static cpumask_t cpus_with_pcps
;
3007 * Make sure nobody triggers this path before mm_percpu_wq is fully
3010 if (WARN_ON_ONCE(!mm_percpu_wq
))
3014 * Do not drain if one is already in progress unless it's specific to
3015 * a zone. Such callers are primarily CMA and memory hotplug and need
3016 * the drain to be complete when the call returns.
3018 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
3021 mutex_lock(&pcpu_drain_mutex
);
3025 * We don't care about racing with CPU hotplug event
3026 * as offline notification will cause the notified
3027 * cpu to drain that CPU pcps and on_each_cpu_mask
3028 * disables preemption as part of its processing
3030 for_each_online_cpu(cpu
) {
3031 struct per_cpu_pageset
*pcp
;
3033 bool has_pcps
= false;
3036 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3040 for_each_populated_zone(z
) {
3041 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
3042 if (pcp
->pcp
.count
) {
3050 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
3052 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
3055 for_each_cpu(cpu
, &cpus_with_pcps
) {
3056 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
3059 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
3060 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
3062 for_each_cpu(cpu
, &cpus_with_pcps
)
3063 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
3065 mutex_unlock(&pcpu_drain_mutex
);
3068 #ifdef CONFIG_HIBERNATION
3071 * Touch the watchdog for every WD_PAGE_COUNT pages.
3073 #define WD_PAGE_COUNT (128*1024)
3075 void mark_free_pages(struct zone
*zone
)
3077 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
3078 unsigned long flags
;
3079 unsigned int order
, t
;
3082 if (zone_is_empty(zone
))
3085 spin_lock_irqsave(&zone
->lock
, flags
);
3087 max_zone_pfn
= zone_end_pfn(zone
);
3088 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
3089 if (pfn_valid(pfn
)) {
3090 page
= pfn_to_page(pfn
);
3092 if (!--page_count
) {
3093 touch_nmi_watchdog();
3094 page_count
= WD_PAGE_COUNT
;
3097 if (page_zone(page
) != zone
)
3100 if (!swsusp_page_is_forbidden(page
))
3101 swsusp_unset_page_free(page
);
3104 for_each_migratetype_order(order
, t
) {
3105 list_for_each_entry(page
,
3106 &zone
->free_area
[order
].free_list
[t
], lru
) {
3109 pfn
= page_to_pfn(page
);
3110 for (i
= 0; i
< (1UL << order
); i
++) {
3111 if (!--page_count
) {
3112 touch_nmi_watchdog();
3113 page_count
= WD_PAGE_COUNT
;
3115 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
3119 spin_unlock_irqrestore(&zone
->lock
, flags
);
3121 #endif /* CONFIG_PM */
3123 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
3127 if (!free_pcp_prepare(page
))
3130 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3131 set_pcppage_migratetype(page
, migratetype
);
3135 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
3137 struct zone
*zone
= page_zone(page
);
3138 struct per_cpu_pages
*pcp
;
3141 migratetype
= get_pcppage_migratetype(page
);
3142 __count_vm_event(PGFREE
);
3145 * We only track unmovable, reclaimable and movable on pcp lists.
3146 * Free ISOLATE pages back to the allocator because they are being
3147 * offlined but treat HIGHATOMIC as movable pages so we can get those
3148 * areas back if necessary. Otherwise, we may have to free
3149 * excessively into the page allocator
3151 if (migratetype
>= MIGRATE_PCPTYPES
) {
3152 if (unlikely(is_migrate_isolate(migratetype
))) {
3153 free_one_page(zone
, page
, pfn
, 0, migratetype
);
3156 migratetype
= MIGRATE_MOVABLE
;
3159 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3160 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
3162 if (pcp
->count
>= pcp
->high
) {
3163 unsigned long batch
= READ_ONCE(pcp
->batch
);
3164 free_pcppages_bulk(zone
, batch
, pcp
);
3169 * Free a 0-order page
3171 void free_unref_page(struct page
*page
)
3173 unsigned long flags
;
3174 unsigned long pfn
= page_to_pfn(page
);
3176 if (!free_unref_page_prepare(page
, pfn
))
3179 local_irq_save(flags
);
3180 free_unref_page_commit(page
, pfn
);
3181 local_irq_restore(flags
);
3185 * Free a list of 0-order pages
3187 void free_unref_page_list(struct list_head
*list
)
3189 struct page
*page
, *next
;
3190 unsigned long flags
, pfn
;
3191 int batch_count
= 0;
3193 /* Prepare pages for freeing */
3194 list_for_each_entry_safe(page
, next
, list
, lru
) {
3195 pfn
= page_to_pfn(page
);
3196 if (!free_unref_page_prepare(page
, pfn
))
3197 list_del(&page
->lru
);
3198 set_page_private(page
, pfn
);
3201 local_irq_save(flags
);
3202 list_for_each_entry_safe(page
, next
, list
, lru
) {
3203 unsigned long pfn
= page_private(page
);
3205 set_page_private(page
, 0);
3206 trace_mm_page_free_batched(page
);
3207 free_unref_page_commit(page
, pfn
);
3210 * Guard against excessive IRQ disabled times when we get
3211 * a large list of pages to free.
3213 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3214 local_irq_restore(flags
);
3216 local_irq_save(flags
);
3219 local_irq_restore(flags
);
3223 * split_page takes a non-compound higher-order page, and splits it into
3224 * n (1<<order) sub-pages: page[0..n]
3225 * Each sub-page must be freed individually.
3227 * Note: this is probably too low level an operation for use in drivers.
3228 * Please consult with lkml before using this in your driver.
3230 void split_page(struct page
*page
, unsigned int order
)
3234 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3235 VM_BUG_ON_PAGE(!page_count(page
), page
);
3237 for (i
= 1; i
< (1 << order
); i
++)
3238 set_page_refcounted(page
+ i
);
3239 split_page_owner(page
, 1 << order
);
3241 EXPORT_SYMBOL_GPL(split_page
);
3243 int __isolate_free_page(struct page
*page
, unsigned int order
)
3245 unsigned long watermark
;
3249 BUG_ON(!PageBuddy(page
));
3251 zone
= page_zone(page
);
3252 mt
= get_pageblock_migratetype(page
);
3254 if (!is_migrate_isolate(mt
)) {
3256 * Obey watermarks as if the page was being allocated. We can
3257 * emulate a high-order watermark check with a raised order-0
3258 * watermark, because we already know our high-order page
3261 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3262 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3265 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3268 /* Remove page from free list */
3270 del_page_from_free_list(page
, zone
, order
);
3273 * Set the pageblock if the isolated page is at least half of a
3276 if (order
>= pageblock_order
- 1) {
3277 struct page
*endpage
= page
+ (1 << order
) - 1;
3278 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3279 int mt
= get_pageblock_migratetype(page
);
3280 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3281 && !is_migrate_highatomic(mt
))
3282 set_pageblock_migratetype(page
,
3288 return 1UL << order
;
3292 * __putback_isolated_page - Return a now-isolated page back where we got it
3293 * @page: Page that was isolated
3294 * @order: Order of the isolated page
3295 * @mt: The page's pageblock's migratetype
3297 * This function is meant to return a page pulled from the free lists via
3298 * __isolate_free_page back to the free lists they were pulled from.
3300 void __putback_isolated_page(struct page
*page
, unsigned int order
, int mt
)
3302 struct zone
*zone
= page_zone(page
);
3304 /* zone lock should be held when this function is called */
3305 lockdep_assert_held(&zone
->lock
);
3307 /* Return isolated page to tail of freelist. */
3308 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
,
3309 FPI_SKIP_REPORT_NOTIFY
);
3313 * Update NUMA hit/miss statistics
3315 * Must be called with interrupts disabled.
3317 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
3320 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3322 /* skip numa counters update if numa stats is disabled */
3323 if (!static_branch_likely(&vm_numa_stat_key
))
3326 if (zone_to_nid(z
) != numa_node_id())
3327 local_stat
= NUMA_OTHER
;
3329 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3330 __inc_numa_state(z
, NUMA_HIT
);
3332 __inc_numa_state(z
, NUMA_MISS
);
3333 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3335 __inc_numa_state(z
, local_stat
);
3339 /* Remove page from the per-cpu list, caller must protect the list */
3340 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3341 unsigned int alloc_flags
,
3342 struct per_cpu_pages
*pcp
,
3343 struct list_head
*list
)
3348 if (list_empty(list
)) {
3349 pcp
->count
+= rmqueue_bulk(zone
, 0,
3351 migratetype
, alloc_flags
);
3352 if (unlikely(list_empty(list
)))
3356 page
= list_first_entry(list
, struct page
, lru
);
3357 list_del(&page
->lru
);
3359 } while (check_new_pcp(page
));
3364 /* Lock and remove page from the per-cpu list */
3365 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3366 struct zone
*zone
, gfp_t gfp_flags
,
3367 int migratetype
, unsigned int alloc_flags
)
3369 struct per_cpu_pages
*pcp
;
3370 struct list_head
*list
;
3372 unsigned long flags
;
3374 local_irq_save(flags
);
3375 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3376 list
= &pcp
->lists
[migratetype
];
3377 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3379 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3380 zone_statistics(preferred_zone
, zone
);
3382 local_irq_restore(flags
);
3387 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3390 struct page
*rmqueue(struct zone
*preferred_zone
,
3391 struct zone
*zone
, unsigned int order
,
3392 gfp_t gfp_flags
, unsigned int alloc_flags
,
3395 unsigned long flags
;
3398 if (likely(order
== 0)) {
3400 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3401 * we need to skip it when CMA area isn't allowed.
3403 if (!IS_ENABLED(CONFIG_CMA
) || alloc_flags
& ALLOC_CMA
||
3404 migratetype
!= MIGRATE_MOVABLE
) {
3405 page
= rmqueue_pcplist(preferred_zone
, zone
, gfp_flags
,
3406 migratetype
, alloc_flags
);
3412 * We most definitely don't want callers attempting to
3413 * allocate greater than order-1 page units with __GFP_NOFAIL.
3415 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3416 spin_lock_irqsave(&zone
->lock
, flags
);
3421 * order-0 request can reach here when the pcplist is skipped
3422 * due to non-CMA allocation context. HIGHATOMIC area is
3423 * reserved for high-order atomic allocation, so order-0
3424 * request should skip it.
3426 if (order
> 0 && alloc_flags
& ALLOC_HARDER
) {
3427 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3429 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3432 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3433 } while (page
&& check_new_pages(page
, order
));
3434 spin_unlock(&zone
->lock
);
3437 __mod_zone_freepage_state(zone
, -(1 << order
),
3438 get_pcppage_migratetype(page
));
3440 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3441 zone_statistics(preferred_zone
, zone
);
3442 local_irq_restore(flags
);
3445 /* Separate test+clear to avoid unnecessary atomics */
3446 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3447 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3448 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3451 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3455 local_irq_restore(flags
);
3459 #ifdef CONFIG_FAIL_PAGE_ALLOC
3462 struct fault_attr attr
;
3464 bool ignore_gfp_highmem
;
3465 bool ignore_gfp_reclaim
;
3467 } fail_page_alloc
= {
3468 .attr
= FAULT_ATTR_INITIALIZER
,
3469 .ignore_gfp_reclaim
= true,
3470 .ignore_gfp_highmem
= true,
3474 static int __init
setup_fail_page_alloc(char *str
)
3476 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3478 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3480 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3482 if (order
< fail_page_alloc
.min_order
)
3484 if (gfp_mask
& __GFP_NOFAIL
)
3486 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3488 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3489 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3492 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3495 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3497 static int __init
fail_page_alloc_debugfs(void)
3499 umode_t mode
= S_IFREG
| 0600;
3502 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3503 &fail_page_alloc
.attr
);
3505 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3506 &fail_page_alloc
.ignore_gfp_reclaim
);
3507 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3508 &fail_page_alloc
.ignore_gfp_highmem
);
3509 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3514 late_initcall(fail_page_alloc_debugfs
);
3516 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3518 #else /* CONFIG_FAIL_PAGE_ALLOC */
3520 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3525 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3527 static noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3529 return __should_fail_alloc_page(gfp_mask
, order
);
3531 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3533 static inline long __zone_watermark_unusable_free(struct zone
*z
,
3534 unsigned int order
, unsigned int alloc_flags
)
3536 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3537 long unusable_free
= (1 << order
) - 1;
3540 * If the caller does not have rights to ALLOC_HARDER then subtract
3541 * the high-atomic reserves. This will over-estimate the size of the
3542 * atomic reserve but it avoids a search.
3544 if (likely(!alloc_harder
))
3545 unusable_free
+= z
->nr_reserved_highatomic
;
3548 /* If allocation can't use CMA areas don't use free CMA pages */
3549 if (!(alloc_flags
& ALLOC_CMA
))
3550 unusable_free
+= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3553 return unusable_free
;
3557 * Return true if free base pages are above 'mark'. For high-order checks it
3558 * will return true of the order-0 watermark is reached and there is at least
3559 * one free page of a suitable size. Checking now avoids taking the zone lock
3560 * to check in the allocation paths if no pages are free.
3562 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3563 int highest_zoneidx
, unsigned int alloc_flags
,
3568 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3570 /* free_pages may go negative - that's OK */
3571 free_pages
-= __zone_watermark_unusable_free(z
, order
, alloc_flags
);
3573 if (alloc_flags
& ALLOC_HIGH
)
3576 if (unlikely(alloc_harder
)) {
3578 * OOM victims can try even harder than normal ALLOC_HARDER
3579 * users on the grounds that it's definitely going to be in
3580 * the exit path shortly and free memory. Any allocation it
3581 * makes during the free path will be small and short-lived.
3583 if (alloc_flags
& ALLOC_OOM
)
3590 * Check watermarks for an order-0 allocation request. If these
3591 * are not met, then a high-order request also cannot go ahead
3592 * even if a suitable page happened to be free.
3594 if (free_pages
<= min
+ z
->lowmem_reserve
[highest_zoneidx
])
3597 /* If this is an order-0 request then the watermark is fine */
3601 /* For a high-order request, check at least one suitable page is free */
3602 for (o
= order
; o
< MAX_ORDER
; o
++) {
3603 struct free_area
*area
= &z
->free_area
[o
];
3609 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3610 if (!free_area_empty(area
, mt
))
3615 if ((alloc_flags
& ALLOC_CMA
) &&
3616 !free_area_empty(area
, MIGRATE_CMA
)) {
3620 if (alloc_harder
&& !free_area_empty(area
, MIGRATE_HIGHATOMIC
))
3626 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3627 int highest_zoneidx
, unsigned int alloc_flags
)
3629 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3630 zone_page_state(z
, NR_FREE_PAGES
));
3633 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3634 unsigned long mark
, int highest_zoneidx
,
3635 unsigned int alloc_flags
, gfp_t gfp_mask
)
3639 free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3642 * Fast check for order-0 only. If this fails then the reserves
3643 * need to be calculated.
3648 fast_free
= free_pages
;
3649 fast_free
-= __zone_watermark_unusable_free(z
, 0, alloc_flags
);
3650 if (fast_free
> mark
+ z
->lowmem_reserve
[highest_zoneidx
])
3654 if (__zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3658 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3659 * when checking the min watermark. The min watermark is the
3660 * point where boosting is ignored so that kswapd is woken up
3661 * when below the low watermark.
3663 if (unlikely(!order
&& (gfp_mask
& __GFP_ATOMIC
) && z
->watermark_boost
3664 && ((alloc_flags
& ALLOC_WMARK_MASK
) == WMARK_MIN
))) {
3665 mark
= z
->_watermark
[WMARK_MIN
];
3666 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
,
3667 alloc_flags
, free_pages
);
3673 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3674 unsigned long mark
, int highest_zoneidx
)
3676 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3678 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3679 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3681 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, 0,
3686 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3688 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3689 node_reclaim_distance
;
3691 #else /* CONFIG_NUMA */
3692 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3696 #endif /* CONFIG_NUMA */
3699 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3700 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3701 * premature use of a lower zone may cause lowmem pressure problems that
3702 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3703 * probably too small. It only makes sense to spread allocations to avoid
3704 * fragmentation between the Normal and DMA32 zones.
3706 static inline unsigned int
3707 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3709 unsigned int alloc_flags
;
3712 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3715 alloc_flags
= (__force
int) (gfp_mask
& __GFP_KSWAPD_RECLAIM
);
3717 #ifdef CONFIG_ZONE_DMA32
3721 if (zone_idx(zone
) != ZONE_NORMAL
)
3725 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3726 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3727 * on UMA that if Normal is populated then so is DMA32.
3729 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3730 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3733 alloc_flags
|= ALLOC_NOFRAGMENT
;
3734 #endif /* CONFIG_ZONE_DMA32 */
3738 static inline unsigned int current_alloc_flags(gfp_t gfp_mask
,
3739 unsigned int alloc_flags
)
3742 unsigned int pflags
= current
->flags
;
3744 if (!(pflags
& PF_MEMALLOC_NOCMA
) &&
3745 gfp_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3746 alloc_flags
|= ALLOC_CMA
;
3753 * get_page_from_freelist goes through the zonelist trying to allocate
3756 static struct page
*
3757 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3758 const struct alloc_context
*ac
)
3762 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3767 * Scan zonelist, looking for a zone with enough free.
3768 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3770 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3771 z
= ac
->preferred_zoneref
;
3772 for_next_zone_zonelist_nodemask(zone
, z
, ac
->highest_zoneidx
,
3777 if (cpusets_enabled() &&
3778 (alloc_flags
& ALLOC_CPUSET
) &&
3779 !__cpuset_zone_allowed(zone
, gfp_mask
))
3782 * When allocating a page cache page for writing, we
3783 * want to get it from a node that is within its dirty
3784 * limit, such that no single node holds more than its
3785 * proportional share of globally allowed dirty pages.
3786 * The dirty limits take into account the node's
3787 * lowmem reserves and high watermark so that kswapd
3788 * should be able to balance it without having to
3789 * write pages from its LRU list.
3791 * XXX: For now, allow allocations to potentially
3792 * exceed the per-node dirty limit in the slowpath
3793 * (spread_dirty_pages unset) before going into reclaim,
3794 * which is important when on a NUMA setup the allowed
3795 * nodes are together not big enough to reach the
3796 * global limit. The proper fix for these situations
3797 * will require awareness of nodes in the
3798 * dirty-throttling and the flusher threads.
3800 if (ac
->spread_dirty_pages
) {
3801 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3804 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3805 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3810 if (no_fallback
&& nr_online_nodes
> 1 &&
3811 zone
!= ac
->preferred_zoneref
->zone
) {
3815 * If moving to a remote node, retry but allow
3816 * fragmenting fallbacks. Locality is more important
3817 * than fragmentation avoidance.
3819 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3820 if (zone_to_nid(zone
) != local_nid
) {
3821 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3826 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3827 if (!zone_watermark_fast(zone
, order
, mark
,
3828 ac
->highest_zoneidx
, alloc_flags
,
3832 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3834 * Watermark failed for this zone, but see if we can
3835 * grow this zone if it contains deferred pages.
3837 if (static_branch_unlikely(&deferred_pages
)) {
3838 if (_deferred_grow_zone(zone
, order
))
3842 /* Checked here to keep the fast path fast */
3843 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3844 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3847 if (node_reclaim_mode
== 0 ||
3848 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3851 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3853 case NODE_RECLAIM_NOSCAN
:
3856 case NODE_RECLAIM_FULL
:
3857 /* scanned but unreclaimable */
3860 /* did we reclaim enough */
3861 if (zone_watermark_ok(zone
, order
, mark
,
3862 ac
->highest_zoneidx
, alloc_flags
))
3870 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3871 gfp_mask
, alloc_flags
, ac
->migratetype
);
3873 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3876 * If this is a high-order atomic allocation then check
3877 * if the pageblock should be reserved for the future
3879 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3880 reserve_highatomic_pageblock(page
, zone
, order
);
3884 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3885 /* Try again if zone has deferred pages */
3886 if (static_branch_unlikely(&deferred_pages
)) {
3887 if (_deferred_grow_zone(zone
, order
))
3895 * It's possible on a UMA machine to get through all zones that are
3896 * fragmented. If avoiding fragmentation, reset and try again.
3899 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3906 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3908 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3911 * This documents exceptions given to allocations in certain
3912 * contexts that are allowed to allocate outside current's set
3915 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3916 if (tsk_is_oom_victim(current
) ||
3917 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3918 filter
&= ~SHOW_MEM_FILTER_NODES
;
3919 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3920 filter
&= ~SHOW_MEM_FILTER_NODES
;
3922 show_mem(filter
, nodemask
);
3925 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3927 struct va_format vaf
;
3929 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
3931 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3934 va_start(args
, fmt
);
3937 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3938 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3939 nodemask_pr_args(nodemask
));
3942 cpuset_print_current_mems_allowed();
3945 warn_alloc_show_mem(gfp_mask
, nodemask
);
3948 static inline struct page
*
3949 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3950 unsigned int alloc_flags
,
3951 const struct alloc_context
*ac
)
3955 page
= get_page_from_freelist(gfp_mask
, order
,
3956 alloc_flags
|ALLOC_CPUSET
, ac
);
3958 * fallback to ignore cpuset restriction if our nodes
3962 page
= get_page_from_freelist(gfp_mask
, order
,
3968 static inline struct page
*
3969 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3970 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3972 struct oom_control oc
= {
3973 .zonelist
= ac
->zonelist
,
3974 .nodemask
= ac
->nodemask
,
3976 .gfp_mask
= gfp_mask
,
3981 *did_some_progress
= 0;
3984 * Acquire the oom lock. If that fails, somebody else is
3985 * making progress for us.
3987 if (!mutex_trylock(&oom_lock
)) {
3988 *did_some_progress
= 1;
3989 schedule_timeout_uninterruptible(1);
3994 * Go through the zonelist yet one more time, keep very high watermark
3995 * here, this is only to catch a parallel oom killing, we must fail if
3996 * we're still under heavy pressure. But make sure that this reclaim
3997 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3998 * allocation which will never fail due to oom_lock already held.
4000 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
4001 ~__GFP_DIRECT_RECLAIM
, order
,
4002 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
4006 /* Coredumps can quickly deplete all memory reserves */
4007 if (current
->flags
& PF_DUMPCORE
)
4009 /* The OOM killer will not help higher order allocs */
4010 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4013 * We have already exhausted all our reclaim opportunities without any
4014 * success so it is time to admit defeat. We will skip the OOM killer
4015 * because it is very likely that the caller has a more reasonable
4016 * fallback than shooting a random task.
4018 * The OOM killer may not free memory on a specific node.
4020 if (gfp_mask
& (__GFP_RETRY_MAYFAIL
| __GFP_THISNODE
))
4022 /* The OOM killer does not needlessly kill tasks for lowmem */
4023 if (ac
->highest_zoneidx
< ZONE_NORMAL
)
4025 if (pm_suspended_storage())
4028 * XXX: GFP_NOFS allocations should rather fail than rely on
4029 * other request to make a forward progress.
4030 * We are in an unfortunate situation where out_of_memory cannot
4031 * do much for this context but let's try it to at least get
4032 * access to memory reserved if the current task is killed (see
4033 * out_of_memory). Once filesystems are ready to handle allocation
4034 * failures more gracefully we should just bail out here.
4037 /* Exhausted what can be done so it's blame time */
4038 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
4039 *did_some_progress
= 1;
4042 * Help non-failing allocations by giving them access to memory
4045 if (gfp_mask
& __GFP_NOFAIL
)
4046 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
4047 ALLOC_NO_WATERMARKS
, ac
);
4050 mutex_unlock(&oom_lock
);
4055 * Maximum number of compaction retries wit a progress before OOM
4056 * killer is consider as the only way to move forward.
4058 #define MAX_COMPACT_RETRIES 16
4060 #ifdef CONFIG_COMPACTION
4061 /* Try memory compaction for high-order allocations before reclaim */
4062 static struct page
*
4063 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4064 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4065 enum compact_priority prio
, enum compact_result
*compact_result
)
4067 struct page
*page
= NULL
;
4068 unsigned long pflags
;
4069 unsigned int noreclaim_flag
;
4074 psi_memstall_enter(&pflags
);
4075 noreclaim_flag
= memalloc_noreclaim_save();
4077 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
4080 memalloc_noreclaim_restore(noreclaim_flag
);
4081 psi_memstall_leave(&pflags
);
4084 * At least in one zone compaction wasn't deferred or skipped, so let's
4085 * count a compaction stall
4087 count_vm_event(COMPACTSTALL
);
4089 /* Prep a captured page if available */
4091 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
4093 /* Try get a page from the freelist if available */
4095 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4098 struct zone
*zone
= page_zone(page
);
4100 zone
->compact_blockskip_flush
= false;
4101 compaction_defer_reset(zone
, order
, true);
4102 count_vm_event(COMPACTSUCCESS
);
4107 * It's bad if compaction run occurs and fails. The most likely reason
4108 * is that pages exist, but not enough to satisfy watermarks.
4110 count_vm_event(COMPACTFAIL
);
4118 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
4119 enum compact_result compact_result
,
4120 enum compact_priority
*compact_priority
,
4121 int *compaction_retries
)
4123 int max_retries
= MAX_COMPACT_RETRIES
;
4126 int retries
= *compaction_retries
;
4127 enum compact_priority priority
= *compact_priority
;
4132 if (compaction_made_progress(compact_result
))
4133 (*compaction_retries
)++;
4136 * compaction considers all the zone as desperately out of memory
4137 * so it doesn't really make much sense to retry except when the
4138 * failure could be caused by insufficient priority
4140 if (compaction_failed(compact_result
))
4141 goto check_priority
;
4144 * compaction was skipped because there are not enough order-0 pages
4145 * to work with, so we retry only if it looks like reclaim can help.
4147 if (compaction_needs_reclaim(compact_result
)) {
4148 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
4153 * make sure the compaction wasn't deferred or didn't bail out early
4154 * due to locks contention before we declare that we should give up.
4155 * But the next retry should use a higher priority if allowed, so
4156 * we don't just keep bailing out endlessly.
4158 if (compaction_withdrawn(compact_result
)) {
4159 goto check_priority
;
4163 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4164 * costly ones because they are de facto nofail and invoke OOM
4165 * killer to move on while costly can fail and users are ready
4166 * to cope with that. 1/4 retries is rather arbitrary but we
4167 * would need much more detailed feedback from compaction to
4168 * make a better decision.
4170 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4172 if (*compaction_retries
<= max_retries
) {
4178 * Make sure there are attempts at the highest priority if we exhausted
4179 * all retries or failed at the lower priorities.
4182 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
4183 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
4185 if (*compact_priority
> min_priority
) {
4186 (*compact_priority
)--;
4187 *compaction_retries
= 0;
4191 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4195 static inline struct page
*
4196 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4197 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4198 enum compact_priority prio
, enum compact_result
*compact_result
)
4200 *compact_result
= COMPACT_SKIPPED
;
4205 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4206 enum compact_result compact_result
,
4207 enum compact_priority
*compact_priority
,
4208 int *compaction_retries
)
4213 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4217 * There are setups with compaction disabled which would prefer to loop
4218 * inside the allocator rather than hit the oom killer prematurely.
4219 * Let's give them a good hope and keep retrying while the order-0
4220 * watermarks are OK.
4222 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4223 ac
->highest_zoneidx
, ac
->nodemask
) {
4224 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4225 ac
->highest_zoneidx
, alloc_flags
))
4230 #endif /* CONFIG_COMPACTION */
4232 #ifdef CONFIG_LOCKDEP
4233 static struct lockdep_map __fs_reclaim_map
=
4234 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4236 static bool __need_fs_reclaim(gfp_t gfp_mask
)
4238 gfp_mask
= current_gfp_context(gfp_mask
);
4240 /* no reclaim without waiting on it */
4241 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4244 /* this guy won't enter reclaim */
4245 if (current
->flags
& PF_MEMALLOC
)
4248 /* We're only interested __GFP_FS allocations for now */
4249 if (!(gfp_mask
& __GFP_FS
))
4252 if (gfp_mask
& __GFP_NOLOCKDEP
)
4258 void __fs_reclaim_acquire(void)
4260 lock_map_acquire(&__fs_reclaim_map
);
4263 void __fs_reclaim_release(void)
4265 lock_map_release(&__fs_reclaim_map
);
4268 void fs_reclaim_acquire(gfp_t gfp_mask
)
4270 if (__need_fs_reclaim(gfp_mask
))
4271 __fs_reclaim_acquire();
4273 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4275 void fs_reclaim_release(gfp_t gfp_mask
)
4277 if (__need_fs_reclaim(gfp_mask
))
4278 __fs_reclaim_release();
4280 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4283 /* Perform direct synchronous page reclaim */
4284 static unsigned long
4285 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4286 const struct alloc_context
*ac
)
4288 unsigned int noreclaim_flag
;
4289 unsigned long pflags
, progress
;
4293 /* We now go into synchronous reclaim */
4294 cpuset_memory_pressure_bump();
4295 psi_memstall_enter(&pflags
);
4296 fs_reclaim_acquire(gfp_mask
);
4297 noreclaim_flag
= memalloc_noreclaim_save();
4299 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4302 memalloc_noreclaim_restore(noreclaim_flag
);
4303 fs_reclaim_release(gfp_mask
);
4304 psi_memstall_leave(&pflags
);
4311 /* The really slow allocator path where we enter direct reclaim */
4312 static inline struct page
*
4313 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4314 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4315 unsigned long *did_some_progress
)
4317 struct page
*page
= NULL
;
4318 bool drained
= false;
4320 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4321 if (unlikely(!(*did_some_progress
)))
4325 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4328 * If an allocation failed after direct reclaim, it could be because
4329 * pages are pinned on the per-cpu lists or in high alloc reserves.
4330 * Shrink them and try again
4332 if (!page
&& !drained
) {
4333 unreserve_highatomic_pageblock(ac
, false);
4334 drain_all_pages(NULL
);
4342 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4343 const struct alloc_context
*ac
)
4347 pg_data_t
*last_pgdat
= NULL
;
4348 enum zone_type highest_zoneidx
= ac
->highest_zoneidx
;
4350 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, highest_zoneidx
,
4352 if (last_pgdat
!= zone
->zone_pgdat
)
4353 wakeup_kswapd(zone
, gfp_mask
, order
, highest_zoneidx
);
4354 last_pgdat
= zone
->zone_pgdat
;
4358 static inline unsigned int
4359 gfp_to_alloc_flags(gfp_t gfp_mask
)
4361 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4364 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4365 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4366 * to save two branches.
4368 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4369 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM
!= (__force gfp_t
) ALLOC_KSWAPD
);
4372 * The caller may dip into page reserves a bit more if the caller
4373 * cannot run direct reclaim, or if the caller has realtime scheduling
4374 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4375 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4377 alloc_flags
|= (__force
int)
4378 (gfp_mask
& (__GFP_HIGH
| __GFP_KSWAPD_RECLAIM
));
4380 if (gfp_mask
& __GFP_ATOMIC
) {
4382 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4383 * if it can't schedule.
4385 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4386 alloc_flags
|= ALLOC_HARDER
;
4388 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4389 * comment for __cpuset_node_allowed().
4391 alloc_flags
&= ~ALLOC_CPUSET
;
4392 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4393 alloc_flags
|= ALLOC_HARDER
;
4395 alloc_flags
= current_alloc_flags(gfp_mask
, alloc_flags
);
4400 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4402 if (!tsk_is_oom_victim(tsk
))
4406 * !MMU doesn't have oom reaper so give access to memory reserves
4407 * only to the thread with TIF_MEMDIE set
4409 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4416 * Distinguish requests which really need access to full memory
4417 * reserves from oom victims which can live with a portion of it
4419 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4421 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4423 if (gfp_mask
& __GFP_MEMALLOC
)
4424 return ALLOC_NO_WATERMARKS
;
4425 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4426 return ALLOC_NO_WATERMARKS
;
4427 if (!in_interrupt()) {
4428 if (current
->flags
& PF_MEMALLOC
)
4429 return ALLOC_NO_WATERMARKS
;
4430 else if (oom_reserves_allowed(current
))
4437 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4439 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4443 * Checks whether it makes sense to retry the reclaim to make a forward progress
4444 * for the given allocation request.
4446 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4447 * without success, or when we couldn't even meet the watermark if we
4448 * reclaimed all remaining pages on the LRU lists.
4450 * Returns true if a retry is viable or false to enter the oom path.
4453 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4454 struct alloc_context
*ac
, int alloc_flags
,
4455 bool did_some_progress
, int *no_progress_loops
)
4462 * Costly allocations might have made a progress but this doesn't mean
4463 * their order will become available due to high fragmentation so
4464 * always increment the no progress counter for them
4466 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4467 *no_progress_loops
= 0;
4469 (*no_progress_loops
)++;
4472 * Make sure we converge to OOM if we cannot make any progress
4473 * several times in the row.
4475 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4476 /* Before OOM, exhaust highatomic_reserve */
4477 return unreserve_highatomic_pageblock(ac
, true);
4481 * Keep reclaiming pages while there is a chance this will lead
4482 * somewhere. If none of the target zones can satisfy our allocation
4483 * request even if all reclaimable pages are considered then we are
4484 * screwed and have to go OOM.
4486 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4487 ac
->highest_zoneidx
, ac
->nodemask
) {
4488 unsigned long available
;
4489 unsigned long reclaimable
;
4490 unsigned long min_wmark
= min_wmark_pages(zone
);
4493 available
= reclaimable
= zone_reclaimable_pages(zone
);
4494 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4497 * Would the allocation succeed if we reclaimed all
4498 * reclaimable pages?
4500 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4501 ac
->highest_zoneidx
, alloc_flags
, available
);
4502 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4503 available
, min_wmark
, *no_progress_loops
, wmark
);
4506 * If we didn't make any progress and have a lot of
4507 * dirty + writeback pages then we should wait for
4508 * an IO to complete to slow down the reclaim and
4509 * prevent from pre mature OOM
4511 if (!did_some_progress
) {
4512 unsigned long write_pending
;
4514 write_pending
= zone_page_state_snapshot(zone
,
4515 NR_ZONE_WRITE_PENDING
);
4517 if (2 * write_pending
> reclaimable
) {
4518 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4530 * Memory allocation/reclaim might be called from a WQ context and the
4531 * current implementation of the WQ concurrency control doesn't
4532 * recognize that a particular WQ is congested if the worker thread is
4533 * looping without ever sleeping. Therefore we have to do a short sleep
4534 * here rather than calling cond_resched().
4536 if (current
->flags
& PF_WQ_WORKER
)
4537 schedule_timeout_uninterruptible(1);
4544 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4547 * It's possible that cpuset's mems_allowed and the nodemask from
4548 * mempolicy don't intersect. This should be normally dealt with by
4549 * policy_nodemask(), but it's possible to race with cpuset update in
4550 * such a way the check therein was true, and then it became false
4551 * before we got our cpuset_mems_cookie here.
4552 * This assumes that for all allocations, ac->nodemask can come only
4553 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4554 * when it does not intersect with the cpuset restrictions) or the
4555 * caller can deal with a violated nodemask.
4557 if (cpusets_enabled() && ac
->nodemask
&&
4558 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4559 ac
->nodemask
= NULL
;
4564 * When updating a task's mems_allowed or mempolicy nodemask, it is
4565 * possible to race with parallel threads in such a way that our
4566 * allocation can fail while the mask is being updated. If we are about
4567 * to fail, check if the cpuset changed during allocation and if so,
4570 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4576 static inline struct page
*
4577 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4578 struct alloc_context
*ac
)
4580 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4581 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4582 struct page
*page
= NULL
;
4583 unsigned int alloc_flags
;
4584 unsigned long did_some_progress
;
4585 enum compact_priority compact_priority
;
4586 enum compact_result compact_result
;
4587 int compaction_retries
;
4588 int no_progress_loops
;
4589 unsigned int cpuset_mems_cookie
;
4593 * We also sanity check to catch abuse of atomic reserves being used by
4594 * callers that are not in atomic context.
4596 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4597 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4598 gfp_mask
&= ~__GFP_ATOMIC
;
4601 compaction_retries
= 0;
4602 no_progress_loops
= 0;
4603 compact_priority
= DEF_COMPACT_PRIORITY
;
4604 cpuset_mems_cookie
= read_mems_allowed_begin();
4607 * The fast path uses conservative alloc_flags to succeed only until
4608 * kswapd needs to be woken up, and to avoid the cost of setting up
4609 * alloc_flags precisely. So we do that now.
4611 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4614 * We need to recalculate the starting point for the zonelist iterator
4615 * because we might have used different nodemask in the fast path, or
4616 * there was a cpuset modification and we are retrying - otherwise we
4617 * could end up iterating over non-eligible zones endlessly.
4619 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4620 ac
->highest_zoneidx
, ac
->nodemask
);
4621 if (!ac
->preferred_zoneref
->zone
)
4624 if (alloc_flags
& ALLOC_KSWAPD
)
4625 wake_all_kswapds(order
, gfp_mask
, ac
);
4628 * The adjusted alloc_flags might result in immediate success, so try
4631 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4636 * For costly allocations, try direct compaction first, as it's likely
4637 * that we have enough base pages and don't need to reclaim. For non-
4638 * movable high-order allocations, do that as well, as compaction will
4639 * try prevent permanent fragmentation by migrating from blocks of the
4641 * Don't try this for allocations that are allowed to ignore
4642 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4644 if (can_direct_reclaim
&&
4646 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4647 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4648 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4650 INIT_COMPACT_PRIORITY
,
4656 * Checks for costly allocations with __GFP_NORETRY, which
4657 * includes some THP page fault allocations
4659 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4661 * If allocating entire pageblock(s) and compaction
4662 * failed because all zones are below low watermarks
4663 * or is prohibited because it recently failed at this
4664 * order, fail immediately unless the allocator has
4665 * requested compaction and reclaim retry.
4668 * - potentially very expensive because zones are far
4669 * below their low watermarks or this is part of very
4670 * bursty high order allocations,
4671 * - not guaranteed to help because isolate_freepages()
4672 * may not iterate over freed pages as part of its
4674 * - unlikely to make entire pageblocks free on its
4677 if (compact_result
== COMPACT_SKIPPED
||
4678 compact_result
== COMPACT_DEFERRED
)
4682 * Looks like reclaim/compaction is worth trying, but
4683 * sync compaction could be very expensive, so keep
4684 * using async compaction.
4686 compact_priority
= INIT_COMPACT_PRIORITY
;
4691 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4692 if (alloc_flags
& ALLOC_KSWAPD
)
4693 wake_all_kswapds(order
, gfp_mask
, ac
);
4695 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4697 alloc_flags
= current_alloc_flags(gfp_mask
, reserve_flags
);
4700 * Reset the nodemask and zonelist iterators if memory policies can be
4701 * ignored. These allocations are high priority and system rather than
4704 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4705 ac
->nodemask
= NULL
;
4706 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4707 ac
->highest_zoneidx
, ac
->nodemask
);
4710 /* Attempt with potentially adjusted zonelist and alloc_flags */
4711 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4715 /* Caller is not willing to reclaim, we can't balance anything */
4716 if (!can_direct_reclaim
)
4719 /* Avoid recursion of direct reclaim */
4720 if (current
->flags
& PF_MEMALLOC
)
4723 /* Try direct reclaim and then allocating */
4724 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4725 &did_some_progress
);
4729 /* Try direct compaction and then allocating */
4730 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4731 compact_priority
, &compact_result
);
4735 /* Do not loop if specifically requested */
4736 if (gfp_mask
& __GFP_NORETRY
)
4740 * Do not retry costly high order allocations unless they are
4741 * __GFP_RETRY_MAYFAIL
4743 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4746 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4747 did_some_progress
> 0, &no_progress_loops
))
4751 * It doesn't make any sense to retry for the compaction if the order-0
4752 * reclaim is not able to make any progress because the current
4753 * implementation of the compaction depends on the sufficient amount
4754 * of free memory (see __compaction_suitable)
4756 if (did_some_progress
> 0 &&
4757 should_compact_retry(ac
, order
, alloc_flags
,
4758 compact_result
, &compact_priority
,
4759 &compaction_retries
))
4763 /* Deal with possible cpuset update races before we start OOM killing */
4764 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4767 /* Reclaim has failed us, start killing things */
4768 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4772 /* Avoid allocations with no watermarks from looping endlessly */
4773 if (tsk_is_oom_victim(current
) &&
4774 (alloc_flags
& ALLOC_OOM
||
4775 (gfp_mask
& __GFP_NOMEMALLOC
)))
4778 /* Retry as long as the OOM killer is making progress */
4779 if (did_some_progress
) {
4780 no_progress_loops
= 0;
4785 /* Deal with possible cpuset update races before we fail */
4786 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4790 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4793 if (gfp_mask
& __GFP_NOFAIL
) {
4795 * All existing users of the __GFP_NOFAIL are blockable, so warn
4796 * of any new users that actually require GFP_NOWAIT
4798 if (WARN_ON_ONCE(!can_direct_reclaim
))
4802 * PF_MEMALLOC request from this context is rather bizarre
4803 * because we cannot reclaim anything and only can loop waiting
4804 * for somebody to do a work for us
4806 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4809 * non failing costly orders are a hard requirement which we
4810 * are not prepared for much so let's warn about these users
4811 * so that we can identify them and convert them to something
4814 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4817 * Help non-failing allocations by giving them access to memory
4818 * reserves but do not use ALLOC_NO_WATERMARKS because this
4819 * could deplete whole memory reserves which would just make
4820 * the situation worse
4822 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4830 warn_alloc(gfp_mask
, ac
->nodemask
,
4831 "page allocation failure: order:%u", order
);
4836 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4837 int preferred_nid
, nodemask_t
*nodemask
,
4838 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4839 unsigned int *alloc_flags
)
4841 ac
->highest_zoneidx
= gfp_zone(gfp_mask
);
4842 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4843 ac
->nodemask
= nodemask
;
4844 ac
->migratetype
= gfp_migratetype(gfp_mask
);
4846 if (cpusets_enabled()) {
4847 *alloc_mask
|= __GFP_HARDWALL
;
4849 * When we are in the interrupt context, it is irrelevant
4850 * to the current task context. It means that any node ok.
4852 if (!in_interrupt() && !ac
->nodemask
)
4853 ac
->nodemask
= &cpuset_current_mems_allowed
;
4855 *alloc_flags
|= ALLOC_CPUSET
;
4858 fs_reclaim_acquire(gfp_mask
);
4859 fs_reclaim_release(gfp_mask
);
4861 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4863 if (should_fail_alloc_page(gfp_mask
, order
))
4866 *alloc_flags
= current_alloc_flags(gfp_mask
, *alloc_flags
);
4868 /* Dirty zone balancing only done in the fast path */
4869 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4872 * The preferred zone is used for statistics but crucially it is
4873 * also used as the starting point for the zonelist iterator. It
4874 * may get reset for allocations that ignore memory policies.
4876 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4877 ac
->highest_zoneidx
, ac
->nodemask
);
4883 * This is the 'heart' of the zoned buddy allocator.
4886 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4887 nodemask_t
*nodemask
)
4890 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4891 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4892 struct alloc_context ac
= { };
4895 * There are several places where we assume that the order value is sane
4896 * so bail out early if the request is out of bound.
4898 if (unlikely(order
>= MAX_ORDER
)) {
4899 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4903 gfp_mask
&= gfp_allowed_mask
;
4904 alloc_mask
= gfp_mask
;
4905 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4909 * Forbid the first pass from falling back to types that fragment
4910 * memory until all local zones are considered.
4912 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4914 /* First allocation attempt */
4915 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4920 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4921 * resp. GFP_NOIO which has to be inherited for all allocation requests
4922 * from a particular context which has been marked by
4923 * memalloc_no{fs,io}_{save,restore}.
4925 alloc_mask
= current_gfp_context(gfp_mask
);
4926 ac
.spread_dirty_pages
= false;
4929 * Restore the original nodemask if it was potentially replaced with
4930 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4932 ac
.nodemask
= nodemask
;
4934 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4937 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4938 unlikely(__memcg_kmem_charge_page(page
, gfp_mask
, order
) != 0)) {
4939 __free_pages(page
, order
);
4943 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4947 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4950 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4951 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4952 * you need to access high mem.
4954 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4958 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4961 return (unsigned long) page_address(page
);
4963 EXPORT_SYMBOL(__get_free_pages
);
4965 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4967 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4969 EXPORT_SYMBOL(get_zeroed_page
);
4971 static inline void free_the_page(struct page
*page
, unsigned int order
)
4973 if (order
== 0) /* Via pcp? */
4974 free_unref_page(page
);
4976 __free_pages_ok(page
, order
);
4979 void __free_pages(struct page
*page
, unsigned int order
)
4981 if (put_page_testzero(page
))
4982 free_the_page(page
, order
);
4983 else if (!PageHead(page
))
4985 free_the_page(page
+ (1 << order
), order
);
4987 EXPORT_SYMBOL(__free_pages
);
4989 void free_pages(unsigned long addr
, unsigned int order
)
4992 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4993 __free_pages(virt_to_page((void *)addr
), order
);
4997 EXPORT_SYMBOL(free_pages
);
5001 * An arbitrary-length arbitrary-offset area of memory which resides
5002 * within a 0 or higher order page. Multiple fragments within that page
5003 * are individually refcounted, in the page's reference counter.
5005 * The page_frag functions below provide a simple allocation framework for
5006 * page fragments. This is used by the network stack and network device
5007 * drivers to provide a backing region of memory for use as either an
5008 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5010 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
5013 struct page
*page
= NULL
;
5014 gfp_t gfp
= gfp_mask
;
5016 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5017 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
5019 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
5020 PAGE_FRAG_CACHE_MAX_ORDER
);
5021 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
5023 if (unlikely(!page
))
5024 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
5026 nc
->va
= page
? page_address(page
) : NULL
;
5031 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
5033 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
5035 if (page_ref_sub_and_test(page
, count
))
5036 free_the_page(page
, compound_order(page
));
5038 EXPORT_SYMBOL(__page_frag_cache_drain
);
5040 void *page_frag_alloc(struct page_frag_cache
*nc
,
5041 unsigned int fragsz
, gfp_t gfp_mask
)
5043 unsigned int size
= PAGE_SIZE
;
5047 if (unlikely(!nc
->va
)) {
5049 page
= __page_frag_cache_refill(nc
, gfp_mask
);
5053 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5054 /* if size can vary use size else just use PAGE_SIZE */
5057 /* Even if we own the page, we do not use atomic_set().
5058 * This would break get_page_unless_zero() users.
5060 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
5062 /* reset page count bias and offset to start of new frag */
5063 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
5064 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
5068 offset
= nc
->offset
- fragsz
;
5069 if (unlikely(offset
< 0)) {
5070 page
= virt_to_page(nc
->va
);
5072 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
5075 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5076 /* if size can vary use size else just use PAGE_SIZE */
5079 /* OK, page count is 0, we can safely set it */
5080 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
5082 /* reset page count bias and offset to start of new frag */
5083 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
5084 offset
= size
- fragsz
;
5088 nc
->offset
= offset
;
5090 return nc
->va
+ offset
;
5092 EXPORT_SYMBOL(page_frag_alloc
);
5095 * Frees a page fragment allocated out of either a compound or order 0 page.
5097 void page_frag_free(void *addr
)
5099 struct page
*page
= virt_to_head_page(addr
);
5101 if (unlikely(put_page_testzero(page
)))
5102 free_the_page(page
, compound_order(page
));
5104 EXPORT_SYMBOL(page_frag_free
);
5106 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
5110 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
5111 unsigned long used
= addr
+ PAGE_ALIGN(size
);
5113 split_page(virt_to_page((void *)addr
), order
);
5114 while (used
< alloc_end
) {
5119 return (void *)addr
;
5123 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5124 * @size: the number of bytes to allocate
5125 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5127 * This function is similar to alloc_pages(), except that it allocates the
5128 * minimum number of pages to satisfy the request. alloc_pages() can only
5129 * allocate memory in power-of-two pages.
5131 * This function is also limited by MAX_ORDER.
5133 * Memory allocated by this function must be released by free_pages_exact().
5135 * Return: pointer to the allocated area or %NULL in case of error.
5137 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
5139 unsigned int order
= get_order(size
);
5142 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5143 gfp_mask
&= ~__GFP_COMP
;
5145 addr
= __get_free_pages(gfp_mask
, order
);
5146 return make_alloc_exact(addr
, order
, size
);
5148 EXPORT_SYMBOL(alloc_pages_exact
);
5151 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5153 * @nid: the preferred node ID where memory should be allocated
5154 * @size: the number of bytes to allocate
5155 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5157 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5160 * Return: pointer to the allocated area or %NULL in case of error.
5162 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
5164 unsigned int order
= get_order(size
);
5167 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5168 gfp_mask
&= ~__GFP_COMP
;
5170 p
= alloc_pages_node(nid
, gfp_mask
, order
);
5173 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5177 * free_pages_exact - release memory allocated via alloc_pages_exact()
5178 * @virt: the value returned by alloc_pages_exact.
5179 * @size: size of allocation, same value as passed to alloc_pages_exact().
5181 * Release the memory allocated by a previous call to alloc_pages_exact.
5183 void free_pages_exact(void *virt
, size_t size
)
5185 unsigned long addr
= (unsigned long)virt
;
5186 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5188 while (addr
< end
) {
5193 EXPORT_SYMBOL(free_pages_exact
);
5196 * nr_free_zone_pages - count number of pages beyond high watermark
5197 * @offset: The zone index of the highest zone
5199 * nr_free_zone_pages() counts the number of pages which are beyond the
5200 * high watermark within all zones at or below a given zone index. For each
5201 * zone, the number of pages is calculated as:
5203 * nr_free_zone_pages = managed_pages - high_pages
5205 * Return: number of pages beyond high watermark.
5207 static unsigned long nr_free_zone_pages(int offset
)
5212 /* Just pick one node, since fallback list is circular */
5213 unsigned long sum
= 0;
5215 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5217 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5218 unsigned long size
= zone_managed_pages(zone
);
5219 unsigned long high
= high_wmark_pages(zone
);
5228 * nr_free_buffer_pages - count number of pages beyond high watermark
5230 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5231 * watermark within ZONE_DMA and ZONE_NORMAL.
5233 * Return: number of pages beyond high watermark within ZONE_DMA and
5236 unsigned long nr_free_buffer_pages(void)
5238 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5240 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5242 static inline void show_node(struct zone
*zone
)
5244 if (IS_ENABLED(CONFIG_NUMA
))
5245 printk("Node %d ", zone_to_nid(zone
));
5248 long si_mem_available(void)
5251 unsigned long pagecache
;
5252 unsigned long wmark_low
= 0;
5253 unsigned long pages
[NR_LRU_LISTS
];
5254 unsigned long reclaimable
;
5258 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5259 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5262 wmark_low
+= low_wmark_pages(zone
);
5265 * Estimate the amount of memory available for userspace allocations,
5266 * without causing swapping.
5268 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5271 * Not all the page cache can be freed, otherwise the system will
5272 * start swapping. Assume at least half of the page cache, or the
5273 * low watermark worth of cache, needs to stay.
5275 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5276 pagecache
-= min(pagecache
/ 2, wmark_low
);
5277 available
+= pagecache
;
5280 * Part of the reclaimable slab and other kernel memory consists of
5281 * items that are in use, and cannot be freed. Cap this estimate at the
5284 reclaimable
= global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
) +
5285 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5286 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5292 EXPORT_SYMBOL_GPL(si_mem_available
);
5294 void si_meminfo(struct sysinfo
*val
)
5296 val
->totalram
= totalram_pages();
5297 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5298 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5299 val
->bufferram
= nr_blockdev_pages();
5300 val
->totalhigh
= totalhigh_pages();
5301 val
->freehigh
= nr_free_highpages();
5302 val
->mem_unit
= PAGE_SIZE
;
5305 EXPORT_SYMBOL(si_meminfo
);
5308 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5310 int zone_type
; /* needs to be signed */
5311 unsigned long managed_pages
= 0;
5312 unsigned long managed_highpages
= 0;
5313 unsigned long free_highpages
= 0;
5314 pg_data_t
*pgdat
= NODE_DATA(nid
);
5316 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5317 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5318 val
->totalram
= managed_pages
;
5319 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5320 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5321 #ifdef CONFIG_HIGHMEM
5322 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5323 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5325 if (is_highmem(zone
)) {
5326 managed_highpages
+= zone_managed_pages(zone
);
5327 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5330 val
->totalhigh
= managed_highpages
;
5331 val
->freehigh
= free_highpages
;
5333 val
->totalhigh
= managed_highpages
;
5334 val
->freehigh
= free_highpages
;
5336 val
->mem_unit
= PAGE_SIZE
;
5341 * Determine whether the node should be displayed or not, depending on whether
5342 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5344 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5346 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5350 * no node mask - aka implicit memory numa policy. Do not bother with
5351 * the synchronization - read_mems_allowed_begin - because we do not
5352 * have to be precise here.
5355 nodemask
= &cpuset_current_mems_allowed
;
5357 return !node_isset(nid
, *nodemask
);
5360 #define K(x) ((x) << (PAGE_SHIFT-10))
5362 static void show_migration_types(unsigned char type
)
5364 static const char types
[MIGRATE_TYPES
] = {
5365 [MIGRATE_UNMOVABLE
] = 'U',
5366 [MIGRATE_MOVABLE
] = 'M',
5367 [MIGRATE_RECLAIMABLE
] = 'E',
5368 [MIGRATE_HIGHATOMIC
] = 'H',
5370 [MIGRATE_CMA
] = 'C',
5372 #ifdef CONFIG_MEMORY_ISOLATION
5373 [MIGRATE_ISOLATE
] = 'I',
5376 char tmp
[MIGRATE_TYPES
+ 1];
5380 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5381 if (type
& (1 << i
))
5386 printk(KERN_CONT
"(%s) ", tmp
);
5390 * Show free area list (used inside shift_scroll-lock stuff)
5391 * We also calculate the percentage fragmentation. We do this by counting the
5392 * memory on each free list with the exception of the first item on the list.
5395 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5398 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5400 unsigned long free_pcp
= 0;
5405 for_each_populated_zone(zone
) {
5406 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5409 for_each_online_cpu(cpu
)
5410 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5413 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5414 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5415 " unevictable:%lu dirty:%lu writeback:%lu\n"
5416 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5417 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5418 " free:%lu free_pcp:%lu free_cma:%lu\n",
5419 global_node_page_state(NR_ACTIVE_ANON
),
5420 global_node_page_state(NR_INACTIVE_ANON
),
5421 global_node_page_state(NR_ISOLATED_ANON
),
5422 global_node_page_state(NR_ACTIVE_FILE
),
5423 global_node_page_state(NR_INACTIVE_FILE
),
5424 global_node_page_state(NR_ISOLATED_FILE
),
5425 global_node_page_state(NR_UNEVICTABLE
),
5426 global_node_page_state(NR_FILE_DIRTY
),
5427 global_node_page_state(NR_WRITEBACK
),
5428 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
),
5429 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B
),
5430 global_node_page_state(NR_FILE_MAPPED
),
5431 global_node_page_state(NR_SHMEM
),
5432 global_zone_page_state(NR_PAGETABLE
),
5433 global_zone_page_state(NR_BOUNCE
),
5434 global_zone_page_state(NR_FREE_PAGES
),
5436 global_zone_page_state(NR_FREE_CMA_PAGES
));
5438 for_each_online_pgdat(pgdat
) {
5439 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5443 " active_anon:%lukB"
5444 " inactive_anon:%lukB"
5445 " active_file:%lukB"
5446 " inactive_file:%lukB"
5447 " unevictable:%lukB"
5448 " isolated(anon):%lukB"
5449 " isolated(file):%lukB"
5454 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5456 " shmem_pmdmapped: %lukB"
5459 " writeback_tmp:%lukB"
5460 " kernel_stack:%lukB"
5461 #ifdef CONFIG_SHADOW_CALL_STACK
5462 " shadow_call_stack:%lukB"
5464 " all_unreclaimable? %s"
5467 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5468 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5469 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5470 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5471 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5472 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5473 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5474 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5475 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5476 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5477 K(node_page_state(pgdat
, NR_SHMEM
)),
5478 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5479 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
5480 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
5482 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
5484 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5485 node_page_state(pgdat
, NR_KERNEL_STACK_KB
),
5486 #ifdef CONFIG_SHADOW_CALL_STACK
5487 node_page_state(pgdat
, NR_KERNEL_SCS_KB
),
5489 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5493 for_each_populated_zone(zone
) {
5496 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5500 for_each_online_cpu(cpu
)
5501 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5510 " reserved_highatomic:%luKB"
5511 " active_anon:%lukB"
5512 " inactive_anon:%lukB"
5513 " active_file:%lukB"
5514 " inactive_file:%lukB"
5515 " unevictable:%lukB"
5516 " writepending:%lukB"
5527 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5528 K(min_wmark_pages(zone
)),
5529 K(low_wmark_pages(zone
)),
5530 K(high_wmark_pages(zone
)),
5531 K(zone
->nr_reserved_highatomic
),
5532 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5533 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5534 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5535 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5536 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5537 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5538 K(zone
->present_pages
),
5539 K(zone_managed_pages(zone
)),
5540 K(zone_page_state(zone
, NR_MLOCK
)),
5541 K(zone_page_state(zone
, NR_PAGETABLE
)),
5542 K(zone_page_state(zone
, NR_BOUNCE
)),
5544 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5545 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5546 printk("lowmem_reserve[]:");
5547 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5548 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5549 printk(KERN_CONT
"\n");
5552 for_each_populated_zone(zone
) {
5554 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5555 unsigned char types
[MAX_ORDER
];
5557 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5560 printk(KERN_CONT
"%s: ", zone
->name
);
5562 spin_lock_irqsave(&zone
->lock
, flags
);
5563 for (order
= 0; order
< MAX_ORDER
; order
++) {
5564 struct free_area
*area
= &zone
->free_area
[order
];
5567 nr
[order
] = area
->nr_free
;
5568 total
+= nr
[order
] << order
;
5571 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5572 if (!free_area_empty(area
, type
))
5573 types
[order
] |= 1 << type
;
5576 spin_unlock_irqrestore(&zone
->lock
, flags
);
5577 for (order
= 0; order
< MAX_ORDER
; order
++) {
5578 printk(KERN_CONT
"%lu*%lukB ",
5579 nr
[order
], K(1UL) << order
);
5581 show_migration_types(types
[order
]);
5583 printk(KERN_CONT
"= %lukB\n", K(total
));
5586 hugetlb_show_meminfo();
5588 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5590 show_swap_cache_info();
5593 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5595 zoneref
->zone
= zone
;
5596 zoneref
->zone_idx
= zone_idx(zone
);
5600 * Builds allocation fallback zone lists.
5602 * Add all populated zones of a node to the zonelist.
5604 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5607 enum zone_type zone_type
= MAX_NR_ZONES
;
5612 zone
= pgdat
->node_zones
+ zone_type
;
5613 if (managed_zone(zone
)) {
5614 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5615 check_highest_zone(zone_type
);
5617 } while (zone_type
);
5624 static int __parse_numa_zonelist_order(char *s
)
5627 * We used to support different zonlists modes but they turned
5628 * out to be just not useful. Let's keep the warning in place
5629 * if somebody still use the cmd line parameter so that we do
5630 * not fail it silently
5632 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5633 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5639 char numa_zonelist_order
[] = "Node";
5642 * sysctl handler for numa_zonelist_order
5644 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5645 void *buffer
, size_t *length
, loff_t
*ppos
)
5648 return __parse_numa_zonelist_order(buffer
);
5649 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5653 #define MAX_NODE_LOAD (nr_online_nodes)
5654 static int node_load
[MAX_NUMNODES
];
5657 * find_next_best_node - find the next node that should appear in a given node's fallback list
5658 * @node: node whose fallback list we're appending
5659 * @used_node_mask: nodemask_t of already used nodes
5661 * We use a number of factors to determine which is the next node that should
5662 * appear on a given node's fallback list. The node should not have appeared
5663 * already in @node's fallback list, and it should be the next closest node
5664 * according to the distance array (which contains arbitrary distance values
5665 * from each node to each node in the system), and should also prefer nodes
5666 * with no CPUs, since presumably they'll have very little allocation pressure
5667 * on them otherwise.
5669 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5671 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5674 int min_val
= INT_MAX
;
5675 int best_node
= NUMA_NO_NODE
;
5677 /* Use the local node if we haven't already */
5678 if (!node_isset(node
, *used_node_mask
)) {
5679 node_set(node
, *used_node_mask
);
5683 for_each_node_state(n
, N_MEMORY
) {
5685 /* Don't want a node to appear more than once */
5686 if (node_isset(n
, *used_node_mask
))
5689 /* Use the distance array to find the distance */
5690 val
= node_distance(node
, n
);
5692 /* Penalize nodes under us ("prefer the next node") */
5695 /* Give preference to headless and unused nodes */
5696 if (!cpumask_empty(cpumask_of_node(n
)))
5697 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5699 /* Slight preference for less loaded node */
5700 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5701 val
+= node_load
[n
];
5703 if (val
< min_val
) {
5710 node_set(best_node
, *used_node_mask
);
5717 * Build zonelists ordered by node and zones within node.
5718 * This results in maximum locality--normal zone overflows into local
5719 * DMA zone, if any--but risks exhausting DMA zone.
5721 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5724 struct zoneref
*zonerefs
;
5727 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5729 for (i
= 0; i
< nr_nodes
; i
++) {
5732 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5734 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5735 zonerefs
+= nr_zones
;
5737 zonerefs
->zone
= NULL
;
5738 zonerefs
->zone_idx
= 0;
5742 * Build gfp_thisnode zonelists
5744 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5746 struct zoneref
*zonerefs
;
5749 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5750 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5751 zonerefs
+= nr_zones
;
5752 zonerefs
->zone
= NULL
;
5753 zonerefs
->zone_idx
= 0;
5757 * Build zonelists ordered by zone and nodes within zones.
5758 * This results in conserving DMA zone[s] until all Normal memory is
5759 * exhausted, but results in overflowing to remote node while memory
5760 * may still exist in local DMA zone.
5763 static void build_zonelists(pg_data_t
*pgdat
)
5765 static int node_order
[MAX_NUMNODES
];
5766 int node
, load
, nr_nodes
= 0;
5767 nodemask_t used_mask
= NODE_MASK_NONE
;
5768 int local_node
, prev_node
;
5770 /* NUMA-aware ordering of nodes */
5771 local_node
= pgdat
->node_id
;
5772 load
= nr_online_nodes
;
5773 prev_node
= local_node
;
5775 memset(node_order
, 0, sizeof(node_order
));
5776 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5778 * We don't want to pressure a particular node.
5779 * So adding penalty to the first node in same
5780 * distance group to make it round-robin.
5782 if (node_distance(local_node
, node
) !=
5783 node_distance(local_node
, prev_node
))
5784 node_load
[node
] = load
;
5786 node_order
[nr_nodes
++] = node
;
5791 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5792 build_thisnode_zonelists(pgdat
);
5795 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5797 * Return node id of node used for "local" allocations.
5798 * I.e., first node id of first zone in arg node's generic zonelist.
5799 * Used for initializing percpu 'numa_mem', which is used primarily
5800 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5802 int local_memory_node(int node
)
5806 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5807 gfp_zone(GFP_KERNEL
),
5809 return zone_to_nid(z
->zone
);
5813 static void setup_min_unmapped_ratio(void);
5814 static void setup_min_slab_ratio(void);
5815 #else /* CONFIG_NUMA */
5817 static void build_zonelists(pg_data_t
*pgdat
)
5819 int node
, local_node
;
5820 struct zoneref
*zonerefs
;
5823 local_node
= pgdat
->node_id
;
5825 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5826 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5827 zonerefs
+= nr_zones
;
5830 * Now we build the zonelist so that it contains the zones
5831 * of all the other nodes.
5832 * We don't want to pressure a particular node, so when
5833 * building the zones for node N, we make sure that the
5834 * zones coming right after the local ones are those from
5835 * node N+1 (modulo N)
5837 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5838 if (!node_online(node
))
5840 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5841 zonerefs
+= nr_zones
;
5843 for (node
= 0; node
< local_node
; node
++) {
5844 if (!node_online(node
))
5846 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5847 zonerefs
+= nr_zones
;
5850 zonerefs
->zone
= NULL
;
5851 zonerefs
->zone_idx
= 0;
5854 #endif /* CONFIG_NUMA */
5857 * Boot pageset table. One per cpu which is going to be used for all
5858 * zones and all nodes. The parameters will be set in such a way
5859 * that an item put on a list will immediately be handed over to
5860 * the buddy list. This is safe since pageset manipulation is done
5861 * with interrupts disabled.
5863 * The boot_pagesets must be kept even after bootup is complete for
5864 * unused processors and/or zones. They do play a role for bootstrapping
5865 * hotplugged processors.
5867 * zoneinfo_show() and maybe other functions do
5868 * not check if the processor is online before following the pageset pointer.
5869 * Other parts of the kernel may not check if the zone is available.
5871 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5872 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5873 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5875 static void __build_all_zonelists(void *data
)
5878 int __maybe_unused cpu
;
5879 pg_data_t
*self
= data
;
5880 static DEFINE_SPINLOCK(lock
);
5885 memset(node_load
, 0, sizeof(node_load
));
5889 * This node is hotadded and no memory is yet present. So just
5890 * building zonelists is fine - no need to touch other nodes.
5892 if (self
&& !node_online(self
->node_id
)) {
5893 build_zonelists(self
);
5895 for_each_online_node(nid
) {
5896 pg_data_t
*pgdat
= NODE_DATA(nid
);
5898 build_zonelists(pgdat
);
5901 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5903 * We now know the "local memory node" for each node--
5904 * i.e., the node of the first zone in the generic zonelist.
5905 * Set up numa_mem percpu variable for on-line cpus. During
5906 * boot, only the boot cpu should be on-line; we'll init the
5907 * secondary cpus' numa_mem as they come on-line. During
5908 * node/memory hotplug, we'll fixup all on-line cpus.
5910 for_each_online_cpu(cpu
)
5911 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5918 static noinline
void __init
5919 build_all_zonelists_init(void)
5923 __build_all_zonelists(NULL
);
5926 * Initialize the boot_pagesets that are going to be used
5927 * for bootstrapping processors. The real pagesets for
5928 * each zone will be allocated later when the per cpu
5929 * allocator is available.
5931 * boot_pagesets are used also for bootstrapping offline
5932 * cpus if the system is already booted because the pagesets
5933 * are needed to initialize allocators on a specific cpu too.
5934 * F.e. the percpu allocator needs the page allocator which
5935 * needs the percpu allocator in order to allocate its pagesets
5936 * (a chicken-egg dilemma).
5938 for_each_possible_cpu(cpu
)
5939 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5941 mminit_verify_zonelist();
5942 cpuset_init_current_mems_allowed();
5946 * unless system_state == SYSTEM_BOOTING.
5948 * __ref due to call of __init annotated helper build_all_zonelists_init
5949 * [protected by SYSTEM_BOOTING].
5951 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5953 unsigned long vm_total_pages
;
5955 if (system_state
== SYSTEM_BOOTING
) {
5956 build_all_zonelists_init();
5958 __build_all_zonelists(pgdat
);
5959 /* cpuset refresh routine should be here */
5961 /* Get the number of free pages beyond high watermark in all zones. */
5962 vm_total_pages
= nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5964 * Disable grouping by mobility if the number of pages in the
5965 * system is too low to allow the mechanism to work. It would be
5966 * more accurate, but expensive to check per-zone. This check is
5967 * made on memory-hotadd so a system can start with mobility
5968 * disabled and enable it later
5970 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5971 page_group_by_mobility_disabled
= 1;
5973 page_group_by_mobility_disabled
= 0;
5975 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5977 page_group_by_mobility_disabled
? "off" : "on",
5980 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5984 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5985 static bool __meminit
5986 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
5988 static struct memblock_region
*r
;
5990 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5991 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
5992 for_each_mem_region(r
) {
5993 if (*pfn
< memblock_region_memory_end_pfn(r
))
5997 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
5998 memblock_is_mirror(r
)) {
5999 *pfn
= memblock_region_memory_end_pfn(r
);
6007 * Initially all pages are reserved - free ones are freed
6008 * up by memblock_free_all() once the early boot process is
6009 * done. Non-atomic initialization, single-pass.
6011 * All aligned pageblocks are initialized to the specified migratetype
6012 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6013 * zone stats (e.g., nr_isolate_pageblock) are touched.
6015 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
6016 unsigned long start_pfn
,
6017 enum meminit_context context
,
6018 struct vmem_altmap
*altmap
, int migratetype
)
6020 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
6023 if (highest_memmap_pfn
< end_pfn
- 1)
6024 highest_memmap_pfn
= end_pfn
- 1;
6026 #ifdef CONFIG_ZONE_DEVICE
6028 * Honor reservation requested by the driver for this ZONE_DEVICE
6029 * memory. We limit the total number of pages to initialize to just
6030 * those that might contain the memory mapping. We will defer the
6031 * ZONE_DEVICE page initialization until after we have released
6034 if (zone
== ZONE_DEVICE
) {
6038 if (start_pfn
== altmap
->base_pfn
)
6039 start_pfn
+= altmap
->reserve
;
6040 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
6044 for (pfn
= start_pfn
; pfn
< end_pfn
; ) {
6046 * There can be holes in boot-time mem_map[]s handed to this
6047 * function. They do not exist on hotplugged memory.
6049 if (context
== MEMINIT_EARLY
) {
6050 if (overlap_memmap_init(zone
, &pfn
))
6052 if (defer_init(nid
, pfn
, end_pfn
))
6056 page
= pfn_to_page(pfn
);
6057 __init_single_page(page
, pfn
, zone
, nid
);
6058 if (context
== MEMINIT_HOTPLUG
)
6059 __SetPageReserved(page
);
6062 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6063 * such that unmovable allocations won't be scattered all
6064 * over the place during system boot.
6066 if (IS_ALIGNED(pfn
, pageblock_nr_pages
)) {
6067 set_pageblock_migratetype(page
, migratetype
);
6074 #ifdef CONFIG_ZONE_DEVICE
6075 void __ref
memmap_init_zone_device(struct zone
*zone
,
6076 unsigned long start_pfn
,
6077 unsigned long nr_pages
,
6078 struct dev_pagemap
*pgmap
)
6080 unsigned long pfn
, end_pfn
= start_pfn
+ nr_pages
;
6081 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6082 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
6083 unsigned long zone_idx
= zone_idx(zone
);
6084 unsigned long start
= jiffies
;
6085 int nid
= pgdat
->node_id
;
6087 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
6091 * The call to memmap_init_zone should have already taken care
6092 * of the pages reserved for the memmap, so we can just jump to
6093 * the end of that region and start processing the device pages.
6096 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
6097 nr_pages
= end_pfn
- start_pfn
;
6100 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
6101 struct page
*page
= pfn_to_page(pfn
);
6103 __init_single_page(page
, pfn
, zone_idx
, nid
);
6106 * Mark page reserved as it will need to wait for onlining
6107 * phase for it to be fully associated with a zone.
6109 * We can use the non-atomic __set_bit operation for setting
6110 * the flag as we are still initializing the pages.
6112 __SetPageReserved(page
);
6115 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6116 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6117 * ever freed or placed on a driver-private list.
6119 page
->pgmap
= pgmap
;
6120 page
->zone_device_data
= NULL
;
6123 * Mark the block movable so that blocks are reserved for
6124 * movable at startup. This will force kernel allocations
6125 * to reserve their blocks rather than leaking throughout
6126 * the address space during boot when many long-lived
6127 * kernel allocations are made.
6129 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6130 * because this is done early in section_activate()
6132 if (IS_ALIGNED(pfn
, pageblock_nr_pages
)) {
6133 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6138 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6139 nr_pages
, jiffies_to_msecs(jiffies
- start
));
6143 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6145 unsigned int order
, t
;
6146 for_each_migratetype_order(order
, t
) {
6147 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6148 zone
->free_area
[order
].nr_free
= 0;
6152 void __meminit __weak
memmap_init(unsigned long size
, int nid
,
6154 unsigned long range_start_pfn
)
6156 unsigned long start_pfn
, end_pfn
;
6157 unsigned long range_end_pfn
= range_start_pfn
+ size
;
6160 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6161 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6162 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6164 if (end_pfn
> start_pfn
) {
6165 size
= end_pfn
- start_pfn
;
6166 memmap_init_zone(size
, nid
, zone
, start_pfn
,
6167 MEMINIT_EARLY
, NULL
, MIGRATE_MOVABLE
);
6172 static int zone_batchsize(struct zone
*zone
)
6178 * The per-cpu-pages pools are set to around 1000th of the
6181 batch
= zone_managed_pages(zone
) / 1024;
6182 /* But no more than a meg. */
6183 if (batch
* PAGE_SIZE
> 1024 * 1024)
6184 batch
= (1024 * 1024) / PAGE_SIZE
;
6185 batch
/= 4; /* We effectively *= 4 below */
6190 * Clamp the batch to a 2^n - 1 value. Having a power
6191 * of 2 value was found to be more likely to have
6192 * suboptimal cache aliasing properties in some cases.
6194 * For example if 2 tasks are alternately allocating
6195 * batches of pages, one task can end up with a lot
6196 * of pages of one half of the possible page colors
6197 * and the other with pages of the other colors.
6199 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6204 /* The deferral and batching of frees should be suppressed under NOMMU
6207 * The problem is that NOMMU needs to be able to allocate large chunks
6208 * of contiguous memory as there's no hardware page translation to
6209 * assemble apparent contiguous memory from discontiguous pages.
6211 * Queueing large contiguous runs of pages for batching, however,
6212 * causes the pages to actually be freed in smaller chunks. As there
6213 * can be a significant delay between the individual batches being
6214 * recycled, this leads to the once large chunks of space being
6215 * fragmented and becoming unavailable for high-order allocations.
6222 * pcp->high and pcp->batch values are related and dependent on one another:
6223 * ->batch must never be higher then ->high.
6224 * The following function updates them in a safe manner without read side
6227 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6228 * those fields changing asynchronously (acording to the above rule).
6230 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6231 * outside of boot time (or some other assurance that no concurrent updaters
6234 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6235 unsigned long batch
)
6237 /* start with a fail safe value for batch */
6241 /* Update high, then batch, in order */
6248 /* a companion to pageset_set_high() */
6249 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
6251 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
6254 static void pageset_init(struct per_cpu_pageset
*p
)
6256 struct per_cpu_pages
*pcp
;
6259 memset(p
, 0, sizeof(*p
));
6262 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
6263 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
6266 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
6269 pageset_set_batch(p
, batch
);
6273 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6274 * to the value high for the pageset p.
6276 static void pageset_set_high(struct per_cpu_pageset
*p
,
6279 unsigned long batch
= max(1UL, high
/ 4);
6280 if ((high
/ 4) > (PAGE_SHIFT
* 8))
6281 batch
= PAGE_SHIFT
* 8;
6283 pageset_update(&p
->pcp
, high
, batch
);
6286 static void pageset_set_high_and_batch(struct zone
*zone
,
6287 struct per_cpu_pageset
*pcp
)
6289 if (percpu_pagelist_fraction
)
6290 pageset_set_high(pcp
,
6291 (zone_managed_pages(zone
) /
6292 percpu_pagelist_fraction
));
6294 pageset_set_batch(pcp
, zone_batchsize(zone
));
6297 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
6299 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
6302 pageset_set_high_and_batch(zone
, pcp
);
6305 void __meminit
setup_zone_pageset(struct zone
*zone
)
6308 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
6309 for_each_possible_cpu(cpu
)
6310 zone_pageset_init(zone
, cpu
);
6314 * Allocate per cpu pagesets and initialize them.
6315 * Before this call only boot pagesets were available.
6317 void __init
setup_per_cpu_pageset(void)
6319 struct pglist_data
*pgdat
;
6321 int __maybe_unused cpu
;
6323 for_each_populated_zone(zone
)
6324 setup_zone_pageset(zone
);
6328 * Unpopulated zones continue using the boot pagesets.
6329 * The numa stats for these pagesets need to be reset.
6330 * Otherwise, they will end up skewing the stats of
6331 * the nodes these zones are associated with.
6333 for_each_possible_cpu(cpu
) {
6334 struct per_cpu_pageset
*pcp
= &per_cpu(boot_pageset
, cpu
);
6335 memset(pcp
->vm_numa_stat_diff
, 0,
6336 sizeof(pcp
->vm_numa_stat_diff
));
6340 for_each_online_pgdat(pgdat
)
6341 pgdat
->per_cpu_nodestats
=
6342 alloc_percpu(struct per_cpu_nodestat
);
6345 static __meminit
void zone_pcp_init(struct zone
*zone
)
6348 * per cpu subsystem is not up at this point. The following code
6349 * relies on the ability of the linker to provide the
6350 * offset of a (static) per cpu variable into the per cpu area.
6352 zone
->pageset
= &boot_pageset
;
6354 if (populated_zone(zone
))
6355 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
6356 zone
->name
, zone
->present_pages
,
6357 zone_batchsize(zone
));
6360 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6361 unsigned long zone_start_pfn
,
6364 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6365 int zone_idx
= zone_idx(zone
) + 1;
6367 if (zone_idx
> pgdat
->nr_zones
)
6368 pgdat
->nr_zones
= zone_idx
;
6370 zone
->zone_start_pfn
= zone_start_pfn
;
6372 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6373 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6375 (unsigned long)zone_idx(zone
),
6376 zone_start_pfn
, (zone_start_pfn
+ size
));
6378 zone_init_free_lists(zone
);
6379 zone
->initialized
= 1;
6383 * get_pfn_range_for_nid - Return the start and end page frames for a node
6384 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6385 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6386 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6388 * It returns the start and end page frame of a node based on information
6389 * provided by memblock_set_node(). If called for a node
6390 * with no available memory, a warning is printed and the start and end
6393 void __init
get_pfn_range_for_nid(unsigned int nid
,
6394 unsigned long *start_pfn
, unsigned long *end_pfn
)
6396 unsigned long this_start_pfn
, this_end_pfn
;
6402 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6403 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6404 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6407 if (*start_pfn
== -1UL)
6412 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6413 * assumption is made that zones within a node are ordered in monotonic
6414 * increasing memory addresses so that the "highest" populated zone is used
6416 static void __init
find_usable_zone_for_movable(void)
6419 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6420 if (zone_index
== ZONE_MOVABLE
)
6423 if (arch_zone_highest_possible_pfn
[zone_index
] >
6424 arch_zone_lowest_possible_pfn
[zone_index
])
6428 VM_BUG_ON(zone_index
== -1);
6429 movable_zone
= zone_index
;
6433 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6434 * because it is sized independent of architecture. Unlike the other zones,
6435 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6436 * in each node depending on the size of each node and how evenly kernelcore
6437 * is distributed. This helper function adjusts the zone ranges
6438 * provided by the architecture for a given node by using the end of the
6439 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6440 * zones within a node are in order of monotonic increases memory addresses
6442 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6443 unsigned long zone_type
,
6444 unsigned long node_start_pfn
,
6445 unsigned long node_end_pfn
,
6446 unsigned long *zone_start_pfn
,
6447 unsigned long *zone_end_pfn
)
6449 /* Only adjust if ZONE_MOVABLE is on this node */
6450 if (zone_movable_pfn
[nid
]) {
6451 /* Size ZONE_MOVABLE */
6452 if (zone_type
== ZONE_MOVABLE
) {
6453 *zone_start_pfn
= zone_movable_pfn
[nid
];
6454 *zone_end_pfn
= min(node_end_pfn
,
6455 arch_zone_highest_possible_pfn
[movable_zone
]);
6457 /* Adjust for ZONE_MOVABLE starting within this range */
6458 } else if (!mirrored_kernelcore
&&
6459 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6460 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6461 *zone_end_pfn
= zone_movable_pfn
[nid
];
6463 /* Check if this whole range is within ZONE_MOVABLE */
6464 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6465 *zone_start_pfn
= *zone_end_pfn
;
6470 * Return the number of pages a zone spans in a node, including holes
6471 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6473 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6474 unsigned long zone_type
,
6475 unsigned long node_start_pfn
,
6476 unsigned long node_end_pfn
,
6477 unsigned long *zone_start_pfn
,
6478 unsigned long *zone_end_pfn
)
6480 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6481 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6482 /* When hotadd a new node from cpu_up(), the node should be empty */
6483 if (!node_start_pfn
&& !node_end_pfn
)
6486 /* Get the start and end of the zone */
6487 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6488 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6489 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6490 node_start_pfn
, node_end_pfn
,
6491 zone_start_pfn
, zone_end_pfn
);
6493 /* Check that this node has pages within the zone's required range */
6494 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6497 /* Move the zone boundaries inside the node if necessary */
6498 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6499 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6501 /* Return the spanned pages */
6502 return *zone_end_pfn
- *zone_start_pfn
;
6506 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6507 * then all holes in the requested range will be accounted for.
6509 unsigned long __init
__absent_pages_in_range(int nid
,
6510 unsigned long range_start_pfn
,
6511 unsigned long range_end_pfn
)
6513 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6514 unsigned long start_pfn
, end_pfn
;
6517 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6518 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6519 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6520 nr_absent
-= end_pfn
- start_pfn
;
6526 * absent_pages_in_range - Return number of page frames in holes within a range
6527 * @start_pfn: The start PFN to start searching for holes
6528 * @end_pfn: The end PFN to stop searching for holes
6530 * Return: the number of pages frames in memory holes within a range.
6532 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6533 unsigned long end_pfn
)
6535 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6538 /* Return the number of page frames in holes in a zone on a node */
6539 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6540 unsigned long zone_type
,
6541 unsigned long node_start_pfn
,
6542 unsigned long node_end_pfn
)
6544 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6545 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6546 unsigned long zone_start_pfn
, zone_end_pfn
;
6547 unsigned long nr_absent
;
6549 /* When hotadd a new node from cpu_up(), the node should be empty */
6550 if (!node_start_pfn
&& !node_end_pfn
)
6553 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6554 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6556 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6557 node_start_pfn
, node_end_pfn
,
6558 &zone_start_pfn
, &zone_end_pfn
);
6559 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6562 * ZONE_MOVABLE handling.
6563 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6566 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6567 unsigned long start_pfn
, end_pfn
;
6568 struct memblock_region
*r
;
6570 for_each_mem_region(r
) {
6571 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6572 zone_start_pfn
, zone_end_pfn
);
6573 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6574 zone_start_pfn
, zone_end_pfn
);
6576 if (zone_type
== ZONE_MOVABLE
&&
6577 memblock_is_mirror(r
))
6578 nr_absent
+= end_pfn
- start_pfn
;
6580 if (zone_type
== ZONE_NORMAL
&&
6581 !memblock_is_mirror(r
))
6582 nr_absent
+= end_pfn
- start_pfn
;
6589 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6590 unsigned long node_start_pfn
,
6591 unsigned long node_end_pfn
)
6593 unsigned long realtotalpages
= 0, totalpages
= 0;
6596 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6597 struct zone
*zone
= pgdat
->node_zones
+ i
;
6598 unsigned long zone_start_pfn
, zone_end_pfn
;
6599 unsigned long spanned
, absent
;
6600 unsigned long size
, real_size
;
6602 spanned
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6607 absent
= zone_absent_pages_in_node(pgdat
->node_id
, i
,
6612 real_size
= size
- absent
;
6615 zone
->zone_start_pfn
= zone_start_pfn
;
6617 zone
->zone_start_pfn
= 0;
6618 zone
->spanned_pages
= size
;
6619 zone
->present_pages
= real_size
;
6622 realtotalpages
+= real_size
;
6625 pgdat
->node_spanned_pages
= totalpages
;
6626 pgdat
->node_present_pages
= realtotalpages
;
6627 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6631 #ifndef CONFIG_SPARSEMEM
6633 * Calculate the size of the zone->blockflags rounded to an unsigned long
6634 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6635 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6636 * round what is now in bits to nearest long in bits, then return it in
6639 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6641 unsigned long usemapsize
;
6643 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6644 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6645 usemapsize
= usemapsize
>> pageblock_order
;
6646 usemapsize
*= NR_PAGEBLOCK_BITS
;
6647 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6649 return usemapsize
/ 8;
6652 static void __ref
setup_usemap(struct pglist_data
*pgdat
,
6654 unsigned long zone_start_pfn
,
6655 unsigned long zonesize
)
6657 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
6658 zone
->pageblock_flags
= NULL
;
6660 zone
->pageblock_flags
=
6661 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
6663 if (!zone
->pageblock_flags
)
6664 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6665 usemapsize
, zone
->name
, pgdat
->node_id
);
6669 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6670 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6671 #endif /* CONFIG_SPARSEMEM */
6673 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6675 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6676 void __init
set_pageblock_order(void)
6680 /* Check that pageblock_nr_pages has not already been setup */
6681 if (pageblock_order
)
6684 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6685 order
= HUGETLB_PAGE_ORDER
;
6687 order
= MAX_ORDER
- 1;
6690 * Assume the largest contiguous order of interest is a huge page.
6691 * This value may be variable depending on boot parameters on IA64 and
6694 pageblock_order
= order
;
6696 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6699 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6700 * is unused as pageblock_order is set at compile-time. See
6701 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6704 void __init
set_pageblock_order(void)
6708 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6710 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6711 unsigned long present_pages
)
6713 unsigned long pages
= spanned_pages
;
6716 * Provide a more accurate estimation if there are holes within
6717 * the zone and SPARSEMEM is in use. If there are holes within the
6718 * zone, each populated memory region may cost us one or two extra
6719 * memmap pages due to alignment because memmap pages for each
6720 * populated regions may not be naturally aligned on page boundary.
6721 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6723 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6724 IS_ENABLED(CONFIG_SPARSEMEM
))
6725 pages
= present_pages
;
6727 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6730 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6731 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6733 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
6735 spin_lock_init(&ds_queue
->split_queue_lock
);
6736 INIT_LIST_HEAD(&ds_queue
->split_queue
);
6737 ds_queue
->split_queue_len
= 0;
6740 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6743 #ifdef CONFIG_COMPACTION
6744 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6746 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6749 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6752 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6754 pgdat_resize_init(pgdat
);
6756 pgdat_init_split_queue(pgdat
);
6757 pgdat_init_kcompactd(pgdat
);
6759 init_waitqueue_head(&pgdat
->kswapd_wait
);
6760 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6762 pgdat_page_ext_init(pgdat
);
6763 spin_lock_init(&pgdat
->lru_lock
);
6764 lruvec_init(&pgdat
->__lruvec
);
6767 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6768 unsigned long remaining_pages
)
6770 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6771 zone_set_nid(zone
, nid
);
6772 zone
->name
= zone_names
[idx
];
6773 zone
->zone_pgdat
= NODE_DATA(nid
);
6774 spin_lock_init(&zone
->lock
);
6775 zone_seqlock_init(zone
);
6776 zone_pcp_init(zone
);
6780 * Set up the zone data structures
6781 * - init pgdat internals
6782 * - init all zones belonging to this node
6784 * NOTE: this function is only called during memory hotplug
6786 #ifdef CONFIG_MEMORY_HOTPLUG
6787 void __ref
free_area_init_core_hotplug(int nid
)
6790 pg_data_t
*pgdat
= NODE_DATA(nid
);
6792 pgdat_init_internals(pgdat
);
6793 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6794 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6799 * Set up the zone data structures:
6800 * - mark all pages reserved
6801 * - mark all memory queues empty
6802 * - clear the memory bitmaps
6804 * NOTE: pgdat should get zeroed by caller.
6805 * NOTE: this function is only called during early init.
6807 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
6810 int nid
= pgdat
->node_id
;
6812 pgdat_init_internals(pgdat
);
6813 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6815 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6816 struct zone
*zone
= pgdat
->node_zones
+ j
;
6817 unsigned long size
, freesize
, memmap_pages
;
6818 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6820 size
= zone
->spanned_pages
;
6821 freesize
= zone
->present_pages
;
6824 * Adjust freesize so that it accounts for how much memory
6825 * is used by this zone for memmap. This affects the watermark
6826 * and per-cpu initialisations
6828 memmap_pages
= calc_memmap_size(size
, freesize
);
6829 if (!is_highmem_idx(j
)) {
6830 if (freesize
>= memmap_pages
) {
6831 freesize
-= memmap_pages
;
6834 " %s zone: %lu pages used for memmap\n",
6835 zone_names
[j
], memmap_pages
);
6837 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6838 zone_names
[j
], memmap_pages
, freesize
);
6841 /* Account for reserved pages */
6842 if (j
== 0 && freesize
> dma_reserve
) {
6843 freesize
-= dma_reserve
;
6844 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6845 zone_names
[0], dma_reserve
);
6848 if (!is_highmem_idx(j
))
6849 nr_kernel_pages
+= freesize
;
6850 /* Charge for highmem memmap if there are enough kernel pages */
6851 else if (nr_kernel_pages
> memmap_pages
* 2)
6852 nr_kernel_pages
-= memmap_pages
;
6853 nr_all_pages
+= freesize
;
6856 * Set an approximate value for lowmem here, it will be adjusted
6857 * when the bootmem allocator frees pages into the buddy system.
6858 * And all highmem pages will be managed by the buddy system.
6860 zone_init_internals(zone
, j
, nid
, freesize
);
6865 set_pageblock_order();
6866 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6867 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6868 memmap_init(size
, nid
, j
, zone_start_pfn
);
6872 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6873 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6875 unsigned long __maybe_unused start
= 0;
6876 unsigned long __maybe_unused offset
= 0;
6878 /* Skip empty nodes */
6879 if (!pgdat
->node_spanned_pages
)
6882 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6883 offset
= pgdat
->node_start_pfn
- start
;
6884 /* ia64 gets its own node_mem_map, before this, without bootmem */
6885 if (!pgdat
->node_mem_map
) {
6886 unsigned long size
, end
;
6890 * The zone's endpoints aren't required to be MAX_ORDER
6891 * aligned but the node_mem_map endpoints must be in order
6892 * for the buddy allocator to function correctly.
6894 end
= pgdat_end_pfn(pgdat
);
6895 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6896 size
= (end
- start
) * sizeof(struct page
);
6897 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
6900 panic("Failed to allocate %ld bytes for node %d memory map\n",
6901 size
, pgdat
->node_id
);
6902 pgdat
->node_mem_map
= map
+ offset
;
6904 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6905 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6906 (unsigned long)pgdat
->node_mem_map
);
6907 #ifndef CONFIG_NEED_MULTIPLE_NODES
6909 * With no DISCONTIG, the global mem_map is just set as node 0's
6911 if (pgdat
== NODE_DATA(0)) {
6912 mem_map
= NODE_DATA(0)->node_mem_map
;
6913 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6919 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6920 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6922 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6923 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
6925 pgdat
->first_deferred_pfn
= ULONG_MAX
;
6928 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
6931 static void __init
free_area_init_node(int nid
)
6933 pg_data_t
*pgdat
= NODE_DATA(nid
);
6934 unsigned long start_pfn
= 0;
6935 unsigned long end_pfn
= 0;
6937 /* pg_data_t should be reset to zero when it's allocated */
6938 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_highest_zoneidx
);
6940 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6942 pgdat
->node_id
= nid
;
6943 pgdat
->node_start_pfn
= start_pfn
;
6944 pgdat
->per_cpu_nodestats
= NULL
;
6946 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6947 (u64
)start_pfn
<< PAGE_SHIFT
,
6948 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6949 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
);
6951 alloc_node_mem_map(pgdat
);
6952 pgdat_set_deferred_range(pgdat
);
6954 free_area_init_core(pgdat
);
6957 void __init
free_area_init_memoryless_node(int nid
)
6959 free_area_init_node(nid
);
6962 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6964 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6965 * PageReserved(). Return the number of struct pages that were initialized.
6967 static u64 __init
init_unavailable_range(unsigned long spfn
, unsigned long epfn
)
6972 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6973 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6974 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6975 + pageblock_nr_pages
- 1;
6979 * Use a fake node/zone (0) for now. Some of these pages
6980 * (in memblock.reserved but not in memblock.memory) will
6981 * get re-initialized via reserve_bootmem_region() later.
6983 __init_single_page(pfn_to_page(pfn
), pfn
, 0, 0);
6984 __SetPageReserved(pfn_to_page(pfn
));
6992 * Only struct pages that are backed by physical memory are zeroed and
6993 * initialized by going through __init_single_page(). But, there are some
6994 * struct pages which are reserved in memblock allocator and their fields
6995 * may be accessed (for example page_to_pfn() on some configuration accesses
6996 * flags). We must explicitly initialize those struct pages.
6998 * This function also addresses a similar issue where struct pages are left
6999 * uninitialized because the physical address range is not covered by
7000 * memblock.memory or memblock.reserved. That could happen when memblock
7001 * layout is manually configured via memmap=, or when the highest physical
7002 * address (max_pfn) does not end on a section boundary.
7004 static void __init
init_unavailable_mem(void)
7006 phys_addr_t start
, end
;
7008 phys_addr_t next
= 0;
7011 * Loop through unavailable ranges not covered by memblock.memory.
7014 for_each_mem_range(i
, &start
, &end
) {
7016 pgcnt
+= init_unavailable_range(PFN_DOWN(next
),
7022 * Early sections always have a fully populated memmap for the whole
7023 * section - see pfn_valid(). If the last section has holes at the
7024 * end and that section is marked "online", the memmap will be
7025 * considered initialized. Make sure that memmap has a well defined
7028 pgcnt
+= init_unavailable_range(PFN_DOWN(next
),
7029 round_up(max_pfn
, PAGES_PER_SECTION
));
7032 * Struct pages that do not have backing memory. This could be because
7033 * firmware is using some of this memory, or for some other reasons.
7036 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt
);
7039 static inline void __init
init_unavailable_mem(void)
7042 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7044 #if MAX_NUMNODES > 1
7046 * Figure out the number of possible node ids.
7048 void __init
setup_nr_node_ids(void)
7050 unsigned int highest
;
7052 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
7053 nr_node_ids
= highest
+ 1;
7058 * node_map_pfn_alignment - determine the maximum internode alignment
7060 * This function should be called after node map is populated and sorted.
7061 * It calculates the maximum power of two alignment which can distinguish
7064 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7065 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7066 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7067 * shifted, 1GiB is enough and this function will indicate so.
7069 * This is used to test whether pfn -> nid mapping of the chosen memory
7070 * model has fine enough granularity to avoid incorrect mapping for the
7071 * populated node map.
7073 * Return: the determined alignment in pfn's. 0 if there is no alignment
7074 * requirement (single node).
7076 unsigned long __init
node_map_pfn_alignment(void)
7078 unsigned long accl_mask
= 0, last_end
= 0;
7079 unsigned long start
, end
, mask
;
7080 int last_nid
= NUMA_NO_NODE
;
7083 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
7084 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
7091 * Start with a mask granular enough to pin-point to the
7092 * start pfn and tick off bits one-by-one until it becomes
7093 * too coarse to separate the current node from the last.
7095 mask
= ~((1 << __ffs(start
)) - 1);
7096 while (mask
&& last_end
<= (start
& (mask
<< 1)))
7099 /* accumulate all internode masks */
7103 /* convert mask to number of pages */
7104 return ~accl_mask
+ 1;
7108 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7110 * Return: the minimum PFN based on information provided via
7111 * memblock_set_node().
7113 unsigned long __init
find_min_pfn_with_active_regions(void)
7115 return PHYS_PFN(memblock_start_of_DRAM());
7119 * early_calculate_totalpages()
7120 * Sum pages in active regions for movable zone.
7121 * Populate N_MEMORY for calculating usable_nodes.
7123 static unsigned long __init
early_calculate_totalpages(void)
7125 unsigned long totalpages
= 0;
7126 unsigned long start_pfn
, end_pfn
;
7129 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7130 unsigned long pages
= end_pfn
- start_pfn
;
7132 totalpages
+= pages
;
7134 node_set_state(nid
, N_MEMORY
);
7140 * Find the PFN the Movable zone begins in each node. Kernel memory
7141 * is spread evenly between nodes as long as the nodes have enough
7142 * memory. When they don't, some nodes will have more kernelcore than
7145 static void __init
find_zone_movable_pfns_for_nodes(void)
7148 unsigned long usable_startpfn
;
7149 unsigned long kernelcore_node
, kernelcore_remaining
;
7150 /* save the state before borrow the nodemask */
7151 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
7152 unsigned long totalpages
= early_calculate_totalpages();
7153 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
7154 struct memblock_region
*r
;
7156 /* Need to find movable_zone earlier when movable_node is specified. */
7157 find_usable_zone_for_movable();
7160 * If movable_node is specified, ignore kernelcore and movablecore
7163 if (movable_node_is_enabled()) {
7164 for_each_mem_region(r
) {
7165 if (!memblock_is_hotpluggable(r
))
7168 nid
= memblock_get_region_node(r
);
7170 usable_startpfn
= PFN_DOWN(r
->base
);
7171 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7172 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7180 * If kernelcore=mirror is specified, ignore movablecore option
7182 if (mirrored_kernelcore
) {
7183 bool mem_below_4gb_not_mirrored
= false;
7185 for_each_mem_region(r
) {
7186 if (memblock_is_mirror(r
))
7189 nid
= memblock_get_region_node(r
);
7191 usable_startpfn
= memblock_region_memory_base_pfn(r
);
7193 if (usable_startpfn
< 0x100000) {
7194 mem_below_4gb_not_mirrored
= true;
7198 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7199 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7203 if (mem_below_4gb_not_mirrored
)
7204 pr_warn("This configuration results in unmirrored kernel memory.\n");
7210 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7211 * amount of necessary memory.
7213 if (required_kernelcore_percent
)
7214 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
7216 if (required_movablecore_percent
)
7217 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
7221 * If movablecore= was specified, calculate what size of
7222 * kernelcore that corresponds so that memory usable for
7223 * any allocation type is evenly spread. If both kernelcore
7224 * and movablecore are specified, then the value of kernelcore
7225 * will be used for required_kernelcore if it's greater than
7226 * what movablecore would have allowed.
7228 if (required_movablecore
) {
7229 unsigned long corepages
;
7232 * Round-up so that ZONE_MOVABLE is at least as large as what
7233 * was requested by the user
7235 required_movablecore
=
7236 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
7237 required_movablecore
= min(totalpages
, required_movablecore
);
7238 corepages
= totalpages
- required_movablecore
;
7240 required_kernelcore
= max(required_kernelcore
, corepages
);
7244 * If kernelcore was not specified or kernelcore size is larger
7245 * than totalpages, there is no ZONE_MOVABLE.
7247 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
7250 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7251 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
7254 /* Spread kernelcore memory as evenly as possible throughout nodes */
7255 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7256 for_each_node_state(nid
, N_MEMORY
) {
7257 unsigned long start_pfn
, end_pfn
;
7260 * Recalculate kernelcore_node if the division per node
7261 * now exceeds what is necessary to satisfy the requested
7262 * amount of memory for the kernel
7264 if (required_kernelcore
< kernelcore_node
)
7265 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7268 * As the map is walked, we track how much memory is usable
7269 * by the kernel using kernelcore_remaining. When it is
7270 * 0, the rest of the node is usable by ZONE_MOVABLE
7272 kernelcore_remaining
= kernelcore_node
;
7274 /* Go through each range of PFNs within this node */
7275 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7276 unsigned long size_pages
;
7278 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
7279 if (start_pfn
>= end_pfn
)
7282 /* Account for what is only usable for kernelcore */
7283 if (start_pfn
< usable_startpfn
) {
7284 unsigned long kernel_pages
;
7285 kernel_pages
= min(end_pfn
, usable_startpfn
)
7288 kernelcore_remaining
-= min(kernel_pages
,
7289 kernelcore_remaining
);
7290 required_kernelcore
-= min(kernel_pages
,
7291 required_kernelcore
);
7293 /* Continue if range is now fully accounted */
7294 if (end_pfn
<= usable_startpfn
) {
7297 * Push zone_movable_pfn to the end so
7298 * that if we have to rebalance
7299 * kernelcore across nodes, we will
7300 * not double account here
7302 zone_movable_pfn
[nid
] = end_pfn
;
7305 start_pfn
= usable_startpfn
;
7309 * The usable PFN range for ZONE_MOVABLE is from
7310 * start_pfn->end_pfn. Calculate size_pages as the
7311 * number of pages used as kernelcore
7313 size_pages
= end_pfn
- start_pfn
;
7314 if (size_pages
> kernelcore_remaining
)
7315 size_pages
= kernelcore_remaining
;
7316 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7319 * Some kernelcore has been met, update counts and
7320 * break if the kernelcore for this node has been
7323 required_kernelcore
-= min(required_kernelcore
,
7325 kernelcore_remaining
-= size_pages
;
7326 if (!kernelcore_remaining
)
7332 * If there is still required_kernelcore, we do another pass with one
7333 * less node in the count. This will push zone_movable_pfn[nid] further
7334 * along on the nodes that still have memory until kernelcore is
7338 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7342 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7343 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7344 zone_movable_pfn
[nid
] =
7345 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7348 /* restore the node_state */
7349 node_states
[N_MEMORY
] = saved_node_state
;
7352 /* Any regular or high memory on that node ? */
7353 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7355 enum zone_type zone_type
;
7357 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7358 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7359 if (populated_zone(zone
)) {
7360 if (IS_ENABLED(CONFIG_HIGHMEM
))
7361 node_set_state(nid
, N_HIGH_MEMORY
);
7362 if (zone_type
<= ZONE_NORMAL
)
7363 node_set_state(nid
, N_NORMAL_MEMORY
);
7370 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7371 * such cases we allow max_zone_pfn sorted in the descending order
7373 bool __weak
arch_has_descending_max_zone_pfns(void)
7379 * free_area_init - Initialise all pg_data_t and zone data
7380 * @max_zone_pfn: an array of max PFNs for each zone
7382 * This will call free_area_init_node() for each active node in the system.
7383 * Using the page ranges provided by memblock_set_node(), the size of each
7384 * zone in each node and their holes is calculated. If the maximum PFN
7385 * between two adjacent zones match, it is assumed that the zone is empty.
7386 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7387 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7388 * starts where the previous one ended. For example, ZONE_DMA32 starts
7389 * at arch_max_dma_pfn.
7391 void __init
free_area_init(unsigned long *max_zone_pfn
)
7393 unsigned long start_pfn
, end_pfn
;
7397 /* Record where the zone boundaries are */
7398 memset(arch_zone_lowest_possible_pfn
, 0,
7399 sizeof(arch_zone_lowest_possible_pfn
));
7400 memset(arch_zone_highest_possible_pfn
, 0,
7401 sizeof(arch_zone_highest_possible_pfn
));
7403 start_pfn
= find_min_pfn_with_active_regions();
7404 descending
= arch_has_descending_max_zone_pfns();
7406 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7408 zone
= MAX_NR_ZONES
- i
- 1;
7412 if (zone
== ZONE_MOVABLE
)
7415 end_pfn
= max(max_zone_pfn
[zone
], start_pfn
);
7416 arch_zone_lowest_possible_pfn
[zone
] = start_pfn
;
7417 arch_zone_highest_possible_pfn
[zone
] = end_pfn
;
7419 start_pfn
= end_pfn
;
7422 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7423 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7424 find_zone_movable_pfns_for_nodes();
7426 /* Print out the zone ranges */
7427 pr_info("Zone ranges:\n");
7428 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7429 if (i
== ZONE_MOVABLE
)
7431 pr_info(" %-8s ", zone_names
[i
]);
7432 if (arch_zone_lowest_possible_pfn
[i
] ==
7433 arch_zone_highest_possible_pfn
[i
])
7436 pr_cont("[mem %#018Lx-%#018Lx]\n",
7437 (u64
)arch_zone_lowest_possible_pfn
[i
]
7439 ((u64
)arch_zone_highest_possible_pfn
[i
]
7440 << PAGE_SHIFT
) - 1);
7443 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7444 pr_info("Movable zone start for each node\n");
7445 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7446 if (zone_movable_pfn
[i
])
7447 pr_info(" Node %d: %#018Lx\n", i
,
7448 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7452 * Print out the early node map, and initialize the
7453 * subsection-map relative to active online memory ranges to
7454 * enable future "sub-section" extensions of the memory map.
7456 pr_info("Early memory node ranges\n");
7457 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7458 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7459 (u64
)start_pfn
<< PAGE_SHIFT
,
7460 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7461 subsection_map_init(start_pfn
, end_pfn
- start_pfn
);
7464 /* Initialise every node */
7465 mminit_verify_pageflags_layout();
7466 setup_nr_node_ids();
7467 init_unavailable_mem();
7468 for_each_online_node(nid
) {
7469 pg_data_t
*pgdat
= NODE_DATA(nid
);
7470 free_area_init_node(nid
);
7472 /* Any memory on that node */
7473 if (pgdat
->node_present_pages
)
7474 node_set_state(nid
, N_MEMORY
);
7475 check_for_memory(pgdat
, nid
);
7479 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7480 unsigned long *percent
)
7482 unsigned long long coremem
;
7488 /* Value may be a percentage of total memory, otherwise bytes */
7489 coremem
= simple_strtoull(p
, &endptr
, 0);
7490 if (*endptr
== '%') {
7491 /* Paranoid check for percent values greater than 100 */
7492 WARN_ON(coremem
> 100);
7496 coremem
= memparse(p
, &p
);
7497 /* Paranoid check that UL is enough for the coremem value */
7498 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7500 *core
= coremem
>> PAGE_SHIFT
;
7507 * kernelcore=size sets the amount of memory for use for allocations that
7508 * cannot be reclaimed or migrated.
7510 static int __init
cmdline_parse_kernelcore(char *p
)
7512 /* parse kernelcore=mirror */
7513 if (parse_option_str(p
, "mirror")) {
7514 mirrored_kernelcore
= true;
7518 return cmdline_parse_core(p
, &required_kernelcore
,
7519 &required_kernelcore_percent
);
7523 * movablecore=size sets the amount of memory for use for allocations that
7524 * can be reclaimed or migrated.
7526 static int __init
cmdline_parse_movablecore(char *p
)
7528 return cmdline_parse_core(p
, &required_movablecore
,
7529 &required_movablecore_percent
);
7532 early_param("kernelcore", cmdline_parse_kernelcore
);
7533 early_param("movablecore", cmdline_parse_movablecore
);
7535 void adjust_managed_page_count(struct page
*page
, long count
)
7537 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7538 totalram_pages_add(count
);
7539 #ifdef CONFIG_HIGHMEM
7540 if (PageHighMem(page
))
7541 totalhigh_pages_add(count
);
7544 EXPORT_SYMBOL(adjust_managed_page_count
);
7546 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7549 unsigned long pages
= 0;
7551 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7552 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7553 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7554 struct page
*page
= virt_to_page(pos
);
7555 void *direct_map_addr
;
7558 * 'direct_map_addr' might be different from 'pos'
7559 * because some architectures' virt_to_page()
7560 * work with aliases. Getting the direct map
7561 * address ensures that we get a _writeable_
7562 * alias for the memset().
7564 direct_map_addr
= page_address(page
);
7565 if ((unsigned int)poison
<= 0xFF)
7566 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7568 free_reserved_page(page
);
7572 pr_info("Freeing %s memory: %ldK\n",
7573 s
, pages
<< (PAGE_SHIFT
- 10));
7578 #ifdef CONFIG_HIGHMEM
7579 void free_highmem_page(struct page
*page
)
7581 __free_reserved_page(page
);
7582 totalram_pages_inc();
7583 atomic_long_inc(&page_zone(page
)->managed_pages
);
7584 totalhigh_pages_inc();
7589 void __init
mem_init_print_info(const char *str
)
7591 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7592 unsigned long init_code_size
, init_data_size
;
7594 physpages
= get_num_physpages();
7595 codesize
= _etext
- _stext
;
7596 datasize
= _edata
- _sdata
;
7597 rosize
= __end_rodata
- __start_rodata
;
7598 bss_size
= __bss_stop
- __bss_start
;
7599 init_data_size
= __init_end
- __init_begin
;
7600 init_code_size
= _einittext
- _sinittext
;
7603 * Detect special cases and adjust section sizes accordingly:
7604 * 1) .init.* may be embedded into .data sections
7605 * 2) .init.text.* may be out of [__init_begin, __init_end],
7606 * please refer to arch/tile/kernel/vmlinux.lds.S.
7607 * 3) .rodata.* may be embedded into .text or .data sections.
7609 #define adj_init_size(start, end, size, pos, adj) \
7611 if (start <= pos && pos < end && size > adj) \
7615 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7616 _sinittext
, init_code_size
);
7617 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7618 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7619 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7620 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7622 #undef adj_init_size
7624 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7625 #ifdef CONFIG_HIGHMEM
7629 nr_free_pages() << (PAGE_SHIFT
- 10),
7630 physpages
<< (PAGE_SHIFT
- 10),
7631 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7632 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7633 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7634 totalcma_pages
<< (PAGE_SHIFT
- 10),
7635 #ifdef CONFIG_HIGHMEM
7636 totalhigh_pages() << (PAGE_SHIFT
- 10),
7638 str
? ", " : "", str
? str
: "");
7642 * set_dma_reserve - set the specified number of pages reserved in the first zone
7643 * @new_dma_reserve: The number of pages to mark reserved
7645 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7646 * In the DMA zone, a significant percentage may be consumed by kernel image
7647 * and other unfreeable allocations which can skew the watermarks badly. This
7648 * function may optionally be used to account for unfreeable pages in the
7649 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7650 * smaller per-cpu batchsize.
7652 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7654 dma_reserve
= new_dma_reserve
;
7657 static int page_alloc_cpu_dead(unsigned int cpu
)
7660 lru_add_drain_cpu(cpu
);
7664 * Spill the event counters of the dead processor
7665 * into the current processors event counters.
7666 * This artificially elevates the count of the current
7669 vm_events_fold_cpu(cpu
);
7672 * Zero the differential counters of the dead processor
7673 * so that the vm statistics are consistent.
7675 * This is only okay since the processor is dead and cannot
7676 * race with what we are doing.
7678 cpu_vm_stats_fold(cpu
);
7683 int hashdist
= HASHDIST_DEFAULT
;
7685 static int __init
set_hashdist(char *str
)
7689 hashdist
= simple_strtoul(str
, &str
, 0);
7692 __setup("hashdist=", set_hashdist
);
7695 void __init
page_alloc_init(void)
7700 if (num_node_state(N_MEMORY
) == 1)
7704 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7705 "mm/page_alloc:dead", NULL
,
7706 page_alloc_cpu_dead
);
7711 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7712 * or min_free_kbytes changes.
7714 static void calculate_totalreserve_pages(void)
7716 struct pglist_data
*pgdat
;
7717 unsigned long reserve_pages
= 0;
7718 enum zone_type i
, j
;
7720 for_each_online_pgdat(pgdat
) {
7722 pgdat
->totalreserve_pages
= 0;
7724 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7725 struct zone
*zone
= pgdat
->node_zones
+ i
;
7727 unsigned long managed_pages
= zone_managed_pages(zone
);
7729 /* Find valid and maximum lowmem_reserve in the zone */
7730 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7731 if (zone
->lowmem_reserve
[j
] > max
)
7732 max
= zone
->lowmem_reserve
[j
];
7735 /* we treat the high watermark as reserved pages. */
7736 max
+= high_wmark_pages(zone
);
7738 if (max
> managed_pages
)
7739 max
= managed_pages
;
7741 pgdat
->totalreserve_pages
+= max
;
7743 reserve_pages
+= max
;
7746 totalreserve_pages
= reserve_pages
;
7750 * setup_per_zone_lowmem_reserve - called whenever
7751 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7752 * has a correct pages reserved value, so an adequate number of
7753 * pages are left in the zone after a successful __alloc_pages().
7755 static void setup_per_zone_lowmem_reserve(void)
7757 struct pglist_data
*pgdat
;
7758 enum zone_type j
, idx
;
7760 for_each_online_pgdat(pgdat
) {
7761 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7762 struct zone
*zone
= pgdat
->node_zones
+ j
;
7763 unsigned long managed_pages
= zone_managed_pages(zone
);
7765 zone
->lowmem_reserve
[j
] = 0;
7769 struct zone
*lower_zone
;
7772 lower_zone
= pgdat
->node_zones
+ idx
;
7774 if (!sysctl_lowmem_reserve_ratio
[idx
] ||
7775 !zone_managed_pages(lower_zone
)) {
7776 lower_zone
->lowmem_reserve
[j
] = 0;
7779 lower_zone
->lowmem_reserve
[j
] =
7780 managed_pages
/ sysctl_lowmem_reserve_ratio
[idx
];
7782 managed_pages
+= zone_managed_pages(lower_zone
);
7787 /* update totalreserve_pages */
7788 calculate_totalreserve_pages();
7791 static void __setup_per_zone_wmarks(void)
7793 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7794 unsigned long lowmem_pages
= 0;
7796 unsigned long flags
;
7798 /* Calculate total number of !ZONE_HIGHMEM pages */
7799 for_each_zone(zone
) {
7800 if (!is_highmem(zone
))
7801 lowmem_pages
+= zone_managed_pages(zone
);
7804 for_each_zone(zone
) {
7807 spin_lock_irqsave(&zone
->lock
, flags
);
7808 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7809 do_div(tmp
, lowmem_pages
);
7810 if (is_highmem(zone
)) {
7812 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7813 * need highmem pages, so cap pages_min to a small
7816 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7817 * deltas control async page reclaim, and so should
7818 * not be capped for highmem.
7820 unsigned long min_pages
;
7822 min_pages
= zone_managed_pages(zone
) / 1024;
7823 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7824 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7827 * If it's a lowmem zone, reserve a number of pages
7828 * proportionate to the zone's size.
7830 zone
->_watermark
[WMARK_MIN
] = tmp
;
7834 * Set the kswapd watermarks distance according to the
7835 * scale factor in proportion to available memory, but
7836 * ensure a minimum size on small systems.
7838 tmp
= max_t(u64
, tmp
>> 2,
7839 mult_frac(zone_managed_pages(zone
),
7840 watermark_scale_factor
, 10000));
7842 zone
->watermark_boost
= 0;
7843 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7844 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7846 spin_unlock_irqrestore(&zone
->lock
, flags
);
7849 /* update totalreserve_pages */
7850 calculate_totalreserve_pages();
7854 * setup_per_zone_wmarks - called when min_free_kbytes changes
7855 * or when memory is hot-{added|removed}
7857 * Ensures that the watermark[min,low,high] values for each zone are set
7858 * correctly with respect to min_free_kbytes.
7860 void setup_per_zone_wmarks(void)
7862 static DEFINE_SPINLOCK(lock
);
7865 __setup_per_zone_wmarks();
7870 * Initialise min_free_kbytes.
7872 * For small machines we want it small (128k min). For large machines
7873 * we want it large (256MB max). But it is not linear, because network
7874 * bandwidth does not increase linearly with machine size. We use
7876 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7877 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7893 int __meminit
init_per_zone_wmark_min(void)
7895 unsigned long lowmem_kbytes
;
7896 int new_min_free_kbytes
;
7898 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7899 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7901 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7902 min_free_kbytes
= new_min_free_kbytes
;
7903 if (min_free_kbytes
< 128)
7904 min_free_kbytes
= 128;
7905 if (min_free_kbytes
> 262144)
7906 min_free_kbytes
= 262144;
7908 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7909 new_min_free_kbytes
, user_min_free_kbytes
);
7911 setup_per_zone_wmarks();
7912 refresh_zone_stat_thresholds();
7913 setup_per_zone_lowmem_reserve();
7916 setup_min_unmapped_ratio();
7917 setup_min_slab_ratio();
7920 khugepaged_min_free_kbytes_update();
7924 postcore_initcall(init_per_zone_wmark_min
)
7927 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7928 * that we can call two helper functions whenever min_free_kbytes
7931 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7932 void *buffer
, size_t *length
, loff_t
*ppos
)
7936 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7941 user_min_free_kbytes
= min_free_kbytes
;
7942 setup_per_zone_wmarks();
7947 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7948 void *buffer
, size_t *length
, loff_t
*ppos
)
7952 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7957 setup_per_zone_wmarks();
7963 static void setup_min_unmapped_ratio(void)
7968 for_each_online_pgdat(pgdat
)
7969 pgdat
->min_unmapped_pages
= 0;
7972 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
7973 sysctl_min_unmapped_ratio
) / 100;
7977 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7978 void *buffer
, size_t *length
, loff_t
*ppos
)
7982 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7986 setup_min_unmapped_ratio();
7991 static void setup_min_slab_ratio(void)
7996 for_each_online_pgdat(pgdat
)
7997 pgdat
->min_slab_pages
= 0;
8000 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
8001 sysctl_min_slab_ratio
) / 100;
8004 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8005 void *buffer
, size_t *length
, loff_t
*ppos
)
8009 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8013 setup_min_slab_ratio();
8020 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8021 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8022 * whenever sysctl_lowmem_reserve_ratio changes.
8024 * The reserve ratio obviously has absolutely no relation with the
8025 * minimum watermarks. The lowmem reserve ratio can only make sense
8026 * if in function of the boot time zone sizes.
8028 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8029 void *buffer
, size_t *length
, loff_t
*ppos
)
8033 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8035 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
8036 if (sysctl_lowmem_reserve_ratio
[i
] < 1)
8037 sysctl_lowmem_reserve_ratio
[i
] = 0;
8040 setup_per_zone_lowmem_reserve();
8044 static void __zone_pcp_update(struct zone
*zone
)
8048 for_each_possible_cpu(cpu
)
8049 pageset_set_high_and_batch(zone
,
8050 per_cpu_ptr(zone
->pageset
, cpu
));
8054 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8055 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8056 * pagelist can have before it gets flushed back to buddy allocator.
8058 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
8059 void *buffer
, size_t *length
, loff_t
*ppos
)
8062 int old_percpu_pagelist_fraction
;
8065 mutex_lock(&pcp_batch_high_lock
);
8066 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
8068 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8069 if (!write
|| ret
< 0)
8072 /* Sanity checking to avoid pcp imbalance */
8073 if (percpu_pagelist_fraction
&&
8074 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
8075 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
8081 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
8084 for_each_populated_zone(zone
)
8085 __zone_pcp_update(zone
);
8087 mutex_unlock(&pcp_batch_high_lock
);
8091 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8093 * Returns the number of pages that arch has reserved but
8094 * is not known to alloc_large_system_hash().
8096 static unsigned long __init
arch_reserved_kernel_pages(void)
8103 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8104 * machines. As memory size is increased the scale is also increased but at
8105 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8106 * quadruples the scale is increased by one, which means the size of hash table
8107 * only doubles, instead of quadrupling as well.
8108 * Because 32-bit systems cannot have large physical memory, where this scaling
8109 * makes sense, it is disabled on such platforms.
8111 #if __BITS_PER_LONG > 32
8112 #define ADAPT_SCALE_BASE (64ul << 30)
8113 #define ADAPT_SCALE_SHIFT 2
8114 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8118 * allocate a large system hash table from bootmem
8119 * - it is assumed that the hash table must contain an exact power-of-2
8120 * quantity of entries
8121 * - limit is the number of hash buckets, not the total allocation size
8123 void *__init
alloc_large_system_hash(const char *tablename
,
8124 unsigned long bucketsize
,
8125 unsigned long numentries
,
8128 unsigned int *_hash_shift
,
8129 unsigned int *_hash_mask
,
8130 unsigned long low_limit
,
8131 unsigned long high_limit
)
8133 unsigned long long max
= high_limit
;
8134 unsigned long log2qty
, size
;
8139 /* allow the kernel cmdline to have a say */
8141 /* round applicable memory size up to nearest megabyte */
8142 numentries
= nr_kernel_pages
;
8143 numentries
-= arch_reserved_kernel_pages();
8145 /* It isn't necessary when PAGE_SIZE >= 1MB */
8146 if (PAGE_SHIFT
< 20)
8147 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
8149 #if __BITS_PER_LONG > 32
8151 unsigned long adapt
;
8153 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
8154 adapt
<<= ADAPT_SCALE_SHIFT
)
8159 /* limit to 1 bucket per 2^scale bytes of low memory */
8160 if (scale
> PAGE_SHIFT
)
8161 numentries
>>= (scale
- PAGE_SHIFT
);
8163 numentries
<<= (PAGE_SHIFT
- scale
);
8165 /* Make sure we've got at least a 0-order allocation.. */
8166 if (unlikely(flags
& HASH_SMALL
)) {
8167 /* Makes no sense without HASH_EARLY */
8168 WARN_ON(!(flags
& HASH_EARLY
));
8169 if (!(numentries
>> *_hash_shift
)) {
8170 numentries
= 1UL << *_hash_shift
;
8171 BUG_ON(!numentries
);
8173 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
8174 numentries
= PAGE_SIZE
/ bucketsize
;
8176 numentries
= roundup_pow_of_two(numentries
);
8178 /* limit allocation size to 1/16 total memory by default */
8180 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
8181 do_div(max
, bucketsize
);
8183 max
= min(max
, 0x80000000ULL
);
8185 if (numentries
< low_limit
)
8186 numentries
= low_limit
;
8187 if (numentries
> max
)
8190 log2qty
= ilog2(numentries
);
8192 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
8195 size
= bucketsize
<< log2qty
;
8196 if (flags
& HASH_EARLY
) {
8197 if (flags
& HASH_ZERO
)
8198 table
= memblock_alloc(size
, SMP_CACHE_BYTES
);
8200 table
= memblock_alloc_raw(size
,
8202 } else if (get_order(size
) >= MAX_ORDER
|| hashdist
) {
8203 table
= __vmalloc(size
, gfp_flags
);
8207 * If bucketsize is not a power-of-two, we may free
8208 * some pages at the end of hash table which
8209 * alloc_pages_exact() automatically does
8211 table
= alloc_pages_exact(size
, gfp_flags
);
8212 kmemleak_alloc(table
, size
, 1, gfp_flags
);
8214 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
8217 panic("Failed to allocate %s hash table\n", tablename
);
8219 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8220 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
,
8221 virt
? "vmalloc" : "linear");
8224 *_hash_shift
= log2qty
;
8226 *_hash_mask
= (1 << log2qty
) - 1;
8232 * This function checks whether pageblock includes unmovable pages or not.
8234 * PageLRU check without isolation or lru_lock could race so that
8235 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8236 * check without lock_page also may miss some movable non-lru pages at
8237 * race condition. So you can't expect this function should be exact.
8239 * Returns a page without holding a reference. If the caller wants to
8240 * dereference that page (e.g., dumping), it has to make sure that it
8241 * cannot get removed (e.g., via memory unplug) concurrently.
8244 struct page
*has_unmovable_pages(struct zone
*zone
, struct page
*page
,
8245 int migratetype
, int flags
)
8247 unsigned long iter
= 0;
8248 unsigned long pfn
= page_to_pfn(page
);
8249 unsigned long offset
= pfn
% pageblock_nr_pages
;
8251 if (is_migrate_cma_page(page
)) {
8253 * CMA allocations (alloc_contig_range) really need to mark
8254 * isolate CMA pageblocks even when they are not movable in fact
8255 * so consider them movable here.
8257 if (is_migrate_cma(migratetype
))
8263 for (; iter
< pageblock_nr_pages
- offset
; iter
++) {
8264 if (!pfn_valid_within(pfn
+ iter
))
8267 page
= pfn_to_page(pfn
+ iter
);
8270 * Both, bootmem allocations and memory holes are marked
8271 * PG_reserved and are unmovable. We can even have unmovable
8272 * allocations inside ZONE_MOVABLE, for example when
8273 * specifying "movablecore".
8275 if (PageReserved(page
))
8279 * If the zone is movable and we have ruled out all reserved
8280 * pages then it should be reasonably safe to assume the rest
8283 if (zone_idx(zone
) == ZONE_MOVABLE
)
8287 * Hugepages are not in LRU lists, but they're movable.
8288 * THPs are on the LRU, but need to be counted as #small pages.
8289 * We need not scan over tail pages because we don't
8290 * handle each tail page individually in migration.
8292 if (PageHuge(page
) || PageTransCompound(page
)) {
8293 struct page
*head
= compound_head(page
);
8294 unsigned int skip_pages
;
8296 if (PageHuge(page
)) {
8297 if (!hugepage_migration_supported(page_hstate(head
)))
8299 } else if (!PageLRU(head
) && !__PageMovable(head
)) {
8303 skip_pages
= compound_nr(head
) - (page
- head
);
8304 iter
+= skip_pages
- 1;
8309 * We can't use page_count without pin a page
8310 * because another CPU can free compound page.
8311 * This check already skips compound tails of THP
8312 * because their page->_refcount is zero at all time.
8314 if (!page_ref_count(page
)) {
8315 if (PageBuddy(page
))
8316 iter
+= (1 << page_order(page
)) - 1;
8321 * The HWPoisoned page may be not in buddy system, and
8322 * page_count() is not 0.
8324 if ((flags
& MEMORY_OFFLINE
) && PageHWPoison(page
))
8328 * We treat all PageOffline() pages as movable when offlining
8329 * to give drivers a chance to decrement their reference count
8330 * in MEM_GOING_OFFLINE in order to indicate that these pages
8331 * can be offlined as there are no direct references anymore.
8332 * For actually unmovable PageOffline() where the driver does
8333 * not support this, we will fail later when trying to actually
8334 * move these pages that still have a reference count > 0.
8335 * (false negatives in this function only)
8337 if ((flags
& MEMORY_OFFLINE
) && PageOffline(page
))
8340 if (__PageMovable(page
) || PageLRU(page
))
8344 * If there are RECLAIMABLE pages, we need to check
8345 * it. But now, memory offline itself doesn't call
8346 * shrink_node_slabs() and it still to be fixed.
8353 #ifdef CONFIG_CONTIG_ALLOC
8354 static unsigned long pfn_max_align_down(unsigned long pfn
)
8356 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8357 pageblock_nr_pages
) - 1);
8360 static unsigned long pfn_max_align_up(unsigned long pfn
)
8362 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8363 pageblock_nr_pages
));
8366 /* [start, end) must belong to a single zone. */
8367 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8368 unsigned long start
, unsigned long end
)
8370 /* This function is based on compact_zone() from compaction.c. */
8371 unsigned int nr_reclaimed
;
8372 unsigned long pfn
= start
;
8373 unsigned int tries
= 0;
8375 struct migration_target_control mtc
= {
8376 .nid
= zone_to_nid(cc
->zone
),
8377 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
8382 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8383 if (fatal_signal_pending(current
)) {
8388 if (list_empty(&cc
->migratepages
)) {
8389 cc
->nr_migratepages
= 0;
8390 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8396 } else if (++tries
== 5) {
8397 ret
= ret
< 0 ? ret
: -EBUSY
;
8401 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8403 cc
->nr_migratepages
-= nr_reclaimed
;
8405 ret
= migrate_pages(&cc
->migratepages
, alloc_migration_target
,
8406 NULL
, (unsigned long)&mtc
, cc
->mode
, MR_CONTIG_RANGE
);
8409 putback_movable_pages(&cc
->migratepages
);
8416 * alloc_contig_range() -- tries to allocate given range of pages
8417 * @start: start PFN to allocate
8418 * @end: one-past-the-last PFN to allocate
8419 * @migratetype: migratetype of the underlaying pageblocks (either
8420 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8421 * in range must have the same migratetype and it must
8422 * be either of the two.
8423 * @gfp_mask: GFP mask to use during compaction
8425 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8426 * aligned. The PFN range must belong to a single zone.
8428 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8429 * pageblocks in the range. Once isolated, the pageblocks should not
8430 * be modified by others.
8432 * Return: zero on success or negative error code. On success all
8433 * pages which PFN is in [start, end) are allocated for the caller and
8434 * need to be freed with free_contig_range().
8436 int alloc_contig_range(unsigned long start
, unsigned long end
,
8437 unsigned migratetype
, gfp_t gfp_mask
)
8439 unsigned long outer_start
, outer_end
;
8443 struct compact_control cc
= {
8444 .nr_migratepages
= 0,
8446 .zone
= page_zone(pfn_to_page(start
)),
8447 .mode
= MIGRATE_SYNC
,
8448 .ignore_skip_hint
= true,
8449 .no_set_skip_hint
= true,
8450 .gfp_mask
= current_gfp_context(gfp_mask
),
8451 .alloc_contig
= true,
8453 INIT_LIST_HEAD(&cc
.migratepages
);
8456 * What we do here is we mark all pageblocks in range as
8457 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8458 * have different sizes, and due to the way page allocator
8459 * work, we align the range to biggest of the two pages so
8460 * that page allocator won't try to merge buddies from
8461 * different pageblocks and change MIGRATE_ISOLATE to some
8462 * other migration type.
8464 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8465 * migrate the pages from an unaligned range (ie. pages that
8466 * we are interested in). This will put all the pages in
8467 * range back to page allocator as MIGRATE_ISOLATE.
8469 * When this is done, we take the pages in range from page
8470 * allocator removing them from the buddy system. This way
8471 * page allocator will never consider using them.
8473 * This lets us mark the pageblocks back as
8474 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8475 * aligned range but not in the unaligned, original range are
8476 * put back to page allocator so that buddy can use them.
8479 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8480 pfn_max_align_up(end
), migratetype
, 0);
8485 * In case of -EBUSY, we'd like to know which page causes problem.
8486 * So, just fall through. test_pages_isolated() has a tracepoint
8487 * which will report the busy page.
8489 * It is possible that busy pages could become available before
8490 * the call to test_pages_isolated, and the range will actually be
8491 * allocated. So, if we fall through be sure to clear ret so that
8492 * -EBUSY is not accidentally used or returned to caller.
8494 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8495 if (ret
&& ret
!= -EBUSY
)
8500 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8501 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8502 * more, all pages in [start, end) are free in page allocator.
8503 * What we are going to do is to allocate all pages from
8504 * [start, end) (that is remove them from page allocator).
8506 * The only problem is that pages at the beginning and at the
8507 * end of interesting range may be not aligned with pages that
8508 * page allocator holds, ie. they can be part of higher order
8509 * pages. Because of this, we reserve the bigger range and
8510 * once this is done free the pages we are not interested in.
8512 * We don't have to hold zone->lock here because the pages are
8513 * isolated thus they won't get removed from buddy.
8516 lru_add_drain_all();
8519 outer_start
= start
;
8520 while (!PageBuddy(pfn_to_page(outer_start
))) {
8521 if (++order
>= MAX_ORDER
) {
8522 outer_start
= start
;
8525 outer_start
&= ~0UL << order
;
8528 if (outer_start
!= start
) {
8529 order
= page_order(pfn_to_page(outer_start
));
8532 * outer_start page could be small order buddy page and
8533 * it doesn't include start page. Adjust outer_start
8534 * in this case to report failed page properly
8535 * on tracepoint in test_pages_isolated()
8537 if (outer_start
+ (1UL << order
) <= start
)
8538 outer_start
= start
;
8541 /* Make sure the range is really isolated. */
8542 if (test_pages_isolated(outer_start
, end
, 0)) {
8543 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8544 __func__
, outer_start
, end
);
8549 /* Grab isolated pages from freelists. */
8550 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8556 /* Free head and tail (if any) */
8557 if (start
!= outer_start
)
8558 free_contig_range(outer_start
, start
- outer_start
);
8559 if (end
!= outer_end
)
8560 free_contig_range(end
, outer_end
- end
);
8563 undo_isolate_page_range(pfn_max_align_down(start
),
8564 pfn_max_align_up(end
), migratetype
);
8567 EXPORT_SYMBOL(alloc_contig_range
);
8569 static int __alloc_contig_pages(unsigned long start_pfn
,
8570 unsigned long nr_pages
, gfp_t gfp_mask
)
8572 unsigned long end_pfn
= start_pfn
+ nr_pages
;
8574 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
8578 static bool pfn_range_valid_contig(struct zone
*z
, unsigned long start_pfn
,
8579 unsigned long nr_pages
)
8581 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
8584 for (i
= start_pfn
; i
< end_pfn
; i
++) {
8585 page
= pfn_to_online_page(i
);
8589 if (page_zone(page
) != z
)
8592 if (PageReserved(page
))
8595 if (page_count(page
) > 0)
8604 static bool zone_spans_last_pfn(const struct zone
*zone
,
8605 unsigned long start_pfn
, unsigned long nr_pages
)
8607 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
8609 return zone_spans_pfn(zone
, last_pfn
);
8613 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8614 * @nr_pages: Number of contiguous pages to allocate
8615 * @gfp_mask: GFP mask to limit search and used during compaction
8617 * @nodemask: Mask for other possible nodes
8619 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8620 * on an applicable zonelist to find a contiguous pfn range which can then be
8621 * tried for allocation with alloc_contig_range(). This routine is intended
8622 * for allocation requests which can not be fulfilled with the buddy allocator.
8624 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8625 * power of two then the alignment is guaranteed to be to the given nr_pages
8626 * (e.g. 1GB request would be aligned to 1GB).
8628 * Allocated pages can be freed with free_contig_range() or by manually calling
8629 * __free_page() on each allocated page.
8631 * Return: pointer to contiguous pages on success, or NULL if not successful.
8633 struct page
*alloc_contig_pages(unsigned long nr_pages
, gfp_t gfp_mask
,
8634 int nid
, nodemask_t
*nodemask
)
8636 unsigned long ret
, pfn
, flags
;
8637 struct zonelist
*zonelist
;
8641 zonelist
= node_zonelist(nid
, gfp_mask
);
8642 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
8643 gfp_zone(gfp_mask
), nodemask
) {
8644 spin_lock_irqsave(&zone
->lock
, flags
);
8646 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
8647 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
8648 if (pfn_range_valid_contig(zone
, pfn
, nr_pages
)) {
8650 * We release the zone lock here because
8651 * alloc_contig_range() will also lock the zone
8652 * at some point. If there's an allocation
8653 * spinning on this lock, it may win the race
8654 * and cause alloc_contig_range() to fail...
8656 spin_unlock_irqrestore(&zone
->lock
, flags
);
8657 ret
= __alloc_contig_pages(pfn
, nr_pages
,
8660 return pfn_to_page(pfn
);
8661 spin_lock_irqsave(&zone
->lock
, flags
);
8665 spin_unlock_irqrestore(&zone
->lock
, flags
);
8669 #endif /* CONFIG_CONTIG_ALLOC */
8671 void free_contig_range(unsigned long pfn
, unsigned int nr_pages
)
8673 unsigned int count
= 0;
8675 for (; nr_pages
--; pfn
++) {
8676 struct page
*page
= pfn_to_page(pfn
);
8678 count
+= page_count(page
) != 1;
8681 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8683 EXPORT_SYMBOL(free_contig_range
);
8686 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8687 * page high values need to be recalulated.
8689 void __meminit
zone_pcp_update(struct zone
*zone
)
8691 mutex_lock(&pcp_batch_high_lock
);
8692 __zone_pcp_update(zone
);
8693 mutex_unlock(&pcp_batch_high_lock
);
8696 void zone_pcp_reset(struct zone
*zone
)
8698 unsigned long flags
;
8700 struct per_cpu_pageset
*pset
;
8702 /* avoid races with drain_pages() */
8703 local_irq_save(flags
);
8704 if (zone
->pageset
!= &boot_pageset
) {
8705 for_each_online_cpu(cpu
) {
8706 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8707 drain_zonestat(zone
, pset
);
8709 free_percpu(zone
->pageset
);
8710 zone
->pageset
= &boot_pageset
;
8712 local_irq_restore(flags
);
8715 #ifdef CONFIG_MEMORY_HOTREMOVE
8717 * All pages in the range must be in a single zone, must not contain holes,
8718 * must span full sections, and must be isolated before calling this function.
8720 void __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8722 unsigned long pfn
= start_pfn
;
8726 unsigned long flags
;
8728 offline_mem_sections(pfn
, end_pfn
);
8729 zone
= page_zone(pfn_to_page(pfn
));
8730 spin_lock_irqsave(&zone
->lock
, flags
);
8731 while (pfn
< end_pfn
) {
8732 page
= pfn_to_page(pfn
);
8734 * The HWPoisoned page may be not in buddy system, and
8735 * page_count() is not 0.
8737 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8742 * At this point all remaining PageOffline() pages have a
8743 * reference count of 0 and can simply be skipped.
8745 if (PageOffline(page
)) {
8746 BUG_ON(page_count(page
));
8747 BUG_ON(PageBuddy(page
));
8752 BUG_ON(page_count(page
));
8753 BUG_ON(!PageBuddy(page
));
8754 order
= page_order(page
);
8755 del_page_from_free_list(page
, zone
, order
);
8756 pfn
+= (1 << order
);
8758 spin_unlock_irqrestore(&zone
->lock
, flags
);
8762 bool is_free_buddy_page(struct page
*page
)
8764 struct zone
*zone
= page_zone(page
);
8765 unsigned long pfn
= page_to_pfn(page
);
8766 unsigned long flags
;
8769 spin_lock_irqsave(&zone
->lock
, flags
);
8770 for (order
= 0; order
< MAX_ORDER
; order
++) {
8771 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8773 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
8776 spin_unlock_irqrestore(&zone
->lock
, flags
);
8778 return order
< MAX_ORDER
;
8781 #ifdef CONFIG_MEMORY_FAILURE
8783 * Break down a higher-order page in sub-pages, and keep our target out of
8786 static void break_down_buddy_pages(struct zone
*zone
, struct page
*page
,
8787 struct page
*target
, int low
, int high
,
8790 unsigned long size
= 1 << high
;
8791 struct page
*current_buddy
, *next_page
;
8793 while (high
> low
) {
8797 if (target
>= &page
[size
]) {
8798 next_page
= page
+ size
;
8799 current_buddy
= page
;
8802 current_buddy
= page
+ size
;
8805 if (set_page_guard(zone
, current_buddy
, high
, migratetype
))
8808 if (current_buddy
!= target
) {
8809 add_to_free_list(current_buddy
, zone
, high
, migratetype
);
8810 set_page_order(current_buddy
, high
);
8817 * Take a page that will be marked as poisoned off the buddy allocator.
8819 bool take_page_off_buddy(struct page
*page
)
8821 struct zone
*zone
= page_zone(page
);
8822 unsigned long pfn
= page_to_pfn(page
);
8823 unsigned long flags
;
8827 spin_lock_irqsave(&zone
->lock
, flags
);
8828 for (order
= 0; order
< MAX_ORDER
; order
++) {
8829 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8830 int buddy_order
= page_order(page_head
);
8832 if (PageBuddy(page_head
) && buddy_order
>= order
) {
8833 unsigned long pfn_head
= page_to_pfn(page_head
);
8834 int migratetype
= get_pfnblock_migratetype(page_head
,
8837 del_page_from_free_list(page_head
, zone
, buddy_order
);
8838 break_down_buddy_pages(zone
, page_head
, page
, 0,
8839 buddy_order
, migratetype
);
8843 if (page_count(page_head
) > 0)
8846 spin_unlock_irqrestore(&zone
->lock
, flags
);