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/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
74 #include <linux/buffer_head.h>
75 #include <asm/sections.h>
76 #include <asm/tlbflush.h>
77 #include <asm/div64.h>
80 #include "page_reporting.h"
82 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
83 typedef int __bitwise fpi_t
;
85 /* No special request */
86 #define FPI_NONE ((__force fpi_t)0)
89 * Skip free page reporting notification for the (possibly merged) page.
90 * This does not hinder free page reporting from grabbing the page,
91 * reporting it and marking it "reported" - it only skips notifying
92 * the free page reporting infrastructure about a newly freed page. For
93 * example, used when temporarily pulling a page from a freelist and
94 * putting it back unmodified.
96 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
99 * Place the (possibly merged) page to the tail of the freelist. Will ignore
100 * page shuffling (relevant code - e.g., memory onlining - is expected to
101 * shuffle the whole zone).
103 * Note: No code should rely on this flag for correctness - it's purely
104 * to allow for optimizations when handing back either fresh pages
105 * (memory onlining) or untouched pages (page isolation, free page
108 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
111 * Don't poison memory with KASAN (only for the tag-based modes).
112 * During boot, all non-reserved memblock memory is exposed to page_alloc.
113 * Poisoning all that memory lengthens boot time, especially on systems with
114 * large amount of RAM. This flag is used to skip that poisoning.
115 * This is only done for the tag-based KASAN modes, as those are able to
116 * detect memory corruptions with the memory tags assigned by default.
117 * All memory allocated normally after boot gets poisoned as usual.
119 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
121 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
122 static DEFINE_MUTEX(pcp_batch_high_lock
);
123 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
128 static DEFINE_PER_CPU(struct pagesets
, pagesets
) = {
129 .lock
= INIT_LOCAL_LOCK(lock
),
132 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
133 DEFINE_PER_CPU(int, numa_node
);
134 EXPORT_PER_CPU_SYMBOL(numa_node
);
137 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
139 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
141 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
142 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
143 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
144 * defined in <linux/topology.h>.
146 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
147 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
150 /* work_structs for global per-cpu drains */
153 struct work_struct work
;
155 static DEFINE_MUTEX(pcpu_drain_mutex
);
156 static DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
158 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
159 volatile unsigned long latent_entropy __latent_entropy
;
160 EXPORT_SYMBOL(latent_entropy
);
164 * Array of node states.
166 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
167 [N_POSSIBLE
] = NODE_MASK_ALL
,
168 [N_ONLINE
] = { { [0] = 1UL } },
170 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
171 #ifdef CONFIG_HIGHMEM
172 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
174 [N_MEMORY
] = { { [0] = 1UL } },
175 [N_CPU
] = { { [0] = 1UL } },
178 EXPORT_SYMBOL(node_states
);
180 atomic_long_t _totalram_pages __read_mostly
;
181 EXPORT_SYMBOL(_totalram_pages
);
182 unsigned long totalreserve_pages __read_mostly
;
183 unsigned long totalcma_pages __read_mostly
;
185 int percpu_pagelist_high_fraction
;
186 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
187 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON
, init_on_alloc
);
188 EXPORT_SYMBOL(init_on_alloc
);
190 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON
, init_on_free
);
191 EXPORT_SYMBOL(init_on_free
);
193 static bool _init_on_alloc_enabled_early __read_mostly
194 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON
);
195 static int __init
early_init_on_alloc(char *buf
)
198 return kstrtobool(buf
, &_init_on_alloc_enabled_early
);
200 early_param("init_on_alloc", early_init_on_alloc
);
202 static bool _init_on_free_enabled_early __read_mostly
203 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON
);
204 static int __init
early_init_on_free(char *buf
)
206 return kstrtobool(buf
, &_init_on_free_enabled_early
);
208 early_param("init_on_free", early_init_on_free
);
211 * A cached value of the page's pageblock's migratetype, used when the page is
212 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
213 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
214 * Also the migratetype set in the page does not necessarily match the pcplist
215 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
216 * other index - this ensures that it will be put on the correct CMA freelist.
218 static inline int get_pcppage_migratetype(struct page
*page
)
223 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
225 page
->index
= migratetype
;
228 #ifdef CONFIG_PM_SLEEP
230 * The following functions are used by the suspend/hibernate code to temporarily
231 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
232 * while devices are suspended. To avoid races with the suspend/hibernate code,
233 * they should always be called with system_transition_mutex held
234 * (gfp_allowed_mask also should only be modified with system_transition_mutex
235 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
236 * with that modification).
239 static gfp_t saved_gfp_mask
;
241 void pm_restore_gfp_mask(void)
243 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
244 if (saved_gfp_mask
) {
245 gfp_allowed_mask
= saved_gfp_mask
;
250 void pm_restrict_gfp_mask(void)
252 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
253 WARN_ON(saved_gfp_mask
);
254 saved_gfp_mask
= gfp_allowed_mask
;
255 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
258 bool pm_suspended_storage(void)
260 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
264 #endif /* CONFIG_PM_SLEEP */
266 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
267 unsigned int pageblock_order __read_mostly
;
270 static void __free_pages_ok(struct page
*page
, unsigned int order
,
274 * results with 256, 32 in the lowmem_reserve sysctl:
275 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
276 * 1G machine -> (16M dma, 784M normal, 224M high)
277 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
278 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
279 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
281 * TBD: should special case ZONE_DMA32 machines here - in those we normally
282 * don't need any ZONE_NORMAL reservation
284 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
285 #ifdef CONFIG_ZONE_DMA
288 #ifdef CONFIG_ZONE_DMA32
292 #ifdef CONFIG_HIGHMEM
298 static char * const zone_names
[MAX_NR_ZONES
] = {
299 #ifdef CONFIG_ZONE_DMA
302 #ifdef CONFIG_ZONE_DMA32
306 #ifdef CONFIG_HIGHMEM
310 #ifdef CONFIG_ZONE_DEVICE
315 const char * const migratetype_names
[MIGRATE_TYPES
] = {
323 #ifdef CONFIG_MEMORY_ISOLATION
328 compound_page_dtor
* const compound_page_dtors
[NR_COMPOUND_DTORS
] = {
329 [NULL_COMPOUND_DTOR
] = NULL
,
330 [COMPOUND_PAGE_DTOR
] = free_compound_page
,
331 #ifdef CONFIG_HUGETLB_PAGE
332 [HUGETLB_PAGE_DTOR
] = free_huge_page
,
334 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
335 [TRANSHUGE_PAGE_DTOR
] = free_transhuge_page
,
339 int min_free_kbytes
= 1024;
340 int user_min_free_kbytes
= -1;
341 int watermark_boost_factor __read_mostly
= 15000;
342 int watermark_scale_factor
= 10;
344 static unsigned long nr_kernel_pages __initdata
;
345 static unsigned long nr_all_pages __initdata
;
346 static unsigned long dma_reserve __initdata
;
348 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
349 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
350 static unsigned long required_kernelcore __initdata
;
351 static unsigned long required_kernelcore_percent __initdata
;
352 static unsigned long required_movablecore __initdata
;
353 static unsigned long required_movablecore_percent __initdata
;
354 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
355 static bool mirrored_kernelcore __meminitdata
;
357 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
359 EXPORT_SYMBOL(movable_zone
);
362 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
363 unsigned int nr_online_nodes __read_mostly
= 1;
364 EXPORT_SYMBOL(nr_node_ids
);
365 EXPORT_SYMBOL(nr_online_nodes
);
368 int page_group_by_mobility_disabled __read_mostly
;
370 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
372 * During boot we initialize deferred pages on-demand, as needed, but once
373 * page_alloc_init_late() has finished, the deferred pages are all initialized,
374 * and we can permanently disable that path.
376 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
379 * Calling kasan_poison_pages() only after deferred memory initialization
380 * has completed. Poisoning pages during deferred memory init will greatly
381 * lengthen the process and cause problem in large memory systems as the
382 * deferred pages initialization is done with interrupt disabled.
384 * Assuming that there will be no reference to those newly initialized
385 * pages before they are ever allocated, this should have no effect on
386 * KASAN memory tracking as the poison will be properly inserted at page
387 * allocation time. The only corner case is when pages are allocated by
388 * on-demand allocation and then freed again before the deferred pages
389 * initialization is done, but this is not likely to happen.
391 static inline bool should_skip_kasan_poison(struct page
*page
, fpi_t fpi_flags
)
393 return static_branch_unlikely(&deferred_pages
) ||
394 (!IS_ENABLED(CONFIG_KASAN_GENERIC
) &&
395 (fpi_flags
& FPI_SKIP_KASAN_POISON
)) ||
396 PageSkipKASanPoison(page
);
399 /* Returns true if the struct page for the pfn is uninitialised */
400 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
402 int nid
= early_pfn_to_nid(pfn
);
404 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
411 * Returns true when the remaining initialisation should be deferred until
412 * later in the boot cycle when it can be parallelised.
414 static bool __meminit
415 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
417 static unsigned long prev_end_pfn
, nr_initialised
;
420 * prev_end_pfn static that contains the end of previous zone
421 * No need to protect because called very early in boot before smp_init.
423 if (prev_end_pfn
!= end_pfn
) {
424 prev_end_pfn
= end_pfn
;
428 /* Always populate low zones for address-constrained allocations */
429 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
432 if (NODE_DATA(nid
)->first_deferred_pfn
!= ULONG_MAX
)
435 * We start only with one section of pages, more pages are added as
436 * needed until the rest of deferred pages are initialized.
439 if ((nr_initialised
> PAGES_PER_SECTION
) &&
440 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
441 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
447 static inline bool should_skip_kasan_poison(struct page
*page
, fpi_t fpi_flags
)
449 return (!IS_ENABLED(CONFIG_KASAN_GENERIC
) &&
450 (fpi_flags
& FPI_SKIP_KASAN_POISON
)) ||
451 PageSkipKASanPoison(page
);
454 static inline bool early_page_uninitialised(unsigned long pfn
)
459 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
465 /* Return a pointer to the bitmap storing bits affecting a block of pages */
466 static inline unsigned long *get_pageblock_bitmap(const struct page
*page
,
469 #ifdef CONFIG_SPARSEMEM
470 return section_to_usemap(__pfn_to_section(pfn
));
472 return page_zone(page
)->pageblock_flags
;
473 #endif /* CONFIG_SPARSEMEM */
476 static inline int pfn_to_bitidx(const struct page
*page
, unsigned long pfn
)
478 #ifdef CONFIG_SPARSEMEM
479 pfn
&= (PAGES_PER_SECTION
-1);
481 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
482 #endif /* CONFIG_SPARSEMEM */
483 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
486 static __always_inline
487 unsigned long __get_pfnblock_flags_mask(const struct page
*page
,
491 unsigned long *bitmap
;
492 unsigned long bitidx
, word_bitidx
;
495 bitmap
= get_pageblock_bitmap(page
, pfn
);
496 bitidx
= pfn_to_bitidx(page
, pfn
);
497 word_bitidx
= bitidx
/ BITS_PER_LONG
;
498 bitidx
&= (BITS_PER_LONG
-1);
500 word
= bitmap
[word_bitidx
];
501 return (word
>> bitidx
) & mask
;
505 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
506 * @page: The page within the block of interest
507 * @pfn: The target page frame number
508 * @mask: mask of bits that the caller is interested in
510 * Return: pageblock_bits flags
512 unsigned long get_pfnblock_flags_mask(const struct page
*page
,
513 unsigned long pfn
, unsigned long mask
)
515 return __get_pfnblock_flags_mask(page
, pfn
, mask
);
518 static __always_inline
int get_pfnblock_migratetype(const struct page
*page
,
521 return __get_pfnblock_flags_mask(page
, pfn
, MIGRATETYPE_MASK
);
525 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
526 * @page: The page within the block of interest
527 * @flags: The flags to set
528 * @pfn: The target page frame number
529 * @mask: mask of bits that the caller is interested in
531 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
535 unsigned long *bitmap
;
536 unsigned long bitidx
, word_bitidx
;
537 unsigned long old_word
, word
;
539 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
540 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
542 bitmap
= get_pageblock_bitmap(page
, pfn
);
543 bitidx
= pfn_to_bitidx(page
, pfn
);
544 word_bitidx
= bitidx
/ BITS_PER_LONG
;
545 bitidx
&= (BITS_PER_LONG
-1);
547 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
552 word
= READ_ONCE(bitmap
[word_bitidx
]);
554 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
555 if (word
== old_word
)
561 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
563 if (unlikely(page_group_by_mobility_disabled
&&
564 migratetype
< MIGRATE_PCPTYPES
))
565 migratetype
= MIGRATE_UNMOVABLE
;
567 set_pfnblock_flags_mask(page
, (unsigned long)migratetype
,
568 page_to_pfn(page
), MIGRATETYPE_MASK
);
571 #ifdef CONFIG_DEBUG_VM
572 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
576 unsigned long pfn
= page_to_pfn(page
);
577 unsigned long sp
, start_pfn
;
580 seq
= zone_span_seqbegin(zone
);
581 start_pfn
= zone
->zone_start_pfn
;
582 sp
= zone
->spanned_pages
;
583 if (!zone_spans_pfn(zone
, pfn
))
585 } while (zone_span_seqretry(zone
, seq
));
588 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
589 pfn
, zone_to_nid(zone
), zone
->name
,
590 start_pfn
, start_pfn
+ sp
);
595 static int page_is_consistent(struct zone
*zone
, struct page
*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
);
654 /* Leave bad fields for debug, except PageBuddy could make trouble */
655 page_mapcount_reset(page
); /* remove PageBuddy */
656 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
659 static inline unsigned int order_to_pindex(int migratetype
, int order
)
663 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
664 if (order
> PAGE_ALLOC_COSTLY_ORDER
) {
665 VM_BUG_ON(order
!= pageblock_order
);
666 base
= PAGE_ALLOC_COSTLY_ORDER
+ 1;
669 VM_BUG_ON(order
> PAGE_ALLOC_COSTLY_ORDER
);
672 return (MIGRATE_PCPTYPES
* base
) + migratetype
;
675 static inline int pindex_to_order(unsigned int pindex
)
677 int order
= pindex
/ MIGRATE_PCPTYPES
;
679 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
680 if (order
> PAGE_ALLOC_COSTLY_ORDER
) {
681 order
= pageblock_order
;
682 VM_BUG_ON(order
!= pageblock_order
);
685 VM_BUG_ON(order
> PAGE_ALLOC_COSTLY_ORDER
);
691 static inline bool pcp_allowed_order(unsigned int order
)
693 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
695 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
696 if (order
== pageblock_order
)
702 static inline void free_the_page(struct page
*page
, unsigned int order
)
704 if (pcp_allowed_order(order
)) /* Via pcp? */
705 free_unref_page(page
, order
);
707 __free_pages_ok(page
, order
, FPI_NONE
);
711 * Higher-order pages are called "compound pages". They are structured thusly:
713 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
715 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
716 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
718 * The first tail page's ->compound_dtor holds the offset in array of compound
719 * page destructors. See compound_page_dtors.
721 * The first tail page's ->compound_order holds the order of allocation.
722 * This usage means that zero-order pages may not be compound.
725 void free_compound_page(struct page
*page
)
727 mem_cgroup_uncharge(page
);
728 free_the_page(page
, compound_order(page
));
731 void prep_compound_page(struct page
*page
, unsigned int order
)
734 int nr_pages
= 1 << order
;
737 for (i
= 1; i
< nr_pages
; i
++) {
738 struct page
*p
= page
+ i
;
739 p
->mapping
= TAIL_MAPPING
;
740 set_compound_head(p
, page
);
743 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
744 set_compound_order(page
, order
);
745 atomic_set(compound_mapcount_ptr(page
), -1);
746 if (hpage_pincount_available(page
))
747 atomic_set(compound_pincount_ptr(page
), 0);
750 #ifdef CONFIG_DEBUG_PAGEALLOC
751 unsigned int _debug_guardpage_minorder
;
753 bool _debug_pagealloc_enabled_early __read_mostly
754 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
755 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
756 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
757 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
759 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
761 static int __init
early_debug_pagealloc(char *buf
)
763 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
765 early_param("debug_pagealloc", early_debug_pagealloc
);
767 static int __init
debug_guardpage_minorder_setup(char *buf
)
771 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
772 pr_err("Bad debug_guardpage_minorder value\n");
775 _debug_guardpage_minorder
= res
;
776 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
779 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
781 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
782 unsigned int order
, int migratetype
)
784 if (!debug_guardpage_enabled())
787 if (order
>= debug_guardpage_minorder())
790 __SetPageGuard(page
);
791 INIT_LIST_HEAD(&page
->lru
);
792 set_page_private(page
, order
);
793 /* Guard pages are not available for any usage */
794 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
799 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
800 unsigned int order
, int migratetype
)
802 if (!debug_guardpage_enabled())
805 __ClearPageGuard(page
);
807 set_page_private(page
, 0);
808 if (!is_migrate_isolate(migratetype
))
809 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
812 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
813 unsigned int order
, int migratetype
) { return false; }
814 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
815 unsigned int order
, int migratetype
) {}
819 * Enable static keys related to various memory debugging and hardening options.
820 * Some override others, and depend on early params that are evaluated in the
821 * order of appearance. So we need to first gather the full picture of what was
822 * enabled, and then make decisions.
824 void init_mem_debugging_and_hardening(void)
826 bool page_poisoning_requested
= false;
828 #ifdef CONFIG_PAGE_POISONING
830 * Page poisoning is debug page alloc for some arches. If
831 * either of those options are enabled, enable poisoning.
833 if (page_poisoning_enabled() ||
834 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC
) &&
835 debug_pagealloc_enabled())) {
836 static_branch_enable(&_page_poisoning_enabled
);
837 page_poisoning_requested
= true;
841 if ((_init_on_alloc_enabled_early
|| _init_on_free_enabled_early
) &&
842 page_poisoning_requested
) {
843 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
844 "will take precedence over init_on_alloc and init_on_free\n");
845 _init_on_alloc_enabled_early
= false;
846 _init_on_free_enabled_early
= false;
849 if (_init_on_alloc_enabled_early
)
850 static_branch_enable(&init_on_alloc
);
852 static_branch_disable(&init_on_alloc
);
854 if (_init_on_free_enabled_early
)
855 static_branch_enable(&init_on_free
);
857 static_branch_disable(&init_on_free
);
859 #ifdef CONFIG_DEBUG_PAGEALLOC
860 if (!debug_pagealloc_enabled())
863 static_branch_enable(&_debug_pagealloc_enabled
);
865 if (!debug_guardpage_minorder())
868 static_branch_enable(&_debug_guardpage_enabled
);
872 static inline void set_buddy_order(struct page
*page
, unsigned int order
)
874 set_page_private(page
, order
);
875 __SetPageBuddy(page
);
879 * This function checks whether a page is free && is the buddy
880 * we can coalesce a page and its buddy if
881 * (a) the buddy is not in a hole (check before calling!) &&
882 * (b) the buddy is in the buddy system &&
883 * (c) a page and its buddy have the same order &&
884 * (d) a page and its buddy are in the same zone.
886 * For recording whether a page is in the buddy system, we set PageBuddy.
887 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
889 * For recording page's order, we use page_private(page).
891 static inline bool page_is_buddy(struct page
*page
, struct page
*buddy
,
894 if (!page_is_guard(buddy
) && !PageBuddy(buddy
))
897 if (buddy_order(buddy
) != order
)
901 * zone check is done late to avoid uselessly calculating
902 * zone/node ids for pages that could never merge.
904 if (page_zone_id(page
) != page_zone_id(buddy
))
907 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
912 #ifdef CONFIG_COMPACTION
913 static inline struct capture_control
*task_capc(struct zone
*zone
)
915 struct capture_control
*capc
= current
->capture_control
;
917 return unlikely(capc
) &&
918 !(current
->flags
& PF_KTHREAD
) &&
920 capc
->cc
->zone
== zone
? capc
: NULL
;
924 compaction_capture(struct capture_control
*capc
, struct page
*page
,
925 int order
, int migratetype
)
927 if (!capc
|| order
!= capc
->cc
->order
)
930 /* Do not accidentally pollute CMA or isolated regions*/
931 if (is_migrate_cma(migratetype
) ||
932 is_migrate_isolate(migratetype
))
936 * Do not let lower order allocations pollute a movable pageblock.
937 * This might let an unmovable request use a reclaimable pageblock
938 * and vice-versa but no more than normal fallback logic which can
939 * have trouble finding a high-order free page.
941 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
949 static inline struct capture_control
*task_capc(struct zone
*zone
)
955 compaction_capture(struct capture_control
*capc
, struct page
*page
,
956 int order
, int migratetype
)
960 #endif /* CONFIG_COMPACTION */
962 /* Used for pages not on another list */
963 static inline void add_to_free_list(struct page
*page
, struct zone
*zone
,
964 unsigned int order
, int migratetype
)
966 struct free_area
*area
= &zone
->free_area
[order
];
968 list_add(&page
->lru
, &area
->free_list
[migratetype
]);
972 /* Used for pages not on another list */
973 static inline void add_to_free_list_tail(struct page
*page
, struct zone
*zone
,
974 unsigned int order
, int migratetype
)
976 struct free_area
*area
= &zone
->free_area
[order
];
978 list_add_tail(&page
->lru
, &area
->free_list
[migratetype
]);
983 * Used for pages which are on another list. Move the pages to the tail
984 * of the list - so the moved pages won't immediately be considered for
985 * allocation again (e.g., optimization for memory onlining).
987 static inline void move_to_free_list(struct page
*page
, struct zone
*zone
,
988 unsigned int order
, int migratetype
)
990 struct free_area
*area
= &zone
->free_area
[order
];
992 list_move_tail(&page
->lru
, &area
->free_list
[migratetype
]);
995 static inline void del_page_from_free_list(struct page
*page
, struct zone
*zone
,
998 /* clear reported state and update reported page count */
999 if (page_reported(page
))
1000 __ClearPageReported(page
);
1002 list_del(&page
->lru
);
1003 __ClearPageBuddy(page
);
1004 set_page_private(page
, 0);
1005 zone
->free_area
[order
].nr_free
--;
1009 * If this is not the largest possible page, check if the buddy
1010 * of the next-highest order is free. If it is, it's possible
1011 * that pages are being freed that will coalesce soon. In case,
1012 * that is happening, add the free page to the tail of the list
1013 * so it's less likely to be used soon and more likely to be merged
1014 * as a higher order page
1017 buddy_merge_likely(unsigned long pfn
, unsigned long buddy_pfn
,
1018 struct page
*page
, unsigned int order
)
1020 struct page
*higher_page
, *higher_buddy
;
1021 unsigned long combined_pfn
;
1023 if (order
>= MAX_ORDER
- 2)
1026 combined_pfn
= buddy_pfn
& pfn
;
1027 higher_page
= page
+ (combined_pfn
- pfn
);
1028 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
1029 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
1031 return page_is_buddy(higher_page
, higher_buddy
, order
+ 1);
1035 * Freeing function for a buddy system allocator.
1037 * The concept of a buddy system is to maintain direct-mapped table
1038 * (containing bit values) for memory blocks of various "orders".
1039 * The bottom level table contains the map for the smallest allocatable
1040 * units of memory (here, pages), and each level above it describes
1041 * pairs of units from the levels below, hence, "buddies".
1042 * At a high level, all that happens here is marking the table entry
1043 * at the bottom level available, and propagating the changes upward
1044 * as necessary, plus some accounting needed to play nicely with other
1045 * parts of the VM system.
1046 * At each level, we keep a list of pages, which are heads of continuous
1047 * free pages of length of (1 << order) and marked with PageBuddy.
1048 * Page's order is recorded in page_private(page) field.
1049 * So when we are allocating or freeing one, we can derive the state of the
1050 * other. That is, if we allocate a small block, and both were
1051 * free, the remainder of the region must be split into blocks.
1052 * If a block is freed, and its buddy is also free, then this
1053 * triggers coalescing into a block of larger size.
1058 static inline void __free_one_page(struct page
*page
,
1060 struct zone
*zone
, unsigned int order
,
1061 int migratetype
, fpi_t fpi_flags
)
1063 struct capture_control
*capc
= task_capc(zone
);
1064 unsigned long buddy_pfn
;
1065 unsigned long combined_pfn
;
1066 unsigned int max_order
;
1070 max_order
= min_t(unsigned int, MAX_ORDER
- 1, pageblock_order
);
1072 VM_BUG_ON(!zone_is_initialized(zone
));
1073 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
1075 VM_BUG_ON(migratetype
== -1);
1076 if (likely(!is_migrate_isolate(migratetype
)))
1077 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
1079 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
1080 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1083 while (order
< max_order
) {
1084 if (compaction_capture(capc
, page
, order
, migratetype
)) {
1085 __mod_zone_freepage_state(zone
, -(1 << order
),
1089 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
1090 buddy
= page
+ (buddy_pfn
- pfn
);
1092 if (!page_is_buddy(page
, buddy
, order
))
1095 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1096 * merge with it and move up one order.
1098 if (page_is_guard(buddy
))
1099 clear_page_guard(zone
, buddy
, order
, migratetype
);
1101 del_page_from_free_list(buddy
, zone
, order
);
1102 combined_pfn
= buddy_pfn
& pfn
;
1103 page
= page
+ (combined_pfn
- pfn
);
1107 if (order
< MAX_ORDER
- 1) {
1108 /* If we are here, it means order is >= pageblock_order.
1109 * We want to prevent merge between freepages on isolate
1110 * pageblock and normal pageblock. Without this, pageblock
1111 * isolation could cause incorrect freepage or CMA accounting.
1113 * We don't want to hit this code for the more frequent
1114 * low-order merging.
1116 if (unlikely(has_isolate_pageblock(zone
))) {
1119 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
1120 buddy
= page
+ (buddy_pfn
- pfn
);
1121 buddy_mt
= get_pageblock_migratetype(buddy
);
1123 if (migratetype
!= buddy_mt
1124 && (is_migrate_isolate(migratetype
) ||
1125 is_migrate_isolate(buddy_mt
)))
1128 max_order
= order
+ 1;
1129 goto continue_merging
;
1133 set_buddy_order(page
, order
);
1135 if (fpi_flags
& FPI_TO_TAIL
)
1137 else if (is_shuffle_order(order
))
1138 to_tail
= shuffle_pick_tail();
1140 to_tail
= buddy_merge_likely(pfn
, buddy_pfn
, page
, order
);
1143 add_to_free_list_tail(page
, zone
, order
, migratetype
);
1145 add_to_free_list(page
, zone
, order
, migratetype
);
1147 /* Notify page reporting subsystem of freed page */
1148 if (!(fpi_flags
& FPI_SKIP_REPORT_NOTIFY
))
1149 page_reporting_notify_free(order
);
1153 * A bad page could be due to a number of fields. Instead of multiple branches,
1154 * try and check multiple fields with one check. The caller must do a detailed
1155 * check if necessary.
1157 static inline bool page_expected_state(struct page
*page
,
1158 unsigned long check_flags
)
1160 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1163 if (unlikely((unsigned long)page
->mapping
|
1164 page_ref_count(page
) |
1168 (page
->flags
& check_flags
)))
1174 static const char *page_bad_reason(struct page
*page
, unsigned long flags
)
1176 const char *bad_reason
= NULL
;
1178 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1179 bad_reason
= "nonzero mapcount";
1180 if (unlikely(page
->mapping
!= NULL
))
1181 bad_reason
= "non-NULL mapping";
1182 if (unlikely(page_ref_count(page
) != 0))
1183 bad_reason
= "nonzero _refcount";
1184 if (unlikely(page
->flags
& flags
)) {
1185 if (flags
== PAGE_FLAGS_CHECK_AT_PREP
)
1186 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1188 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1191 if (unlikely(page
->memcg_data
))
1192 bad_reason
= "page still charged to cgroup";
1197 static void check_free_page_bad(struct page
*page
)
1200 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_FREE
));
1203 static inline int check_free_page(struct page
*page
)
1205 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1208 /* Something has gone sideways, find it */
1209 check_free_page_bad(page
);
1213 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1218 * We rely page->lru.next never has bit 0 set, unless the page
1219 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1221 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1223 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1227 switch (page
- head_page
) {
1229 /* the first tail page: ->mapping may be compound_mapcount() */
1230 if (unlikely(compound_mapcount(page
))) {
1231 bad_page(page
, "nonzero compound_mapcount");
1237 * the second tail page: ->mapping is
1238 * deferred_list.next -- ignore value.
1242 if (page
->mapping
!= TAIL_MAPPING
) {
1243 bad_page(page
, "corrupted mapping in tail page");
1248 if (unlikely(!PageTail(page
))) {
1249 bad_page(page
, "PageTail not set");
1252 if (unlikely(compound_head(page
) != head_page
)) {
1253 bad_page(page
, "compound_head not consistent");
1258 page
->mapping
= NULL
;
1259 clear_compound_head(page
);
1263 static void kernel_init_free_pages(struct page
*page
, int numpages
, bool zero_tags
)
1268 for (i
= 0; i
< numpages
; i
++)
1269 tag_clear_highpage(page
+ i
);
1273 /* s390's use of memset() could override KASAN redzones. */
1274 kasan_disable_current();
1275 for (i
= 0; i
< numpages
; i
++) {
1276 u8 tag
= page_kasan_tag(page
+ i
);
1277 page_kasan_tag_reset(page
+ i
);
1278 clear_highpage(page
+ i
);
1279 page_kasan_tag_set(page
+ i
, tag
);
1281 kasan_enable_current();
1284 static __always_inline
bool free_pages_prepare(struct page
*page
,
1285 unsigned int order
, bool check_free
, fpi_t fpi_flags
)
1288 bool skip_kasan_poison
= should_skip_kasan_poison(page
, fpi_flags
);
1290 VM_BUG_ON_PAGE(PageTail(page
), page
);
1292 trace_mm_page_free(page
, order
);
1294 if (unlikely(PageHWPoison(page
)) && !order
) {
1296 * Do not let hwpoison pages hit pcplists/buddy
1297 * Untie memcg state and reset page's owner
1299 if (memcg_kmem_enabled() && PageMemcgKmem(page
))
1300 __memcg_kmem_uncharge_page(page
, order
);
1301 reset_page_owner(page
, order
);
1306 * Check tail pages before head page information is cleared to
1307 * avoid checking PageCompound for order-0 pages.
1309 if (unlikely(order
)) {
1310 bool compound
= PageCompound(page
);
1313 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1316 ClearPageDoubleMap(page
);
1317 for (i
= 1; i
< (1 << order
); i
++) {
1319 bad
+= free_tail_pages_check(page
, page
+ i
);
1320 if (unlikely(check_free_page(page
+ i
))) {
1324 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1327 if (PageMappingFlags(page
))
1328 page
->mapping
= NULL
;
1329 if (memcg_kmem_enabled() && PageMemcgKmem(page
))
1330 __memcg_kmem_uncharge_page(page
, order
);
1332 bad
+= check_free_page(page
);
1336 page_cpupid_reset_last(page
);
1337 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1338 reset_page_owner(page
, order
);
1340 if (!PageHighMem(page
)) {
1341 debug_check_no_locks_freed(page_address(page
),
1342 PAGE_SIZE
<< order
);
1343 debug_check_no_obj_freed(page_address(page
),
1344 PAGE_SIZE
<< order
);
1347 kernel_poison_pages(page
, 1 << order
);
1350 * As memory initialization might be integrated into KASAN,
1351 * kasan_free_pages and kernel_init_free_pages must be
1352 * kept together to avoid discrepancies in behavior.
1354 * With hardware tag-based KASAN, memory tags must be set before the
1355 * page becomes unavailable via debug_pagealloc or arch_free_page.
1357 if (kasan_has_integrated_init()) {
1358 if (!skip_kasan_poison
)
1359 kasan_free_pages(page
, order
);
1361 bool init
= want_init_on_free();
1364 kernel_init_free_pages(page
, 1 << order
, false);
1365 if (!skip_kasan_poison
)
1366 kasan_poison_pages(page
, order
, init
);
1370 * arch_free_page() can make the page's contents inaccessible. s390
1371 * does this. So nothing which can access the page's contents should
1372 * happen after this.
1374 arch_free_page(page
, order
);
1376 debug_pagealloc_unmap_pages(page
, 1 << order
);
1381 #ifdef CONFIG_DEBUG_VM
1383 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1384 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1385 * moved from pcp lists to free lists.
1387 static bool free_pcp_prepare(struct page
*page
, unsigned int order
)
1389 return free_pages_prepare(page
, order
, true, FPI_NONE
);
1392 static bool bulkfree_pcp_prepare(struct page
*page
)
1394 if (debug_pagealloc_enabled_static())
1395 return check_free_page(page
);
1401 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1402 * moving from pcp lists to free list in order to reduce overhead. With
1403 * debug_pagealloc enabled, they are checked also immediately when being freed
1406 static bool free_pcp_prepare(struct page
*page
, unsigned int order
)
1408 if (debug_pagealloc_enabled_static())
1409 return free_pages_prepare(page
, order
, true, FPI_NONE
);
1411 return free_pages_prepare(page
, order
, false, FPI_NONE
);
1414 static bool bulkfree_pcp_prepare(struct page
*page
)
1416 return check_free_page(page
);
1418 #endif /* CONFIG_DEBUG_VM */
1420 static inline void prefetch_buddy(struct page
*page
)
1422 unsigned long pfn
= page_to_pfn(page
);
1423 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1424 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1430 * Frees a number of pages from the PCP lists
1431 * Assumes all pages on list are in same zone, and of same order.
1432 * count is the number of pages to free.
1434 * If the zone was previously in an "all pages pinned" state then look to
1435 * see if this freeing clears that state.
1437 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1438 * pinned" detection logic.
1440 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1441 struct per_cpu_pages
*pcp
)
1447 int prefetch_nr
= READ_ONCE(pcp
->batch
);
1448 bool isolated_pageblocks
;
1449 struct page
*page
, *tmp
;
1453 * Ensure proper count is passed which otherwise would stuck in the
1454 * below while (list_empty(list)) loop.
1456 count
= min(pcp
->count
, count
);
1458 struct list_head
*list
;
1461 * Remove pages from lists in a round-robin fashion. A
1462 * batch_free count is maintained that is incremented when an
1463 * empty list is encountered. This is so more pages are freed
1464 * off fuller lists instead of spinning excessively around empty
1469 if (++pindex
== NR_PCP_LISTS
)
1471 list
= &pcp
->lists
[pindex
];
1472 } while (list_empty(list
));
1474 /* This is the only non-empty list. Free them all. */
1475 if (batch_free
== NR_PCP_LISTS
)
1478 order
= pindex_to_order(pindex
);
1479 BUILD_BUG_ON(MAX_ORDER
>= (1<<NR_PCP_ORDER_WIDTH
));
1481 page
= list_last_entry(list
, struct page
, lru
);
1482 /* must delete to avoid corrupting pcp list */
1483 list_del(&page
->lru
);
1484 nr_freed
+= 1 << order
;
1485 count
-= 1 << order
;
1487 if (bulkfree_pcp_prepare(page
))
1490 /* Encode order with the migratetype */
1491 page
->index
<<= NR_PCP_ORDER_WIDTH
;
1492 page
->index
|= order
;
1494 list_add_tail(&page
->lru
, &head
);
1497 * We are going to put the page back to the global
1498 * pool, prefetch its buddy to speed up later access
1499 * under zone->lock. It is believed the overhead of
1500 * an additional test and calculating buddy_pfn here
1501 * can be offset by reduced memory latency later. To
1502 * avoid excessive prefetching due to large count, only
1503 * prefetch buddy for the first pcp->batch nr of pages.
1506 prefetch_buddy(page
);
1509 } while (count
> 0 && --batch_free
&& !list_empty(list
));
1511 pcp
->count
-= nr_freed
;
1514 * local_lock_irq held so equivalent to spin_lock_irqsave for
1515 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1517 spin_lock(&zone
->lock
);
1518 isolated_pageblocks
= has_isolate_pageblock(zone
);
1521 * Use safe version since after __free_one_page(),
1522 * page->lru.next will not point to original list.
1524 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1525 int mt
= get_pcppage_migratetype(page
);
1527 /* mt has been encoded with the order (see above) */
1528 order
= mt
& NR_PCP_ORDER_MASK
;
1529 mt
>>= NR_PCP_ORDER_WIDTH
;
1531 /* MIGRATE_ISOLATE page should not go to pcplists */
1532 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1533 /* Pageblock could have been isolated meanwhile */
1534 if (unlikely(isolated_pageblocks
))
1535 mt
= get_pageblock_migratetype(page
);
1537 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
, FPI_NONE
);
1538 trace_mm_page_pcpu_drain(page
, order
, mt
);
1540 spin_unlock(&zone
->lock
);
1543 static void free_one_page(struct zone
*zone
,
1544 struct page
*page
, unsigned long pfn
,
1546 int migratetype
, fpi_t fpi_flags
)
1548 unsigned long flags
;
1550 spin_lock_irqsave(&zone
->lock
, flags
);
1551 if (unlikely(has_isolate_pageblock(zone
) ||
1552 is_migrate_isolate(migratetype
))) {
1553 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1555 __free_one_page(page
, pfn
, zone
, order
, migratetype
, fpi_flags
);
1556 spin_unlock_irqrestore(&zone
->lock
, flags
);
1559 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1560 unsigned long zone
, int nid
)
1562 mm_zero_struct_page(page
);
1563 set_page_links(page
, zone
, nid
, pfn
);
1564 init_page_count(page
);
1565 page_mapcount_reset(page
);
1566 page_cpupid_reset_last(page
);
1567 page_kasan_tag_reset(page
);
1569 INIT_LIST_HEAD(&page
->lru
);
1570 #ifdef WANT_PAGE_VIRTUAL
1571 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1572 if (!is_highmem_idx(zone
))
1573 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1577 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1578 static void __meminit
init_reserved_page(unsigned long pfn
)
1583 if (!early_page_uninitialised(pfn
))
1586 nid
= early_pfn_to_nid(pfn
);
1587 pgdat
= NODE_DATA(nid
);
1589 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1590 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1592 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1595 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1598 static inline void init_reserved_page(unsigned long pfn
)
1601 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1604 * Initialised pages do not have PageReserved set. This function is
1605 * called for each range allocated by the bootmem allocator and
1606 * marks the pages PageReserved. The remaining valid pages are later
1607 * sent to the buddy page allocator.
1609 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1611 unsigned long start_pfn
= PFN_DOWN(start
);
1612 unsigned long end_pfn
= PFN_UP(end
);
1614 for (; start_pfn
< end_pfn
; start_pfn
++) {
1615 if (pfn_valid(start_pfn
)) {
1616 struct page
*page
= pfn_to_page(start_pfn
);
1618 init_reserved_page(start_pfn
);
1620 /* Avoid false-positive PageTail() */
1621 INIT_LIST_HEAD(&page
->lru
);
1624 * no need for atomic set_bit because the struct
1625 * page is not visible yet so nobody should
1628 __SetPageReserved(page
);
1633 static void __free_pages_ok(struct page
*page
, unsigned int order
,
1636 unsigned long flags
;
1638 unsigned long pfn
= page_to_pfn(page
);
1639 struct zone
*zone
= page_zone(page
);
1641 if (!free_pages_prepare(page
, order
, true, fpi_flags
))
1644 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1646 spin_lock_irqsave(&zone
->lock
, flags
);
1647 if (unlikely(has_isolate_pageblock(zone
) ||
1648 is_migrate_isolate(migratetype
))) {
1649 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1651 __free_one_page(page
, pfn
, zone
, order
, migratetype
, fpi_flags
);
1652 spin_unlock_irqrestore(&zone
->lock
, flags
);
1654 __count_vm_events(PGFREE
, 1 << order
);
1657 void __free_pages_core(struct page
*page
, unsigned int order
)
1659 unsigned int nr_pages
= 1 << order
;
1660 struct page
*p
= page
;
1664 * When initializing the memmap, __init_single_page() sets the refcount
1665 * of all pages to 1 ("allocated"/"not free"). We have to set the
1666 * refcount of all involved pages to 0.
1669 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1671 __ClearPageReserved(p
);
1672 set_page_count(p
, 0);
1674 __ClearPageReserved(p
);
1675 set_page_count(p
, 0);
1677 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1680 * Bypass PCP and place fresh pages right to the tail, primarily
1681 * relevant for memory onlining.
1683 __free_pages_ok(page
, order
, FPI_TO_TAIL
| FPI_SKIP_KASAN_POISON
);
1689 * During memory init memblocks map pfns to nids. The search is expensive and
1690 * this caches recent lookups. The implementation of __early_pfn_to_nid
1691 * treats start/end as pfns.
1693 struct mminit_pfnnid_cache
{
1694 unsigned long last_start
;
1695 unsigned long last_end
;
1699 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1702 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1704 static int __meminit
__early_pfn_to_nid(unsigned long pfn
,
1705 struct mminit_pfnnid_cache
*state
)
1707 unsigned long start_pfn
, end_pfn
;
1710 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
1711 return state
->last_nid
;
1713 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
1714 if (nid
!= NUMA_NO_NODE
) {
1715 state
->last_start
= start_pfn
;
1716 state
->last_end
= end_pfn
;
1717 state
->last_nid
= nid
;
1723 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1725 static DEFINE_SPINLOCK(early_pfn_lock
);
1728 spin_lock(&early_pfn_lock
);
1729 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1731 nid
= first_online_node
;
1732 spin_unlock(&early_pfn_lock
);
1736 #endif /* CONFIG_NUMA */
1738 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1741 if (early_page_uninitialised(pfn
))
1743 __free_pages_core(page
, order
);
1747 * Check that the whole (or subset of) a pageblock given by the interval of
1748 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1749 * with the migration of free compaction scanner.
1751 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1753 * It's possible on some configurations to have a setup like node0 node1 node0
1754 * i.e. it's possible that all pages within a zones range of pages do not
1755 * belong to a single zone. We assume that a border between node0 and node1
1756 * can occur within a single pageblock, but not a node0 node1 node0
1757 * interleaving within a single pageblock. It is therefore sufficient to check
1758 * the first and last page of a pageblock and avoid checking each individual
1759 * page in a pageblock.
1761 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1762 unsigned long end_pfn
, struct zone
*zone
)
1764 struct page
*start_page
;
1765 struct page
*end_page
;
1767 /* end_pfn is one past the range we are checking */
1770 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1773 start_page
= pfn_to_online_page(start_pfn
);
1777 if (page_zone(start_page
) != zone
)
1780 end_page
= pfn_to_page(end_pfn
);
1782 /* This gives a shorter code than deriving page_zone(end_page) */
1783 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1789 void set_zone_contiguous(struct zone
*zone
)
1791 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1792 unsigned long block_end_pfn
;
1794 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1795 for (; block_start_pfn
< zone_end_pfn(zone
);
1796 block_start_pfn
= block_end_pfn
,
1797 block_end_pfn
+= pageblock_nr_pages
) {
1799 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1801 if (!__pageblock_pfn_to_page(block_start_pfn
,
1802 block_end_pfn
, zone
))
1807 /* We confirm that there is no hole */
1808 zone
->contiguous
= true;
1811 void clear_zone_contiguous(struct zone
*zone
)
1813 zone
->contiguous
= false;
1816 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1817 static void __init
deferred_free_range(unsigned long pfn
,
1818 unsigned long nr_pages
)
1826 page
= pfn_to_page(pfn
);
1828 /* Free a large naturally-aligned chunk if possible */
1829 if (nr_pages
== pageblock_nr_pages
&&
1830 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1831 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1832 __free_pages_core(page
, pageblock_order
);
1836 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1837 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1838 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1839 __free_pages_core(page
, 0);
1843 /* Completion tracking for deferred_init_memmap() threads */
1844 static atomic_t pgdat_init_n_undone __initdata
;
1845 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1847 static inline void __init
pgdat_init_report_one_done(void)
1849 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1850 complete(&pgdat_init_all_done_comp
);
1854 * Returns true if page needs to be initialized or freed to buddy allocator.
1856 * First we check if pfn is valid on architectures where it is possible to have
1857 * holes within pageblock_nr_pages. On systems where it is not possible, this
1858 * function is optimized out.
1860 * Then, we check if a current large page is valid by only checking the validity
1863 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1865 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1871 * Free pages to buddy allocator. Try to free aligned pages in
1872 * pageblock_nr_pages sizes.
1874 static void __init
deferred_free_pages(unsigned long pfn
,
1875 unsigned long end_pfn
)
1877 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1878 unsigned long nr_free
= 0;
1880 for (; pfn
< end_pfn
; pfn
++) {
1881 if (!deferred_pfn_valid(pfn
)) {
1882 deferred_free_range(pfn
- nr_free
, nr_free
);
1884 } else if (!(pfn
& nr_pgmask
)) {
1885 deferred_free_range(pfn
- nr_free
, nr_free
);
1891 /* Free the last block of pages to allocator */
1892 deferred_free_range(pfn
- nr_free
, nr_free
);
1896 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1897 * by performing it only once every pageblock_nr_pages.
1898 * Return number of pages initialized.
1900 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1902 unsigned long end_pfn
)
1904 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1905 int nid
= zone_to_nid(zone
);
1906 unsigned long nr_pages
= 0;
1907 int zid
= zone_idx(zone
);
1908 struct page
*page
= NULL
;
1910 for (; pfn
< end_pfn
; pfn
++) {
1911 if (!deferred_pfn_valid(pfn
)) {
1914 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1915 page
= pfn_to_page(pfn
);
1919 __init_single_page(page
, pfn
, zid
, nid
);
1926 * This function is meant to pre-load the iterator for the zone init.
1927 * Specifically it walks through the ranges until we are caught up to the
1928 * first_init_pfn value and exits there. If we never encounter the value we
1929 * return false indicating there are no valid ranges left.
1932 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1933 unsigned long *spfn
, unsigned long *epfn
,
1934 unsigned long first_init_pfn
)
1939 * Start out by walking through the ranges in this zone that have
1940 * already been initialized. We don't need to do anything with them
1941 * so we just need to flush them out of the system.
1943 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1944 if (*epfn
<= first_init_pfn
)
1946 if (*spfn
< first_init_pfn
)
1947 *spfn
= first_init_pfn
;
1956 * Initialize and free pages. We do it in two loops: first we initialize
1957 * struct page, then free to buddy allocator, because while we are
1958 * freeing pages we can access pages that are ahead (computing buddy
1959 * page in __free_one_page()).
1961 * In order to try and keep some memory in the cache we have the loop
1962 * broken along max page order boundaries. This way we will not cause
1963 * any issues with the buddy page computation.
1965 static unsigned long __init
1966 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1967 unsigned long *end_pfn
)
1969 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1970 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1971 unsigned long nr_pages
= 0;
1974 /* First we loop through and initialize the page values */
1975 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1978 if (mo_pfn
<= *start_pfn
)
1981 t
= min(mo_pfn
, *end_pfn
);
1982 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1984 if (mo_pfn
< *end_pfn
) {
1985 *start_pfn
= mo_pfn
;
1990 /* Reset values and now loop through freeing pages as needed */
1993 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1999 t
= min(mo_pfn
, epfn
);
2000 deferred_free_pages(spfn
, t
);
2010 deferred_init_memmap_chunk(unsigned long start_pfn
, unsigned long end_pfn
,
2013 unsigned long spfn
, epfn
;
2014 struct zone
*zone
= arg
;
2017 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
, start_pfn
);
2020 * Initialize and free pages in MAX_ORDER sized increments so that we
2021 * can avoid introducing any issues with the buddy allocator.
2023 while (spfn
< end_pfn
) {
2024 deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
2029 /* An arch may override for more concurrency. */
2031 deferred_page_init_max_threads(const struct cpumask
*node_cpumask
)
2036 /* Initialise remaining memory on a node */
2037 static int __init
deferred_init_memmap(void *data
)
2039 pg_data_t
*pgdat
= data
;
2040 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2041 unsigned long spfn
= 0, epfn
= 0;
2042 unsigned long first_init_pfn
, flags
;
2043 unsigned long start
= jiffies
;
2045 int zid
, max_threads
;
2048 /* Bind memory initialisation thread to a local node if possible */
2049 if (!cpumask_empty(cpumask
))
2050 set_cpus_allowed_ptr(current
, cpumask
);
2052 pgdat_resize_lock(pgdat
, &flags
);
2053 first_init_pfn
= pgdat
->first_deferred_pfn
;
2054 if (first_init_pfn
== ULONG_MAX
) {
2055 pgdat_resize_unlock(pgdat
, &flags
);
2056 pgdat_init_report_one_done();
2060 /* Sanity check boundaries */
2061 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
2062 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
2063 pgdat
->first_deferred_pfn
= ULONG_MAX
;
2066 * Once we unlock here, the zone cannot be grown anymore, thus if an
2067 * interrupt thread must allocate this early in boot, zone must be
2068 * pre-grown prior to start of deferred page initialization.
2070 pgdat_resize_unlock(pgdat
, &flags
);
2072 /* Only the highest zone is deferred so find it */
2073 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
2074 zone
= pgdat
->node_zones
+ zid
;
2075 if (first_init_pfn
< zone_end_pfn(zone
))
2079 /* If the zone is empty somebody else may have cleared out the zone */
2080 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
2084 max_threads
= deferred_page_init_max_threads(cpumask
);
2086 while (spfn
< epfn
) {
2087 unsigned long epfn_align
= ALIGN(epfn
, PAGES_PER_SECTION
);
2088 struct padata_mt_job job
= {
2089 .thread_fn
= deferred_init_memmap_chunk
,
2092 .size
= epfn_align
- spfn
,
2093 .align
= PAGES_PER_SECTION
,
2094 .min_chunk
= PAGES_PER_SECTION
,
2095 .max_threads
= max_threads
,
2098 padata_do_multithreaded(&job
);
2099 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
2103 /* Sanity check that the next zone really is unpopulated */
2104 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
2106 pr_info("node %d deferred pages initialised in %ums\n",
2107 pgdat
->node_id
, jiffies_to_msecs(jiffies
- start
));
2109 pgdat_init_report_one_done();
2114 * If this zone has deferred pages, try to grow it by initializing enough
2115 * deferred pages to satisfy the allocation specified by order, rounded up to
2116 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2117 * of SECTION_SIZE bytes by initializing struct pages in increments of
2118 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2120 * Return true when zone was grown, otherwise return false. We return true even
2121 * when we grow less than requested, to let the caller decide if there are
2122 * enough pages to satisfy the allocation.
2124 * Note: We use noinline because this function is needed only during boot, and
2125 * it is called from a __ref function _deferred_grow_zone. This way we are
2126 * making sure that it is not inlined into permanent text section.
2128 static noinline
bool __init
2129 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
2131 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
2132 pg_data_t
*pgdat
= zone
->zone_pgdat
;
2133 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
2134 unsigned long spfn
, epfn
, flags
;
2135 unsigned long nr_pages
= 0;
2138 /* Only the last zone may have deferred pages */
2139 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
2142 pgdat_resize_lock(pgdat
, &flags
);
2145 * If someone grew this zone while we were waiting for spinlock, return
2146 * true, as there might be enough pages already.
2148 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
2149 pgdat_resize_unlock(pgdat
, &flags
);
2153 /* If the zone is empty somebody else may have cleared out the zone */
2154 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
2155 first_deferred_pfn
)) {
2156 pgdat
->first_deferred_pfn
= ULONG_MAX
;
2157 pgdat_resize_unlock(pgdat
, &flags
);
2158 /* Retry only once. */
2159 return first_deferred_pfn
!= ULONG_MAX
;
2163 * Initialize and free pages in MAX_ORDER sized increments so
2164 * that we can avoid introducing any issues with the buddy
2167 while (spfn
< epfn
) {
2168 /* update our first deferred PFN for this section */
2169 first_deferred_pfn
= spfn
;
2171 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
2172 touch_nmi_watchdog();
2174 /* We should only stop along section boundaries */
2175 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
2178 /* If our quota has been met we can stop here */
2179 if (nr_pages
>= nr_pages_needed
)
2183 pgdat
->first_deferred_pfn
= spfn
;
2184 pgdat_resize_unlock(pgdat
, &flags
);
2186 return nr_pages
> 0;
2190 * deferred_grow_zone() is __init, but it is called from
2191 * get_page_from_freelist() during early boot until deferred_pages permanently
2192 * disables this call. This is why we have refdata wrapper to avoid warning,
2193 * and to ensure that the function body gets unloaded.
2196 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
2198 return deferred_grow_zone(zone
, order
);
2201 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2203 void __init
page_alloc_init_late(void)
2208 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2210 /* There will be num_node_state(N_MEMORY) threads */
2211 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
2212 for_each_node_state(nid
, N_MEMORY
) {
2213 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
2216 /* Block until all are initialised */
2217 wait_for_completion(&pgdat_init_all_done_comp
);
2220 * We initialized the rest of the deferred pages. Permanently disable
2221 * on-demand struct page initialization.
2223 static_branch_disable(&deferred_pages
);
2225 /* Reinit limits that are based on free pages after the kernel is up */
2226 files_maxfiles_init();
2231 /* Discard memblock private memory */
2234 for_each_node_state(nid
, N_MEMORY
)
2235 shuffle_free_memory(NODE_DATA(nid
));
2237 for_each_populated_zone(zone
)
2238 set_zone_contiguous(zone
);
2242 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2243 void __init
init_cma_reserved_pageblock(struct page
*page
)
2245 unsigned i
= pageblock_nr_pages
;
2246 struct page
*p
= page
;
2249 __ClearPageReserved(p
);
2250 set_page_count(p
, 0);
2253 set_pageblock_migratetype(page
, MIGRATE_CMA
);
2255 if (pageblock_order
>= MAX_ORDER
) {
2256 i
= pageblock_nr_pages
;
2259 set_page_refcounted(p
);
2260 __free_pages(p
, MAX_ORDER
- 1);
2261 p
+= MAX_ORDER_NR_PAGES
;
2262 } while (i
-= MAX_ORDER_NR_PAGES
);
2264 set_page_refcounted(page
);
2265 __free_pages(page
, pageblock_order
);
2268 adjust_managed_page_count(page
, pageblock_nr_pages
);
2269 page_zone(page
)->cma_pages
+= pageblock_nr_pages
;
2274 * The order of subdivision here is critical for the IO subsystem.
2275 * Please do not alter this order without good reasons and regression
2276 * testing. Specifically, as large blocks of memory are subdivided,
2277 * the order in which smaller blocks are delivered depends on the order
2278 * they're subdivided in this function. This is the primary factor
2279 * influencing the order in which pages are delivered to the IO
2280 * subsystem according to empirical testing, and this is also justified
2281 * by considering the behavior of a buddy system containing a single
2282 * large block of memory acted on by a series of small allocations.
2283 * This behavior is a critical factor in sglist merging's success.
2287 static inline void expand(struct zone
*zone
, struct page
*page
,
2288 int low
, int high
, int migratetype
)
2290 unsigned long size
= 1 << high
;
2292 while (high
> low
) {
2295 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2298 * Mark as guard pages (or page), that will allow to
2299 * merge back to allocator when buddy will be freed.
2300 * Corresponding page table entries will not be touched,
2301 * pages will stay not present in virtual address space
2303 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2306 add_to_free_list(&page
[size
], zone
, high
, migratetype
);
2307 set_buddy_order(&page
[size
], high
);
2311 static void check_new_page_bad(struct page
*page
)
2313 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2314 /* Don't complain about hwpoisoned pages */
2315 page_mapcount_reset(page
); /* remove PageBuddy */
2320 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_PREP
));
2324 * This page is about to be returned from the page allocator
2326 static inline int check_new_page(struct page
*page
)
2328 if (likely(page_expected_state(page
,
2329 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2332 check_new_page_bad(page
);
2336 #ifdef CONFIG_DEBUG_VM
2338 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2339 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2340 * also checked when pcp lists are refilled from the free lists.
2342 static inline bool check_pcp_refill(struct page
*page
)
2344 if (debug_pagealloc_enabled_static())
2345 return check_new_page(page
);
2350 static inline bool check_new_pcp(struct page
*page
)
2352 return check_new_page(page
);
2356 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2357 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2358 * enabled, they are also checked when being allocated from the pcp lists.
2360 static inline bool check_pcp_refill(struct page
*page
)
2362 return check_new_page(page
);
2364 static inline bool check_new_pcp(struct page
*page
)
2366 if (debug_pagealloc_enabled_static())
2367 return check_new_page(page
);
2371 #endif /* CONFIG_DEBUG_VM */
2373 static bool check_new_pages(struct page
*page
, unsigned int order
)
2376 for (i
= 0; i
< (1 << order
); i
++) {
2377 struct page
*p
= page
+ i
;
2379 if (unlikely(check_new_page(p
)))
2386 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2389 set_page_private(page
, 0);
2390 set_page_refcounted(page
);
2392 arch_alloc_page(page
, order
);
2393 debug_pagealloc_map_pages(page
, 1 << order
);
2396 * Page unpoisoning must happen before memory initialization.
2397 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2398 * allocations and the page unpoisoning code will complain.
2400 kernel_unpoison_pages(page
, 1 << order
);
2403 * As memory initialization might be integrated into KASAN,
2404 * kasan_alloc_pages and kernel_init_free_pages must be
2405 * kept together to avoid discrepancies in behavior.
2407 if (kasan_has_integrated_init()) {
2408 kasan_alloc_pages(page
, order
, gfp_flags
);
2410 bool init
= !want_init_on_free() && want_init_on_alloc(gfp_flags
);
2412 kasan_unpoison_pages(page
, order
, init
);
2414 kernel_init_free_pages(page
, 1 << order
,
2415 gfp_flags
& __GFP_ZEROTAGS
);
2418 set_page_owner(page
, order
, gfp_flags
);
2421 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2422 unsigned int alloc_flags
)
2424 post_alloc_hook(page
, order
, gfp_flags
);
2426 if (order
&& (gfp_flags
& __GFP_COMP
))
2427 prep_compound_page(page
, order
);
2430 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2431 * allocate the page. The expectation is that the caller is taking
2432 * steps that will free more memory. The caller should avoid the page
2433 * being used for !PFMEMALLOC purposes.
2435 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2436 set_page_pfmemalloc(page
);
2438 clear_page_pfmemalloc(page
);
2442 * Go through the free lists for the given migratetype and remove
2443 * the smallest available page from the freelists
2445 static __always_inline
2446 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2449 unsigned int current_order
;
2450 struct free_area
*area
;
2453 /* Find a page of the appropriate size in the preferred list */
2454 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2455 area
= &(zone
->free_area
[current_order
]);
2456 page
= get_page_from_free_area(area
, migratetype
);
2459 del_page_from_free_list(page
, zone
, current_order
);
2460 expand(zone
, page
, order
, current_order
, migratetype
);
2461 set_pcppage_migratetype(page
, migratetype
);
2470 * This array describes the order lists are fallen back to when
2471 * the free lists for the desirable migrate type are depleted
2473 static int fallbacks
[MIGRATE_TYPES
][3] = {
2474 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2475 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2476 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2478 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2480 #ifdef CONFIG_MEMORY_ISOLATION
2481 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2486 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2489 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2492 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2493 unsigned int order
) { return NULL
; }
2497 * Move the free pages in a range to the freelist tail of the requested type.
2498 * Note that start_page and end_pages are not aligned on a pageblock
2499 * boundary. If alignment is required, use move_freepages_block()
2501 static int move_freepages(struct zone
*zone
,
2502 unsigned long start_pfn
, unsigned long end_pfn
,
2503 int migratetype
, int *num_movable
)
2508 int pages_moved
= 0;
2510 for (pfn
= start_pfn
; pfn
<= end_pfn
;) {
2511 page
= pfn_to_page(pfn
);
2512 if (!PageBuddy(page
)) {
2514 * We assume that pages that could be isolated for
2515 * migration are movable. But we don't actually try
2516 * isolating, as that would be expensive.
2519 (PageLRU(page
) || __PageMovable(page
)))
2525 /* Make sure we are not inadvertently changing nodes */
2526 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2527 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2529 order
= buddy_order(page
);
2530 move_to_free_list(page
, zone
, order
, migratetype
);
2532 pages_moved
+= 1 << order
;
2538 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2539 int migratetype
, int *num_movable
)
2541 unsigned long start_pfn
, end_pfn
, pfn
;
2546 pfn
= page_to_pfn(page
);
2547 start_pfn
= pfn
& ~(pageblock_nr_pages
- 1);
2548 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2550 /* Do not cross zone boundaries */
2551 if (!zone_spans_pfn(zone
, start_pfn
))
2553 if (!zone_spans_pfn(zone
, end_pfn
))
2556 return move_freepages(zone
, start_pfn
, end_pfn
, migratetype
,
2560 static void change_pageblock_range(struct page
*pageblock_page
,
2561 int start_order
, int migratetype
)
2563 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2565 while (nr_pageblocks
--) {
2566 set_pageblock_migratetype(pageblock_page
, migratetype
);
2567 pageblock_page
+= pageblock_nr_pages
;
2572 * When we are falling back to another migratetype during allocation, try to
2573 * steal extra free pages from the same pageblocks to satisfy further
2574 * allocations, instead of polluting multiple pageblocks.
2576 * If we are stealing a relatively large buddy page, it is likely there will
2577 * be more free pages in the pageblock, so try to steal them all. For
2578 * reclaimable and unmovable allocations, we steal regardless of page size,
2579 * as fragmentation caused by those allocations polluting movable pageblocks
2580 * is worse than movable allocations stealing from unmovable and reclaimable
2583 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2586 * Leaving this order check is intended, although there is
2587 * relaxed order check in next check. The reason is that
2588 * we can actually steal whole pageblock if this condition met,
2589 * but, below check doesn't guarantee it and that is just heuristic
2590 * so could be changed anytime.
2592 if (order
>= pageblock_order
)
2595 if (order
>= pageblock_order
/ 2 ||
2596 start_mt
== MIGRATE_RECLAIMABLE
||
2597 start_mt
== MIGRATE_UNMOVABLE
||
2598 page_group_by_mobility_disabled
)
2604 static inline bool boost_watermark(struct zone
*zone
)
2606 unsigned long max_boost
;
2608 if (!watermark_boost_factor
)
2611 * Don't bother in zones that are unlikely to produce results.
2612 * On small machines, including kdump capture kernels running
2613 * in a small area, boosting the watermark can cause an out of
2614 * memory situation immediately.
2616 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
2619 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2620 watermark_boost_factor
, 10000);
2623 * high watermark may be uninitialised if fragmentation occurs
2624 * very early in boot so do not boost. We do not fall
2625 * through and boost by pageblock_nr_pages as failing
2626 * allocations that early means that reclaim is not going
2627 * to help and it may even be impossible to reclaim the
2628 * boosted watermark resulting in a hang.
2633 max_boost
= max(pageblock_nr_pages
, max_boost
);
2635 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2642 * This function implements actual steal behaviour. If order is large enough,
2643 * we can steal whole pageblock. If not, we first move freepages in this
2644 * pageblock to our migratetype and determine how many already-allocated pages
2645 * are there in the pageblock with a compatible migratetype. If at least half
2646 * of pages are free or compatible, we can change migratetype of the pageblock
2647 * itself, so pages freed in the future will be put on the correct free list.
2649 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2650 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2652 unsigned int current_order
= buddy_order(page
);
2653 int free_pages
, movable_pages
, alike_pages
;
2656 old_block_type
= get_pageblock_migratetype(page
);
2659 * This can happen due to races and we want to prevent broken
2660 * highatomic accounting.
2662 if (is_migrate_highatomic(old_block_type
))
2665 /* Take ownership for orders >= pageblock_order */
2666 if (current_order
>= pageblock_order
) {
2667 change_pageblock_range(page
, current_order
, start_type
);
2672 * Boost watermarks to increase reclaim pressure to reduce the
2673 * likelihood of future fallbacks. Wake kswapd now as the node
2674 * may be balanced overall and kswapd will not wake naturally.
2676 if (boost_watermark(zone
) && (alloc_flags
& ALLOC_KSWAPD
))
2677 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2679 /* We are not allowed to try stealing from the whole block */
2683 free_pages
= move_freepages_block(zone
, page
, start_type
,
2686 * Determine how many pages are compatible with our allocation.
2687 * For movable allocation, it's the number of movable pages which
2688 * we just obtained. For other types it's a bit more tricky.
2690 if (start_type
== MIGRATE_MOVABLE
) {
2691 alike_pages
= movable_pages
;
2694 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2695 * to MOVABLE pageblock, consider all non-movable pages as
2696 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2697 * vice versa, be conservative since we can't distinguish the
2698 * exact migratetype of non-movable pages.
2700 if (old_block_type
== MIGRATE_MOVABLE
)
2701 alike_pages
= pageblock_nr_pages
2702 - (free_pages
+ movable_pages
);
2707 /* moving whole block can fail due to zone boundary conditions */
2712 * If a sufficient number of pages in the block are either free or of
2713 * comparable migratability as our allocation, claim the whole block.
2715 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2716 page_group_by_mobility_disabled
)
2717 set_pageblock_migratetype(page
, start_type
);
2722 move_to_free_list(page
, zone
, current_order
, start_type
);
2726 * Check whether there is a suitable fallback freepage with requested order.
2727 * If only_stealable is true, this function returns fallback_mt only if
2728 * we can steal other freepages all together. This would help to reduce
2729 * fragmentation due to mixed migratetype pages in one pageblock.
2731 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2732 int migratetype
, bool only_stealable
, bool *can_steal
)
2737 if (area
->nr_free
== 0)
2742 fallback_mt
= fallbacks
[migratetype
][i
];
2743 if (fallback_mt
== MIGRATE_TYPES
)
2746 if (free_area_empty(area
, fallback_mt
))
2749 if (can_steal_fallback(order
, migratetype
))
2752 if (!only_stealable
)
2763 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2764 * there are no empty page blocks that contain a page with a suitable order
2766 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2767 unsigned int alloc_order
)
2770 unsigned long max_managed
, flags
;
2773 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2774 * Check is race-prone but harmless.
2776 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2777 if (zone
->nr_reserved_highatomic
>= max_managed
)
2780 spin_lock_irqsave(&zone
->lock
, flags
);
2782 /* Recheck the nr_reserved_highatomic limit under the lock */
2783 if (zone
->nr_reserved_highatomic
>= max_managed
)
2787 mt
= get_pageblock_migratetype(page
);
2788 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2789 && !is_migrate_cma(mt
)) {
2790 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2791 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2792 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2796 spin_unlock_irqrestore(&zone
->lock
, flags
);
2800 * Used when an allocation is about to fail under memory pressure. This
2801 * potentially hurts the reliability of high-order allocations when under
2802 * intense memory pressure but failed atomic allocations should be easier
2803 * to recover from than an OOM.
2805 * If @force is true, try to unreserve a pageblock even though highatomic
2806 * pageblock is exhausted.
2808 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2811 struct zonelist
*zonelist
= ac
->zonelist
;
2812 unsigned long flags
;
2819 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->highest_zoneidx
,
2822 * Preserve at least one pageblock unless memory pressure
2825 if (!force
&& zone
->nr_reserved_highatomic
<=
2829 spin_lock_irqsave(&zone
->lock
, flags
);
2830 for (order
= 0; order
< MAX_ORDER
; order
++) {
2831 struct free_area
*area
= &(zone
->free_area
[order
]);
2833 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2838 * In page freeing path, migratetype change is racy so
2839 * we can counter several free pages in a pageblock
2840 * in this loop although we changed the pageblock type
2841 * from highatomic to ac->migratetype. So we should
2842 * adjust the count once.
2844 if (is_migrate_highatomic_page(page
)) {
2846 * It should never happen but changes to
2847 * locking could inadvertently allow a per-cpu
2848 * drain to add pages to MIGRATE_HIGHATOMIC
2849 * while unreserving so be safe and watch for
2852 zone
->nr_reserved_highatomic
-= min(
2854 zone
->nr_reserved_highatomic
);
2858 * Convert to ac->migratetype and avoid the normal
2859 * pageblock stealing heuristics. Minimally, the caller
2860 * is doing the work and needs the pages. More
2861 * importantly, if the block was always converted to
2862 * MIGRATE_UNMOVABLE or another type then the number
2863 * of pageblocks that cannot be completely freed
2866 set_pageblock_migratetype(page
, ac
->migratetype
);
2867 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2870 spin_unlock_irqrestore(&zone
->lock
, flags
);
2874 spin_unlock_irqrestore(&zone
->lock
, flags
);
2881 * Try finding a free buddy page on the fallback list and put it on the free
2882 * list of requested migratetype, possibly along with other pages from the same
2883 * block, depending on fragmentation avoidance heuristics. Returns true if
2884 * fallback was found so that __rmqueue_smallest() can grab it.
2886 * The use of signed ints for order and current_order is a deliberate
2887 * deviation from the rest of this file, to make the for loop
2888 * condition simpler.
2890 static __always_inline
bool
2891 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2892 unsigned int alloc_flags
)
2894 struct free_area
*area
;
2896 int min_order
= order
;
2902 * Do not steal pages from freelists belonging to other pageblocks
2903 * i.e. orders < pageblock_order. If there are no local zones free,
2904 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2906 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2907 min_order
= pageblock_order
;
2910 * Find the largest available free page in the other list. This roughly
2911 * approximates finding the pageblock with the most free pages, which
2912 * would be too costly to do exactly.
2914 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2916 area
= &(zone
->free_area
[current_order
]);
2917 fallback_mt
= find_suitable_fallback(area
, current_order
,
2918 start_migratetype
, false, &can_steal
);
2919 if (fallback_mt
== -1)
2923 * We cannot steal all free pages from the pageblock and the
2924 * requested migratetype is movable. In that case it's better to
2925 * steal and split the smallest available page instead of the
2926 * largest available page, because even if the next movable
2927 * allocation falls back into a different pageblock than this
2928 * one, it won't cause permanent fragmentation.
2930 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2931 && current_order
> order
)
2940 for (current_order
= order
; current_order
< MAX_ORDER
;
2942 area
= &(zone
->free_area
[current_order
]);
2943 fallback_mt
= find_suitable_fallback(area
, current_order
,
2944 start_migratetype
, false, &can_steal
);
2945 if (fallback_mt
!= -1)
2950 * This should not happen - we already found a suitable fallback
2951 * when looking for the largest page.
2953 VM_BUG_ON(current_order
== MAX_ORDER
);
2956 page
= get_page_from_free_area(area
, fallback_mt
);
2958 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2961 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2962 start_migratetype
, fallback_mt
);
2969 * Do the hard work of removing an element from the buddy allocator.
2970 * Call me with the zone->lock already held.
2972 static __always_inline
struct page
*
2973 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2974 unsigned int alloc_flags
)
2978 if (IS_ENABLED(CONFIG_CMA
)) {
2980 * Balance movable allocations between regular and CMA areas by
2981 * allocating from CMA when over half of the zone's free memory
2982 * is in the CMA area.
2984 if (alloc_flags
& ALLOC_CMA
&&
2985 zone_page_state(zone
, NR_FREE_CMA_PAGES
) >
2986 zone_page_state(zone
, NR_FREE_PAGES
) / 2) {
2987 page
= __rmqueue_cma_fallback(zone
, order
);
2993 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2994 if (unlikely(!page
)) {
2995 if (alloc_flags
& ALLOC_CMA
)
2996 page
= __rmqueue_cma_fallback(zone
, order
);
2998 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
3004 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3009 * Obtain a specified number of elements from the buddy allocator, all under
3010 * a single hold of the lock, for efficiency. Add them to the supplied list.
3011 * Returns the number of new pages which were placed at *list.
3013 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
3014 unsigned long count
, struct list_head
*list
,
3015 int migratetype
, unsigned int alloc_flags
)
3017 int i
, allocated
= 0;
3020 * local_lock_irq held so equivalent to spin_lock_irqsave for
3021 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3023 spin_lock(&zone
->lock
);
3024 for (i
= 0; i
< count
; ++i
) {
3025 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
3027 if (unlikely(page
== NULL
))
3030 if (unlikely(check_pcp_refill(page
)))
3034 * Split buddy pages returned by expand() are received here in
3035 * physical page order. The page is added to the tail of
3036 * caller's list. From the callers perspective, the linked list
3037 * is ordered by page number under some conditions. This is
3038 * useful for IO devices that can forward direction from the
3039 * head, thus also in the physical page order. This is useful
3040 * for IO devices that can merge IO requests if the physical
3041 * pages are ordered properly.
3043 list_add_tail(&page
->lru
, list
);
3045 if (is_migrate_cma(get_pcppage_migratetype(page
)))
3046 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
3051 * i pages were removed from the buddy list even if some leak due
3052 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3053 * on i. Do not confuse with 'allocated' which is the number of
3054 * pages added to the pcp list.
3056 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
3057 spin_unlock(&zone
->lock
);
3063 * Called from the vmstat counter updater to drain pagesets of this
3064 * currently executing processor on remote nodes after they have
3067 * Note that this function must be called with the thread pinned to
3068 * a single processor.
3070 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
3072 unsigned long flags
;
3073 int to_drain
, batch
;
3075 local_lock_irqsave(&pagesets
.lock
, flags
);
3076 batch
= READ_ONCE(pcp
->batch
);
3077 to_drain
= min(pcp
->count
, batch
);
3079 free_pcppages_bulk(zone
, to_drain
, pcp
);
3080 local_unlock_irqrestore(&pagesets
.lock
, flags
);
3085 * Drain pcplists of the indicated processor and zone.
3087 * The processor must either be the current processor and the
3088 * thread pinned to the current processor or a processor that
3091 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
3093 unsigned long flags
;
3094 struct per_cpu_pages
*pcp
;
3096 local_lock_irqsave(&pagesets
.lock
, flags
);
3098 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
3100 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3102 local_unlock_irqrestore(&pagesets
.lock
, flags
);
3106 * Drain pcplists of all zones on the indicated processor.
3108 * The processor must either be the current processor and the
3109 * thread pinned to the current processor or a processor that
3112 static void drain_pages(unsigned int cpu
)
3116 for_each_populated_zone(zone
) {
3117 drain_pages_zone(cpu
, zone
);
3122 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3124 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3125 * the single zone's pages.
3127 void drain_local_pages(struct zone
*zone
)
3129 int cpu
= smp_processor_id();
3132 drain_pages_zone(cpu
, zone
);
3137 static void drain_local_pages_wq(struct work_struct
*work
)
3139 struct pcpu_drain
*drain
;
3141 drain
= container_of(work
, struct pcpu_drain
, work
);
3144 * drain_all_pages doesn't use proper cpu hotplug protection so
3145 * we can race with cpu offline when the WQ can move this from
3146 * a cpu pinned worker to an unbound one. We can operate on a different
3147 * cpu which is alright but we also have to make sure to not move to
3151 drain_local_pages(drain
->zone
);
3156 * The implementation of drain_all_pages(), exposing an extra parameter to
3157 * drain on all cpus.
3159 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3160 * not empty. The check for non-emptiness can however race with a free to
3161 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3162 * that need the guarantee that every CPU has drained can disable the
3163 * optimizing racy check.
3165 static void __drain_all_pages(struct zone
*zone
, bool force_all_cpus
)
3170 * Allocate in the BSS so we won't require allocation in
3171 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3173 static cpumask_t cpus_with_pcps
;
3176 * Make sure nobody triggers this path before mm_percpu_wq is fully
3179 if (WARN_ON_ONCE(!mm_percpu_wq
))
3183 * Do not drain if one is already in progress unless it's specific to
3184 * a zone. Such callers are primarily CMA and memory hotplug and need
3185 * the drain to be complete when the call returns.
3187 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
3190 mutex_lock(&pcpu_drain_mutex
);
3194 * We don't care about racing with CPU hotplug event
3195 * as offline notification will cause the notified
3196 * cpu to drain that CPU pcps and on_each_cpu_mask
3197 * disables preemption as part of its processing
3199 for_each_online_cpu(cpu
) {
3200 struct per_cpu_pages
*pcp
;
3202 bool has_pcps
= false;
3204 if (force_all_cpus
) {
3206 * The pcp.count check is racy, some callers need a
3207 * guarantee that no cpu is missed.
3211 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
3215 for_each_populated_zone(z
) {
3216 pcp
= per_cpu_ptr(z
->per_cpu_pageset
, cpu
);
3225 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
3227 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
3230 for_each_cpu(cpu
, &cpus_with_pcps
) {
3231 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
3234 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
3235 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
3237 for_each_cpu(cpu
, &cpus_with_pcps
)
3238 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
3240 mutex_unlock(&pcpu_drain_mutex
);
3244 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3246 * When zone parameter is non-NULL, spill just the single zone's pages.
3248 * Note that this can be extremely slow as the draining happens in a workqueue.
3250 void drain_all_pages(struct zone
*zone
)
3252 __drain_all_pages(zone
, false);
3255 #ifdef CONFIG_HIBERNATION
3258 * Touch the watchdog for every WD_PAGE_COUNT pages.
3260 #define WD_PAGE_COUNT (128*1024)
3262 void mark_free_pages(struct zone
*zone
)
3264 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
3265 unsigned long flags
;
3266 unsigned int order
, t
;
3269 if (zone_is_empty(zone
))
3272 spin_lock_irqsave(&zone
->lock
, flags
);
3274 max_zone_pfn
= zone_end_pfn(zone
);
3275 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
3276 if (pfn_valid(pfn
)) {
3277 page
= pfn_to_page(pfn
);
3279 if (!--page_count
) {
3280 touch_nmi_watchdog();
3281 page_count
= WD_PAGE_COUNT
;
3284 if (page_zone(page
) != zone
)
3287 if (!swsusp_page_is_forbidden(page
))
3288 swsusp_unset_page_free(page
);
3291 for_each_migratetype_order(order
, t
) {
3292 list_for_each_entry(page
,
3293 &zone
->free_area
[order
].free_list
[t
], lru
) {
3296 pfn
= page_to_pfn(page
);
3297 for (i
= 0; i
< (1UL << order
); i
++) {
3298 if (!--page_count
) {
3299 touch_nmi_watchdog();
3300 page_count
= WD_PAGE_COUNT
;
3302 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
3306 spin_unlock_irqrestore(&zone
->lock
, flags
);
3308 #endif /* CONFIG_PM */
3310 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
,
3315 if (!free_pcp_prepare(page
, order
))
3318 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3319 set_pcppage_migratetype(page
, migratetype
);
3323 static int nr_pcp_free(struct per_cpu_pages
*pcp
, int high
, int batch
)
3325 int min_nr_free
, max_nr_free
;
3327 /* Check for PCP disabled or boot pageset */
3328 if (unlikely(high
< batch
))
3331 /* Leave at least pcp->batch pages on the list */
3332 min_nr_free
= batch
;
3333 max_nr_free
= high
- batch
;
3336 * Double the number of pages freed each time there is subsequent
3337 * freeing of pages without any allocation.
3339 batch
<<= pcp
->free_factor
;
3340 if (batch
< max_nr_free
)
3342 batch
= clamp(batch
, min_nr_free
, max_nr_free
);
3347 static int nr_pcp_high(struct per_cpu_pages
*pcp
, struct zone
*zone
)
3349 int high
= READ_ONCE(pcp
->high
);
3351 if (unlikely(!high
))
3354 if (!test_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
))
3358 * If reclaim is active, limit the number of pages that can be
3359 * stored on pcp lists
3361 return min(READ_ONCE(pcp
->batch
) << 2, high
);
3364 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
,
3365 int migratetype
, unsigned int order
)
3367 struct zone
*zone
= page_zone(page
);
3368 struct per_cpu_pages
*pcp
;
3372 __count_vm_event(PGFREE
);
3373 pcp
= this_cpu_ptr(zone
->per_cpu_pageset
);
3374 pindex
= order_to_pindex(migratetype
, order
);
3375 list_add(&page
->lru
, &pcp
->lists
[pindex
]);
3376 pcp
->count
+= 1 << order
;
3377 high
= nr_pcp_high(pcp
, zone
);
3378 if (pcp
->count
>= high
) {
3379 int batch
= READ_ONCE(pcp
->batch
);
3381 free_pcppages_bulk(zone
, nr_pcp_free(pcp
, high
, batch
), pcp
);
3388 void free_unref_page(struct page
*page
, unsigned int order
)
3390 unsigned long flags
;
3391 unsigned long pfn
= page_to_pfn(page
);
3394 if (!free_unref_page_prepare(page
, pfn
, order
))
3398 * We only track unmovable, reclaimable and movable on pcp lists.
3399 * Place ISOLATE pages on the isolated list because they are being
3400 * offlined but treat HIGHATOMIC as movable pages so we can get those
3401 * areas back if necessary. Otherwise, we may have to free
3402 * excessively into the page allocator
3404 migratetype
= get_pcppage_migratetype(page
);
3405 if (unlikely(migratetype
>= MIGRATE_PCPTYPES
)) {
3406 if (unlikely(is_migrate_isolate(migratetype
))) {
3407 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
, FPI_NONE
);
3410 migratetype
= MIGRATE_MOVABLE
;
3413 local_lock_irqsave(&pagesets
.lock
, flags
);
3414 free_unref_page_commit(page
, pfn
, migratetype
, order
);
3415 local_unlock_irqrestore(&pagesets
.lock
, flags
);
3419 * Free a list of 0-order pages
3421 void free_unref_page_list(struct list_head
*list
)
3423 struct page
*page
, *next
;
3424 unsigned long flags
, pfn
;
3425 int batch_count
= 0;
3428 /* Prepare pages for freeing */
3429 list_for_each_entry_safe(page
, next
, list
, lru
) {
3430 pfn
= page_to_pfn(page
);
3431 if (!free_unref_page_prepare(page
, pfn
, 0))
3432 list_del(&page
->lru
);
3435 * Free isolated pages directly to the allocator, see
3436 * comment in free_unref_page.
3438 migratetype
= get_pcppage_migratetype(page
);
3439 if (unlikely(is_migrate_isolate(migratetype
))) {
3440 list_del(&page
->lru
);
3441 free_one_page(page_zone(page
), page
, pfn
, 0, migratetype
, FPI_NONE
);
3445 set_page_private(page
, pfn
);
3448 local_lock_irqsave(&pagesets
.lock
, flags
);
3449 list_for_each_entry_safe(page
, next
, list
, lru
) {
3450 pfn
= page_private(page
);
3451 set_page_private(page
, 0);
3454 * Non-isolated types over MIGRATE_PCPTYPES get added
3455 * to the MIGRATE_MOVABLE pcp list.
3457 migratetype
= get_pcppage_migratetype(page
);
3458 if (unlikely(migratetype
>= MIGRATE_PCPTYPES
))
3459 migratetype
= MIGRATE_MOVABLE
;
3461 trace_mm_page_free_batched(page
);
3462 free_unref_page_commit(page
, pfn
, migratetype
, 0);
3465 * Guard against excessive IRQ disabled times when we get
3466 * a large list of pages to free.
3468 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3469 local_unlock_irqrestore(&pagesets
.lock
, flags
);
3471 local_lock_irqsave(&pagesets
.lock
, flags
);
3474 local_unlock_irqrestore(&pagesets
.lock
, flags
);
3478 * split_page takes a non-compound higher-order page, and splits it into
3479 * n (1<<order) sub-pages: page[0..n]
3480 * Each sub-page must be freed individually.
3482 * Note: this is probably too low level an operation for use in drivers.
3483 * Please consult with lkml before using this in your driver.
3485 void split_page(struct page
*page
, unsigned int order
)
3489 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3490 VM_BUG_ON_PAGE(!page_count(page
), page
);
3492 for (i
= 1; i
< (1 << order
); i
++)
3493 set_page_refcounted(page
+ i
);
3494 split_page_owner(page
, 1 << order
);
3495 split_page_memcg(page
, 1 << order
);
3497 EXPORT_SYMBOL_GPL(split_page
);
3499 int __isolate_free_page(struct page
*page
, unsigned int order
)
3501 unsigned long watermark
;
3505 BUG_ON(!PageBuddy(page
));
3507 zone
= page_zone(page
);
3508 mt
= get_pageblock_migratetype(page
);
3510 if (!is_migrate_isolate(mt
)) {
3512 * Obey watermarks as if the page was being allocated. We can
3513 * emulate a high-order watermark check with a raised order-0
3514 * watermark, because we already know our high-order page
3517 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3518 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3521 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3524 /* Remove page from free list */
3526 del_page_from_free_list(page
, zone
, order
);
3529 * Set the pageblock if the isolated page is at least half of a
3532 if (order
>= pageblock_order
- 1) {
3533 struct page
*endpage
= page
+ (1 << order
) - 1;
3534 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3535 int mt
= get_pageblock_migratetype(page
);
3536 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3537 && !is_migrate_highatomic(mt
))
3538 set_pageblock_migratetype(page
,
3544 return 1UL << order
;
3548 * __putback_isolated_page - Return a now-isolated page back where we got it
3549 * @page: Page that was isolated
3550 * @order: Order of the isolated page
3551 * @mt: The page's pageblock's migratetype
3553 * This function is meant to return a page pulled from the free lists via
3554 * __isolate_free_page back to the free lists they were pulled from.
3556 void __putback_isolated_page(struct page
*page
, unsigned int order
, int mt
)
3558 struct zone
*zone
= page_zone(page
);
3560 /* zone lock should be held when this function is called */
3561 lockdep_assert_held(&zone
->lock
);
3563 /* Return isolated page to tail of freelist. */
3564 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
,
3565 FPI_SKIP_REPORT_NOTIFY
| FPI_TO_TAIL
);
3569 * Update NUMA hit/miss statistics
3571 * Must be called with interrupts disabled.
3573 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
3577 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3579 /* skip numa counters update if numa stats is disabled */
3580 if (!static_branch_likely(&vm_numa_stat_key
))
3583 if (zone_to_nid(z
) != numa_node_id())
3584 local_stat
= NUMA_OTHER
;
3586 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3587 __count_numa_events(z
, NUMA_HIT
, nr_account
);
3589 __count_numa_events(z
, NUMA_MISS
, nr_account
);
3590 __count_numa_events(preferred_zone
, NUMA_FOREIGN
, nr_account
);
3592 __count_numa_events(z
, local_stat
, nr_account
);
3596 /* Remove page from the per-cpu list, caller must protect the list */
3598 struct page
*__rmqueue_pcplist(struct zone
*zone
, unsigned int order
,
3600 unsigned int alloc_flags
,
3601 struct per_cpu_pages
*pcp
,
3602 struct list_head
*list
)
3607 if (list_empty(list
)) {
3608 int batch
= READ_ONCE(pcp
->batch
);
3612 * Scale batch relative to order if batch implies
3613 * free pages can be stored on the PCP. Batch can
3614 * be 1 for small zones or for boot pagesets which
3615 * should never store free pages as the pages may
3616 * belong to arbitrary zones.
3619 batch
= max(batch
>> order
, 2);
3620 alloced
= rmqueue_bulk(zone
, order
,
3622 migratetype
, alloc_flags
);
3624 pcp
->count
+= alloced
<< order
;
3625 if (unlikely(list_empty(list
)))
3629 page
= list_first_entry(list
, struct page
, lru
);
3630 list_del(&page
->lru
);
3631 pcp
->count
-= 1 << order
;
3632 } while (check_new_pcp(page
));
3637 /* Lock and remove page from the per-cpu list */
3638 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3639 struct zone
*zone
, unsigned int order
,
3640 gfp_t gfp_flags
, int migratetype
,
3641 unsigned int alloc_flags
)
3643 struct per_cpu_pages
*pcp
;
3644 struct list_head
*list
;
3646 unsigned long flags
;
3648 local_lock_irqsave(&pagesets
.lock
, flags
);
3651 * On allocation, reduce the number of pages that are batch freed.
3652 * See nr_pcp_free() where free_factor is increased for subsequent
3655 pcp
= this_cpu_ptr(zone
->per_cpu_pageset
);
3656 pcp
->free_factor
>>= 1;
3657 list
= &pcp
->lists
[order_to_pindex(migratetype
, order
)];
3658 page
= __rmqueue_pcplist(zone
, order
, migratetype
, alloc_flags
, pcp
, list
);
3659 local_unlock_irqrestore(&pagesets
.lock
, flags
);
3661 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3662 zone_statistics(preferred_zone
, zone
, 1);
3668 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3671 struct page
*rmqueue(struct zone
*preferred_zone
,
3672 struct zone
*zone
, unsigned int order
,
3673 gfp_t gfp_flags
, unsigned int alloc_flags
,
3676 unsigned long flags
;
3679 if (likely(pcp_allowed_order(order
))) {
3681 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3682 * we need to skip it when CMA area isn't allowed.
3684 if (!IS_ENABLED(CONFIG_CMA
) || alloc_flags
& ALLOC_CMA
||
3685 migratetype
!= MIGRATE_MOVABLE
) {
3686 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
3687 gfp_flags
, migratetype
, alloc_flags
);
3693 * We most definitely don't want callers attempting to
3694 * allocate greater than order-1 page units with __GFP_NOFAIL.
3696 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3697 spin_lock_irqsave(&zone
->lock
, flags
);
3702 * order-0 request can reach here when the pcplist is skipped
3703 * due to non-CMA allocation context. HIGHATOMIC area is
3704 * reserved for high-order atomic allocation, so order-0
3705 * request should skip it.
3707 if (order
> 0 && alloc_flags
& ALLOC_HARDER
) {
3708 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3710 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3713 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3714 } while (page
&& check_new_pages(page
, order
));
3718 __mod_zone_freepage_state(zone
, -(1 << order
),
3719 get_pcppage_migratetype(page
));
3720 spin_unlock_irqrestore(&zone
->lock
, flags
);
3722 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3723 zone_statistics(preferred_zone
, zone
, 1);
3726 /* Separate test+clear to avoid unnecessary atomics */
3727 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3728 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3729 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3732 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3736 spin_unlock_irqrestore(&zone
->lock
, flags
);
3740 #ifdef CONFIG_FAIL_PAGE_ALLOC
3743 struct fault_attr attr
;
3745 bool ignore_gfp_highmem
;
3746 bool ignore_gfp_reclaim
;
3748 } fail_page_alloc
= {
3749 .attr
= FAULT_ATTR_INITIALIZER
,
3750 .ignore_gfp_reclaim
= true,
3751 .ignore_gfp_highmem
= true,
3755 static int __init
setup_fail_page_alloc(char *str
)
3757 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3759 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3761 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3763 if (order
< fail_page_alloc
.min_order
)
3765 if (gfp_mask
& __GFP_NOFAIL
)
3767 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3769 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3770 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3773 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3776 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3778 static int __init
fail_page_alloc_debugfs(void)
3780 umode_t mode
= S_IFREG
| 0600;
3783 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3784 &fail_page_alloc
.attr
);
3786 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3787 &fail_page_alloc
.ignore_gfp_reclaim
);
3788 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3789 &fail_page_alloc
.ignore_gfp_highmem
);
3790 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3795 late_initcall(fail_page_alloc_debugfs
);
3797 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3799 #else /* CONFIG_FAIL_PAGE_ALLOC */
3801 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3806 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3808 noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3810 return __should_fail_alloc_page(gfp_mask
, order
);
3812 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3814 static inline long __zone_watermark_unusable_free(struct zone
*z
,
3815 unsigned int order
, unsigned int alloc_flags
)
3817 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3818 long unusable_free
= (1 << order
) - 1;
3821 * If the caller does not have rights to ALLOC_HARDER then subtract
3822 * the high-atomic reserves. This will over-estimate the size of the
3823 * atomic reserve but it avoids a search.
3825 if (likely(!alloc_harder
))
3826 unusable_free
+= z
->nr_reserved_highatomic
;
3829 /* If allocation can't use CMA areas don't use free CMA pages */
3830 if (!(alloc_flags
& ALLOC_CMA
))
3831 unusable_free
+= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3834 return unusable_free
;
3838 * Return true if free base pages are above 'mark'. For high-order checks it
3839 * will return true of the order-0 watermark is reached and there is at least
3840 * one free page of a suitable size. Checking now avoids taking the zone lock
3841 * to check in the allocation paths if no pages are free.
3843 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3844 int highest_zoneidx
, unsigned int alloc_flags
,
3849 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3851 /* free_pages may go negative - that's OK */
3852 free_pages
-= __zone_watermark_unusable_free(z
, order
, alloc_flags
);
3854 if (alloc_flags
& ALLOC_HIGH
)
3857 if (unlikely(alloc_harder
)) {
3859 * OOM victims can try even harder than normal ALLOC_HARDER
3860 * users on the grounds that it's definitely going to be in
3861 * the exit path shortly and free memory. Any allocation it
3862 * makes during the free path will be small and short-lived.
3864 if (alloc_flags
& ALLOC_OOM
)
3871 * Check watermarks for an order-0 allocation request. If these
3872 * are not met, then a high-order request also cannot go ahead
3873 * even if a suitable page happened to be free.
3875 if (free_pages
<= min
+ z
->lowmem_reserve
[highest_zoneidx
])
3878 /* If this is an order-0 request then the watermark is fine */
3882 /* For a high-order request, check at least one suitable page is free */
3883 for (o
= order
; o
< MAX_ORDER
; o
++) {
3884 struct free_area
*area
= &z
->free_area
[o
];
3890 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3891 if (!free_area_empty(area
, mt
))
3896 if ((alloc_flags
& ALLOC_CMA
) &&
3897 !free_area_empty(area
, MIGRATE_CMA
)) {
3901 if (alloc_harder
&& !free_area_empty(area
, MIGRATE_HIGHATOMIC
))
3907 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3908 int highest_zoneidx
, unsigned int alloc_flags
)
3910 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3911 zone_page_state(z
, NR_FREE_PAGES
));
3914 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3915 unsigned long mark
, int highest_zoneidx
,
3916 unsigned int alloc_flags
, gfp_t gfp_mask
)
3920 free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3923 * Fast check for order-0 only. If this fails then the reserves
3924 * need to be calculated.
3929 fast_free
= free_pages
;
3930 fast_free
-= __zone_watermark_unusable_free(z
, 0, alloc_flags
);
3931 if (fast_free
> mark
+ z
->lowmem_reserve
[highest_zoneidx
])
3935 if (__zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3939 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3940 * when checking the min watermark. The min watermark is the
3941 * point where boosting is ignored so that kswapd is woken up
3942 * when below the low watermark.
3944 if (unlikely(!order
&& (gfp_mask
& __GFP_ATOMIC
) && z
->watermark_boost
3945 && ((alloc_flags
& ALLOC_WMARK_MASK
) == WMARK_MIN
))) {
3946 mark
= z
->_watermark
[WMARK_MIN
];
3947 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
,
3948 alloc_flags
, free_pages
);
3954 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3955 unsigned long mark
, int highest_zoneidx
)
3957 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3959 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3960 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3962 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, 0,
3967 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3969 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3970 node_reclaim_distance
;
3972 #else /* CONFIG_NUMA */
3973 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3977 #endif /* CONFIG_NUMA */
3980 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3981 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3982 * premature use of a lower zone may cause lowmem pressure problems that
3983 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3984 * probably too small. It only makes sense to spread allocations to avoid
3985 * fragmentation between the Normal and DMA32 zones.
3987 static inline unsigned int
3988 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3990 unsigned int alloc_flags
;
3993 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3996 alloc_flags
= (__force
int) (gfp_mask
& __GFP_KSWAPD_RECLAIM
);
3998 #ifdef CONFIG_ZONE_DMA32
4002 if (zone_idx(zone
) != ZONE_NORMAL
)
4006 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4007 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4008 * on UMA that if Normal is populated then so is DMA32.
4010 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
4011 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
4014 alloc_flags
|= ALLOC_NOFRAGMENT
;
4015 #endif /* CONFIG_ZONE_DMA32 */
4019 /* Must be called after current_gfp_context() which can change gfp_mask */
4020 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask
,
4021 unsigned int alloc_flags
)
4024 if (gfp_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
4025 alloc_flags
|= ALLOC_CMA
;
4031 * get_page_from_freelist goes through the zonelist trying to allocate
4034 static struct page
*
4035 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
4036 const struct alloc_context
*ac
)
4040 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
4045 * Scan zonelist, looking for a zone with enough free.
4046 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4048 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
4049 z
= ac
->preferred_zoneref
;
4050 for_next_zone_zonelist_nodemask(zone
, z
, ac
->highest_zoneidx
,
4055 if (cpusets_enabled() &&
4056 (alloc_flags
& ALLOC_CPUSET
) &&
4057 !__cpuset_zone_allowed(zone
, gfp_mask
))
4060 * When allocating a page cache page for writing, we
4061 * want to get it from a node that is within its dirty
4062 * limit, such that no single node holds more than its
4063 * proportional share of globally allowed dirty pages.
4064 * The dirty limits take into account the node's
4065 * lowmem reserves and high watermark so that kswapd
4066 * should be able to balance it without having to
4067 * write pages from its LRU list.
4069 * XXX: For now, allow allocations to potentially
4070 * exceed the per-node dirty limit in the slowpath
4071 * (spread_dirty_pages unset) before going into reclaim,
4072 * which is important when on a NUMA setup the allowed
4073 * nodes are together not big enough to reach the
4074 * global limit. The proper fix for these situations
4075 * will require awareness of nodes in the
4076 * dirty-throttling and the flusher threads.
4078 if (ac
->spread_dirty_pages
) {
4079 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
4082 if (!node_dirty_ok(zone
->zone_pgdat
)) {
4083 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
4088 if (no_fallback
&& nr_online_nodes
> 1 &&
4089 zone
!= ac
->preferred_zoneref
->zone
) {
4093 * If moving to a remote node, retry but allow
4094 * fragmenting fallbacks. Locality is more important
4095 * than fragmentation avoidance.
4097 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
4098 if (zone_to_nid(zone
) != local_nid
) {
4099 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
4104 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
4105 if (!zone_watermark_fast(zone
, order
, mark
,
4106 ac
->highest_zoneidx
, alloc_flags
,
4110 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4112 * Watermark failed for this zone, but see if we can
4113 * grow this zone if it contains deferred pages.
4115 if (static_branch_unlikely(&deferred_pages
)) {
4116 if (_deferred_grow_zone(zone
, order
))
4120 /* Checked here to keep the fast path fast */
4121 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
4122 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
4125 if (!node_reclaim_enabled() ||
4126 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
4129 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
4131 case NODE_RECLAIM_NOSCAN
:
4134 case NODE_RECLAIM_FULL
:
4135 /* scanned but unreclaimable */
4138 /* did we reclaim enough */
4139 if (zone_watermark_ok(zone
, order
, mark
,
4140 ac
->highest_zoneidx
, alloc_flags
))
4148 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
4149 gfp_mask
, alloc_flags
, ac
->migratetype
);
4151 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
4154 * If this is a high-order atomic allocation then check
4155 * if the pageblock should be reserved for the future
4157 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
4158 reserve_highatomic_pageblock(page
, zone
, order
);
4162 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4163 /* Try again if zone has deferred pages */
4164 if (static_branch_unlikely(&deferred_pages
)) {
4165 if (_deferred_grow_zone(zone
, order
))
4173 * It's possible on a UMA machine to get through all zones that are
4174 * fragmented. If avoiding fragmentation, reset and try again.
4177 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
4184 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
4186 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
4189 * This documents exceptions given to allocations in certain
4190 * contexts that are allowed to allocate outside current's set
4193 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4194 if (tsk_is_oom_victim(current
) ||
4195 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
4196 filter
&= ~SHOW_MEM_FILTER_NODES
;
4197 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4198 filter
&= ~SHOW_MEM_FILTER_NODES
;
4200 show_mem(filter
, nodemask
);
4203 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
4205 struct va_format vaf
;
4207 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
4209 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
4212 va_start(args
, fmt
);
4215 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4216 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
4217 nodemask_pr_args(nodemask
));
4220 cpuset_print_current_mems_allowed();
4223 warn_alloc_show_mem(gfp_mask
, nodemask
);
4226 static inline struct page
*
4227 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
4228 unsigned int alloc_flags
,
4229 const struct alloc_context
*ac
)
4233 page
= get_page_from_freelist(gfp_mask
, order
,
4234 alloc_flags
|ALLOC_CPUSET
, ac
);
4236 * fallback to ignore cpuset restriction if our nodes
4240 page
= get_page_from_freelist(gfp_mask
, order
,
4246 static inline struct page
*
4247 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
4248 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
4250 struct oom_control oc
= {
4251 .zonelist
= ac
->zonelist
,
4252 .nodemask
= ac
->nodemask
,
4254 .gfp_mask
= gfp_mask
,
4259 *did_some_progress
= 0;
4262 * Acquire the oom lock. If that fails, somebody else is
4263 * making progress for us.
4265 if (!mutex_trylock(&oom_lock
)) {
4266 *did_some_progress
= 1;
4267 schedule_timeout_uninterruptible(1);
4272 * Go through the zonelist yet one more time, keep very high watermark
4273 * here, this is only to catch a parallel oom killing, we must fail if
4274 * we're still under heavy pressure. But make sure that this reclaim
4275 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4276 * allocation which will never fail due to oom_lock already held.
4278 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
4279 ~__GFP_DIRECT_RECLAIM
, order
,
4280 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
4284 /* Coredumps can quickly deplete all memory reserves */
4285 if (current
->flags
& PF_DUMPCORE
)
4287 /* The OOM killer will not help higher order allocs */
4288 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4291 * We have already exhausted all our reclaim opportunities without any
4292 * success so it is time to admit defeat. We will skip the OOM killer
4293 * because it is very likely that the caller has a more reasonable
4294 * fallback than shooting a random task.
4296 * The OOM killer may not free memory on a specific node.
4298 if (gfp_mask
& (__GFP_RETRY_MAYFAIL
| __GFP_THISNODE
))
4300 /* The OOM killer does not needlessly kill tasks for lowmem */
4301 if (ac
->highest_zoneidx
< ZONE_NORMAL
)
4303 if (pm_suspended_storage())
4306 * XXX: GFP_NOFS allocations should rather fail than rely on
4307 * other request to make a forward progress.
4308 * We are in an unfortunate situation where out_of_memory cannot
4309 * do much for this context but let's try it to at least get
4310 * access to memory reserved if the current task is killed (see
4311 * out_of_memory). Once filesystems are ready to handle allocation
4312 * failures more gracefully we should just bail out here.
4315 /* Exhausted what can be done so it's blame time */
4316 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
4317 *did_some_progress
= 1;
4320 * Help non-failing allocations by giving them access to memory
4323 if (gfp_mask
& __GFP_NOFAIL
)
4324 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
4325 ALLOC_NO_WATERMARKS
, ac
);
4328 mutex_unlock(&oom_lock
);
4333 * Maximum number of compaction retries with a progress before OOM
4334 * killer is consider as the only way to move forward.
4336 #define MAX_COMPACT_RETRIES 16
4338 #ifdef CONFIG_COMPACTION
4339 /* Try memory compaction for high-order allocations before reclaim */
4340 static struct page
*
4341 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4342 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4343 enum compact_priority prio
, enum compact_result
*compact_result
)
4345 struct page
*page
= NULL
;
4346 unsigned long pflags
;
4347 unsigned int noreclaim_flag
;
4352 psi_memstall_enter(&pflags
);
4353 noreclaim_flag
= memalloc_noreclaim_save();
4355 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
4358 memalloc_noreclaim_restore(noreclaim_flag
);
4359 psi_memstall_leave(&pflags
);
4361 if (*compact_result
== COMPACT_SKIPPED
)
4364 * At least in one zone compaction wasn't deferred or skipped, so let's
4365 * count a compaction stall
4367 count_vm_event(COMPACTSTALL
);
4369 /* Prep a captured page if available */
4371 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
4373 /* Try get a page from the freelist if available */
4375 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4378 struct zone
*zone
= page_zone(page
);
4380 zone
->compact_blockskip_flush
= false;
4381 compaction_defer_reset(zone
, order
, true);
4382 count_vm_event(COMPACTSUCCESS
);
4387 * It's bad if compaction run occurs and fails. The most likely reason
4388 * is that pages exist, but not enough to satisfy watermarks.
4390 count_vm_event(COMPACTFAIL
);
4398 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
4399 enum compact_result compact_result
,
4400 enum compact_priority
*compact_priority
,
4401 int *compaction_retries
)
4403 int max_retries
= MAX_COMPACT_RETRIES
;
4406 int retries
= *compaction_retries
;
4407 enum compact_priority priority
= *compact_priority
;
4412 if (fatal_signal_pending(current
))
4415 if (compaction_made_progress(compact_result
))
4416 (*compaction_retries
)++;
4419 * compaction considers all the zone as desperately out of memory
4420 * so it doesn't really make much sense to retry except when the
4421 * failure could be caused by insufficient priority
4423 if (compaction_failed(compact_result
))
4424 goto check_priority
;
4427 * compaction was skipped because there are not enough order-0 pages
4428 * to work with, so we retry only if it looks like reclaim can help.
4430 if (compaction_needs_reclaim(compact_result
)) {
4431 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
4436 * make sure the compaction wasn't deferred or didn't bail out early
4437 * due to locks contention before we declare that we should give up.
4438 * But the next retry should use a higher priority if allowed, so
4439 * we don't just keep bailing out endlessly.
4441 if (compaction_withdrawn(compact_result
)) {
4442 goto check_priority
;
4446 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4447 * costly ones because they are de facto nofail and invoke OOM
4448 * killer to move on while costly can fail and users are ready
4449 * to cope with that. 1/4 retries is rather arbitrary but we
4450 * would need much more detailed feedback from compaction to
4451 * make a better decision.
4453 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4455 if (*compaction_retries
<= max_retries
) {
4461 * Make sure there are attempts at the highest priority if we exhausted
4462 * all retries or failed at the lower priorities.
4465 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
4466 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
4468 if (*compact_priority
> min_priority
) {
4469 (*compact_priority
)--;
4470 *compaction_retries
= 0;
4474 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4478 static inline struct page
*
4479 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4480 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4481 enum compact_priority prio
, enum compact_result
*compact_result
)
4483 *compact_result
= COMPACT_SKIPPED
;
4488 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4489 enum compact_result compact_result
,
4490 enum compact_priority
*compact_priority
,
4491 int *compaction_retries
)
4496 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4500 * There are setups with compaction disabled which would prefer to loop
4501 * inside the allocator rather than hit the oom killer prematurely.
4502 * Let's give them a good hope and keep retrying while the order-0
4503 * watermarks are OK.
4505 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4506 ac
->highest_zoneidx
, ac
->nodemask
) {
4507 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4508 ac
->highest_zoneidx
, alloc_flags
))
4513 #endif /* CONFIG_COMPACTION */
4515 #ifdef CONFIG_LOCKDEP
4516 static struct lockdep_map __fs_reclaim_map
=
4517 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4519 static bool __need_reclaim(gfp_t gfp_mask
)
4521 /* no reclaim without waiting on it */
4522 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4525 /* this guy won't enter reclaim */
4526 if (current
->flags
& PF_MEMALLOC
)
4529 if (gfp_mask
& __GFP_NOLOCKDEP
)
4535 void __fs_reclaim_acquire(void)
4537 lock_map_acquire(&__fs_reclaim_map
);
4540 void __fs_reclaim_release(void)
4542 lock_map_release(&__fs_reclaim_map
);
4545 void fs_reclaim_acquire(gfp_t gfp_mask
)
4547 gfp_mask
= current_gfp_context(gfp_mask
);
4549 if (__need_reclaim(gfp_mask
)) {
4550 if (gfp_mask
& __GFP_FS
)
4551 __fs_reclaim_acquire();
4553 #ifdef CONFIG_MMU_NOTIFIER
4554 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
4555 lock_map_release(&__mmu_notifier_invalidate_range_start_map
);
4560 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4562 void fs_reclaim_release(gfp_t gfp_mask
)
4564 gfp_mask
= current_gfp_context(gfp_mask
);
4566 if (__need_reclaim(gfp_mask
)) {
4567 if (gfp_mask
& __GFP_FS
)
4568 __fs_reclaim_release();
4571 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4574 /* Perform direct synchronous page reclaim */
4575 static unsigned long
4576 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4577 const struct alloc_context
*ac
)
4579 unsigned int noreclaim_flag
;
4580 unsigned long pflags
, progress
;
4584 /* We now go into synchronous reclaim */
4585 cpuset_memory_pressure_bump();
4586 psi_memstall_enter(&pflags
);
4587 fs_reclaim_acquire(gfp_mask
);
4588 noreclaim_flag
= memalloc_noreclaim_save();
4590 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4593 memalloc_noreclaim_restore(noreclaim_flag
);
4594 fs_reclaim_release(gfp_mask
);
4595 psi_memstall_leave(&pflags
);
4602 /* The really slow allocator path where we enter direct reclaim */
4603 static inline struct page
*
4604 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4605 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4606 unsigned long *did_some_progress
)
4608 struct page
*page
= NULL
;
4609 bool drained
= false;
4611 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4612 if (unlikely(!(*did_some_progress
)))
4616 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4619 * If an allocation failed after direct reclaim, it could be because
4620 * pages are pinned on the per-cpu lists or in high alloc reserves.
4621 * Shrink them and try again
4623 if (!page
&& !drained
) {
4624 unreserve_highatomic_pageblock(ac
, false);
4625 drain_all_pages(NULL
);
4633 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4634 const struct alloc_context
*ac
)
4638 pg_data_t
*last_pgdat
= NULL
;
4639 enum zone_type highest_zoneidx
= ac
->highest_zoneidx
;
4641 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, highest_zoneidx
,
4643 if (last_pgdat
!= zone
->zone_pgdat
)
4644 wakeup_kswapd(zone
, gfp_mask
, order
, highest_zoneidx
);
4645 last_pgdat
= zone
->zone_pgdat
;
4649 static inline unsigned int
4650 gfp_to_alloc_flags(gfp_t gfp_mask
)
4652 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4655 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4656 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4657 * to save two branches.
4659 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4660 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM
!= (__force gfp_t
) ALLOC_KSWAPD
);
4663 * The caller may dip into page reserves a bit more if the caller
4664 * cannot run direct reclaim, or if the caller has realtime scheduling
4665 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4666 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4668 alloc_flags
|= (__force
int)
4669 (gfp_mask
& (__GFP_HIGH
| __GFP_KSWAPD_RECLAIM
));
4671 if (gfp_mask
& __GFP_ATOMIC
) {
4673 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4674 * if it can't schedule.
4676 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4677 alloc_flags
|= ALLOC_HARDER
;
4679 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4680 * comment for __cpuset_node_allowed().
4682 alloc_flags
&= ~ALLOC_CPUSET
;
4683 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4684 alloc_flags
|= ALLOC_HARDER
;
4686 alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, alloc_flags
);
4691 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4693 if (!tsk_is_oom_victim(tsk
))
4697 * !MMU doesn't have oom reaper so give access to memory reserves
4698 * only to the thread with TIF_MEMDIE set
4700 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4707 * Distinguish requests which really need access to full memory
4708 * reserves from oom victims which can live with a portion of it
4710 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4712 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4714 if (gfp_mask
& __GFP_MEMALLOC
)
4715 return ALLOC_NO_WATERMARKS
;
4716 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4717 return ALLOC_NO_WATERMARKS
;
4718 if (!in_interrupt()) {
4719 if (current
->flags
& PF_MEMALLOC
)
4720 return ALLOC_NO_WATERMARKS
;
4721 else if (oom_reserves_allowed(current
))
4728 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4730 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4734 * Checks whether it makes sense to retry the reclaim to make a forward progress
4735 * for the given allocation request.
4737 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4738 * without success, or when we couldn't even meet the watermark if we
4739 * reclaimed all remaining pages on the LRU lists.
4741 * Returns true if a retry is viable or false to enter the oom path.
4744 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4745 struct alloc_context
*ac
, int alloc_flags
,
4746 bool did_some_progress
, int *no_progress_loops
)
4753 * Costly allocations might have made a progress but this doesn't mean
4754 * their order will become available due to high fragmentation so
4755 * always increment the no progress counter for them
4757 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4758 *no_progress_loops
= 0;
4760 (*no_progress_loops
)++;
4763 * Make sure we converge to OOM if we cannot make any progress
4764 * several times in the row.
4766 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4767 /* Before OOM, exhaust highatomic_reserve */
4768 return unreserve_highatomic_pageblock(ac
, true);
4772 * Keep reclaiming pages while there is a chance this will lead
4773 * somewhere. If none of the target zones can satisfy our allocation
4774 * request even if all reclaimable pages are considered then we are
4775 * screwed and have to go OOM.
4777 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4778 ac
->highest_zoneidx
, ac
->nodemask
) {
4779 unsigned long available
;
4780 unsigned long reclaimable
;
4781 unsigned long min_wmark
= min_wmark_pages(zone
);
4784 available
= reclaimable
= zone_reclaimable_pages(zone
);
4785 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4788 * Would the allocation succeed if we reclaimed all
4789 * reclaimable pages?
4791 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4792 ac
->highest_zoneidx
, alloc_flags
, available
);
4793 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4794 available
, min_wmark
, *no_progress_loops
, wmark
);
4797 * If we didn't make any progress and have a lot of
4798 * dirty + writeback pages then we should wait for
4799 * an IO to complete to slow down the reclaim and
4800 * prevent from pre mature OOM
4802 if (!did_some_progress
) {
4803 unsigned long write_pending
;
4805 write_pending
= zone_page_state_snapshot(zone
,
4806 NR_ZONE_WRITE_PENDING
);
4808 if (2 * write_pending
> reclaimable
) {
4809 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4821 * Memory allocation/reclaim might be called from a WQ context and the
4822 * current implementation of the WQ concurrency control doesn't
4823 * recognize that a particular WQ is congested if the worker thread is
4824 * looping without ever sleeping. Therefore we have to do a short sleep
4825 * here rather than calling cond_resched().
4827 if (current
->flags
& PF_WQ_WORKER
)
4828 schedule_timeout_uninterruptible(1);
4835 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4838 * It's possible that cpuset's mems_allowed and the nodemask from
4839 * mempolicy don't intersect. This should be normally dealt with by
4840 * policy_nodemask(), but it's possible to race with cpuset update in
4841 * such a way the check therein was true, and then it became false
4842 * before we got our cpuset_mems_cookie here.
4843 * This assumes that for all allocations, ac->nodemask can come only
4844 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4845 * when it does not intersect with the cpuset restrictions) or the
4846 * caller can deal with a violated nodemask.
4848 if (cpusets_enabled() && ac
->nodemask
&&
4849 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4850 ac
->nodemask
= NULL
;
4855 * When updating a task's mems_allowed or mempolicy nodemask, it is
4856 * possible to race with parallel threads in such a way that our
4857 * allocation can fail while the mask is being updated. If we are about
4858 * to fail, check if the cpuset changed during allocation and if so,
4861 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4867 static inline struct page
*
4868 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4869 struct alloc_context
*ac
)
4871 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4872 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4873 struct page
*page
= NULL
;
4874 unsigned int alloc_flags
;
4875 unsigned long did_some_progress
;
4876 enum compact_priority compact_priority
;
4877 enum compact_result compact_result
;
4878 int compaction_retries
;
4879 int no_progress_loops
;
4880 unsigned int cpuset_mems_cookie
;
4884 * We also sanity check to catch abuse of atomic reserves being used by
4885 * callers that are not in atomic context.
4887 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4888 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4889 gfp_mask
&= ~__GFP_ATOMIC
;
4892 compaction_retries
= 0;
4893 no_progress_loops
= 0;
4894 compact_priority
= DEF_COMPACT_PRIORITY
;
4895 cpuset_mems_cookie
= read_mems_allowed_begin();
4898 * The fast path uses conservative alloc_flags to succeed only until
4899 * kswapd needs to be woken up, and to avoid the cost of setting up
4900 * alloc_flags precisely. So we do that now.
4902 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4905 * We need to recalculate the starting point for the zonelist iterator
4906 * because we might have used different nodemask in the fast path, or
4907 * there was a cpuset modification and we are retrying - otherwise we
4908 * could end up iterating over non-eligible zones endlessly.
4910 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4911 ac
->highest_zoneidx
, ac
->nodemask
);
4912 if (!ac
->preferred_zoneref
->zone
)
4915 if (alloc_flags
& ALLOC_KSWAPD
)
4916 wake_all_kswapds(order
, gfp_mask
, ac
);
4919 * The adjusted alloc_flags might result in immediate success, so try
4922 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4927 * For costly allocations, try direct compaction first, as it's likely
4928 * that we have enough base pages and don't need to reclaim. For non-
4929 * movable high-order allocations, do that as well, as compaction will
4930 * try prevent permanent fragmentation by migrating from blocks of the
4932 * Don't try this for allocations that are allowed to ignore
4933 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4935 if (can_direct_reclaim
&&
4937 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4938 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4939 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4941 INIT_COMPACT_PRIORITY
,
4947 * Checks for costly allocations with __GFP_NORETRY, which
4948 * includes some THP page fault allocations
4950 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4952 * If allocating entire pageblock(s) and compaction
4953 * failed because all zones are below low watermarks
4954 * or is prohibited because it recently failed at this
4955 * order, fail immediately unless the allocator has
4956 * requested compaction and reclaim retry.
4959 * - potentially very expensive because zones are far
4960 * below their low watermarks or this is part of very
4961 * bursty high order allocations,
4962 * - not guaranteed to help because isolate_freepages()
4963 * may not iterate over freed pages as part of its
4965 * - unlikely to make entire pageblocks free on its
4968 if (compact_result
== COMPACT_SKIPPED
||
4969 compact_result
== COMPACT_DEFERRED
)
4973 * Looks like reclaim/compaction is worth trying, but
4974 * sync compaction could be very expensive, so keep
4975 * using async compaction.
4977 compact_priority
= INIT_COMPACT_PRIORITY
;
4982 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4983 if (alloc_flags
& ALLOC_KSWAPD
)
4984 wake_all_kswapds(order
, gfp_mask
, ac
);
4986 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4988 alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, reserve_flags
);
4991 * Reset the nodemask and zonelist iterators if memory policies can be
4992 * ignored. These allocations are high priority and system rather than
4995 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4996 ac
->nodemask
= NULL
;
4997 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4998 ac
->highest_zoneidx
, ac
->nodemask
);
5001 /* Attempt with potentially adjusted zonelist and alloc_flags */
5002 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
5006 /* Caller is not willing to reclaim, we can't balance anything */
5007 if (!can_direct_reclaim
)
5010 /* Avoid recursion of direct reclaim */
5011 if (current
->flags
& PF_MEMALLOC
)
5014 /* Try direct reclaim and then allocating */
5015 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
5016 &did_some_progress
);
5020 /* Try direct compaction and then allocating */
5021 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
5022 compact_priority
, &compact_result
);
5026 /* Do not loop if specifically requested */
5027 if (gfp_mask
& __GFP_NORETRY
)
5031 * Do not retry costly high order allocations unless they are
5032 * __GFP_RETRY_MAYFAIL
5034 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
5037 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
5038 did_some_progress
> 0, &no_progress_loops
))
5042 * It doesn't make any sense to retry for the compaction if the order-0
5043 * reclaim is not able to make any progress because the current
5044 * implementation of the compaction depends on the sufficient amount
5045 * of free memory (see __compaction_suitable)
5047 if (did_some_progress
> 0 &&
5048 should_compact_retry(ac
, order
, alloc_flags
,
5049 compact_result
, &compact_priority
,
5050 &compaction_retries
))
5054 /* Deal with possible cpuset update races before we start OOM killing */
5055 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
5058 /* Reclaim has failed us, start killing things */
5059 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
5063 /* Avoid allocations with no watermarks from looping endlessly */
5064 if (tsk_is_oom_victim(current
) &&
5065 (alloc_flags
& ALLOC_OOM
||
5066 (gfp_mask
& __GFP_NOMEMALLOC
)))
5069 /* Retry as long as the OOM killer is making progress */
5070 if (did_some_progress
) {
5071 no_progress_loops
= 0;
5076 /* Deal with possible cpuset update races before we fail */
5077 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
5081 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5084 if (gfp_mask
& __GFP_NOFAIL
) {
5086 * All existing users of the __GFP_NOFAIL are blockable, so warn
5087 * of any new users that actually require GFP_NOWAIT
5089 if (WARN_ON_ONCE(!can_direct_reclaim
))
5093 * PF_MEMALLOC request from this context is rather bizarre
5094 * because we cannot reclaim anything and only can loop waiting
5095 * for somebody to do a work for us
5097 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
5100 * non failing costly orders are a hard requirement which we
5101 * are not prepared for much so let's warn about these users
5102 * so that we can identify them and convert them to something
5105 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
5108 * Help non-failing allocations by giving them access to memory
5109 * reserves but do not use ALLOC_NO_WATERMARKS because this
5110 * could deplete whole memory reserves which would just make
5111 * the situation worse
5113 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
5121 warn_alloc(gfp_mask
, ac
->nodemask
,
5122 "page allocation failure: order:%u", order
);
5127 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
5128 int preferred_nid
, nodemask_t
*nodemask
,
5129 struct alloc_context
*ac
, gfp_t
*alloc_gfp
,
5130 unsigned int *alloc_flags
)
5132 ac
->highest_zoneidx
= gfp_zone(gfp_mask
);
5133 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
5134 ac
->nodemask
= nodemask
;
5135 ac
->migratetype
= gfp_migratetype(gfp_mask
);
5137 if (cpusets_enabled()) {
5138 *alloc_gfp
|= __GFP_HARDWALL
;
5140 * When we are in the interrupt context, it is irrelevant
5141 * to the current task context. It means that any node ok.
5143 if (!in_interrupt() && !ac
->nodemask
)
5144 ac
->nodemask
= &cpuset_current_mems_allowed
;
5146 *alloc_flags
|= ALLOC_CPUSET
;
5149 fs_reclaim_acquire(gfp_mask
);
5150 fs_reclaim_release(gfp_mask
);
5152 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
5154 if (should_fail_alloc_page(gfp_mask
, order
))
5157 *alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, *alloc_flags
);
5159 /* Dirty zone balancing only done in the fast path */
5160 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
5163 * The preferred zone is used for statistics but crucially it is
5164 * also used as the starting point for the zonelist iterator. It
5165 * may get reset for allocations that ignore memory policies.
5167 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
5168 ac
->highest_zoneidx
, ac
->nodemask
);
5174 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5175 * @gfp: GFP flags for the allocation
5176 * @preferred_nid: The preferred NUMA node ID to allocate from
5177 * @nodemask: Set of nodes to allocate from, may be NULL
5178 * @nr_pages: The number of pages desired on the list or array
5179 * @page_list: Optional list to store the allocated pages
5180 * @page_array: Optional array to store the pages
5182 * This is a batched version of the page allocator that attempts to
5183 * allocate nr_pages quickly. Pages are added to page_list if page_list
5184 * is not NULL, otherwise it is assumed that the page_array is valid.
5186 * For lists, nr_pages is the number of pages that should be allocated.
5188 * For arrays, only NULL elements are populated with pages and nr_pages
5189 * is the maximum number of pages that will be stored in the array.
5191 * Returns the number of pages on the list or array.
5193 unsigned long __alloc_pages_bulk(gfp_t gfp
, int preferred_nid
,
5194 nodemask_t
*nodemask
, int nr_pages
,
5195 struct list_head
*page_list
,
5196 struct page
**page_array
)
5199 unsigned long flags
;
5202 struct per_cpu_pages
*pcp
;
5203 struct list_head
*pcp_list
;
5204 struct alloc_context ac
;
5206 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
5207 int nr_populated
= 0, nr_account
= 0;
5210 * Skip populated array elements to determine if any pages need
5211 * to be allocated before disabling IRQs.
5213 while (page_array
&& nr_populated
< nr_pages
&& page_array
[nr_populated
])
5216 /* No pages requested? */
5217 if (unlikely(nr_pages
<= 0))
5220 /* Already populated array? */
5221 if (unlikely(page_array
&& nr_pages
- nr_populated
== 0))
5224 /* Use the single page allocator for one page. */
5225 if (nr_pages
- nr_populated
== 1)
5228 #ifdef CONFIG_PAGE_OWNER
5230 * PAGE_OWNER may recurse into the allocator to allocate space to
5231 * save the stack with pagesets.lock held. Releasing/reacquiring
5232 * removes much of the performance benefit of bulk allocation so
5233 * force the caller to allocate one page at a time as it'll have
5234 * similar performance to added complexity to the bulk allocator.
5236 if (static_branch_unlikely(&page_owner_inited
))
5240 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5241 gfp
&= gfp_allowed_mask
;
5243 if (!prepare_alloc_pages(gfp
, 0, preferred_nid
, nodemask
, &ac
, &alloc_gfp
, &alloc_flags
))
5247 /* Find an allowed local zone that meets the low watermark. */
5248 for_each_zone_zonelist_nodemask(zone
, z
, ac
.zonelist
, ac
.highest_zoneidx
, ac
.nodemask
) {
5251 if (cpusets_enabled() && (alloc_flags
& ALLOC_CPUSET
) &&
5252 !__cpuset_zone_allowed(zone
, gfp
)) {
5256 if (nr_online_nodes
> 1 && zone
!= ac
.preferred_zoneref
->zone
&&
5257 zone_to_nid(zone
) != zone_to_nid(ac
.preferred_zoneref
->zone
)) {
5261 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
) + nr_pages
;
5262 if (zone_watermark_fast(zone
, 0, mark
,
5263 zonelist_zone_idx(ac
.preferred_zoneref
),
5264 alloc_flags
, gfp
)) {
5270 * If there are no allowed local zones that meets the watermarks then
5271 * try to allocate a single page and reclaim if necessary.
5273 if (unlikely(!zone
))
5276 /* Attempt the batch allocation */
5277 local_lock_irqsave(&pagesets
.lock
, flags
);
5278 pcp
= this_cpu_ptr(zone
->per_cpu_pageset
);
5279 pcp_list
= &pcp
->lists
[order_to_pindex(ac
.migratetype
, 0)];
5281 while (nr_populated
< nr_pages
) {
5283 /* Skip existing pages */
5284 if (page_array
&& page_array
[nr_populated
]) {
5289 page
= __rmqueue_pcplist(zone
, 0, ac
.migratetype
, alloc_flags
,
5291 if (unlikely(!page
)) {
5292 /* Try and get at least one page */
5299 prep_new_page(page
, 0, gfp
, 0);
5301 list_add(&page
->lru
, page_list
);
5303 page_array
[nr_populated
] = page
;
5307 local_unlock_irqrestore(&pagesets
.lock
, flags
);
5309 __count_zid_vm_events(PGALLOC
, zone_idx(zone
), nr_account
);
5310 zone_statistics(ac
.preferred_zoneref
->zone
, zone
, nr_account
);
5313 return nr_populated
;
5316 local_unlock_irqrestore(&pagesets
.lock
, flags
);
5319 page
= __alloc_pages(gfp
, 0, preferred_nid
, nodemask
);
5322 list_add(&page
->lru
, page_list
);
5324 page_array
[nr_populated
] = page
;
5330 EXPORT_SYMBOL_GPL(__alloc_pages_bulk
);
5333 * This is the 'heart' of the zoned buddy allocator.
5335 struct page
*__alloc_pages(gfp_t gfp
, unsigned int order
, int preferred_nid
,
5336 nodemask_t
*nodemask
)
5339 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
5340 gfp_t alloc_gfp
; /* The gfp_t that was actually used for allocation */
5341 struct alloc_context ac
= { };
5344 * There are several places where we assume that the order value is sane
5345 * so bail out early if the request is out of bound.
5347 if (unlikely(order
>= MAX_ORDER
)) {
5348 WARN_ON_ONCE(!(gfp
& __GFP_NOWARN
));
5352 gfp
&= gfp_allowed_mask
;
5354 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5355 * resp. GFP_NOIO which has to be inherited for all allocation requests
5356 * from a particular context which has been marked by
5357 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5358 * movable zones are not used during allocation.
5360 gfp
= current_gfp_context(gfp
);
5362 if (!prepare_alloc_pages(gfp
, order
, preferred_nid
, nodemask
, &ac
,
5363 &alloc_gfp
, &alloc_flags
))
5367 * Forbid the first pass from falling back to types that fragment
5368 * memory until all local zones are considered.
5370 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp
);
5372 /* First allocation attempt */
5373 page
= get_page_from_freelist(alloc_gfp
, order
, alloc_flags
, &ac
);
5378 ac
.spread_dirty_pages
= false;
5381 * Restore the original nodemask if it was potentially replaced with
5382 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5384 ac
.nodemask
= nodemask
;
5386 page
= __alloc_pages_slowpath(alloc_gfp
, order
, &ac
);
5389 if (memcg_kmem_enabled() && (gfp
& __GFP_ACCOUNT
) && page
&&
5390 unlikely(__memcg_kmem_charge_page(page
, gfp
, order
) != 0)) {
5391 __free_pages(page
, order
);
5395 trace_mm_page_alloc(page
, order
, alloc_gfp
, ac
.migratetype
);
5399 EXPORT_SYMBOL(__alloc_pages
);
5402 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5403 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5404 * you need to access high mem.
5406 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
5410 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
5413 return (unsigned long) page_address(page
);
5415 EXPORT_SYMBOL(__get_free_pages
);
5417 unsigned long get_zeroed_page(gfp_t gfp_mask
)
5419 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
5421 EXPORT_SYMBOL(get_zeroed_page
);
5424 * __free_pages - Free pages allocated with alloc_pages().
5425 * @page: The page pointer returned from alloc_pages().
5426 * @order: The order of the allocation.
5428 * This function can free multi-page allocations that are not compound
5429 * pages. It does not check that the @order passed in matches that of
5430 * the allocation, so it is easy to leak memory. Freeing more memory
5431 * than was allocated will probably emit a warning.
5433 * If the last reference to this page is speculative, it will be released
5434 * by put_page() which only frees the first page of a non-compound
5435 * allocation. To prevent the remaining pages from being leaked, we free
5436 * the subsequent pages here. If you want to use the page's reference
5437 * count to decide when to free the allocation, you should allocate a
5438 * compound page, and use put_page() instead of __free_pages().
5440 * Context: May be called in interrupt context or while holding a normal
5441 * spinlock, but not in NMI context or while holding a raw spinlock.
5443 void __free_pages(struct page
*page
, unsigned int order
)
5445 if (put_page_testzero(page
))
5446 free_the_page(page
, order
);
5447 else if (!PageHead(page
))
5449 free_the_page(page
+ (1 << order
), order
);
5451 EXPORT_SYMBOL(__free_pages
);
5453 void free_pages(unsigned long addr
, unsigned int order
)
5456 VM_BUG_ON(!virt_addr_valid((void *)addr
));
5457 __free_pages(virt_to_page((void *)addr
), order
);
5461 EXPORT_SYMBOL(free_pages
);
5465 * An arbitrary-length arbitrary-offset area of memory which resides
5466 * within a 0 or higher order page. Multiple fragments within that page
5467 * are individually refcounted, in the page's reference counter.
5469 * The page_frag functions below provide a simple allocation framework for
5470 * page fragments. This is used by the network stack and network device
5471 * drivers to provide a backing region of memory for use as either an
5472 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5474 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
5477 struct page
*page
= NULL
;
5478 gfp_t gfp
= gfp_mask
;
5480 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5481 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
5483 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
5484 PAGE_FRAG_CACHE_MAX_ORDER
);
5485 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
5487 if (unlikely(!page
))
5488 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
5490 nc
->va
= page
? page_address(page
) : NULL
;
5495 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
5497 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
5499 if (page_ref_sub_and_test(page
, count
))
5500 free_the_page(page
, compound_order(page
));
5502 EXPORT_SYMBOL(__page_frag_cache_drain
);
5504 void *page_frag_alloc_align(struct page_frag_cache
*nc
,
5505 unsigned int fragsz
, gfp_t gfp_mask
,
5506 unsigned int align_mask
)
5508 unsigned int size
= PAGE_SIZE
;
5512 if (unlikely(!nc
->va
)) {
5514 page
= __page_frag_cache_refill(nc
, gfp_mask
);
5518 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5519 /* if size can vary use size else just use PAGE_SIZE */
5522 /* Even if we own the page, we do not use atomic_set().
5523 * This would break get_page_unless_zero() users.
5525 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
5527 /* reset page count bias and offset to start of new frag */
5528 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
5529 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
5533 offset
= nc
->offset
- fragsz
;
5534 if (unlikely(offset
< 0)) {
5535 page
= virt_to_page(nc
->va
);
5537 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
5540 if (unlikely(nc
->pfmemalloc
)) {
5541 free_the_page(page
, compound_order(page
));
5545 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5546 /* if size can vary use size else just use PAGE_SIZE */
5549 /* OK, page count is 0, we can safely set it */
5550 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
5552 /* reset page count bias and offset to start of new frag */
5553 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
5554 offset
= size
- fragsz
;
5558 offset
&= align_mask
;
5559 nc
->offset
= offset
;
5561 return nc
->va
+ offset
;
5563 EXPORT_SYMBOL(page_frag_alloc_align
);
5566 * Frees a page fragment allocated out of either a compound or order 0 page.
5568 void page_frag_free(void *addr
)
5570 struct page
*page
= virt_to_head_page(addr
);
5572 if (unlikely(put_page_testzero(page
)))
5573 free_the_page(page
, compound_order(page
));
5575 EXPORT_SYMBOL(page_frag_free
);
5577 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
5581 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
5582 unsigned long used
= addr
+ PAGE_ALIGN(size
);
5584 split_page(virt_to_page((void *)addr
), order
);
5585 while (used
< alloc_end
) {
5590 return (void *)addr
;
5594 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5595 * @size: the number of bytes to allocate
5596 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5598 * This function is similar to alloc_pages(), except that it allocates the
5599 * minimum number of pages to satisfy the request. alloc_pages() can only
5600 * allocate memory in power-of-two pages.
5602 * This function is also limited by MAX_ORDER.
5604 * Memory allocated by this function must be released by free_pages_exact().
5606 * Return: pointer to the allocated area or %NULL in case of error.
5608 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
5610 unsigned int order
= get_order(size
);
5613 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5614 gfp_mask
&= ~__GFP_COMP
;
5616 addr
= __get_free_pages(gfp_mask
, order
);
5617 return make_alloc_exact(addr
, order
, size
);
5619 EXPORT_SYMBOL(alloc_pages_exact
);
5622 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5624 * @nid: the preferred node ID where memory should be allocated
5625 * @size: the number of bytes to allocate
5626 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5628 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5631 * Return: pointer to the allocated area or %NULL in case of error.
5633 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
5635 unsigned int order
= get_order(size
);
5638 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5639 gfp_mask
&= ~__GFP_COMP
;
5641 p
= alloc_pages_node(nid
, gfp_mask
, order
);
5644 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5648 * free_pages_exact - release memory allocated via alloc_pages_exact()
5649 * @virt: the value returned by alloc_pages_exact.
5650 * @size: size of allocation, same value as passed to alloc_pages_exact().
5652 * Release the memory allocated by a previous call to alloc_pages_exact.
5654 void free_pages_exact(void *virt
, size_t size
)
5656 unsigned long addr
= (unsigned long)virt
;
5657 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5659 while (addr
< end
) {
5664 EXPORT_SYMBOL(free_pages_exact
);
5667 * nr_free_zone_pages - count number of pages beyond high watermark
5668 * @offset: The zone index of the highest zone
5670 * nr_free_zone_pages() counts the number of pages which are beyond the
5671 * high watermark within all zones at or below a given zone index. For each
5672 * zone, the number of pages is calculated as:
5674 * nr_free_zone_pages = managed_pages - high_pages
5676 * Return: number of pages beyond high watermark.
5678 static unsigned long nr_free_zone_pages(int offset
)
5683 /* Just pick one node, since fallback list is circular */
5684 unsigned long sum
= 0;
5686 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5688 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5689 unsigned long size
= zone_managed_pages(zone
);
5690 unsigned long high
= high_wmark_pages(zone
);
5699 * nr_free_buffer_pages - count number of pages beyond high watermark
5701 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5702 * watermark within ZONE_DMA and ZONE_NORMAL.
5704 * Return: number of pages beyond high watermark within ZONE_DMA and
5707 unsigned long nr_free_buffer_pages(void)
5709 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5711 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5713 static inline void show_node(struct zone
*zone
)
5715 if (IS_ENABLED(CONFIG_NUMA
))
5716 printk("Node %d ", zone_to_nid(zone
));
5719 long si_mem_available(void)
5722 unsigned long pagecache
;
5723 unsigned long wmark_low
= 0;
5724 unsigned long pages
[NR_LRU_LISTS
];
5725 unsigned long reclaimable
;
5729 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5730 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5733 wmark_low
+= low_wmark_pages(zone
);
5736 * Estimate the amount of memory available for userspace allocations,
5737 * without causing swapping.
5739 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5742 * Not all the page cache can be freed, otherwise the system will
5743 * start swapping. Assume at least half of the page cache, or the
5744 * low watermark worth of cache, needs to stay.
5746 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5747 pagecache
-= min(pagecache
/ 2, wmark_low
);
5748 available
+= pagecache
;
5751 * Part of the reclaimable slab and other kernel memory consists of
5752 * items that are in use, and cannot be freed. Cap this estimate at the
5755 reclaimable
= global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
) +
5756 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5757 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5763 EXPORT_SYMBOL_GPL(si_mem_available
);
5765 void si_meminfo(struct sysinfo
*val
)
5767 val
->totalram
= totalram_pages();
5768 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5769 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5770 val
->bufferram
= nr_blockdev_pages();
5771 val
->totalhigh
= totalhigh_pages();
5772 val
->freehigh
= nr_free_highpages();
5773 val
->mem_unit
= PAGE_SIZE
;
5776 EXPORT_SYMBOL(si_meminfo
);
5779 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5781 int zone_type
; /* needs to be signed */
5782 unsigned long managed_pages
= 0;
5783 unsigned long managed_highpages
= 0;
5784 unsigned long free_highpages
= 0;
5785 pg_data_t
*pgdat
= NODE_DATA(nid
);
5787 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5788 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5789 val
->totalram
= managed_pages
;
5790 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5791 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5792 #ifdef CONFIG_HIGHMEM
5793 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5794 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5796 if (is_highmem(zone
)) {
5797 managed_highpages
+= zone_managed_pages(zone
);
5798 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5801 val
->totalhigh
= managed_highpages
;
5802 val
->freehigh
= free_highpages
;
5804 val
->totalhigh
= managed_highpages
;
5805 val
->freehigh
= free_highpages
;
5807 val
->mem_unit
= PAGE_SIZE
;
5812 * Determine whether the node should be displayed or not, depending on whether
5813 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5815 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5817 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5821 * no node mask - aka implicit memory numa policy. Do not bother with
5822 * the synchronization - read_mems_allowed_begin - because we do not
5823 * have to be precise here.
5826 nodemask
= &cpuset_current_mems_allowed
;
5828 return !node_isset(nid
, *nodemask
);
5831 #define K(x) ((x) << (PAGE_SHIFT-10))
5833 static void show_migration_types(unsigned char type
)
5835 static const char types
[MIGRATE_TYPES
] = {
5836 [MIGRATE_UNMOVABLE
] = 'U',
5837 [MIGRATE_MOVABLE
] = 'M',
5838 [MIGRATE_RECLAIMABLE
] = 'E',
5839 [MIGRATE_HIGHATOMIC
] = 'H',
5841 [MIGRATE_CMA
] = 'C',
5843 #ifdef CONFIG_MEMORY_ISOLATION
5844 [MIGRATE_ISOLATE
] = 'I',
5847 char tmp
[MIGRATE_TYPES
+ 1];
5851 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5852 if (type
& (1 << i
))
5857 printk(KERN_CONT
"(%s) ", tmp
);
5861 * Show free area list (used inside shift_scroll-lock stuff)
5862 * We also calculate the percentage fragmentation. We do this by counting the
5863 * memory on each free list with the exception of the first item on the list.
5866 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5869 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5871 unsigned long free_pcp
= 0;
5876 for_each_populated_zone(zone
) {
5877 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5880 for_each_online_cpu(cpu
)
5881 free_pcp
+= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
)->count
;
5884 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5885 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5886 " unevictable:%lu dirty:%lu writeback:%lu\n"
5887 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5888 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5889 " free:%lu free_pcp:%lu free_cma:%lu\n",
5890 global_node_page_state(NR_ACTIVE_ANON
),
5891 global_node_page_state(NR_INACTIVE_ANON
),
5892 global_node_page_state(NR_ISOLATED_ANON
),
5893 global_node_page_state(NR_ACTIVE_FILE
),
5894 global_node_page_state(NR_INACTIVE_FILE
),
5895 global_node_page_state(NR_ISOLATED_FILE
),
5896 global_node_page_state(NR_UNEVICTABLE
),
5897 global_node_page_state(NR_FILE_DIRTY
),
5898 global_node_page_state(NR_WRITEBACK
),
5899 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
),
5900 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B
),
5901 global_node_page_state(NR_FILE_MAPPED
),
5902 global_node_page_state(NR_SHMEM
),
5903 global_node_page_state(NR_PAGETABLE
),
5904 global_zone_page_state(NR_BOUNCE
),
5905 global_zone_page_state(NR_FREE_PAGES
),
5907 global_zone_page_state(NR_FREE_CMA_PAGES
));
5909 for_each_online_pgdat(pgdat
) {
5910 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5914 " active_anon:%lukB"
5915 " inactive_anon:%lukB"
5916 " active_file:%lukB"
5917 " inactive_file:%lukB"
5918 " unevictable:%lukB"
5919 " isolated(anon):%lukB"
5920 " isolated(file):%lukB"
5925 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5927 " shmem_pmdmapped: %lukB"
5930 " writeback_tmp:%lukB"
5931 " kernel_stack:%lukB"
5932 #ifdef CONFIG_SHADOW_CALL_STACK
5933 " shadow_call_stack:%lukB"
5936 " all_unreclaimable? %s"
5939 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5940 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5941 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5942 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5943 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5944 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5945 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5946 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5947 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5948 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5949 K(node_page_state(pgdat
, NR_SHMEM
)),
5950 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5951 K(node_page_state(pgdat
, NR_SHMEM_THPS
)),
5952 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)),
5953 K(node_page_state(pgdat
, NR_ANON_THPS
)),
5955 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5956 node_page_state(pgdat
, NR_KERNEL_STACK_KB
),
5957 #ifdef CONFIG_SHADOW_CALL_STACK
5958 node_page_state(pgdat
, NR_KERNEL_SCS_KB
),
5960 K(node_page_state(pgdat
, NR_PAGETABLE
)),
5961 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5965 for_each_populated_zone(zone
) {
5968 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5972 for_each_online_cpu(cpu
)
5973 free_pcp
+= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
)->count
;
5982 " reserved_highatomic:%luKB"
5983 " active_anon:%lukB"
5984 " inactive_anon:%lukB"
5985 " active_file:%lukB"
5986 " inactive_file:%lukB"
5987 " unevictable:%lukB"
5988 " writepending:%lukB"
5998 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5999 K(min_wmark_pages(zone
)),
6000 K(low_wmark_pages(zone
)),
6001 K(high_wmark_pages(zone
)),
6002 K(zone
->nr_reserved_highatomic
),
6003 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
6004 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
6005 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
6006 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
6007 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
6008 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
6009 K(zone
->present_pages
),
6010 K(zone_managed_pages(zone
)),
6011 K(zone_page_state(zone
, NR_MLOCK
)),
6012 K(zone_page_state(zone
, NR_BOUNCE
)),
6014 K(this_cpu_read(zone
->per_cpu_pageset
->count
)),
6015 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
6016 printk("lowmem_reserve[]:");
6017 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
6018 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
6019 printk(KERN_CONT
"\n");
6022 for_each_populated_zone(zone
) {
6024 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
6025 unsigned char types
[MAX_ORDER
];
6027 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
6030 printk(KERN_CONT
"%s: ", zone
->name
);
6032 spin_lock_irqsave(&zone
->lock
, flags
);
6033 for (order
= 0; order
< MAX_ORDER
; order
++) {
6034 struct free_area
*area
= &zone
->free_area
[order
];
6037 nr
[order
] = area
->nr_free
;
6038 total
+= nr
[order
] << order
;
6041 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
6042 if (!free_area_empty(area
, type
))
6043 types
[order
] |= 1 << type
;
6046 spin_unlock_irqrestore(&zone
->lock
, flags
);
6047 for (order
= 0; order
< MAX_ORDER
; order
++) {
6048 printk(KERN_CONT
"%lu*%lukB ",
6049 nr
[order
], K(1UL) << order
);
6051 show_migration_types(types
[order
]);
6053 printk(KERN_CONT
"= %lukB\n", K(total
));
6056 hugetlb_show_meminfo();
6058 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
6060 show_swap_cache_info();
6063 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
6065 zoneref
->zone
= zone
;
6066 zoneref
->zone_idx
= zone_idx(zone
);
6070 * Builds allocation fallback zone lists.
6072 * Add all populated zones of a node to the zonelist.
6074 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
6077 enum zone_type zone_type
= MAX_NR_ZONES
;
6082 zone
= pgdat
->node_zones
+ zone_type
;
6083 if (managed_zone(zone
)) {
6084 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
6085 check_highest_zone(zone_type
);
6087 } while (zone_type
);
6094 static int __parse_numa_zonelist_order(char *s
)
6097 * We used to support different zonelists modes but they turned
6098 * out to be just not useful. Let's keep the warning in place
6099 * if somebody still use the cmd line parameter so that we do
6100 * not fail it silently
6102 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
6103 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
6109 char numa_zonelist_order
[] = "Node";
6112 * sysctl handler for numa_zonelist_order
6114 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
6115 void *buffer
, size_t *length
, loff_t
*ppos
)
6118 return __parse_numa_zonelist_order(buffer
);
6119 return proc_dostring(table
, write
, buffer
, length
, ppos
);
6123 #define MAX_NODE_LOAD (nr_online_nodes)
6124 static int node_load
[MAX_NUMNODES
];
6127 * find_next_best_node - find the next node that should appear in a given node's fallback list
6128 * @node: node whose fallback list we're appending
6129 * @used_node_mask: nodemask_t of already used nodes
6131 * We use a number of factors to determine which is the next node that should
6132 * appear on a given node's fallback list. The node should not have appeared
6133 * already in @node's fallback list, and it should be the next closest node
6134 * according to the distance array (which contains arbitrary distance values
6135 * from each node to each node in the system), and should also prefer nodes
6136 * with no CPUs, since presumably they'll have very little allocation pressure
6137 * on them otherwise.
6139 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6141 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
6144 int min_val
= INT_MAX
;
6145 int best_node
= NUMA_NO_NODE
;
6147 /* Use the local node if we haven't already */
6148 if (!node_isset(node
, *used_node_mask
)) {
6149 node_set(node
, *used_node_mask
);
6153 for_each_node_state(n
, N_MEMORY
) {
6155 /* Don't want a node to appear more than once */
6156 if (node_isset(n
, *used_node_mask
))
6159 /* Use the distance array to find the distance */
6160 val
= node_distance(node
, n
);
6162 /* Penalize nodes under us ("prefer the next node") */
6165 /* Give preference to headless and unused nodes */
6166 if (!cpumask_empty(cpumask_of_node(n
)))
6167 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
6169 /* Slight preference for less loaded node */
6170 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
6171 val
+= node_load
[n
];
6173 if (val
< min_val
) {
6180 node_set(best_node
, *used_node_mask
);
6187 * Build zonelists ordered by node and zones within node.
6188 * This results in maximum locality--normal zone overflows into local
6189 * DMA zone, if any--but risks exhausting DMA zone.
6191 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
6194 struct zoneref
*zonerefs
;
6197 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
6199 for (i
= 0; i
< nr_nodes
; i
++) {
6202 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
6204 nr_zones
= build_zonerefs_node(node
, zonerefs
);
6205 zonerefs
+= nr_zones
;
6207 zonerefs
->zone
= NULL
;
6208 zonerefs
->zone_idx
= 0;
6212 * Build gfp_thisnode zonelists
6214 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
6216 struct zoneref
*zonerefs
;
6219 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
6220 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
6221 zonerefs
+= nr_zones
;
6222 zonerefs
->zone
= NULL
;
6223 zonerefs
->zone_idx
= 0;
6227 * Build zonelists ordered by zone and nodes within zones.
6228 * This results in conserving DMA zone[s] until all Normal memory is
6229 * exhausted, but results in overflowing to remote node while memory
6230 * may still exist in local DMA zone.
6233 static void build_zonelists(pg_data_t
*pgdat
)
6235 static int node_order
[MAX_NUMNODES
];
6236 int node
, load
, nr_nodes
= 0;
6237 nodemask_t used_mask
= NODE_MASK_NONE
;
6238 int local_node
, prev_node
;
6240 /* NUMA-aware ordering of nodes */
6241 local_node
= pgdat
->node_id
;
6242 load
= nr_online_nodes
;
6243 prev_node
= local_node
;
6245 memset(node_order
, 0, sizeof(node_order
));
6246 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
6248 * We don't want to pressure a particular node.
6249 * So adding penalty to the first node in same
6250 * distance group to make it round-robin.
6252 if (node_distance(local_node
, node
) !=
6253 node_distance(local_node
, prev_node
))
6254 node_load
[node
] = load
;
6256 node_order
[nr_nodes
++] = node
;
6261 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
6262 build_thisnode_zonelists(pgdat
);
6265 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6267 * Return node id of node used for "local" allocations.
6268 * I.e., first node id of first zone in arg node's generic zonelist.
6269 * Used for initializing percpu 'numa_mem', which is used primarily
6270 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6272 int local_memory_node(int node
)
6276 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
6277 gfp_zone(GFP_KERNEL
),
6279 return zone_to_nid(z
->zone
);
6283 static void setup_min_unmapped_ratio(void);
6284 static void setup_min_slab_ratio(void);
6285 #else /* CONFIG_NUMA */
6287 static void build_zonelists(pg_data_t
*pgdat
)
6289 int node
, local_node
;
6290 struct zoneref
*zonerefs
;
6293 local_node
= pgdat
->node_id
;
6295 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
6296 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
6297 zonerefs
+= nr_zones
;
6300 * Now we build the zonelist so that it contains the zones
6301 * of all the other nodes.
6302 * We don't want to pressure a particular node, so when
6303 * building the zones for node N, we make sure that the
6304 * zones coming right after the local ones are those from
6305 * node N+1 (modulo N)
6307 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
6308 if (!node_online(node
))
6310 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
6311 zonerefs
+= nr_zones
;
6313 for (node
= 0; node
< local_node
; node
++) {
6314 if (!node_online(node
))
6316 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
6317 zonerefs
+= nr_zones
;
6320 zonerefs
->zone
= NULL
;
6321 zonerefs
->zone_idx
= 0;
6324 #endif /* CONFIG_NUMA */
6327 * Boot pageset table. One per cpu which is going to be used for all
6328 * zones and all nodes. The parameters will be set in such a way
6329 * that an item put on a list will immediately be handed over to
6330 * the buddy list. This is safe since pageset manipulation is done
6331 * with interrupts disabled.
6333 * The boot_pagesets must be kept even after bootup is complete for
6334 * unused processors and/or zones. They do play a role for bootstrapping
6335 * hotplugged processors.
6337 * zoneinfo_show() and maybe other functions do
6338 * not check if the processor is online before following the pageset pointer.
6339 * Other parts of the kernel may not check if the zone is available.
6341 static void per_cpu_pages_init(struct per_cpu_pages
*pcp
, struct per_cpu_zonestat
*pzstats
);
6342 /* These effectively disable the pcplists in the boot pageset completely */
6343 #define BOOT_PAGESET_HIGH 0
6344 #define BOOT_PAGESET_BATCH 1
6345 static DEFINE_PER_CPU(struct per_cpu_pages
, boot_pageset
);
6346 static DEFINE_PER_CPU(struct per_cpu_zonestat
, boot_zonestats
);
6347 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
6349 static void __build_all_zonelists(void *data
)
6352 int __maybe_unused cpu
;
6353 pg_data_t
*self
= data
;
6354 static DEFINE_SPINLOCK(lock
);
6359 memset(node_load
, 0, sizeof(node_load
));
6363 * This node is hotadded and no memory is yet present. So just
6364 * building zonelists is fine - no need to touch other nodes.
6366 if (self
&& !node_online(self
->node_id
)) {
6367 build_zonelists(self
);
6369 for_each_online_node(nid
) {
6370 pg_data_t
*pgdat
= NODE_DATA(nid
);
6372 build_zonelists(pgdat
);
6375 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6377 * We now know the "local memory node" for each node--
6378 * i.e., the node of the first zone in the generic zonelist.
6379 * Set up numa_mem percpu variable for on-line cpus. During
6380 * boot, only the boot cpu should be on-line; we'll init the
6381 * secondary cpus' numa_mem as they come on-line. During
6382 * node/memory hotplug, we'll fixup all on-line cpus.
6384 for_each_online_cpu(cpu
)
6385 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
6392 static noinline
void __init
6393 build_all_zonelists_init(void)
6397 __build_all_zonelists(NULL
);
6400 * Initialize the boot_pagesets that are going to be used
6401 * for bootstrapping processors. The real pagesets for
6402 * each zone will be allocated later when the per cpu
6403 * allocator is available.
6405 * boot_pagesets are used also for bootstrapping offline
6406 * cpus if the system is already booted because the pagesets
6407 * are needed to initialize allocators on a specific cpu too.
6408 * F.e. the percpu allocator needs the page allocator which
6409 * needs the percpu allocator in order to allocate its pagesets
6410 * (a chicken-egg dilemma).
6412 for_each_possible_cpu(cpu
)
6413 per_cpu_pages_init(&per_cpu(boot_pageset
, cpu
), &per_cpu(boot_zonestats
, cpu
));
6415 mminit_verify_zonelist();
6416 cpuset_init_current_mems_allowed();
6420 * unless system_state == SYSTEM_BOOTING.
6422 * __ref due to call of __init annotated helper build_all_zonelists_init
6423 * [protected by SYSTEM_BOOTING].
6425 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
6427 unsigned long vm_total_pages
;
6429 if (system_state
== SYSTEM_BOOTING
) {
6430 build_all_zonelists_init();
6432 __build_all_zonelists(pgdat
);
6433 /* cpuset refresh routine should be here */
6435 /* Get the number of free pages beyond high watermark in all zones. */
6436 vm_total_pages
= nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
6438 * Disable grouping by mobility if the number of pages in the
6439 * system is too low to allow the mechanism to work. It would be
6440 * more accurate, but expensive to check per-zone. This check is
6441 * made on memory-hotadd so a system can start with mobility
6442 * disabled and enable it later
6444 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
6445 page_group_by_mobility_disabled
= 1;
6447 page_group_by_mobility_disabled
= 0;
6449 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6451 page_group_by_mobility_disabled
? "off" : "on",
6454 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
6458 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6459 static bool __meminit
6460 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
6462 static struct memblock_region
*r
;
6464 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
6465 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
6466 for_each_mem_region(r
) {
6467 if (*pfn
< memblock_region_memory_end_pfn(r
))
6471 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
6472 memblock_is_mirror(r
)) {
6473 *pfn
= memblock_region_memory_end_pfn(r
);
6481 * Initially all pages are reserved - free ones are freed
6482 * up by memblock_free_all() once the early boot process is
6483 * done. Non-atomic initialization, single-pass.
6485 * All aligned pageblocks are initialized to the specified migratetype
6486 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6487 * zone stats (e.g., nr_isolate_pageblock) are touched.
6489 void __meminit
memmap_init_range(unsigned long size
, int nid
, unsigned long zone
,
6490 unsigned long start_pfn
, unsigned long zone_end_pfn
,
6491 enum meminit_context context
,
6492 struct vmem_altmap
*altmap
, int migratetype
)
6494 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
6497 if (highest_memmap_pfn
< end_pfn
- 1)
6498 highest_memmap_pfn
= end_pfn
- 1;
6500 #ifdef CONFIG_ZONE_DEVICE
6502 * Honor reservation requested by the driver for this ZONE_DEVICE
6503 * memory. We limit the total number of pages to initialize to just
6504 * those that might contain the memory mapping. We will defer the
6505 * ZONE_DEVICE page initialization until after we have released
6508 if (zone
== ZONE_DEVICE
) {
6512 if (start_pfn
== altmap
->base_pfn
)
6513 start_pfn
+= altmap
->reserve
;
6514 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
6518 for (pfn
= start_pfn
; pfn
< end_pfn
; ) {
6520 * There can be holes in boot-time mem_map[]s handed to this
6521 * function. They do not exist on hotplugged memory.
6523 if (context
== MEMINIT_EARLY
) {
6524 if (overlap_memmap_init(zone
, &pfn
))
6526 if (defer_init(nid
, pfn
, zone_end_pfn
))
6530 page
= pfn_to_page(pfn
);
6531 __init_single_page(page
, pfn
, zone
, nid
);
6532 if (context
== MEMINIT_HOTPLUG
)
6533 __SetPageReserved(page
);
6536 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6537 * such that unmovable allocations won't be scattered all
6538 * over the place during system boot.
6540 if (IS_ALIGNED(pfn
, pageblock_nr_pages
)) {
6541 set_pageblock_migratetype(page
, migratetype
);
6548 #ifdef CONFIG_ZONE_DEVICE
6549 void __ref
memmap_init_zone_device(struct zone
*zone
,
6550 unsigned long start_pfn
,
6551 unsigned long nr_pages
,
6552 struct dev_pagemap
*pgmap
)
6554 unsigned long pfn
, end_pfn
= start_pfn
+ nr_pages
;
6555 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6556 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
6557 unsigned long zone_idx
= zone_idx(zone
);
6558 unsigned long start
= jiffies
;
6559 int nid
= pgdat
->node_id
;
6561 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
6565 * The call to memmap_init should have already taken care
6566 * of the pages reserved for the memmap, so we can just jump to
6567 * the end of that region and start processing the device pages.
6570 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
6571 nr_pages
= end_pfn
- start_pfn
;
6574 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
6575 struct page
*page
= pfn_to_page(pfn
);
6577 __init_single_page(page
, pfn
, zone_idx
, nid
);
6580 * Mark page reserved as it will need to wait for onlining
6581 * phase for it to be fully associated with a zone.
6583 * We can use the non-atomic __set_bit operation for setting
6584 * the flag as we are still initializing the pages.
6586 __SetPageReserved(page
);
6589 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6590 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6591 * ever freed or placed on a driver-private list.
6593 page
->pgmap
= pgmap
;
6594 page
->zone_device_data
= NULL
;
6597 * Mark the block movable so that blocks are reserved for
6598 * movable at startup. This will force kernel allocations
6599 * to reserve their blocks rather than leaking throughout
6600 * the address space during boot when many long-lived
6601 * kernel allocations are made.
6603 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6604 * because this is done early in section_activate()
6606 if (IS_ALIGNED(pfn
, pageblock_nr_pages
)) {
6607 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6612 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6613 nr_pages
, jiffies_to_msecs(jiffies
- start
));
6617 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6619 unsigned int order
, t
;
6620 for_each_migratetype_order(order
, t
) {
6621 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6622 zone
->free_area
[order
].nr_free
= 0;
6626 #if !defined(CONFIG_FLATMEM)
6628 * Only struct pages that correspond to ranges defined by memblock.memory
6629 * are zeroed and initialized by going through __init_single_page() during
6630 * memmap_init_zone_range().
6632 * But, there could be struct pages that correspond to holes in
6633 * memblock.memory. This can happen because of the following reasons:
6634 * - physical memory bank size is not necessarily the exact multiple of the
6635 * arbitrary section size
6636 * - early reserved memory may not be listed in memblock.memory
6637 * - memory layouts defined with memmap= kernel parameter may not align
6638 * nicely with memmap sections
6640 * Explicitly initialize those struct pages so that:
6641 * - PG_Reserved is set
6642 * - zone and node links point to zone and node that span the page if the
6643 * hole is in the middle of a zone
6644 * - zone and node links point to adjacent zone/node if the hole falls on
6645 * the zone boundary; the pages in such holes will be prepended to the
6646 * zone/node above the hole except for the trailing pages in the last
6647 * section that will be appended to the zone/node below.
6649 static void __init
init_unavailable_range(unsigned long spfn
,
6656 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6657 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6658 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6659 + pageblock_nr_pages
- 1;
6662 __init_single_page(pfn_to_page(pfn
), pfn
, zone
, node
);
6663 __SetPageReserved(pfn_to_page(pfn
));
6668 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6669 node
, zone_names
[zone
], pgcnt
);
6672 static inline void init_unavailable_range(unsigned long spfn
,
6679 static void __init
memmap_init_zone_range(struct zone
*zone
,
6680 unsigned long start_pfn
,
6681 unsigned long end_pfn
,
6682 unsigned long *hole_pfn
)
6684 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6685 unsigned long zone_end_pfn
= zone_start_pfn
+ zone
->spanned_pages
;
6686 int nid
= zone_to_nid(zone
), zone_id
= zone_idx(zone
);
6688 start_pfn
= clamp(start_pfn
, zone_start_pfn
, zone_end_pfn
);
6689 end_pfn
= clamp(end_pfn
, zone_start_pfn
, zone_end_pfn
);
6691 if (start_pfn
>= end_pfn
)
6694 memmap_init_range(end_pfn
- start_pfn
, nid
, zone_id
, start_pfn
,
6695 zone_end_pfn
, MEMINIT_EARLY
, NULL
, MIGRATE_MOVABLE
);
6697 if (*hole_pfn
< start_pfn
)
6698 init_unavailable_range(*hole_pfn
, start_pfn
, zone_id
, nid
);
6700 *hole_pfn
= end_pfn
;
6703 static void __init
memmap_init(void)
6705 unsigned long start_pfn
, end_pfn
;
6706 unsigned long hole_pfn
= 0;
6707 int i
, j
, zone_id
, nid
;
6709 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6710 struct pglist_data
*node
= NODE_DATA(nid
);
6712 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6713 struct zone
*zone
= node
->node_zones
+ j
;
6715 if (!populated_zone(zone
))
6718 memmap_init_zone_range(zone
, start_pfn
, end_pfn
,
6724 #ifdef CONFIG_SPARSEMEM
6726 * Initialize the memory map for hole in the range [memory_end,
6728 * Append the pages in this hole to the highest zone in the last
6730 * The call to init_unavailable_range() is outside the ifdef to
6731 * silence the compiler warining about zone_id set but not used;
6732 * for FLATMEM it is a nop anyway
6734 end_pfn
= round_up(end_pfn
, PAGES_PER_SECTION
);
6735 if (hole_pfn
< end_pfn
)
6737 init_unavailable_range(hole_pfn
, end_pfn
, zone_id
, nid
);
6740 static int zone_batchsize(struct zone
*zone
)
6746 * The number of pages to batch allocate is either ~0.1%
6747 * of the zone or 1MB, whichever is smaller. The batch
6748 * size is striking a balance between allocation latency
6749 * and zone lock contention.
6751 batch
= min(zone_managed_pages(zone
) >> 10, (1024 * 1024) / PAGE_SIZE
);
6752 batch
/= 4; /* We effectively *= 4 below */
6757 * Clamp the batch to a 2^n - 1 value. Having a power
6758 * of 2 value was found to be more likely to have
6759 * suboptimal cache aliasing properties in some cases.
6761 * For example if 2 tasks are alternately allocating
6762 * batches of pages, one task can end up with a lot
6763 * of pages of one half of the possible page colors
6764 * and the other with pages of the other colors.
6766 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6771 /* The deferral and batching of frees should be suppressed under NOMMU
6774 * The problem is that NOMMU needs to be able to allocate large chunks
6775 * of contiguous memory as there's no hardware page translation to
6776 * assemble apparent contiguous memory from discontiguous pages.
6778 * Queueing large contiguous runs of pages for batching, however,
6779 * causes the pages to actually be freed in smaller chunks. As there
6780 * can be a significant delay between the individual batches being
6781 * recycled, this leads to the once large chunks of space being
6782 * fragmented and becoming unavailable for high-order allocations.
6788 static int zone_highsize(struct zone
*zone
, int batch
, int cpu_online
)
6793 unsigned long total_pages
;
6795 if (!percpu_pagelist_high_fraction
) {
6797 * By default, the high value of the pcp is based on the zone
6798 * low watermark so that if they are full then background
6799 * reclaim will not be started prematurely.
6801 total_pages
= low_wmark_pages(zone
);
6804 * If percpu_pagelist_high_fraction is configured, the high
6805 * value is based on a fraction of the managed pages in the
6808 total_pages
= zone_managed_pages(zone
) / percpu_pagelist_high_fraction
;
6812 * Split the high value across all online CPUs local to the zone. Note
6813 * that early in boot that CPUs may not be online yet and that during
6814 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6815 * onlined. For memory nodes that have no CPUs, split pcp->high across
6816 * all online CPUs to mitigate the risk that reclaim is triggered
6817 * prematurely due to pages stored on pcp lists.
6819 nr_split_cpus
= cpumask_weight(cpumask_of_node(zone_to_nid(zone
))) + cpu_online
;
6821 nr_split_cpus
= num_online_cpus();
6822 high
= total_pages
/ nr_split_cpus
;
6825 * Ensure high is at least batch*4. The multiple is based on the
6826 * historical relationship between high and batch.
6828 high
= max(high
, batch
<< 2);
6837 * pcp->high and pcp->batch values are related and generally batch is lower
6838 * than high. They are also related to pcp->count such that count is lower
6839 * than high, and as soon as it reaches high, the pcplist is flushed.
6841 * However, guaranteeing these relations at all times would require e.g. write
6842 * barriers here but also careful usage of read barriers at the read side, and
6843 * thus be prone to error and bad for performance. Thus the update only prevents
6844 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6845 * can cope with those fields changing asynchronously, and fully trust only the
6846 * pcp->count field on the local CPU with interrupts disabled.
6848 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6849 * outside of boot time (or some other assurance that no concurrent updaters
6852 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6853 unsigned long batch
)
6855 WRITE_ONCE(pcp
->batch
, batch
);
6856 WRITE_ONCE(pcp
->high
, high
);
6859 static void per_cpu_pages_init(struct per_cpu_pages
*pcp
, struct per_cpu_zonestat
*pzstats
)
6863 memset(pcp
, 0, sizeof(*pcp
));
6864 memset(pzstats
, 0, sizeof(*pzstats
));
6866 for (pindex
= 0; pindex
< NR_PCP_LISTS
; pindex
++)
6867 INIT_LIST_HEAD(&pcp
->lists
[pindex
]);
6870 * Set batch and high values safe for a boot pageset. A true percpu
6871 * pageset's initialization will update them subsequently. Here we don't
6872 * need to be as careful as pageset_update() as nobody can access the
6875 pcp
->high
= BOOT_PAGESET_HIGH
;
6876 pcp
->batch
= BOOT_PAGESET_BATCH
;
6877 pcp
->free_factor
= 0;
6880 static void __zone_set_pageset_high_and_batch(struct zone
*zone
, unsigned long high
,
6881 unsigned long batch
)
6883 struct per_cpu_pages
*pcp
;
6886 for_each_possible_cpu(cpu
) {
6887 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
6888 pageset_update(pcp
, high
, batch
);
6893 * Calculate and set new high and batch values for all per-cpu pagesets of a
6894 * zone based on the zone's size.
6896 static void zone_set_pageset_high_and_batch(struct zone
*zone
, int cpu_online
)
6898 int new_high
, new_batch
;
6900 new_batch
= max(1, zone_batchsize(zone
));
6901 new_high
= zone_highsize(zone
, new_batch
, cpu_online
);
6903 if (zone
->pageset_high
== new_high
&&
6904 zone
->pageset_batch
== new_batch
)
6907 zone
->pageset_high
= new_high
;
6908 zone
->pageset_batch
= new_batch
;
6910 __zone_set_pageset_high_and_batch(zone
, new_high
, new_batch
);
6913 void __meminit
setup_zone_pageset(struct zone
*zone
)
6917 /* Size may be 0 on !SMP && !NUMA */
6918 if (sizeof(struct per_cpu_zonestat
) > 0)
6919 zone
->per_cpu_zonestats
= alloc_percpu(struct per_cpu_zonestat
);
6921 zone
->per_cpu_pageset
= alloc_percpu(struct per_cpu_pages
);
6922 for_each_possible_cpu(cpu
) {
6923 struct per_cpu_pages
*pcp
;
6924 struct per_cpu_zonestat
*pzstats
;
6926 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
6927 pzstats
= per_cpu_ptr(zone
->per_cpu_zonestats
, cpu
);
6928 per_cpu_pages_init(pcp
, pzstats
);
6931 zone_set_pageset_high_and_batch(zone
, 0);
6935 * Allocate per cpu pagesets and initialize them.
6936 * Before this call only boot pagesets were available.
6938 void __init
setup_per_cpu_pageset(void)
6940 struct pglist_data
*pgdat
;
6942 int __maybe_unused cpu
;
6944 for_each_populated_zone(zone
)
6945 setup_zone_pageset(zone
);
6949 * Unpopulated zones continue using the boot pagesets.
6950 * The numa stats for these pagesets need to be reset.
6951 * Otherwise, they will end up skewing the stats of
6952 * the nodes these zones are associated with.
6954 for_each_possible_cpu(cpu
) {
6955 struct per_cpu_zonestat
*pzstats
= &per_cpu(boot_zonestats
, cpu
);
6956 memset(pzstats
->vm_numa_event
, 0,
6957 sizeof(pzstats
->vm_numa_event
));
6961 for_each_online_pgdat(pgdat
)
6962 pgdat
->per_cpu_nodestats
=
6963 alloc_percpu(struct per_cpu_nodestat
);
6966 static __meminit
void zone_pcp_init(struct zone
*zone
)
6969 * per cpu subsystem is not up at this point. The following code
6970 * relies on the ability of the linker to provide the
6971 * offset of a (static) per cpu variable into the per cpu area.
6973 zone
->per_cpu_pageset
= &boot_pageset
;
6974 zone
->per_cpu_zonestats
= &boot_zonestats
;
6975 zone
->pageset_high
= BOOT_PAGESET_HIGH
;
6976 zone
->pageset_batch
= BOOT_PAGESET_BATCH
;
6978 if (populated_zone(zone
))
6979 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone
->name
,
6980 zone
->present_pages
, zone_batchsize(zone
));
6983 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6984 unsigned long zone_start_pfn
,
6987 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6988 int zone_idx
= zone_idx(zone
) + 1;
6990 if (zone_idx
> pgdat
->nr_zones
)
6991 pgdat
->nr_zones
= zone_idx
;
6993 zone
->zone_start_pfn
= zone_start_pfn
;
6995 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6996 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6998 (unsigned long)zone_idx(zone
),
6999 zone_start_pfn
, (zone_start_pfn
+ size
));
7001 zone_init_free_lists(zone
);
7002 zone
->initialized
= 1;
7006 * get_pfn_range_for_nid - Return the start and end page frames for a node
7007 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7008 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7009 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7011 * It returns the start and end page frame of a node based on information
7012 * provided by memblock_set_node(). If called for a node
7013 * with no available memory, a warning is printed and the start and end
7016 void __init
get_pfn_range_for_nid(unsigned int nid
,
7017 unsigned long *start_pfn
, unsigned long *end_pfn
)
7019 unsigned long this_start_pfn
, this_end_pfn
;
7025 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
7026 *start_pfn
= min(*start_pfn
, this_start_pfn
);
7027 *end_pfn
= max(*end_pfn
, this_end_pfn
);
7030 if (*start_pfn
== -1UL)
7035 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7036 * assumption is made that zones within a node are ordered in monotonic
7037 * increasing memory addresses so that the "highest" populated zone is used
7039 static void __init
find_usable_zone_for_movable(void)
7042 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
7043 if (zone_index
== ZONE_MOVABLE
)
7046 if (arch_zone_highest_possible_pfn
[zone_index
] >
7047 arch_zone_lowest_possible_pfn
[zone_index
])
7051 VM_BUG_ON(zone_index
== -1);
7052 movable_zone
= zone_index
;
7056 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7057 * because it is sized independent of architecture. Unlike the other zones,
7058 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7059 * in each node depending on the size of each node and how evenly kernelcore
7060 * is distributed. This helper function adjusts the zone ranges
7061 * provided by the architecture for a given node by using the end of the
7062 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7063 * zones within a node are in order of monotonic increases memory addresses
7065 static void __init
adjust_zone_range_for_zone_movable(int nid
,
7066 unsigned long zone_type
,
7067 unsigned long node_start_pfn
,
7068 unsigned long node_end_pfn
,
7069 unsigned long *zone_start_pfn
,
7070 unsigned long *zone_end_pfn
)
7072 /* Only adjust if ZONE_MOVABLE is on this node */
7073 if (zone_movable_pfn
[nid
]) {
7074 /* Size ZONE_MOVABLE */
7075 if (zone_type
== ZONE_MOVABLE
) {
7076 *zone_start_pfn
= zone_movable_pfn
[nid
];
7077 *zone_end_pfn
= min(node_end_pfn
,
7078 arch_zone_highest_possible_pfn
[movable_zone
]);
7080 /* Adjust for ZONE_MOVABLE starting within this range */
7081 } else if (!mirrored_kernelcore
&&
7082 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
7083 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
7084 *zone_end_pfn
= zone_movable_pfn
[nid
];
7086 /* Check if this whole range is within ZONE_MOVABLE */
7087 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
7088 *zone_start_pfn
= *zone_end_pfn
;
7093 * Return the number of pages a zone spans in a node, including holes
7094 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7096 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
7097 unsigned long zone_type
,
7098 unsigned long node_start_pfn
,
7099 unsigned long node_end_pfn
,
7100 unsigned long *zone_start_pfn
,
7101 unsigned long *zone_end_pfn
)
7103 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
7104 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
7105 /* When hotadd a new node from cpu_up(), the node should be empty */
7106 if (!node_start_pfn
&& !node_end_pfn
)
7109 /* Get the start and end of the zone */
7110 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
7111 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
7112 adjust_zone_range_for_zone_movable(nid
, zone_type
,
7113 node_start_pfn
, node_end_pfn
,
7114 zone_start_pfn
, zone_end_pfn
);
7116 /* Check that this node has pages within the zone's required range */
7117 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
7120 /* Move the zone boundaries inside the node if necessary */
7121 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
7122 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
7124 /* Return the spanned pages */
7125 return *zone_end_pfn
- *zone_start_pfn
;
7129 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7130 * then all holes in the requested range will be accounted for.
7132 unsigned long __init
__absent_pages_in_range(int nid
,
7133 unsigned long range_start_pfn
,
7134 unsigned long range_end_pfn
)
7136 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
7137 unsigned long start_pfn
, end_pfn
;
7140 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7141 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
7142 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
7143 nr_absent
-= end_pfn
- start_pfn
;
7149 * absent_pages_in_range - Return number of page frames in holes within a range
7150 * @start_pfn: The start PFN to start searching for holes
7151 * @end_pfn: The end PFN to stop searching for holes
7153 * Return: the number of pages frames in memory holes within a range.
7155 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
7156 unsigned long end_pfn
)
7158 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
7161 /* Return the number of page frames in holes in a zone on a node */
7162 static unsigned long __init
zone_absent_pages_in_node(int nid
,
7163 unsigned long zone_type
,
7164 unsigned long node_start_pfn
,
7165 unsigned long node_end_pfn
)
7167 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
7168 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
7169 unsigned long zone_start_pfn
, zone_end_pfn
;
7170 unsigned long nr_absent
;
7172 /* When hotadd a new node from cpu_up(), the node should be empty */
7173 if (!node_start_pfn
&& !node_end_pfn
)
7176 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
7177 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
7179 adjust_zone_range_for_zone_movable(nid
, zone_type
,
7180 node_start_pfn
, node_end_pfn
,
7181 &zone_start_pfn
, &zone_end_pfn
);
7182 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
7185 * ZONE_MOVABLE handling.
7186 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7189 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
7190 unsigned long start_pfn
, end_pfn
;
7191 struct memblock_region
*r
;
7193 for_each_mem_region(r
) {
7194 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
7195 zone_start_pfn
, zone_end_pfn
);
7196 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
7197 zone_start_pfn
, zone_end_pfn
);
7199 if (zone_type
== ZONE_MOVABLE
&&
7200 memblock_is_mirror(r
))
7201 nr_absent
+= end_pfn
- start_pfn
;
7203 if (zone_type
== ZONE_NORMAL
&&
7204 !memblock_is_mirror(r
))
7205 nr_absent
+= end_pfn
- start_pfn
;
7212 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
7213 unsigned long node_start_pfn
,
7214 unsigned long node_end_pfn
)
7216 unsigned long realtotalpages
= 0, totalpages
= 0;
7219 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7220 struct zone
*zone
= pgdat
->node_zones
+ i
;
7221 unsigned long zone_start_pfn
, zone_end_pfn
;
7222 unsigned long spanned
, absent
;
7223 unsigned long size
, real_size
;
7225 spanned
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
7230 absent
= zone_absent_pages_in_node(pgdat
->node_id
, i
,
7235 real_size
= size
- absent
;
7238 zone
->zone_start_pfn
= zone_start_pfn
;
7240 zone
->zone_start_pfn
= 0;
7241 zone
->spanned_pages
= size
;
7242 zone
->present_pages
= real_size
;
7243 #if defined(CONFIG_MEMORY_HOTPLUG)
7244 zone
->present_early_pages
= real_size
;
7248 realtotalpages
+= real_size
;
7251 pgdat
->node_spanned_pages
= totalpages
;
7252 pgdat
->node_present_pages
= realtotalpages
;
7253 pr_debug("On node %d totalpages: %lu\n", pgdat
->node_id
, realtotalpages
);
7256 #ifndef CONFIG_SPARSEMEM
7258 * Calculate the size of the zone->blockflags rounded to an unsigned long
7259 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7260 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7261 * round what is now in bits to nearest long in bits, then return it in
7264 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
7266 unsigned long usemapsize
;
7268 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
7269 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
7270 usemapsize
= usemapsize
>> pageblock_order
;
7271 usemapsize
*= NR_PAGEBLOCK_BITS
;
7272 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
7274 return usemapsize
/ 8;
7277 static void __ref
setup_usemap(struct zone
*zone
)
7279 unsigned long usemapsize
= usemap_size(zone
->zone_start_pfn
,
7280 zone
->spanned_pages
);
7281 zone
->pageblock_flags
= NULL
;
7283 zone
->pageblock_flags
=
7284 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
7286 if (!zone
->pageblock_flags
)
7287 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7288 usemapsize
, zone
->name
, zone_to_nid(zone
));
7292 static inline void setup_usemap(struct zone
*zone
) {}
7293 #endif /* CONFIG_SPARSEMEM */
7295 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7297 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7298 void __init
set_pageblock_order(void)
7302 /* Check that pageblock_nr_pages has not already been setup */
7303 if (pageblock_order
)
7306 if (HPAGE_SHIFT
> PAGE_SHIFT
)
7307 order
= HUGETLB_PAGE_ORDER
;
7309 order
= MAX_ORDER
- 1;
7312 * Assume the largest contiguous order of interest is a huge page.
7313 * This value may be variable depending on boot parameters on IA64 and
7316 pageblock_order
= order
;
7318 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7321 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7322 * is unused as pageblock_order is set at compile-time. See
7323 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7326 void __init
set_pageblock_order(void)
7330 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7332 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
7333 unsigned long present_pages
)
7335 unsigned long pages
= spanned_pages
;
7338 * Provide a more accurate estimation if there are holes within
7339 * the zone and SPARSEMEM is in use. If there are holes within the
7340 * zone, each populated memory region may cost us one or two extra
7341 * memmap pages due to alignment because memmap pages for each
7342 * populated regions may not be naturally aligned on page boundary.
7343 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7345 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
7346 IS_ENABLED(CONFIG_SPARSEMEM
))
7347 pages
= present_pages
;
7349 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
7352 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7353 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
7355 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
7357 spin_lock_init(&ds_queue
->split_queue_lock
);
7358 INIT_LIST_HEAD(&ds_queue
->split_queue
);
7359 ds_queue
->split_queue_len
= 0;
7362 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
7365 #ifdef CONFIG_COMPACTION
7366 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
7368 init_waitqueue_head(&pgdat
->kcompactd_wait
);
7371 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
7374 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
7376 pgdat_resize_init(pgdat
);
7378 pgdat_init_split_queue(pgdat
);
7379 pgdat_init_kcompactd(pgdat
);
7381 init_waitqueue_head(&pgdat
->kswapd_wait
);
7382 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
7384 pgdat_page_ext_init(pgdat
);
7385 lruvec_init(&pgdat
->__lruvec
);
7388 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
7389 unsigned long remaining_pages
)
7391 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
7392 zone_set_nid(zone
, nid
);
7393 zone
->name
= zone_names
[idx
];
7394 zone
->zone_pgdat
= NODE_DATA(nid
);
7395 spin_lock_init(&zone
->lock
);
7396 zone_seqlock_init(zone
);
7397 zone_pcp_init(zone
);
7401 * Set up the zone data structures
7402 * - init pgdat internals
7403 * - init all zones belonging to this node
7405 * NOTE: this function is only called during memory hotplug
7407 #ifdef CONFIG_MEMORY_HOTPLUG
7408 void __ref
free_area_init_core_hotplug(int nid
)
7411 pg_data_t
*pgdat
= NODE_DATA(nid
);
7413 pgdat_init_internals(pgdat
);
7414 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
7415 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
7420 * Set up the zone data structures:
7421 * - mark all pages reserved
7422 * - mark all memory queues empty
7423 * - clear the memory bitmaps
7425 * NOTE: pgdat should get zeroed by caller.
7426 * NOTE: this function is only called during early init.
7428 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
7431 int nid
= pgdat
->node_id
;
7433 pgdat_init_internals(pgdat
);
7434 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
7436 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7437 struct zone
*zone
= pgdat
->node_zones
+ j
;
7438 unsigned long size
, freesize
, memmap_pages
;
7440 size
= zone
->spanned_pages
;
7441 freesize
= zone
->present_pages
;
7444 * Adjust freesize so that it accounts for how much memory
7445 * is used by this zone for memmap. This affects the watermark
7446 * and per-cpu initialisations
7448 memmap_pages
= calc_memmap_size(size
, freesize
);
7449 if (!is_highmem_idx(j
)) {
7450 if (freesize
>= memmap_pages
) {
7451 freesize
-= memmap_pages
;
7453 pr_debug(" %s zone: %lu pages used for memmap\n",
7454 zone_names
[j
], memmap_pages
);
7456 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7457 zone_names
[j
], memmap_pages
, freesize
);
7460 /* Account for reserved pages */
7461 if (j
== 0 && freesize
> dma_reserve
) {
7462 freesize
-= dma_reserve
;
7463 pr_debug(" %s zone: %lu pages reserved\n", zone_names
[0], dma_reserve
);
7466 if (!is_highmem_idx(j
))
7467 nr_kernel_pages
+= freesize
;
7468 /* Charge for highmem memmap if there are enough kernel pages */
7469 else if (nr_kernel_pages
> memmap_pages
* 2)
7470 nr_kernel_pages
-= memmap_pages
;
7471 nr_all_pages
+= freesize
;
7474 * Set an approximate value for lowmem here, it will be adjusted
7475 * when the bootmem allocator frees pages into the buddy system.
7476 * And all highmem pages will be managed by the buddy system.
7478 zone_init_internals(zone
, j
, nid
, freesize
);
7483 set_pageblock_order();
7485 init_currently_empty_zone(zone
, zone
->zone_start_pfn
, size
);
7489 #ifdef CONFIG_FLATMEM
7490 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
7492 unsigned long __maybe_unused start
= 0;
7493 unsigned long __maybe_unused offset
= 0;
7495 /* Skip empty nodes */
7496 if (!pgdat
->node_spanned_pages
)
7499 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
7500 offset
= pgdat
->node_start_pfn
- start
;
7501 /* ia64 gets its own node_mem_map, before this, without bootmem */
7502 if (!pgdat
->node_mem_map
) {
7503 unsigned long size
, end
;
7507 * The zone's endpoints aren't required to be MAX_ORDER
7508 * aligned but the node_mem_map endpoints must be in order
7509 * for the buddy allocator to function correctly.
7511 end
= pgdat_end_pfn(pgdat
);
7512 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
7513 size
= (end
- start
) * sizeof(struct page
);
7514 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
7517 panic("Failed to allocate %ld bytes for node %d memory map\n",
7518 size
, pgdat
->node_id
);
7519 pgdat
->node_mem_map
= map
+ offset
;
7521 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7522 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
7523 (unsigned long)pgdat
->node_mem_map
);
7526 * With no DISCONTIG, the global mem_map is just set as node 0's
7528 if (pgdat
== NODE_DATA(0)) {
7529 mem_map
= NODE_DATA(0)->node_mem_map
;
7530 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
7536 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
7537 #endif /* CONFIG_FLATMEM */
7539 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7540 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
7542 pgdat
->first_deferred_pfn
= ULONG_MAX
;
7545 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
7548 static void __init
free_area_init_node(int nid
)
7550 pg_data_t
*pgdat
= NODE_DATA(nid
);
7551 unsigned long start_pfn
= 0;
7552 unsigned long end_pfn
= 0;
7554 /* pg_data_t should be reset to zero when it's allocated */
7555 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_highest_zoneidx
);
7557 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
7559 pgdat
->node_id
= nid
;
7560 pgdat
->node_start_pfn
= start_pfn
;
7561 pgdat
->per_cpu_nodestats
= NULL
;
7563 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
7564 (u64
)start_pfn
<< PAGE_SHIFT
,
7565 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
7566 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
);
7568 alloc_node_mem_map(pgdat
);
7569 pgdat_set_deferred_range(pgdat
);
7571 free_area_init_core(pgdat
);
7574 void __init
free_area_init_memoryless_node(int nid
)
7576 free_area_init_node(nid
);
7579 #if MAX_NUMNODES > 1
7581 * Figure out the number of possible node ids.
7583 void __init
setup_nr_node_ids(void)
7585 unsigned int highest
;
7587 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
7588 nr_node_ids
= highest
+ 1;
7593 * node_map_pfn_alignment - determine the maximum internode alignment
7595 * This function should be called after node map is populated and sorted.
7596 * It calculates the maximum power of two alignment which can distinguish
7599 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7600 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7601 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7602 * shifted, 1GiB is enough and this function will indicate so.
7604 * This is used to test whether pfn -> nid mapping of the chosen memory
7605 * model has fine enough granularity to avoid incorrect mapping for the
7606 * populated node map.
7608 * Return: the determined alignment in pfn's. 0 if there is no alignment
7609 * requirement (single node).
7611 unsigned long __init
node_map_pfn_alignment(void)
7613 unsigned long accl_mask
= 0, last_end
= 0;
7614 unsigned long start
, end
, mask
;
7615 int last_nid
= NUMA_NO_NODE
;
7618 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
7619 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
7626 * Start with a mask granular enough to pin-point to the
7627 * start pfn and tick off bits one-by-one until it becomes
7628 * too coarse to separate the current node from the last.
7630 mask
= ~((1 << __ffs(start
)) - 1);
7631 while (mask
&& last_end
<= (start
& (mask
<< 1)))
7634 /* accumulate all internode masks */
7638 /* convert mask to number of pages */
7639 return ~accl_mask
+ 1;
7643 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7645 * Return: the minimum PFN based on information provided via
7646 * memblock_set_node().
7648 unsigned long __init
find_min_pfn_with_active_regions(void)
7650 return PHYS_PFN(memblock_start_of_DRAM());
7654 * early_calculate_totalpages()
7655 * Sum pages in active regions for movable zone.
7656 * Populate N_MEMORY for calculating usable_nodes.
7658 static unsigned long __init
early_calculate_totalpages(void)
7660 unsigned long totalpages
= 0;
7661 unsigned long start_pfn
, end_pfn
;
7664 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7665 unsigned long pages
= end_pfn
- start_pfn
;
7667 totalpages
+= pages
;
7669 node_set_state(nid
, N_MEMORY
);
7675 * Find the PFN the Movable zone begins in each node. Kernel memory
7676 * is spread evenly between nodes as long as the nodes have enough
7677 * memory. When they don't, some nodes will have more kernelcore than
7680 static void __init
find_zone_movable_pfns_for_nodes(void)
7683 unsigned long usable_startpfn
;
7684 unsigned long kernelcore_node
, kernelcore_remaining
;
7685 /* save the state before borrow the nodemask */
7686 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
7687 unsigned long totalpages
= early_calculate_totalpages();
7688 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
7689 struct memblock_region
*r
;
7691 /* Need to find movable_zone earlier when movable_node is specified. */
7692 find_usable_zone_for_movable();
7695 * If movable_node is specified, ignore kernelcore and movablecore
7698 if (movable_node_is_enabled()) {
7699 for_each_mem_region(r
) {
7700 if (!memblock_is_hotpluggable(r
))
7703 nid
= memblock_get_region_node(r
);
7705 usable_startpfn
= PFN_DOWN(r
->base
);
7706 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7707 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7715 * If kernelcore=mirror is specified, ignore movablecore option
7717 if (mirrored_kernelcore
) {
7718 bool mem_below_4gb_not_mirrored
= false;
7720 for_each_mem_region(r
) {
7721 if (memblock_is_mirror(r
))
7724 nid
= memblock_get_region_node(r
);
7726 usable_startpfn
= memblock_region_memory_base_pfn(r
);
7728 if (usable_startpfn
< 0x100000) {
7729 mem_below_4gb_not_mirrored
= true;
7733 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7734 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7738 if (mem_below_4gb_not_mirrored
)
7739 pr_warn("This configuration results in unmirrored kernel memory.\n");
7745 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7746 * amount of necessary memory.
7748 if (required_kernelcore_percent
)
7749 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
7751 if (required_movablecore_percent
)
7752 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
7756 * If movablecore= was specified, calculate what size of
7757 * kernelcore that corresponds so that memory usable for
7758 * any allocation type is evenly spread. If both kernelcore
7759 * and movablecore are specified, then the value of kernelcore
7760 * will be used for required_kernelcore if it's greater than
7761 * what movablecore would have allowed.
7763 if (required_movablecore
) {
7764 unsigned long corepages
;
7767 * Round-up so that ZONE_MOVABLE is at least as large as what
7768 * was requested by the user
7770 required_movablecore
=
7771 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
7772 required_movablecore
= min(totalpages
, required_movablecore
);
7773 corepages
= totalpages
- required_movablecore
;
7775 required_kernelcore
= max(required_kernelcore
, corepages
);
7779 * If kernelcore was not specified or kernelcore size is larger
7780 * than totalpages, there is no ZONE_MOVABLE.
7782 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
7785 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7786 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
7789 /* Spread kernelcore memory as evenly as possible throughout nodes */
7790 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7791 for_each_node_state(nid
, N_MEMORY
) {
7792 unsigned long start_pfn
, end_pfn
;
7795 * Recalculate kernelcore_node if the division per node
7796 * now exceeds what is necessary to satisfy the requested
7797 * amount of memory for the kernel
7799 if (required_kernelcore
< kernelcore_node
)
7800 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7803 * As the map is walked, we track how much memory is usable
7804 * by the kernel using kernelcore_remaining. When it is
7805 * 0, the rest of the node is usable by ZONE_MOVABLE
7807 kernelcore_remaining
= kernelcore_node
;
7809 /* Go through each range of PFNs within this node */
7810 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7811 unsigned long size_pages
;
7813 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
7814 if (start_pfn
>= end_pfn
)
7817 /* Account for what is only usable for kernelcore */
7818 if (start_pfn
< usable_startpfn
) {
7819 unsigned long kernel_pages
;
7820 kernel_pages
= min(end_pfn
, usable_startpfn
)
7823 kernelcore_remaining
-= min(kernel_pages
,
7824 kernelcore_remaining
);
7825 required_kernelcore
-= min(kernel_pages
,
7826 required_kernelcore
);
7828 /* Continue if range is now fully accounted */
7829 if (end_pfn
<= usable_startpfn
) {
7832 * Push zone_movable_pfn to the end so
7833 * that if we have to rebalance
7834 * kernelcore across nodes, we will
7835 * not double account here
7837 zone_movable_pfn
[nid
] = end_pfn
;
7840 start_pfn
= usable_startpfn
;
7844 * The usable PFN range for ZONE_MOVABLE is from
7845 * start_pfn->end_pfn. Calculate size_pages as the
7846 * number of pages used as kernelcore
7848 size_pages
= end_pfn
- start_pfn
;
7849 if (size_pages
> kernelcore_remaining
)
7850 size_pages
= kernelcore_remaining
;
7851 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7854 * Some kernelcore has been met, update counts and
7855 * break if the kernelcore for this node has been
7858 required_kernelcore
-= min(required_kernelcore
,
7860 kernelcore_remaining
-= size_pages
;
7861 if (!kernelcore_remaining
)
7867 * If there is still required_kernelcore, we do another pass with one
7868 * less node in the count. This will push zone_movable_pfn[nid] further
7869 * along on the nodes that still have memory until kernelcore is
7873 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7877 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7878 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7879 zone_movable_pfn
[nid
] =
7880 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7883 /* restore the node_state */
7884 node_states
[N_MEMORY
] = saved_node_state
;
7887 /* Any regular or high memory on that node ? */
7888 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7890 enum zone_type zone_type
;
7892 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7893 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7894 if (populated_zone(zone
)) {
7895 if (IS_ENABLED(CONFIG_HIGHMEM
))
7896 node_set_state(nid
, N_HIGH_MEMORY
);
7897 if (zone_type
<= ZONE_NORMAL
)
7898 node_set_state(nid
, N_NORMAL_MEMORY
);
7905 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7906 * such cases we allow max_zone_pfn sorted in the descending order
7908 bool __weak
arch_has_descending_max_zone_pfns(void)
7914 * free_area_init - Initialise all pg_data_t and zone data
7915 * @max_zone_pfn: an array of max PFNs for each zone
7917 * This will call free_area_init_node() for each active node in the system.
7918 * Using the page ranges provided by memblock_set_node(), the size of each
7919 * zone in each node and their holes is calculated. If the maximum PFN
7920 * between two adjacent zones match, it is assumed that the zone is empty.
7921 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7922 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7923 * starts where the previous one ended. For example, ZONE_DMA32 starts
7924 * at arch_max_dma_pfn.
7926 void __init
free_area_init(unsigned long *max_zone_pfn
)
7928 unsigned long start_pfn
, end_pfn
;
7932 /* Record where the zone boundaries are */
7933 memset(arch_zone_lowest_possible_pfn
, 0,
7934 sizeof(arch_zone_lowest_possible_pfn
));
7935 memset(arch_zone_highest_possible_pfn
, 0,
7936 sizeof(arch_zone_highest_possible_pfn
));
7938 start_pfn
= find_min_pfn_with_active_regions();
7939 descending
= arch_has_descending_max_zone_pfns();
7941 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7943 zone
= MAX_NR_ZONES
- i
- 1;
7947 if (zone
== ZONE_MOVABLE
)
7950 end_pfn
= max(max_zone_pfn
[zone
], start_pfn
);
7951 arch_zone_lowest_possible_pfn
[zone
] = start_pfn
;
7952 arch_zone_highest_possible_pfn
[zone
] = end_pfn
;
7954 start_pfn
= end_pfn
;
7957 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7958 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7959 find_zone_movable_pfns_for_nodes();
7961 /* Print out the zone ranges */
7962 pr_info("Zone ranges:\n");
7963 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7964 if (i
== ZONE_MOVABLE
)
7966 pr_info(" %-8s ", zone_names
[i
]);
7967 if (arch_zone_lowest_possible_pfn
[i
] ==
7968 arch_zone_highest_possible_pfn
[i
])
7971 pr_cont("[mem %#018Lx-%#018Lx]\n",
7972 (u64
)arch_zone_lowest_possible_pfn
[i
]
7974 ((u64
)arch_zone_highest_possible_pfn
[i
]
7975 << PAGE_SHIFT
) - 1);
7978 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7979 pr_info("Movable zone start for each node\n");
7980 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7981 if (zone_movable_pfn
[i
])
7982 pr_info(" Node %d: %#018Lx\n", i
,
7983 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7987 * Print out the early node map, and initialize the
7988 * subsection-map relative to active online memory ranges to
7989 * enable future "sub-section" extensions of the memory map.
7991 pr_info("Early memory node ranges\n");
7992 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7993 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7994 (u64
)start_pfn
<< PAGE_SHIFT
,
7995 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7996 subsection_map_init(start_pfn
, end_pfn
- start_pfn
);
7999 /* Initialise every node */
8000 mminit_verify_pageflags_layout();
8001 setup_nr_node_ids();
8002 for_each_online_node(nid
) {
8003 pg_data_t
*pgdat
= NODE_DATA(nid
);
8004 free_area_init_node(nid
);
8006 /* Any memory on that node */
8007 if (pgdat
->node_present_pages
)
8008 node_set_state(nid
, N_MEMORY
);
8009 check_for_memory(pgdat
, nid
);
8015 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
8016 unsigned long *percent
)
8018 unsigned long long coremem
;
8024 /* Value may be a percentage of total memory, otherwise bytes */
8025 coremem
= simple_strtoull(p
, &endptr
, 0);
8026 if (*endptr
== '%') {
8027 /* Paranoid check for percent values greater than 100 */
8028 WARN_ON(coremem
> 100);
8032 coremem
= memparse(p
, &p
);
8033 /* Paranoid check that UL is enough for the coremem value */
8034 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
8036 *core
= coremem
>> PAGE_SHIFT
;
8043 * kernelcore=size sets the amount of memory for use for allocations that
8044 * cannot be reclaimed or migrated.
8046 static int __init
cmdline_parse_kernelcore(char *p
)
8048 /* parse kernelcore=mirror */
8049 if (parse_option_str(p
, "mirror")) {
8050 mirrored_kernelcore
= true;
8054 return cmdline_parse_core(p
, &required_kernelcore
,
8055 &required_kernelcore_percent
);
8059 * movablecore=size sets the amount of memory for use for allocations that
8060 * can be reclaimed or migrated.
8062 static int __init
cmdline_parse_movablecore(char *p
)
8064 return cmdline_parse_core(p
, &required_movablecore
,
8065 &required_movablecore_percent
);
8068 early_param("kernelcore", cmdline_parse_kernelcore
);
8069 early_param("movablecore", cmdline_parse_movablecore
);
8071 void adjust_managed_page_count(struct page
*page
, long count
)
8073 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
8074 totalram_pages_add(count
);
8075 #ifdef CONFIG_HIGHMEM
8076 if (PageHighMem(page
))
8077 totalhigh_pages_add(count
);
8080 EXPORT_SYMBOL(adjust_managed_page_count
);
8082 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
8085 unsigned long pages
= 0;
8087 start
= (void *)PAGE_ALIGN((unsigned long)start
);
8088 end
= (void *)((unsigned long)end
& PAGE_MASK
);
8089 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
8090 struct page
*page
= virt_to_page(pos
);
8091 void *direct_map_addr
;
8094 * 'direct_map_addr' might be different from 'pos'
8095 * because some architectures' virt_to_page()
8096 * work with aliases. Getting the direct map
8097 * address ensures that we get a _writeable_
8098 * alias for the memset().
8100 direct_map_addr
= page_address(page
);
8102 * Perform a kasan-unchecked memset() since this memory
8103 * has not been initialized.
8105 direct_map_addr
= kasan_reset_tag(direct_map_addr
);
8106 if ((unsigned int)poison
<= 0xFF)
8107 memset(direct_map_addr
, poison
, PAGE_SIZE
);
8109 free_reserved_page(page
);
8113 pr_info("Freeing %s memory: %ldK\n",
8114 s
, pages
<< (PAGE_SHIFT
- 10));
8119 void __init
mem_init_print_info(void)
8121 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
8122 unsigned long init_code_size
, init_data_size
;
8124 physpages
= get_num_physpages();
8125 codesize
= _etext
- _stext
;
8126 datasize
= _edata
- _sdata
;
8127 rosize
= __end_rodata
- __start_rodata
;
8128 bss_size
= __bss_stop
- __bss_start
;
8129 init_data_size
= __init_end
- __init_begin
;
8130 init_code_size
= _einittext
- _sinittext
;
8133 * Detect special cases and adjust section sizes accordingly:
8134 * 1) .init.* may be embedded into .data sections
8135 * 2) .init.text.* may be out of [__init_begin, __init_end],
8136 * please refer to arch/tile/kernel/vmlinux.lds.S.
8137 * 3) .rodata.* may be embedded into .text or .data sections.
8139 #define adj_init_size(start, end, size, pos, adj) \
8141 if (start <= pos && pos < end && size > adj) \
8145 adj_init_size(__init_begin
, __init_end
, init_data_size
,
8146 _sinittext
, init_code_size
);
8147 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
8148 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
8149 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
8150 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
8152 #undef adj_init_size
8154 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8155 #ifdef CONFIG_HIGHMEM
8159 nr_free_pages() << (PAGE_SHIFT
- 10),
8160 physpages
<< (PAGE_SHIFT
- 10),
8161 codesize
>> 10, datasize
>> 10, rosize
>> 10,
8162 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
8163 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
8164 totalcma_pages
<< (PAGE_SHIFT
- 10)
8165 #ifdef CONFIG_HIGHMEM
8166 , totalhigh_pages() << (PAGE_SHIFT
- 10)
8172 * set_dma_reserve - set the specified number of pages reserved in the first zone
8173 * @new_dma_reserve: The number of pages to mark reserved
8175 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8176 * In the DMA zone, a significant percentage may be consumed by kernel image
8177 * and other unfreeable allocations which can skew the watermarks badly. This
8178 * function may optionally be used to account for unfreeable pages in the
8179 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8180 * smaller per-cpu batchsize.
8182 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
8184 dma_reserve
= new_dma_reserve
;
8187 static int page_alloc_cpu_dead(unsigned int cpu
)
8191 lru_add_drain_cpu(cpu
);
8195 * Spill the event counters of the dead processor
8196 * into the current processors event counters.
8197 * This artificially elevates the count of the current
8200 vm_events_fold_cpu(cpu
);
8203 * Zero the differential counters of the dead processor
8204 * so that the vm statistics are consistent.
8206 * This is only okay since the processor is dead and cannot
8207 * race with what we are doing.
8209 cpu_vm_stats_fold(cpu
);
8211 for_each_populated_zone(zone
)
8212 zone_pcp_update(zone
, 0);
8217 static int page_alloc_cpu_online(unsigned int cpu
)
8221 for_each_populated_zone(zone
)
8222 zone_pcp_update(zone
, 1);
8227 int hashdist
= HASHDIST_DEFAULT
;
8229 static int __init
set_hashdist(char *str
)
8233 hashdist
= simple_strtoul(str
, &str
, 0);
8236 __setup("hashdist=", set_hashdist
);
8239 void __init
page_alloc_init(void)
8244 if (num_node_state(N_MEMORY
) == 1)
8248 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC
,
8249 "mm/page_alloc:pcp",
8250 page_alloc_cpu_online
,
8251 page_alloc_cpu_dead
);
8256 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8257 * or min_free_kbytes changes.
8259 static void calculate_totalreserve_pages(void)
8261 struct pglist_data
*pgdat
;
8262 unsigned long reserve_pages
= 0;
8263 enum zone_type i
, j
;
8265 for_each_online_pgdat(pgdat
) {
8267 pgdat
->totalreserve_pages
= 0;
8269 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
8270 struct zone
*zone
= pgdat
->node_zones
+ i
;
8272 unsigned long managed_pages
= zone_managed_pages(zone
);
8274 /* Find valid and maximum lowmem_reserve in the zone */
8275 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
8276 if (zone
->lowmem_reserve
[j
] > max
)
8277 max
= zone
->lowmem_reserve
[j
];
8280 /* we treat the high watermark as reserved pages. */
8281 max
+= high_wmark_pages(zone
);
8283 if (max
> managed_pages
)
8284 max
= managed_pages
;
8286 pgdat
->totalreserve_pages
+= max
;
8288 reserve_pages
+= max
;
8291 totalreserve_pages
= reserve_pages
;
8295 * setup_per_zone_lowmem_reserve - called whenever
8296 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8297 * has a correct pages reserved value, so an adequate number of
8298 * pages are left in the zone after a successful __alloc_pages().
8300 static void setup_per_zone_lowmem_reserve(void)
8302 struct pglist_data
*pgdat
;
8303 enum zone_type i
, j
;
8305 for_each_online_pgdat(pgdat
) {
8306 for (i
= 0; i
< MAX_NR_ZONES
- 1; i
++) {
8307 struct zone
*zone
= &pgdat
->node_zones
[i
];
8308 int ratio
= sysctl_lowmem_reserve_ratio
[i
];
8309 bool clear
= !ratio
|| !zone_managed_pages(zone
);
8310 unsigned long managed_pages
= 0;
8312 for (j
= i
+ 1; j
< MAX_NR_ZONES
; j
++) {
8313 struct zone
*upper_zone
= &pgdat
->node_zones
[j
];
8315 managed_pages
+= zone_managed_pages(upper_zone
);
8318 zone
->lowmem_reserve
[j
] = 0;
8320 zone
->lowmem_reserve
[j
] = managed_pages
/ ratio
;
8325 /* update totalreserve_pages */
8326 calculate_totalreserve_pages();
8329 static void __setup_per_zone_wmarks(void)
8331 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
8332 unsigned long lowmem_pages
= 0;
8334 unsigned long flags
;
8336 /* Calculate total number of !ZONE_HIGHMEM pages */
8337 for_each_zone(zone
) {
8338 if (!is_highmem(zone
))
8339 lowmem_pages
+= zone_managed_pages(zone
);
8342 for_each_zone(zone
) {
8345 spin_lock_irqsave(&zone
->lock
, flags
);
8346 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
8347 do_div(tmp
, lowmem_pages
);
8348 if (is_highmem(zone
)) {
8350 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8351 * need highmem pages, so cap pages_min to a small
8354 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8355 * deltas control async page reclaim, and so should
8356 * not be capped for highmem.
8358 unsigned long min_pages
;
8360 min_pages
= zone_managed_pages(zone
) / 1024;
8361 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
8362 zone
->_watermark
[WMARK_MIN
] = min_pages
;
8365 * If it's a lowmem zone, reserve a number of pages
8366 * proportionate to the zone's size.
8368 zone
->_watermark
[WMARK_MIN
] = tmp
;
8372 * Set the kswapd watermarks distance according to the
8373 * scale factor in proportion to available memory, but
8374 * ensure a minimum size on small systems.
8376 tmp
= max_t(u64
, tmp
>> 2,
8377 mult_frac(zone_managed_pages(zone
),
8378 watermark_scale_factor
, 10000));
8380 zone
->watermark_boost
= 0;
8381 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
8382 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
8384 spin_unlock_irqrestore(&zone
->lock
, flags
);
8387 /* update totalreserve_pages */
8388 calculate_totalreserve_pages();
8392 * setup_per_zone_wmarks - called when min_free_kbytes changes
8393 * or when memory is hot-{added|removed}
8395 * Ensures that the watermark[min,low,high] values for each zone are set
8396 * correctly with respect to min_free_kbytes.
8398 void setup_per_zone_wmarks(void)
8401 static DEFINE_SPINLOCK(lock
);
8404 __setup_per_zone_wmarks();
8408 * The watermark size have changed so update the pcpu batch
8409 * and high limits or the limits may be inappropriate.
8412 zone_pcp_update(zone
, 0);
8416 * Initialise min_free_kbytes.
8418 * For small machines we want it small (128k min). For large machines
8419 * we want it large (256MB max). But it is not linear, because network
8420 * bandwidth does not increase linearly with machine size. We use
8422 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8423 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8439 int __meminit
init_per_zone_wmark_min(void)
8441 unsigned long lowmem_kbytes
;
8442 int new_min_free_kbytes
;
8444 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
8445 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
8447 if (new_min_free_kbytes
> user_min_free_kbytes
) {
8448 min_free_kbytes
= new_min_free_kbytes
;
8449 if (min_free_kbytes
< 128)
8450 min_free_kbytes
= 128;
8451 if (min_free_kbytes
> 262144)
8452 min_free_kbytes
= 262144;
8454 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8455 new_min_free_kbytes
, user_min_free_kbytes
);
8457 setup_per_zone_wmarks();
8458 refresh_zone_stat_thresholds();
8459 setup_per_zone_lowmem_reserve();
8462 setup_min_unmapped_ratio();
8463 setup_min_slab_ratio();
8466 khugepaged_min_free_kbytes_update();
8470 postcore_initcall(init_per_zone_wmark_min
)
8473 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8474 * that we can call two helper functions whenever min_free_kbytes
8477 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
8478 void *buffer
, size_t *length
, loff_t
*ppos
)
8482 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8487 user_min_free_kbytes
= min_free_kbytes
;
8488 setup_per_zone_wmarks();
8493 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
8494 void *buffer
, size_t *length
, loff_t
*ppos
)
8498 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8503 setup_per_zone_wmarks();
8509 static void setup_min_unmapped_ratio(void)
8514 for_each_online_pgdat(pgdat
)
8515 pgdat
->min_unmapped_pages
= 0;
8518 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
8519 sysctl_min_unmapped_ratio
) / 100;
8523 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8524 void *buffer
, size_t *length
, loff_t
*ppos
)
8528 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8532 setup_min_unmapped_ratio();
8537 static void setup_min_slab_ratio(void)
8542 for_each_online_pgdat(pgdat
)
8543 pgdat
->min_slab_pages
= 0;
8546 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
8547 sysctl_min_slab_ratio
) / 100;
8550 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8551 void *buffer
, size_t *length
, loff_t
*ppos
)
8555 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8559 setup_min_slab_ratio();
8566 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8567 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8568 * whenever sysctl_lowmem_reserve_ratio changes.
8570 * The reserve ratio obviously has absolutely no relation with the
8571 * minimum watermarks. The lowmem reserve ratio can only make sense
8572 * if in function of the boot time zone sizes.
8574 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8575 void *buffer
, size_t *length
, loff_t
*ppos
)
8579 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8581 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
8582 if (sysctl_lowmem_reserve_ratio
[i
] < 1)
8583 sysctl_lowmem_reserve_ratio
[i
] = 0;
8586 setup_per_zone_lowmem_reserve();
8591 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8592 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8593 * pagelist can have before it gets flushed back to buddy allocator.
8595 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table
*table
,
8596 int write
, void *buffer
, size_t *length
, loff_t
*ppos
)
8599 int old_percpu_pagelist_high_fraction
;
8602 mutex_lock(&pcp_batch_high_lock
);
8603 old_percpu_pagelist_high_fraction
= percpu_pagelist_high_fraction
;
8605 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8606 if (!write
|| ret
< 0)
8609 /* Sanity checking to avoid pcp imbalance */
8610 if (percpu_pagelist_high_fraction
&&
8611 percpu_pagelist_high_fraction
< MIN_PERCPU_PAGELIST_HIGH_FRACTION
) {
8612 percpu_pagelist_high_fraction
= old_percpu_pagelist_high_fraction
;
8618 if (percpu_pagelist_high_fraction
== old_percpu_pagelist_high_fraction
)
8621 for_each_populated_zone(zone
)
8622 zone_set_pageset_high_and_batch(zone
, 0);
8624 mutex_unlock(&pcp_batch_high_lock
);
8628 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8630 * Returns the number of pages that arch has reserved but
8631 * is not known to alloc_large_system_hash().
8633 static unsigned long __init
arch_reserved_kernel_pages(void)
8640 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8641 * machines. As memory size is increased the scale is also increased but at
8642 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8643 * quadruples the scale is increased by one, which means the size of hash table
8644 * only doubles, instead of quadrupling as well.
8645 * Because 32-bit systems cannot have large physical memory, where this scaling
8646 * makes sense, it is disabled on such platforms.
8648 #if __BITS_PER_LONG > 32
8649 #define ADAPT_SCALE_BASE (64ul << 30)
8650 #define ADAPT_SCALE_SHIFT 2
8651 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8655 * allocate a large system hash table from bootmem
8656 * - it is assumed that the hash table must contain an exact power-of-2
8657 * quantity of entries
8658 * - limit is the number of hash buckets, not the total allocation size
8660 void *__init
alloc_large_system_hash(const char *tablename
,
8661 unsigned long bucketsize
,
8662 unsigned long numentries
,
8665 unsigned int *_hash_shift
,
8666 unsigned int *_hash_mask
,
8667 unsigned long low_limit
,
8668 unsigned long high_limit
)
8670 unsigned long long max
= high_limit
;
8671 unsigned long log2qty
, size
;
8677 /* allow the kernel cmdline to have a say */
8679 /* round applicable memory size up to nearest megabyte */
8680 numentries
= nr_kernel_pages
;
8681 numentries
-= arch_reserved_kernel_pages();
8683 /* It isn't necessary when PAGE_SIZE >= 1MB */
8684 if (PAGE_SHIFT
< 20)
8685 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
8687 #if __BITS_PER_LONG > 32
8689 unsigned long adapt
;
8691 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
8692 adapt
<<= ADAPT_SCALE_SHIFT
)
8697 /* limit to 1 bucket per 2^scale bytes of low memory */
8698 if (scale
> PAGE_SHIFT
)
8699 numentries
>>= (scale
- PAGE_SHIFT
);
8701 numentries
<<= (PAGE_SHIFT
- scale
);
8703 /* Make sure we've got at least a 0-order allocation.. */
8704 if (unlikely(flags
& HASH_SMALL
)) {
8705 /* Makes no sense without HASH_EARLY */
8706 WARN_ON(!(flags
& HASH_EARLY
));
8707 if (!(numentries
>> *_hash_shift
)) {
8708 numentries
= 1UL << *_hash_shift
;
8709 BUG_ON(!numentries
);
8711 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
8712 numentries
= PAGE_SIZE
/ bucketsize
;
8714 numentries
= roundup_pow_of_two(numentries
);
8716 /* limit allocation size to 1/16 total memory by default */
8718 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
8719 do_div(max
, bucketsize
);
8721 max
= min(max
, 0x80000000ULL
);
8723 if (numentries
< low_limit
)
8724 numentries
= low_limit
;
8725 if (numentries
> max
)
8728 log2qty
= ilog2(numentries
);
8730 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
8733 size
= bucketsize
<< log2qty
;
8734 if (flags
& HASH_EARLY
) {
8735 if (flags
& HASH_ZERO
)
8736 table
= memblock_alloc(size
, SMP_CACHE_BYTES
);
8738 table
= memblock_alloc_raw(size
,
8740 } else if (get_order(size
) >= MAX_ORDER
|| hashdist
) {
8741 table
= __vmalloc(size
, gfp_flags
);
8743 huge
= is_vm_area_hugepages(table
);
8746 * If bucketsize is not a power-of-two, we may free
8747 * some pages at the end of hash table which
8748 * alloc_pages_exact() automatically does
8750 table
= alloc_pages_exact(size
, gfp_flags
);
8751 kmemleak_alloc(table
, size
, 1, gfp_flags
);
8753 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
8756 panic("Failed to allocate %s hash table\n", tablename
);
8758 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8759 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
,
8760 virt
? (huge
? "vmalloc hugepage" : "vmalloc") : "linear");
8763 *_hash_shift
= log2qty
;
8765 *_hash_mask
= (1 << log2qty
) - 1;
8771 * This function checks whether pageblock includes unmovable pages or not.
8773 * PageLRU check without isolation or lru_lock could race so that
8774 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8775 * check without lock_page also may miss some movable non-lru pages at
8776 * race condition. So you can't expect this function should be exact.
8778 * Returns a page without holding a reference. If the caller wants to
8779 * dereference that page (e.g., dumping), it has to make sure that it
8780 * cannot get removed (e.g., via memory unplug) concurrently.
8783 struct page
*has_unmovable_pages(struct zone
*zone
, struct page
*page
,
8784 int migratetype
, int flags
)
8786 unsigned long iter
= 0;
8787 unsigned long pfn
= page_to_pfn(page
);
8788 unsigned long offset
= pfn
% pageblock_nr_pages
;
8790 if (is_migrate_cma_page(page
)) {
8792 * CMA allocations (alloc_contig_range) really need to mark
8793 * isolate CMA pageblocks even when they are not movable in fact
8794 * so consider them movable here.
8796 if (is_migrate_cma(migratetype
))
8802 for (; iter
< pageblock_nr_pages
- offset
; iter
++) {
8803 page
= pfn_to_page(pfn
+ iter
);
8806 * Both, bootmem allocations and memory holes are marked
8807 * PG_reserved and are unmovable. We can even have unmovable
8808 * allocations inside ZONE_MOVABLE, for example when
8809 * specifying "movablecore".
8811 if (PageReserved(page
))
8815 * If the zone is movable and we have ruled out all reserved
8816 * pages then it should be reasonably safe to assume the rest
8819 if (zone_idx(zone
) == ZONE_MOVABLE
)
8823 * Hugepages are not in LRU lists, but they're movable.
8824 * THPs are on the LRU, but need to be counted as #small pages.
8825 * We need not scan over tail pages because we don't
8826 * handle each tail page individually in migration.
8828 if (PageHuge(page
) || PageTransCompound(page
)) {
8829 struct page
*head
= compound_head(page
);
8830 unsigned int skip_pages
;
8832 if (PageHuge(page
)) {
8833 if (!hugepage_migration_supported(page_hstate(head
)))
8835 } else if (!PageLRU(head
) && !__PageMovable(head
)) {
8839 skip_pages
= compound_nr(head
) - (page
- head
);
8840 iter
+= skip_pages
- 1;
8845 * We can't use page_count without pin a page
8846 * because another CPU can free compound page.
8847 * This check already skips compound tails of THP
8848 * because their page->_refcount is zero at all time.
8850 if (!page_ref_count(page
)) {
8851 if (PageBuddy(page
))
8852 iter
+= (1 << buddy_order(page
)) - 1;
8857 * The HWPoisoned page may be not in buddy system, and
8858 * page_count() is not 0.
8860 if ((flags
& MEMORY_OFFLINE
) && PageHWPoison(page
))
8864 * We treat all PageOffline() pages as movable when offlining
8865 * to give drivers a chance to decrement their reference count
8866 * in MEM_GOING_OFFLINE in order to indicate that these pages
8867 * can be offlined as there are no direct references anymore.
8868 * For actually unmovable PageOffline() where the driver does
8869 * not support this, we will fail later when trying to actually
8870 * move these pages that still have a reference count > 0.
8871 * (false negatives in this function only)
8873 if ((flags
& MEMORY_OFFLINE
) && PageOffline(page
))
8876 if (__PageMovable(page
) || PageLRU(page
))
8880 * If there are RECLAIMABLE pages, we need to check
8881 * it. But now, memory offline itself doesn't call
8882 * shrink_node_slabs() and it still to be fixed.
8889 #ifdef CONFIG_CONTIG_ALLOC
8890 static unsigned long pfn_max_align_down(unsigned long pfn
)
8892 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8893 pageblock_nr_pages
) - 1);
8896 static unsigned long pfn_max_align_up(unsigned long pfn
)
8898 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8899 pageblock_nr_pages
));
8902 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8903 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8904 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8905 static void alloc_contig_dump_pages(struct list_head
*page_list
)
8907 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor
, "migrate failure");
8909 if (DYNAMIC_DEBUG_BRANCH(descriptor
)) {
8913 list_for_each_entry(page
, page_list
, lru
)
8914 dump_page(page
, "migration failure");
8918 static inline void alloc_contig_dump_pages(struct list_head
*page_list
)
8923 /* [start, end) must belong to a single zone. */
8924 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8925 unsigned long start
, unsigned long end
)
8927 /* This function is based on compact_zone() from compaction.c. */
8928 unsigned int nr_reclaimed
;
8929 unsigned long pfn
= start
;
8930 unsigned int tries
= 0;
8932 struct migration_target_control mtc
= {
8933 .nid
= zone_to_nid(cc
->zone
),
8934 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
8937 lru_cache_disable();
8939 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8940 if (fatal_signal_pending(current
)) {
8945 if (list_empty(&cc
->migratepages
)) {
8946 cc
->nr_migratepages
= 0;
8947 ret
= isolate_migratepages_range(cc
, pfn
, end
);
8948 if (ret
&& ret
!= -EAGAIN
)
8950 pfn
= cc
->migrate_pfn
;
8952 } else if (++tries
== 5) {
8957 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8959 cc
->nr_migratepages
-= nr_reclaimed
;
8961 ret
= migrate_pages(&cc
->migratepages
, alloc_migration_target
,
8962 NULL
, (unsigned long)&mtc
, cc
->mode
, MR_CONTIG_RANGE
);
8965 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8966 * to retry again over this error, so do the same here.
8975 alloc_contig_dump_pages(&cc
->migratepages
);
8976 putback_movable_pages(&cc
->migratepages
);
8983 * alloc_contig_range() -- tries to allocate given range of pages
8984 * @start: start PFN to allocate
8985 * @end: one-past-the-last PFN to allocate
8986 * @migratetype: migratetype of the underlying pageblocks (either
8987 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8988 * in range must have the same migratetype and it must
8989 * be either of the two.
8990 * @gfp_mask: GFP mask to use during compaction
8992 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8993 * aligned. The PFN range must belong to a single zone.
8995 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8996 * pageblocks in the range. Once isolated, the pageblocks should not
8997 * be modified by others.
8999 * Return: zero on success or negative error code. On success all
9000 * pages which PFN is in [start, end) are allocated for the caller and
9001 * need to be freed with free_contig_range().
9003 int alloc_contig_range(unsigned long start
, unsigned long end
,
9004 unsigned migratetype
, gfp_t gfp_mask
)
9006 unsigned long outer_start
, outer_end
;
9010 struct compact_control cc
= {
9011 .nr_migratepages
= 0,
9013 .zone
= page_zone(pfn_to_page(start
)),
9014 .mode
= MIGRATE_SYNC
,
9015 .ignore_skip_hint
= true,
9016 .no_set_skip_hint
= true,
9017 .gfp_mask
= current_gfp_context(gfp_mask
),
9018 .alloc_contig
= true,
9020 INIT_LIST_HEAD(&cc
.migratepages
);
9023 * What we do here is we mark all pageblocks in range as
9024 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9025 * have different sizes, and due to the way page allocator
9026 * work, we align the range to biggest of the two pages so
9027 * that page allocator won't try to merge buddies from
9028 * different pageblocks and change MIGRATE_ISOLATE to some
9029 * other migration type.
9031 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9032 * migrate the pages from an unaligned range (ie. pages that
9033 * we are interested in). This will put all the pages in
9034 * range back to page allocator as MIGRATE_ISOLATE.
9036 * When this is done, we take the pages in range from page
9037 * allocator removing them from the buddy system. This way
9038 * page allocator will never consider using them.
9040 * This lets us mark the pageblocks back as
9041 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9042 * aligned range but not in the unaligned, original range are
9043 * put back to page allocator so that buddy can use them.
9046 ret
= start_isolate_page_range(pfn_max_align_down(start
),
9047 pfn_max_align_up(end
), migratetype
, 0);
9051 drain_all_pages(cc
.zone
);
9054 * In case of -EBUSY, we'd like to know which page causes problem.
9055 * So, just fall through. test_pages_isolated() has a tracepoint
9056 * which will report the busy page.
9058 * It is possible that busy pages could become available before
9059 * the call to test_pages_isolated, and the range will actually be
9060 * allocated. So, if we fall through be sure to clear ret so that
9061 * -EBUSY is not accidentally used or returned to caller.
9063 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
9064 if (ret
&& ret
!= -EBUSY
)
9069 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9070 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9071 * more, all pages in [start, end) are free in page allocator.
9072 * What we are going to do is to allocate all pages from
9073 * [start, end) (that is remove them from page allocator).
9075 * The only problem is that pages at the beginning and at the
9076 * end of interesting range may be not aligned with pages that
9077 * page allocator holds, ie. they can be part of higher order
9078 * pages. Because of this, we reserve the bigger range and
9079 * once this is done free the pages we are not interested in.
9081 * We don't have to hold zone->lock here because the pages are
9082 * isolated thus they won't get removed from buddy.
9086 outer_start
= start
;
9087 while (!PageBuddy(pfn_to_page(outer_start
))) {
9088 if (++order
>= MAX_ORDER
) {
9089 outer_start
= start
;
9092 outer_start
&= ~0UL << order
;
9095 if (outer_start
!= start
) {
9096 order
= buddy_order(pfn_to_page(outer_start
));
9099 * outer_start page could be small order buddy page and
9100 * it doesn't include start page. Adjust outer_start
9101 * in this case to report failed page properly
9102 * on tracepoint in test_pages_isolated()
9104 if (outer_start
+ (1UL << order
) <= start
)
9105 outer_start
= start
;
9108 /* Make sure the range is really isolated. */
9109 if (test_pages_isolated(outer_start
, end
, 0)) {
9114 /* Grab isolated pages from freelists. */
9115 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
9121 /* Free head and tail (if any) */
9122 if (start
!= outer_start
)
9123 free_contig_range(outer_start
, start
- outer_start
);
9124 if (end
!= outer_end
)
9125 free_contig_range(end
, outer_end
- end
);
9128 undo_isolate_page_range(pfn_max_align_down(start
),
9129 pfn_max_align_up(end
), migratetype
);
9132 EXPORT_SYMBOL(alloc_contig_range
);
9134 static int __alloc_contig_pages(unsigned long start_pfn
,
9135 unsigned long nr_pages
, gfp_t gfp_mask
)
9137 unsigned long end_pfn
= start_pfn
+ nr_pages
;
9139 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
9143 static bool pfn_range_valid_contig(struct zone
*z
, unsigned long start_pfn
,
9144 unsigned long nr_pages
)
9146 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
9149 for (i
= start_pfn
; i
< end_pfn
; i
++) {
9150 page
= pfn_to_online_page(i
);
9154 if (page_zone(page
) != z
)
9157 if (PageReserved(page
))
9163 static bool zone_spans_last_pfn(const struct zone
*zone
,
9164 unsigned long start_pfn
, unsigned long nr_pages
)
9166 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
9168 return zone_spans_pfn(zone
, last_pfn
);
9172 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9173 * @nr_pages: Number of contiguous pages to allocate
9174 * @gfp_mask: GFP mask to limit search and used during compaction
9176 * @nodemask: Mask for other possible nodes
9178 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9179 * on an applicable zonelist to find a contiguous pfn range which can then be
9180 * tried for allocation with alloc_contig_range(). This routine is intended
9181 * for allocation requests which can not be fulfilled with the buddy allocator.
9183 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9184 * power of two then the alignment is guaranteed to be to the given nr_pages
9185 * (e.g. 1GB request would be aligned to 1GB).
9187 * Allocated pages can be freed with free_contig_range() or by manually calling
9188 * __free_page() on each allocated page.
9190 * Return: pointer to contiguous pages on success, or NULL if not successful.
9192 struct page
*alloc_contig_pages(unsigned long nr_pages
, gfp_t gfp_mask
,
9193 int nid
, nodemask_t
*nodemask
)
9195 unsigned long ret
, pfn
, flags
;
9196 struct zonelist
*zonelist
;
9200 zonelist
= node_zonelist(nid
, gfp_mask
);
9201 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
9202 gfp_zone(gfp_mask
), nodemask
) {
9203 spin_lock_irqsave(&zone
->lock
, flags
);
9205 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
9206 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
9207 if (pfn_range_valid_contig(zone
, pfn
, nr_pages
)) {
9209 * We release the zone lock here because
9210 * alloc_contig_range() will also lock the zone
9211 * at some point. If there's an allocation
9212 * spinning on this lock, it may win the race
9213 * and cause alloc_contig_range() to fail...
9215 spin_unlock_irqrestore(&zone
->lock
, flags
);
9216 ret
= __alloc_contig_pages(pfn
, nr_pages
,
9219 return pfn_to_page(pfn
);
9220 spin_lock_irqsave(&zone
->lock
, flags
);
9224 spin_unlock_irqrestore(&zone
->lock
, flags
);
9228 #endif /* CONFIG_CONTIG_ALLOC */
9230 void free_contig_range(unsigned long pfn
, unsigned long nr_pages
)
9232 unsigned long count
= 0;
9234 for (; nr_pages
--; pfn
++) {
9235 struct page
*page
= pfn_to_page(pfn
);
9237 count
+= page_count(page
) != 1;
9240 WARN(count
!= 0, "%lu pages are still in use!\n", count
);
9242 EXPORT_SYMBOL(free_contig_range
);
9245 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9246 * page high values need to be recalculated.
9248 void zone_pcp_update(struct zone
*zone
, int cpu_online
)
9250 mutex_lock(&pcp_batch_high_lock
);
9251 zone_set_pageset_high_and_batch(zone
, cpu_online
);
9252 mutex_unlock(&pcp_batch_high_lock
);
9256 * Effectively disable pcplists for the zone by setting the high limit to 0
9257 * and draining all cpus. A concurrent page freeing on another CPU that's about
9258 * to put the page on pcplist will either finish before the drain and the page
9259 * will be drained, or observe the new high limit and skip the pcplist.
9261 * Must be paired with a call to zone_pcp_enable().
9263 void zone_pcp_disable(struct zone
*zone
)
9265 mutex_lock(&pcp_batch_high_lock
);
9266 __zone_set_pageset_high_and_batch(zone
, 0, 1);
9267 __drain_all_pages(zone
, true);
9270 void zone_pcp_enable(struct zone
*zone
)
9272 __zone_set_pageset_high_and_batch(zone
, zone
->pageset_high
, zone
->pageset_batch
);
9273 mutex_unlock(&pcp_batch_high_lock
);
9276 void zone_pcp_reset(struct zone
*zone
)
9279 struct per_cpu_zonestat
*pzstats
;
9281 if (zone
->per_cpu_pageset
!= &boot_pageset
) {
9282 for_each_online_cpu(cpu
) {
9283 pzstats
= per_cpu_ptr(zone
->per_cpu_zonestats
, cpu
);
9284 drain_zonestat(zone
, pzstats
);
9286 free_percpu(zone
->per_cpu_pageset
);
9287 free_percpu(zone
->per_cpu_zonestats
);
9288 zone
->per_cpu_pageset
= &boot_pageset
;
9289 zone
->per_cpu_zonestats
= &boot_zonestats
;
9293 #ifdef CONFIG_MEMORY_HOTREMOVE
9295 * All pages in the range must be in a single zone, must not contain holes,
9296 * must span full sections, and must be isolated before calling this function.
9298 void __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
9300 unsigned long pfn
= start_pfn
;
9304 unsigned long flags
;
9306 offline_mem_sections(pfn
, end_pfn
);
9307 zone
= page_zone(pfn_to_page(pfn
));
9308 spin_lock_irqsave(&zone
->lock
, flags
);
9309 while (pfn
< end_pfn
) {
9310 page
= pfn_to_page(pfn
);
9312 * The HWPoisoned page may be not in buddy system, and
9313 * page_count() is not 0.
9315 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
9320 * At this point all remaining PageOffline() pages have a
9321 * reference count of 0 and can simply be skipped.
9323 if (PageOffline(page
)) {
9324 BUG_ON(page_count(page
));
9325 BUG_ON(PageBuddy(page
));
9330 BUG_ON(page_count(page
));
9331 BUG_ON(!PageBuddy(page
));
9332 order
= buddy_order(page
);
9333 del_page_from_free_list(page
, zone
, order
);
9334 pfn
+= (1 << order
);
9336 spin_unlock_irqrestore(&zone
->lock
, flags
);
9340 bool is_free_buddy_page(struct page
*page
)
9342 struct zone
*zone
= page_zone(page
);
9343 unsigned long pfn
= page_to_pfn(page
);
9344 unsigned long flags
;
9347 spin_lock_irqsave(&zone
->lock
, flags
);
9348 for (order
= 0; order
< MAX_ORDER
; order
++) {
9349 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
9351 if (PageBuddy(page_head
) && buddy_order(page_head
) >= order
)
9354 spin_unlock_irqrestore(&zone
->lock
, flags
);
9356 return order
< MAX_ORDER
;
9359 #ifdef CONFIG_MEMORY_FAILURE
9361 * Break down a higher-order page in sub-pages, and keep our target out of
9364 static void break_down_buddy_pages(struct zone
*zone
, struct page
*page
,
9365 struct page
*target
, int low
, int high
,
9368 unsigned long size
= 1 << high
;
9369 struct page
*current_buddy
, *next_page
;
9371 while (high
> low
) {
9375 if (target
>= &page
[size
]) {
9376 next_page
= page
+ size
;
9377 current_buddy
= page
;
9380 current_buddy
= page
+ size
;
9383 if (set_page_guard(zone
, current_buddy
, high
, migratetype
))
9386 if (current_buddy
!= target
) {
9387 add_to_free_list(current_buddy
, zone
, high
, migratetype
);
9388 set_buddy_order(current_buddy
, high
);
9395 * Take a page that will be marked as poisoned off the buddy allocator.
9397 bool take_page_off_buddy(struct page
*page
)
9399 struct zone
*zone
= page_zone(page
);
9400 unsigned long pfn
= page_to_pfn(page
);
9401 unsigned long flags
;
9405 spin_lock_irqsave(&zone
->lock
, flags
);
9406 for (order
= 0; order
< MAX_ORDER
; order
++) {
9407 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
9408 int page_order
= buddy_order(page_head
);
9410 if (PageBuddy(page_head
) && page_order
>= order
) {
9411 unsigned long pfn_head
= page_to_pfn(page_head
);
9412 int migratetype
= get_pfnblock_migratetype(page_head
,
9415 del_page_from_free_list(page_head
, zone
, page_order
);
9416 break_down_buddy_pages(zone
, page_head
, page
, 0,
9417 page_order
, migratetype
);
9418 if (!is_migrate_isolate(migratetype
))
9419 __mod_zone_freepage_state(zone
, -1, migratetype
);
9423 if (page_count(page_head
) > 0)
9426 spin_unlock_irqrestore(&zone
->lock
, flags
);