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/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/kmsan.h>
31 #include <linux/module.h>
32 #include <linux/suspend.h>
33 #include <linux/pagevec.h>
34 #include <linux/blkdev.h>
35 #include <linux/slab.h>
36 #include <linux/ratelimit.h>
37 #include <linux/oom.h>
38 #include <linux/topology.h>
39 #include <linux/sysctl.h>
40 #include <linux/cpu.h>
41 #include <linux/cpuset.h>
42 #include <linux/memory_hotplug.h>
43 #include <linux/nodemask.h>
44 #include <linux/vmalloc.h>
45 #include <linux/vmstat.h>
46 #include <linux/mempolicy.h>
47 #include <linux/memremap.h>
48 #include <linux/stop_machine.h>
49 #include <linux/random.h>
50 #include <linux/sort.h>
51 #include <linux/pfn.h>
52 #include <linux/backing-dev.h>
53 #include <linux/fault-inject.h>
54 #include <linux/page-isolation.h>
55 #include <linux/debugobjects.h>
56 #include <linux/kmemleak.h>
57 #include <linux/compaction.h>
58 #include <trace/events/kmem.h>
59 #include <trace/events/oom.h>
60 #include <linux/prefetch.h>
61 #include <linux/mm_inline.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/migrate.h>
64 #include <linux/hugetlb.h>
65 #include <linux/sched/rt.h>
66 #include <linux/sched/mm.h>
67 #include <linux/page_owner.h>
68 #include <linux/page_table_check.h>
69 #include <linux/kthread.h>
70 #include <linux/memcontrol.h>
71 #include <linux/ftrace.h>
72 #include <linux/lockdep.h>
73 #include <linux/nmi.h>
74 #include <linux/psi.h>
75 #include <linux/khugepaged.h>
76 #include <linux/delayacct.h>
77 #include <asm/sections.h>
78 #include <asm/tlbflush.h>
79 #include <asm/div64.h>
82 #include "page_reporting.h"
85 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
86 typedef int __bitwise fpi_t
;
88 /* No special request */
89 #define FPI_NONE ((__force fpi_t)0)
92 * Skip free page reporting notification for the (possibly merged) page.
93 * This does not hinder free page reporting from grabbing the page,
94 * reporting it and marking it "reported" - it only skips notifying
95 * the free page reporting infrastructure about a newly freed page. For
96 * example, used when temporarily pulling a page from a freelist and
97 * putting it back unmodified.
99 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
102 * Place the (possibly merged) page to the tail of the freelist. Will ignore
103 * page shuffling (relevant code - e.g., memory onlining - is expected to
104 * shuffle the whole zone).
106 * Note: No code should rely on this flag for correctness - it's purely
107 * to allow for optimizations when handing back either fresh pages
108 * (memory onlining) or untouched pages (page isolation, free page
111 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
113 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
114 static DEFINE_MUTEX(pcp_batch_high_lock
);
115 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
117 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
119 * On SMP, spin_trylock is sufficient protection.
120 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
122 #define pcp_trylock_prepare(flags) do { } while (0)
123 #define pcp_trylock_finish(flag) do { } while (0)
126 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
127 #define pcp_trylock_prepare(flags) local_irq_save(flags)
128 #define pcp_trylock_finish(flags) local_irq_restore(flags)
132 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
133 * a migration causing the wrong PCP to be locked and remote memory being
134 * potentially allocated, pin the task to the CPU for the lookup+lock.
135 * preempt_disable is used on !RT because it is faster than migrate_disable.
136 * migrate_disable is used on RT because otherwise RT spinlock usage is
137 * interfered with and a high priority task cannot preempt the allocator.
139 #ifndef CONFIG_PREEMPT_RT
140 #define pcpu_task_pin() preempt_disable()
141 #define pcpu_task_unpin() preempt_enable()
143 #define pcpu_task_pin() migrate_disable()
144 #define pcpu_task_unpin() migrate_enable()
148 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
149 * Return value should be used with equivalent unlock helper.
151 #define pcpu_spin_lock(type, member, ptr) \
155 _ret = this_cpu_ptr(ptr); \
156 spin_lock(&_ret->member); \
160 #define pcpu_spin_trylock(type, member, ptr) \
164 _ret = this_cpu_ptr(ptr); \
165 if (!spin_trylock(&_ret->member)) { \
172 #define pcpu_spin_unlock(member, ptr) \
174 spin_unlock(&ptr->member); \
178 /* struct per_cpu_pages specific helpers. */
179 #define pcp_spin_lock(ptr) \
180 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
182 #define pcp_spin_trylock(ptr) \
183 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
185 #define pcp_spin_unlock(ptr) \
186 pcpu_spin_unlock(lock, ptr)
188 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
189 DEFINE_PER_CPU(int, numa_node
);
190 EXPORT_PER_CPU_SYMBOL(numa_node
);
193 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
195 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
197 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
198 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
199 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
200 * defined in <linux/topology.h>.
202 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
203 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
206 static DEFINE_MUTEX(pcpu_drain_mutex
);
208 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
209 volatile unsigned long latent_entropy __latent_entropy
;
210 EXPORT_SYMBOL(latent_entropy
);
214 * Array of node states.
216 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
217 [N_POSSIBLE
] = NODE_MASK_ALL
,
218 [N_ONLINE
] = { { [0] = 1UL } },
220 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
221 #ifdef CONFIG_HIGHMEM
222 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
224 [N_MEMORY
] = { { [0] = 1UL } },
225 [N_CPU
] = { { [0] = 1UL } },
228 EXPORT_SYMBOL(node_states
);
230 atomic_long_t _totalram_pages __read_mostly
;
231 EXPORT_SYMBOL(_totalram_pages
);
232 unsigned long totalreserve_pages __read_mostly
;
233 unsigned long totalcma_pages __read_mostly
;
235 int percpu_pagelist_high_fraction
;
236 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
237 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON
, init_on_alloc
);
238 EXPORT_SYMBOL(init_on_alloc
);
240 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON
, init_on_free
);
241 EXPORT_SYMBOL(init_on_free
);
244 * A cached value of the page's pageblock's migratetype, used when the page is
245 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
246 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
247 * Also the migratetype set in the page does not necessarily match the pcplist
248 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
249 * other index - this ensures that it will be put on the correct CMA freelist.
251 static inline int get_pcppage_migratetype(struct page
*page
)
256 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
258 page
->index
= migratetype
;
261 #ifdef CONFIG_PM_SLEEP
263 * The following functions are used by the suspend/hibernate code to temporarily
264 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
265 * while devices are suspended. To avoid races with the suspend/hibernate code,
266 * they should always be called with system_transition_mutex held
267 * (gfp_allowed_mask also should only be modified with system_transition_mutex
268 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
269 * with that modification).
272 static gfp_t saved_gfp_mask
;
274 void pm_restore_gfp_mask(void)
276 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
277 if (saved_gfp_mask
) {
278 gfp_allowed_mask
= saved_gfp_mask
;
283 void pm_restrict_gfp_mask(void)
285 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
286 WARN_ON(saved_gfp_mask
);
287 saved_gfp_mask
= gfp_allowed_mask
;
288 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
291 bool pm_suspended_storage(void)
293 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
297 #endif /* CONFIG_PM_SLEEP */
299 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
300 unsigned int pageblock_order __read_mostly
;
303 static void __free_pages_ok(struct page
*page
, unsigned int order
,
307 * results with 256, 32 in the lowmem_reserve sysctl:
308 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
309 * 1G machine -> (16M dma, 784M normal, 224M high)
310 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
311 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
312 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
314 * TBD: should special case ZONE_DMA32 machines here - in those we normally
315 * don't need any ZONE_NORMAL reservation
317 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
318 #ifdef CONFIG_ZONE_DMA
321 #ifdef CONFIG_ZONE_DMA32
325 #ifdef CONFIG_HIGHMEM
331 char * const zone_names
[MAX_NR_ZONES
] = {
332 #ifdef CONFIG_ZONE_DMA
335 #ifdef CONFIG_ZONE_DMA32
339 #ifdef CONFIG_HIGHMEM
343 #ifdef CONFIG_ZONE_DEVICE
348 const char * const migratetype_names
[MIGRATE_TYPES
] = {
356 #ifdef CONFIG_MEMORY_ISOLATION
361 compound_page_dtor
* const compound_page_dtors
[NR_COMPOUND_DTORS
] = {
362 [NULL_COMPOUND_DTOR
] = NULL
,
363 [COMPOUND_PAGE_DTOR
] = free_compound_page
,
364 #ifdef CONFIG_HUGETLB_PAGE
365 [HUGETLB_PAGE_DTOR
] = free_huge_page
,
367 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
368 [TRANSHUGE_PAGE_DTOR
] = free_transhuge_page
,
372 int min_free_kbytes
= 1024;
373 int user_min_free_kbytes
= -1;
374 int watermark_boost_factor __read_mostly
= 15000;
375 int watermark_scale_factor
= 10;
377 bool mirrored_kernelcore __initdata_memblock
;
379 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
381 EXPORT_SYMBOL(movable_zone
);
384 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
385 unsigned int nr_online_nodes __read_mostly
= 1;
386 EXPORT_SYMBOL(nr_node_ids
);
387 EXPORT_SYMBOL(nr_online_nodes
);
390 int page_group_by_mobility_disabled __read_mostly
;
392 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
394 * During boot we initialize deferred pages on-demand, as needed, but once
395 * page_alloc_init_late() has finished, the deferred pages are all initialized,
396 * and we can permanently disable that path.
398 DEFINE_STATIC_KEY_TRUE(deferred_pages
);
400 static inline bool deferred_pages_enabled(void)
402 return static_branch_unlikely(&deferred_pages
);
406 * deferred_grow_zone() is __init, but it is called from
407 * get_page_from_freelist() during early boot until deferred_pages permanently
408 * disables this call. This is why we have refdata wrapper to avoid warning,
409 * and to ensure that the function body gets unloaded.
412 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
414 return deferred_grow_zone(zone
, order
);
417 static inline bool deferred_pages_enabled(void)
421 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
423 /* Return a pointer to the bitmap storing bits affecting a block of pages */
424 static inline unsigned long *get_pageblock_bitmap(const struct page
*page
,
427 #ifdef CONFIG_SPARSEMEM
428 return section_to_usemap(__pfn_to_section(pfn
));
430 return page_zone(page
)->pageblock_flags
;
431 #endif /* CONFIG_SPARSEMEM */
434 static inline int pfn_to_bitidx(const struct page
*page
, unsigned long pfn
)
436 #ifdef CONFIG_SPARSEMEM
437 pfn
&= (PAGES_PER_SECTION
-1);
439 pfn
= pfn
- pageblock_start_pfn(page_zone(page
)->zone_start_pfn
);
440 #endif /* CONFIG_SPARSEMEM */
441 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
444 static __always_inline
445 unsigned long __get_pfnblock_flags_mask(const struct page
*page
,
449 unsigned long *bitmap
;
450 unsigned long bitidx
, word_bitidx
;
453 bitmap
= get_pageblock_bitmap(page
, pfn
);
454 bitidx
= pfn_to_bitidx(page
, pfn
);
455 word_bitidx
= bitidx
/ BITS_PER_LONG
;
456 bitidx
&= (BITS_PER_LONG
-1);
458 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
459 * a consistent read of the memory array, so that results, even though
460 * racy, are not corrupted.
462 word
= READ_ONCE(bitmap
[word_bitidx
]);
463 return (word
>> bitidx
) & mask
;
467 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
468 * @page: The page within the block of interest
469 * @pfn: The target page frame number
470 * @mask: mask of bits that the caller is interested in
472 * Return: pageblock_bits flags
474 unsigned long get_pfnblock_flags_mask(const struct page
*page
,
475 unsigned long pfn
, unsigned long mask
)
477 return __get_pfnblock_flags_mask(page
, pfn
, mask
);
480 static __always_inline
int get_pfnblock_migratetype(const struct page
*page
,
483 return __get_pfnblock_flags_mask(page
, pfn
, MIGRATETYPE_MASK
);
487 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
488 * @page: The page within the block of interest
489 * @flags: The flags to set
490 * @pfn: The target page frame number
491 * @mask: mask of bits that the caller is interested in
493 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
497 unsigned long *bitmap
;
498 unsigned long bitidx
, word_bitidx
;
501 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
502 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
504 bitmap
= get_pageblock_bitmap(page
, pfn
);
505 bitidx
= pfn_to_bitidx(page
, pfn
);
506 word_bitidx
= bitidx
/ BITS_PER_LONG
;
507 bitidx
&= (BITS_PER_LONG
-1);
509 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
514 word
= READ_ONCE(bitmap
[word_bitidx
]);
516 } while (!try_cmpxchg(&bitmap
[word_bitidx
], &word
, (word
& ~mask
) | flags
));
519 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
521 if (unlikely(page_group_by_mobility_disabled
&&
522 migratetype
< MIGRATE_PCPTYPES
))
523 migratetype
= MIGRATE_UNMOVABLE
;
525 set_pfnblock_flags_mask(page
, (unsigned long)migratetype
,
526 page_to_pfn(page
), MIGRATETYPE_MASK
);
529 #ifdef CONFIG_DEBUG_VM
530 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
534 unsigned long pfn
= page_to_pfn(page
);
535 unsigned long sp
, start_pfn
;
538 seq
= zone_span_seqbegin(zone
);
539 start_pfn
= zone
->zone_start_pfn
;
540 sp
= zone
->spanned_pages
;
541 if (!zone_spans_pfn(zone
, pfn
))
543 } while (zone_span_seqretry(zone
, seq
));
546 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
547 pfn
, zone_to_nid(zone
), zone
->name
,
548 start_pfn
, start_pfn
+ sp
);
553 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
555 if (zone
!= page_zone(page
))
561 * Temporary debugging check for pages not lying within a given zone.
563 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
565 if (page_outside_zone_boundaries(zone
, page
))
567 if (!page_is_consistent(zone
, page
))
573 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
579 static void bad_page(struct page
*page
, const char *reason
)
581 static unsigned long resume
;
582 static unsigned long nr_shown
;
583 static unsigned long nr_unshown
;
586 * Allow a burst of 60 reports, then keep quiet for that minute;
587 * or allow a steady drip of one report per second.
589 if (nr_shown
== 60) {
590 if (time_before(jiffies
, resume
)) {
596 "BUG: Bad page state: %lu messages suppressed\n",
603 resume
= jiffies
+ 60 * HZ
;
605 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
606 current
->comm
, page_to_pfn(page
));
607 dump_page(page
, reason
);
612 /* Leave bad fields for debug, except PageBuddy could make trouble */
613 page_mapcount_reset(page
); /* remove PageBuddy */
614 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
617 static inline unsigned int order_to_pindex(int migratetype
, int order
)
621 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
622 if (order
> PAGE_ALLOC_COSTLY_ORDER
) {
623 VM_BUG_ON(order
!= pageblock_order
);
624 return NR_LOWORDER_PCP_LISTS
;
627 VM_BUG_ON(order
> PAGE_ALLOC_COSTLY_ORDER
);
630 return (MIGRATE_PCPTYPES
* base
) + migratetype
;
633 static inline int pindex_to_order(unsigned int pindex
)
635 int order
= pindex
/ MIGRATE_PCPTYPES
;
637 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
638 if (pindex
== NR_LOWORDER_PCP_LISTS
)
639 order
= pageblock_order
;
641 VM_BUG_ON(order
> PAGE_ALLOC_COSTLY_ORDER
);
647 static inline bool pcp_allowed_order(unsigned int order
)
649 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
651 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
652 if (order
== pageblock_order
)
658 static inline void free_the_page(struct page
*page
, unsigned int order
)
660 if (pcp_allowed_order(order
)) /* Via pcp? */
661 free_unref_page(page
, order
);
663 __free_pages_ok(page
, order
, FPI_NONE
);
667 * Higher-order pages are called "compound pages". They are structured thusly:
669 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
671 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
672 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
674 * The first tail page's ->compound_dtor holds the offset in array of compound
675 * page destructors. See compound_page_dtors.
677 * The first tail page's ->compound_order holds the order of allocation.
678 * This usage means that zero-order pages may not be compound.
681 void free_compound_page(struct page
*page
)
683 mem_cgroup_uncharge(page_folio(page
));
684 free_the_page(page
, compound_order(page
));
687 void prep_compound_page(struct page
*page
, unsigned int order
)
690 int nr_pages
= 1 << order
;
693 for (i
= 1; i
< nr_pages
; i
++)
694 prep_compound_tail(page
, i
);
696 prep_compound_head(page
, order
);
699 void destroy_large_folio(struct folio
*folio
)
701 enum compound_dtor_id dtor
= folio
->_folio_dtor
;
703 VM_BUG_ON_FOLIO(dtor
>= NR_COMPOUND_DTORS
, folio
);
704 compound_page_dtors
[dtor
](&folio
->page
);
707 #ifdef CONFIG_DEBUG_PAGEALLOC
708 unsigned int _debug_guardpage_minorder
;
710 bool _debug_pagealloc_enabled_early __read_mostly
711 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
712 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
713 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
714 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
716 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
718 static int __init
early_debug_pagealloc(char *buf
)
720 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
722 early_param("debug_pagealloc", early_debug_pagealloc
);
724 static int __init
debug_guardpage_minorder_setup(char *buf
)
728 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
729 pr_err("Bad debug_guardpage_minorder value\n");
732 _debug_guardpage_minorder
= res
;
733 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
736 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
738 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
739 unsigned int order
, int migratetype
)
741 if (!debug_guardpage_enabled())
744 if (order
>= debug_guardpage_minorder())
747 __SetPageGuard(page
);
748 INIT_LIST_HEAD(&page
->buddy_list
);
749 set_page_private(page
, order
);
750 /* Guard pages are not available for any usage */
751 if (!is_migrate_isolate(migratetype
))
752 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
757 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
758 unsigned int order
, int migratetype
)
760 if (!debug_guardpage_enabled())
763 __ClearPageGuard(page
);
765 set_page_private(page
, 0);
766 if (!is_migrate_isolate(migratetype
))
767 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
770 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
771 unsigned int order
, int migratetype
) { return false; }
772 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
773 unsigned int order
, int migratetype
) {}
776 static inline void set_buddy_order(struct page
*page
, unsigned int order
)
778 set_page_private(page
, order
);
779 __SetPageBuddy(page
);
782 #ifdef CONFIG_COMPACTION
783 static inline struct capture_control
*task_capc(struct zone
*zone
)
785 struct capture_control
*capc
= current
->capture_control
;
787 return unlikely(capc
) &&
788 !(current
->flags
& PF_KTHREAD
) &&
790 capc
->cc
->zone
== zone
? capc
: NULL
;
794 compaction_capture(struct capture_control
*capc
, struct page
*page
,
795 int order
, int migratetype
)
797 if (!capc
|| order
!= capc
->cc
->order
)
800 /* Do not accidentally pollute CMA or isolated regions*/
801 if (is_migrate_cma(migratetype
) ||
802 is_migrate_isolate(migratetype
))
806 * Do not let lower order allocations pollute a movable pageblock.
807 * This might let an unmovable request use a reclaimable pageblock
808 * and vice-versa but no more than normal fallback logic which can
809 * have trouble finding a high-order free page.
811 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
819 static inline struct capture_control
*task_capc(struct zone
*zone
)
825 compaction_capture(struct capture_control
*capc
, struct page
*page
,
826 int order
, int migratetype
)
830 #endif /* CONFIG_COMPACTION */
832 /* Used for pages not on another list */
833 static inline void add_to_free_list(struct page
*page
, struct zone
*zone
,
834 unsigned int order
, int migratetype
)
836 struct free_area
*area
= &zone
->free_area
[order
];
838 list_add(&page
->buddy_list
, &area
->free_list
[migratetype
]);
842 /* Used for pages not on another list */
843 static inline void add_to_free_list_tail(struct page
*page
, struct zone
*zone
,
844 unsigned int order
, int migratetype
)
846 struct free_area
*area
= &zone
->free_area
[order
];
848 list_add_tail(&page
->buddy_list
, &area
->free_list
[migratetype
]);
853 * Used for pages which are on another list. Move the pages to the tail
854 * of the list - so the moved pages won't immediately be considered for
855 * allocation again (e.g., optimization for memory onlining).
857 static inline void move_to_free_list(struct page
*page
, struct zone
*zone
,
858 unsigned int order
, int migratetype
)
860 struct free_area
*area
= &zone
->free_area
[order
];
862 list_move_tail(&page
->buddy_list
, &area
->free_list
[migratetype
]);
865 static inline void del_page_from_free_list(struct page
*page
, struct zone
*zone
,
868 /* clear reported state and update reported page count */
869 if (page_reported(page
))
870 __ClearPageReported(page
);
872 list_del(&page
->buddy_list
);
873 __ClearPageBuddy(page
);
874 set_page_private(page
, 0);
875 zone
->free_area
[order
].nr_free
--;
878 static inline struct page
*get_page_from_free_area(struct free_area
*area
,
881 return list_first_entry_or_null(&area
->free_list
[migratetype
],
886 * If this is not the largest possible page, check if the buddy
887 * of the next-highest order is free. If it is, it's possible
888 * that pages are being freed that will coalesce soon. In case,
889 * that is happening, add the free page to the tail of the list
890 * so it's less likely to be used soon and more likely to be merged
891 * as a higher order page
894 buddy_merge_likely(unsigned long pfn
, unsigned long buddy_pfn
,
895 struct page
*page
, unsigned int order
)
897 unsigned long higher_page_pfn
;
898 struct page
*higher_page
;
900 if (order
>= MAX_ORDER
- 1)
903 higher_page_pfn
= buddy_pfn
& pfn
;
904 higher_page
= page
+ (higher_page_pfn
- pfn
);
906 return find_buddy_page_pfn(higher_page
, higher_page_pfn
, order
+ 1,
911 * Freeing function for a buddy system allocator.
913 * The concept of a buddy system is to maintain direct-mapped table
914 * (containing bit values) for memory blocks of various "orders".
915 * The bottom level table contains the map for the smallest allocatable
916 * units of memory (here, pages), and each level above it describes
917 * pairs of units from the levels below, hence, "buddies".
918 * At a high level, all that happens here is marking the table entry
919 * at the bottom level available, and propagating the changes upward
920 * as necessary, plus some accounting needed to play nicely with other
921 * parts of the VM system.
922 * At each level, we keep a list of pages, which are heads of continuous
923 * free pages of length of (1 << order) and marked with PageBuddy.
924 * Page's order is recorded in page_private(page) field.
925 * So when we are allocating or freeing one, we can derive the state of the
926 * other. That is, if we allocate a small block, and both were
927 * free, the remainder of the region must be split into blocks.
928 * If a block is freed, and its buddy is also free, then this
929 * triggers coalescing into a block of larger size.
934 static inline void __free_one_page(struct page
*page
,
936 struct zone
*zone
, unsigned int order
,
937 int migratetype
, fpi_t fpi_flags
)
939 struct capture_control
*capc
= task_capc(zone
);
940 unsigned long buddy_pfn
= 0;
941 unsigned long combined_pfn
;
945 VM_BUG_ON(!zone_is_initialized(zone
));
946 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
948 VM_BUG_ON(migratetype
== -1);
949 if (likely(!is_migrate_isolate(migratetype
)))
950 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
952 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
953 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
955 while (order
< MAX_ORDER
) {
956 if (compaction_capture(capc
, page
, order
, migratetype
)) {
957 __mod_zone_freepage_state(zone
, -(1 << order
),
962 buddy
= find_buddy_page_pfn(page
, pfn
, order
, &buddy_pfn
);
966 if (unlikely(order
>= pageblock_order
)) {
968 * We want to prevent merge between freepages on pageblock
969 * without fallbacks and normal pageblock. Without this,
970 * pageblock isolation could cause incorrect freepage or CMA
971 * accounting or HIGHATOMIC accounting.
973 int buddy_mt
= get_pageblock_migratetype(buddy
);
975 if (migratetype
!= buddy_mt
976 && (!migratetype_is_mergeable(migratetype
) ||
977 !migratetype_is_mergeable(buddy_mt
)))
982 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
983 * merge with it and move up one order.
985 if (page_is_guard(buddy
))
986 clear_page_guard(zone
, buddy
, order
, migratetype
);
988 del_page_from_free_list(buddy
, zone
, order
);
989 combined_pfn
= buddy_pfn
& pfn
;
990 page
= page
+ (combined_pfn
- pfn
);
996 set_buddy_order(page
, order
);
998 if (fpi_flags
& FPI_TO_TAIL
)
1000 else if (is_shuffle_order(order
))
1001 to_tail
= shuffle_pick_tail();
1003 to_tail
= buddy_merge_likely(pfn
, buddy_pfn
, page
, order
);
1006 add_to_free_list_tail(page
, zone
, order
, migratetype
);
1008 add_to_free_list(page
, zone
, order
, migratetype
);
1010 /* Notify page reporting subsystem of freed page */
1011 if (!(fpi_flags
& FPI_SKIP_REPORT_NOTIFY
))
1012 page_reporting_notify_free(order
);
1016 * split_free_page() -- split a free page at split_pfn_offset
1017 * @free_page: the original free page
1018 * @order: the order of the page
1019 * @split_pfn_offset: split offset within the page
1021 * Return -ENOENT if the free page is changed, otherwise 0
1023 * It is used when the free page crosses two pageblocks with different migratetypes
1024 * at split_pfn_offset within the page. The split free page will be put into
1025 * separate migratetype lists afterwards. Otherwise, the function achieves
1028 int split_free_page(struct page
*free_page
,
1029 unsigned int order
, unsigned long split_pfn_offset
)
1031 struct zone
*zone
= page_zone(free_page
);
1032 unsigned long free_page_pfn
= page_to_pfn(free_page
);
1034 unsigned long flags
;
1035 int free_page_order
;
1039 if (split_pfn_offset
== 0)
1042 spin_lock_irqsave(&zone
->lock
, flags
);
1044 if (!PageBuddy(free_page
) || buddy_order(free_page
) != order
) {
1049 mt
= get_pageblock_migratetype(free_page
);
1050 if (likely(!is_migrate_isolate(mt
)))
1051 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1053 del_page_from_free_list(free_page
, zone
, order
);
1054 for (pfn
= free_page_pfn
;
1055 pfn
< free_page_pfn
+ (1UL << order
);) {
1056 int mt
= get_pfnblock_migratetype(pfn_to_page(pfn
), pfn
);
1058 free_page_order
= min_t(unsigned int,
1059 pfn
? __ffs(pfn
) : order
,
1060 __fls(split_pfn_offset
));
1061 __free_one_page(pfn_to_page(pfn
), pfn
, zone
, free_page_order
,
1063 pfn
+= 1UL << free_page_order
;
1064 split_pfn_offset
-= (1UL << free_page_order
);
1065 /* we have done the first part, now switch to second part */
1066 if (split_pfn_offset
== 0)
1067 split_pfn_offset
= (1UL << order
) - (pfn
- free_page_pfn
);
1070 spin_unlock_irqrestore(&zone
->lock
, flags
);
1074 * A bad page could be due to a number of fields. Instead of multiple branches,
1075 * try and check multiple fields with one check. The caller must do a detailed
1076 * check if necessary.
1078 static inline bool page_expected_state(struct page
*page
,
1079 unsigned long check_flags
)
1081 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1084 if (unlikely((unsigned long)page
->mapping
|
1085 page_ref_count(page
) |
1089 (page
->flags
& check_flags
)))
1095 static const char *page_bad_reason(struct page
*page
, unsigned long flags
)
1097 const char *bad_reason
= NULL
;
1099 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1100 bad_reason
= "nonzero mapcount";
1101 if (unlikely(page
->mapping
!= NULL
))
1102 bad_reason
= "non-NULL mapping";
1103 if (unlikely(page_ref_count(page
) != 0))
1104 bad_reason
= "nonzero _refcount";
1105 if (unlikely(page
->flags
& flags
)) {
1106 if (flags
== PAGE_FLAGS_CHECK_AT_PREP
)
1107 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1109 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1112 if (unlikely(page
->memcg_data
))
1113 bad_reason
= "page still charged to cgroup";
1118 static void free_page_is_bad_report(struct page
*page
)
1121 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_FREE
));
1124 static inline bool free_page_is_bad(struct page
*page
)
1126 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1129 /* Something has gone sideways, find it */
1130 free_page_is_bad_report(page
);
1134 static int free_tail_page_prepare(struct page
*head_page
, struct page
*page
)
1136 struct folio
*folio
= (struct folio
*)head_page
;
1140 * We rely page->lru.next never has bit 0 set, unless the page
1141 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1143 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1145 if (!static_branch_unlikely(&check_pages_enabled
)) {
1149 switch (page
- head_page
) {
1151 /* the first tail page: these may be in place of ->mapping */
1152 if (unlikely(folio_entire_mapcount(folio
))) {
1153 bad_page(page
, "nonzero entire_mapcount");
1156 if (unlikely(atomic_read(&folio
->_nr_pages_mapped
))) {
1157 bad_page(page
, "nonzero nr_pages_mapped");
1160 if (unlikely(atomic_read(&folio
->_pincount
))) {
1161 bad_page(page
, "nonzero pincount");
1167 * the second tail page: ->mapping is
1168 * deferred_list.next -- ignore value.
1172 if (page
->mapping
!= TAIL_MAPPING
) {
1173 bad_page(page
, "corrupted mapping in tail page");
1178 if (unlikely(!PageTail(page
))) {
1179 bad_page(page
, "PageTail not set");
1182 if (unlikely(compound_head(page
) != head_page
)) {
1183 bad_page(page
, "compound_head not consistent");
1188 page
->mapping
= NULL
;
1189 clear_compound_head(page
);
1194 * Skip KASAN memory poisoning when either:
1196 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1197 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1198 * using page tags instead (see below).
1199 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1200 * that error detection is disabled for accesses via the page address.
1202 * Pages will have match-all tags in the following circumstances:
1204 * 1. Pages are being initialized for the first time, including during deferred
1205 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1206 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1207 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1208 * 3. The allocation was excluded from being checked due to sampling,
1209 * see the call to kasan_unpoison_pages.
1211 * Poisoning pages during deferred memory init will greatly lengthen the
1212 * process and cause problem in large memory systems as the deferred pages
1213 * initialization is done with interrupt disabled.
1215 * Assuming that there will be no reference to those newly initialized
1216 * pages before they are ever allocated, this should have no effect on
1217 * KASAN memory tracking as the poison will be properly inserted at page
1218 * allocation time. The only corner case is when pages are allocated by
1219 * on-demand allocation and then freed again before the deferred pages
1220 * initialization is done, but this is not likely to happen.
1222 static inline bool should_skip_kasan_poison(struct page
*page
, fpi_t fpi_flags
)
1224 if (IS_ENABLED(CONFIG_KASAN_GENERIC
))
1225 return deferred_pages_enabled();
1227 return page_kasan_tag(page
) == 0xff;
1230 static void kernel_init_pages(struct page
*page
, int numpages
)
1234 /* s390's use of memset() could override KASAN redzones. */
1235 kasan_disable_current();
1236 for (i
= 0; i
< numpages
; i
++)
1237 clear_highpage_kasan_tagged(page
+ i
);
1238 kasan_enable_current();
1241 static __always_inline
bool free_pages_prepare(struct page
*page
,
1242 unsigned int order
, fpi_t fpi_flags
)
1245 bool skip_kasan_poison
= should_skip_kasan_poison(page
, fpi_flags
);
1246 bool init
= want_init_on_free();
1248 VM_BUG_ON_PAGE(PageTail(page
), page
);
1250 trace_mm_page_free(page
, order
);
1251 kmsan_free_page(page
, order
);
1253 if (unlikely(PageHWPoison(page
)) && !order
) {
1255 * Do not let hwpoison pages hit pcplists/buddy
1256 * Untie memcg state and reset page's owner
1258 if (memcg_kmem_online() && PageMemcgKmem(page
))
1259 __memcg_kmem_uncharge_page(page
, order
);
1260 reset_page_owner(page
, order
);
1261 page_table_check_free(page
, order
);
1266 * Check tail pages before head page information is cleared to
1267 * avoid checking PageCompound for order-0 pages.
1269 if (unlikely(order
)) {
1270 bool compound
= PageCompound(page
);
1273 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1276 ClearPageHasHWPoisoned(page
);
1277 for (i
= 1; i
< (1 << order
); i
++) {
1279 bad
+= free_tail_page_prepare(page
, page
+ i
);
1280 if (is_check_pages_enabled()) {
1281 if (free_page_is_bad(page
+ i
)) {
1286 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1289 if (PageMappingFlags(page
))
1290 page
->mapping
= NULL
;
1291 if (memcg_kmem_online() && PageMemcgKmem(page
))
1292 __memcg_kmem_uncharge_page(page
, order
);
1293 if (is_check_pages_enabled()) {
1294 if (free_page_is_bad(page
))
1300 page_cpupid_reset_last(page
);
1301 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1302 reset_page_owner(page
, order
);
1303 page_table_check_free(page
, order
);
1305 if (!PageHighMem(page
)) {
1306 debug_check_no_locks_freed(page_address(page
),
1307 PAGE_SIZE
<< order
);
1308 debug_check_no_obj_freed(page_address(page
),
1309 PAGE_SIZE
<< order
);
1312 kernel_poison_pages(page
, 1 << order
);
1315 * As memory initialization might be integrated into KASAN,
1316 * KASAN poisoning and memory initialization code must be
1317 * kept together to avoid discrepancies in behavior.
1319 * With hardware tag-based KASAN, memory tags must be set before the
1320 * page becomes unavailable via debug_pagealloc or arch_free_page.
1322 if (!skip_kasan_poison
) {
1323 kasan_poison_pages(page
, order
, init
);
1325 /* Memory is already initialized if KASAN did it internally. */
1326 if (kasan_has_integrated_init())
1330 kernel_init_pages(page
, 1 << order
);
1333 * arch_free_page() can make the page's contents inaccessible. s390
1334 * does this. So nothing which can access the page's contents should
1335 * happen after this.
1337 arch_free_page(page
, order
);
1339 debug_pagealloc_unmap_pages(page
, 1 << order
);
1345 * Frees a number of pages from the PCP lists
1346 * Assumes all pages on list are in same zone.
1347 * count is the number of pages to free.
1349 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1350 struct per_cpu_pages
*pcp
,
1353 unsigned long flags
;
1355 int max_pindex
= NR_PCP_LISTS
- 1;
1357 bool isolated_pageblocks
;
1361 * Ensure proper count is passed which otherwise would stuck in the
1362 * below while (list_empty(list)) loop.
1364 count
= min(pcp
->count
, count
);
1366 /* Ensure requested pindex is drained first. */
1367 pindex
= pindex
- 1;
1369 spin_lock_irqsave(&zone
->lock
, flags
);
1370 isolated_pageblocks
= has_isolate_pageblock(zone
);
1373 struct list_head
*list
;
1376 /* Remove pages from lists in a round-robin fashion. */
1378 if (++pindex
> max_pindex
)
1379 pindex
= min_pindex
;
1380 list
= &pcp
->lists
[pindex
];
1381 if (!list_empty(list
))
1384 if (pindex
== max_pindex
)
1386 if (pindex
== min_pindex
)
1390 order
= pindex_to_order(pindex
);
1391 nr_pages
= 1 << order
;
1395 page
= list_last_entry(list
, struct page
, pcp_list
);
1396 mt
= get_pcppage_migratetype(page
);
1398 /* must delete to avoid corrupting pcp list */
1399 list_del(&page
->pcp_list
);
1401 pcp
->count
-= nr_pages
;
1403 /* MIGRATE_ISOLATE page should not go to pcplists */
1404 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1405 /* Pageblock could have been isolated meanwhile */
1406 if (unlikely(isolated_pageblocks
))
1407 mt
= get_pageblock_migratetype(page
);
1409 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
, FPI_NONE
);
1410 trace_mm_page_pcpu_drain(page
, order
, mt
);
1411 } while (count
> 0 && !list_empty(list
));
1414 spin_unlock_irqrestore(&zone
->lock
, flags
);
1417 static void free_one_page(struct zone
*zone
,
1418 struct page
*page
, unsigned long pfn
,
1420 int migratetype
, fpi_t fpi_flags
)
1422 unsigned long flags
;
1424 spin_lock_irqsave(&zone
->lock
, flags
);
1425 if (unlikely(has_isolate_pageblock(zone
) ||
1426 is_migrate_isolate(migratetype
))) {
1427 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1429 __free_one_page(page
, pfn
, zone
, order
, migratetype
, fpi_flags
);
1430 spin_unlock_irqrestore(&zone
->lock
, flags
);
1433 static void __free_pages_ok(struct page
*page
, unsigned int order
,
1436 unsigned long flags
;
1438 unsigned long pfn
= page_to_pfn(page
);
1439 struct zone
*zone
= page_zone(page
);
1441 if (!free_pages_prepare(page
, order
, fpi_flags
))
1445 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1446 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1447 * This will reduce the lock holding time.
1449 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1451 spin_lock_irqsave(&zone
->lock
, flags
);
1452 if (unlikely(has_isolate_pageblock(zone
) ||
1453 is_migrate_isolate(migratetype
))) {
1454 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1456 __free_one_page(page
, pfn
, zone
, order
, migratetype
, fpi_flags
);
1457 spin_unlock_irqrestore(&zone
->lock
, flags
);
1459 __count_vm_events(PGFREE
, 1 << order
);
1462 void __free_pages_core(struct page
*page
, unsigned int order
)
1464 unsigned int nr_pages
= 1 << order
;
1465 struct page
*p
= page
;
1469 * When initializing the memmap, __init_single_page() sets the refcount
1470 * of all pages to 1 ("allocated"/"not free"). We have to set the
1471 * refcount of all involved pages to 0.
1474 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1476 __ClearPageReserved(p
);
1477 set_page_count(p
, 0);
1479 __ClearPageReserved(p
);
1480 set_page_count(p
, 0);
1482 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1485 * Bypass PCP and place fresh pages right to the tail, primarily
1486 * relevant for memory onlining.
1488 __free_pages_ok(page
, order
, FPI_TO_TAIL
);
1492 * Check that the whole (or subset of) a pageblock given by the interval of
1493 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1494 * with the migration of free compaction scanner.
1496 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1498 * It's possible on some configurations to have a setup like node0 node1 node0
1499 * i.e. it's possible that all pages within a zones range of pages do not
1500 * belong to a single zone. We assume that a border between node0 and node1
1501 * can occur within a single pageblock, but not a node0 node1 node0
1502 * interleaving within a single pageblock. It is therefore sufficient to check
1503 * the first and last page of a pageblock and avoid checking each individual
1504 * page in a pageblock.
1506 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1507 unsigned long end_pfn
, struct zone
*zone
)
1509 struct page
*start_page
;
1510 struct page
*end_page
;
1512 /* end_pfn is one past the range we are checking */
1515 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1518 start_page
= pfn_to_online_page(start_pfn
);
1522 if (page_zone(start_page
) != zone
)
1525 end_page
= pfn_to_page(end_pfn
);
1527 /* This gives a shorter code than deriving page_zone(end_page) */
1528 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1534 void set_zone_contiguous(struct zone
*zone
)
1536 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1537 unsigned long block_end_pfn
;
1539 block_end_pfn
= pageblock_end_pfn(block_start_pfn
);
1540 for (; block_start_pfn
< zone_end_pfn(zone
);
1541 block_start_pfn
= block_end_pfn
,
1542 block_end_pfn
+= pageblock_nr_pages
) {
1544 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1546 if (!__pageblock_pfn_to_page(block_start_pfn
,
1547 block_end_pfn
, zone
))
1552 /* We confirm that there is no hole */
1553 zone
->contiguous
= true;
1556 void clear_zone_contiguous(struct zone
*zone
)
1558 zone
->contiguous
= false;
1562 * The order of subdivision here is critical for the IO subsystem.
1563 * Please do not alter this order without good reasons and regression
1564 * testing. Specifically, as large blocks of memory are subdivided,
1565 * the order in which smaller blocks are delivered depends on the order
1566 * they're subdivided in this function. This is the primary factor
1567 * influencing the order in which pages are delivered to the IO
1568 * subsystem according to empirical testing, and this is also justified
1569 * by considering the behavior of a buddy system containing a single
1570 * large block of memory acted on by a series of small allocations.
1571 * This behavior is a critical factor in sglist merging's success.
1575 static inline void expand(struct zone
*zone
, struct page
*page
,
1576 int low
, int high
, int migratetype
)
1578 unsigned long size
= 1 << high
;
1580 while (high
> low
) {
1583 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1586 * Mark as guard pages (or page), that will allow to
1587 * merge back to allocator when buddy will be freed.
1588 * Corresponding page table entries will not be touched,
1589 * pages will stay not present in virtual address space
1591 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1594 add_to_free_list(&page
[size
], zone
, high
, migratetype
);
1595 set_buddy_order(&page
[size
], high
);
1599 static void check_new_page_bad(struct page
*page
)
1601 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1602 /* Don't complain about hwpoisoned pages */
1603 page_mapcount_reset(page
); /* remove PageBuddy */
1608 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_PREP
));
1612 * This page is about to be returned from the page allocator
1614 static int check_new_page(struct page
*page
)
1616 if (likely(page_expected_state(page
,
1617 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1620 check_new_page_bad(page
);
1624 static inline bool check_new_pages(struct page
*page
, unsigned int order
)
1626 if (is_check_pages_enabled()) {
1627 for (int i
= 0; i
< (1 << order
); i
++) {
1628 struct page
*p
= page
+ i
;
1630 if (check_new_page(p
))
1638 static inline bool should_skip_kasan_unpoison(gfp_t flags
)
1640 /* Don't skip if a software KASAN mode is enabled. */
1641 if (IS_ENABLED(CONFIG_KASAN_GENERIC
) ||
1642 IS_ENABLED(CONFIG_KASAN_SW_TAGS
))
1645 /* Skip, if hardware tag-based KASAN is not enabled. */
1646 if (!kasan_hw_tags_enabled())
1650 * With hardware tag-based KASAN enabled, skip if this has been
1651 * requested via __GFP_SKIP_KASAN.
1653 return flags
& __GFP_SKIP_KASAN
;
1656 static inline bool should_skip_init(gfp_t flags
)
1658 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1659 if (!kasan_hw_tags_enabled())
1662 /* For hardware tag-based KASAN, skip if requested. */
1663 return (flags
& __GFP_SKIP_ZERO
);
1666 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1669 bool init
= !want_init_on_free() && want_init_on_alloc(gfp_flags
) &&
1670 !should_skip_init(gfp_flags
);
1671 bool zero_tags
= init
&& (gfp_flags
& __GFP_ZEROTAGS
);
1674 set_page_private(page
, 0);
1675 set_page_refcounted(page
);
1677 arch_alloc_page(page
, order
);
1678 debug_pagealloc_map_pages(page
, 1 << order
);
1681 * Page unpoisoning must happen before memory initialization.
1682 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1683 * allocations and the page unpoisoning code will complain.
1685 kernel_unpoison_pages(page
, 1 << order
);
1688 * As memory initialization might be integrated into KASAN,
1689 * KASAN unpoisoning and memory initializion code must be
1690 * kept together to avoid discrepancies in behavior.
1694 * If memory tags should be zeroed
1695 * (which happens only when memory should be initialized as well).
1698 /* Initialize both memory and memory tags. */
1699 for (i
= 0; i
!= 1 << order
; ++i
)
1700 tag_clear_highpage(page
+ i
);
1702 /* Take note that memory was initialized by the loop above. */
1705 if (!should_skip_kasan_unpoison(gfp_flags
) &&
1706 kasan_unpoison_pages(page
, order
, init
)) {
1707 /* Take note that memory was initialized by KASAN. */
1708 if (kasan_has_integrated_init())
1712 * If memory tags have not been set by KASAN, reset the page
1713 * tags to ensure page_address() dereferencing does not fault.
1715 for (i
= 0; i
!= 1 << order
; ++i
)
1716 page_kasan_tag_reset(page
+ i
);
1718 /* If memory is still not initialized, initialize it now. */
1720 kernel_init_pages(page
, 1 << order
);
1722 set_page_owner(page
, order
, gfp_flags
);
1723 page_table_check_alloc(page
, order
);
1726 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1727 unsigned int alloc_flags
)
1729 post_alloc_hook(page
, order
, gfp_flags
);
1731 if (order
&& (gfp_flags
& __GFP_COMP
))
1732 prep_compound_page(page
, order
);
1735 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1736 * allocate the page. The expectation is that the caller is taking
1737 * steps that will free more memory. The caller should avoid the page
1738 * being used for !PFMEMALLOC purposes.
1740 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1741 set_page_pfmemalloc(page
);
1743 clear_page_pfmemalloc(page
);
1747 * Go through the free lists for the given migratetype and remove
1748 * the smallest available page from the freelists
1750 static __always_inline
1751 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1754 unsigned int current_order
;
1755 struct free_area
*area
;
1758 /* Find a page of the appropriate size in the preferred list */
1759 for (current_order
= order
; current_order
<= MAX_ORDER
; ++current_order
) {
1760 area
= &(zone
->free_area
[current_order
]);
1761 page
= get_page_from_free_area(area
, migratetype
);
1764 del_page_from_free_list(page
, zone
, current_order
);
1765 expand(zone
, page
, order
, current_order
, migratetype
);
1766 set_pcppage_migratetype(page
, migratetype
);
1767 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
,
1768 pcp_allowed_order(order
) &&
1769 migratetype
< MIGRATE_PCPTYPES
);
1778 * This array describes the order lists are fallen back to when
1779 * the free lists for the desirable migrate type are depleted
1781 * The other migratetypes do not have fallbacks.
1783 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_PCPTYPES
- 1] = {
1784 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
},
1785 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
},
1786 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
},
1790 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1793 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1796 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1797 unsigned int order
) { return NULL
; }
1801 * Move the free pages in a range to the freelist tail of the requested type.
1802 * Note that start_page and end_pages are not aligned on a pageblock
1803 * boundary. If alignment is required, use move_freepages_block()
1805 static int move_freepages(struct zone
*zone
,
1806 unsigned long start_pfn
, unsigned long end_pfn
,
1807 int migratetype
, int *num_movable
)
1812 int pages_moved
= 0;
1814 for (pfn
= start_pfn
; pfn
<= end_pfn
;) {
1815 page
= pfn_to_page(pfn
);
1816 if (!PageBuddy(page
)) {
1818 * We assume that pages that could be isolated for
1819 * migration are movable. But we don't actually try
1820 * isolating, as that would be expensive.
1823 (PageLRU(page
) || __PageMovable(page
)))
1829 /* Make sure we are not inadvertently changing nodes */
1830 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1831 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
1833 order
= buddy_order(page
);
1834 move_to_free_list(page
, zone
, order
, migratetype
);
1836 pages_moved
+= 1 << order
;
1842 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1843 int migratetype
, int *num_movable
)
1845 unsigned long start_pfn
, end_pfn
, pfn
;
1850 pfn
= page_to_pfn(page
);
1851 start_pfn
= pageblock_start_pfn(pfn
);
1852 end_pfn
= pageblock_end_pfn(pfn
) - 1;
1854 /* Do not cross zone boundaries */
1855 if (!zone_spans_pfn(zone
, start_pfn
))
1857 if (!zone_spans_pfn(zone
, end_pfn
))
1860 return move_freepages(zone
, start_pfn
, end_pfn
, migratetype
,
1864 static void change_pageblock_range(struct page
*pageblock_page
,
1865 int start_order
, int migratetype
)
1867 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1869 while (nr_pageblocks
--) {
1870 set_pageblock_migratetype(pageblock_page
, migratetype
);
1871 pageblock_page
+= pageblock_nr_pages
;
1876 * When we are falling back to another migratetype during allocation, try to
1877 * steal extra free pages from the same pageblocks to satisfy further
1878 * allocations, instead of polluting multiple pageblocks.
1880 * If we are stealing a relatively large buddy page, it is likely there will
1881 * be more free pages in the pageblock, so try to steal them all. For
1882 * reclaimable and unmovable allocations, we steal regardless of page size,
1883 * as fragmentation caused by those allocations polluting movable pageblocks
1884 * is worse than movable allocations stealing from unmovable and reclaimable
1887 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1890 * Leaving this order check is intended, although there is
1891 * relaxed order check in next check. The reason is that
1892 * we can actually steal whole pageblock if this condition met,
1893 * but, below check doesn't guarantee it and that is just heuristic
1894 * so could be changed anytime.
1896 if (order
>= pageblock_order
)
1899 if (order
>= pageblock_order
/ 2 ||
1900 start_mt
== MIGRATE_RECLAIMABLE
||
1901 start_mt
== MIGRATE_UNMOVABLE
||
1902 page_group_by_mobility_disabled
)
1908 static inline bool boost_watermark(struct zone
*zone
)
1910 unsigned long max_boost
;
1912 if (!watermark_boost_factor
)
1915 * Don't bother in zones that are unlikely to produce results.
1916 * On small machines, including kdump capture kernels running
1917 * in a small area, boosting the watermark can cause an out of
1918 * memory situation immediately.
1920 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
1923 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
1924 watermark_boost_factor
, 10000);
1927 * high watermark may be uninitialised if fragmentation occurs
1928 * very early in boot so do not boost. We do not fall
1929 * through and boost by pageblock_nr_pages as failing
1930 * allocations that early means that reclaim is not going
1931 * to help and it may even be impossible to reclaim the
1932 * boosted watermark resulting in a hang.
1937 max_boost
= max(pageblock_nr_pages
, max_boost
);
1939 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
1946 * This function implements actual steal behaviour. If order is large enough,
1947 * we can steal whole pageblock. If not, we first move freepages in this
1948 * pageblock to our migratetype and determine how many already-allocated pages
1949 * are there in the pageblock with a compatible migratetype. If at least half
1950 * of pages are free or compatible, we can change migratetype of the pageblock
1951 * itself, so pages freed in the future will be put on the correct free list.
1953 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1954 unsigned int alloc_flags
, int start_type
, bool whole_block
)
1956 unsigned int current_order
= buddy_order(page
);
1957 int free_pages
, movable_pages
, alike_pages
;
1960 old_block_type
= get_pageblock_migratetype(page
);
1963 * This can happen due to races and we want to prevent broken
1964 * highatomic accounting.
1966 if (is_migrate_highatomic(old_block_type
))
1969 /* Take ownership for orders >= pageblock_order */
1970 if (current_order
>= pageblock_order
) {
1971 change_pageblock_range(page
, current_order
, start_type
);
1976 * Boost watermarks to increase reclaim pressure to reduce the
1977 * likelihood of future fallbacks. Wake kswapd now as the node
1978 * may be balanced overall and kswapd will not wake naturally.
1980 if (boost_watermark(zone
) && (alloc_flags
& ALLOC_KSWAPD
))
1981 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
1983 /* We are not allowed to try stealing from the whole block */
1987 free_pages
= move_freepages_block(zone
, page
, start_type
,
1990 * Determine how many pages are compatible with our allocation.
1991 * For movable allocation, it's the number of movable pages which
1992 * we just obtained. For other types it's a bit more tricky.
1994 if (start_type
== MIGRATE_MOVABLE
) {
1995 alike_pages
= movable_pages
;
1998 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1999 * to MOVABLE pageblock, consider all non-movable pages as
2000 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2001 * vice versa, be conservative since we can't distinguish the
2002 * exact migratetype of non-movable pages.
2004 if (old_block_type
== MIGRATE_MOVABLE
)
2005 alike_pages
= pageblock_nr_pages
2006 - (free_pages
+ movable_pages
);
2011 /* moving whole block can fail due to zone boundary conditions */
2016 * If a sufficient number of pages in the block are either free or of
2017 * comparable migratability as our allocation, claim the whole block.
2019 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2020 page_group_by_mobility_disabled
)
2021 set_pageblock_migratetype(page
, start_type
);
2026 move_to_free_list(page
, zone
, current_order
, start_type
);
2030 * Check whether there is a suitable fallback freepage with requested order.
2031 * If only_stealable is true, this function returns fallback_mt only if
2032 * we can steal other freepages all together. This would help to reduce
2033 * fragmentation due to mixed migratetype pages in one pageblock.
2035 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2036 int migratetype
, bool only_stealable
, bool *can_steal
)
2041 if (area
->nr_free
== 0)
2045 for (i
= 0; i
< MIGRATE_PCPTYPES
- 1 ; i
++) {
2046 fallback_mt
= fallbacks
[migratetype
][i
];
2047 if (free_area_empty(area
, fallback_mt
))
2050 if (can_steal_fallback(order
, migratetype
))
2053 if (!only_stealable
)
2064 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2065 * there are no empty page blocks that contain a page with a suitable order
2067 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2068 unsigned int alloc_order
)
2071 unsigned long max_managed
, flags
;
2074 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2075 * Check is race-prone but harmless.
2077 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2078 if (zone
->nr_reserved_highatomic
>= max_managed
)
2081 spin_lock_irqsave(&zone
->lock
, flags
);
2083 /* Recheck the nr_reserved_highatomic limit under the lock */
2084 if (zone
->nr_reserved_highatomic
>= max_managed
)
2088 mt
= get_pageblock_migratetype(page
);
2089 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2090 if (migratetype_is_mergeable(mt
)) {
2091 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2092 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2093 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2097 spin_unlock_irqrestore(&zone
->lock
, flags
);
2101 * Used when an allocation is about to fail under memory pressure. This
2102 * potentially hurts the reliability of high-order allocations when under
2103 * intense memory pressure but failed atomic allocations should be easier
2104 * to recover from than an OOM.
2106 * If @force is true, try to unreserve a pageblock even though highatomic
2107 * pageblock is exhausted.
2109 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2112 struct zonelist
*zonelist
= ac
->zonelist
;
2113 unsigned long flags
;
2120 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->highest_zoneidx
,
2123 * Preserve at least one pageblock unless memory pressure
2126 if (!force
&& zone
->nr_reserved_highatomic
<=
2130 spin_lock_irqsave(&zone
->lock
, flags
);
2131 for (order
= 0; order
<= MAX_ORDER
; order
++) {
2132 struct free_area
*area
= &(zone
->free_area
[order
]);
2134 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2139 * In page freeing path, migratetype change is racy so
2140 * we can counter several free pages in a pageblock
2141 * in this loop although we changed the pageblock type
2142 * from highatomic to ac->migratetype. So we should
2143 * adjust the count once.
2145 if (is_migrate_highatomic_page(page
)) {
2147 * It should never happen but changes to
2148 * locking could inadvertently allow a per-cpu
2149 * drain to add pages to MIGRATE_HIGHATOMIC
2150 * while unreserving so be safe and watch for
2153 zone
->nr_reserved_highatomic
-= min(
2155 zone
->nr_reserved_highatomic
);
2159 * Convert to ac->migratetype and avoid the normal
2160 * pageblock stealing heuristics. Minimally, the caller
2161 * is doing the work and needs the pages. More
2162 * importantly, if the block was always converted to
2163 * MIGRATE_UNMOVABLE or another type then the number
2164 * of pageblocks that cannot be completely freed
2167 set_pageblock_migratetype(page
, ac
->migratetype
);
2168 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2171 spin_unlock_irqrestore(&zone
->lock
, flags
);
2175 spin_unlock_irqrestore(&zone
->lock
, flags
);
2182 * Try finding a free buddy page on the fallback list and put it on the free
2183 * list of requested migratetype, possibly along with other pages from the same
2184 * block, depending on fragmentation avoidance heuristics. Returns true if
2185 * fallback was found so that __rmqueue_smallest() can grab it.
2187 * The use of signed ints for order and current_order is a deliberate
2188 * deviation from the rest of this file, to make the for loop
2189 * condition simpler.
2191 static __always_inline
bool
2192 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2193 unsigned int alloc_flags
)
2195 struct free_area
*area
;
2197 int min_order
= order
;
2203 * Do not steal pages from freelists belonging to other pageblocks
2204 * i.e. orders < pageblock_order. If there are no local zones free,
2205 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2207 if (order
< pageblock_order
&& alloc_flags
& ALLOC_NOFRAGMENT
)
2208 min_order
= pageblock_order
;
2211 * Find the largest available free page in the other list. This roughly
2212 * approximates finding the pageblock with the most free pages, which
2213 * would be too costly to do exactly.
2215 for (current_order
= MAX_ORDER
; current_order
>= min_order
;
2217 area
= &(zone
->free_area
[current_order
]);
2218 fallback_mt
= find_suitable_fallback(area
, current_order
,
2219 start_migratetype
, false, &can_steal
);
2220 if (fallback_mt
== -1)
2224 * We cannot steal all free pages from the pageblock and the
2225 * requested migratetype is movable. In that case it's better to
2226 * steal and split the smallest available page instead of the
2227 * largest available page, because even if the next movable
2228 * allocation falls back into a different pageblock than this
2229 * one, it won't cause permanent fragmentation.
2231 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2232 && current_order
> order
)
2241 for (current_order
= order
; current_order
<= MAX_ORDER
;
2243 area
= &(zone
->free_area
[current_order
]);
2244 fallback_mt
= find_suitable_fallback(area
, current_order
,
2245 start_migratetype
, false, &can_steal
);
2246 if (fallback_mt
!= -1)
2251 * This should not happen - we already found a suitable fallback
2252 * when looking for the largest page.
2254 VM_BUG_ON(current_order
> MAX_ORDER
);
2257 page
= get_page_from_free_area(area
, fallback_mt
);
2259 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2262 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2263 start_migratetype
, fallback_mt
);
2270 * Do the hard work of removing an element from the buddy allocator.
2271 * Call me with the zone->lock already held.
2273 static __always_inline
struct page
*
2274 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2275 unsigned int alloc_flags
)
2279 if (IS_ENABLED(CONFIG_CMA
)) {
2281 * Balance movable allocations between regular and CMA areas by
2282 * allocating from CMA when over half of the zone's free memory
2283 * is in the CMA area.
2285 if (alloc_flags
& ALLOC_CMA
&&
2286 zone_page_state(zone
, NR_FREE_CMA_PAGES
) >
2287 zone_page_state(zone
, NR_FREE_PAGES
) / 2) {
2288 page
= __rmqueue_cma_fallback(zone
, order
);
2294 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2295 if (unlikely(!page
)) {
2296 if (alloc_flags
& ALLOC_CMA
)
2297 page
= __rmqueue_cma_fallback(zone
, order
);
2299 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2307 * Obtain a specified number of elements from the buddy allocator, all under
2308 * a single hold of the lock, for efficiency. Add them to the supplied list.
2309 * Returns the number of new pages which were placed at *list.
2311 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2312 unsigned long count
, struct list_head
*list
,
2313 int migratetype
, unsigned int alloc_flags
)
2315 unsigned long flags
;
2318 spin_lock_irqsave(&zone
->lock
, flags
);
2319 for (i
= 0; i
< count
; ++i
) {
2320 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2322 if (unlikely(page
== NULL
))
2326 * Split buddy pages returned by expand() are received here in
2327 * physical page order. The page is added to the tail of
2328 * caller's list. From the callers perspective, the linked list
2329 * is ordered by page number under some conditions. This is
2330 * useful for IO devices that can forward direction from the
2331 * head, thus also in the physical page order. This is useful
2332 * for IO devices that can merge IO requests if the physical
2333 * pages are ordered properly.
2335 list_add_tail(&page
->pcp_list
, list
);
2336 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2337 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2341 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2342 spin_unlock_irqrestore(&zone
->lock
, flags
);
2349 * Called from the vmstat counter updater to drain pagesets of this
2350 * currently executing processor on remote nodes after they have
2353 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2355 int to_drain
, batch
;
2357 batch
= READ_ONCE(pcp
->batch
);
2358 to_drain
= min(pcp
->count
, batch
);
2360 spin_lock(&pcp
->lock
);
2361 free_pcppages_bulk(zone
, to_drain
, pcp
, 0);
2362 spin_unlock(&pcp
->lock
);
2368 * Drain pcplists of the indicated processor and zone.
2370 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2372 struct per_cpu_pages
*pcp
;
2374 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
2376 spin_lock(&pcp
->lock
);
2377 free_pcppages_bulk(zone
, pcp
->count
, pcp
, 0);
2378 spin_unlock(&pcp
->lock
);
2383 * Drain pcplists of all zones on the indicated processor.
2385 static void drain_pages(unsigned int cpu
)
2389 for_each_populated_zone(zone
) {
2390 drain_pages_zone(cpu
, zone
);
2395 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2397 void drain_local_pages(struct zone
*zone
)
2399 int cpu
= smp_processor_id();
2402 drain_pages_zone(cpu
, zone
);
2408 * The implementation of drain_all_pages(), exposing an extra parameter to
2409 * drain on all cpus.
2411 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2412 * not empty. The check for non-emptiness can however race with a free to
2413 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2414 * that need the guarantee that every CPU has drained can disable the
2415 * optimizing racy check.
2417 static void __drain_all_pages(struct zone
*zone
, bool force_all_cpus
)
2422 * Allocate in the BSS so we won't require allocation in
2423 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2425 static cpumask_t cpus_with_pcps
;
2428 * Do not drain if one is already in progress unless it's specific to
2429 * a zone. Such callers are primarily CMA and memory hotplug and need
2430 * the drain to be complete when the call returns.
2432 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2435 mutex_lock(&pcpu_drain_mutex
);
2439 * We don't care about racing with CPU hotplug event
2440 * as offline notification will cause the notified
2441 * cpu to drain that CPU pcps and on_each_cpu_mask
2442 * disables preemption as part of its processing
2444 for_each_online_cpu(cpu
) {
2445 struct per_cpu_pages
*pcp
;
2447 bool has_pcps
= false;
2449 if (force_all_cpus
) {
2451 * The pcp.count check is racy, some callers need a
2452 * guarantee that no cpu is missed.
2456 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
2460 for_each_populated_zone(z
) {
2461 pcp
= per_cpu_ptr(z
->per_cpu_pageset
, cpu
);
2470 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2472 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2475 for_each_cpu(cpu
, &cpus_with_pcps
) {
2477 drain_pages_zone(cpu
, zone
);
2482 mutex_unlock(&pcpu_drain_mutex
);
2486 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2488 * When zone parameter is non-NULL, spill just the single zone's pages.
2490 void drain_all_pages(struct zone
*zone
)
2492 __drain_all_pages(zone
, false);
2495 #ifdef CONFIG_HIBERNATION
2498 * Touch the watchdog for every WD_PAGE_COUNT pages.
2500 #define WD_PAGE_COUNT (128*1024)
2502 void mark_free_pages(struct zone
*zone
)
2504 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2505 unsigned long flags
;
2506 unsigned int order
, t
;
2509 if (zone_is_empty(zone
))
2512 spin_lock_irqsave(&zone
->lock
, flags
);
2514 max_zone_pfn
= zone_end_pfn(zone
);
2515 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2516 if (pfn_valid(pfn
)) {
2517 page
= pfn_to_page(pfn
);
2519 if (!--page_count
) {
2520 touch_nmi_watchdog();
2521 page_count
= WD_PAGE_COUNT
;
2524 if (page_zone(page
) != zone
)
2527 if (!swsusp_page_is_forbidden(page
))
2528 swsusp_unset_page_free(page
);
2531 for_each_migratetype_order(order
, t
) {
2532 list_for_each_entry(page
,
2533 &zone
->free_area
[order
].free_list
[t
], buddy_list
) {
2536 pfn
= page_to_pfn(page
);
2537 for (i
= 0; i
< (1UL << order
); i
++) {
2538 if (!--page_count
) {
2539 touch_nmi_watchdog();
2540 page_count
= WD_PAGE_COUNT
;
2542 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2546 spin_unlock_irqrestore(&zone
->lock
, flags
);
2548 #endif /* CONFIG_PM */
2550 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
,
2555 if (!free_pages_prepare(page
, order
, FPI_NONE
))
2558 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2559 set_pcppage_migratetype(page
, migratetype
);
2563 static int nr_pcp_free(struct per_cpu_pages
*pcp
, int high
, int batch
,
2566 int min_nr_free
, max_nr_free
;
2568 /* Free everything if batch freeing high-order pages. */
2569 if (unlikely(free_high
))
2572 /* Check for PCP disabled or boot pageset */
2573 if (unlikely(high
< batch
))
2576 /* Leave at least pcp->batch pages on the list */
2577 min_nr_free
= batch
;
2578 max_nr_free
= high
- batch
;
2581 * Double the number of pages freed each time there is subsequent
2582 * freeing of pages without any allocation.
2584 batch
<<= pcp
->free_factor
;
2585 if (batch
< max_nr_free
)
2587 batch
= clamp(batch
, min_nr_free
, max_nr_free
);
2592 static int nr_pcp_high(struct per_cpu_pages
*pcp
, struct zone
*zone
,
2595 int high
= READ_ONCE(pcp
->high
);
2597 if (unlikely(!high
|| free_high
))
2600 if (!test_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
))
2604 * If reclaim is active, limit the number of pages that can be
2605 * stored on pcp lists
2607 return min(READ_ONCE(pcp
->batch
) << 2, high
);
2610 static void free_unref_page_commit(struct zone
*zone
, struct per_cpu_pages
*pcp
,
2611 struct page
*page
, int migratetype
,
2618 __count_vm_events(PGFREE
, 1 << order
);
2619 pindex
= order_to_pindex(migratetype
, order
);
2620 list_add(&page
->pcp_list
, &pcp
->lists
[pindex
]);
2621 pcp
->count
+= 1 << order
;
2624 * As high-order pages other than THP's stored on PCP can contribute
2625 * to fragmentation, limit the number stored when PCP is heavily
2626 * freeing without allocation. The remainder after bulk freeing
2627 * stops will be drained from vmstat refresh context.
2629 free_high
= (pcp
->free_factor
&& order
&& order
<= PAGE_ALLOC_COSTLY_ORDER
);
2631 high
= nr_pcp_high(pcp
, zone
, free_high
);
2632 if (pcp
->count
>= high
) {
2633 int batch
= READ_ONCE(pcp
->batch
);
2635 free_pcppages_bulk(zone
, nr_pcp_free(pcp
, high
, batch
, free_high
), pcp
, pindex
);
2642 void free_unref_page(struct page
*page
, unsigned int order
)
2644 unsigned long __maybe_unused UP_flags
;
2645 struct per_cpu_pages
*pcp
;
2647 unsigned long pfn
= page_to_pfn(page
);
2650 if (!free_unref_page_prepare(page
, pfn
, order
))
2654 * We only track unmovable, reclaimable and movable on pcp lists.
2655 * Place ISOLATE pages on the isolated list because they are being
2656 * offlined but treat HIGHATOMIC as movable pages so we can get those
2657 * areas back if necessary. Otherwise, we may have to free
2658 * excessively into the page allocator
2660 migratetype
= get_pcppage_migratetype(page
);
2661 if (unlikely(migratetype
>= MIGRATE_PCPTYPES
)) {
2662 if (unlikely(is_migrate_isolate(migratetype
))) {
2663 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
, FPI_NONE
);
2666 migratetype
= MIGRATE_MOVABLE
;
2669 zone
= page_zone(page
);
2670 pcp_trylock_prepare(UP_flags
);
2671 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
2673 free_unref_page_commit(zone
, pcp
, page
, migratetype
, order
);
2674 pcp_spin_unlock(pcp
);
2676 free_one_page(zone
, page
, pfn
, order
, migratetype
, FPI_NONE
);
2678 pcp_trylock_finish(UP_flags
);
2682 * Free a list of 0-order pages
2684 void free_unref_page_list(struct list_head
*list
)
2686 unsigned long __maybe_unused UP_flags
;
2687 struct page
*page
, *next
;
2688 struct per_cpu_pages
*pcp
= NULL
;
2689 struct zone
*locked_zone
= NULL
;
2690 int batch_count
= 0;
2693 /* Prepare pages for freeing */
2694 list_for_each_entry_safe(page
, next
, list
, lru
) {
2695 unsigned long pfn
= page_to_pfn(page
);
2696 if (!free_unref_page_prepare(page
, pfn
, 0)) {
2697 list_del(&page
->lru
);
2702 * Free isolated pages directly to the allocator, see
2703 * comment in free_unref_page.
2705 migratetype
= get_pcppage_migratetype(page
);
2706 if (unlikely(is_migrate_isolate(migratetype
))) {
2707 list_del(&page
->lru
);
2708 free_one_page(page_zone(page
), page
, pfn
, 0, migratetype
, FPI_NONE
);
2713 list_for_each_entry_safe(page
, next
, list
, lru
) {
2714 struct zone
*zone
= page_zone(page
);
2716 list_del(&page
->lru
);
2717 migratetype
= get_pcppage_migratetype(page
);
2720 * Either different zone requiring a different pcp lock or
2721 * excessive lock hold times when freeing a large list of
2724 if (zone
!= locked_zone
|| batch_count
== SWAP_CLUSTER_MAX
) {
2726 pcp_spin_unlock(pcp
);
2727 pcp_trylock_finish(UP_flags
);
2733 * trylock is necessary as pages may be getting freed
2734 * from IRQ or SoftIRQ context after an IO completion.
2736 pcp_trylock_prepare(UP_flags
);
2737 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
2738 if (unlikely(!pcp
)) {
2739 pcp_trylock_finish(UP_flags
);
2740 free_one_page(zone
, page
, page_to_pfn(page
),
2741 0, migratetype
, FPI_NONE
);
2749 * Non-isolated types over MIGRATE_PCPTYPES get added
2750 * to the MIGRATE_MOVABLE pcp list.
2752 if (unlikely(migratetype
>= MIGRATE_PCPTYPES
))
2753 migratetype
= MIGRATE_MOVABLE
;
2755 trace_mm_page_free_batched(page
);
2756 free_unref_page_commit(zone
, pcp
, page
, migratetype
, 0);
2761 pcp_spin_unlock(pcp
);
2762 pcp_trylock_finish(UP_flags
);
2767 * split_page takes a non-compound higher-order page, and splits it into
2768 * n (1<<order) sub-pages: page[0..n]
2769 * Each sub-page must be freed individually.
2771 * Note: this is probably too low level an operation for use in drivers.
2772 * Please consult with lkml before using this in your driver.
2774 void split_page(struct page
*page
, unsigned int order
)
2778 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2779 VM_BUG_ON_PAGE(!page_count(page
), page
);
2781 for (i
= 1; i
< (1 << order
); i
++)
2782 set_page_refcounted(page
+ i
);
2783 split_page_owner(page
, 1 << order
);
2784 split_page_memcg(page
, 1 << order
);
2786 EXPORT_SYMBOL_GPL(split_page
);
2788 int __isolate_free_page(struct page
*page
, unsigned int order
)
2790 struct zone
*zone
= page_zone(page
);
2791 int mt
= get_pageblock_migratetype(page
);
2793 if (!is_migrate_isolate(mt
)) {
2794 unsigned long watermark
;
2796 * Obey watermarks as if the page was being allocated. We can
2797 * emulate a high-order watermark check with a raised order-0
2798 * watermark, because we already know our high-order page
2801 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
2802 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2805 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2808 del_page_from_free_list(page
, zone
, order
);
2811 * Set the pageblock if the isolated page is at least half of a
2814 if (order
>= pageblock_order
- 1) {
2815 struct page
*endpage
= page
+ (1 << order
) - 1;
2816 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2817 int mt
= get_pageblock_migratetype(page
);
2819 * Only change normal pageblocks (i.e., they can merge
2822 if (migratetype_is_mergeable(mt
))
2823 set_pageblock_migratetype(page
,
2828 return 1UL << order
;
2832 * __putback_isolated_page - Return a now-isolated page back where we got it
2833 * @page: Page that was isolated
2834 * @order: Order of the isolated page
2835 * @mt: The page's pageblock's migratetype
2837 * This function is meant to return a page pulled from the free lists via
2838 * __isolate_free_page back to the free lists they were pulled from.
2840 void __putback_isolated_page(struct page
*page
, unsigned int order
, int mt
)
2842 struct zone
*zone
= page_zone(page
);
2844 /* zone lock should be held when this function is called */
2845 lockdep_assert_held(&zone
->lock
);
2847 /* Return isolated page to tail of freelist. */
2848 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
,
2849 FPI_SKIP_REPORT_NOTIFY
| FPI_TO_TAIL
);
2853 * Update NUMA hit/miss statistics
2855 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2859 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2861 /* skip numa counters update if numa stats is disabled */
2862 if (!static_branch_likely(&vm_numa_stat_key
))
2865 if (zone_to_nid(z
) != numa_node_id())
2866 local_stat
= NUMA_OTHER
;
2868 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
2869 __count_numa_events(z
, NUMA_HIT
, nr_account
);
2871 __count_numa_events(z
, NUMA_MISS
, nr_account
);
2872 __count_numa_events(preferred_zone
, NUMA_FOREIGN
, nr_account
);
2874 __count_numa_events(z
, local_stat
, nr_account
);
2878 static __always_inline
2879 struct page
*rmqueue_buddy(struct zone
*preferred_zone
, struct zone
*zone
,
2880 unsigned int order
, unsigned int alloc_flags
,
2884 unsigned long flags
;
2888 spin_lock_irqsave(&zone
->lock
, flags
);
2890 * order-0 request can reach here when the pcplist is skipped
2891 * due to non-CMA allocation context. HIGHATOMIC area is
2892 * reserved for high-order atomic allocation, so order-0
2893 * request should skip it.
2895 if (alloc_flags
& ALLOC_HIGHATOMIC
)
2896 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2898 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
2901 * If the allocation fails, allow OOM handling access
2902 * to HIGHATOMIC reserves as failing now is worse than
2903 * failing a high-order atomic allocation in the
2906 if (!page
&& (alloc_flags
& ALLOC_OOM
))
2907 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2910 spin_unlock_irqrestore(&zone
->lock
, flags
);
2914 __mod_zone_freepage_state(zone
, -(1 << order
),
2915 get_pcppage_migratetype(page
));
2916 spin_unlock_irqrestore(&zone
->lock
, flags
);
2917 } while (check_new_pages(page
, order
));
2919 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2920 zone_statistics(preferred_zone
, zone
, 1);
2925 /* Remove page from the per-cpu list, caller must protect the list */
2927 struct page
*__rmqueue_pcplist(struct zone
*zone
, unsigned int order
,
2929 unsigned int alloc_flags
,
2930 struct per_cpu_pages
*pcp
,
2931 struct list_head
*list
)
2936 if (list_empty(list
)) {
2937 int batch
= READ_ONCE(pcp
->batch
);
2941 * Scale batch relative to order if batch implies
2942 * free pages can be stored on the PCP. Batch can
2943 * be 1 for small zones or for boot pagesets which
2944 * should never store free pages as the pages may
2945 * belong to arbitrary zones.
2948 batch
= max(batch
>> order
, 2);
2949 alloced
= rmqueue_bulk(zone
, order
,
2951 migratetype
, alloc_flags
);
2953 pcp
->count
+= alloced
<< order
;
2954 if (unlikely(list_empty(list
)))
2958 page
= list_first_entry(list
, struct page
, pcp_list
);
2959 list_del(&page
->pcp_list
);
2960 pcp
->count
-= 1 << order
;
2961 } while (check_new_pages(page
, order
));
2966 /* Lock and remove page from the per-cpu list */
2967 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2968 struct zone
*zone
, unsigned int order
,
2969 int migratetype
, unsigned int alloc_flags
)
2971 struct per_cpu_pages
*pcp
;
2972 struct list_head
*list
;
2974 unsigned long __maybe_unused UP_flags
;
2976 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2977 pcp_trylock_prepare(UP_flags
);
2978 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
2980 pcp_trylock_finish(UP_flags
);
2985 * On allocation, reduce the number of pages that are batch freed.
2986 * See nr_pcp_free() where free_factor is increased for subsequent
2989 pcp
->free_factor
>>= 1;
2990 list
= &pcp
->lists
[order_to_pindex(migratetype
, order
)];
2991 page
= __rmqueue_pcplist(zone
, order
, migratetype
, alloc_flags
, pcp
, list
);
2992 pcp_spin_unlock(pcp
);
2993 pcp_trylock_finish(UP_flags
);
2995 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2996 zone_statistics(preferred_zone
, zone
, 1);
3002 * Allocate a page from the given zone.
3003 * Use pcplists for THP or "cheap" high-order allocations.
3007 * Do not instrument rmqueue() with KMSAN. This function may call
3008 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3009 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3010 * may call rmqueue() again, which will result in a deadlock.
3012 __no_sanitize_memory
3014 struct page
*rmqueue(struct zone
*preferred_zone
,
3015 struct zone
*zone
, unsigned int order
,
3016 gfp_t gfp_flags
, unsigned int alloc_flags
,
3022 * We most definitely don't want callers attempting to
3023 * allocate greater than order-1 page units with __GFP_NOFAIL.
3025 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3027 if (likely(pcp_allowed_order(order
))) {
3029 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3030 * we need to skip it when CMA area isn't allowed.
3032 if (!IS_ENABLED(CONFIG_CMA
) || alloc_flags
& ALLOC_CMA
||
3033 migratetype
!= MIGRATE_MOVABLE
) {
3034 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
3035 migratetype
, alloc_flags
);
3041 page
= rmqueue_buddy(preferred_zone
, zone
, order
, alloc_flags
,
3045 /* Separate test+clear to avoid unnecessary atomics */
3046 if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
))) {
3047 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3048 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3051 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3055 #ifdef CONFIG_FAIL_PAGE_ALLOC
3058 struct fault_attr attr
;
3060 bool ignore_gfp_highmem
;
3061 bool ignore_gfp_reclaim
;
3063 } fail_page_alloc
= {
3064 .attr
= FAULT_ATTR_INITIALIZER
,
3065 .ignore_gfp_reclaim
= true,
3066 .ignore_gfp_highmem
= true,
3070 static int __init
setup_fail_page_alloc(char *str
)
3072 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3074 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3076 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3080 if (order
< fail_page_alloc
.min_order
)
3082 if (gfp_mask
& __GFP_NOFAIL
)
3084 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3086 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3087 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3090 /* See comment in __should_failslab() */
3091 if (gfp_mask
& __GFP_NOWARN
)
3092 flags
|= FAULT_NOWARN
;
3094 return should_fail_ex(&fail_page_alloc
.attr
, 1 << order
, flags
);
3097 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3099 static int __init
fail_page_alloc_debugfs(void)
3101 umode_t mode
= S_IFREG
| 0600;
3104 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3105 &fail_page_alloc
.attr
);
3107 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3108 &fail_page_alloc
.ignore_gfp_reclaim
);
3109 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3110 &fail_page_alloc
.ignore_gfp_highmem
);
3111 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3116 late_initcall(fail_page_alloc_debugfs
);
3118 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3120 #else /* CONFIG_FAIL_PAGE_ALLOC */
3122 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3127 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3129 noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3131 return __should_fail_alloc_page(gfp_mask
, order
);
3133 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3135 static inline long __zone_watermark_unusable_free(struct zone
*z
,
3136 unsigned int order
, unsigned int alloc_flags
)
3138 long unusable_free
= (1 << order
) - 1;
3141 * If the caller does not have rights to reserves below the min
3142 * watermark then subtract the high-atomic reserves. This will
3143 * over-estimate the size of the atomic reserve but it avoids a search.
3145 if (likely(!(alloc_flags
& ALLOC_RESERVES
)))
3146 unusable_free
+= z
->nr_reserved_highatomic
;
3149 /* If allocation can't use CMA areas don't use free CMA pages */
3150 if (!(alloc_flags
& ALLOC_CMA
))
3151 unusable_free
+= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3154 return unusable_free
;
3158 * Return true if free base pages are above 'mark'. For high-order checks it
3159 * will return true of the order-0 watermark is reached and there is at least
3160 * one free page of a suitable size. Checking now avoids taking the zone lock
3161 * to check in the allocation paths if no pages are free.
3163 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3164 int highest_zoneidx
, unsigned int alloc_flags
,
3170 /* free_pages may go negative - that's OK */
3171 free_pages
-= __zone_watermark_unusable_free(z
, order
, alloc_flags
);
3173 if (unlikely(alloc_flags
& ALLOC_RESERVES
)) {
3175 * __GFP_HIGH allows access to 50% of the min reserve as well
3178 if (alloc_flags
& ALLOC_MIN_RESERVE
) {
3182 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3183 * access more reserves than just __GFP_HIGH. Other
3184 * non-blocking allocations requests such as GFP_NOWAIT
3185 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3186 * access to the min reserve.
3188 if (alloc_flags
& ALLOC_NON_BLOCK
)
3193 * OOM victims can try even harder than the normal reserve
3194 * users on the grounds that it's definitely going to be in
3195 * the exit path shortly and free memory. Any allocation it
3196 * makes during the free path will be small and short-lived.
3198 if (alloc_flags
& ALLOC_OOM
)
3203 * Check watermarks for an order-0 allocation request. If these
3204 * are not met, then a high-order request also cannot go ahead
3205 * even if a suitable page happened to be free.
3207 if (free_pages
<= min
+ z
->lowmem_reserve
[highest_zoneidx
])
3210 /* If this is an order-0 request then the watermark is fine */
3214 /* For a high-order request, check at least one suitable page is free */
3215 for (o
= order
; o
<= MAX_ORDER
; o
++) {
3216 struct free_area
*area
= &z
->free_area
[o
];
3222 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3223 if (!free_area_empty(area
, mt
))
3228 if ((alloc_flags
& ALLOC_CMA
) &&
3229 !free_area_empty(area
, MIGRATE_CMA
)) {
3233 if ((alloc_flags
& (ALLOC_HIGHATOMIC
|ALLOC_OOM
)) &&
3234 !free_area_empty(area
, MIGRATE_HIGHATOMIC
)) {
3241 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3242 int highest_zoneidx
, unsigned int alloc_flags
)
3244 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3245 zone_page_state(z
, NR_FREE_PAGES
));
3248 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3249 unsigned long mark
, int highest_zoneidx
,
3250 unsigned int alloc_flags
, gfp_t gfp_mask
)
3254 free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3257 * Fast check for order-0 only. If this fails then the reserves
3258 * need to be calculated.
3264 usable_free
= free_pages
;
3265 reserved
= __zone_watermark_unusable_free(z
, 0, alloc_flags
);
3267 /* reserved may over estimate high-atomic reserves. */
3268 usable_free
-= min(usable_free
, reserved
);
3269 if (usable_free
> mark
+ z
->lowmem_reserve
[highest_zoneidx
])
3273 if (__zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3278 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3279 * when checking the min watermark. The min watermark is the
3280 * point where boosting is ignored so that kswapd is woken up
3281 * when below the low watermark.
3283 if (unlikely(!order
&& (alloc_flags
& ALLOC_MIN_RESERVE
) && z
->watermark_boost
3284 && ((alloc_flags
& ALLOC_WMARK_MASK
) == WMARK_MIN
))) {
3285 mark
= z
->_watermark
[WMARK_MIN
];
3286 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
,
3287 alloc_flags
, free_pages
);
3293 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3294 unsigned long mark
, int highest_zoneidx
)
3296 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3298 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3299 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3301 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, 0,
3306 int __read_mostly node_reclaim_distance
= RECLAIM_DISTANCE
;
3308 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3310 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3311 node_reclaim_distance
;
3313 #else /* CONFIG_NUMA */
3314 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3318 #endif /* CONFIG_NUMA */
3321 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3322 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3323 * premature use of a lower zone may cause lowmem pressure problems that
3324 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3325 * probably too small. It only makes sense to spread allocations to avoid
3326 * fragmentation between the Normal and DMA32 zones.
3328 static inline unsigned int
3329 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3331 unsigned int alloc_flags
;
3334 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3337 alloc_flags
= (__force
int) (gfp_mask
& __GFP_KSWAPD_RECLAIM
);
3339 #ifdef CONFIG_ZONE_DMA32
3343 if (zone_idx(zone
) != ZONE_NORMAL
)
3347 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3348 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3349 * on UMA that if Normal is populated then so is DMA32.
3351 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3352 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3355 alloc_flags
|= ALLOC_NOFRAGMENT
;
3356 #endif /* CONFIG_ZONE_DMA32 */
3360 /* Must be called after current_gfp_context() which can change gfp_mask */
3361 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask
,
3362 unsigned int alloc_flags
)
3365 if (gfp_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3366 alloc_flags
|= ALLOC_CMA
;
3372 * get_page_from_freelist goes through the zonelist trying to allocate
3375 static struct page
*
3376 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3377 const struct alloc_context
*ac
)
3381 struct pglist_data
*last_pgdat
= NULL
;
3382 bool last_pgdat_dirty_ok
= false;
3387 * Scan zonelist, looking for a zone with enough free.
3388 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3390 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3391 z
= ac
->preferred_zoneref
;
3392 for_next_zone_zonelist_nodemask(zone
, z
, ac
->highest_zoneidx
,
3397 if (cpusets_enabled() &&
3398 (alloc_flags
& ALLOC_CPUSET
) &&
3399 !__cpuset_zone_allowed(zone
, gfp_mask
))
3402 * When allocating a page cache page for writing, we
3403 * want to get it from a node that is within its dirty
3404 * limit, such that no single node holds more than its
3405 * proportional share of globally allowed dirty pages.
3406 * The dirty limits take into account the node's
3407 * lowmem reserves and high watermark so that kswapd
3408 * should be able to balance it without having to
3409 * write pages from its LRU list.
3411 * XXX: For now, allow allocations to potentially
3412 * exceed the per-node dirty limit in the slowpath
3413 * (spread_dirty_pages unset) before going into reclaim,
3414 * which is important when on a NUMA setup the allowed
3415 * nodes are together not big enough to reach the
3416 * global limit. The proper fix for these situations
3417 * will require awareness of nodes in the
3418 * dirty-throttling and the flusher threads.
3420 if (ac
->spread_dirty_pages
) {
3421 if (last_pgdat
!= zone
->zone_pgdat
) {
3422 last_pgdat
= zone
->zone_pgdat
;
3423 last_pgdat_dirty_ok
= node_dirty_ok(zone
->zone_pgdat
);
3426 if (!last_pgdat_dirty_ok
)
3430 if (no_fallback
&& nr_online_nodes
> 1 &&
3431 zone
!= ac
->preferred_zoneref
->zone
) {
3435 * If moving to a remote node, retry but allow
3436 * fragmenting fallbacks. Locality is more important
3437 * than fragmentation avoidance.
3439 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3440 if (zone_to_nid(zone
) != local_nid
) {
3441 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3446 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3447 if (!zone_watermark_fast(zone
, order
, mark
,
3448 ac
->highest_zoneidx
, alloc_flags
,
3452 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3454 * Watermark failed for this zone, but see if we can
3455 * grow this zone if it contains deferred pages.
3457 if (deferred_pages_enabled()) {
3458 if (_deferred_grow_zone(zone
, order
))
3462 /* Checked here to keep the fast path fast */
3463 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3464 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3467 if (!node_reclaim_enabled() ||
3468 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3471 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3473 case NODE_RECLAIM_NOSCAN
:
3476 case NODE_RECLAIM_FULL
:
3477 /* scanned but unreclaimable */
3480 /* did we reclaim enough */
3481 if (zone_watermark_ok(zone
, order
, mark
,
3482 ac
->highest_zoneidx
, alloc_flags
))
3490 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3491 gfp_mask
, alloc_flags
, ac
->migratetype
);
3493 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3496 * If this is a high-order atomic allocation then check
3497 * if the pageblock should be reserved for the future
3499 if (unlikely(alloc_flags
& ALLOC_HIGHATOMIC
))
3500 reserve_highatomic_pageblock(page
, zone
, order
);
3504 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3505 /* Try again if zone has deferred pages */
3506 if (deferred_pages_enabled()) {
3507 if (_deferred_grow_zone(zone
, order
))
3515 * It's possible on a UMA machine to get through all zones that are
3516 * fragmented. If avoiding fragmentation, reset and try again.
3519 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3526 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3528 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3531 * This documents exceptions given to allocations in certain
3532 * contexts that are allowed to allocate outside current's set
3535 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3536 if (tsk_is_oom_victim(current
) ||
3537 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3538 filter
&= ~SHOW_MEM_FILTER_NODES
;
3539 if (!in_task() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3540 filter
&= ~SHOW_MEM_FILTER_NODES
;
3542 __show_mem(filter
, nodemask
, gfp_zone(gfp_mask
));
3545 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3547 struct va_format vaf
;
3549 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
3551 if ((gfp_mask
& __GFP_NOWARN
) ||
3552 !__ratelimit(&nopage_rs
) ||
3553 ((gfp_mask
& __GFP_DMA
) && !has_managed_dma()))
3556 va_start(args
, fmt
);
3559 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3560 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3561 nodemask_pr_args(nodemask
));
3564 cpuset_print_current_mems_allowed();
3567 warn_alloc_show_mem(gfp_mask
, nodemask
);
3570 static inline struct page
*
3571 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3572 unsigned int alloc_flags
,
3573 const struct alloc_context
*ac
)
3577 page
= get_page_from_freelist(gfp_mask
, order
,
3578 alloc_flags
|ALLOC_CPUSET
, ac
);
3580 * fallback to ignore cpuset restriction if our nodes
3584 page
= get_page_from_freelist(gfp_mask
, order
,
3590 static inline struct page
*
3591 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3592 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3594 struct oom_control oc
= {
3595 .zonelist
= ac
->zonelist
,
3596 .nodemask
= ac
->nodemask
,
3598 .gfp_mask
= gfp_mask
,
3603 *did_some_progress
= 0;
3606 * Acquire the oom lock. If that fails, somebody else is
3607 * making progress for us.
3609 if (!mutex_trylock(&oom_lock
)) {
3610 *did_some_progress
= 1;
3611 schedule_timeout_uninterruptible(1);
3616 * Go through the zonelist yet one more time, keep very high watermark
3617 * here, this is only to catch a parallel oom killing, we must fail if
3618 * we're still under heavy pressure. But make sure that this reclaim
3619 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3620 * allocation which will never fail due to oom_lock already held.
3622 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3623 ~__GFP_DIRECT_RECLAIM
, order
,
3624 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3628 /* Coredumps can quickly deplete all memory reserves */
3629 if (current
->flags
& PF_DUMPCORE
)
3631 /* The OOM killer will not help higher order allocs */
3632 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3635 * We have already exhausted all our reclaim opportunities without any
3636 * success so it is time to admit defeat. We will skip the OOM killer
3637 * because it is very likely that the caller has a more reasonable
3638 * fallback than shooting a random task.
3640 * The OOM killer may not free memory on a specific node.
3642 if (gfp_mask
& (__GFP_RETRY_MAYFAIL
| __GFP_THISNODE
))
3644 /* The OOM killer does not needlessly kill tasks for lowmem */
3645 if (ac
->highest_zoneidx
< ZONE_NORMAL
)
3647 if (pm_suspended_storage())
3650 * XXX: GFP_NOFS allocations should rather fail than rely on
3651 * other request to make a forward progress.
3652 * We are in an unfortunate situation where out_of_memory cannot
3653 * do much for this context but let's try it to at least get
3654 * access to memory reserved if the current task is killed (see
3655 * out_of_memory). Once filesystems are ready to handle allocation
3656 * failures more gracefully we should just bail out here.
3659 /* Exhausted what can be done so it's blame time */
3660 if (out_of_memory(&oc
) ||
3661 WARN_ON_ONCE_GFP(gfp_mask
& __GFP_NOFAIL
, gfp_mask
)) {
3662 *did_some_progress
= 1;
3665 * Help non-failing allocations by giving them access to memory
3668 if (gfp_mask
& __GFP_NOFAIL
)
3669 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3670 ALLOC_NO_WATERMARKS
, ac
);
3673 mutex_unlock(&oom_lock
);
3678 * Maximum number of compaction retries with a progress before OOM
3679 * killer is consider as the only way to move forward.
3681 #define MAX_COMPACT_RETRIES 16
3683 #ifdef CONFIG_COMPACTION
3684 /* Try memory compaction for high-order allocations before reclaim */
3685 static struct page
*
3686 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3687 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3688 enum compact_priority prio
, enum compact_result
*compact_result
)
3690 struct page
*page
= NULL
;
3691 unsigned long pflags
;
3692 unsigned int noreclaim_flag
;
3697 psi_memstall_enter(&pflags
);
3698 delayacct_compact_start();
3699 noreclaim_flag
= memalloc_noreclaim_save();
3701 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3704 memalloc_noreclaim_restore(noreclaim_flag
);
3705 psi_memstall_leave(&pflags
);
3706 delayacct_compact_end();
3708 if (*compact_result
== COMPACT_SKIPPED
)
3711 * At least in one zone compaction wasn't deferred or skipped, so let's
3712 * count a compaction stall
3714 count_vm_event(COMPACTSTALL
);
3716 /* Prep a captured page if available */
3718 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3720 /* Try get a page from the freelist if available */
3722 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3725 struct zone
*zone
= page_zone(page
);
3727 zone
->compact_blockskip_flush
= false;
3728 compaction_defer_reset(zone
, order
, true);
3729 count_vm_event(COMPACTSUCCESS
);
3734 * It's bad if compaction run occurs and fails. The most likely reason
3735 * is that pages exist, but not enough to satisfy watermarks.
3737 count_vm_event(COMPACTFAIL
);
3745 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3746 enum compact_result compact_result
,
3747 enum compact_priority
*compact_priority
,
3748 int *compaction_retries
)
3750 int max_retries
= MAX_COMPACT_RETRIES
;
3753 int retries
= *compaction_retries
;
3754 enum compact_priority priority
= *compact_priority
;
3759 if (fatal_signal_pending(current
))
3762 if (compaction_made_progress(compact_result
))
3763 (*compaction_retries
)++;
3766 * compaction considers all the zone as desperately out of memory
3767 * so it doesn't really make much sense to retry except when the
3768 * failure could be caused by insufficient priority
3770 if (compaction_failed(compact_result
))
3771 goto check_priority
;
3774 * compaction was skipped because there are not enough order-0 pages
3775 * to work with, so we retry only if it looks like reclaim can help.
3777 if (compaction_needs_reclaim(compact_result
)) {
3778 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3783 * make sure the compaction wasn't deferred or didn't bail out early
3784 * due to locks contention before we declare that we should give up.
3785 * But the next retry should use a higher priority if allowed, so
3786 * we don't just keep bailing out endlessly.
3788 if (compaction_withdrawn(compact_result
)) {
3789 goto check_priority
;
3793 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3794 * costly ones because they are de facto nofail and invoke OOM
3795 * killer to move on while costly can fail and users are ready
3796 * to cope with that. 1/4 retries is rather arbitrary but we
3797 * would need much more detailed feedback from compaction to
3798 * make a better decision.
3800 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3802 if (*compaction_retries
<= max_retries
) {
3808 * Make sure there are attempts at the highest priority if we exhausted
3809 * all retries or failed at the lower priorities.
3812 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3813 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3815 if (*compact_priority
> min_priority
) {
3816 (*compact_priority
)--;
3817 *compaction_retries
= 0;
3821 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3825 static inline struct page
*
3826 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3827 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3828 enum compact_priority prio
, enum compact_result
*compact_result
)
3830 *compact_result
= COMPACT_SKIPPED
;
3835 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3836 enum compact_result compact_result
,
3837 enum compact_priority
*compact_priority
,
3838 int *compaction_retries
)
3843 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3847 * There are setups with compaction disabled which would prefer to loop
3848 * inside the allocator rather than hit the oom killer prematurely.
3849 * Let's give them a good hope and keep retrying while the order-0
3850 * watermarks are OK.
3852 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3853 ac
->highest_zoneidx
, ac
->nodemask
) {
3854 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3855 ac
->highest_zoneidx
, alloc_flags
))
3860 #endif /* CONFIG_COMPACTION */
3862 #ifdef CONFIG_LOCKDEP
3863 static struct lockdep_map __fs_reclaim_map
=
3864 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3866 static bool __need_reclaim(gfp_t gfp_mask
)
3868 /* no reclaim without waiting on it */
3869 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3872 /* this guy won't enter reclaim */
3873 if (current
->flags
& PF_MEMALLOC
)
3876 if (gfp_mask
& __GFP_NOLOCKDEP
)
3882 void __fs_reclaim_acquire(unsigned long ip
)
3884 lock_acquire_exclusive(&__fs_reclaim_map
, 0, 0, NULL
, ip
);
3887 void __fs_reclaim_release(unsigned long ip
)
3889 lock_release(&__fs_reclaim_map
, ip
);
3892 void fs_reclaim_acquire(gfp_t gfp_mask
)
3894 gfp_mask
= current_gfp_context(gfp_mask
);
3896 if (__need_reclaim(gfp_mask
)) {
3897 if (gfp_mask
& __GFP_FS
)
3898 __fs_reclaim_acquire(_RET_IP_
);
3900 #ifdef CONFIG_MMU_NOTIFIER
3901 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
3902 lock_map_release(&__mmu_notifier_invalidate_range_start_map
);
3907 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3909 void fs_reclaim_release(gfp_t gfp_mask
)
3911 gfp_mask
= current_gfp_context(gfp_mask
);
3913 if (__need_reclaim(gfp_mask
)) {
3914 if (gfp_mask
& __GFP_FS
)
3915 __fs_reclaim_release(_RET_IP_
);
3918 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3922 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3923 * have been rebuilt so allocation retries. Reader side does not lock and
3924 * retries the allocation if zonelist changes. Writer side is protected by the
3925 * embedded spin_lock.
3927 static DEFINE_SEQLOCK(zonelist_update_seq
);
3929 static unsigned int zonelist_iter_begin(void)
3931 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE
))
3932 return read_seqbegin(&zonelist_update_seq
);
3937 static unsigned int check_retry_zonelist(unsigned int seq
)
3939 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE
))
3940 return read_seqretry(&zonelist_update_seq
, seq
);
3945 /* Perform direct synchronous page reclaim */
3946 static unsigned long
3947 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3948 const struct alloc_context
*ac
)
3950 unsigned int noreclaim_flag
;
3951 unsigned long progress
;
3955 /* We now go into synchronous reclaim */
3956 cpuset_memory_pressure_bump();
3957 fs_reclaim_acquire(gfp_mask
);
3958 noreclaim_flag
= memalloc_noreclaim_save();
3960 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3963 memalloc_noreclaim_restore(noreclaim_flag
);
3964 fs_reclaim_release(gfp_mask
);
3971 /* The really slow allocator path where we enter direct reclaim */
3972 static inline struct page
*
3973 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3974 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3975 unsigned long *did_some_progress
)
3977 struct page
*page
= NULL
;
3978 unsigned long pflags
;
3979 bool drained
= false;
3981 psi_memstall_enter(&pflags
);
3982 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3983 if (unlikely(!(*did_some_progress
)))
3987 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3990 * If an allocation failed after direct reclaim, it could be because
3991 * pages are pinned on the per-cpu lists or in high alloc reserves.
3992 * Shrink them and try again
3994 if (!page
&& !drained
) {
3995 unreserve_highatomic_pageblock(ac
, false);
3996 drain_all_pages(NULL
);
4001 psi_memstall_leave(&pflags
);
4006 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4007 const struct alloc_context
*ac
)
4011 pg_data_t
*last_pgdat
= NULL
;
4012 enum zone_type highest_zoneidx
= ac
->highest_zoneidx
;
4014 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, highest_zoneidx
,
4016 if (!managed_zone(zone
))
4018 if (last_pgdat
!= zone
->zone_pgdat
) {
4019 wakeup_kswapd(zone
, gfp_mask
, order
, highest_zoneidx
);
4020 last_pgdat
= zone
->zone_pgdat
;
4025 static inline unsigned int
4026 gfp_to_alloc_flags(gfp_t gfp_mask
, unsigned int order
)
4028 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4031 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
4032 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4033 * to save two branches.
4035 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_MIN_RESERVE
);
4036 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM
!= (__force gfp_t
) ALLOC_KSWAPD
);
4039 * The caller may dip into page reserves a bit more if the caller
4040 * cannot run direct reclaim, or if the caller has realtime scheduling
4041 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4042 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
4044 alloc_flags
|= (__force
int)
4045 (gfp_mask
& (__GFP_HIGH
| __GFP_KSWAPD_RECLAIM
));
4047 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
)) {
4049 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4050 * if it can't schedule.
4052 if (!(gfp_mask
& __GFP_NOMEMALLOC
)) {
4053 alloc_flags
|= ALLOC_NON_BLOCK
;
4056 alloc_flags
|= ALLOC_HIGHATOMIC
;
4060 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
4061 * GFP_ATOMIC) rather than fail, see the comment for
4062 * cpuset_node_allowed().
4064 if (alloc_flags
& ALLOC_MIN_RESERVE
)
4065 alloc_flags
&= ~ALLOC_CPUSET
;
4066 } else if (unlikely(rt_task(current
)) && in_task())
4067 alloc_flags
|= ALLOC_MIN_RESERVE
;
4069 alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, alloc_flags
);
4074 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4076 if (!tsk_is_oom_victim(tsk
))
4080 * !MMU doesn't have oom reaper so give access to memory reserves
4081 * only to the thread with TIF_MEMDIE set
4083 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4090 * Distinguish requests which really need access to full memory
4091 * reserves from oom victims which can live with a portion of it
4093 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4095 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4097 if (gfp_mask
& __GFP_MEMALLOC
)
4098 return ALLOC_NO_WATERMARKS
;
4099 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4100 return ALLOC_NO_WATERMARKS
;
4101 if (!in_interrupt()) {
4102 if (current
->flags
& PF_MEMALLOC
)
4103 return ALLOC_NO_WATERMARKS
;
4104 else if (oom_reserves_allowed(current
))
4111 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4113 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4117 * Checks whether it makes sense to retry the reclaim to make a forward progress
4118 * for the given allocation request.
4120 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4121 * without success, or when we couldn't even meet the watermark if we
4122 * reclaimed all remaining pages on the LRU lists.
4124 * Returns true if a retry is viable or false to enter the oom path.
4127 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4128 struct alloc_context
*ac
, int alloc_flags
,
4129 bool did_some_progress
, int *no_progress_loops
)
4136 * Costly allocations might have made a progress but this doesn't mean
4137 * their order will become available due to high fragmentation so
4138 * always increment the no progress counter for them
4140 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4141 *no_progress_loops
= 0;
4143 (*no_progress_loops
)++;
4146 * Make sure we converge to OOM if we cannot make any progress
4147 * several times in the row.
4149 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4150 /* Before OOM, exhaust highatomic_reserve */
4151 return unreserve_highatomic_pageblock(ac
, true);
4155 * Keep reclaiming pages while there is a chance this will lead
4156 * somewhere. If none of the target zones can satisfy our allocation
4157 * request even if all reclaimable pages are considered then we are
4158 * screwed and have to go OOM.
4160 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4161 ac
->highest_zoneidx
, ac
->nodemask
) {
4162 unsigned long available
;
4163 unsigned long reclaimable
;
4164 unsigned long min_wmark
= min_wmark_pages(zone
);
4167 available
= reclaimable
= zone_reclaimable_pages(zone
);
4168 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4171 * Would the allocation succeed if we reclaimed all
4172 * reclaimable pages?
4174 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4175 ac
->highest_zoneidx
, alloc_flags
, available
);
4176 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4177 available
, min_wmark
, *no_progress_loops
, wmark
);
4185 * Memory allocation/reclaim might be called from a WQ context and the
4186 * current implementation of the WQ concurrency control doesn't
4187 * recognize that a particular WQ is congested if the worker thread is
4188 * looping without ever sleeping. Therefore we have to do a short sleep
4189 * here rather than calling cond_resched().
4191 if (current
->flags
& PF_WQ_WORKER
)
4192 schedule_timeout_uninterruptible(1);
4199 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4202 * It's possible that cpuset's mems_allowed and the nodemask from
4203 * mempolicy don't intersect. This should be normally dealt with by
4204 * policy_nodemask(), but it's possible to race with cpuset update in
4205 * such a way the check therein was true, and then it became false
4206 * before we got our cpuset_mems_cookie here.
4207 * This assumes that for all allocations, ac->nodemask can come only
4208 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4209 * when it does not intersect with the cpuset restrictions) or the
4210 * caller can deal with a violated nodemask.
4212 if (cpusets_enabled() && ac
->nodemask
&&
4213 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4214 ac
->nodemask
= NULL
;
4219 * When updating a task's mems_allowed or mempolicy nodemask, it is
4220 * possible to race with parallel threads in such a way that our
4221 * allocation can fail while the mask is being updated. If we are about
4222 * to fail, check if the cpuset changed during allocation and if so,
4225 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4231 static inline struct page
*
4232 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4233 struct alloc_context
*ac
)
4235 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4236 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4237 struct page
*page
= NULL
;
4238 unsigned int alloc_flags
;
4239 unsigned long did_some_progress
;
4240 enum compact_priority compact_priority
;
4241 enum compact_result compact_result
;
4242 int compaction_retries
;
4243 int no_progress_loops
;
4244 unsigned int cpuset_mems_cookie
;
4245 unsigned int zonelist_iter_cookie
;
4249 compaction_retries
= 0;
4250 no_progress_loops
= 0;
4251 compact_priority
= DEF_COMPACT_PRIORITY
;
4252 cpuset_mems_cookie
= read_mems_allowed_begin();
4253 zonelist_iter_cookie
= zonelist_iter_begin();
4256 * The fast path uses conservative alloc_flags to succeed only until
4257 * kswapd needs to be woken up, and to avoid the cost of setting up
4258 * alloc_flags precisely. So we do that now.
4260 alloc_flags
= gfp_to_alloc_flags(gfp_mask
, order
);
4263 * We need to recalculate the starting point for the zonelist iterator
4264 * because we might have used different nodemask in the fast path, or
4265 * there was a cpuset modification and we are retrying - otherwise we
4266 * could end up iterating over non-eligible zones endlessly.
4268 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4269 ac
->highest_zoneidx
, ac
->nodemask
);
4270 if (!ac
->preferred_zoneref
->zone
)
4274 * Check for insane configurations where the cpuset doesn't contain
4275 * any suitable zone to satisfy the request - e.g. non-movable
4276 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4278 if (cpusets_insane_config() && (gfp_mask
& __GFP_HARDWALL
)) {
4279 struct zoneref
*z
= first_zones_zonelist(ac
->zonelist
,
4280 ac
->highest_zoneidx
,
4281 &cpuset_current_mems_allowed
);
4286 if (alloc_flags
& ALLOC_KSWAPD
)
4287 wake_all_kswapds(order
, gfp_mask
, ac
);
4290 * The adjusted alloc_flags might result in immediate success, so try
4293 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4298 * For costly allocations, try direct compaction first, as it's likely
4299 * that we have enough base pages and don't need to reclaim. For non-
4300 * movable high-order allocations, do that as well, as compaction will
4301 * try prevent permanent fragmentation by migrating from blocks of the
4303 * Don't try this for allocations that are allowed to ignore
4304 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4306 if (can_direct_reclaim
&&
4308 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4309 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4310 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4312 INIT_COMPACT_PRIORITY
,
4318 * Checks for costly allocations with __GFP_NORETRY, which
4319 * includes some THP page fault allocations
4321 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4323 * If allocating entire pageblock(s) and compaction
4324 * failed because all zones are below low watermarks
4325 * or is prohibited because it recently failed at this
4326 * order, fail immediately unless the allocator has
4327 * requested compaction and reclaim retry.
4330 * - potentially very expensive because zones are far
4331 * below their low watermarks or this is part of very
4332 * bursty high order allocations,
4333 * - not guaranteed to help because isolate_freepages()
4334 * may not iterate over freed pages as part of its
4336 * - unlikely to make entire pageblocks free on its
4339 if (compact_result
== COMPACT_SKIPPED
||
4340 compact_result
== COMPACT_DEFERRED
)
4344 * Looks like reclaim/compaction is worth trying, but
4345 * sync compaction could be very expensive, so keep
4346 * using async compaction.
4348 compact_priority
= INIT_COMPACT_PRIORITY
;
4353 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4354 if (alloc_flags
& ALLOC_KSWAPD
)
4355 wake_all_kswapds(order
, gfp_mask
, ac
);
4357 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4359 alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, reserve_flags
) |
4360 (alloc_flags
& ALLOC_KSWAPD
);
4363 * Reset the nodemask and zonelist iterators if memory policies can be
4364 * ignored. These allocations are high priority and system rather than
4367 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4368 ac
->nodemask
= NULL
;
4369 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4370 ac
->highest_zoneidx
, ac
->nodemask
);
4373 /* Attempt with potentially adjusted zonelist and alloc_flags */
4374 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4378 /* Caller is not willing to reclaim, we can't balance anything */
4379 if (!can_direct_reclaim
)
4382 /* Avoid recursion of direct reclaim */
4383 if (current
->flags
& PF_MEMALLOC
)
4386 /* Try direct reclaim and then allocating */
4387 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4388 &did_some_progress
);
4392 /* Try direct compaction and then allocating */
4393 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4394 compact_priority
, &compact_result
);
4398 /* Do not loop if specifically requested */
4399 if (gfp_mask
& __GFP_NORETRY
)
4403 * Do not retry costly high order allocations unless they are
4404 * __GFP_RETRY_MAYFAIL
4406 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4409 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4410 did_some_progress
> 0, &no_progress_loops
))
4414 * It doesn't make any sense to retry for the compaction if the order-0
4415 * reclaim is not able to make any progress because the current
4416 * implementation of the compaction depends on the sufficient amount
4417 * of free memory (see __compaction_suitable)
4419 if (did_some_progress
> 0 &&
4420 should_compact_retry(ac
, order
, alloc_flags
,
4421 compact_result
, &compact_priority
,
4422 &compaction_retries
))
4427 * Deal with possible cpuset update races or zonelist updates to avoid
4428 * a unnecessary OOM kill.
4430 if (check_retry_cpuset(cpuset_mems_cookie
, ac
) ||
4431 check_retry_zonelist(zonelist_iter_cookie
))
4434 /* Reclaim has failed us, start killing things */
4435 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4439 /* Avoid allocations with no watermarks from looping endlessly */
4440 if (tsk_is_oom_victim(current
) &&
4441 (alloc_flags
& ALLOC_OOM
||
4442 (gfp_mask
& __GFP_NOMEMALLOC
)))
4445 /* Retry as long as the OOM killer is making progress */
4446 if (did_some_progress
) {
4447 no_progress_loops
= 0;
4453 * Deal with possible cpuset update races or zonelist updates to avoid
4454 * a unnecessary OOM kill.
4456 if (check_retry_cpuset(cpuset_mems_cookie
, ac
) ||
4457 check_retry_zonelist(zonelist_iter_cookie
))
4461 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4464 if (gfp_mask
& __GFP_NOFAIL
) {
4466 * All existing users of the __GFP_NOFAIL are blockable, so warn
4467 * of any new users that actually require GFP_NOWAIT
4469 if (WARN_ON_ONCE_GFP(!can_direct_reclaim
, gfp_mask
))
4473 * PF_MEMALLOC request from this context is rather bizarre
4474 * because we cannot reclaim anything and only can loop waiting
4475 * for somebody to do a work for us
4477 WARN_ON_ONCE_GFP(current
->flags
& PF_MEMALLOC
, gfp_mask
);
4480 * non failing costly orders are a hard requirement which we
4481 * are not prepared for much so let's warn about these users
4482 * so that we can identify them and convert them to something
4485 WARN_ON_ONCE_GFP(costly_order
, gfp_mask
);
4488 * Help non-failing allocations by giving some access to memory
4489 * reserves normally used for high priority non-blocking
4490 * allocations but do not use ALLOC_NO_WATERMARKS because this
4491 * could deplete whole memory reserves which would just make
4492 * the situation worse.
4494 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_MIN_RESERVE
, ac
);
4502 warn_alloc(gfp_mask
, ac
->nodemask
,
4503 "page allocation failure: order:%u", order
);
4508 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4509 int preferred_nid
, nodemask_t
*nodemask
,
4510 struct alloc_context
*ac
, gfp_t
*alloc_gfp
,
4511 unsigned int *alloc_flags
)
4513 ac
->highest_zoneidx
= gfp_zone(gfp_mask
);
4514 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4515 ac
->nodemask
= nodemask
;
4516 ac
->migratetype
= gfp_migratetype(gfp_mask
);
4518 if (cpusets_enabled()) {
4519 *alloc_gfp
|= __GFP_HARDWALL
;
4521 * When we are in the interrupt context, it is irrelevant
4522 * to the current task context. It means that any node ok.
4524 if (in_task() && !ac
->nodemask
)
4525 ac
->nodemask
= &cpuset_current_mems_allowed
;
4527 *alloc_flags
|= ALLOC_CPUSET
;
4530 might_alloc(gfp_mask
);
4532 if (should_fail_alloc_page(gfp_mask
, order
))
4535 *alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, *alloc_flags
);
4537 /* Dirty zone balancing only done in the fast path */
4538 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4541 * The preferred zone is used for statistics but crucially it is
4542 * also used as the starting point for the zonelist iterator. It
4543 * may get reset for allocations that ignore memory policies.
4545 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4546 ac
->highest_zoneidx
, ac
->nodemask
);
4552 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4553 * @gfp: GFP flags for the allocation
4554 * @preferred_nid: The preferred NUMA node ID to allocate from
4555 * @nodemask: Set of nodes to allocate from, may be NULL
4556 * @nr_pages: The number of pages desired on the list or array
4557 * @page_list: Optional list to store the allocated pages
4558 * @page_array: Optional array to store the pages
4560 * This is a batched version of the page allocator that attempts to
4561 * allocate nr_pages quickly. Pages are added to page_list if page_list
4562 * is not NULL, otherwise it is assumed that the page_array is valid.
4564 * For lists, nr_pages is the number of pages that should be allocated.
4566 * For arrays, only NULL elements are populated with pages and nr_pages
4567 * is the maximum number of pages that will be stored in the array.
4569 * Returns the number of pages on the list or array.
4571 unsigned long __alloc_pages_bulk(gfp_t gfp
, int preferred_nid
,
4572 nodemask_t
*nodemask
, int nr_pages
,
4573 struct list_head
*page_list
,
4574 struct page
**page_array
)
4577 unsigned long __maybe_unused UP_flags
;
4580 struct per_cpu_pages
*pcp
;
4581 struct list_head
*pcp_list
;
4582 struct alloc_context ac
;
4584 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4585 int nr_populated
= 0, nr_account
= 0;
4588 * Skip populated array elements to determine if any pages need
4589 * to be allocated before disabling IRQs.
4591 while (page_array
&& nr_populated
< nr_pages
&& page_array
[nr_populated
])
4594 /* No pages requested? */
4595 if (unlikely(nr_pages
<= 0))
4598 /* Already populated array? */
4599 if (unlikely(page_array
&& nr_pages
- nr_populated
== 0))
4602 /* Bulk allocator does not support memcg accounting. */
4603 if (memcg_kmem_online() && (gfp
& __GFP_ACCOUNT
))
4606 /* Use the single page allocator for one page. */
4607 if (nr_pages
- nr_populated
== 1)
4610 #ifdef CONFIG_PAGE_OWNER
4612 * PAGE_OWNER may recurse into the allocator to allocate space to
4613 * save the stack with pagesets.lock held. Releasing/reacquiring
4614 * removes much of the performance benefit of bulk allocation so
4615 * force the caller to allocate one page at a time as it'll have
4616 * similar performance to added complexity to the bulk allocator.
4618 if (static_branch_unlikely(&page_owner_inited
))
4622 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4623 gfp
&= gfp_allowed_mask
;
4625 if (!prepare_alloc_pages(gfp
, 0, preferred_nid
, nodemask
, &ac
, &alloc_gfp
, &alloc_flags
))
4629 /* Find an allowed local zone that meets the low watermark. */
4630 for_each_zone_zonelist_nodemask(zone
, z
, ac
.zonelist
, ac
.highest_zoneidx
, ac
.nodemask
) {
4633 if (cpusets_enabled() && (alloc_flags
& ALLOC_CPUSET
) &&
4634 !__cpuset_zone_allowed(zone
, gfp
)) {
4638 if (nr_online_nodes
> 1 && zone
!= ac
.preferred_zoneref
->zone
&&
4639 zone_to_nid(zone
) != zone_to_nid(ac
.preferred_zoneref
->zone
)) {
4643 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
) + nr_pages
;
4644 if (zone_watermark_fast(zone
, 0, mark
,
4645 zonelist_zone_idx(ac
.preferred_zoneref
),
4646 alloc_flags
, gfp
)) {
4652 * If there are no allowed local zones that meets the watermarks then
4653 * try to allocate a single page and reclaim if necessary.
4655 if (unlikely(!zone
))
4658 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4659 pcp_trylock_prepare(UP_flags
);
4660 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
4664 /* Attempt the batch allocation */
4665 pcp_list
= &pcp
->lists
[order_to_pindex(ac
.migratetype
, 0)];
4666 while (nr_populated
< nr_pages
) {
4668 /* Skip existing pages */
4669 if (page_array
&& page_array
[nr_populated
]) {
4674 page
= __rmqueue_pcplist(zone
, 0, ac
.migratetype
, alloc_flags
,
4676 if (unlikely(!page
)) {
4677 /* Try and allocate at least one page */
4679 pcp_spin_unlock(pcp
);
4686 prep_new_page(page
, 0, gfp
, 0);
4688 list_add(&page
->lru
, page_list
);
4690 page_array
[nr_populated
] = page
;
4694 pcp_spin_unlock(pcp
);
4695 pcp_trylock_finish(UP_flags
);
4697 __count_zid_vm_events(PGALLOC
, zone_idx(zone
), nr_account
);
4698 zone_statistics(ac
.preferred_zoneref
->zone
, zone
, nr_account
);
4701 return nr_populated
;
4704 pcp_trylock_finish(UP_flags
);
4707 page
= __alloc_pages(gfp
, 0, preferred_nid
, nodemask
);
4710 list_add(&page
->lru
, page_list
);
4712 page_array
[nr_populated
] = page
;
4718 EXPORT_SYMBOL_GPL(__alloc_pages_bulk
);
4721 * This is the 'heart' of the zoned buddy allocator.
4723 struct page
*__alloc_pages(gfp_t gfp
, unsigned int order
, int preferred_nid
,
4724 nodemask_t
*nodemask
)
4727 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4728 gfp_t alloc_gfp
; /* The gfp_t that was actually used for allocation */
4729 struct alloc_context ac
= { };
4732 * There are several places where we assume that the order value is sane
4733 * so bail out early if the request is out of bound.
4735 if (WARN_ON_ONCE_GFP(order
> MAX_ORDER
, gfp
))
4738 gfp
&= gfp_allowed_mask
;
4740 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4741 * resp. GFP_NOIO which has to be inherited for all allocation requests
4742 * from a particular context which has been marked by
4743 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4744 * movable zones are not used during allocation.
4746 gfp
= current_gfp_context(gfp
);
4748 if (!prepare_alloc_pages(gfp
, order
, preferred_nid
, nodemask
, &ac
,
4749 &alloc_gfp
, &alloc_flags
))
4753 * Forbid the first pass from falling back to types that fragment
4754 * memory until all local zones are considered.
4756 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp
);
4758 /* First allocation attempt */
4759 page
= get_page_from_freelist(alloc_gfp
, order
, alloc_flags
, &ac
);
4764 ac
.spread_dirty_pages
= false;
4767 * Restore the original nodemask if it was potentially replaced with
4768 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4770 ac
.nodemask
= nodemask
;
4772 page
= __alloc_pages_slowpath(alloc_gfp
, order
, &ac
);
4775 if (memcg_kmem_online() && (gfp
& __GFP_ACCOUNT
) && page
&&
4776 unlikely(__memcg_kmem_charge_page(page
, gfp
, order
) != 0)) {
4777 __free_pages(page
, order
);
4781 trace_mm_page_alloc(page
, order
, alloc_gfp
, ac
.migratetype
);
4782 kmsan_alloc_page(page
, order
, alloc_gfp
);
4786 EXPORT_SYMBOL(__alloc_pages
);
4788 struct folio
*__folio_alloc(gfp_t gfp
, unsigned int order
, int preferred_nid
,
4789 nodemask_t
*nodemask
)
4791 struct page
*page
= __alloc_pages(gfp
| __GFP_COMP
, order
,
4792 preferred_nid
, nodemask
);
4794 if (page
&& order
> 1)
4795 prep_transhuge_page(page
);
4796 return (struct folio
*)page
;
4798 EXPORT_SYMBOL(__folio_alloc
);
4801 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4802 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4803 * you need to access high mem.
4805 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4809 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4812 return (unsigned long) page_address(page
);
4814 EXPORT_SYMBOL(__get_free_pages
);
4816 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4818 return __get_free_page(gfp_mask
| __GFP_ZERO
);
4820 EXPORT_SYMBOL(get_zeroed_page
);
4823 * __free_pages - Free pages allocated with alloc_pages().
4824 * @page: The page pointer returned from alloc_pages().
4825 * @order: The order of the allocation.
4827 * This function can free multi-page allocations that are not compound
4828 * pages. It does not check that the @order passed in matches that of
4829 * the allocation, so it is easy to leak memory. Freeing more memory
4830 * than was allocated will probably emit a warning.
4832 * If the last reference to this page is speculative, it will be released
4833 * by put_page() which only frees the first page of a non-compound
4834 * allocation. To prevent the remaining pages from being leaked, we free
4835 * the subsequent pages here. If you want to use the page's reference
4836 * count to decide when to free the allocation, you should allocate a
4837 * compound page, and use put_page() instead of __free_pages().
4839 * Context: May be called in interrupt context or while holding a normal
4840 * spinlock, but not in NMI context or while holding a raw spinlock.
4842 void __free_pages(struct page
*page
, unsigned int order
)
4844 /* get PageHead before we drop reference */
4845 int head
= PageHead(page
);
4847 if (put_page_testzero(page
))
4848 free_the_page(page
, order
);
4851 free_the_page(page
+ (1 << order
), order
);
4853 EXPORT_SYMBOL(__free_pages
);
4855 void free_pages(unsigned long addr
, unsigned int order
)
4858 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4859 __free_pages(virt_to_page((void *)addr
), order
);
4863 EXPORT_SYMBOL(free_pages
);
4867 * An arbitrary-length arbitrary-offset area of memory which resides
4868 * within a 0 or higher order page. Multiple fragments within that page
4869 * are individually refcounted, in the page's reference counter.
4871 * The page_frag functions below provide a simple allocation framework for
4872 * page fragments. This is used by the network stack and network device
4873 * drivers to provide a backing region of memory for use as either an
4874 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4876 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4879 struct page
*page
= NULL
;
4880 gfp_t gfp
= gfp_mask
;
4882 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4883 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4885 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4886 PAGE_FRAG_CACHE_MAX_ORDER
);
4887 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4889 if (unlikely(!page
))
4890 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4892 nc
->va
= page
? page_address(page
) : NULL
;
4897 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4899 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4901 if (page_ref_sub_and_test(page
, count
))
4902 free_the_page(page
, compound_order(page
));
4904 EXPORT_SYMBOL(__page_frag_cache_drain
);
4906 void *page_frag_alloc_align(struct page_frag_cache
*nc
,
4907 unsigned int fragsz
, gfp_t gfp_mask
,
4908 unsigned int align_mask
)
4910 unsigned int size
= PAGE_SIZE
;
4914 if (unlikely(!nc
->va
)) {
4916 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4920 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4921 /* if size can vary use size else just use PAGE_SIZE */
4924 /* Even if we own the page, we do not use atomic_set().
4925 * This would break get_page_unless_zero() users.
4927 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
4929 /* reset page count bias and offset to start of new frag */
4930 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4931 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4935 offset
= nc
->offset
- fragsz
;
4936 if (unlikely(offset
< 0)) {
4937 page
= virt_to_page(nc
->va
);
4939 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4942 if (unlikely(nc
->pfmemalloc
)) {
4943 free_the_page(page
, compound_order(page
));
4947 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4948 /* if size can vary use size else just use PAGE_SIZE */
4951 /* OK, page count is 0, we can safely set it */
4952 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
4954 /* reset page count bias and offset to start of new frag */
4955 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4956 offset
= size
- fragsz
;
4957 if (unlikely(offset
< 0)) {
4959 * The caller is trying to allocate a fragment
4960 * with fragsz > PAGE_SIZE but the cache isn't big
4961 * enough to satisfy the request, this may
4962 * happen in low memory conditions.
4963 * We don't release the cache page because
4964 * it could make memory pressure worse
4965 * so we simply return NULL here.
4972 offset
&= align_mask
;
4973 nc
->offset
= offset
;
4975 return nc
->va
+ offset
;
4977 EXPORT_SYMBOL(page_frag_alloc_align
);
4980 * Frees a page fragment allocated out of either a compound or order 0 page.
4982 void page_frag_free(void *addr
)
4984 struct page
*page
= virt_to_head_page(addr
);
4986 if (unlikely(put_page_testzero(page
)))
4987 free_the_page(page
, compound_order(page
));
4989 EXPORT_SYMBOL(page_frag_free
);
4991 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4995 unsigned long nr
= DIV_ROUND_UP(size
, PAGE_SIZE
);
4996 struct page
*page
= virt_to_page((void *)addr
);
4997 struct page
*last
= page
+ nr
;
4999 split_page_owner(page
, 1 << order
);
5000 split_page_memcg(page
, 1 << order
);
5001 while (page
< --last
)
5002 set_page_refcounted(last
);
5004 last
= page
+ (1UL << order
);
5005 for (page
+= nr
; page
< last
; page
++)
5006 __free_pages_ok(page
, 0, FPI_TO_TAIL
);
5008 return (void *)addr
;
5012 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5013 * @size: the number of bytes to allocate
5014 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5016 * This function is similar to alloc_pages(), except that it allocates the
5017 * minimum number of pages to satisfy the request. alloc_pages() can only
5018 * allocate memory in power-of-two pages.
5020 * This function is also limited by MAX_ORDER.
5022 * Memory allocated by this function must be released by free_pages_exact().
5024 * Return: pointer to the allocated area or %NULL in case of error.
5026 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
5028 unsigned int order
= get_order(size
);
5031 if (WARN_ON_ONCE(gfp_mask
& (__GFP_COMP
| __GFP_HIGHMEM
)))
5032 gfp_mask
&= ~(__GFP_COMP
| __GFP_HIGHMEM
);
5034 addr
= __get_free_pages(gfp_mask
, order
);
5035 return make_alloc_exact(addr
, order
, size
);
5037 EXPORT_SYMBOL(alloc_pages_exact
);
5040 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5042 * @nid: the preferred node ID where memory should be allocated
5043 * @size: the number of bytes to allocate
5044 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5046 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5049 * Return: pointer to the allocated area or %NULL in case of error.
5051 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
5053 unsigned int order
= get_order(size
);
5056 if (WARN_ON_ONCE(gfp_mask
& (__GFP_COMP
| __GFP_HIGHMEM
)))
5057 gfp_mask
&= ~(__GFP_COMP
| __GFP_HIGHMEM
);
5059 p
= alloc_pages_node(nid
, gfp_mask
, order
);
5062 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5066 * free_pages_exact - release memory allocated via alloc_pages_exact()
5067 * @virt: the value returned by alloc_pages_exact.
5068 * @size: size of allocation, same value as passed to alloc_pages_exact().
5070 * Release the memory allocated by a previous call to alloc_pages_exact.
5072 void free_pages_exact(void *virt
, size_t size
)
5074 unsigned long addr
= (unsigned long)virt
;
5075 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5077 while (addr
< end
) {
5082 EXPORT_SYMBOL(free_pages_exact
);
5085 * nr_free_zone_pages - count number of pages beyond high watermark
5086 * @offset: The zone index of the highest zone
5088 * nr_free_zone_pages() counts the number of pages which are beyond the
5089 * high watermark within all zones at or below a given zone index. For each
5090 * zone, the number of pages is calculated as:
5092 * nr_free_zone_pages = managed_pages - high_pages
5094 * Return: number of pages beyond high watermark.
5096 static unsigned long nr_free_zone_pages(int offset
)
5101 /* Just pick one node, since fallback list is circular */
5102 unsigned long sum
= 0;
5104 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5106 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5107 unsigned long size
= zone_managed_pages(zone
);
5108 unsigned long high
= high_wmark_pages(zone
);
5117 * nr_free_buffer_pages - count number of pages beyond high watermark
5119 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5120 * watermark within ZONE_DMA and ZONE_NORMAL.
5122 * Return: number of pages beyond high watermark within ZONE_DMA and
5125 unsigned long nr_free_buffer_pages(void)
5127 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5129 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5131 static inline void show_node(struct zone
*zone
)
5133 if (IS_ENABLED(CONFIG_NUMA
))
5134 printk("Node %d ", zone_to_nid(zone
));
5137 long si_mem_available(void)
5140 unsigned long pagecache
;
5141 unsigned long wmark_low
= 0;
5142 unsigned long pages
[NR_LRU_LISTS
];
5143 unsigned long reclaimable
;
5147 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5148 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5151 wmark_low
+= low_wmark_pages(zone
);
5154 * Estimate the amount of memory available for userspace allocations,
5155 * without causing swapping or OOM.
5157 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5160 * Not all the page cache can be freed, otherwise the system will
5161 * start swapping or thrashing. Assume at least half of the page
5162 * cache, or the low watermark worth of cache, needs to stay.
5164 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5165 pagecache
-= min(pagecache
/ 2, wmark_low
);
5166 available
+= pagecache
;
5169 * Part of the reclaimable slab and other kernel memory consists of
5170 * items that are in use, and cannot be freed. Cap this estimate at the
5173 reclaimable
= global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
) +
5174 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5175 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5181 EXPORT_SYMBOL_GPL(si_mem_available
);
5183 void si_meminfo(struct sysinfo
*val
)
5185 val
->totalram
= totalram_pages();
5186 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5187 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5188 val
->bufferram
= nr_blockdev_pages();
5189 val
->totalhigh
= totalhigh_pages();
5190 val
->freehigh
= nr_free_highpages();
5191 val
->mem_unit
= PAGE_SIZE
;
5194 EXPORT_SYMBOL(si_meminfo
);
5197 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5199 int zone_type
; /* needs to be signed */
5200 unsigned long managed_pages
= 0;
5201 unsigned long managed_highpages
= 0;
5202 unsigned long free_highpages
= 0;
5203 pg_data_t
*pgdat
= NODE_DATA(nid
);
5205 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5206 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5207 val
->totalram
= managed_pages
;
5208 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5209 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5210 #ifdef CONFIG_HIGHMEM
5211 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5212 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5214 if (is_highmem(zone
)) {
5215 managed_highpages
+= zone_managed_pages(zone
);
5216 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5219 val
->totalhigh
= managed_highpages
;
5220 val
->freehigh
= free_highpages
;
5222 val
->totalhigh
= managed_highpages
;
5223 val
->freehigh
= free_highpages
;
5225 val
->mem_unit
= PAGE_SIZE
;
5230 * Determine whether the node should be displayed or not, depending on whether
5231 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5233 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5235 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5239 * no node mask - aka implicit memory numa policy. Do not bother with
5240 * the synchronization - read_mems_allowed_begin - because we do not
5241 * have to be precise here.
5244 nodemask
= &cpuset_current_mems_allowed
;
5246 return !node_isset(nid
, *nodemask
);
5249 static void show_migration_types(unsigned char type
)
5251 static const char types
[MIGRATE_TYPES
] = {
5252 [MIGRATE_UNMOVABLE
] = 'U',
5253 [MIGRATE_MOVABLE
] = 'M',
5254 [MIGRATE_RECLAIMABLE
] = 'E',
5255 [MIGRATE_HIGHATOMIC
] = 'H',
5257 [MIGRATE_CMA
] = 'C',
5259 #ifdef CONFIG_MEMORY_ISOLATION
5260 [MIGRATE_ISOLATE
] = 'I',
5263 char tmp
[MIGRATE_TYPES
+ 1];
5267 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5268 if (type
& (1 << i
))
5273 printk(KERN_CONT
"(%s) ", tmp
);
5276 static bool node_has_managed_zones(pg_data_t
*pgdat
, int max_zone_idx
)
5279 for (zone_idx
= 0; zone_idx
<= max_zone_idx
; zone_idx
++)
5280 if (zone_managed_pages(pgdat
->node_zones
+ zone_idx
))
5286 * Show free area list (used inside shift_scroll-lock stuff)
5287 * We also calculate the percentage fragmentation. We do this by counting the
5288 * memory on each free list with the exception of the first item on the list.
5291 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5294 void __show_free_areas(unsigned int filter
, nodemask_t
*nodemask
, int max_zone_idx
)
5296 unsigned long free_pcp
= 0;
5301 for_each_populated_zone(zone
) {
5302 if (zone_idx(zone
) > max_zone_idx
)
5304 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5307 for_each_online_cpu(cpu
)
5308 free_pcp
+= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
)->count
;
5311 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5312 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5313 " unevictable:%lu dirty:%lu writeback:%lu\n"
5314 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5315 " mapped:%lu shmem:%lu pagetables:%lu\n"
5316 " sec_pagetables:%lu bounce:%lu\n"
5317 " kernel_misc_reclaimable:%lu\n"
5318 " free:%lu free_pcp:%lu free_cma:%lu\n",
5319 global_node_page_state(NR_ACTIVE_ANON
),
5320 global_node_page_state(NR_INACTIVE_ANON
),
5321 global_node_page_state(NR_ISOLATED_ANON
),
5322 global_node_page_state(NR_ACTIVE_FILE
),
5323 global_node_page_state(NR_INACTIVE_FILE
),
5324 global_node_page_state(NR_ISOLATED_FILE
),
5325 global_node_page_state(NR_UNEVICTABLE
),
5326 global_node_page_state(NR_FILE_DIRTY
),
5327 global_node_page_state(NR_WRITEBACK
),
5328 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
),
5329 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B
),
5330 global_node_page_state(NR_FILE_MAPPED
),
5331 global_node_page_state(NR_SHMEM
),
5332 global_node_page_state(NR_PAGETABLE
),
5333 global_node_page_state(NR_SECONDARY_PAGETABLE
),
5334 global_zone_page_state(NR_BOUNCE
),
5335 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
),
5336 global_zone_page_state(NR_FREE_PAGES
),
5338 global_zone_page_state(NR_FREE_CMA_PAGES
));
5340 for_each_online_pgdat(pgdat
) {
5341 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5343 if (!node_has_managed_zones(pgdat
, max_zone_idx
))
5347 " active_anon:%lukB"
5348 " inactive_anon:%lukB"
5349 " active_file:%lukB"
5350 " inactive_file:%lukB"
5351 " unevictable:%lukB"
5352 " isolated(anon):%lukB"
5353 " isolated(file):%lukB"
5358 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5360 " shmem_pmdmapped: %lukB"
5363 " writeback_tmp:%lukB"
5364 " kernel_stack:%lukB"
5365 #ifdef CONFIG_SHADOW_CALL_STACK
5366 " shadow_call_stack:%lukB"
5369 " sec_pagetables:%lukB"
5370 " all_unreclaimable? %s"
5373 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5374 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5375 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5376 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5377 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5378 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5379 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5380 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5381 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5382 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5383 K(node_page_state(pgdat
, NR_SHMEM
)),
5384 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5385 K(node_page_state(pgdat
, NR_SHMEM_THPS
)),
5386 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)),
5387 K(node_page_state(pgdat
, NR_ANON_THPS
)),
5389 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5390 node_page_state(pgdat
, NR_KERNEL_STACK_KB
),
5391 #ifdef CONFIG_SHADOW_CALL_STACK
5392 node_page_state(pgdat
, NR_KERNEL_SCS_KB
),
5394 K(node_page_state(pgdat
, NR_PAGETABLE
)),
5395 K(node_page_state(pgdat
, NR_SECONDARY_PAGETABLE
)),
5396 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5400 for_each_populated_zone(zone
) {
5403 if (zone_idx(zone
) > max_zone_idx
)
5405 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5409 for_each_online_cpu(cpu
)
5410 free_pcp
+= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
)->count
;
5420 " reserved_highatomic:%luKB"
5421 " active_anon:%lukB"
5422 " inactive_anon:%lukB"
5423 " active_file:%lukB"
5424 " inactive_file:%lukB"
5425 " unevictable:%lukB"
5426 " writepending:%lukB"
5436 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5437 K(zone
->watermark_boost
),
5438 K(min_wmark_pages(zone
)),
5439 K(low_wmark_pages(zone
)),
5440 K(high_wmark_pages(zone
)),
5441 K(zone
->nr_reserved_highatomic
),
5442 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5443 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5444 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5445 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5446 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5447 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5448 K(zone
->present_pages
),
5449 K(zone_managed_pages(zone
)),
5450 K(zone_page_state(zone
, NR_MLOCK
)),
5451 K(zone_page_state(zone
, NR_BOUNCE
)),
5453 K(this_cpu_read(zone
->per_cpu_pageset
->count
)),
5454 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5455 printk("lowmem_reserve[]:");
5456 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5457 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5458 printk(KERN_CONT
"\n");
5461 for_each_populated_zone(zone
) {
5463 unsigned long nr
[MAX_ORDER
+ 1], flags
, total
= 0;
5464 unsigned char types
[MAX_ORDER
+ 1];
5466 if (zone_idx(zone
) > max_zone_idx
)
5468 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5471 printk(KERN_CONT
"%s: ", zone
->name
);
5473 spin_lock_irqsave(&zone
->lock
, flags
);
5474 for (order
= 0; order
<= MAX_ORDER
; order
++) {
5475 struct free_area
*area
= &zone
->free_area
[order
];
5478 nr
[order
] = area
->nr_free
;
5479 total
+= nr
[order
] << order
;
5482 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5483 if (!free_area_empty(area
, type
))
5484 types
[order
] |= 1 << type
;
5487 spin_unlock_irqrestore(&zone
->lock
, flags
);
5488 for (order
= 0; order
<= MAX_ORDER
; order
++) {
5489 printk(KERN_CONT
"%lu*%lukB ",
5490 nr
[order
], K(1UL) << order
);
5492 show_migration_types(types
[order
]);
5494 printk(KERN_CONT
"= %lukB\n", K(total
));
5497 for_each_online_node(nid
) {
5498 if (show_mem_node_skip(filter
, nid
, nodemask
))
5500 hugetlb_show_meminfo_node(nid
);
5503 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5505 show_swap_cache_info();
5508 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5510 zoneref
->zone
= zone
;
5511 zoneref
->zone_idx
= zone_idx(zone
);
5515 * Builds allocation fallback zone lists.
5517 * Add all populated zones of a node to the zonelist.
5519 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5522 enum zone_type zone_type
= MAX_NR_ZONES
;
5527 zone
= pgdat
->node_zones
+ zone_type
;
5528 if (populated_zone(zone
)) {
5529 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5530 check_highest_zone(zone_type
);
5532 } while (zone_type
);
5539 static int __parse_numa_zonelist_order(char *s
)
5542 * We used to support different zonelists modes but they turned
5543 * out to be just not useful. Let's keep the warning in place
5544 * if somebody still use the cmd line parameter so that we do
5545 * not fail it silently
5547 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5548 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5554 char numa_zonelist_order
[] = "Node";
5557 * sysctl handler for numa_zonelist_order
5559 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5560 void *buffer
, size_t *length
, loff_t
*ppos
)
5563 return __parse_numa_zonelist_order(buffer
);
5564 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5568 static int node_load
[MAX_NUMNODES
];
5571 * find_next_best_node - find the next node that should appear in a given node's fallback list
5572 * @node: node whose fallback list we're appending
5573 * @used_node_mask: nodemask_t of already used nodes
5575 * We use a number of factors to determine which is the next node that should
5576 * appear on a given node's fallback list. The node should not have appeared
5577 * already in @node's fallback list, and it should be the next closest node
5578 * according to the distance array (which contains arbitrary distance values
5579 * from each node to each node in the system), and should also prefer nodes
5580 * with no CPUs, since presumably they'll have very little allocation pressure
5581 * on them otherwise.
5583 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5585 int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5588 int min_val
= INT_MAX
;
5589 int best_node
= NUMA_NO_NODE
;
5591 /* Use the local node if we haven't already */
5592 if (!node_isset(node
, *used_node_mask
)) {
5593 node_set(node
, *used_node_mask
);
5597 for_each_node_state(n
, N_MEMORY
) {
5599 /* Don't want a node to appear more than once */
5600 if (node_isset(n
, *used_node_mask
))
5603 /* Use the distance array to find the distance */
5604 val
= node_distance(node
, n
);
5606 /* Penalize nodes under us ("prefer the next node") */
5609 /* Give preference to headless and unused nodes */
5610 if (!cpumask_empty(cpumask_of_node(n
)))
5611 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5613 /* Slight preference for less loaded node */
5614 val
*= MAX_NUMNODES
;
5615 val
+= node_load
[n
];
5617 if (val
< min_val
) {
5624 node_set(best_node
, *used_node_mask
);
5631 * Build zonelists ordered by node and zones within node.
5632 * This results in maximum locality--normal zone overflows into local
5633 * DMA zone, if any--but risks exhausting DMA zone.
5635 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5638 struct zoneref
*zonerefs
;
5641 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5643 for (i
= 0; i
< nr_nodes
; i
++) {
5646 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5648 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5649 zonerefs
+= nr_zones
;
5651 zonerefs
->zone
= NULL
;
5652 zonerefs
->zone_idx
= 0;
5656 * Build gfp_thisnode zonelists
5658 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5660 struct zoneref
*zonerefs
;
5663 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5664 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5665 zonerefs
+= nr_zones
;
5666 zonerefs
->zone
= NULL
;
5667 zonerefs
->zone_idx
= 0;
5671 * Build zonelists ordered by zone and nodes within zones.
5672 * This results in conserving DMA zone[s] until all Normal memory is
5673 * exhausted, but results in overflowing to remote node while memory
5674 * may still exist in local DMA zone.
5677 static void build_zonelists(pg_data_t
*pgdat
)
5679 static int node_order
[MAX_NUMNODES
];
5680 int node
, nr_nodes
= 0;
5681 nodemask_t used_mask
= NODE_MASK_NONE
;
5682 int local_node
, prev_node
;
5684 /* NUMA-aware ordering of nodes */
5685 local_node
= pgdat
->node_id
;
5686 prev_node
= local_node
;
5688 memset(node_order
, 0, sizeof(node_order
));
5689 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5691 * We don't want to pressure a particular node.
5692 * So adding penalty to the first node in same
5693 * distance group to make it round-robin.
5695 if (node_distance(local_node
, node
) !=
5696 node_distance(local_node
, prev_node
))
5697 node_load
[node
] += 1;
5699 node_order
[nr_nodes
++] = node
;
5703 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5704 build_thisnode_zonelists(pgdat
);
5705 pr_info("Fallback order for Node %d: ", local_node
);
5706 for (node
= 0; node
< nr_nodes
; node
++)
5707 pr_cont("%d ", node_order
[node
]);
5711 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5713 * Return node id of node used for "local" allocations.
5714 * I.e., first node id of first zone in arg node's generic zonelist.
5715 * Used for initializing percpu 'numa_mem', which is used primarily
5716 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5718 int local_memory_node(int node
)
5722 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5723 gfp_zone(GFP_KERNEL
),
5725 return zone_to_nid(z
->zone
);
5729 static void setup_min_unmapped_ratio(void);
5730 static void setup_min_slab_ratio(void);
5731 #else /* CONFIG_NUMA */
5733 static void build_zonelists(pg_data_t
*pgdat
)
5735 int node
, local_node
;
5736 struct zoneref
*zonerefs
;
5739 local_node
= pgdat
->node_id
;
5741 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5742 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5743 zonerefs
+= nr_zones
;
5746 * Now we build the zonelist so that it contains the zones
5747 * of all the other nodes.
5748 * We don't want to pressure a particular node, so when
5749 * building the zones for node N, we make sure that the
5750 * zones coming right after the local ones are those from
5751 * node N+1 (modulo N)
5753 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5754 if (!node_online(node
))
5756 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5757 zonerefs
+= nr_zones
;
5759 for (node
= 0; node
< local_node
; node
++) {
5760 if (!node_online(node
))
5762 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5763 zonerefs
+= nr_zones
;
5766 zonerefs
->zone
= NULL
;
5767 zonerefs
->zone_idx
= 0;
5770 #endif /* CONFIG_NUMA */
5773 * Boot pageset table. One per cpu which is going to be used for all
5774 * zones and all nodes. The parameters will be set in such a way
5775 * that an item put on a list will immediately be handed over to
5776 * the buddy list. This is safe since pageset manipulation is done
5777 * with interrupts disabled.
5779 * The boot_pagesets must be kept even after bootup is complete for
5780 * unused processors and/or zones. They do play a role for bootstrapping
5781 * hotplugged processors.
5783 * zoneinfo_show() and maybe other functions do
5784 * not check if the processor is online before following the pageset pointer.
5785 * Other parts of the kernel may not check if the zone is available.
5787 static void per_cpu_pages_init(struct per_cpu_pages
*pcp
, struct per_cpu_zonestat
*pzstats
);
5788 /* These effectively disable the pcplists in the boot pageset completely */
5789 #define BOOT_PAGESET_HIGH 0
5790 #define BOOT_PAGESET_BATCH 1
5791 static DEFINE_PER_CPU(struct per_cpu_pages
, boot_pageset
);
5792 static DEFINE_PER_CPU(struct per_cpu_zonestat
, boot_zonestats
);
5794 static void __build_all_zonelists(void *data
)
5797 int __maybe_unused cpu
;
5798 pg_data_t
*self
= data
;
5799 unsigned long flags
;
5802 * Explicitly disable this CPU's interrupts before taking seqlock
5803 * to prevent any IRQ handler from calling into the page allocator
5804 * (e.g. GFP_ATOMIC) that could hit zonelist_iter_begin and livelock.
5806 local_irq_save(flags
);
5808 * Explicitly disable this CPU's synchronous printk() before taking
5809 * seqlock to prevent any printk() from trying to hold port->lock, for
5810 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5811 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5813 printk_deferred_enter();
5814 write_seqlock(&zonelist_update_seq
);
5817 memset(node_load
, 0, sizeof(node_load
));
5821 * This node is hotadded and no memory is yet present. So just
5822 * building zonelists is fine - no need to touch other nodes.
5824 if (self
&& !node_online(self
->node_id
)) {
5825 build_zonelists(self
);
5828 * All possible nodes have pgdat preallocated
5831 for_each_node(nid
) {
5832 pg_data_t
*pgdat
= NODE_DATA(nid
);
5834 build_zonelists(pgdat
);
5837 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5839 * We now know the "local memory node" for each node--
5840 * i.e., the node of the first zone in the generic zonelist.
5841 * Set up numa_mem percpu variable for on-line cpus. During
5842 * boot, only the boot cpu should be on-line; we'll init the
5843 * secondary cpus' numa_mem as they come on-line. During
5844 * node/memory hotplug, we'll fixup all on-line cpus.
5846 for_each_online_cpu(cpu
)
5847 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5851 write_sequnlock(&zonelist_update_seq
);
5852 printk_deferred_exit();
5853 local_irq_restore(flags
);
5856 static noinline
void __init
5857 build_all_zonelists_init(void)
5861 __build_all_zonelists(NULL
);
5864 * Initialize the boot_pagesets that are going to be used
5865 * for bootstrapping processors. The real pagesets for
5866 * each zone will be allocated later when the per cpu
5867 * allocator is available.
5869 * boot_pagesets are used also for bootstrapping offline
5870 * cpus if the system is already booted because the pagesets
5871 * are needed to initialize allocators on a specific cpu too.
5872 * F.e. the percpu allocator needs the page allocator which
5873 * needs the percpu allocator in order to allocate its pagesets
5874 * (a chicken-egg dilemma).
5876 for_each_possible_cpu(cpu
)
5877 per_cpu_pages_init(&per_cpu(boot_pageset
, cpu
), &per_cpu(boot_zonestats
, cpu
));
5879 mminit_verify_zonelist();
5880 cpuset_init_current_mems_allowed();
5884 * unless system_state == SYSTEM_BOOTING.
5886 * __ref due to call of __init annotated helper build_all_zonelists_init
5887 * [protected by SYSTEM_BOOTING].
5889 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5891 unsigned long vm_total_pages
;
5893 if (system_state
== SYSTEM_BOOTING
) {
5894 build_all_zonelists_init();
5896 __build_all_zonelists(pgdat
);
5897 /* cpuset refresh routine should be here */
5899 /* Get the number of free pages beyond high watermark in all zones. */
5900 vm_total_pages
= nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5902 * Disable grouping by mobility if the number of pages in the
5903 * system is too low to allow the mechanism to work. It would be
5904 * more accurate, but expensive to check per-zone. This check is
5905 * made on memory-hotadd so a system can start with mobility
5906 * disabled and enable it later
5908 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5909 page_group_by_mobility_disabled
= 1;
5911 page_group_by_mobility_disabled
= 0;
5913 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5915 page_group_by_mobility_disabled
? "off" : "on",
5918 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5922 static int zone_batchsize(struct zone
*zone
)
5928 * The number of pages to batch allocate is either ~0.1%
5929 * of the zone or 1MB, whichever is smaller. The batch
5930 * size is striking a balance between allocation latency
5931 * and zone lock contention.
5933 batch
= min(zone_managed_pages(zone
) >> 10, SZ_1M
/ PAGE_SIZE
);
5934 batch
/= 4; /* We effectively *= 4 below */
5939 * Clamp the batch to a 2^n - 1 value. Having a power
5940 * of 2 value was found to be more likely to have
5941 * suboptimal cache aliasing properties in some cases.
5943 * For example if 2 tasks are alternately allocating
5944 * batches of pages, one task can end up with a lot
5945 * of pages of one half of the possible page colors
5946 * and the other with pages of the other colors.
5948 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5953 /* The deferral and batching of frees should be suppressed under NOMMU
5956 * The problem is that NOMMU needs to be able to allocate large chunks
5957 * of contiguous memory as there's no hardware page translation to
5958 * assemble apparent contiguous memory from discontiguous pages.
5960 * Queueing large contiguous runs of pages for batching, however,
5961 * causes the pages to actually be freed in smaller chunks. As there
5962 * can be a significant delay between the individual batches being
5963 * recycled, this leads to the once large chunks of space being
5964 * fragmented and becoming unavailable for high-order allocations.
5970 static int zone_highsize(struct zone
*zone
, int batch
, int cpu_online
)
5975 unsigned long total_pages
;
5977 if (!percpu_pagelist_high_fraction
) {
5979 * By default, the high value of the pcp is based on the zone
5980 * low watermark so that if they are full then background
5981 * reclaim will not be started prematurely.
5983 total_pages
= low_wmark_pages(zone
);
5986 * If percpu_pagelist_high_fraction is configured, the high
5987 * value is based on a fraction of the managed pages in the
5990 total_pages
= zone_managed_pages(zone
) / percpu_pagelist_high_fraction
;
5994 * Split the high value across all online CPUs local to the zone. Note
5995 * that early in boot that CPUs may not be online yet and that during
5996 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5997 * onlined. For memory nodes that have no CPUs, split pcp->high across
5998 * all online CPUs to mitigate the risk that reclaim is triggered
5999 * prematurely due to pages stored on pcp lists.
6001 nr_split_cpus
= cpumask_weight(cpumask_of_node(zone_to_nid(zone
))) + cpu_online
;
6003 nr_split_cpus
= num_online_cpus();
6004 high
= total_pages
/ nr_split_cpus
;
6007 * Ensure high is at least batch*4. The multiple is based on the
6008 * historical relationship between high and batch.
6010 high
= max(high
, batch
<< 2);
6019 * pcp->high and pcp->batch values are related and generally batch is lower
6020 * than high. They are also related to pcp->count such that count is lower
6021 * than high, and as soon as it reaches high, the pcplist is flushed.
6023 * However, guaranteeing these relations at all times would require e.g. write
6024 * barriers here but also careful usage of read barriers at the read side, and
6025 * thus be prone to error and bad for performance. Thus the update only prevents
6026 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6027 * can cope with those fields changing asynchronously, and fully trust only the
6028 * pcp->count field on the local CPU with interrupts disabled.
6030 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6031 * outside of boot time (or some other assurance that no concurrent updaters
6034 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6035 unsigned long batch
)
6037 WRITE_ONCE(pcp
->batch
, batch
);
6038 WRITE_ONCE(pcp
->high
, high
);
6041 static void per_cpu_pages_init(struct per_cpu_pages
*pcp
, struct per_cpu_zonestat
*pzstats
)
6045 memset(pcp
, 0, sizeof(*pcp
));
6046 memset(pzstats
, 0, sizeof(*pzstats
));
6048 spin_lock_init(&pcp
->lock
);
6049 for (pindex
= 0; pindex
< NR_PCP_LISTS
; pindex
++)
6050 INIT_LIST_HEAD(&pcp
->lists
[pindex
]);
6053 * Set batch and high values safe for a boot pageset. A true percpu
6054 * pageset's initialization will update them subsequently. Here we don't
6055 * need to be as careful as pageset_update() as nobody can access the
6058 pcp
->high
= BOOT_PAGESET_HIGH
;
6059 pcp
->batch
= BOOT_PAGESET_BATCH
;
6060 pcp
->free_factor
= 0;
6063 static void __zone_set_pageset_high_and_batch(struct zone
*zone
, unsigned long high
,
6064 unsigned long batch
)
6066 struct per_cpu_pages
*pcp
;
6069 for_each_possible_cpu(cpu
) {
6070 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
6071 pageset_update(pcp
, high
, batch
);
6076 * Calculate and set new high and batch values for all per-cpu pagesets of a
6077 * zone based on the zone's size.
6079 static void zone_set_pageset_high_and_batch(struct zone
*zone
, int cpu_online
)
6081 int new_high
, new_batch
;
6083 new_batch
= max(1, zone_batchsize(zone
));
6084 new_high
= zone_highsize(zone
, new_batch
, cpu_online
);
6086 if (zone
->pageset_high
== new_high
&&
6087 zone
->pageset_batch
== new_batch
)
6090 zone
->pageset_high
= new_high
;
6091 zone
->pageset_batch
= new_batch
;
6093 __zone_set_pageset_high_and_batch(zone
, new_high
, new_batch
);
6096 void __meminit
setup_zone_pageset(struct zone
*zone
)
6100 /* Size may be 0 on !SMP && !NUMA */
6101 if (sizeof(struct per_cpu_zonestat
) > 0)
6102 zone
->per_cpu_zonestats
= alloc_percpu(struct per_cpu_zonestat
);
6104 zone
->per_cpu_pageset
= alloc_percpu(struct per_cpu_pages
);
6105 for_each_possible_cpu(cpu
) {
6106 struct per_cpu_pages
*pcp
;
6107 struct per_cpu_zonestat
*pzstats
;
6109 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
6110 pzstats
= per_cpu_ptr(zone
->per_cpu_zonestats
, cpu
);
6111 per_cpu_pages_init(pcp
, pzstats
);
6114 zone_set_pageset_high_and_batch(zone
, 0);
6118 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6119 * page high values need to be recalculated.
6121 static void zone_pcp_update(struct zone
*zone
, int cpu_online
)
6123 mutex_lock(&pcp_batch_high_lock
);
6124 zone_set_pageset_high_and_batch(zone
, cpu_online
);
6125 mutex_unlock(&pcp_batch_high_lock
);
6129 * Allocate per cpu pagesets and initialize them.
6130 * Before this call only boot pagesets were available.
6132 void __init
setup_per_cpu_pageset(void)
6134 struct pglist_data
*pgdat
;
6136 int __maybe_unused cpu
;
6138 for_each_populated_zone(zone
)
6139 setup_zone_pageset(zone
);
6143 * Unpopulated zones continue using the boot pagesets.
6144 * The numa stats for these pagesets need to be reset.
6145 * Otherwise, they will end up skewing the stats of
6146 * the nodes these zones are associated with.
6148 for_each_possible_cpu(cpu
) {
6149 struct per_cpu_zonestat
*pzstats
= &per_cpu(boot_zonestats
, cpu
);
6150 memset(pzstats
->vm_numa_event
, 0,
6151 sizeof(pzstats
->vm_numa_event
));
6155 for_each_online_pgdat(pgdat
)
6156 pgdat
->per_cpu_nodestats
=
6157 alloc_percpu(struct per_cpu_nodestat
);
6160 __meminit
void zone_pcp_init(struct zone
*zone
)
6163 * per cpu subsystem is not up at this point. The following code
6164 * relies on the ability of the linker to provide the
6165 * offset of a (static) per cpu variable into the per cpu area.
6167 zone
->per_cpu_pageset
= &boot_pageset
;
6168 zone
->per_cpu_zonestats
= &boot_zonestats
;
6169 zone
->pageset_high
= BOOT_PAGESET_HIGH
;
6170 zone
->pageset_batch
= BOOT_PAGESET_BATCH
;
6172 if (populated_zone(zone
))
6173 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone
->name
,
6174 zone
->present_pages
, zone_batchsize(zone
));
6177 void adjust_managed_page_count(struct page
*page
, long count
)
6179 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
6180 totalram_pages_add(count
);
6181 #ifdef CONFIG_HIGHMEM
6182 if (PageHighMem(page
))
6183 totalhigh_pages_add(count
);
6186 EXPORT_SYMBOL(adjust_managed_page_count
);
6188 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
6191 unsigned long pages
= 0;
6193 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6194 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6195 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6196 struct page
*page
= virt_to_page(pos
);
6197 void *direct_map_addr
;
6200 * 'direct_map_addr' might be different from 'pos'
6201 * because some architectures' virt_to_page()
6202 * work with aliases. Getting the direct map
6203 * address ensures that we get a _writeable_
6204 * alias for the memset().
6206 direct_map_addr
= page_address(page
);
6208 * Perform a kasan-unchecked memset() since this memory
6209 * has not been initialized.
6211 direct_map_addr
= kasan_reset_tag(direct_map_addr
);
6212 if ((unsigned int)poison
<= 0xFF)
6213 memset(direct_map_addr
, poison
, PAGE_SIZE
);
6215 free_reserved_page(page
);
6219 pr_info("Freeing %s memory: %ldK\n", s
, K(pages
));
6224 static int page_alloc_cpu_dead(unsigned int cpu
)
6228 lru_add_drain_cpu(cpu
);
6229 mlock_drain_remote(cpu
);
6233 * Spill the event counters of the dead processor
6234 * into the current processors event counters.
6235 * This artificially elevates the count of the current
6238 vm_events_fold_cpu(cpu
);
6241 * Zero the differential counters of the dead processor
6242 * so that the vm statistics are consistent.
6244 * This is only okay since the processor is dead and cannot
6245 * race with what we are doing.
6247 cpu_vm_stats_fold(cpu
);
6249 for_each_populated_zone(zone
)
6250 zone_pcp_update(zone
, 0);
6255 static int page_alloc_cpu_online(unsigned int cpu
)
6259 for_each_populated_zone(zone
)
6260 zone_pcp_update(zone
, 1);
6264 void __init
page_alloc_init_cpuhp(void)
6268 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC
,
6269 "mm/page_alloc:pcp",
6270 page_alloc_cpu_online
,
6271 page_alloc_cpu_dead
);
6276 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6277 * or min_free_kbytes changes.
6279 static void calculate_totalreserve_pages(void)
6281 struct pglist_data
*pgdat
;
6282 unsigned long reserve_pages
= 0;
6283 enum zone_type i
, j
;
6285 for_each_online_pgdat(pgdat
) {
6287 pgdat
->totalreserve_pages
= 0;
6289 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6290 struct zone
*zone
= pgdat
->node_zones
+ i
;
6292 unsigned long managed_pages
= zone_managed_pages(zone
);
6294 /* Find valid and maximum lowmem_reserve in the zone */
6295 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6296 if (zone
->lowmem_reserve
[j
] > max
)
6297 max
= zone
->lowmem_reserve
[j
];
6300 /* we treat the high watermark as reserved pages. */
6301 max
+= high_wmark_pages(zone
);
6303 if (max
> managed_pages
)
6304 max
= managed_pages
;
6306 pgdat
->totalreserve_pages
+= max
;
6308 reserve_pages
+= max
;
6311 totalreserve_pages
= reserve_pages
;
6315 * setup_per_zone_lowmem_reserve - called whenever
6316 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6317 * has a correct pages reserved value, so an adequate number of
6318 * pages are left in the zone after a successful __alloc_pages().
6320 static void setup_per_zone_lowmem_reserve(void)
6322 struct pglist_data
*pgdat
;
6323 enum zone_type i
, j
;
6325 for_each_online_pgdat(pgdat
) {
6326 for (i
= 0; i
< MAX_NR_ZONES
- 1; i
++) {
6327 struct zone
*zone
= &pgdat
->node_zones
[i
];
6328 int ratio
= sysctl_lowmem_reserve_ratio
[i
];
6329 bool clear
= !ratio
|| !zone_managed_pages(zone
);
6330 unsigned long managed_pages
= 0;
6332 for (j
= i
+ 1; j
< MAX_NR_ZONES
; j
++) {
6333 struct zone
*upper_zone
= &pgdat
->node_zones
[j
];
6335 managed_pages
+= zone_managed_pages(upper_zone
);
6338 zone
->lowmem_reserve
[j
] = 0;
6340 zone
->lowmem_reserve
[j
] = managed_pages
/ ratio
;
6345 /* update totalreserve_pages */
6346 calculate_totalreserve_pages();
6349 static void __setup_per_zone_wmarks(void)
6351 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6352 unsigned long lowmem_pages
= 0;
6354 unsigned long flags
;
6356 /* Calculate total number of !ZONE_HIGHMEM pages */
6357 for_each_zone(zone
) {
6358 if (!is_highmem(zone
))
6359 lowmem_pages
+= zone_managed_pages(zone
);
6362 for_each_zone(zone
) {
6365 spin_lock_irqsave(&zone
->lock
, flags
);
6366 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
6367 do_div(tmp
, lowmem_pages
);
6368 if (is_highmem(zone
)) {
6370 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6371 * need highmem pages, so cap pages_min to a small
6374 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6375 * deltas control async page reclaim, and so should
6376 * not be capped for highmem.
6378 unsigned long min_pages
;
6380 min_pages
= zone_managed_pages(zone
) / 1024;
6381 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6382 zone
->_watermark
[WMARK_MIN
] = min_pages
;
6385 * If it's a lowmem zone, reserve a number of pages
6386 * proportionate to the zone's size.
6388 zone
->_watermark
[WMARK_MIN
] = tmp
;
6392 * Set the kswapd watermarks distance according to the
6393 * scale factor in proportion to available memory, but
6394 * ensure a minimum size on small systems.
6396 tmp
= max_t(u64
, tmp
>> 2,
6397 mult_frac(zone_managed_pages(zone
),
6398 watermark_scale_factor
, 10000));
6400 zone
->watermark_boost
= 0;
6401 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6402 zone
->_watermark
[WMARK_HIGH
] = low_wmark_pages(zone
) + tmp
;
6403 zone
->_watermark
[WMARK_PROMO
] = high_wmark_pages(zone
) + tmp
;
6405 spin_unlock_irqrestore(&zone
->lock
, flags
);
6408 /* update totalreserve_pages */
6409 calculate_totalreserve_pages();
6413 * setup_per_zone_wmarks - called when min_free_kbytes changes
6414 * or when memory is hot-{added|removed}
6416 * Ensures that the watermark[min,low,high] values for each zone are set
6417 * correctly with respect to min_free_kbytes.
6419 void setup_per_zone_wmarks(void)
6422 static DEFINE_SPINLOCK(lock
);
6425 __setup_per_zone_wmarks();
6429 * The watermark size have changed so update the pcpu batch
6430 * and high limits or the limits may be inappropriate.
6433 zone_pcp_update(zone
, 0);
6437 * Initialise min_free_kbytes.
6439 * For small machines we want it small (128k min). For large machines
6440 * we want it large (256MB max). But it is not linear, because network
6441 * bandwidth does not increase linearly with machine size. We use
6443 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6444 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6460 void calculate_min_free_kbytes(void)
6462 unsigned long lowmem_kbytes
;
6463 int new_min_free_kbytes
;
6465 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6466 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6468 if (new_min_free_kbytes
> user_min_free_kbytes
)
6469 min_free_kbytes
= clamp(new_min_free_kbytes
, 128, 262144);
6471 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6472 new_min_free_kbytes
, user_min_free_kbytes
);
6476 int __meminit
init_per_zone_wmark_min(void)
6478 calculate_min_free_kbytes();
6479 setup_per_zone_wmarks();
6480 refresh_zone_stat_thresholds();
6481 setup_per_zone_lowmem_reserve();
6484 setup_min_unmapped_ratio();
6485 setup_min_slab_ratio();
6488 khugepaged_min_free_kbytes_update();
6492 postcore_initcall(init_per_zone_wmark_min
)
6495 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6496 * that we can call two helper functions whenever min_free_kbytes
6499 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6500 void *buffer
, size_t *length
, loff_t
*ppos
)
6504 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6509 user_min_free_kbytes
= min_free_kbytes
;
6510 setup_per_zone_wmarks();
6515 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6516 void *buffer
, size_t *length
, loff_t
*ppos
)
6520 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6525 setup_per_zone_wmarks();
6531 static void setup_min_unmapped_ratio(void)
6536 for_each_online_pgdat(pgdat
)
6537 pgdat
->min_unmapped_pages
= 0;
6540 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
6541 sysctl_min_unmapped_ratio
) / 100;
6545 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6546 void *buffer
, size_t *length
, loff_t
*ppos
)
6550 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6554 setup_min_unmapped_ratio();
6559 static void setup_min_slab_ratio(void)
6564 for_each_online_pgdat(pgdat
)
6565 pgdat
->min_slab_pages
= 0;
6568 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
6569 sysctl_min_slab_ratio
) / 100;
6572 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6573 void *buffer
, size_t *length
, loff_t
*ppos
)
6577 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6581 setup_min_slab_ratio();
6588 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6589 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6590 * whenever sysctl_lowmem_reserve_ratio changes.
6592 * The reserve ratio obviously has absolutely no relation with the
6593 * minimum watermarks. The lowmem reserve ratio can only make sense
6594 * if in function of the boot time zone sizes.
6596 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6597 void *buffer
, size_t *length
, loff_t
*ppos
)
6601 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6603 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6604 if (sysctl_lowmem_reserve_ratio
[i
] < 1)
6605 sysctl_lowmem_reserve_ratio
[i
] = 0;
6608 setup_per_zone_lowmem_reserve();
6613 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6614 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6615 * pagelist can have before it gets flushed back to buddy allocator.
6617 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table
*table
,
6618 int write
, void *buffer
, size_t *length
, loff_t
*ppos
)
6621 int old_percpu_pagelist_high_fraction
;
6624 mutex_lock(&pcp_batch_high_lock
);
6625 old_percpu_pagelist_high_fraction
= percpu_pagelist_high_fraction
;
6627 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6628 if (!write
|| ret
< 0)
6631 /* Sanity checking to avoid pcp imbalance */
6632 if (percpu_pagelist_high_fraction
&&
6633 percpu_pagelist_high_fraction
< MIN_PERCPU_PAGELIST_HIGH_FRACTION
) {
6634 percpu_pagelist_high_fraction
= old_percpu_pagelist_high_fraction
;
6640 if (percpu_pagelist_high_fraction
== old_percpu_pagelist_high_fraction
)
6643 for_each_populated_zone(zone
)
6644 zone_set_pageset_high_and_batch(zone
, 0);
6646 mutex_unlock(&pcp_batch_high_lock
);
6650 #ifdef CONFIG_CONTIG_ALLOC
6651 #if defined(CONFIG_DYNAMIC_DEBUG) || \
6652 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
6653 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6654 static void alloc_contig_dump_pages(struct list_head
*page_list
)
6656 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor
, "migrate failure");
6658 if (DYNAMIC_DEBUG_BRANCH(descriptor
)) {
6662 list_for_each_entry(page
, page_list
, lru
)
6663 dump_page(page
, "migration failure");
6667 static inline void alloc_contig_dump_pages(struct list_head
*page_list
)
6672 /* [start, end) must belong to a single zone. */
6673 int __alloc_contig_migrate_range(struct compact_control
*cc
,
6674 unsigned long start
, unsigned long end
)
6676 /* This function is based on compact_zone() from compaction.c. */
6677 unsigned int nr_reclaimed
;
6678 unsigned long pfn
= start
;
6679 unsigned int tries
= 0;
6681 struct migration_target_control mtc
= {
6682 .nid
= zone_to_nid(cc
->zone
),
6683 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
6686 lru_cache_disable();
6688 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6689 if (fatal_signal_pending(current
)) {
6694 if (list_empty(&cc
->migratepages
)) {
6695 cc
->nr_migratepages
= 0;
6696 ret
= isolate_migratepages_range(cc
, pfn
, end
);
6697 if (ret
&& ret
!= -EAGAIN
)
6699 pfn
= cc
->migrate_pfn
;
6701 } else if (++tries
== 5) {
6706 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6708 cc
->nr_migratepages
-= nr_reclaimed
;
6710 ret
= migrate_pages(&cc
->migratepages
, alloc_migration_target
,
6711 NULL
, (unsigned long)&mtc
, cc
->mode
, MR_CONTIG_RANGE
, NULL
);
6714 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6715 * to retry again over this error, so do the same here.
6723 if (!(cc
->gfp_mask
& __GFP_NOWARN
) && ret
== -EBUSY
)
6724 alloc_contig_dump_pages(&cc
->migratepages
);
6725 putback_movable_pages(&cc
->migratepages
);
6732 * alloc_contig_range() -- tries to allocate given range of pages
6733 * @start: start PFN to allocate
6734 * @end: one-past-the-last PFN to allocate
6735 * @migratetype: migratetype of the underlying pageblocks (either
6736 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6737 * in range must have the same migratetype and it must
6738 * be either of the two.
6739 * @gfp_mask: GFP mask to use during compaction
6741 * The PFN range does not have to be pageblock aligned. The PFN range must
6742 * belong to a single zone.
6744 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6745 * pageblocks in the range. Once isolated, the pageblocks should not
6746 * be modified by others.
6748 * Return: zero on success or negative error code. On success all
6749 * pages which PFN is in [start, end) are allocated for the caller and
6750 * need to be freed with free_contig_range().
6752 int alloc_contig_range(unsigned long start
, unsigned long end
,
6753 unsigned migratetype
, gfp_t gfp_mask
)
6755 unsigned long outer_start
, outer_end
;
6759 struct compact_control cc
= {
6760 .nr_migratepages
= 0,
6762 .zone
= page_zone(pfn_to_page(start
)),
6763 .mode
= MIGRATE_SYNC
,
6764 .ignore_skip_hint
= true,
6765 .no_set_skip_hint
= true,
6766 .gfp_mask
= current_gfp_context(gfp_mask
),
6767 .alloc_contig
= true,
6769 INIT_LIST_HEAD(&cc
.migratepages
);
6772 * What we do here is we mark all pageblocks in range as
6773 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6774 * have different sizes, and due to the way page allocator
6775 * work, start_isolate_page_range() has special handlings for this.
6777 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6778 * migrate the pages from an unaligned range (ie. pages that
6779 * we are interested in). This will put all the pages in
6780 * range back to page allocator as MIGRATE_ISOLATE.
6782 * When this is done, we take the pages in range from page
6783 * allocator removing them from the buddy system. This way
6784 * page allocator will never consider using them.
6786 * This lets us mark the pageblocks back as
6787 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6788 * aligned range but not in the unaligned, original range are
6789 * put back to page allocator so that buddy can use them.
6792 ret
= start_isolate_page_range(start
, end
, migratetype
, 0, gfp_mask
);
6796 drain_all_pages(cc
.zone
);
6799 * In case of -EBUSY, we'd like to know which page causes problem.
6800 * So, just fall through. test_pages_isolated() has a tracepoint
6801 * which will report the busy page.
6803 * It is possible that busy pages could become available before
6804 * the call to test_pages_isolated, and the range will actually be
6805 * allocated. So, if we fall through be sure to clear ret so that
6806 * -EBUSY is not accidentally used or returned to caller.
6808 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6809 if (ret
&& ret
!= -EBUSY
)
6814 * Pages from [start, end) are within a pageblock_nr_pages
6815 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6816 * more, all pages in [start, end) are free in page allocator.
6817 * What we are going to do is to allocate all pages from
6818 * [start, end) (that is remove them from page allocator).
6820 * The only problem is that pages at the beginning and at the
6821 * end of interesting range may be not aligned with pages that
6822 * page allocator holds, ie. they can be part of higher order
6823 * pages. Because of this, we reserve the bigger range and
6824 * once this is done free the pages we are not interested in.
6826 * We don't have to hold zone->lock here because the pages are
6827 * isolated thus they won't get removed from buddy.
6831 outer_start
= start
;
6832 while (!PageBuddy(pfn_to_page(outer_start
))) {
6833 if (++order
> MAX_ORDER
) {
6834 outer_start
= start
;
6837 outer_start
&= ~0UL << order
;
6840 if (outer_start
!= start
) {
6841 order
= buddy_order(pfn_to_page(outer_start
));
6844 * outer_start page could be small order buddy page and
6845 * it doesn't include start page. Adjust outer_start
6846 * in this case to report failed page properly
6847 * on tracepoint in test_pages_isolated()
6849 if (outer_start
+ (1UL << order
) <= start
)
6850 outer_start
= start
;
6853 /* Make sure the range is really isolated. */
6854 if (test_pages_isolated(outer_start
, end
, 0)) {
6859 /* Grab isolated pages from freelists. */
6860 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6866 /* Free head and tail (if any) */
6867 if (start
!= outer_start
)
6868 free_contig_range(outer_start
, start
- outer_start
);
6869 if (end
!= outer_end
)
6870 free_contig_range(end
, outer_end
- end
);
6873 undo_isolate_page_range(start
, end
, migratetype
);
6876 EXPORT_SYMBOL(alloc_contig_range
);
6878 static int __alloc_contig_pages(unsigned long start_pfn
,
6879 unsigned long nr_pages
, gfp_t gfp_mask
)
6881 unsigned long end_pfn
= start_pfn
+ nr_pages
;
6883 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
6887 static bool pfn_range_valid_contig(struct zone
*z
, unsigned long start_pfn
,
6888 unsigned long nr_pages
)
6890 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
6893 for (i
= start_pfn
; i
< end_pfn
; i
++) {
6894 page
= pfn_to_online_page(i
);
6898 if (page_zone(page
) != z
)
6901 if (PageReserved(page
))
6910 static bool zone_spans_last_pfn(const struct zone
*zone
,
6911 unsigned long start_pfn
, unsigned long nr_pages
)
6913 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
6915 return zone_spans_pfn(zone
, last_pfn
);
6919 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6920 * @nr_pages: Number of contiguous pages to allocate
6921 * @gfp_mask: GFP mask to limit search and used during compaction
6923 * @nodemask: Mask for other possible nodes
6925 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6926 * on an applicable zonelist to find a contiguous pfn range which can then be
6927 * tried for allocation with alloc_contig_range(). This routine is intended
6928 * for allocation requests which can not be fulfilled with the buddy allocator.
6930 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6931 * power of two, then allocated range is also guaranteed to be aligned to same
6932 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6934 * Allocated pages can be freed with free_contig_range() or by manually calling
6935 * __free_page() on each allocated page.
6937 * Return: pointer to contiguous pages on success, or NULL if not successful.
6939 struct page
*alloc_contig_pages(unsigned long nr_pages
, gfp_t gfp_mask
,
6940 int nid
, nodemask_t
*nodemask
)
6942 unsigned long ret
, pfn
, flags
;
6943 struct zonelist
*zonelist
;
6947 zonelist
= node_zonelist(nid
, gfp_mask
);
6948 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
6949 gfp_zone(gfp_mask
), nodemask
) {
6950 spin_lock_irqsave(&zone
->lock
, flags
);
6952 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
6953 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
6954 if (pfn_range_valid_contig(zone
, pfn
, nr_pages
)) {
6956 * We release the zone lock here because
6957 * alloc_contig_range() will also lock the zone
6958 * at some point. If there's an allocation
6959 * spinning on this lock, it may win the race
6960 * and cause alloc_contig_range() to fail...
6962 spin_unlock_irqrestore(&zone
->lock
, flags
);
6963 ret
= __alloc_contig_pages(pfn
, nr_pages
,
6966 return pfn_to_page(pfn
);
6967 spin_lock_irqsave(&zone
->lock
, flags
);
6971 spin_unlock_irqrestore(&zone
->lock
, flags
);
6975 #endif /* CONFIG_CONTIG_ALLOC */
6977 void free_contig_range(unsigned long pfn
, unsigned long nr_pages
)
6979 unsigned long count
= 0;
6981 for (; nr_pages
--; pfn
++) {
6982 struct page
*page
= pfn_to_page(pfn
);
6984 count
+= page_count(page
) != 1;
6987 WARN(count
!= 0, "%lu pages are still in use!\n", count
);
6989 EXPORT_SYMBOL(free_contig_range
);
6992 * Effectively disable pcplists for the zone by setting the high limit to 0
6993 * and draining all cpus. A concurrent page freeing on another CPU that's about
6994 * to put the page on pcplist will either finish before the drain and the page
6995 * will be drained, or observe the new high limit and skip the pcplist.
6997 * Must be paired with a call to zone_pcp_enable().
6999 void zone_pcp_disable(struct zone
*zone
)
7001 mutex_lock(&pcp_batch_high_lock
);
7002 __zone_set_pageset_high_and_batch(zone
, 0, 1);
7003 __drain_all_pages(zone
, true);
7006 void zone_pcp_enable(struct zone
*zone
)
7008 __zone_set_pageset_high_and_batch(zone
, zone
->pageset_high
, zone
->pageset_batch
);
7009 mutex_unlock(&pcp_batch_high_lock
);
7012 void zone_pcp_reset(struct zone
*zone
)
7015 struct per_cpu_zonestat
*pzstats
;
7017 if (zone
->per_cpu_pageset
!= &boot_pageset
) {
7018 for_each_online_cpu(cpu
) {
7019 pzstats
= per_cpu_ptr(zone
->per_cpu_zonestats
, cpu
);
7020 drain_zonestat(zone
, pzstats
);
7022 free_percpu(zone
->per_cpu_pageset
);
7023 zone
->per_cpu_pageset
= &boot_pageset
;
7024 if (zone
->per_cpu_zonestats
!= &boot_zonestats
) {
7025 free_percpu(zone
->per_cpu_zonestats
);
7026 zone
->per_cpu_zonestats
= &boot_zonestats
;
7031 #ifdef CONFIG_MEMORY_HOTREMOVE
7033 * All pages in the range must be in a single zone, must not contain holes,
7034 * must span full sections, and must be isolated before calling this function.
7036 void __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7038 unsigned long pfn
= start_pfn
;
7042 unsigned long flags
;
7044 offline_mem_sections(pfn
, end_pfn
);
7045 zone
= page_zone(pfn_to_page(pfn
));
7046 spin_lock_irqsave(&zone
->lock
, flags
);
7047 while (pfn
< end_pfn
) {
7048 page
= pfn_to_page(pfn
);
7050 * The HWPoisoned page may be not in buddy system, and
7051 * page_count() is not 0.
7053 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7058 * At this point all remaining PageOffline() pages have a
7059 * reference count of 0 and can simply be skipped.
7061 if (PageOffline(page
)) {
7062 BUG_ON(page_count(page
));
7063 BUG_ON(PageBuddy(page
));
7068 BUG_ON(page_count(page
));
7069 BUG_ON(!PageBuddy(page
));
7070 order
= buddy_order(page
);
7071 del_page_from_free_list(page
, zone
, order
);
7072 pfn
+= (1 << order
);
7074 spin_unlock_irqrestore(&zone
->lock
, flags
);
7079 * This function returns a stable result only if called under zone lock.
7081 bool is_free_buddy_page(struct page
*page
)
7083 unsigned long pfn
= page_to_pfn(page
);
7086 for (order
= 0; order
<= MAX_ORDER
; order
++) {
7087 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7089 if (PageBuddy(page_head
) &&
7090 buddy_order_unsafe(page_head
) >= order
)
7094 return order
<= MAX_ORDER
;
7096 EXPORT_SYMBOL(is_free_buddy_page
);
7098 #ifdef CONFIG_MEMORY_FAILURE
7100 * Break down a higher-order page in sub-pages, and keep our target out of
7103 static void break_down_buddy_pages(struct zone
*zone
, struct page
*page
,
7104 struct page
*target
, int low
, int high
,
7107 unsigned long size
= 1 << high
;
7108 struct page
*current_buddy
, *next_page
;
7110 while (high
> low
) {
7114 if (target
>= &page
[size
]) {
7115 next_page
= page
+ size
;
7116 current_buddy
= page
;
7119 current_buddy
= page
+ size
;
7122 if (set_page_guard(zone
, current_buddy
, high
, migratetype
))
7125 if (current_buddy
!= target
) {
7126 add_to_free_list(current_buddy
, zone
, high
, migratetype
);
7127 set_buddy_order(current_buddy
, high
);
7134 * Take a page that will be marked as poisoned off the buddy allocator.
7136 bool take_page_off_buddy(struct page
*page
)
7138 struct zone
*zone
= page_zone(page
);
7139 unsigned long pfn
= page_to_pfn(page
);
7140 unsigned long flags
;
7144 spin_lock_irqsave(&zone
->lock
, flags
);
7145 for (order
= 0; order
<= MAX_ORDER
; order
++) {
7146 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7147 int page_order
= buddy_order(page_head
);
7149 if (PageBuddy(page_head
) && page_order
>= order
) {
7150 unsigned long pfn_head
= page_to_pfn(page_head
);
7151 int migratetype
= get_pfnblock_migratetype(page_head
,
7154 del_page_from_free_list(page_head
, zone
, page_order
);
7155 break_down_buddy_pages(zone
, page_head
, page
, 0,
7156 page_order
, migratetype
);
7157 SetPageHWPoisonTakenOff(page
);
7158 if (!is_migrate_isolate(migratetype
))
7159 __mod_zone_freepage_state(zone
, -1, migratetype
);
7163 if (page_count(page_head
) > 0)
7166 spin_unlock_irqrestore(&zone
->lock
, flags
);
7171 * Cancel takeoff done by take_page_off_buddy().
7173 bool put_page_back_buddy(struct page
*page
)
7175 struct zone
*zone
= page_zone(page
);
7176 unsigned long pfn
= page_to_pfn(page
);
7177 unsigned long flags
;
7178 int migratetype
= get_pfnblock_migratetype(page
, pfn
);
7181 spin_lock_irqsave(&zone
->lock
, flags
);
7182 if (put_page_testzero(page
)) {
7183 ClearPageHWPoisonTakenOff(page
);
7184 __free_one_page(page
, pfn
, zone
, 0, migratetype
, FPI_NONE
);
7185 if (TestClearPageHWPoison(page
)) {
7189 spin_unlock_irqrestore(&zone
->lock
, flags
);
7195 #ifdef CONFIG_ZONE_DMA
7196 bool has_managed_dma(void)
7198 struct pglist_data
*pgdat
;
7200 for_each_online_pgdat(pgdat
) {
7201 struct zone
*zone
= &pgdat
->node_zones
[ZONE_DMA
];
7203 if (managed_zone(zone
))
7208 #endif /* CONFIG_ZONE_DMA */