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/interrupt.h>
22 #include <linux/jiffies.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/kasan.h>
26 #include <linux/kmsan.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/ratelimit.h>
30 #include <linux/oom.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmstat.h>
38 #include <linux/fault-inject.h>
39 #include <linux/compaction.h>
40 #include <trace/events/kmem.h>
41 #include <trace/events/oom.h>
42 #include <linux/prefetch.h>
43 #include <linux/mm_inline.h>
44 #include <linux/mmu_notifier.h>
45 #include <linux/migrate.h>
46 #include <linux/sched/mm.h>
47 #include <linux/page_owner.h>
48 #include <linux/page_table_check.h>
49 #include <linux/memcontrol.h>
50 #include <linux/ftrace.h>
51 #include <linux/lockdep.h>
52 #include <linux/psi.h>
53 #include <linux/khugepaged.h>
54 #include <linux/delayacct.h>
55 #include <asm/div64.h>
58 #include "page_reporting.h"
60 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
61 typedef int __bitwise fpi_t
;
63 /* No special request */
64 #define FPI_NONE ((__force fpi_t)0)
67 * Skip free page reporting notification for the (possibly merged) page.
68 * This does not hinder free page reporting from grabbing the page,
69 * reporting it and marking it "reported" - it only skips notifying
70 * the free page reporting infrastructure about a newly freed page. For
71 * example, used when temporarily pulling a page from a freelist and
72 * putting it back unmodified.
74 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
77 * Place the (possibly merged) page to the tail of the freelist. Will ignore
78 * page shuffling (relevant code - e.g., memory onlining - is expected to
79 * shuffle the whole zone).
81 * Note: No code should rely on this flag for correctness - it's purely
82 * to allow for optimizations when handing back either fresh pages
83 * (memory onlining) or untouched pages (page isolation, free page
86 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
88 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
89 static DEFINE_MUTEX(pcp_batch_high_lock
);
90 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
92 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
94 * On SMP, spin_trylock is sufficient protection.
95 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
97 #define pcp_trylock_prepare(flags) do { } while (0)
98 #define pcp_trylock_finish(flag) do { } while (0)
101 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
102 #define pcp_trylock_prepare(flags) local_irq_save(flags)
103 #define pcp_trylock_finish(flags) local_irq_restore(flags)
107 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
108 * a migration causing the wrong PCP to be locked and remote memory being
109 * potentially allocated, pin the task to the CPU for the lookup+lock.
110 * preempt_disable is used on !RT because it is faster than migrate_disable.
111 * migrate_disable is used on RT because otherwise RT spinlock usage is
112 * interfered with and a high priority task cannot preempt the allocator.
114 #ifndef CONFIG_PREEMPT_RT
115 #define pcpu_task_pin() preempt_disable()
116 #define pcpu_task_unpin() preempt_enable()
118 #define pcpu_task_pin() migrate_disable()
119 #define pcpu_task_unpin() migrate_enable()
123 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
124 * Return value should be used with equivalent unlock helper.
126 #define pcpu_spin_lock(type, member, ptr) \
130 _ret = this_cpu_ptr(ptr); \
131 spin_lock(&_ret->member); \
135 #define pcpu_spin_trylock(type, member, ptr) \
139 _ret = this_cpu_ptr(ptr); \
140 if (!spin_trylock(&_ret->member)) { \
147 #define pcpu_spin_unlock(member, ptr) \
149 spin_unlock(&ptr->member); \
153 /* struct per_cpu_pages specific helpers. */
154 #define pcp_spin_lock(ptr) \
155 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
157 #define pcp_spin_trylock(ptr) \
158 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
160 #define pcp_spin_unlock(ptr) \
161 pcpu_spin_unlock(lock, ptr)
163 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
164 DEFINE_PER_CPU(int, numa_node
);
165 EXPORT_PER_CPU_SYMBOL(numa_node
);
168 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
170 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
172 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
173 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
174 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
175 * defined in <linux/topology.h>.
177 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
178 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
181 static DEFINE_MUTEX(pcpu_drain_mutex
);
183 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
184 volatile unsigned long latent_entropy __latent_entropy
;
185 EXPORT_SYMBOL(latent_entropy
);
189 * Array of node states.
191 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
192 [N_POSSIBLE
] = NODE_MASK_ALL
,
193 [N_ONLINE
] = { { [0] = 1UL } },
195 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
196 #ifdef CONFIG_HIGHMEM
197 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
199 [N_MEMORY
] = { { [0] = 1UL } },
200 [N_CPU
] = { { [0] = 1UL } },
203 EXPORT_SYMBOL(node_states
);
205 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
208 * A cached value of the page's pageblock's migratetype, used when the page is
209 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
210 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
211 * Also the migratetype set in the page does not necessarily match the pcplist
212 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
213 * other index - this ensures that it will be put on the correct CMA freelist.
215 static inline int get_pcppage_migratetype(struct page
*page
)
220 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
222 page
->index
= migratetype
;
225 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
226 unsigned int pageblock_order __read_mostly
;
229 static void __free_pages_ok(struct page
*page
, unsigned int order
,
233 * results with 256, 32 in the lowmem_reserve sysctl:
234 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
235 * 1G machine -> (16M dma, 784M normal, 224M high)
236 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
237 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
238 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
240 * TBD: should special case ZONE_DMA32 machines here - in those we normally
241 * don't need any ZONE_NORMAL reservation
243 static int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
244 #ifdef CONFIG_ZONE_DMA
247 #ifdef CONFIG_ZONE_DMA32
251 #ifdef CONFIG_HIGHMEM
257 char * const zone_names
[MAX_NR_ZONES
] = {
258 #ifdef CONFIG_ZONE_DMA
261 #ifdef CONFIG_ZONE_DMA32
265 #ifdef CONFIG_HIGHMEM
269 #ifdef CONFIG_ZONE_DEVICE
274 const char * const migratetype_names
[MIGRATE_TYPES
] = {
282 #ifdef CONFIG_MEMORY_ISOLATION
287 int min_free_kbytes
= 1024;
288 int user_min_free_kbytes
= -1;
289 static int watermark_boost_factor __read_mostly
= 15000;
290 static int watermark_scale_factor
= 10;
292 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
294 EXPORT_SYMBOL(movable_zone
);
297 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
298 unsigned int nr_online_nodes __read_mostly
= 1;
299 EXPORT_SYMBOL(nr_node_ids
);
300 EXPORT_SYMBOL(nr_online_nodes
);
303 static bool page_contains_unaccepted(struct page
*page
, unsigned int order
);
304 static void accept_page(struct page
*page
, unsigned int order
);
305 static bool try_to_accept_memory(struct zone
*zone
, unsigned int order
);
306 static inline bool has_unaccepted_memory(void);
307 static bool __free_unaccepted(struct page
*page
);
309 int page_group_by_mobility_disabled __read_mostly
;
311 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
313 * During boot we initialize deferred pages on-demand, as needed, but once
314 * page_alloc_init_late() has finished, the deferred pages are all initialized,
315 * and we can permanently disable that path.
317 DEFINE_STATIC_KEY_TRUE(deferred_pages
);
319 static inline bool deferred_pages_enabled(void)
321 return static_branch_unlikely(&deferred_pages
);
325 * deferred_grow_zone() is __init, but it is called from
326 * get_page_from_freelist() during early boot until deferred_pages permanently
327 * disables this call. This is why we have refdata wrapper to avoid warning,
328 * and to ensure that the function body gets unloaded.
331 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
333 return deferred_grow_zone(zone
, order
);
336 static inline bool deferred_pages_enabled(void)
340 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
342 /* Return a pointer to the bitmap storing bits affecting a block of pages */
343 static inline unsigned long *get_pageblock_bitmap(const struct page
*page
,
346 #ifdef CONFIG_SPARSEMEM
347 return section_to_usemap(__pfn_to_section(pfn
));
349 return page_zone(page
)->pageblock_flags
;
350 #endif /* CONFIG_SPARSEMEM */
353 static inline int pfn_to_bitidx(const struct page
*page
, unsigned long pfn
)
355 #ifdef CONFIG_SPARSEMEM
356 pfn
&= (PAGES_PER_SECTION
-1);
358 pfn
= pfn
- pageblock_start_pfn(page_zone(page
)->zone_start_pfn
);
359 #endif /* CONFIG_SPARSEMEM */
360 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
364 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
365 * @page: The page within the block of interest
366 * @pfn: The target page frame number
367 * @mask: mask of bits that the caller is interested in
369 * Return: pageblock_bits flags
371 unsigned long get_pfnblock_flags_mask(const struct page
*page
,
372 unsigned long pfn
, unsigned long mask
)
374 unsigned long *bitmap
;
375 unsigned long bitidx
, word_bitidx
;
378 bitmap
= get_pageblock_bitmap(page
, pfn
);
379 bitidx
= pfn_to_bitidx(page
, pfn
);
380 word_bitidx
= bitidx
/ BITS_PER_LONG
;
381 bitidx
&= (BITS_PER_LONG
-1);
383 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
384 * a consistent read of the memory array, so that results, even though
385 * racy, are not corrupted.
387 word
= READ_ONCE(bitmap
[word_bitidx
]);
388 return (word
>> bitidx
) & mask
;
391 static __always_inline
int get_pfnblock_migratetype(const struct page
*page
,
394 return get_pfnblock_flags_mask(page
, pfn
, MIGRATETYPE_MASK
);
398 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
399 * @page: The page within the block of interest
400 * @flags: The flags to set
401 * @pfn: The target page frame number
402 * @mask: mask of bits that the caller is interested in
404 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
408 unsigned long *bitmap
;
409 unsigned long bitidx
, word_bitidx
;
412 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
413 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
415 bitmap
= get_pageblock_bitmap(page
, pfn
);
416 bitidx
= pfn_to_bitidx(page
, pfn
);
417 word_bitidx
= bitidx
/ BITS_PER_LONG
;
418 bitidx
&= (BITS_PER_LONG
-1);
420 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
425 word
= READ_ONCE(bitmap
[word_bitidx
]);
427 } while (!try_cmpxchg(&bitmap
[word_bitidx
], &word
, (word
& ~mask
) | flags
));
430 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
432 if (unlikely(page_group_by_mobility_disabled
&&
433 migratetype
< MIGRATE_PCPTYPES
))
434 migratetype
= MIGRATE_UNMOVABLE
;
436 set_pfnblock_flags_mask(page
, (unsigned long)migratetype
,
437 page_to_pfn(page
), MIGRATETYPE_MASK
);
440 #ifdef CONFIG_DEBUG_VM
441 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
445 unsigned long pfn
= page_to_pfn(page
);
446 unsigned long sp
, start_pfn
;
449 seq
= zone_span_seqbegin(zone
);
450 start_pfn
= zone
->zone_start_pfn
;
451 sp
= zone
->spanned_pages
;
452 ret
= !zone_spans_pfn(zone
, pfn
);
453 } while (zone_span_seqretry(zone
, seq
));
456 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
457 pfn
, zone_to_nid(zone
), zone
->name
,
458 start_pfn
, start_pfn
+ sp
);
464 * Temporary debugging check for pages not lying within a given zone.
466 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
468 if (page_outside_zone_boundaries(zone
, page
))
470 if (zone
!= page_zone(page
))
476 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
482 static void bad_page(struct page
*page
, const char *reason
)
484 static unsigned long resume
;
485 static unsigned long nr_shown
;
486 static unsigned long nr_unshown
;
489 * Allow a burst of 60 reports, then keep quiet for that minute;
490 * or allow a steady drip of one report per second.
492 if (nr_shown
== 60) {
493 if (time_before(jiffies
, resume
)) {
499 "BUG: Bad page state: %lu messages suppressed\n",
506 resume
= jiffies
+ 60 * HZ
;
508 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
509 current
->comm
, page_to_pfn(page
));
510 dump_page(page
, reason
);
515 /* Leave bad fields for debug, except PageBuddy could make trouble */
516 page_mapcount_reset(page
); /* remove PageBuddy */
517 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
520 static inline unsigned int order_to_pindex(int migratetype
, int order
)
522 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
523 if (order
> PAGE_ALLOC_COSTLY_ORDER
) {
524 VM_BUG_ON(order
!= pageblock_order
);
525 return NR_LOWORDER_PCP_LISTS
;
528 VM_BUG_ON(order
> PAGE_ALLOC_COSTLY_ORDER
);
531 return (MIGRATE_PCPTYPES
* order
) + migratetype
;
534 static inline int pindex_to_order(unsigned int pindex
)
536 int order
= pindex
/ MIGRATE_PCPTYPES
;
538 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
539 if (pindex
== NR_LOWORDER_PCP_LISTS
)
540 order
= pageblock_order
;
542 VM_BUG_ON(order
> PAGE_ALLOC_COSTLY_ORDER
);
548 static inline bool pcp_allowed_order(unsigned int order
)
550 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
552 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
553 if (order
== pageblock_order
)
559 static inline void free_the_page(struct page
*page
, unsigned int order
)
561 if (pcp_allowed_order(order
)) /* Via pcp? */
562 free_unref_page(page
, order
);
564 __free_pages_ok(page
, order
, FPI_NONE
);
568 * Higher-order pages are called "compound pages". They are structured thusly:
570 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
572 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
573 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
575 * The first tail page's ->compound_order holds the order of allocation.
576 * This usage means that zero-order pages may not be compound.
579 void prep_compound_page(struct page
*page
, unsigned int order
)
582 int nr_pages
= 1 << order
;
585 for (i
= 1; i
< nr_pages
; i
++)
586 prep_compound_tail(page
, i
);
588 prep_compound_head(page
, order
);
591 void destroy_large_folio(struct folio
*folio
)
593 if (folio_test_hugetlb(folio
)) {
594 free_huge_folio(folio
);
598 if (folio_test_large_rmappable(folio
))
599 folio_undo_large_rmappable(folio
);
601 mem_cgroup_uncharge(folio
);
602 free_the_page(&folio
->page
, folio_order(folio
));
605 static inline void set_buddy_order(struct page
*page
, unsigned int order
)
607 set_page_private(page
, order
);
608 __SetPageBuddy(page
);
611 #ifdef CONFIG_COMPACTION
612 static inline struct capture_control
*task_capc(struct zone
*zone
)
614 struct capture_control
*capc
= current
->capture_control
;
616 return unlikely(capc
) &&
617 !(current
->flags
& PF_KTHREAD
) &&
619 capc
->cc
->zone
== zone
? capc
: NULL
;
623 compaction_capture(struct capture_control
*capc
, struct page
*page
,
624 int order
, int migratetype
)
626 if (!capc
|| order
!= capc
->cc
->order
)
629 /* Do not accidentally pollute CMA or isolated regions*/
630 if (is_migrate_cma(migratetype
) ||
631 is_migrate_isolate(migratetype
))
635 * Do not let lower order allocations pollute a movable pageblock.
636 * This might let an unmovable request use a reclaimable pageblock
637 * and vice-versa but no more than normal fallback logic which can
638 * have trouble finding a high-order free page.
640 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
648 static inline struct capture_control
*task_capc(struct zone
*zone
)
654 compaction_capture(struct capture_control
*capc
, struct page
*page
,
655 int order
, int migratetype
)
659 #endif /* CONFIG_COMPACTION */
661 /* Used for pages not on another list */
662 static inline void add_to_free_list(struct page
*page
, struct zone
*zone
,
663 unsigned int order
, int migratetype
)
665 struct free_area
*area
= &zone
->free_area
[order
];
667 list_add(&page
->buddy_list
, &area
->free_list
[migratetype
]);
671 /* Used for pages not on another list */
672 static inline void add_to_free_list_tail(struct page
*page
, struct zone
*zone
,
673 unsigned int order
, int migratetype
)
675 struct free_area
*area
= &zone
->free_area
[order
];
677 list_add_tail(&page
->buddy_list
, &area
->free_list
[migratetype
]);
682 * Used for pages which are on another list. Move the pages to the tail
683 * of the list - so the moved pages won't immediately be considered for
684 * allocation again (e.g., optimization for memory onlining).
686 static inline void move_to_free_list(struct page
*page
, struct zone
*zone
,
687 unsigned int order
, int migratetype
)
689 struct free_area
*area
= &zone
->free_area
[order
];
691 list_move_tail(&page
->buddy_list
, &area
->free_list
[migratetype
]);
694 static inline void del_page_from_free_list(struct page
*page
, struct zone
*zone
,
697 /* clear reported state and update reported page count */
698 if (page_reported(page
))
699 __ClearPageReported(page
);
701 list_del(&page
->buddy_list
);
702 __ClearPageBuddy(page
);
703 set_page_private(page
, 0);
704 zone
->free_area
[order
].nr_free
--;
707 static inline struct page
*get_page_from_free_area(struct free_area
*area
,
710 return list_first_entry_or_null(&area
->free_list
[migratetype
],
711 struct page
, buddy_list
);
715 * If this is not the largest possible page, check if the buddy
716 * of the next-highest order is free. If it is, it's possible
717 * that pages are being freed that will coalesce soon. In case,
718 * that is happening, add the free page to the tail of the list
719 * so it's less likely to be used soon and more likely to be merged
720 * as a higher order page
723 buddy_merge_likely(unsigned long pfn
, unsigned long buddy_pfn
,
724 struct page
*page
, unsigned int order
)
726 unsigned long higher_page_pfn
;
727 struct page
*higher_page
;
729 if (order
>= MAX_ORDER
- 1)
732 higher_page_pfn
= buddy_pfn
& pfn
;
733 higher_page
= page
+ (higher_page_pfn
- pfn
);
735 return find_buddy_page_pfn(higher_page
, higher_page_pfn
, order
+ 1,
740 * Freeing function for a buddy system allocator.
742 * The concept of a buddy system is to maintain direct-mapped table
743 * (containing bit values) for memory blocks of various "orders".
744 * The bottom level table contains the map for the smallest allocatable
745 * units of memory (here, pages), and each level above it describes
746 * pairs of units from the levels below, hence, "buddies".
747 * At a high level, all that happens here is marking the table entry
748 * at the bottom level available, and propagating the changes upward
749 * as necessary, plus some accounting needed to play nicely with other
750 * parts of the VM system.
751 * At each level, we keep a list of pages, which are heads of continuous
752 * free pages of length of (1 << order) and marked with PageBuddy.
753 * Page's order is recorded in page_private(page) field.
754 * So when we are allocating or freeing one, we can derive the state of the
755 * other. That is, if we allocate a small block, and both were
756 * free, the remainder of the region must be split into blocks.
757 * If a block is freed, and its buddy is also free, then this
758 * triggers coalescing into a block of larger size.
763 static inline void __free_one_page(struct page
*page
,
765 struct zone
*zone
, unsigned int order
,
766 int migratetype
, fpi_t fpi_flags
)
768 struct capture_control
*capc
= task_capc(zone
);
769 unsigned long buddy_pfn
= 0;
770 unsigned long combined_pfn
;
774 VM_BUG_ON(!zone_is_initialized(zone
));
775 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
777 VM_BUG_ON(migratetype
== -1);
778 if (likely(!is_migrate_isolate(migratetype
)))
779 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
781 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
782 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
784 while (order
< MAX_ORDER
) {
785 if (compaction_capture(capc
, page
, order
, migratetype
)) {
786 __mod_zone_freepage_state(zone
, -(1 << order
),
791 buddy
= find_buddy_page_pfn(page
, pfn
, order
, &buddy_pfn
);
795 if (unlikely(order
>= pageblock_order
)) {
797 * We want to prevent merge between freepages on pageblock
798 * without fallbacks and normal pageblock. Without this,
799 * pageblock isolation could cause incorrect freepage or CMA
800 * accounting or HIGHATOMIC accounting.
802 int buddy_mt
= get_pfnblock_migratetype(buddy
, buddy_pfn
);
804 if (migratetype
!= buddy_mt
805 && (!migratetype_is_mergeable(migratetype
) ||
806 !migratetype_is_mergeable(buddy_mt
)))
811 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
812 * merge with it and move up one order.
814 if (page_is_guard(buddy
))
815 clear_page_guard(zone
, buddy
, order
, migratetype
);
817 del_page_from_free_list(buddy
, zone
, order
);
818 combined_pfn
= buddy_pfn
& pfn
;
819 page
= page
+ (combined_pfn
- pfn
);
825 set_buddy_order(page
, order
);
827 if (fpi_flags
& FPI_TO_TAIL
)
829 else if (is_shuffle_order(order
))
830 to_tail
= shuffle_pick_tail();
832 to_tail
= buddy_merge_likely(pfn
, buddy_pfn
, page
, order
);
835 add_to_free_list_tail(page
, zone
, order
, migratetype
);
837 add_to_free_list(page
, zone
, order
, migratetype
);
839 /* Notify page reporting subsystem of freed page */
840 if (!(fpi_flags
& FPI_SKIP_REPORT_NOTIFY
))
841 page_reporting_notify_free(order
);
845 * split_free_page() -- split a free page at split_pfn_offset
846 * @free_page: the original free page
847 * @order: the order of the page
848 * @split_pfn_offset: split offset within the page
850 * Return -ENOENT if the free page is changed, otherwise 0
852 * It is used when the free page crosses two pageblocks with different migratetypes
853 * at split_pfn_offset within the page. The split free page will be put into
854 * separate migratetype lists afterwards. Otherwise, the function achieves
857 int split_free_page(struct page
*free_page
,
858 unsigned int order
, unsigned long split_pfn_offset
)
860 struct zone
*zone
= page_zone(free_page
);
861 unsigned long free_page_pfn
= page_to_pfn(free_page
);
868 if (split_pfn_offset
== 0)
871 spin_lock_irqsave(&zone
->lock
, flags
);
873 if (!PageBuddy(free_page
) || buddy_order(free_page
) != order
) {
878 mt
= get_pfnblock_migratetype(free_page
, free_page_pfn
);
879 if (likely(!is_migrate_isolate(mt
)))
880 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
882 del_page_from_free_list(free_page
, zone
, order
);
883 for (pfn
= free_page_pfn
;
884 pfn
< free_page_pfn
+ (1UL << order
);) {
885 int mt
= get_pfnblock_migratetype(pfn_to_page(pfn
), pfn
);
887 free_page_order
= min_t(unsigned int,
888 pfn
? __ffs(pfn
) : order
,
889 __fls(split_pfn_offset
));
890 __free_one_page(pfn_to_page(pfn
), pfn
, zone
, free_page_order
,
892 pfn
+= 1UL << free_page_order
;
893 split_pfn_offset
-= (1UL << free_page_order
);
894 /* we have done the first part, now switch to second part */
895 if (split_pfn_offset
== 0)
896 split_pfn_offset
= (1UL << order
) - (pfn
- free_page_pfn
);
899 spin_unlock_irqrestore(&zone
->lock
, flags
);
903 * A bad page could be due to a number of fields. Instead of multiple branches,
904 * try and check multiple fields with one check. The caller must do a detailed
905 * check if necessary.
907 static inline bool page_expected_state(struct page
*page
,
908 unsigned long check_flags
)
910 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
913 if (unlikely((unsigned long)page
->mapping
|
914 page_ref_count(page
) |
918 (page
->flags
& check_flags
)))
924 static const char *page_bad_reason(struct page
*page
, unsigned long flags
)
926 const char *bad_reason
= NULL
;
928 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
929 bad_reason
= "nonzero mapcount";
930 if (unlikely(page
->mapping
!= NULL
))
931 bad_reason
= "non-NULL mapping";
932 if (unlikely(page_ref_count(page
) != 0))
933 bad_reason
= "nonzero _refcount";
934 if (unlikely(page
->flags
& flags
)) {
935 if (flags
== PAGE_FLAGS_CHECK_AT_PREP
)
936 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
938 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
941 if (unlikely(page
->memcg_data
))
942 bad_reason
= "page still charged to cgroup";
947 static void free_page_is_bad_report(struct page
*page
)
950 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_FREE
));
953 static inline bool free_page_is_bad(struct page
*page
)
955 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
958 /* Something has gone sideways, find it */
959 free_page_is_bad_report(page
);
963 static inline bool is_check_pages_enabled(void)
965 return static_branch_unlikely(&check_pages_enabled
);
968 static int free_tail_page_prepare(struct page
*head_page
, struct page
*page
)
970 struct folio
*folio
= (struct folio
*)head_page
;
974 * We rely page->lru.next never has bit 0 set, unless the page
975 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
977 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
979 if (!is_check_pages_enabled()) {
983 switch (page
- head_page
) {
985 /* the first tail page: these may be in place of ->mapping */
986 if (unlikely(folio_entire_mapcount(folio
))) {
987 bad_page(page
, "nonzero entire_mapcount");
990 if (unlikely(atomic_read(&folio
->_nr_pages_mapped
))) {
991 bad_page(page
, "nonzero nr_pages_mapped");
994 if (unlikely(atomic_read(&folio
->_pincount
))) {
995 bad_page(page
, "nonzero pincount");
1001 * the second tail page: ->mapping is
1002 * deferred_list.next -- ignore value.
1006 if (page
->mapping
!= TAIL_MAPPING
) {
1007 bad_page(page
, "corrupted mapping in tail page");
1012 if (unlikely(!PageTail(page
))) {
1013 bad_page(page
, "PageTail not set");
1016 if (unlikely(compound_head(page
) != head_page
)) {
1017 bad_page(page
, "compound_head not consistent");
1022 page
->mapping
= NULL
;
1023 clear_compound_head(page
);
1028 * Skip KASAN memory poisoning when either:
1030 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1031 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1032 * using page tags instead (see below).
1033 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1034 * that error detection is disabled for accesses via the page address.
1036 * Pages will have match-all tags in the following circumstances:
1038 * 1. Pages are being initialized for the first time, including during deferred
1039 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1040 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1041 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1042 * 3. The allocation was excluded from being checked due to sampling,
1043 * see the call to kasan_unpoison_pages.
1045 * Poisoning pages during deferred memory init will greatly lengthen the
1046 * process and cause problem in large memory systems as the deferred pages
1047 * initialization is done with interrupt disabled.
1049 * Assuming that there will be no reference to those newly initialized
1050 * pages before they are ever allocated, this should have no effect on
1051 * KASAN memory tracking as the poison will be properly inserted at page
1052 * allocation time. The only corner case is when pages are allocated by
1053 * on-demand allocation and then freed again before the deferred pages
1054 * initialization is done, but this is not likely to happen.
1056 static inline bool should_skip_kasan_poison(struct page
*page
, fpi_t fpi_flags
)
1058 if (IS_ENABLED(CONFIG_KASAN_GENERIC
))
1059 return deferred_pages_enabled();
1061 return page_kasan_tag(page
) == 0xff;
1064 static void kernel_init_pages(struct page
*page
, int numpages
)
1068 /* s390's use of memset() could override KASAN redzones. */
1069 kasan_disable_current();
1070 for (i
= 0; i
< numpages
; i
++)
1071 clear_highpage_kasan_tagged(page
+ i
);
1072 kasan_enable_current();
1075 static __always_inline
bool free_pages_prepare(struct page
*page
,
1076 unsigned int order
, fpi_t fpi_flags
)
1079 bool skip_kasan_poison
= should_skip_kasan_poison(page
, fpi_flags
);
1080 bool init
= want_init_on_free();
1082 VM_BUG_ON_PAGE(PageTail(page
), page
);
1084 trace_mm_page_free(page
, order
);
1085 kmsan_free_page(page
, order
);
1087 if (unlikely(PageHWPoison(page
)) && !order
) {
1089 * Do not let hwpoison pages hit pcplists/buddy
1090 * Untie memcg state and reset page's owner
1092 if (memcg_kmem_online() && PageMemcgKmem(page
))
1093 __memcg_kmem_uncharge_page(page
, order
);
1094 reset_page_owner(page
, order
);
1095 page_table_check_free(page
, order
);
1100 * Check tail pages before head page information is cleared to
1101 * avoid checking PageCompound for order-0 pages.
1103 if (unlikely(order
)) {
1104 bool compound
= PageCompound(page
);
1107 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1110 page
[1].flags
&= ~PAGE_FLAGS_SECOND
;
1111 for (i
= 1; i
< (1 << order
); i
++) {
1113 bad
+= free_tail_page_prepare(page
, page
+ i
);
1114 if (is_check_pages_enabled()) {
1115 if (free_page_is_bad(page
+ i
)) {
1120 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1123 if (PageMappingFlags(page
))
1124 page
->mapping
= NULL
;
1125 if (memcg_kmem_online() && PageMemcgKmem(page
))
1126 __memcg_kmem_uncharge_page(page
, order
);
1127 if (is_check_pages_enabled()) {
1128 if (free_page_is_bad(page
))
1134 page_cpupid_reset_last(page
);
1135 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1136 reset_page_owner(page
, order
);
1137 page_table_check_free(page
, order
);
1139 if (!PageHighMem(page
)) {
1140 debug_check_no_locks_freed(page_address(page
),
1141 PAGE_SIZE
<< order
);
1142 debug_check_no_obj_freed(page_address(page
),
1143 PAGE_SIZE
<< order
);
1146 kernel_poison_pages(page
, 1 << order
);
1149 * As memory initialization might be integrated into KASAN,
1150 * KASAN poisoning and memory initialization code must be
1151 * kept together to avoid discrepancies in behavior.
1153 * With hardware tag-based KASAN, memory tags must be set before the
1154 * page becomes unavailable via debug_pagealloc or arch_free_page.
1156 if (!skip_kasan_poison
) {
1157 kasan_poison_pages(page
, order
, init
);
1159 /* Memory is already initialized if KASAN did it internally. */
1160 if (kasan_has_integrated_init())
1164 kernel_init_pages(page
, 1 << order
);
1167 * arch_free_page() can make the page's contents inaccessible. s390
1168 * does this. So nothing which can access the page's contents should
1169 * happen after this.
1171 arch_free_page(page
, order
);
1173 debug_pagealloc_unmap_pages(page
, 1 << order
);
1179 * Frees a number of pages from the PCP lists
1180 * Assumes all pages on list are in same zone.
1181 * count is the number of pages to free.
1183 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1184 struct per_cpu_pages
*pcp
,
1187 unsigned long flags
;
1189 bool isolated_pageblocks
;
1193 * Ensure proper count is passed which otherwise would stuck in the
1194 * below while (list_empty(list)) loop.
1196 count
= min(pcp
->count
, count
);
1198 /* Ensure requested pindex is drained first. */
1199 pindex
= pindex
- 1;
1201 spin_lock_irqsave(&zone
->lock
, flags
);
1202 isolated_pageblocks
= has_isolate_pageblock(zone
);
1205 struct list_head
*list
;
1208 /* Remove pages from lists in a round-robin fashion. */
1210 if (++pindex
> NR_PCP_LISTS
- 1)
1212 list
= &pcp
->lists
[pindex
];
1213 } while (list_empty(list
));
1215 order
= pindex_to_order(pindex
);
1216 nr_pages
= 1 << order
;
1220 page
= list_last_entry(list
, struct page
, pcp_list
);
1221 mt
= get_pcppage_migratetype(page
);
1223 /* must delete to avoid corrupting pcp list */
1224 list_del(&page
->pcp_list
);
1226 pcp
->count
-= nr_pages
;
1228 /* MIGRATE_ISOLATE page should not go to pcplists */
1229 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1230 /* Pageblock could have been isolated meanwhile */
1231 if (unlikely(isolated_pageblocks
))
1232 mt
= get_pageblock_migratetype(page
);
1234 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
, FPI_NONE
);
1235 trace_mm_page_pcpu_drain(page
, order
, mt
);
1236 } while (count
> 0 && !list_empty(list
));
1239 spin_unlock_irqrestore(&zone
->lock
, flags
);
1242 static void free_one_page(struct zone
*zone
,
1243 struct page
*page
, unsigned long pfn
,
1245 int migratetype
, fpi_t fpi_flags
)
1247 unsigned long flags
;
1249 spin_lock_irqsave(&zone
->lock
, flags
);
1250 if (unlikely(has_isolate_pageblock(zone
) ||
1251 is_migrate_isolate(migratetype
))) {
1252 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1254 __free_one_page(page
, pfn
, zone
, order
, migratetype
, fpi_flags
);
1255 spin_unlock_irqrestore(&zone
->lock
, flags
);
1258 static void __free_pages_ok(struct page
*page
, unsigned int order
,
1261 unsigned long flags
;
1263 unsigned long pfn
= page_to_pfn(page
);
1264 struct zone
*zone
= page_zone(page
);
1266 if (!free_pages_prepare(page
, order
, fpi_flags
))
1270 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1271 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1272 * This will reduce the lock holding time.
1274 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1276 spin_lock_irqsave(&zone
->lock
, flags
);
1277 if (unlikely(has_isolate_pageblock(zone
) ||
1278 is_migrate_isolate(migratetype
))) {
1279 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1281 __free_one_page(page
, pfn
, zone
, order
, migratetype
, fpi_flags
);
1282 spin_unlock_irqrestore(&zone
->lock
, flags
);
1284 __count_vm_events(PGFREE
, 1 << order
);
1287 void __free_pages_core(struct page
*page
, unsigned int order
)
1289 unsigned int nr_pages
= 1 << order
;
1290 struct page
*p
= page
;
1294 * When initializing the memmap, __init_single_page() sets the refcount
1295 * of all pages to 1 ("allocated"/"not free"). We have to set the
1296 * refcount of all involved pages to 0.
1299 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1301 __ClearPageReserved(p
);
1302 set_page_count(p
, 0);
1304 __ClearPageReserved(p
);
1305 set_page_count(p
, 0);
1307 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1309 if (page_contains_unaccepted(page
, order
)) {
1310 if (order
== MAX_ORDER
&& __free_unaccepted(page
))
1313 accept_page(page
, order
);
1317 * Bypass PCP and place fresh pages right to the tail, primarily
1318 * relevant for memory onlining.
1320 __free_pages_ok(page
, order
, FPI_TO_TAIL
);
1324 * Check that the whole (or subset of) a pageblock given by the interval of
1325 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1326 * with the migration of free compaction scanner.
1328 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1330 * It's possible on some configurations to have a setup like node0 node1 node0
1331 * i.e. it's possible that all pages within a zones range of pages do not
1332 * belong to a single zone. We assume that a border between node0 and node1
1333 * can occur within a single pageblock, but not a node0 node1 node0
1334 * interleaving within a single pageblock. It is therefore sufficient to check
1335 * the first and last page of a pageblock and avoid checking each individual
1336 * page in a pageblock.
1338 * Note: the function may return non-NULL struct page even for a page block
1339 * which contains a memory hole (i.e. there is no physical memory for a subset
1340 * of the pfn range). For example, if the pageblock order is MAX_ORDER, which
1341 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1342 * even though the start pfn is online and valid. This should be safe most of
1343 * the time because struct pages are still initialized via init_unavailable_range()
1344 * and pfn walkers shouldn't touch any physical memory range for which they do
1345 * not recognize any specific metadata in struct pages.
1347 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1348 unsigned long end_pfn
, struct zone
*zone
)
1350 struct page
*start_page
;
1351 struct page
*end_page
;
1353 /* end_pfn is one past the range we are checking */
1356 if (!pfn_valid(end_pfn
))
1359 start_page
= pfn_to_online_page(start_pfn
);
1363 if (page_zone(start_page
) != zone
)
1366 end_page
= pfn_to_page(end_pfn
);
1368 /* This gives a shorter code than deriving page_zone(end_page) */
1369 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1376 * The order of subdivision here is critical for the IO subsystem.
1377 * Please do not alter this order without good reasons and regression
1378 * testing. Specifically, as large blocks of memory are subdivided,
1379 * the order in which smaller blocks are delivered depends on the order
1380 * they're subdivided in this function. This is the primary factor
1381 * influencing the order in which pages are delivered to the IO
1382 * subsystem according to empirical testing, and this is also justified
1383 * by considering the behavior of a buddy system containing a single
1384 * large block of memory acted on by a series of small allocations.
1385 * This behavior is a critical factor in sglist merging's success.
1389 static inline void expand(struct zone
*zone
, struct page
*page
,
1390 int low
, int high
, int migratetype
)
1392 unsigned long size
= 1 << high
;
1394 while (high
> low
) {
1397 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1400 * Mark as guard pages (or page), that will allow to
1401 * merge back to allocator when buddy will be freed.
1402 * Corresponding page table entries will not be touched,
1403 * pages will stay not present in virtual address space
1405 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1408 add_to_free_list(&page
[size
], zone
, high
, migratetype
);
1409 set_buddy_order(&page
[size
], high
);
1413 static void check_new_page_bad(struct page
*page
)
1415 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1416 /* Don't complain about hwpoisoned pages */
1417 page_mapcount_reset(page
); /* remove PageBuddy */
1422 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_PREP
));
1426 * This page is about to be returned from the page allocator
1428 static int check_new_page(struct page
*page
)
1430 if (likely(page_expected_state(page
,
1431 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1434 check_new_page_bad(page
);
1438 static inline bool check_new_pages(struct page
*page
, unsigned int order
)
1440 if (is_check_pages_enabled()) {
1441 for (int i
= 0; i
< (1 << order
); i
++) {
1442 struct page
*p
= page
+ i
;
1444 if (check_new_page(p
))
1452 static inline bool should_skip_kasan_unpoison(gfp_t flags
)
1454 /* Don't skip if a software KASAN mode is enabled. */
1455 if (IS_ENABLED(CONFIG_KASAN_GENERIC
) ||
1456 IS_ENABLED(CONFIG_KASAN_SW_TAGS
))
1459 /* Skip, if hardware tag-based KASAN is not enabled. */
1460 if (!kasan_hw_tags_enabled())
1464 * With hardware tag-based KASAN enabled, skip if this has been
1465 * requested via __GFP_SKIP_KASAN.
1467 return flags
& __GFP_SKIP_KASAN
;
1470 static inline bool should_skip_init(gfp_t flags
)
1472 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1473 if (!kasan_hw_tags_enabled())
1476 /* For hardware tag-based KASAN, skip if requested. */
1477 return (flags
& __GFP_SKIP_ZERO
);
1480 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1483 bool init
= !want_init_on_free() && want_init_on_alloc(gfp_flags
) &&
1484 !should_skip_init(gfp_flags
);
1485 bool zero_tags
= init
&& (gfp_flags
& __GFP_ZEROTAGS
);
1488 set_page_private(page
, 0);
1489 set_page_refcounted(page
);
1491 arch_alloc_page(page
, order
);
1492 debug_pagealloc_map_pages(page
, 1 << order
);
1495 * Page unpoisoning must happen before memory initialization.
1496 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1497 * allocations and the page unpoisoning code will complain.
1499 kernel_unpoison_pages(page
, 1 << order
);
1502 * As memory initialization might be integrated into KASAN,
1503 * KASAN unpoisoning and memory initializion code must be
1504 * kept together to avoid discrepancies in behavior.
1508 * If memory tags should be zeroed
1509 * (which happens only when memory should be initialized as well).
1512 /* Initialize both memory and memory tags. */
1513 for (i
= 0; i
!= 1 << order
; ++i
)
1514 tag_clear_highpage(page
+ i
);
1516 /* Take note that memory was initialized by the loop above. */
1519 if (!should_skip_kasan_unpoison(gfp_flags
) &&
1520 kasan_unpoison_pages(page
, order
, init
)) {
1521 /* Take note that memory was initialized by KASAN. */
1522 if (kasan_has_integrated_init())
1526 * If memory tags have not been set by KASAN, reset the page
1527 * tags to ensure page_address() dereferencing does not fault.
1529 for (i
= 0; i
!= 1 << order
; ++i
)
1530 page_kasan_tag_reset(page
+ i
);
1532 /* If memory is still not initialized, initialize it now. */
1534 kernel_init_pages(page
, 1 << order
);
1536 set_page_owner(page
, order
, gfp_flags
);
1537 page_table_check_alloc(page
, order
);
1540 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1541 unsigned int alloc_flags
)
1543 post_alloc_hook(page
, order
, gfp_flags
);
1545 if (order
&& (gfp_flags
& __GFP_COMP
))
1546 prep_compound_page(page
, order
);
1549 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1550 * allocate the page. The expectation is that the caller is taking
1551 * steps that will free more memory. The caller should avoid the page
1552 * being used for !PFMEMALLOC purposes.
1554 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1555 set_page_pfmemalloc(page
);
1557 clear_page_pfmemalloc(page
);
1561 * Go through the free lists for the given migratetype and remove
1562 * the smallest available page from the freelists
1564 static __always_inline
1565 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1568 unsigned int current_order
;
1569 struct free_area
*area
;
1572 /* Find a page of the appropriate size in the preferred list */
1573 for (current_order
= order
; current_order
<= MAX_ORDER
; ++current_order
) {
1574 area
= &(zone
->free_area
[current_order
]);
1575 page
= get_page_from_free_area(area
, migratetype
);
1578 del_page_from_free_list(page
, zone
, current_order
);
1579 expand(zone
, page
, order
, current_order
, migratetype
);
1580 set_pcppage_migratetype(page
, migratetype
);
1581 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
,
1582 pcp_allowed_order(order
) &&
1583 migratetype
< MIGRATE_PCPTYPES
);
1592 * This array describes the order lists are fallen back to when
1593 * the free lists for the desirable migrate type are depleted
1595 * The other migratetypes do not have fallbacks.
1597 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_PCPTYPES
- 1] = {
1598 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
},
1599 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
},
1600 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
},
1604 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1607 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1610 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1611 unsigned int order
) { return NULL
; }
1615 * Move the free pages in a range to the freelist tail of the requested type.
1616 * Note that start_page and end_pages are not aligned on a pageblock
1617 * boundary. If alignment is required, use move_freepages_block()
1619 static int move_freepages(struct zone
*zone
,
1620 unsigned long start_pfn
, unsigned long end_pfn
,
1621 int migratetype
, int *num_movable
)
1626 int pages_moved
= 0;
1628 for (pfn
= start_pfn
; pfn
<= end_pfn
;) {
1629 page
= pfn_to_page(pfn
);
1630 if (!PageBuddy(page
)) {
1632 * We assume that pages that could be isolated for
1633 * migration are movable. But we don't actually try
1634 * isolating, as that would be expensive.
1637 (PageLRU(page
) || __PageMovable(page
)))
1643 /* Make sure we are not inadvertently changing nodes */
1644 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1645 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
1647 order
= buddy_order(page
);
1648 move_to_free_list(page
, zone
, order
, migratetype
);
1650 pages_moved
+= 1 << order
;
1656 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1657 int migratetype
, int *num_movable
)
1659 unsigned long start_pfn
, end_pfn
, pfn
;
1664 pfn
= page_to_pfn(page
);
1665 start_pfn
= pageblock_start_pfn(pfn
);
1666 end_pfn
= pageblock_end_pfn(pfn
) - 1;
1668 /* Do not cross zone boundaries */
1669 if (!zone_spans_pfn(zone
, start_pfn
))
1671 if (!zone_spans_pfn(zone
, end_pfn
))
1674 return move_freepages(zone
, start_pfn
, end_pfn
, migratetype
,
1678 static void change_pageblock_range(struct page
*pageblock_page
,
1679 int start_order
, int migratetype
)
1681 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1683 while (nr_pageblocks
--) {
1684 set_pageblock_migratetype(pageblock_page
, migratetype
);
1685 pageblock_page
+= pageblock_nr_pages
;
1690 * When we are falling back to another migratetype during allocation, try to
1691 * steal extra free pages from the same pageblocks to satisfy further
1692 * allocations, instead of polluting multiple pageblocks.
1694 * If we are stealing a relatively large buddy page, it is likely there will
1695 * be more free pages in the pageblock, so try to steal them all. For
1696 * reclaimable and unmovable allocations, we steal regardless of page size,
1697 * as fragmentation caused by those allocations polluting movable pageblocks
1698 * is worse than movable allocations stealing from unmovable and reclaimable
1701 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1704 * Leaving this order check is intended, although there is
1705 * relaxed order check in next check. The reason is that
1706 * we can actually steal whole pageblock if this condition met,
1707 * but, below check doesn't guarantee it and that is just heuristic
1708 * so could be changed anytime.
1710 if (order
>= pageblock_order
)
1713 if (order
>= pageblock_order
/ 2 ||
1714 start_mt
== MIGRATE_RECLAIMABLE
||
1715 start_mt
== MIGRATE_UNMOVABLE
||
1716 page_group_by_mobility_disabled
)
1722 static inline bool boost_watermark(struct zone
*zone
)
1724 unsigned long max_boost
;
1726 if (!watermark_boost_factor
)
1729 * Don't bother in zones that are unlikely to produce results.
1730 * On small machines, including kdump capture kernels running
1731 * in a small area, boosting the watermark can cause an out of
1732 * memory situation immediately.
1734 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
1737 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
1738 watermark_boost_factor
, 10000);
1741 * high watermark may be uninitialised if fragmentation occurs
1742 * very early in boot so do not boost. We do not fall
1743 * through and boost by pageblock_nr_pages as failing
1744 * allocations that early means that reclaim is not going
1745 * to help and it may even be impossible to reclaim the
1746 * boosted watermark resulting in a hang.
1751 max_boost
= max(pageblock_nr_pages
, max_boost
);
1753 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
1760 * This function implements actual steal behaviour. If order is large enough,
1761 * we can steal whole pageblock. If not, we first move freepages in this
1762 * pageblock to our migratetype and determine how many already-allocated pages
1763 * are there in the pageblock with a compatible migratetype. If at least half
1764 * of pages are free or compatible, we can change migratetype of the pageblock
1765 * itself, so pages freed in the future will be put on the correct free list.
1767 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1768 unsigned int alloc_flags
, int start_type
, bool whole_block
)
1770 unsigned int current_order
= buddy_order(page
);
1771 int free_pages
, movable_pages
, alike_pages
;
1774 old_block_type
= get_pageblock_migratetype(page
);
1777 * This can happen due to races and we want to prevent broken
1778 * highatomic accounting.
1780 if (is_migrate_highatomic(old_block_type
))
1783 /* Take ownership for orders >= pageblock_order */
1784 if (current_order
>= pageblock_order
) {
1785 change_pageblock_range(page
, current_order
, start_type
);
1790 * Boost watermarks to increase reclaim pressure to reduce the
1791 * likelihood of future fallbacks. Wake kswapd now as the node
1792 * may be balanced overall and kswapd will not wake naturally.
1794 if (boost_watermark(zone
) && (alloc_flags
& ALLOC_KSWAPD
))
1795 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
1797 /* We are not allowed to try stealing from the whole block */
1801 free_pages
= move_freepages_block(zone
, page
, start_type
,
1803 /* moving whole block can fail due to zone boundary conditions */
1808 * Determine how many pages are compatible with our allocation.
1809 * For movable allocation, it's the number of movable pages which
1810 * we just obtained. For other types it's a bit more tricky.
1812 if (start_type
== MIGRATE_MOVABLE
) {
1813 alike_pages
= movable_pages
;
1816 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1817 * to MOVABLE pageblock, consider all non-movable pages as
1818 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1819 * vice versa, be conservative since we can't distinguish the
1820 * exact migratetype of non-movable pages.
1822 if (old_block_type
== MIGRATE_MOVABLE
)
1823 alike_pages
= pageblock_nr_pages
1824 - (free_pages
+ movable_pages
);
1829 * If a sufficient number of pages in the block are either free or of
1830 * compatible migratability as our allocation, claim the whole block.
1832 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
1833 page_group_by_mobility_disabled
)
1834 set_pageblock_migratetype(page
, start_type
);
1839 move_to_free_list(page
, zone
, current_order
, start_type
);
1843 * Check whether there is a suitable fallback freepage with requested order.
1844 * If only_stealable is true, this function returns fallback_mt only if
1845 * we can steal other freepages all together. This would help to reduce
1846 * fragmentation due to mixed migratetype pages in one pageblock.
1848 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1849 int migratetype
, bool only_stealable
, bool *can_steal
)
1854 if (area
->nr_free
== 0)
1858 for (i
= 0; i
< MIGRATE_PCPTYPES
- 1 ; i
++) {
1859 fallback_mt
= fallbacks
[migratetype
][i
];
1860 if (free_area_empty(area
, fallback_mt
))
1863 if (can_steal_fallback(order
, migratetype
))
1866 if (!only_stealable
)
1877 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1878 * there are no empty page blocks that contain a page with a suitable order
1880 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
)
1883 unsigned long max_managed
, flags
;
1886 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1887 * Check is race-prone but harmless.
1889 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
1890 if (zone
->nr_reserved_highatomic
>= max_managed
)
1893 spin_lock_irqsave(&zone
->lock
, flags
);
1895 /* Recheck the nr_reserved_highatomic limit under the lock */
1896 if (zone
->nr_reserved_highatomic
>= max_managed
)
1900 mt
= get_pageblock_migratetype(page
);
1901 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1902 if (migratetype_is_mergeable(mt
)) {
1903 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
1904 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
1905 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
1909 spin_unlock_irqrestore(&zone
->lock
, flags
);
1913 * Used when an allocation is about to fail under memory pressure. This
1914 * potentially hurts the reliability of high-order allocations when under
1915 * intense memory pressure but failed atomic allocations should be easier
1916 * to recover from than an OOM.
1918 * If @force is true, try to unreserve a pageblock even though highatomic
1919 * pageblock is exhausted.
1921 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
1924 struct zonelist
*zonelist
= ac
->zonelist
;
1925 unsigned long flags
;
1932 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->highest_zoneidx
,
1935 * Preserve at least one pageblock unless memory pressure
1938 if (!force
&& zone
->nr_reserved_highatomic
<=
1942 spin_lock_irqsave(&zone
->lock
, flags
);
1943 for (order
= 0; order
<= MAX_ORDER
; order
++) {
1944 struct free_area
*area
= &(zone
->free_area
[order
]);
1946 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
1951 * In page freeing path, migratetype change is racy so
1952 * we can counter several free pages in a pageblock
1953 * in this loop although we changed the pageblock type
1954 * from highatomic to ac->migratetype. So we should
1955 * adjust the count once.
1957 if (is_migrate_highatomic_page(page
)) {
1959 * It should never happen but changes to
1960 * locking could inadvertently allow a per-cpu
1961 * drain to add pages to MIGRATE_HIGHATOMIC
1962 * while unreserving so be safe and watch for
1965 zone
->nr_reserved_highatomic
-= min(
1967 zone
->nr_reserved_highatomic
);
1971 * Convert to ac->migratetype and avoid the normal
1972 * pageblock stealing heuristics. Minimally, the caller
1973 * is doing the work and needs the pages. More
1974 * importantly, if the block was always converted to
1975 * MIGRATE_UNMOVABLE or another type then the number
1976 * of pageblocks that cannot be completely freed
1979 set_pageblock_migratetype(page
, ac
->migratetype
);
1980 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
1983 spin_unlock_irqrestore(&zone
->lock
, flags
);
1987 spin_unlock_irqrestore(&zone
->lock
, flags
);
1994 * Try finding a free buddy page on the fallback list and put it on the free
1995 * list of requested migratetype, possibly along with other pages from the same
1996 * block, depending on fragmentation avoidance heuristics. Returns true if
1997 * fallback was found so that __rmqueue_smallest() can grab it.
1999 * The use of signed ints for order and current_order is a deliberate
2000 * deviation from the rest of this file, to make the for loop
2001 * condition simpler.
2003 static __always_inline
bool
2004 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2005 unsigned int alloc_flags
)
2007 struct free_area
*area
;
2009 int min_order
= order
;
2015 * Do not steal pages from freelists belonging to other pageblocks
2016 * i.e. orders < pageblock_order. If there are no local zones free,
2017 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2019 if (order
< pageblock_order
&& alloc_flags
& ALLOC_NOFRAGMENT
)
2020 min_order
= pageblock_order
;
2023 * Find the largest available free page in the other list. This roughly
2024 * approximates finding the pageblock with the most free pages, which
2025 * would be too costly to do exactly.
2027 for (current_order
= MAX_ORDER
; current_order
>= min_order
;
2029 area
= &(zone
->free_area
[current_order
]);
2030 fallback_mt
= find_suitable_fallback(area
, current_order
,
2031 start_migratetype
, false, &can_steal
);
2032 if (fallback_mt
== -1)
2036 * We cannot steal all free pages from the pageblock and the
2037 * requested migratetype is movable. In that case it's better to
2038 * steal and split the smallest available page instead of the
2039 * largest available page, because even if the next movable
2040 * allocation falls back into a different pageblock than this
2041 * one, it won't cause permanent fragmentation.
2043 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2044 && current_order
> order
)
2053 for (current_order
= order
; current_order
<= MAX_ORDER
;
2055 area
= &(zone
->free_area
[current_order
]);
2056 fallback_mt
= find_suitable_fallback(area
, current_order
,
2057 start_migratetype
, false, &can_steal
);
2058 if (fallback_mt
!= -1)
2063 * This should not happen - we already found a suitable fallback
2064 * when looking for the largest page.
2066 VM_BUG_ON(current_order
> MAX_ORDER
);
2069 page
= get_page_from_free_area(area
, fallback_mt
);
2071 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2074 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2075 start_migratetype
, fallback_mt
);
2082 * Do the hard work of removing an element from the buddy allocator.
2083 * Call me with the zone->lock already held.
2085 static __always_inline
struct page
*
2086 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2087 unsigned int alloc_flags
)
2091 if (IS_ENABLED(CONFIG_CMA
)) {
2093 * Balance movable allocations between regular and CMA areas by
2094 * allocating from CMA when over half of the zone's free memory
2095 * is in the CMA area.
2097 if (alloc_flags
& ALLOC_CMA
&&
2098 zone_page_state(zone
, NR_FREE_CMA_PAGES
) >
2099 zone_page_state(zone
, NR_FREE_PAGES
) / 2) {
2100 page
= __rmqueue_cma_fallback(zone
, order
);
2106 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2107 if (unlikely(!page
)) {
2108 if (alloc_flags
& ALLOC_CMA
)
2109 page
= __rmqueue_cma_fallback(zone
, order
);
2111 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2119 * Obtain a specified number of elements from the buddy allocator, all under
2120 * a single hold of the lock, for efficiency. Add them to the supplied list.
2121 * Returns the number of new pages which were placed at *list.
2123 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2124 unsigned long count
, struct list_head
*list
,
2125 int migratetype
, unsigned int alloc_flags
)
2127 unsigned long flags
;
2130 spin_lock_irqsave(&zone
->lock
, flags
);
2131 for (i
= 0; i
< count
; ++i
) {
2132 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2134 if (unlikely(page
== NULL
))
2138 * Split buddy pages returned by expand() are received here in
2139 * physical page order. The page is added to the tail of
2140 * caller's list. From the callers perspective, the linked list
2141 * is ordered by page number under some conditions. This is
2142 * useful for IO devices that can forward direction from the
2143 * head, thus also in the physical page order. This is useful
2144 * for IO devices that can merge IO requests if the physical
2145 * pages are ordered properly.
2147 list_add_tail(&page
->pcp_list
, list
);
2148 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2149 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2153 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2154 spin_unlock_irqrestore(&zone
->lock
, flags
);
2161 * Called from the vmstat counter updater to drain pagesets of this
2162 * currently executing processor on remote nodes after they have
2165 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2167 int to_drain
, batch
;
2169 batch
= READ_ONCE(pcp
->batch
);
2170 to_drain
= min(pcp
->count
, batch
);
2172 spin_lock(&pcp
->lock
);
2173 free_pcppages_bulk(zone
, to_drain
, pcp
, 0);
2174 spin_unlock(&pcp
->lock
);
2180 * Drain pcplists of the indicated processor and zone.
2182 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2184 struct per_cpu_pages
*pcp
;
2186 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
2188 spin_lock(&pcp
->lock
);
2189 free_pcppages_bulk(zone
, pcp
->count
, pcp
, 0);
2190 spin_unlock(&pcp
->lock
);
2195 * Drain pcplists of all zones on the indicated processor.
2197 static void drain_pages(unsigned int cpu
)
2201 for_each_populated_zone(zone
) {
2202 drain_pages_zone(cpu
, zone
);
2207 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2209 void drain_local_pages(struct zone
*zone
)
2211 int cpu
= smp_processor_id();
2214 drain_pages_zone(cpu
, zone
);
2220 * The implementation of drain_all_pages(), exposing an extra parameter to
2221 * drain on all cpus.
2223 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2224 * not empty. The check for non-emptiness can however race with a free to
2225 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2226 * that need the guarantee that every CPU has drained can disable the
2227 * optimizing racy check.
2229 static void __drain_all_pages(struct zone
*zone
, bool force_all_cpus
)
2234 * Allocate in the BSS so we won't require allocation in
2235 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2237 static cpumask_t cpus_with_pcps
;
2240 * Do not drain if one is already in progress unless it's specific to
2241 * a zone. Such callers are primarily CMA and memory hotplug and need
2242 * the drain to be complete when the call returns.
2244 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2247 mutex_lock(&pcpu_drain_mutex
);
2251 * We don't care about racing with CPU hotplug event
2252 * as offline notification will cause the notified
2253 * cpu to drain that CPU pcps and on_each_cpu_mask
2254 * disables preemption as part of its processing
2256 for_each_online_cpu(cpu
) {
2257 struct per_cpu_pages
*pcp
;
2259 bool has_pcps
= false;
2261 if (force_all_cpus
) {
2263 * The pcp.count check is racy, some callers need a
2264 * guarantee that no cpu is missed.
2268 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
2272 for_each_populated_zone(z
) {
2273 pcp
= per_cpu_ptr(z
->per_cpu_pageset
, cpu
);
2282 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2284 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2287 for_each_cpu(cpu
, &cpus_with_pcps
) {
2289 drain_pages_zone(cpu
, zone
);
2294 mutex_unlock(&pcpu_drain_mutex
);
2298 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2300 * When zone parameter is non-NULL, spill just the single zone's pages.
2302 void drain_all_pages(struct zone
*zone
)
2304 __drain_all_pages(zone
, false);
2307 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
,
2312 if (!free_pages_prepare(page
, order
, FPI_NONE
))
2315 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2316 set_pcppage_migratetype(page
, migratetype
);
2320 static int nr_pcp_free(struct per_cpu_pages
*pcp
, int high
, bool free_high
)
2322 int min_nr_free
, max_nr_free
;
2323 int batch
= READ_ONCE(pcp
->batch
);
2325 /* Free everything if batch freeing high-order pages. */
2326 if (unlikely(free_high
))
2329 /* Check for PCP disabled or boot pageset */
2330 if (unlikely(high
< batch
))
2333 /* Leave at least pcp->batch pages on the list */
2334 min_nr_free
= batch
;
2335 max_nr_free
= high
- batch
;
2338 * Double the number of pages freed each time there is subsequent
2339 * freeing of pages without any allocation.
2341 batch
<<= pcp
->free_factor
;
2342 if (batch
< max_nr_free
)
2344 batch
= clamp(batch
, min_nr_free
, max_nr_free
);
2349 static int nr_pcp_high(struct per_cpu_pages
*pcp
, struct zone
*zone
,
2352 int high
= READ_ONCE(pcp
->high
);
2354 if (unlikely(!high
|| free_high
))
2357 if (!test_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
))
2361 * If reclaim is active, limit the number of pages that can be
2362 * stored on pcp lists
2364 return min(READ_ONCE(pcp
->batch
) << 2, high
);
2367 static void free_unref_page_commit(struct zone
*zone
, struct per_cpu_pages
*pcp
,
2368 struct page
*page
, int migratetype
,
2375 __count_vm_events(PGFREE
, 1 << order
);
2376 pindex
= order_to_pindex(migratetype
, order
);
2377 list_add(&page
->pcp_list
, &pcp
->lists
[pindex
]);
2378 pcp
->count
+= 1 << order
;
2381 * As high-order pages other than THP's stored on PCP can contribute
2382 * to fragmentation, limit the number stored when PCP is heavily
2383 * freeing without allocation. The remainder after bulk freeing
2384 * stops will be drained from vmstat refresh context.
2386 free_high
= (pcp
->free_factor
&& order
&& order
<= PAGE_ALLOC_COSTLY_ORDER
);
2388 high
= nr_pcp_high(pcp
, zone
, free_high
);
2389 if (pcp
->count
>= high
) {
2390 free_pcppages_bulk(zone
, nr_pcp_free(pcp
, high
, free_high
), pcp
, pindex
);
2397 void free_unref_page(struct page
*page
, unsigned int order
)
2399 unsigned long __maybe_unused UP_flags
;
2400 struct per_cpu_pages
*pcp
;
2402 unsigned long pfn
= page_to_pfn(page
);
2405 if (!free_unref_page_prepare(page
, pfn
, order
))
2409 * We only track unmovable, reclaimable and movable on pcp lists.
2410 * Place ISOLATE pages on the isolated list because they are being
2411 * offlined but treat HIGHATOMIC as movable pages so we can get those
2412 * areas back if necessary. Otherwise, we may have to free
2413 * excessively into the page allocator
2415 migratetype
= get_pcppage_migratetype(page
);
2416 if (unlikely(migratetype
>= MIGRATE_PCPTYPES
)) {
2417 if (unlikely(is_migrate_isolate(migratetype
))) {
2418 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
, FPI_NONE
);
2421 migratetype
= MIGRATE_MOVABLE
;
2424 zone
= page_zone(page
);
2425 pcp_trylock_prepare(UP_flags
);
2426 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
2428 free_unref_page_commit(zone
, pcp
, page
, migratetype
, order
);
2429 pcp_spin_unlock(pcp
);
2431 free_one_page(zone
, page
, pfn
, order
, migratetype
, FPI_NONE
);
2433 pcp_trylock_finish(UP_flags
);
2437 * Free a list of 0-order pages
2439 void free_unref_page_list(struct list_head
*list
)
2441 unsigned long __maybe_unused UP_flags
;
2442 struct page
*page
, *next
;
2443 struct per_cpu_pages
*pcp
= NULL
;
2444 struct zone
*locked_zone
= NULL
;
2445 int batch_count
= 0;
2448 /* Prepare pages for freeing */
2449 list_for_each_entry_safe(page
, next
, list
, lru
) {
2450 unsigned long pfn
= page_to_pfn(page
);
2451 if (!free_unref_page_prepare(page
, pfn
, 0)) {
2452 list_del(&page
->lru
);
2457 * Free isolated pages directly to the allocator, see
2458 * comment in free_unref_page.
2460 migratetype
= get_pcppage_migratetype(page
);
2461 if (unlikely(is_migrate_isolate(migratetype
))) {
2462 list_del(&page
->lru
);
2463 free_one_page(page_zone(page
), page
, pfn
, 0, migratetype
, FPI_NONE
);
2468 list_for_each_entry_safe(page
, next
, list
, lru
) {
2469 struct zone
*zone
= page_zone(page
);
2471 list_del(&page
->lru
);
2472 migratetype
= get_pcppage_migratetype(page
);
2475 * Either different zone requiring a different pcp lock or
2476 * excessive lock hold times when freeing a large list of
2479 if (zone
!= locked_zone
|| batch_count
== SWAP_CLUSTER_MAX
) {
2481 pcp_spin_unlock(pcp
);
2482 pcp_trylock_finish(UP_flags
);
2488 * trylock is necessary as pages may be getting freed
2489 * from IRQ or SoftIRQ context after an IO completion.
2491 pcp_trylock_prepare(UP_flags
);
2492 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
2493 if (unlikely(!pcp
)) {
2494 pcp_trylock_finish(UP_flags
);
2495 free_one_page(zone
, page
, page_to_pfn(page
),
2496 0, migratetype
, FPI_NONE
);
2504 * Non-isolated types over MIGRATE_PCPTYPES get added
2505 * to the MIGRATE_MOVABLE pcp list.
2507 if (unlikely(migratetype
>= MIGRATE_PCPTYPES
))
2508 migratetype
= MIGRATE_MOVABLE
;
2510 trace_mm_page_free_batched(page
);
2511 free_unref_page_commit(zone
, pcp
, page
, migratetype
, 0);
2516 pcp_spin_unlock(pcp
);
2517 pcp_trylock_finish(UP_flags
);
2522 * split_page takes a non-compound higher-order page, and splits it into
2523 * n (1<<order) sub-pages: page[0..n]
2524 * Each sub-page must be freed individually.
2526 * Note: this is probably too low level an operation for use in drivers.
2527 * Please consult with lkml before using this in your driver.
2529 void split_page(struct page
*page
, unsigned int order
)
2533 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2534 VM_BUG_ON_PAGE(!page_count(page
), page
);
2536 for (i
= 1; i
< (1 << order
); i
++)
2537 set_page_refcounted(page
+ i
);
2538 split_page_owner(page
, 1 << order
);
2539 split_page_memcg(page
, 1 << order
);
2541 EXPORT_SYMBOL_GPL(split_page
);
2543 int __isolate_free_page(struct page
*page
, unsigned int order
)
2545 struct zone
*zone
= page_zone(page
);
2546 int mt
= get_pageblock_migratetype(page
);
2548 if (!is_migrate_isolate(mt
)) {
2549 unsigned long watermark
;
2551 * Obey watermarks as if the page was being allocated. We can
2552 * emulate a high-order watermark check with a raised order-0
2553 * watermark, because we already know our high-order page
2556 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
2557 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2560 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2563 del_page_from_free_list(page
, zone
, order
);
2566 * Set the pageblock if the isolated page is at least half of a
2569 if (order
>= pageblock_order
- 1) {
2570 struct page
*endpage
= page
+ (1 << order
) - 1;
2571 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2572 int mt
= get_pageblock_migratetype(page
);
2574 * Only change normal pageblocks (i.e., they can merge
2577 if (migratetype_is_mergeable(mt
))
2578 set_pageblock_migratetype(page
,
2583 return 1UL << order
;
2587 * __putback_isolated_page - Return a now-isolated page back where we got it
2588 * @page: Page that was isolated
2589 * @order: Order of the isolated page
2590 * @mt: The page's pageblock's migratetype
2592 * This function is meant to return a page pulled from the free lists via
2593 * __isolate_free_page back to the free lists they were pulled from.
2595 void __putback_isolated_page(struct page
*page
, unsigned int order
, int mt
)
2597 struct zone
*zone
= page_zone(page
);
2599 /* zone lock should be held when this function is called */
2600 lockdep_assert_held(&zone
->lock
);
2602 /* Return isolated page to tail of freelist. */
2603 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
,
2604 FPI_SKIP_REPORT_NOTIFY
| FPI_TO_TAIL
);
2608 * Update NUMA hit/miss statistics
2610 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2614 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2616 /* skip numa counters update if numa stats is disabled */
2617 if (!static_branch_likely(&vm_numa_stat_key
))
2620 if (zone_to_nid(z
) != numa_node_id())
2621 local_stat
= NUMA_OTHER
;
2623 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
2624 __count_numa_events(z
, NUMA_HIT
, nr_account
);
2626 __count_numa_events(z
, NUMA_MISS
, nr_account
);
2627 __count_numa_events(preferred_zone
, NUMA_FOREIGN
, nr_account
);
2629 __count_numa_events(z
, local_stat
, nr_account
);
2633 static __always_inline
2634 struct page
*rmqueue_buddy(struct zone
*preferred_zone
, struct zone
*zone
,
2635 unsigned int order
, unsigned int alloc_flags
,
2639 unsigned long flags
;
2643 spin_lock_irqsave(&zone
->lock
, flags
);
2644 if (alloc_flags
& ALLOC_HIGHATOMIC
)
2645 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2647 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
2650 * If the allocation fails, allow OOM handling access
2651 * to HIGHATOMIC reserves as failing now is worse than
2652 * failing a high-order atomic allocation in the
2655 if (!page
&& (alloc_flags
& ALLOC_OOM
))
2656 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2659 spin_unlock_irqrestore(&zone
->lock
, flags
);
2663 __mod_zone_freepage_state(zone
, -(1 << order
),
2664 get_pcppage_migratetype(page
));
2665 spin_unlock_irqrestore(&zone
->lock
, flags
);
2666 } while (check_new_pages(page
, order
));
2668 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2669 zone_statistics(preferred_zone
, zone
, 1);
2674 /* Remove page from the per-cpu list, caller must protect the list */
2676 struct page
*__rmqueue_pcplist(struct zone
*zone
, unsigned int order
,
2678 unsigned int alloc_flags
,
2679 struct per_cpu_pages
*pcp
,
2680 struct list_head
*list
)
2685 if (list_empty(list
)) {
2686 int batch
= READ_ONCE(pcp
->batch
);
2690 * Scale batch relative to order if batch implies
2691 * free pages can be stored on the PCP. Batch can
2692 * be 1 for small zones or for boot pagesets which
2693 * should never store free pages as the pages may
2694 * belong to arbitrary zones.
2697 batch
= max(batch
>> order
, 2);
2698 alloced
= rmqueue_bulk(zone
, order
,
2700 migratetype
, alloc_flags
);
2702 pcp
->count
+= alloced
<< order
;
2703 if (unlikely(list_empty(list
)))
2707 page
= list_first_entry(list
, struct page
, pcp_list
);
2708 list_del(&page
->pcp_list
);
2709 pcp
->count
-= 1 << order
;
2710 } while (check_new_pages(page
, order
));
2715 /* Lock and remove page from the per-cpu list */
2716 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2717 struct zone
*zone
, unsigned int order
,
2718 int migratetype
, unsigned int alloc_flags
)
2720 struct per_cpu_pages
*pcp
;
2721 struct list_head
*list
;
2723 unsigned long __maybe_unused UP_flags
;
2725 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2726 pcp_trylock_prepare(UP_flags
);
2727 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
2729 pcp_trylock_finish(UP_flags
);
2734 * On allocation, reduce the number of pages that are batch freed.
2735 * See nr_pcp_free() where free_factor is increased for subsequent
2738 pcp
->free_factor
>>= 1;
2739 list
= &pcp
->lists
[order_to_pindex(migratetype
, order
)];
2740 page
= __rmqueue_pcplist(zone
, order
, migratetype
, alloc_flags
, pcp
, list
);
2741 pcp_spin_unlock(pcp
);
2742 pcp_trylock_finish(UP_flags
);
2744 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2745 zone_statistics(preferred_zone
, zone
, 1);
2751 * Allocate a page from the given zone.
2752 * Use pcplists for THP or "cheap" high-order allocations.
2756 * Do not instrument rmqueue() with KMSAN. This function may call
2757 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2758 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2759 * may call rmqueue() again, which will result in a deadlock.
2761 __no_sanitize_memory
2763 struct page
*rmqueue(struct zone
*preferred_zone
,
2764 struct zone
*zone
, unsigned int order
,
2765 gfp_t gfp_flags
, unsigned int alloc_flags
,
2771 * We most definitely don't want callers attempting to
2772 * allocate greater than order-1 page units with __GFP_NOFAIL.
2774 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2776 if (likely(pcp_allowed_order(order
))) {
2777 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
2778 migratetype
, alloc_flags
);
2783 page
= rmqueue_buddy(preferred_zone
, zone
, order
, alloc_flags
,
2787 /* Separate test+clear to avoid unnecessary atomics */
2788 if ((alloc_flags
& ALLOC_KSWAPD
) &&
2789 unlikely(test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
))) {
2790 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2791 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
2794 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
2798 noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2800 return __should_fail_alloc_page(gfp_mask
, order
);
2802 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
2804 static inline long __zone_watermark_unusable_free(struct zone
*z
,
2805 unsigned int order
, unsigned int alloc_flags
)
2807 long unusable_free
= (1 << order
) - 1;
2810 * If the caller does not have rights to reserves below the min
2811 * watermark then subtract the high-atomic reserves. This will
2812 * over-estimate the size of the atomic reserve but it avoids a search.
2814 if (likely(!(alloc_flags
& ALLOC_RESERVES
)))
2815 unusable_free
+= z
->nr_reserved_highatomic
;
2818 /* If allocation can't use CMA areas don't use free CMA pages */
2819 if (!(alloc_flags
& ALLOC_CMA
))
2820 unusable_free
+= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2822 #ifdef CONFIG_UNACCEPTED_MEMORY
2823 unusable_free
+= zone_page_state(z
, NR_UNACCEPTED
);
2826 return unusable_free
;
2830 * Return true if free base pages are above 'mark'. For high-order checks it
2831 * will return true of the order-0 watermark is reached and there is at least
2832 * one free page of a suitable size. Checking now avoids taking the zone lock
2833 * to check in the allocation paths if no pages are free.
2835 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2836 int highest_zoneidx
, unsigned int alloc_flags
,
2842 /* free_pages may go negative - that's OK */
2843 free_pages
-= __zone_watermark_unusable_free(z
, order
, alloc_flags
);
2845 if (unlikely(alloc_flags
& ALLOC_RESERVES
)) {
2847 * __GFP_HIGH allows access to 50% of the min reserve as well
2850 if (alloc_flags
& ALLOC_MIN_RESERVE
) {
2854 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2855 * access more reserves than just __GFP_HIGH. Other
2856 * non-blocking allocations requests such as GFP_NOWAIT
2857 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2858 * access to the min reserve.
2860 if (alloc_flags
& ALLOC_NON_BLOCK
)
2865 * OOM victims can try even harder than the normal reserve
2866 * users on the grounds that it's definitely going to be in
2867 * the exit path shortly and free memory. Any allocation it
2868 * makes during the free path will be small and short-lived.
2870 if (alloc_flags
& ALLOC_OOM
)
2875 * Check watermarks for an order-0 allocation request. If these
2876 * are not met, then a high-order request also cannot go ahead
2877 * even if a suitable page happened to be free.
2879 if (free_pages
<= min
+ z
->lowmem_reserve
[highest_zoneidx
])
2882 /* If this is an order-0 request then the watermark is fine */
2886 /* For a high-order request, check at least one suitable page is free */
2887 for (o
= order
; o
<= MAX_ORDER
; o
++) {
2888 struct free_area
*area
= &z
->free_area
[o
];
2894 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2895 if (!free_area_empty(area
, mt
))
2900 if ((alloc_flags
& ALLOC_CMA
) &&
2901 !free_area_empty(area
, MIGRATE_CMA
)) {
2905 if ((alloc_flags
& (ALLOC_HIGHATOMIC
|ALLOC_OOM
)) &&
2906 !free_area_empty(area
, MIGRATE_HIGHATOMIC
)) {
2913 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2914 int highest_zoneidx
, unsigned int alloc_flags
)
2916 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
2917 zone_page_state(z
, NR_FREE_PAGES
));
2920 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2921 unsigned long mark
, int highest_zoneidx
,
2922 unsigned int alloc_flags
, gfp_t gfp_mask
)
2926 free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2929 * Fast check for order-0 only. If this fails then the reserves
2930 * need to be calculated.
2936 usable_free
= free_pages
;
2937 reserved
= __zone_watermark_unusable_free(z
, 0, alloc_flags
);
2939 /* reserved may over estimate high-atomic reserves. */
2940 usable_free
-= min(usable_free
, reserved
);
2941 if (usable_free
> mark
+ z
->lowmem_reserve
[highest_zoneidx
])
2945 if (__zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
2950 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
2951 * when checking the min watermark. The min watermark is the
2952 * point where boosting is ignored so that kswapd is woken up
2953 * when below the low watermark.
2955 if (unlikely(!order
&& (alloc_flags
& ALLOC_MIN_RESERVE
) && z
->watermark_boost
2956 && ((alloc_flags
& ALLOC_WMARK_MASK
) == WMARK_MIN
))) {
2957 mark
= z
->_watermark
[WMARK_MIN
];
2958 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
,
2959 alloc_flags
, free_pages
);
2965 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2966 unsigned long mark
, int highest_zoneidx
)
2968 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2970 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2971 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2973 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, 0,
2978 int __read_mostly node_reclaim_distance
= RECLAIM_DISTANCE
;
2980 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2982 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
2983 node_reclaim_distance
;
2985 #else /* CONFIG_NUMA */
2986 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2990 #endif /* CONFIG_NUMA */
2993 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
2994 * fragmentation is subtle. If the preferred zone was HIGHMEM then
2995 * premature use of a lower zone may cause lowmem pressure problems that
2996 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
2997 * probably too small. It only makes sense to spread allocations to avoid
2998 * fragmentation between the Normal and DMA32 zones.
3000 static inline unsigned int
3001 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3003 unsigned int alloc_flags
;
3006 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3009 alloc_flags
= (__force
int) (gfp_mask
& __GFP_KSWAPD_RECLAIM
);
3011 #ifdef CONFIG_ZONE_DMA32
3015 if (zone_idx(zone
) != ZONE_NORMAL
)
3019 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3020 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3021 * on UMA that if Normal is populated then so is DMA32.
3023 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3024 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3027 alloc_flags
|= ALLOC_NOFRAGMENT
;
3028 #endif /* CONFIG_ZONE_DMA32 */
3032 /* Must be called after current_gfp_context() which can change gfp_mask */
3033 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask
,
3034 unsigned int alloc_flags
)
3037 if (gfp_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3038 alloc_flags
|= ALLOC_CMA
;
3044 * get_page_from_freelist goes through the zonelist trying to allocate
3047 static struct page
*
3048 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3049 const struct alloc_context
*ac
)
3053 struct pglist_data
*last_pgdat
= NULL
;
3054 bool last_pgdat_dirty_ok
= false;
3059 * Scan zonelist, looking for a zone with enough free.
3060 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3062 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3063 z
= ac
->preferred_zoneref
;
3064 for_next_zone_zonelist_nodemask(zone
, z
, ac
->highest_zoneidx
,
3069 if (cpusets_enabled() &&
3070 (alloc_flags
& ALLOC_CPUSET
) &&
3071 !__cpuset_zone_allowed(zone
, gfp_mask
))
3074 * When allocating a page cache page for writing, we
3075 * want to get it from a node that is within its dirty
3076 * limit, such that no single node holds more than its
3077 * proportional share of globally allowed dirty pages.
3078 * The dirty limits take into account the node's
3079 * lowmem reserves and high watermark so that kswapd
3080 * should be able to balance it without having to
3081 * write pages from its LRU list.
3083 * XXX: For now, allow allocations to potentially
3084 * exceed the per-node dirty limit in the slowpath
3085 * (spread_dirty_pages unset) before going into reclaim,
3086 * which is important when on a NUMA setup the allowed
3087 * nodes are together not big enough to reach the
3088 * global limit. The proper fix for these situations
3089 * will require awareness of nodes in the
3090 * dirty-throttling and the flusher threads.
3092 if (ac
->spread_dirty_pages
) {
3093 if (last_pgdat
!= zone
->zone_pgdat
) {
3094 last_pgdat
= zone
->zone_pgdat
;
3095 last_pgdat_dirty_ok
= node_dirty_ok(zone
->zone_pgdat
);
3098 if (!last_pgdat_dirty_ok
)
3102 if (no_fallback
&& nr_online_nodes
> 1 &&
3103 zone
!= ac
->preferred_zoneref
->zone
) {
3107 * If moving to a remote node, retry but allow
3108 * fragmenting fallbacks. Locality is more important
3109 * than fragmentation avoidance.
3111 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3112 if (zone_to_nid(zone
) != local_nid
) {
3113 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3118 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3119 if (!zone_watermark_fast(zone
, order
, mark
,
3120 ac
->highest_zoneidx
, alloc_flags
,
3124 if (has_unaccepted_memory()) {
3125 if (try_to_accept_memory(zone
, order
))
3129 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3131 * Watermark failed for this zone, but see if we can
3132 * grow this zone if it contains deferred pages.
3134 if (deferred_pages_enabled()) {
3135 if (_deferred_grow_zone(zone
, order
))
3139 /* Checked here to keep the fast path fast */
3140 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3141 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3144 if (!node_reclaim_enabled() ||
3145 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3148 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3150 case NODE_RECLAIM_NOSCAN
:
3153 case NODE_RECLAIM_FULL
:
3154 /* scanned but unreclaimable */
3157 /* did we reclaim enough */
3158 if (zone_watermark_ok(zone
, order
, mark
,
3159 ac
->highest_zoneidx
, alloc_flags
))
3167 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3168 gfp_mask
, alloc_flags
, ac
->migratetype
);
3170 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3173 * If this is a high-order atomic allocation then check
3174 * if the pageblock should be reserved for the future
3176 if (unlikely(alloc_flags
& ALLOC_HIGHATOMIC
))
3177 reserve_highatomic_pageblock(page
, zone
);
3181 if (has_unaccepted_memory()) {
3182 if (try_to_accept_memory(zone
, order
))
3186 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3187 /* Try again if zone has deferred pages */
3188 if (deferred_pages_enabled()) {
3189 if (_deferred_grow_zone(zone
, order
))
3197 * It's possible on a UMA machine to get through all zones that are
3198 * fragmented. If avoiding fragmentation, reset and try again.
3201 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3208 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3210 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3213 * This documents exceptions given to allocations in certain
3214 * contexts that are allowed to allocate outside current's set
3217 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3218 if (tsk_is_oom_victim(current
) ||
3219 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3220 filter
&= ~SHOW_MEM_FILTER_NODES
;
3221 if (!in_task() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3222 filter
&= ~SHOW_MEM_FILTER_NODES
;
3224 __show_mem(filter
, nodemask
, gfp_zone(gfp_mask
));
3227 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3229 struct va_format vaf
;
3231 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
3233 if ((gfp_mask
& __GFP_NOWARN
) ||
3234 !__ratelimit(&nopage_rs
) ||
3235 ((gfp_mask
& __GFP_DMA
) && !has_managed_dma()))
3238 va_start(args
, fmt
);
3241 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3242 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3243 nodemask_pr_args(nodemask
));
3246 cpuset_print_current_mems_allowed();
3249 warn_alloc_show_mem(gfp_mask
, nodemask
);
3252 static inline struct page
*
3253 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3254 unsigned int alloc_flags
,
3255 const struct alloc_context
*ac
)
3259 page
= get_page_from_freelist(gfp_mask
, order
,
3260 alloc_flags
|ALLOC_CPUSET
, ac
);
3262 * fallback to ignore cpuset restriction if our nodes
3266 page
= get_page_from_freelist(gfp_mask
, order
,
3272 static inline struct page
*
3273 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3274 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3276 struct oom_control oc
= {
3277 .zonelist
= ac
->zonelist
,
3278 .nodemask
= ac
->nodemask
,
3280 .gfp_mask
= gfp_mask
,
3285 *did_some_progress
= 0;
3288 * Acquire the oom lock. If that fails, somebody else is
3289 * making progress for us.
3291 if (!mutex_trylock(&oom_lock
)) {
3292 *did_some_progress
= 1;
3293 schedule_timeout_uninterruptible(1);
3298 * Go through the zonelist yet one more time, keep very high watermark
3299 * here, this is only to catch a parallel oom killing, we must fail if
3300 * we're still under heavy pressure. But make sure that this reclaim
3301 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3302 * allocation which will never fail due to oom_lock already held.
3304 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3305 ~__GFP_DIRECT_RECLAIM
, order
,
3306 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3310 /* Coredumps can quickly deplete all memory reserves */
3311 if (current
->flags
& PF_DUMPCORE
)
3313 /* The OOM killer will not help higher order allocs */
3314 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3317 * We have already exhausted all our reclaim opportunities without any
3318 * success so it is time to admit defeat. We will skip the OOM killer
3319 * because it is very likely that the caller has a more reasonable
3320 * fallback than shooting a random task.
3322 * The OOM killer may not free memory on a specific node.
3324 if (gfp_mask
& (__GFP_RETRY_MAYFAIL
| __GFP_THISNODE
))
3326 /* The OOM killer does not needlessly kill tasks for lowmem */
3327 if (ac
->highest_zoneidx
< ZONE_NORMAL
)
3329 if (pm_suspended_storage())
3332 * XXX: GFP_NOFS allocations should rather fail than rely on
3333 * other request to make a forward progress.
3334 * We are in an unfortunate situation where out_of_memory cannot
3335 * do much for this context but let's try it to at least get
3336 * access to memory reserved if the current task is killed (see
3337 * out_of_memory). Once filesystems are ready to handle allocation
3338 * failures more gracefully we should just bail out here.
3341 /* Exhausted what can be done so it's blame time */
3342 if (out_of_memory(&oc
) ||
3343 WARN_ON_ONCE_GFP(gfp_mask
& __GFP_NOFAIL
, gfp_mask
)) {
3344 *did_some_progress
= 1;
3347 * Help non-failing allocations by giving them access to memory
3350 if (gfp_mask
& __GFP_NOFAIL
)
3351 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3352 ALLOC_NO_WATERMARKS
, ac
);
3355 mutex_unlock(&oom_lock
);
3360 * Maximum number of compaction retries with a progress before OOM
3361 * killer is consider as the only way to move forward.
3363 #define MAX_COMPACT_RETRIES 16
3365 #ifdef CONFIG_COMPACTION
3366 /* Try memory compaction for high-order allocations before reclaim */
3367 static struct page
*
3368 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3369 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3370 enum compact_priority prio
, enum compact_result
*compact_result
)
3372 struct page
*page
= NULL
;
3373 unsigned long pflags
;
3374 unsigned int noreclaim_flag
;
3379 psi_memstall_enter(&pflags
);
3380 delayacct_compact_start();
3381 noreclaim_flag
= memalloc_noreclaim_save();
3383 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3386 memalloc_noreclaim_restore(noreclaim_flag
);
3387 psi_memstall_leave(&pflags
);
3388 delayacct_compact_end();
3390 if (*compact_result
== COMPACT_SKIPPED
)
3393 * At least in one zone compaction wasn't deferred or skipped, so let's
3394 * count a compaction stall
3396 count_vm_event(COMPACTSTALL
);
3398 /* Prep a captured page if available */
3400 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3402 /* Try get a page from the freelist if available */
3404 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3407 struct zone
*zone
= page_zone(page
);
3409 zone
->compact_blockskip_flush
= false;
3410 compaction_defer_reset(zone
, order
, true);
3411 count_vm_event(COMPACTSUCCESS
);
3416 * It's bad if compaction run occurs and fails. The most likely reason
3417 * is that pages exist, but not enough to satisfy watermarks.
3419 count_vm_event(COMPACTFAIL
);
3427 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3428 enum compact_result compact_result
,
3429 enum compact_priority
*compact_priority
,
3430 int *compaction_retries
)
3432 int max_retries
= MAX_COMPACT_RETRIES
;
3435 int retries
= *compaction_retries
;
3436 enum compact_priority priority
= *compact_priority
;
3441 if (fatal_signal_pending(current
))
3445 * Compaction was skipped due to a lack of free order-0
3446 * migration targets. Continue if reclaim can help.
3448 if (compact_result
== COMPACT_SKIPPED
) {
3449 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3454 * Compaction managed to coalesce some page blocks, but the
3455 * allocation failed presumably due to a race. Retry some.
3457 if (compact_result
== COMPACT_SUCCESS
) {
3459 * !costly requests are much more important than
3460 * __GFP_RETRY_MAYFAIL costly ones because they are de
3461 * facto nofail and invoke OOM killer to move on while
3462 * costly can fail and users are ready to cope with
3463 * that. 1/4 retries is rather arbitrary but we would
3464 * need much more detailed feedback from compaction to
3465 * make a better decision.
3467 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3470 if (++(*compaction_retries
) <= max_retries
) {
3477 * Compaction failed. Retry with increasing priority.
3479 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3480 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3482 if (*compact_priority
> min_priority
) {
3483 (*compact_priority
)--;
3484 *compaction_retries
= 0;
3488 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3492 static inline struct page
*
3493 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3494 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3495 enum compact_priority prio
, enum compact_result
*compact_result
)
3497 *compact_result
= COMPACT_SKIPPED
;
3502 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3503 enum compact_result compact_result
,
3504 enum compact_priority
*compact_priority
,
3505 int *compaction_retries
)
3510 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3514 * There are setups with compaction disabled which would prefer to loop
3515 * inside the allocator rather than hit the oom killer prematurely.
3516 * Let's give them a good hope and keep retrying while the order-0
3517 * watermarks are OK.
3519 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3520 ac
->highest_zoneidx
, ac
->nodemask
) {
3521 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3522 ac
->highest_zoneidx
, alloc_flags
))
3527 #endif /* CONFIG_COMPACTION */
3529 #ifdef CONFIG_LOCKDEP
3530 static struct lockdep_map __fs_reclaim_map
=
3531 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3533 static bool __need_reclaim(gfp_t gfp_mask
)
3535 /* no reclaim without waiting on it */
3536 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3539 /* this guy won't enter reclaim */
3540 if (current
->flags
& PF_MEMALLOC
)
3543 if (gfp_mask
& __GFP_NOLOCKDEP
)
3549 void __fs_reclaim_acquire(unsigned long ip
)
3551 lock_acquire_exclusive(&__fs_reclaim_map
, 0, 0, NULL
, ip
);
3554 void __fs_reclaim_release(unsigned long ip
)
3556 lock_release(&__fs_reclaim_map
, ip
);
3559 void fs_reclaim_acquire(gfp_t gfp_mask
)
3561 gfp_mask
= current_gfp_context(gfp_mask
);
3563 if (__need_reclaim(gfp_mask
)) {
3564 if (gfp_mask
& __GFP_FS
)
3565 __fs_reclaim_acquire(_RET_IP_
);
3567 #ifdef CONFIG_MMU_NOTIFIER
3568 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
3569 lock_map_release(&__mmu_notifier_invalidate_range_start_map
);
3574 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3576 void fs_reclaim_release(gfp_t gfp_mask
)
3578 gfp_mask
= current_gfp_context(gfp_mask
);
3580 if (__need_reclaim(gfp_mask
)) {
3581 if (gfp_mask
& __GFP_FS
)
3582 __fs_reclaim_release(_RET_IP_
);
3585 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3589 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3590 * have been rebuilt so allocation retries. Reader side does not lock and
3591 * retries the allocation if zonelist changes. Writer side is protected by the
3592 * embedded spin_lock.
3594 static DEFINE_SEQLOCK(zonelist_update_seq
);
3596 static unsigned int zonelist_iter_begin(void)
3598 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE
))
3599 return read_seqbegin(&zonelist_update_seq
);
3604 static unsigned int check_retry_zonelist(unsigned int seq
)
3606 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE
))
3607 return read_seqretry(&zonelist_update_seq
, seq
);
3612 /* Perform direct synchronous page reclaim */
3613 static unsigned long
3614 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3615 const struct alloc_context
*ac
)
3617 unsigned int noreclaim_flag
;
3618 unsigned long progress
;
3622 /* We now go into synchronous reclaim */
3623 cpuset_memory_pressure_bump();
3624 fs_reclaim_acquire(gfp_mask
);
3625 noreclaim_flag
= memalloc_noreclaim_save();
3627 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3630 memalloc_noreclaim_restore(noreclaim_flag
);
3631 fs_reclaim_release(gfp_mask
);
3638 /* The really slow allocator path where we enter direct reclaim */
3639 static inline struct page
*
3640 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3641 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3642 unsigned long *did_some_progress
)
3644 struct page
*page
= NULL
;
3645 unsigned long pflags
;
3646 bool drained
= false;
3648 psi_memstall_enter(&pflags
);
3649 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3650 if (unlikely(!(*did_some_progress
)))
3654 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3657 * If an allocation failed after direct reclaim, it could be because
3658 * pages are pinned on the per-cpu lists or in high alloc reserves.
3659 * Shrink them and try again
3661 if (!page
&& !drained
) {
3662 unreserve_highatomic_pageblock(ac
, false);
3663 drain_all_pages(NULL
);
3668 psi_memstall_leave(&pflags
);
3673 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
3674 const struct alloc_context
*ac
)
3678 pg_data_t
*last_pgdat
= NULL
;
3679 enum zone_type highest_zoneidx
= ac
->highest_zoneidx
;
3681 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, highest_zoneidx
,
3683 if (!managed_zone(zone
))
3685 if (last_pgdat
!= zone
->zone_pgdat
) {
3686 wakeup_kswapd(zone
, gfp_mask
, order
, highest_zoneidx
);
3687 last_pgdat
= zone
->zone_pgdat
;
3692 static inline unsigned int
3693 gfp_to_alloc_flags(gfp_t gfp_mask
, unsigned int order
)
3695 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3698 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3699 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3700 * to save two branches.
3702 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_MIN_RESERVE
);
3703 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM
!= (__force gfp_t
) ALLOC_KSWAPD
);
3706 * The caller may dip into page reserves a bit more if the caller
3707 * cannot run direct reclaim, or if the caller has realtime scheduling
3708 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3709 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3711 alloc_flags
|= (__force
int)
3712 (gfp_mask
& (__GFP_HIGH
| __GFP_KSWAPD_RECLAIM
));
3714 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
)) {
3716 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3717 * if it can't schedule.
3719 if (!(gfp_mask
& __GFP_NOMEMALLOC
)) {
3720 alloc_flags
|= ALLOC_NON_BLOCK
;
3723 alloc_flags
|= ALLOC_HIGHATOMIC
;
3727 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3728 * GFP_ATOMIC) rather than fail, see the comment for
3729 * cpuset_node_allowed().
3731 if (alloc_flags
& ALLOC_MIN_RESERVE
)
3732 alloc_flags
&= ~ALLOC_CPUSET
;
3733 } else if (unlikely(rt_task(current
)) && in_task())
3734 alloc_flags
|= ALLOC_MIN_RESERVE
;
3736 alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, alloc_flags
);
3741 static bool oom_reserves_allowed(struct task_struct
*tsk
)
3743 if (!tsk_is_oom_victim(tsk
))
3747 * !MMU doesn't have oom reaper so give access to memory reserves
3748 * only to the thread with TIF_MEMDIE set
3750 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
3757 * Distinguish requests which really need access to full memory
3758 * reserves from oom victims which can live with a portion of it
3760 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
3762 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3764 if (gfp_mask
& __GFP_MEMALLOC
)
3765 return ALLOC_NO_WATERMARKS
;
3766 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3767 return ALLOC_NO_WATERMARKS
;
3768 if (!in_interrupt()) {
3769 if (current
->flags
& PF_MEMALLOC
)
3770 return ALLOC_NO_WATERMARKS
;
3771 else if (oom_reserves_allowed(current
))
3778 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3780 return !!__gfp_pfmemalloc_flags(gfp_mask
);
3784 * Checks whether it makes sense to retry the reclaim to make a forward progress
3785 * for the given allocation request.
3787 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3788 * without success, or when we couldn't even meet the watermark if we
3789 * reclaimed all remaining pages on the LRU lists.
3791 * Returns true if a retry is viable or false to enter the oom path.
3794 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3795 struct alloc_context
*ac
, int alloc_flags
,
3796 bool did_some_progress
, int *no_progress_loops
)
3803 * Costly allocations might have made a progress but this doesn't mean
3804 * their order will become available due to high fragmentation so
3805 * always increment the no progress counter for them
3807 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3808 *no_progress_loops
= 0;
3810 (*no_progress_loops
)++;
3813 * Make sure we converge to OOM if we cannot make any progress
3814 * several times in the row.
3816 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3817 /* Before OOM, exhaust highatomic_reserve */
3818 return unreserve_highatomic_pageblock(ac
, true);
3822 * Keep reclaiming pages while there is a chance this will lead
3823 * somewhere. If none of the target zones can satisfy our allocation
3824 * request even if all reclaimable pages are considered then we are
3825 * screwed and have to go OOM.
3827 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3828 ac
->highest_zoneidx
, ac
->nodemask
) {
3829 unsigned long available
;
3830 unsigned long reclaimable
;
3831 unsigned long min_wmark
= min_wmark_pages(zone
);
3834 available
= reclaimable
= zone_reclaimable_pages(zone
);
3835 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3838 * Would the allocation succeed if we reclaimed all
3839 * reclaimable pages?
3841 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3842 ac
->highest_zoneidx
, alloc_flags
, available
);
3843 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3844 available
, min_wmark
, *no_progress_loops
, wmark
);
3852 * Memory allocation/reclaim might be called from a WQ context and the
3853 * current implementation of the WQ concurrency control doesn't
3854 * recognize that a particular WQ is congested if the worker thread is
3855 * looping without ever sleeping. Therefore we have to do a short sleep
3856 * here rather than calling cond_resched().
3858 if (current
->flags
& PF_WQ_WORKER
)
3859 schedule_timeout_uninterruptible(1);
3866 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
3869 * It's possible that cpuset's mems_allowed and the nodemask from
3870 * mempolicy don't intersect. This should be normally dealt with by
3871 * policy_nodemask(), but it's possible to race with cpuset update in
3872 * such a way the check therein was true, and then it became false
3873 * before we got our cpuset_mems_cookie here.
3874 * This assumes that for all allocations, ac->nodemask can come only
3875 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3876 * when it does not intersect with the cpuset restrictions) or the
3877 * caller can deal with a violated nodemask.
3879 if (cpusets_enabled() && ac
->nodemask
&&
3880 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
3881 ac
->nodemask
= NULL
;
3886 * When updating a task's mems_allowed or mempolicy nodemask, it is
3887 * possible to race with parallel threads in such a way that our
3888 * allocation can fail while the mask is being updated. If we are about
3889 * to fail, check if the cpuset changed during allocation and if so,
3892 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3898 static inline struct page
*
3899 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3900 struct alloc_context
*ac
)
3902 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3903 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
3904 struct page
*page
= NULL
;
3905 unsigned int alloc_flags
;
3906 unsigned long did_some_progress
;
3907 enum compact_priority compact_priority
;
3908 enum compact_result compact_result
;
3909 int compaction_retries
;
3910 int no_progress_loops
;
3911 unsigned int cpuset_mems_cookie
;
3912 unsigned int zonelist_iter_cookie
;
3916 compaction_retries
= 0;
3917 no_progress_loops
= 0;
3918 compact_priority
= DEF_COMPACT_PRIORITY
;
3919 cpuset_mems_cookie
= read_mems_allowed_begin();
3920 zonelist_iter_cookie
= zonelist_iter_begin();
3923 * The fast path uses conservative alloc_flags to succeed only until
3924 * kswapd needs to be woken up, and to avoid the cost of setting up
3925 * alloc_flags precisely. So we do that now.
3927 alloc_flags
= gfp_to_alloc_flags(gfp_mask
, order
);
3930 * We need to recalculate the starting point for the zonelist iterator
3931 * because we might have used different nodemask in the fast path, or
3932 * there was a cpuset modification and we are retrying - otherwise we
3933 * could end up iterating over non-eligible zones endlessly.
3935 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3936 ac
->highest_zoneidx
, ac
->nodemask
);
3937 if (!ac
->preferred_zoneref
->zone
)
3941 * Check for insane configurations where the cpuset doesn't contain
3942 * any suitable zone to satisfy the request - e.g. non-movable
3943 * GFP_HIGHUSER allocations from MOVABLE nodes only.
3945 if (cpusets_insane_config() && (gfp_mask
& __GFP_HARDWALL
)) {
3946 struct zoneref
*z
= first_zones_zonelist(ac
->zonelist
,
3947 ac
->highest_zoneidx
,
3948 &cpuset_current_mems_allowed
);
3953 if (alloc_flags
& ALLOC_KSWAPD
)
3954 wake_all_kswapds(order
, gfp_mask
, ac
);
3957 * The adjusted alloc_flags might result in immediate success, so try
3960 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3965 * For costly allocations, try direct compaction first, as it's likely
3966 * that we have enough base pages and don't need to reclaim. For non-
3967 * movable high-order allocations, do that as well, as compaction will
3968 * try prevent permanent fragmentation by migrating from blocks of the
3970 * Don't try this for allocations that are allowed to ignore
3971 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3973 if (can_direct_reclaim
&&
3975 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
3976 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
3977 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3979 INIT_COMPACT_PRIORITY
,
3985 * Checks for costly allocations with __GFP_NORETRY, which
3986 * includes some THP page fault allocations
3988 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
3990 * If allocating entire pageblock(s) and compaction
3991 * failed because all zones are below low watermarks
3992 * or is prohibited because it recently failed at this
3993 * order, fail immediately unless the allocator has
3994 * requested compaction and reclaim retry.
3997 * - potentially very expensive because zones are far
3998 * below their low watermarks or this is part of very
3999 * bursty high order allocations,
4000 * - not guaranteed to help because isolate_freepages()
4001 * may not iterate over freed pages as part of its
4003 * - unlikely to make entire pageblocks free on its
4006 if (compact_result
== COMPACT_SKIPPED
||
4007 compact_result
== COMPACT_DEFERRED
)
4011 * Looks like reclaim/compaction is worth trying, but
4012 * sync compaction could be very expensive, so keep
4013 * using async compaction.
4015 compact_priority
= INIT_COMPACT_PRIORITY
;
4020 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4021 if (alloc_flags
& ALLOC_KSWAPD
)
4022 wake_all_kswapds(order
, gfp_mask
, ac
);
4024 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4026 alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, reserve_flags
) |
4027 (alloc_flags
& ALLOC_KSWAPD
);
4030 * Reset the nodemask and zonelist iterators if memory policies can be
4031 * ignored. These allocations are high priority and system rather than
4034 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4035 ac
->nodemask
= NULL
;
4036 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4037 ac
->highest_zoneidx
, ac
->nodemask
);
4040 /* Attempt with potentially adjusted zonelist and alloc_flags */
4041 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4045 /* Caller is not willing to reclaim, we can't balance anything */
4046 if (!can_direct_reclaim
)
4049 /* Avoid recursion of direct reclaim */
4050 if (current
->flags
& PF_MEMALLOC
)
4053 /* Try direct reclaim and then allocating */
4054 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4055 &did_some_progress
);
4059 /* Try direct compaction and then allocating */
4060 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4061 compact_priority
, &compact_result
);
4065 /* Do not loop if specifically requested */
4066 if (gfp_mask
& __GFP_NORETRY
)
4070 * Do not retry costly high order allocations unless they are
4071 * __GFP_RETRY_MAYFAIL
4073 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4076 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4077 did_some_progress
> 0, &no_progress_loops
))
4081 * It doesn't make any sense to retry for the compaction if the order-0
4082 * reclaim is not able to make any progress because the current
4083 * implementation of the compaction depends on the sufficient amount
4084 * of free memory (see __compaction_suitable)
4086 if (did_some_progress
> 0 &&
4087 should_compact_retry(ac
, order
, alloc_flags
,
4088 compact_result
, &compact_priority
,
4089 &compaction_retries
))
4094 * Deal with possible cpuset update races or zonelist updates to avoid
4095 * a unnecessary OOM kill.
4097 if (check_retry_cpuset(cpuset_mems_cookie
, ac
) ||
4098 check_retry_zonelist(zonelist_iter_cookie
))
4101 /* Reclaim has failed us, start killing things */
4102 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4106 /* Avoid allocations with no watermarks from looping endlessly */
4107 if (tsk_is_oom_victim(current
) &&
4108 (alloc_flags
& ALLOC_OOM
||
4109 (gfp_mask
& __GFP_NOMEMALLOC
)))
4112 /* Retry as long as the OOM killer is making progress */
4113 if (did_some_progress
) {
4114 no_progress_loops
= 0;
4120 * Deal with possible cpuset update races or zonelist updates to avoid
4121 * a unnecessary OOM kill.
4123 if (check_retry_cpuset(cpuset_mems_cookie
, ac
) ||
4124 check_retry_zonelist(zonelist_iter_cookie
))
4128 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4131 if (gfp_mask
& __GFP_NOFAIL
) {
4133 * All existing users of the __GFP_NOFAIL are blockable, so warn
4134 * of any new users that actually require GFP_NOWAIT
4136 if (WARN_ON_ONCE_GFP(!can_direct_reclaim
, gfp_mask
))
4140 * PF_MEMALLOC request from this context is rather bizarre
4141 * because we cannot reclaim anything and only can loop waiting
4142 * for somebody to do a work for us
4144 WARN_ON_ONCE_GFP(current
->flags
& PF_MEMALLOC
, gfp_mask
);
4147 * non failing costly orders are a hard requirement which we
4148 * are not prepared for much so let's warn about these users
4149 * so that we can identify them and convert them to something
4152 WARN_ON_ONCE_GFP(costly_order
, gfp_mask
);
4155 * Help non-failing allocations by giving some access to memory
4156 * reserves normally used for high priority non-blocking
4157 * allocations but do not use ALLOC_NO_WATERMARKS because this
4158 * could deplete whole memory reserves which would just make
4159 * the situation worse.
4161 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_MIN_RESERVE
, ac
);
4169 warn_alloc(gfp_mask
, ac
->nodemask
,
4170 "page allocation failure: order:%u", order
);
4175 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4176 int preferred_nid
, nodemask_t
*nodemask
,
4177 struct alloc_context
*ac
, gfp_t
*alloc_gfp
,
4178 unsigned int *alloc_flags
)
4180 ac
->highest_zoneidx
= gfp_zone(gfp_mask
);
4181 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4182 ac
->nodemask
= nodemask
;
4183 ac
->migratetype
= gfp_migratetype(gfp_mask
);
4185 if (cpusets_enabled()) {
4186 *alloc_gfp
|= __GFP_HARDWALL
;
4188 * When we are in the interrupt context, it is irrelevant
4189 * to the current task context. It means that any node ok.
4191 if (in_task() && !ac
->nodemask
)
4192 ac
->nodemask
= &cpuset_current_mems_allowed
;
4194 *alloc_flags
|= ALLOC_CPUSET
;
4197 might_alloc(gfp_mask
);
4199 if (should_fail_alloc_page(gfp_mask
, order
))
4202 *alloc_flags
= gfp_to_alloc_flags_cma(gfp_mask
, *alloc_flags
);
4204 /* Dirty zone balancing only done in the fast path */
4205 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4208 * The preferred zone is used for statistics but crucially it is
4209 * also used as the starting point for the zonelist iterator. It
4210 * may get reset for allocations that ignore memory policies.
4212 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4213 ac
->highest_zoneidx
, ac
->nodemask
);
4219 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4220 * @gfp: GFP flags for the allocation
4221 * @preferred_nid: The preferred NUMA node ID to allocate from
4222 * @nodemask: Set of nodes to allocate from, may be NULL
4223 * @nr_pages: The number of pages desired on the list or array
4224 * @page_list: Optional list to store the allocated pages
4225 * @page_array: Optional array to store the pages
4227 * This is a batched version of the page allocator that attempts to
4228 * allocate nr_pages quickly. Pages are added to page_list if page_list
4229 * is not NULL, otherwise it is assumed that the page_array is valid.
4231 * For lists, nr_pages is the number of pages that should be allocated.
4233 * For arrays, only NULL elements are populated with pages and nr_pages
4234 * is the maximum number of pages that will be stored in the array.
4236 * Returns the number of pages on the list or array.
4238 unsigned long __alloc_pages_bulk(gfp_t gfp
, int preferred_nid
,
4239 nodemask_t
*nodemask
, int nr_pages
,
4240 struct list_head
*page_list
,
4241 struct page
**page_array
)
4244 unsigned long __maybe_unused UP_flags
;
4247 struct per_cpu_pages
*pcp
;
4248 struct list_head
*pcp_list
;
4249 struct alloc_context ac
;
4251 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4252 int nr_populated
= 0, nr_account
= 0;
4255 * Skip populated array elements to determine if any pages need
4256 * to be allocated before disabling IRQs.
4258 while (page_array
&& nr_populated
< nr_pages
&& page_array
[nr_populated
])
4261 /* No pages requested? */
4262 if (unlikely(nr_pages
<= 0))
4265 /* Already populated array? */
4266 if (unlikely(page_array
&& nr_pages
- nr_populated
== 0))
4269 /* Bulk allocator does not support memcg accounting. */
4270 if (memcg_kmem_online() && (gfp
& __GFP_ACCOUNT
))
4273 /* Use the single page allocator for one page. */
4274 if (nr_pages
- nr_populated
== 1)
4277 #ifdef CONFIG_PAGE_OWNER
4279 * PAGE_OWNER may recurse into the allocator to allocate space to
4280 * save the stack with pagesets.lock held. Releasing/reacquiring
4281 * removes much of the performance benefit of bulk allocation so
4282 * force the caller to allocate one page at a time as it'll have
4283 * similar performance to added complexity to the bulk allocator.
4285 if (static_branch_unlikely(&page_owner_inited
))
4289 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4290 gfp
&= gfp_allowed_mask
;
4292 if (!prepare_alloc_pages(gfp
, 0, preferred_nid
, nodemask
, &ac
, &alloc_gfp
, &alloc_flags
))
4296 /* Find an allowed local zone that meets the low watermark. */
4297 for_each_zone_zonelist_nodemask(zone
, z
, ac
.zonelist
, ac
.highest_zoneidx
, ac
.nodemask
) {
4300 if (cpusets_enabled() && (alloc_flags
& ALLOC_CPUSET
) &&
4301 !__cpuset_zone_allowed(zone
, gfp
)) {
4305 if (nr_online_nodes
> 1 && zone
!= ac
.preferred_zoneref
->zone
&&
4306 zone_to_nid(zone
) != zone_to_nid(ac
.preferred_zoneref
->zone
)) {
4310 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
) + nr_pages
;
4311 if (zone_watermark_fast(zone
, 0, mark
,
4312 zonelist_zone_idx(ac
.preferred_zoneref
),
4313 alloc_flags
, gfp
)) {
4319 * If there are no allowed local zones that meets the watermarks then
4320 * try to allocate a single page and reclaim if necessary.
4322 if (unlikely(!zone
))
4325 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4326 pcp_trylock_prepare(UP_flags
);
4327 pcp
= pcp_spin_trylock(zone
->per_cpu_pageset
);
4331 /* Attempt the batch allocation */
4332 pcp_list
= &pcp
->lists
[order_to_pindex(ac
.migratetype
, 0)];
4333 while (nr_populated
< nr_pages
) {
4335 /* Skip existing pages */
4336 if (page_array
&& page_array
[nr_populated
]) {
4341 page
= __rmqueue_pcplist(zone
, 0, ac
.migratetype
, alloc_flags
,
4343 if (unlikely(!page
)) {
4344 /* Try and allocate at least one page */
4346 pcp_spin_unlock(pcp
);
4353 prep_new_page(page
, 0, gfp
, 0);
4355 list_add(&page
->lru
, page_list
);
4357 page_array
[nr_populated
] = page
;
4361 pcp_spin_unlock(pcp
);
4362 pcp_trylock_finish(UP_flags
);
4364 __count_zid_vm_events(PGALLOC
, zone_idx(zone
), nr_account
);
4365 zone_statistics(ac
.preferred_zoneref
->zone
, zone
, nr_account
);
4368 return nr_populated
;
4371 pcp_trylock_finish(UP_flags
);
4374 page
= __alloc_pages(gfp
, 0, preferred_nid
, nodemask
);
4377 list_add(&page
->lru
, page_list
);
4379 page_array
[nr_populated
] = page
;
4385 EXPORT_SYMBOL_GPL(__alloc_pages_bulk
);
4388 * This is the 'heart' of the zoned buddy allocator.
4390 struct page
*__alloc_pages(gfp_t gfp
, unsigned int order
, int preferred_nid
,
4391 nodemask_t
*nodemask
)
4394 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4395 gfp_t alloc_gfp
; /* The gfp_t that was actually used for allocation */
4396 struct alloc_context ac
= { };
4399 * There are several places where we assume that the order value is sane
4400 * so bail out early if the request is out of bound.
4402 if (WARN_ON_ONCE_GFP(order
> MAX_ORDER
, gfp
))
4405 gfp
&= gfp_allowed_mask
;
4407 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4408 * resp. GFP_NOIO which has to be inherited for all allocation requests
4409 * from a particular context which has been marked by
4410 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4411 * movable zones are not used during allocation.
4413 gfp
= current_gfp_context(gfp
);
4415 if (!prepare_alloc_pages(gfp
, order
, preferred_nid
, nodemask
, &ac
,
4416 &alloc_gfp
, &alloc_flags
))
4420 * Forbid the first pass from falling back to types that fragment
4421 * memory until all local zones are considered.
4423 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp
);
4425 /* First allocation attempt */
4426 page
= get_page_from_freelist(alloc_gfp
, order
, alloc_flags
, &ac
);
4431 ac
.spread_dirty_pages
= false;
4434 * Restore the original nodemask if it was potentially replaced with
4435 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4437 ac
.nodemask
= nodemask
;
4439 page
= __alloc_pages_slowpath(alloc_gfp
, order
, &ac
);
4442 if (memcg_kmem_online() && (gfp
& __GFP_ACCOUNT
) && page
&&
4443 unlikely(__memcg_kmem_charge_page(page
, gfp
, order
) != 0)) {
4444 __free_pages(page
, order
);
4448 trace_mm_page_alloc(page
, order
, alloc_gfp
, ac
.migratetype
);
4449 kmsan_alloc_page(page
, order
, alloc_gfp
);
4453 EXPORT_SYMBOL(__alloc_pages
);
4455 struct folio
*__folio_alloc(gfp_t gfp
, unsigned int order
, int preferred_nid
,
4456 nodemask_t
*nodemask
)
4458 struct page
*page
= __alloc_pages(gfp
| __GFP_COMP
, order
,
4459 preferred_nid
, nodemask
);
4460 struct folio
*folio
= (struct folio
*)page
;
4462 if (folio
&& order
> 1)
4463 folio_prep_large_rmappable(folio
);
4466 EXPORT_SYMBOL(__folio_alloc
);
4469 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4470 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4471 * you need to access high mem.
4473 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4477 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
4480 return (unsigned long) page_address(page
);
4482 EXPORT_SYMBOL(__get_free_pages
);
4484 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4486 return __get_free_page(gfp_mask
| __GFP_ZERO
);
4488 EXPORT_SYMBOL(get_zeroed_page
);
4491 * __free_pages - Free pages allocated with alloc_pages().
4492 * @page: The page pointer returned from alloc_pages().
4493 * @order: The order of the allocation.
4495 * This function can free multi-page allocations that are not compound
4496 * pages. It does not check that the @order passed in matches that of
4497 * the allocation, so it is easy to leak memory. Freeing more memory
4498 * than was allocated will probably emit a warning.
4500 * If the last reference to this page is speculative, it will be released
4501 * by put_page() which only frees the first page of a non-compound
4502 * allocation. To prevent the remaining pages from being leaked, we free
4503 * the subsequent pages here. If you want to use the page's reference
4504 * count to decide when to free the allocation, you should allocate a
4505 * compound page, and use put_page() instead of __free_pages().
4507 * Context: May be called in interrupt context or while holding a normal
4508 * spinlock, but not in NMI context or while holding a raw spinlock.
4510 void __free_pages(struct page
*page
, unsigned int order
)
4512 /* get PageHead before we drop reference */
4513 int head
= PageHead(page
);
4515 if (put_page_testzero(page
))
4516 free_the_page(page
, order
);
4519 free_the_page(page
+ (1 << order
), order
);
4521 EXPORT_SYMBOL(__free_pages
);
4523 void free_pages(unsigned long addr
, unsigned int order
)
4526 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4527 __free_pages(virt_to_page((void *)addr
), order
);
4531 EXPORT_SYMBOL(free_pages
);
4535 * An arbitrary-length arbitrary-offset area of memory which resides
4536 * within a 0 or higher order page. Multiple fragments within that page
4537 * are individually refcounted, in the page's reference counter.
4539 * The page_frag functions below provide a simple allocation framework for
4540 * page fragments. This is used by the network stack and network device
4541 * drivers to provide a backing region of memory for use as either an
4542 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4544 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4547 struct page
*page
= NULL
;
4548 gfp_t gfp
= gfp_mask
;
4550 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4551 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4553 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4554 PAGE_FRAG_CACHE_MAX_ORDER
);
4555 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4557 if (unlikely(!page
))
4558 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4560 nc
->va
= page
? page_address(page
) : NULL
;
4565 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4567 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4569 if (page_ref_sub_and_test(page
, count
))
4570 free_the_page(page
, compound_order(page
));
4572 EXPORT_SYMBOL(__page_frag_cache_drain
);
4574 void *page_frag_alloc_align(struct page_frag_cache
*nc
,
4575 unsigned int fragsz
, gfp_t gfp_mask
,
4576 unsigned int align_mask
)
4578 unsigned int size
= PAGE_SIZE
;
4582 if (unlikely(!nc
->va
)) {
4584 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4588 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4589 /* if size can vary use size else just use PAGE_SIZE */
4592 /* Even if we own the page, we do not use atomic_set().
4593 * This would break get_page_unless_zero() users.
4595 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
4597 /* reset page count bias and offset to start of new frag */
4598 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4599 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4603 offset
= nc
->offset
- fragsz
;
4604 if (unlikely(offset
< 0)) {
4605 page
= virt_to_page(nc
->va
);
4607 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4610 if (unlikely(nc
->pfmemalloc
)) {
4611 free_the_page(page
, compound_order(page
));
4615 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4616 /* if size can vary use size else just use PAGE_SIZE */
4619 /* OK, page count is 0, we can safely set it */
4620 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
4622 /* reset page count bias and offset to start of new frag */
4623 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
4624 offset
= size
- fragsz
;
4625 if (unlikely(offset
< 0)) {
4627 * The caller is trying to allocate a fragment
4628 * with fragsz > PAGE_SIZE but the cache isn't big
4629 * enough to satisfy the request, this may
4630 * happen in low memory conditions.
4631 * We don't release the cache page because
4632 * it could make memory pressure worse
4633 * so we simply return NULL here.
4640 offset
&= align_mask
;
4641 nc
->offset
= offset
;
4643 return nc
->va
+ offset
;
4645 EXPORT_SYMBOL(page_frag_alloc_align
);
4648 * Frees a page fragment allocated out of either a compound or order 0 page.
4650 void page_frag_free(void *addr
)
4652 struct page
*page
= virt_to_head_page(addr
);
4654 if (unlikely(put_page_testzero(page
)))
4655 free_the_page(page
, compound_order(page
));
4657 EXPORT_SYMBOL(page_frag_free
);
4659 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4663 unsigned long nr
= DIV_ROUND_UP(size
, PAGE_SIZE
);
4664 struct page
*page
= virt_to_page((void *)addr
);
4665 struct page
*last
= page
+ nr
;
4667 split_page_owner(page
, 1 << order
);
4668 split_page_memcg(page
, 1 << order
);
4669 while (page
< --last
)
4670 set_page_refcounted(last
);
4672 last
= page
+ (1UL << order
);
4673 for (page
+= nr
; page
< last
; page
++)
4674 __free_pages_ok(page
, 0, FPI_TO_TAIL
);
4676 return (void *)addr
;
4680 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4681 * @size: the number of bytes to allocate
4682 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4684 * This function is similar to alloc_pages(), except that it allocates the
4685 * minimum number of pages to satisfy the request. alloc_pages() can only
4686 * allocate memory in power-of-two pages.
4688 * This function is also limited by MAX_ORDER.
4690 * Memory allocated by this function must be released by free_pages_exact().
4692 * Return: pointer to the allocated area or %NULL in case of error.
4694 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4696 unsigned int order
= get_order(size
);
4699 if (WARN_ON_ONCE(gfp_mask
& (__GFP_COMP
| __GFP_HIGHMEM
)))
4700 gfp_mask
&= ~(__GFP_COMP
| __GFP_HIGHMEM
);
4702 addr
= __get_free_pages(gfp_mask
, order
);
4703 return make_alloc_exact(addr
, order
, size
);
4705 EXPORT_SYMBOL(alloc_pages_exact
);
4708 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4710 * @nid: the preferred node ID where memory should be allocated
4711 * @size: the number of bytes to allocate
4712 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4714 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4717 * Return: pointer to the allocated area or %NULL in case of error.
4719 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4721 unsigned int order
= get_order(size
);
4724 if (WARN_ON_ONCE(gfp_mask
& (__GFP_COMP
| __GFP_HIGHMEM
)))
4725 gfp_mask
&= ~(__GFP_COMP
| __GFP_HIGHMEM
);
4727 p
= alloc_pages_node(nid
, gfp_mask
, order
);
4730 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4734 * free_pages_exact - release memory allocated via alloc_pages_exact()
4735 * @virt: the value returned by alloc_pages_exact.
4736 * @size: size of allocation, same value as passed to alloc_pages_exact().
4738 * Release the memory allocated by a previous call to alloc_pages_exact.
4740 void free_pages_exact(void *virt
, size_t size
)
4742 unsigned long addr
= (unsigned long)virt
;
4743 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4745 while (addr
< end
) {
4750 EXPORT_SYMBOL(free_pages_exact
);
4753 * nr_free_zone_pages - count number of pages beyond high watermark
4754 * @offset: The zone index of the highest zone
4756 * nr_free_zone_pages() counts the number of pages which are beyond the
4757 * high watermark within all zones at or below a given zone index. For each
4758 * zone, the number of pages is calculated as:
4760 * nr_free_zone_pages = managed_pages - high_pages
4762 * Return: number of pages beyond high watermark.
4764 static unsigned long nr_free_zone_pages(int offset
)
4769 /* Just pick one node, since fallback list is circular */
4770 unsigned long sum
= 0;
4772 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4774 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4775 unsigned long size
= zone_managed_pages(zone
);
4776 unsigned long high
= high_wmark_pages(zone
);
4785 * nr_free_buffer_pages - count number of pages beyond high watermark
4787 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4788 * watermark within ZONE_DMA and ZONE_NORMAL.
4790 * Return: number of pages beyond high watermark within ZONE_DMA and
4793 unsigned long nr_free_buffer_pages(void)
4795 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4797 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4799 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4801 zoneref
->zone
= zone
;
4802 zoneref
->zone_idx
= zone_idx(zone
);
4806 * Builds allocation fallback zone lists.
4808 * Add all populated zones of a node to the zonelist.
4810 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
4813 enum zone_type zone_type
= MAX_NR_ZONES
;
4818 zone
= pgdat
->node_zones
+ zone_type
;
4819 if (populated_zone(zone
)) {
4820 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
4821 check_highest_zone(zone_type
);
4823 } while (zone_type
);
4830 static int __parse_numa_zonelist_order(char *s
)
4833 * We used to support different zonelists modes but they turned
4834 * out to be just not useful. Let's keep the warning in place
4835 * if somebody still use the cmd line parameter so that we do
4836 * not fail it silently
4838 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
4839 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
4845 static char numa_zonelist_order
[] = "Node";
4846 #define NUMA_ZONELIST_ORDER_LEN 16
4848 * sysctl handler for numa_zonelist_order
4850 static int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4851 void *buffer
, size_t *length
, loff_t
*ppos
)
4854 return __parse_numa_zonelist_order(buffer
);
4855 return proc_dostring(table
, write
, buffer
, length
, ppos
);
4858 static int node_load
[MAX_NUMNODES
];
4861 * find_next_best_node - find the next node that should appear in a given node's fallback list
4862 * @node: node whose fallback list we're appending
4863 * @used_node_mask: nodemask_t of already used nodes
4865 * We use a number of factors to determine which is the next node that should
4866 * appear on a given node's fallback list. The node should not have appeared
4867 * already in @node's fallback list, and it should be the next closest node
4868 * according to the distance array (which contains arbitrary distance values
4869 * from each node to each node in the system), and should also prefer nodes
4870 * with no CPUs, since presumably they'll have very little allocation pressure
4871 * on them otherwise.
4873 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
4875 int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4878 int min_val
= INT_MAX
;
4879 int best_node
= NUMA_NO_NODE
;
4881 /* Use the local node if we haven't already */
4882 if (!node_isset(node
, *used_node_mask
)) {
4883 node_set(node
, *used_node_mask
);
4887 for_each_node_state(n
, N_MEMORY
) {
4889 /* Don't want a node to appear more than once */
4890 if (node_isset(n
, *used_node_mask
))
4893 /* Use the distance array to find the distance */
4894 val
= node_distance(node
, n
);
4896 /* Penalize nodes under us ("prefer the next node") */
4899 /* Give preference to headless and unused nodes */
4900 if (!cpumask_empty(cpumask_of_node(n
)))
4901 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4903 /* Slight preference for less loaded node */
4904 val
*= MAX_NUMNODES
;
4905 val
+= node_load
[n
];
4907 if (val
< min_val
) {
4914 node_set(best_node
, *used_node_mask
);
4921 * Build zonelists ordered by node and zones within node.
4922 * This results in maximum locality--normal zone overflows into local
4923 * DMA zone, if any--but risks exhausting DMA zone.
4925 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
4928 struct zoneref
*zonerefs
;
4931 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
4933 for (i
= 0; i
< nr_nodes
; i
++) {
4936 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
4938 nr_zones
= build_zonerefs_node(node
, zonerefs
);
4939 zonerefs
+= nr_zones
;
4941 zonerefs
->zone
= NULL
;
4942 zonerefs
->zone_idx
= 0;
4946 * Build gfp_thisnode zonelists
4948 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4950 struct zoneref
*zonerefs
;
4953 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
4954 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
4955 zonerefs
+= nr_zones
;
4956 zonerefs
->zone
= NULL
;
4957 zonerefs
->zone_idx
= 0;
4961 * Build zonelists ordered by zone and nodes within zones.
4962 * This results in conserving DMA zone[s] until all Normal memory is
4963 * exhausted, but results in overflowing to remote node while memory
4964 * may still exist in local DMA zone.
4967 static void build_zonelists(pg_data_t
*pgdat
)
4969 static int node_order
[MAX_NUMNODES
];
4970 int node
, nr_nodes
= 0;
4971 nodemask_t used_mask
= NODE_MASK_NONE
;
4972 int local_node
, prev_node
;
4974 /* NUMA-aware ordering of nodes */
4975 local_node
= pgdat
->node_id
;
4976 prev_node
= local_node
;
4978 memset(node_order
, 0, sizeof(node_order
));
4979 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4981 * We don't want to pressure a particular node.
4982 * So adding penalty to the first node in same
4983 * distance group to make it round-robin.
4985 if (node_distance(local_node
, node
) !=
4986 node_distance(local_node
, prev_node
))
4987 node_load
[node
] += 1;
4989 node_order
[nr_nodes
++] = node
;
4993 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
4994 build_thisnode_zonelists(pgdat
);
4995 pr_info("Fallback order for Node %d: ", local_node
);
4996 for (node
= 0; node
< nr_nodes
; node
++)
4997 pr_cont("%d ", node_order
[node
]);
5001 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5003 * Return node id of node used for "local" allocations.
5004 * I.e., first node id of first zone in arg node's generic zonelist.
5005 * Used for initializing percpu 'numa_mem', which is used primarily
5006 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5008 int local_memory_node(int node
)
5012 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5013 gfp_zone(GFP_KERNEL
),
5015 return zone_to_nid(z
->zone
);
5019 static void setup_min_unmapped_ratio(void);
5020 static void setup_min_slab_ratio(void);
5021 #else /* CONFIG_NUMA */
5023 static void build_zonelists(pg_data_t
*pgdat
)
5025 int node
, local_node
;
5026 struct zoneref
*zonerefs
;
5029 local_node
= pgdat
->node_id
;
5031 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5032 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5033 zonerefs
+= nr_zones
;
5036 * Now we build the zonelist so that it contains the zones
5037 * of all the other nodes.
5038 * We don't want to pressure a particular node, so when
5039 * building the zones for node N, we make sure that the
5040 * zones coming right after the local ones are those from
5041 * node N+1 (modulo N)
5043 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5044 if (!node_online(node
))
5046 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5047 zonerefs
+= nr_zones
;
5049 for (node
= 0; node
< local_node
; node
++) {
5050 if (!node_online(node
))
5052 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5053 zonerefs
+= nr_zones
;
5056 zonerefs
->zone
= NULL
;
5057 zonerefs
->zone_idx
= 0;
5060 #endif /* CONFIG_NUMA */
5063 * Boot pageset table. One per cpu which is going to be used for all
5064 * zones and all nodes. The parameters will be set in such a way
5065 * that an item put on a list will immediately be handed over to
5066 * the buddy list. This is safe since pageset manipulation is done
5067 * with interrupts disabled.
5069 * The boot_pagesets must be kept even after bootup is complete for
5070 * unused processors and/or zones. They do play a role for bootstrapping
5071 * hotplugged processors.
5073 * zoneinfo_show() and maybe other functions do
5074 * not check if the processor is online before following the pageset pointer.
5075 * Other parts of the kernel may not check if the zone is available.
5077 static void per_cpu_pages_init(struct per_cpu_pages
*pcp
, struct per_cpu_zonestat
*pzstats
);
5078 /* These effectively disable the pcplists in the boot pageset completely */
5079 #define BOOT_PAGESET_HIGH 0
5080 #define BOOT_PAGESET_BATCH 1
5081 static DEFINE_PER_CPU(struct per_cpu_pages
, boot_pageset
);
5082 static DEFINE_PER_CPU(struct per_cpu_zonestat
, boot_zonestats
);
5084 static void __build_all_zonelists(void *data
)
5087 int __maybe_unused cpu
;
5088 pg_data_t
*self
= data
;
5089 unsigned long flags
;
5092 * The zonelist_update_seq must be acquired with irqsave because the
5093 * reader can be invoked from IRQ with GFP_ATOMIC.
5095 write_seqlock_irqsave(&zonelist_update_seq
, flags
);
5097 * Also disable synchronous printk() to prevent any printk() from
5098 * trying to hold port->lock, for
5099 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5100 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5102 printk_deferred_enter();
5105 memset(node_load
, 0, sizeof(node_load
));
5109 * This node is hotadded and no memory is yet present. So just
5110 * building zonelists is fine - no need to touch other nodes.
5112 if (self
&& !node_online(self
->node_id
)) {
5113 build_zonelists(self
);
5116 * All possible nodes have pgdat preallocated
5119 for_each_node(nid
) {
5120 pg_data_t
*pgdat
= NODE_DATA(nid
);
5122 build_zonelists(pgdat
);
5125 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5127 * We now know the "local memory node" for each node--
5128 * i.e., the node of the first zone in the generic zonelist.
5129 * Set up numa_mem percpu variable for on-line cpus. During
5130 * boot, only the boot cpu should be on-line; we'll init the
5131 * secondary cpus' numa_mem as they come on-line. During
5132 * node/memory hotplug, we'll fixup all on-line cpus.
5134 for_each_online_cpu(cpu
)
5135 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5139 printk_deferred_exit();
5140 write_sequnlock_irqrestore(&zonelist_update_seq
, flags
);
5143 static noinline
void __init
5144 build_all_zonelists_init(void)
5148 __build_all_zonelists(NULL
);
5151 * Initialize the boot_pagesets that are going to be used
5152 * for bootstrapping processors. The real pagesets for
5153 * each zone will be allocated later when the per cpu
5154 * allocator is available.
5156 * boot_pagesets are used also for bootstrapping offline
5157 * cpus if the system is already booted because the pagesets
5158 * are needed to initialize allocators on a specific cpu too.
5159 * F.e. the percpu allocator needs the page allocator which
5160 * needs the percpu allocator in order to allocate its pagesets
5161 * (a chicken-egg dilemma).
5163 for_each_possible_cpu(cpu
)
5164 per_cpu_pages_init(&per_cpu(boot_pageset
, cpu
), &per_cpu(boot_zonestats
, cpu
));
5166 mminit_verify_zonelist();
5167 cpuset_init_current_mems_allowed();
5171 * unless system_state == SYSTEM_BOOTING.
5173 * __ref due to call of __init annotated helper build_all_zonelists_init
5174 * [protected by SYSTEM_BOOTING].
5176 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5178 unsigned long vm_total_pages
;
5180 if (system_state
== SYSTEM_BOOTING
) {
5181 build_all_zonelists_init();
5183 __build_all_zonelists(pgdat
);
5184 /* cpuset refresh routine should be here */
5186 /* Get the number of free pages beyond high watermark in all zones. */
5187 vm_total_pages
= nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
5189 * Disable grouping by mobility if the number of pages in the
5190 * system is too low to allow the mechanism to work. It would be
5191 * more accurate, but expensive to check per-zone. This check is
5192 * made on memory-hotadd so a system can start with mobility
5193 * disabled and enable it later
5195 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5196 page_group_by_mobility_disabled
= 1;
5198 page_group_by_mobility_disabled
= 0;
5200 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5202 page_group_by_mobility_disabled
? "off" : "on",
5205 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5209 static int zone_batchsize(struct zone
*zone
)
5215 * The number of pages to batch allocate is either ~0.1%
5216 * of the zone or 1MB, whichever is smaller. The batch
5217 * size is striking a balance between allocation latency
5218 * and zone lock contention.
5220 batch
= min(zone_managed_pages(zone
) >> 10, SZ_1M
/ PAGE_SIZE
);
5221 batch
/= 4; /* We effectively *= 4 below */
5226 * Clamp the batch to a 2^n - 1 value. Having a power
5227 * of 2 value was found to be more likely to have
5228 * suboptimal cache aliasing properties in some cases.
5230 * For example if 2 tasks are alternately allocating
5231 * batches of pages, one task can end up with a lot
5232 * of pages of one half of the possible page colors
5233 * and the other with pages of the other colors.
5235 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5240 /* The deferral and batching of frees should be suppressed under NOMMU
5243 * The problem is that NOMMU needs to be able to allocate large chunks
5244 * of contiguous memory as there's no hardware page translation to
5245 * assemble apparent contiguous memory from discontiguous pages.
5247 * Queueing large contiguous runs of pages for batching, however,
5248 * causes the pages to actually be freed in smaller chunks. As there
5249 * can be a significant delay between the individual batches being
5250 * recycled, this leads to the once large chunks of space being
5251 * fragmented and becoming unavailable for high-order allocations.
5257 static int percpu_pagelist_high_fraction
;
5258 static int zone_highsize(struct zone
*zone
, int batch
, int cpu_online
)
5263 unsigned long total_pages
;
5265 if (!percpu_pagelist_high_fraction
) {
5267 * By default, the high value of the pcp is based on the zone
5268 * low watermark so that if they are full then background
5269 * reclaim will not be started prematurely.
5271 total_pages
= low_wmark_pages(zone
);
5274 * If percpu_pagelist_high_fraction is configured, the high
5275 * value is based on a fraction of the managed pages in the
5278 total_pages
= zone_managed_pages(zone
) / percpu_pagelist_high_fraction
;
5282 * Split the high value across all online CPUs local to the zone. Note
5283 * that early in boot that CPUs may not be online yet and that during
5284 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5285 * onlined. For memory nodes that have no CPUs, split pcp->high across
5286 * all online CPUs to mitigate the risk that reclaim is triggered
5287 * prematurely due to pages stored on pcp lists.
5289 nr_split_cpus
= cpumask_weight(cpumask_of_node(zone_to_nid(zone
))) + cpu_online
;
5291 nr_split_cpus
= num_online_cpus();
5292 high
= total_pages
/ nr_split_cpus
;
5295 * Ensure high is at least batch*4. The multiple is based on the
5296 * historical relationship between high and batch.
5298 high
= max(high
, batch
<< 2);
5307 * pcp->high and pcp->batch values are related and generally batch is lower
5308 * than high. They are also related to pcp->count such that count is lower
5309 * than high, and as soon as it reaches high, the pcplist is flushed.
5311 * However, guaranteeing these relations at all times would require e.g. write
5312 * barriers here but also careful usage of read barriers at the read side, and
5313 * thus be prone to error and bad for performance. Thus the update only prevents
5314 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
5315 * can cope with those fields changing asynchronously, and fully trust only the
5316 * pcp->count field on the local CPU with interrupts disabled.
5318 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5319 * outside of boot time (or some other assurance that no concurrent updaters
5322 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5323 unsigned long batch
)
5325 WRITE_ONCE(pcp
->batch
, batch
);
5326 WRITE_ONCE(pcp
->high
, high
);
5329 static void per_cpu_pages_init(struct per_cpu_pages
*pcp
, struct per_cpu_zonestat
*pzstats
)
5333 memset(pcp
, 0, sizeof(*pcp
));
5334 memset(pzstats
, 0, sizeof(*pzstats
));
5336 spin_lock_init(&pcp
->lock
);
5337 for (pindex
= 0; pindex
< NR_PCP_LISTS
; pindex
++)
5338 INIT_LIST_HEAD(&pcp
->lists
[pindex
]);
5341 * Set batch and high values safe for a boot pageset. A true percpu
5342 * pageset's initialization will update them subsequently. Here we don't
5343 * need to be as careful as pageset_update() as nobody can access the
5346 pcp
->high
= BOOT_PAGESET_HIGH
;
5347 pcp
->batch
= BOOT_PAGESET_BATCH
;
5348 pcp
->free_factor
= 0;
5351 static void __zone_set_pageset_high_and_batch(struct zone
*zone
, unsigned long high
,
5352 unsigned long batch
)
5354 struct per_cpu_pages
*pcp
;
5357 for_each_possible_cpu(cpu
) {
5358 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
5359 pageset_update(pcp
, high
, batch
);
5364 * Calculate and set new high and batch values for all per-cpu pagesets of a
5365 * zone based on the zone's size.
5367 static void zone_set_pageset_high_and_batch(struct zone
*zone
, int cpu_online
)
5369 int new_high
, new_batch
;
5371 new_batch
= max(1, zone_batchsize(zone
));
5372 new_high
= zone_highsize(zone
, new_batch
, cpu_online
);
5374 if (zone
->pageset_high
== new_high
&&
5375 zone
->pageset_batch
== new_batch
)
5378 zone
->pageset_high
= new_high
;
5379 zone
->pageset_batch
= new_batch
;
5381 __zone_set_pageset_high_and_batch(zone
, new_high
, new_batch
);
5384 void __meminit
setup_zone_pageset(struct zone
*zone
)
5388 /* Size may be 0 on !SMP && !NUMA */
5389 if (sizeof(struct per_cpu_zonestat
) > 0)
5390 zone
->per_cpu_zonestats
= alloc_percpu(struct per_cpu_zonestat
);
5392 zone
->per_cpu_pageset
= alloc_percpu(struct per_cpu_pages
);
5393 for_each_possible_cpu(cpu
) {
5394 struct per_cpu_pages
*pcp
;
5395 struct per_cpu_zonestat
*pzstats
;
5397 pcp
= per_cpu_ptr(zone
->per_cpu_pageset
, cpu
);
5398 pzstats
= per_cpu_ptr(zone
->per_cpu_zonestats
, cpu
);
5399 per_cpu_pages_init(pcp
, pzstats
);
5402 zone_set_pageset_high_and_batch(zone
, 0);
5406 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5407 * page high values need to be recalculated.
5409 static void zone_pcp_update(struct zone
*zone
, int cpu_online
)
5411 mutex_lock(&pcp_batch_high_lock
);
5412 zone_set_pageset_high_and_batch(zone
, cpu_online
);
5413 mutex_unlock(&pcp_batch_high_lock
);
5417 * Allocate per cpu pagesets and initialize them.
5418 * Before this call only boot pagesets were available.
5420 void __init
setup_per_cpu_pageset(void)
5422 struct pglist_data
*pgdat
;
5424 int __maybe_unused cpu
;
5426 for_each_populated_zone(zone
)
5427 setup_zone_pageset(zone
);
5431 * Unpopulated zones continue using the boot pagesets.
5432 * The numa stats for these pagesets need to be reset.
5433 * Otherwise, they will end up skewing the stats of
5434 * the nodes these zones are associated with.
5436 for_each_possible_cpu(cpu
) {
5437 struct per_cpu_zonestat
*pzstats
= &per_cpu(boot_zonestats
, cpu
);
5438 memset(pzstats
->vm_numa_event
, 0,
5439 sizeof(pzstats
->vm_numa_event
));
5443 for_each_online_pgdat(pgdat
)
5444 pgdat
->per_cpu_nodestats
=
5445 alloc_percpu(struct per_cpu_nodestat
);
5448 __meminit
void zone_pcp_init(struct zone
*zone
)
5451 * per cpu subsystem is not up at this point. The following code
5452 * relies on the ability of the linker to provide the
5453 * offset of a (static) per cpu variable into the per cpu area.
5455 zone
->per_cpu_pageset
= &boot_pageset
;
5456 zone
->per_cpu_zonestats
= &boot_zonestats
;
5457 zone
->pageset_high
= BOOT_PAGESET_HIGH
;
5458 zone
->pageset_batch
= BOOT_PAGESET_BATCH
;
5460 if (populated_zone(zone
))
5461 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone
->name
,
5462 zone
->present_pages
, zone_batchsize(zone
));
5465 void adjust_managed_page_count(struct page
*page
, long count
)
5467 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
5468 totalram_pages_add(count
);
5469 #ifdef CONFIG_HIGHMEM
5470 if (PageHighMem(page
))
5471 totalhigh_pages_add(count
);
5474 EXPORT_SYMBOL(adjust_managed_page_count
);
5476 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
5479 unsigned long pages
= 0;
5481 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5482 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5483 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5484 struct page
*page
= virt_to_page(pos
);
5485 void *direct_map_addr
;
5488 * 'direct_map_addr' might be different from 'pos'
5489 * because some architectures' virt_to_page()
5490 * work with aliases. Getting the direct map
5491 * address ensures that we get a _writeable_
5492 * alias for the memset().
5494 direct_map_addr
= page_address(page
);
5496 * Perform a kasan-unchecked memset() since this memory
5497 * has not been initialized.
5499 direct_map_addr
= kasan_reset_tag(direct_map_addr
);
5500 if ((unsigned int)poison
<= 0xFF)
5501 memset(direct_map_addr
, poison
, PAGE_SIZE
);
5503 free_reserved_page(page
);
5507 pr_info("Freeing %s memory: %ldK\n", s
, K(pages
));
5512 static int page_alloc_cpu_dead(unsigned int cpu
)
5516 lru_add_drain_cpu(cpu
);
5517 mlock_drain_remote(cpu
);
5521 * Spill the event counters of the dead processor
5522 * into the current processors event counters.
5523 * This artificially elevates the count of the current
5526 vm_events_fold_cpu(cpu
);
5529 * Zero the differential counters of the dead processor
5530 * so that the vm statistics are consistent.
5532 * This is only okay since the processor is dead and cannot
5533 * race with what we are doing.
5535 cpu_vm_stats_fold(cpu
);
5537 for_each_populated_zone(zone
)
5538 zone_pcp_update(zone
, 0);
5543 static int page_alloc_cpu_online(unsigned int cpu
)
5547 for_each_populated_zone(zone
)
5548 zone_pcp_update(zone
, 1);
5552 void __init
page_alloc_init_cpuhp(void)
5556 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC
,
5557 "mm/page_alloc:pcp",
5558 page_alloc_cpu_online
,
5559 page_alloc_cpu_dead
);
5564 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5565 * or min_free_kbytes changes.
5567 static void calculate_totalreserve_pages(void)
5569 struct pglist_data
*pgdat
;
5570 unsigned long reserve_pages
= 0;
5571 enum zone_type i
, j
;
5573 for_each_online_pgdat(pgdat
) {
5575 pgdat
->totalreserve_pages
= 0;
5577 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5578 struct zone
*zone
= pgdat
->node_zones
+ i
;
5580 unsigned long managed_pages
= zone_managed_pages(zone
);
5582 /* Find valid and maximum lowmem_reserve in the zone */
5583 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5584 if (zone
->lowmem_reserve
[j
] > max
)
5585 max
= zone
->lowmem_reserve
[j
];
5588 /* we treat the high watermark as reserved pages. */
5589 max
+= high_wmark_pages(zone
);
5591 if (max
> managed_pages
)
5592 max
= managed_pages
;
5594 pgdat
->totalreserve_pages
+= max
;
5596 reserve_pages
+= max
;
5599 totalreserve_pages
= reserve_pages
;
5603 * setup_per_zone_lowmem_reserve - called whenever
5604 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5605 * has a correct pages reserved value, so an adequate number of
5606 * pages are left in the zone after a successful __alloc_pages().
5608 static void setup_per_zone_lowmem_reserve(void)
5610 struct pglist_data
*pgdat
;
5611 enum zone_type i
, j
;
5613 for_each_online_pgdat(pgdat
) {
5614 for (i
= 0; i
< MAX_NR_ZONES
- 1; i
++) {
5615 struct zone
*zone
= &pgdat
->node_zones
[i
];
5616 int ratio
= sysctl_lowmem_reserve_ratio
[i
];
5617 bool clear
= !ratio
|| !zone_managed_pages(zone
);
5618 unsigned long managed_pages
= 0;
5620 for (j
= i
+ 1; j
< MAX_NR_ZONES
; j
++) {
5621 struct zone
*upper_zone
= &pgdat
->node_zones
[j
];
5623 managed_pages
+= zone_managed_pages(upper_zone
);
5626 zone
->lowmem_reserve
[j
] = 0;
5628 zone
->lowmem_reserve
[j
] = managed_pages
/ ratio
;
5633 /* update totalreserve_pages */
5634 calculate_totalreserve_pages();
5637 static void __setup_per_zone_wmarks(void)
5639 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5640 unsigned long lowmem_pages
= 0;
5642 unsigned long flags
;
5644 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5645 for_each_zone(zone
) {
5646 if (!is_highmem(zone
) && zone_idx(zone
) != ZONE_MOVABLE
)
5647 lowmem_pages
+= zone_managed_pages(zone
);
5650 for_each_zone(zone
) {
5653 spin_lock_irqsave(&zone
->lock
, flags
);
5654 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
5655 do_div(tmp
, lowmem_pages
);
5656 if (is_highmem(zone
) || zone_idx(zone
) == ZONE_MOVABLE
) {
5658 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5659 * need highmem and movable zones pages, so cap pages_min
5660 * to a small value here.
5662 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5663 * deltas control async page reclaim, and so should
5664 * not be capped for highmem and movable zones.
5666 unsigned long min_pages
;
5668 min_pages
= zone_managed_pages(zone
) / 1024;
5669 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5670 zone
->_watermark
[WMARK_MIN
] = min_pages
;
5673 * If it's a lowmem zone, reserve a number of pages
5674 * proportionate to the zone's size.
5676 zone
->_watermark
[WMARK_MIN
] = tmp
;
5680 * Set the kswapd watermarks distance according to the
5681 * scale factor in proportion to available memory, but
5682 * ensure a minimum size on small systems.
5684 tmp
= max_t(u64
, tmp
>> 2,
5685 mult_frac(zone_managed_pages(zone
),
5686 watermark_scale_factor
, 10000));
5688 zone
->watermark_boost
= 0;
5689 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
5690 zone
->_watermark
[WMARK_HIGH
] = low_wmark_pages(zone
) + tmp
;
5691 zone
->_watermark
[WMARK_PROMO
] = high_wmark_pages(zone
) + tmp
;
5693 spin_unlock_irqrestore(&zone
->lock
, flags
);
5696 /* update totalreserve_pages */
5697 calculate_totalreserve_pages();
5701 * setup_per_zone_wmarks - called when min_free_kbytes changes
5702 * or when memory is hot-{added|removed}
5704 * Ensures that the watermark[min,low,high] values for each zone are set
5705 * correctly with respect to min_free_kbytes.
5707 void setup_per_zone_wmarks(void)
5710 static DEFINE_SPINLOCK(lock
);
5713 __setup_per_zone_wmarks();
5717 * The watermark size have changed so update the pcpu batch
5718 * and high limits or the limits may be inappropriate.
5721 zone_pcp_update(zone
, 0);
5725 * Initialise min_free_kbytes.
5727 * For small machines we want it small (128k min). For large machines
5728 * we want it large (256MB max). But it is not linear, because network
5729 * bandwidth does not increase linearly with machine size. We use
5731 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5732 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5748 void calculate_min_free_kbytes(void)
5750 unsigned long lowmem_kbytes
;
5751 int new_min_free_kbytes
;
5753 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5754 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5756 if (new_min_free_kbytes
> user_min_free_kbytes
)
5757 min_free_kbytes
= clamp(new_min_free_kbytes
, 128, 262144);
5759 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5760 new_min_free_kbytes
, user_min_free_kbytes
);
5764 int __meminit
init_per_zone_wmark_min(void)
5766 calculate_min_free_kbytes();
5767 setup_per_zone_wmarks();
5768 refresh_zone_stat_thresholds();
5769 setup_per_zone_lowmem_reserve();
5772 setup_min_unmapped_ratio();
5773 setup_min_slab_ratio();
5776 khugepaged_min_free_kbytes_update();
5780 postcore_initcall(init_per_zone_wmark_min
)
5783 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5784 * that we can call two helper functions whenever min_free_kbytes
5787 static int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5788 void *buffer
, size_t *length
, loff_t
*ppos
)
5792 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5797 user_min_free_kbytes
= min_free_kbytes
;
5798 setup_per_zone_wmarks();
5803 static int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
5804 void *buffer
, size_t *length
, loff_t
*ppos
)
5808 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5813 setup_per_zone_wmarks();
5819 static void setup_min_unmapped_ratio(void)
5824 for_each_online_pgdat(pgdat
)
5825 pgdat
->min_unmapped_pages
= 0;
5828 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
5829 sysctl_min_unmapped_ratio
) / 100;
5833 static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5834 void *buffer
, size_t *length
, loff_t
*ppos
)
5838 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5842 setup_min_unmapped_ratio();
5847 static void setup_min_slab_ratio(void)
5852 for_each_online_pgdat(pgdat
)
5853 pgdat
->min_slab_pages
= 0;
5856 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
5857 sysctl_min_slab_ratio
) / 100;
5860 static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5861 void *buffer
, size_t *length
, loff_t
*ppos
)
5865 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5869 setup_min_slab_ratio();
5876 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5877 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5878 * whenever sysctl_lowmem_reserve_ratio changes.
5880 * The reserve ratio obviously has absolutely no relation with the
5881 * minimum watermarks. The lowmem reserve ratio can only make sense
5882 * if in function of the boot time zone sizes.
5884 static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
,
5885 int write
, void *buffer
, size_t *length
, loff_t
*ppos
)
5889 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5891 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5892 if (sysctl_lowmem_reserve_ratio
[i
] < 1)
5893 sysctl_lowmem_reserve_ratio
[i
] = 0;
5896 setup_per_zone_lowmem_reserve();
5901 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
5902 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5903 * pagelist can have before it gets flushed back to buddy allocator.
5905 static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table
*table
,
5906 int write
, void *buffer
, size_t *length
, loff_t
*ppos
)
5909 int old_percpu_pagelist_high_fraction
;
5912 mutex_lock(&pcp_batch_high_lock
);
5913 old_percpu_pagelist_high_fraction
= percpu_pagelist_high_fraction
;
5915 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5916 if (!write
|| ret
< 0)
5919 /* Sanity checking to avoid pcp imbalance */
5920 if (percpu_pagelist_high_fraction
&&
5921 percpu_pagelist_high_fraction
< MIN_PERCPU_PAGELIST_HIGH_FRACTION
) {
5922 percpu_pagelist_high_fraction
= old_percpu_pagelist_high_fraction
;
5928 if (percpu_pagelist_high_fraction
== old_percpu_pagelist_high_fraction
)
5931 for_each_populated_zone(zone
)
5932 zone_set_pageset_high_and_batch(zone
, 0);
5934 mutex_unlock(&pcp_batch_high_lock
);
5938 static struct ctl_table page_alloc_sysctl_table
[] = {
5940 .procname
= "min_free_kbytes",
5941 .data
= &min_free_kbytes
,
5942 .maxlen
= sizeof(min_free_kbytes
),
5944 .proc_handler
= min_free_kbytes_sysctl_handler
,
5945 .extra1
= SYSCTL_ZERO
,
5948 .procname
= "watermark_boost_factor",
5949 .data
= &watermark_boost_factor
,
5950 .maxlen
= sizeof(watermark_boost_factor
),
5952 .proc_handler
= proc_dointvec_minmax
,
5953 .extra1
= SYSCTL_ZERO
,
5956 .procname
= "watermark_scale_factor",
5957 .data
= &watermark_scale_factor
,
5958 .maxlen
= sizeof(watermark_scale_factor
),
5960 .proc_handler
= watermark_scale_factor_sysctl_handler
,
5961 .extra1
= SYSCTL_ONE
,
5962 .extra2
= SYSCTL_THREE_THOUSAND
,
5965 .procname
= "percpu_pagelist_high_fraction",
5966 .data
= &percpu_pagelist_high_fraction
,
5967 .maxlen
= sizeof(percpu_pagelist_high_fraction
),
5969 .proc_handler
= percpu_pagelist_high_fraction_sysctl_handler
,
5970 .extra1
= SYSCTL_ZERO
,
5973 .procname
= "lowmem_reserve_ratio",
5974 .data
= &sysctl_lowmem_reserve_ratio
,
5975 .maxlen
= sizeof(sysctl_lowmem_reserve_ratio
),
5977 .proc_handler
= lowmem_reserve_ratio_sysctl_handler
,
5981 .procname
= "numa_zonelist_order",
5982 .data
= &numa_zonelist_order
,
5983 .maxlen
= NUMA_ZONELIST_ORDER_LEN
,
5985 .proc_handler
= numa_zonelist_order_handler
,
5988 .procname
= "min_unmapped_ratio",
5989 .data
= &sysctl_min_unmapped_ratio
,
5990 .maxlen
= sizeof(sysctl_min_unmapped_ratio
),
5992 .proc_handler
= sysctl_min_unmapped_ratio_sysctl_handler
,
5993 .extra1
= SYSCTL_ZERO
,
5994 .extra2
= SYSCTL_ONE_HUNDRED
,
5997 .procname
= "min_slab_ratio",
5998 .data
= &sysctl_min_slab_ratio
,
5999 .maxlen
= sizeof(sysctl_min_slab_ratio
),
6001 .proc_handler
= sysctl_min_slab_ratio_sysctl_handler
,
6002 .extra1
= SYSCTL_ZERO
,
6003 .extra2
= SYSCTL_ONE_HUNDRED
,
6009 void __init
page_alloc_sysctl_init(void)
6011 register_sysctl_init("vm", page_alloc_sysctl_table
);
6014 #ifdef CONFIG_CONTIG_ALLOC
6015 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6016 static void alloc_contig_dump_pages(struct list_head
*page_list
)
6018 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor
, "migrate failure");
6020 if (DYNAMIC_DEBUG_BRANCH(descriptor
)) {
6024 list_for_each_entry(page
, page_list
, lru
)
6025 dump_page(page
, "migration failure");
6029 /* [start, end) must belong to a single zone. */
6030 int __alloc_contig_migrate_range(struct compact_control
*cc
,
6031 unsigned long start
, unsigned long end
)
6033 /* This function is based on compact_zone() from compaction.c. */
6034 unsigned int nr_reclaimed
;
6035 unsigned long pfn
= start
;
6036 unsigned int tries
= 0;
6038 struct migration_target_control mtc
= {
6039 .nid
= zone_to_nid(cc
->zone
),
6040 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
6043 lru_cache_disable();
6045 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6046 if (fatal_signal_pending(current
)) {
6051 if (list_empty(&cc
->migratepages
)) {
6052 cc
->nr_migratepages
= 0;
6053 ret
= isolate_migratepages_range(cc
, pfn
, end
);
6054 if (ret
&& ret
!= -EAGAIN
)
6056 pfn
= cc
->migrate_pfn
;
6058 } else if (++tries
== 5) {
6063 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6065 cc
->nr_migratepages
-= nr_reclaimed
;
6067 ret
= migrate_pages(&cc
->migratepages
, alloc_migration_target
,
6068 NULL
, (unsigned long)&mtc
, cc
->mode
, MR_CONTIG_RANGE
, NULL
);
6071 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6072 * to retry again over this error, so do the same here.
6080 if (!(cc
->gfp_mask
& __GFP_NOWARN
) && ret
== -EBUSY
)
6081 alloc_contig_dump_pages(&cc
->migratepages
);
6082 putback_movable_pages(&cc
->migratepages
);
6089 * alloc_contig_range() -- tries to allocate given range of pages
6090 * @start: start PFN to allocate
6091 * @end: one-past-the-last PFN to allocate
6092 * @migratetype: migratetype of the underlying pageblocks (either
6093 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6094 * in range must have the same migratetype and it must
6095 * be either of the two.
6096 * @gfp_mask: GFP mask to use during compaction
6098 * The PFN range does not have to be pageblock aligned. The PFN range must
6099 * belong to a single zone.
6101 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6102 * pageblocks in the range. Once isolated, the pageblocks should not
6103 * be modified by others.
6105 * Return: zero on success or negative error code. On success all
6106 * pages which PFN is in [start, end) are allocated for the caller and
6107 * need to be freed with free_contig_range().
6109 int alloc_contig_range(unsigned long start
, unsigned long end
,
6110 unsigned migratetype
, gfp_t gfp_mask
)
6112 unsigned long outer_start
, outer_end
;
6116 struct compact_control cc
= {
6117 .nr_migratepages
= 0,
6119 .zone
= page_zone(pfn_to_page(start
)),
6120 .mode
= MIGRATE_SYNC
,
6121 .ignore_skip_hint
= true,
6122 .no_set_skip_hint
= true,
6123 .gfp_mask
= current_gfp_context(gfp_mask
),
6124 .alloc_contig
= true,
6126 INIT_LIST_HEAD(&cc
.migratepages
);
6129 * What we do here is we mark all pageblocks in range as
6130 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6131 * have different sizes, and due to the way page allocator
6132 * work, start_isolate_page_range() has special handlings for this.
6134 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6135 * migrate the pages from an unaligned range (ie. pages that
6136 * we are interested in). This will put all the pages in
6137 * range back to page allocator as MIGRATE_ISOLATE.
6139 * When this is done, we take the pages in range from page
6140 * allocator removing them from the buddy system. This way
6141 * page allocator will never consider using them.
6143 * This lets us mark the pageblocks back as
6144 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6145 * aligned range but not in the unaligned, original range are
6146 * put back to page allocator so that buddy can use them.
6149 ret
= start_isolate_page_range(start
, end
, migratetype
, 0, gfp_mask
);
6153 drain_all_pages(cc
.zone
);
6156 * In case of -EBUSY, we'd like to know which page causes problem.
6157 * So, just fall through. test_pages_isolated() has a tracepoint
6158 * which will report the busy page.
6160 * It is possible that busy pages could become available before
6161 * the call to test_pages_isolated, and the range will actually be
6162 * allocated. So, if we fall through be sure to clear ret so that
6163 * -EBUSY is not accidentally used or returned to caller.
6165 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6166 if (ret
&& ret
!= -EBUSY
)
6171 * Pages from [start, end) are within a pageblock_nr_pages
6172 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6173 * more, all pages in [start, end) are free in page allocator.
6174 * What we are going to do is to allocate all pages from
6175 * [start, end) (that is remove them from page allocator).
6177 * The only problem is that pages at the beginning and at the
6178 * end of interesting range may be not aligned with pages that
6179 * page allocator holds, ie. they can be part of higher order
6180 * pages. Because of this, we reserve the bigger range and
6181 * once this is done free the pages we are not interested in.
6183 * We don't have to hold zone->lock here because the pages are
6184 * isolated thus they won't get removed from buddy.
6188 outer_start
= start
;
6189 while (!PageBuddy(pfn_to_page(outer_start
))) {
6190 if (++order
> MAX_ORDER
) {
6191 outer_start
= start
;
6194 outer_start
&= ~0UL << order
;
6197 if (outer_start
!= start
) {
6198 order
= buddy_order(pfn_to_page(outer_start
));
6201 * outer_start page could be small order buddy page and
6202 * it doesn't include start page. Adjust outer_start
6203 * in this case to report failed page properly
6204 * on tracepoint in test_pages_isolated()
6206 if (outer_start
+ (1UL << order
) <= start
)
6207 outer_start
= start
;
6210 /* Make sure the range is really isolated. */
6211 if (test_pages_isolated(outer_start
, end
, 0)) {
6216 /* Grab isolated pages from freelists. */
6217 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6223 /* Free head and tail (if any) */
6224 if (start
!= outer_start
)
6225 free_contig_range(outer_start
, start
- outer_start
);
6226 if (end
!= outer_end
)
6227 free_contig_range(end
, outer_end
- end
);
6230 undo_isolate_page_range(start
, end
, migratetype
);
6233 EXPORT_SYMBOL(alloc_contig_range
);
6235 static int __alloc_contig_pages(unsigned long start_pfn
,
6236 unsigned long nr_pages
, gfp_t gfp_mask
)
6238 unsigned long end_pfn
= start_pfn
+ nr_pages
;
6240 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
6244 static bool pfn_range_valid_contig(struct zone
*z
, unsigned long start_pfn
,
6245 unsigned long nr_pages
)
6247 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
6250 for (i
= start_pfn
; i
< end_pfn
; i
++) {
6251 page
= pfn_to_online_page(i
);
6255 if (page_zone(page
) != z
)
6258 if (PageReserved(page
))
6267 static bool zone_spans_last_pfn(const struct zone
*zone
,
6268 unsigned long start_pfn
, unsigned long nr_pages
)
6270 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
6272 return zone_spans_pfn(zone
, last_pfn
);
6276 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6277 * @nr_pages: Number of contiguous pages to allocate
6278 * @gfp_mask: GFP mask to limit search and used during compaction
6280 * @nodemask: Mask for other possible nodes
6282 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6283 * on an applicable zonelist to find a contiguous pfn range which can then be
6284 * tried for allocation with alloc_contig_range(). This routine is intended
6285 * for allocation requests which can not be fulfilled with the buddy allocator.
6287 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6288 * power of two, then allocated range is also guaranteed to be aligned to same
6289 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6291 * Allocated pages can be freed with free_contig_range() or by manually calling
6292 * __free_page() on each allocated page.
6294 * Return: pointer to contiguous pages on success, or NULL if not successful.
6296 struct page
*alloc_contig_pages(unsigned long nr_pages
, gfp_t gfp_mask
,
6297 int nid
, nodemask_t
*nodemask
)
6299 unsigned long ret
, pfn
, flags
;
6300 struct zonelist
*zonelist
;
6304 zonelist
= node_zonelist(nid
, gfp_mask
);
6305 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
6306 gfp_zone(gfp_mask
), nodemask
) {
6307 spin_lock_irqsave(&zone
->lock
, flags
);
6309 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
6310 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
6311 if (pfn_range_valid_contig(zone
, pfn
, nr_pages
)) {
6313 * We release the zone lock here because
6314 * alloc_contig_range() will also lock the zone
6315 * at some point. If there's an allocation
6316 * spinning on this lock, it may win the race
6317 * and cause alloc_contig_range() to fail...
6319 spin_unlock_irqrestore(&zone
->lock
, flags
);
6320 ret
= __alloc_contig_pages(pfn
, nr_pages
,
6323 return pfn_to_page(pfn
);
6324 spin_lock_irqsave(&zone
->lock
, flags
);
6328 spin_unlock_irqrestore(&zone
->lock
, flags
);
6332 #endif /* CONFIG_CONTIG_ALLOC */
6334 void free_contig_range(unsigned long pfn
, unsigned long nr_pages
)
6336 unsigned long count
= 0;
6338 for (; nr_pages
--; pfn
++) {
6339 struct page
*page
= pfn_to_page(pfn
);
6341 count
+= page_count(page
) != 1;
6344 WARN(count
!= 0, "%lu pages are still in use!\n", count
);
6346 EXPORT_SYMBOL(free_contig_range
);
6349 * Effectively disable pcplists for the zone by setting the high limit to 0
6350 * and draining all cpus. A concurrent page freeing on another CPU that's about
6351 * to put the page on pcplist will either finish before the drain and the page
6352 * will be drained, or observe the new high limit and skip the pcplist.
6354 * Must be paired with a call to zone_pcp_enable().
6356 void zone_pcp_disable(struct zone
*zone
)
6358 mutex_lock(&pcp_batch_high_lock
);
6359 __zone_set_pageset_high_and_batch(zone
, 0, 1);
6360 __drain_all_pages(zone
, true);
6363 void zone_pcp_enable(struct zone
*zone
)
6365 __zone_set_pageset_high_and_batch(zone
, zone
->pageset_high
, zone
->pageset_batch
);
6366 mutex_unlock(&pcp_batch_high_lock
);
6369 void zone_pcp_reset(struct zone
*zone
)
6372 struct per_cpu_zonestat
*pzstats
;
6374 if (zone
->per_cpu_pageset
!= &boot_pageset
) {
6375 for_each_online_cpu(cpu
) {
6376 pzstats
= per_cpu_ptr(zone
->per_cpu_zonestats
, cpu
);
6377 drain_zonestat(zone
, pzstats
);
6379 free_percpu(zone
->per_cpu_pageset
);
6380 zone
->per_cpu_pageset
= &boot_pageset
;
6381 if (zone
->per_cpu_zonestats
!= &boot_zonestats
) {
6382 free_percpu(zone
->per_cpu_zonestats
);
6383 zone
->per_cpu_zonestats
= &boot_zonestats
;
6388 #ifdef CONFIG_MEMORY_HOTREMOVE
6390 * All pages in the range must be in a single zone, must not contain holes,
6391 * must span full sections, and must be isolated before calling this function.
6393 void __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6395 unsigned long pfn
= start_pfn
;
6399 unsigned long flags
;
6401 offline_mem_sections(pfn
, end_pfn
);
6402 zone
= page_zone(pfn_to_page(pfn
));
6403 spin_lock_irqsave(&zone
->lock
, flags
);
6404 while (pfn
< end_pfn
) {
6405 page
= pfn_to_page(pfn
);
6407 * The HWPoisoned page may be not in buddy system, and
6408 * page_count() is not 0.
6410 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6415 * At this point all remaining PageOffline() pages have a
6416 * reference count of 0 and can simply be skipped.
6418 if (PageOffline(page
)) {
6419 BUG_ON(page_count(page
));
6420 BUG_ON(PageBuddy(page
));
6425 BUG_ON(page_count(page
));
6426 BUG_ON(!PageBuddy(page
));
6427 order
= buddy_order(page
);
6428 del_page_from_free_list(page
, zone
, order
);
6429 pfn
+= (1 << order
);
6431 spin_unlock_irqrestore(&zone
->lock
, flags
);
6436 * This function returns a stable result only if called under zone lock.
6438 bool is_free_buddy_page(struct page
*page
)
6440 unsigned long pfn
= page_to_pfn(page
);
6443 for (order
= 0; order
<= MAX_ORDER
; order
++) {
6444 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6446 if (PageBuddy(page_head
) &&
6447 buddy_order_unsafe(page_head
) >= order
)
6451 return order
<= MAX_ORDER
;
6453 EXPORT_SYMBOL(is_free_buddy_page
);
6455 #ifdef CONFIG_MEMORY_FAILURE
6457 * Break down a higher-order page in sub-pages, and keep our target out of
6460 static void break_down_buddy_pages(struct zone
*zone
, struct page
*page
,
6461 struct page
*target
, int low
, int high
,
6464 unsigned long size
= 1 << high
;
6465 struct page
*current_buddy
, *next_page
;
6467 while (high
> low
) {
6471 if (target
>= &page
[size
]) {
6472 next_page
= page
+ size
;
6473 current_buddy
= page
;
6476 current_buddy
= page
+ size
;
6479 if (set_page_guard(zone
, current_buddy
, high
, migratetype
))
6482 if (current_buddy
!= target
) {
6483 add_to_free_list(current_buddy
, zone
, high
, migratetype
);
6484 set_buddy_order(current_buddy
, high
);
6491 * Take a page that will be marked as poisoned off the buddy allocator.
6493 bool take_page_off_buddy(struct page
*page
)
6495 struct zone
*zone
= page_zone(page
);
6496 unsigned long pfn
= page_to_pfn(page
);
6497 unsigned long flags
;
6501 spin_lock_irqsave(&zone
->lock
, flags
);
6502 for (order
= 0; order
<= MAX_ORDER
; order
++) {
6503 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6504 int page_order
= buddy_order(page_head
);
6506 if (PageBuddy(page_head
) && page_order
>= order
) {
6507 unsigned long pfn_head
= page_to_pfn(page_head
);
6508 int migratetype
= get_pfnblock_migratetype(page_head
,
6511 del_page_from_free_list(page_head
, zone
, page_order
);
6512 break_down_buddy_pages(zone
, page_head
, page
, 0,
6513 page_order
, migratetype
);
6514 SetPageHWPoisonTakenOff(page
);
6515 if (!is_migrate_isolate(migratetype
))
6516 __mod_zone_freepage_state(zone
, -1, migratetype
);
6520 if (page_count(page_head
) > 0)
6523 spin_unlock_irqrestore(&zone
->lock
, flags
);
6528 * Cancel takeoff done by take_page_off_buddy().
6530 bool put_page_back_buddy(struct page
*page
)
6532 struct zone
*zone
= page_zone(page
);
6533 unsigned long pfn
= page_to_pfn(page
);
6534 unsigned long flags
;
6535 int migratetype
= get_pfnblock_migratetype(page
, pfn
);
6538 spin_lock_irqsave(&zone
->lock
, flags
);
6539 if (put_page_testzero(page
)) {
6540 ClearPageHWPoisonTakenOff(page
);
6541 __free_one_page(page
, pfn
, zone
, 0, migratetype
, FPI_NONE
);
6542 if (TestClearPageHWPoison(page
)) {
6546 spin_unlock_irqrestore(&zone
->lock
, flags
);
6552 #ifdef CONFIG_ZONE_DMA
6553 bool has_managed_dma(void)
6555 struct pglist_data
*pgdat
;
6557 for_each_online_pgdat(pgdat
) {
6558 struct zone
*zone
= &pgdat
->node_zones
[ZONE_DMA
];
6560 if (managed_zone(zone
))
6565 #endif /* CONFIG_ZONE_DMA */
6567 #ifdef CONFIG_UNACCEPTED_MEMORY
6569 /* Counts number of zones with unaccepted pages. */
6570 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages
);
6572 static bool lazy_accept
= true;
6574 static int __init
accept_memory_parse(char *p
)
6576 if (!strcmp(p
, "lazy")) {
6579 } else if (!strcmp(p
, "eager")) {
6580 lazy_accept
= false;
6586 early_param("accept_memory", accept_memory_parse
);
6588 static bool page_contains_unaccepted(struct page
*page
, unsigned int order
)
6590 phys_addr_t start
= page_to_phys(page
);
6591 phys_addr_t end
= start
+ (PAGE_SIZE
<< order
);
6593 return range_contains_unaccepted_memory(start
, end
);
6596 static void accept_page(struct page
*page
, unsigned int order
)
6598 phys_addr_t start
= page_to_phys(page
);
6600 accept_memory(start
, start
+ (PAGE_SIZE
<< order
));
6603 static bool try_to_accept_memory_one(struct zone
*zone
)
6605 unsigned long flags
;
6609 if (list_empty(&zone
->unaccepted_pages
))
6612 spin_lock_irqsave(&zone
->lock
, flags
);
6613 page
= list_first_entry_or_null(&zone
->unaccepted_pages
,
6616 spin_unlock_irqrestore(&zone
->lock
, flags
);
6620 list_del(&page
->lru
);
6621 last
= list_empty(&zone
->unaccepted_pages
);
6623 __mod_zone_freepage_state(zone
, -MAX_ORDER_NR_PAGES
, MIGRATE_MOVABLE
);
6624 __mod_zone_page_state(zone
, NR_UNACCEPTED
, -MAX_ORDER_NR_PAGES
);
6625 spin_unlock_irqrestore(&zone
->lock
, flags
);
6627 accept_page(page
, MAX_ORDER
);
6629 __free_pages_ok(page
, MAX_ORDER
, FPI_TO_TAIL
);
6632 static_branch_dec(&zones_with_unaccepted_pages
);
6637 static bool try_to_accept_memory(struct zone
*zone
, unsigned int order
)
6642 /* How much to accept to get to high watermark? */
6643 to_accept
= high_wmark_pages(zone
) -
6644 (zone_page_state(zone
, NR_FREE_PAGES
) -
6645 __zone_watermark_unusable_free(zone
, order
, 0));
6647 /* Accept at least one page */
6649 if (!try_to_accept_memory_one(zone
))
6652 to_accept
-= MAX_ORDER_NR_PAGES
;
6653 } while (to_accept
> 0);
6658 static inline bool has_unaccepted_memory(void)
6660 return static_branch_unlikely(&zones_with_unaccepted_pages
);
6663 static bool __free_unaccepted(struct page
*page
)
6665 struct zone
*zone
= page_zone(page
);
6666 unsigned long flags
;
6672 spin_lock_irqsave(&zone
->lock
, flags
);
6673 first
= list_empty(&zone
->unaccepted_pages
);
6674 list_add_tail(&page
->lru
, &zone
->unaccepted_pages
);
6675 __mod_zone_freepage_state(zone
, MAX_ORDER_NR_PAGES
, MIGRATE_MOVABLE
);
6676 __mod_zone_page_state(zone
, NR_UNACCEPTED
, MAX_ORDER_NR_PAGES
);
6677 spin_unlock_irqrestore(&zone
->lock
, flags
);
6680 static_branch_inc(&zones_with_unaccepted_pages
);
6687 static bool page_contains_unaccepted(struct page
*page
, unsigned int order
)
6692 static void accept_page(struct page
*page
, unsigned int order
)
6696 static bool try_to_accept_memory(struct zone
*zone
, unsigned int order
)
6701 static inline bool has_unaccepted_memory(void)
6706 static bool __free_unaccepted(struct page
*page
)
6712 #endif /* CONFIG_UNACCEPTED_MEMORY */