2 * mm/percpu.c - percpu memory allocator
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 * This file is released under the GPLv2.
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated.
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
30 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
32 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be equal to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
43 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
46 * To use this allocator, arch code should do the followings.
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be
50 * different from the default
52 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
56 #include <linux/bitmap.h>
57 #include <linux/bootmem.h>
58 #include <linux/err.h>
59 #include <linux/list.h>
60 #include <linux/log2.h>
62 #include <linux/module.h>
63 #include <linux/mutex.h>
64 #include <linux/percpu.h>
65 #include <linux/pfn.h>
66 #include <linux/slab.h>
67 #include <linux/spinlock.h>
68 #include <linux/vmalloc.h>
69 #include <linux/workqueue.h>
71 #include <asm/cacheflush.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
76 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
77 #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
80 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
81 #ifndef __addr_to_pcpu_ptr
82 #define __addr_to_pcpu_ptr(addr) \
83 (void __percpu *)((unsigned long)(addr) - \
84 (unsigned long)pcpu_base_addr + \
85 (unsigned long)__per_cpu_start)
87 #ifndef __pcpu_ptr_to_addr
88 #define __pcpu_ptr_to_addr(ptr) \
89 (void __force *)((unsigned long)(ptr) + \
90 (unsigned long)pcpu_base_addr - \
91 (unsigned long)__per_cpu_start)
93 #else /* CONFIG_SMP */
94 /* on UP, it's always identity mapped */
95 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
96 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
97 #endif /* CONFIG_SMP */
100 struct list_head list
; /* linked to pcpu_slot lists */
101 int free_size
; /* free bytes in the chunk */
102 int contig_hint
; /* max contiguous size hint */
103 void *base_addr
; /* base address of this chunk */
104 int map_used
; /* # of map entries used */
105 int map_alloc
; /* # of map entries allocated */
106 int *map
; /* allocation map */
107 void *data
; /* chunk data */
108 bool immutable
; /* no [de]population allowed */
109 unsigned long populated
[]; /* populated bitmap */
112 static int pcpu_unit_pages __read_mostly
;
113 static int pcpu_unit_size __read_mostly
;
114 static int pcpu_nr_units __read_mostly
;
115 static int pcpu_atom_size __read_mostly
;
116 static int pcpu_nr_slots __read_mostly
;
117 static size_t pcpu_chunk_struct_size __read_mostly
;
119 /* cpus with the lowest and highest unit numbers */
120 static unsigned int pcpu_first_unit_cpu __read_mostly
;
121 static unsigned int pcpu_last_unit_cpu __read_mostly
;
123 /* the address of the first chunk which starts with the kernel static area */
124 void *pcpu_base_addr __read_mostly
;
125 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
127 static const int *pcpu_unit_map __read_mostly
; /* cpu -> unit */
128 const unsigned long *pcpu_unit_offsets __read_mostly
; /* cpu -> unit offset */
130 /* group information, used for vm allocation */
131 static int pcpu_nr_groups __read_mostly
;
132 static const unsigned long *pcpu_group_offsets __read_mostly
;
133 static const size_t *pcpu_group_sizes __read_mostly
;
136 * The first chunk which always exists. Note that unlike other
137 * chunks, this one can be allocated and mapped in several different
138 * ways and thus often doesn't live in the vmalloc area.
140 static struct pcpu_chunk
*pcpu_first_chunk
;
143 * Optional reserved chunk. This chunk reserves part of the first
144 * chunk and serves it for reserved allocations. The amount of
145 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
146 * area doesn't exist, the following variables contain NULL and 0
149 static struct pcpu_chunk
*pcpu_reserved_chunk
;
150 static int pcpu_reserved_chunk_limit
;
153 * Synchronization rules.
155 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
156 * protects allocation/reclaim paths, chunks, populated bitmap and
157 * vmalloc mapping. The latter is a spinlock and protects the index
158 * data structures - chunk slots, chunks and area maps in chunks.
160 * During allocation, pcpu_alloc_mutex is kept locked all the time and
161 * pcpu_lock is grabbed and released as necessary. All actual memory
162 * allocations are done using GFP_KERNEL with pcpu_lock released. In
163 * general, percpu memory can't be allocated with irq off but
164 * irqsave/restore are still used in alloc path so that it can be used
165 * from early init path - sched_init() specifically.
167 * Free path accesses and alters only the index data structures, so it
168 * can be safely called from atomic context. When memory needs to be
169 * returned to the system, free path schedules reclaim_work which
170 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
171 * reclaimed, release both locks and frees the chunks. Note that it's
172 * necessary to grab both locks to remove a chunk from circulation as
173 * allocation path might be referencing the chunk with only
174 * pcpu_alloc_mutex locked.
176 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* protects whole alloc and reclaim */
177 static DEFINE_SPINLOCK(pcpu_lock
); /* protects index data structures */
179 static struct list_head
*pcpu_slot __read_mostly
; /* chunk list slots */
181 /* reclaim work to release fully free chunks, scheduled from free path */
182 static void pcpu_reclaim(struct work_struct
*work
);
183 static DECLARE_WORK(pcpu_reclaim_work
, pcpu_reclaim
);
185 static bool pcpu_addr_in_first_chunk(void *addr
)
187 void *first_start
= pcpu_first_chunk
->base_addr
;
189 return addr
>= first_start
&& addr
< first_start
+ pcpu_unit_size
;
192 static bool pcpu_addr_in_reserved_chunk(void *addr
)
194 void *first_start
= pcpu_first_chunk
->base_addr
;
196 return addr
>= first_start
&&
197 addr
< first_start
+ pcpu_reserved_chunk_limit
;
200 static int __pcpu_size_to_slot(int size
)
202 int highbit
= fls(size
); /* size is in bytes */
203 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
206 static int pcpu_size_to_slot(int size
)
208 if (size
== pcpu_unit_size
)
209 return pcpu_nr_slots
- 1;
210 return __pcpu_size_to_slot(size
);
213 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
215 if (chunk
->free_size
< sizeof(int) || chunk
->contig_hint
< sizeof(int))
218 return pcpu_size_to_slot(chunk
->free_size
);
221 /* set the pointer to a chunk in a page struct */
222 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
224 page
->index
= (unsigned long)pcpu
;
227 /* obtain pointer to a chunk from a page struct */
228 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
230 return (struct pcpu_chunk
*)page
->index
;
233 static int __maybe_unused
pcpu_page_idx(unsigned int cpu
, int page_idx
)
235 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
238 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
239 unsigned int cpu
, int page_idx
)
241 return (unsigned long)chunk
->base_addr
+ pcpu_unit_offsets
[cpu
] +
242 (page_idx
<< PAGE_SHIFT
);
245 static void __maybe_unused
pcpu_next_unpop(struct pcpu_chunk
*chunk
,
246 int *rs
, int *re
, int end
)
248 *rs
= find_next_zero_bit(chunk
->populated
, end
, *rs
);
249 *re
= find_next_bit(chunk
->populated
, end
, *rs
+ 1);
252 static void __maybe_unused
pcpu_next_pop(struct pcpu_chunk
*chunk
,
253 int *rs
, int *re
, int end
)
255 *rs
= find_next_bit(chunk
->populated
, end
, *rs
);
256 *re
= find_next_zero_bit(chunk
->populated
, end
, *rs
+ 1);
260 * (Un)populated page region iterators. Iterate over (un)populated
261 * page regions between @start and @end in @chunk. @rs and @re should
262 * be integer variables and will be set to start and end page index of
263 * the current region.
265 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
266 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
268 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
270 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
271 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
273 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
276 * pcpu_mem_alloc - allocate memory
277 * @size: bytes to allocate
279 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
280 * kzalloc() is used; otherwise, vmalloc() is used. The returned
281 * memory is always zeroed.
284 * Does GFP_KERNEL allocation.
287 * Pointer to the allocated area on success, NULL on failure.
289 static void *pcpu_mem_alloc(size_t size
)
291 if (WARN_ON_ONCE(!slab_is_available()))
294 if (size
<= PAGE_SIZE
)
295 return kzalloc(size
, GFP_KERNEL
);
297 return vzalloc(size
);
301 * pcpu_mem_free - free memory
302 * @ptr: memory to free
303 * @size: size of the area
305 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
307 static void pcpu_mem_free(void *ptr
, size_t size
)
309 if (size
<= PAGE_SIZE
)
316 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
317 * @chunk: chunk of interest
318 * @oslot: the previous slot it was on
320 * This function is called after an allocation or free changed @chunk.
321 * New slot according to the changed state is determined and @chunk is
322 * moved to the slot. Note that the reserved chunk is never put on
328 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
330 int nslot
= pcpu_chunk_slot(chunk
);
332 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
334 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
336 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
341 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
342 * @chunk: chunk of interest
344 * Determine whether area map of @chunk needs to be extended to
345 * accomodate a new allocation.
351 * New target map allocation length if extension is necessary, 0
354 static int pcpu_need_to_extend(struct pcpu_chunk
*chunk
)
358 if (chunk
->map_alloc
>= chunk
->map_used
+ 2)
361 new_alloc
= PCPU_DFL_MAP_ALLOC
;
362 while (new_alloc
< chunk
->map_used
+ 2)
369 * pcpu_extend_area_map - extend area map of a chunk
370 * @chunk: chunk of interest
371 * @new_alloc: new target allocation length of the area map
373 * Extend area map of @chunk to have @new_alloc entries.
376 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
379 * 0 on success, -errno on failure.
381 static int pcpu_extend_area_map(struct pcpu_chunk
*chunk
, int new_alloc
)
383 int *old
= NULL
, *new = NULL
;
384 size_t old_size
= 0, new_size
= new_alloc
* sizeof(new[0]);
387 new = pcpu_mem_alloc(new_size
);
391 /* acquire pcpu_lock and switch to new area map */
392 spin_lock_irqsave(&pcpu_lock
, flags
);
394 if (new_alloc
<= chunk
->map_alloc
)
397 old_size
= chunk
->map_alloc
* sizeof(chunk
->map
[0]);
400 memcpy(new, old
, old_size
);
402 chunk
->map_alloc
= new_alloc
;
407 spin_unlock_irqrestore(&pcpu_lock
, flags
);
410 * pcpu_mem_free() might end up calling vfree() which uses
411 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
413 pcpu_mem_free(old
, old_size
);
414 pcpu_mem_free(new, new_size
);
420 * pcpu_split_block - split a map block
421 * @chunk: chunk of interest
422 * @i: index of map block to split
423 * @head: head size in bytes (can be 0)
424 * @tail: tail size in bytes (can be 0)
426 * Split the @i'th map block into two or three blocks. If @head is
427 * non-zero, @head bytes block is inserted before block @i moving it
428 * to @i+1 and reducing its size by @head bytes.
430 * If @tail is non-zero, the target block, which can be @i or @i+1
431 * depending on @head, is reduced by @tail bytes and @tail byte block
432 * is inserted after the target block.
434 * @chunk->map must have enough free slots to accomodate the split.
439 static void pcpu_split_block(struct pcpu_chunk
*chunk
, int i
,
442 int nr_extra
= !!head
+ !!tail
;
444 BUG_ON(chunk
->map_alloc
< chunk
->map_used
+ nr_extra
);
446 /* insert new subblocks */
447 memmove(&chunk
->map
[i
+ nr_extra
], &chunk
->map
[i
],
448 sizeof(chunk
->map
[0]) * (chunk
->map_used
- i
));
449 chunk
->map_used
+= nr_extra
;
452 chunk
->map
[i
+ 1] = chunk
->map
[i
] - head
;
453 chunk
->map
[i
++] = head
;
456 chunk
->map
[i
++] -= tail
;
457 chunk
->map
[i
] = tail
;
462 * pcpu_alloc_area - allocate area from a pcpu_chunk
463 * @chunk: chunk of interest
464 * @size: wanted size in bytes
465 * @align: wanted align
467 * Try to allocate @size bytes area aligned at @align from @chunk.
468 * Note that this function only allocates the offset. It doesn't
469 * populate or map the area.
471 * @chunk->map must have at least two free slots.
477 * Allocated offset in @chunk on success, -1 if no matching area is
480 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int size
, int align
)
482 int oslot
= pcpu_chunk_slot(chunk
);
486 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++])) {
487 bool is_last
= i
+ 1 == chunk
->map_used
;
490 /* extra for alignment requirement */
491 head
= ALIGN(off
, align
) - off
;
492 BUG_ON(i
== 0 && head
!= 0);
494 if (chunk
->map
[i
] < 0)
496 if (chunk
->map
[i
] < head
+ size
) {
497 max_contig
= max(chunk
->map
[i
], max_contig
);
502 * If head is small or the previous block is free,
503 * merge'em. Note that 'small' is defined as smaller
504 * than sizeof(int), which is very small but isn't too
505 * uncommon for percpu allocations.
507 if (head
&& (head
< sizeof(int) || chunk
->map
[i
- 1] > 0)) {
508 if (chunk
->map
[i
- 1] > 0)
509 chunk
->map
[i
- 1] += head
;
511 chunk
->map
[i
- 1] -= head
;
512 chunk
->free_size
-= head
;
514 chunk
->map
[i
] -= head
;
519 /* if tail is small, just keep it around */
520 tail
= chunk
->map
[i
] - head
- size
;
521 if (tail
< sizeof(int))
524 /* split if warranted */
526 pcpu_split_block(chunk
, i
, head
, tail
);
530 max_contig
= max(chunk
->map
[i
- 1], max_contig
);
533 max_contig
= max(chunk
->map
[i
+ 1], max_contig
);
536 /* update hint and mark allocated */
538 chunk
->contig_hint
= max_contig
; /* fully scanned */
540 chunk
->contig_hint
= max(chunk
->contig_hint
,
543 chunk
->free_size
-= chunk
->map
[i
];
544 chunk
->map
[i
] = -chunk
->map
[i
];
546 pcpu_chunk_relocate(chunk
, oslot
);
550 chunk
->contig_hint
= max_contig
; /* fully scanned */
551 pcpu_chunk_relocate(chunk
, oslot
);
553 /* tell the upper layer that this chunk has no matching area */
558 * pcpu_free_area - free area to a pcpu_chunk
559 * @chunk: chunk of interest
560 * @freeme: offset of area to free
562 * Free area starting from @freeme to @chunk. Note that this function
563 * only modifies the allocation map. It doesn't depopulate or unmap
569 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int freeme
)
571 int oslot
= pcpu_chunk_slot(chunk
);
574 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++]))
577 BUG_ON(off
!= freeme
);
578 BUG_ON(chunk
->map
[i
] > 0);
580 chunk
->map
[i
] = -chunk
->map
[i
];
581 chunk
->free_size
+= chunk
->map
[i
];
583 /* merge with previous? */
584 if (i
> 0 && chunk
->map
[i
- 1] >= 0) {
585 chunk
->map
[i
- 1] += chunk
->map
[i
];
587 memmove(&chunk
->map
[i
], &chunk
->map
[i
+ 1],
588 (chunk
->map_used
- i
) * sizeof(chunk
->map
[0]));
591 /* merge with next? */
592 if (i
+ 1 < chunk
->map_used
&& chunk
->map
[i
+ 1] >= 0) {
593 chunk
->map
[i
] += chunk
->map
[i
+ 1];
595 memmove(&chunk
->map
[i
+ 1], &chunk
->map
[i
+ 2],
596 (chunk
->map_used
- (i
+ 1)) * sizeof(chunk
->map
[0]));
599 chunk
->contig_hint
= max(chunk
->map
[i
], chunk
->contig_hint
);
600 pcpu_chunk_relocate(chunk
, oslot
);
603 static struct pcpu_chunk
*pcpu_alloc_chunk(void)
605 struct pcpu_chunk
*chunk
;
607 chunk
= pcpu_mem_alloc(pcpu_chunk_struct_size
);
611 chunk
->map
= pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC
* sizeof(chunk
->map
[0]));
617 chunk
->map_alloc
= PCPU_DFL_MAP_ALLOC
;
618 chunk
->map
[chunk
->map_used
++] = pcpu_unit_size
;
620 INIT_LIST_HEAD(&chunk
->list
);
621 chunk
->free_size
= pcpu_unit_size
;
622 chunk
->contig_hint
= pcpu_unit_size
;
627 static void pcpu_free_chunk(struct pcpu_chunk
*chunk
)
631 pcpu_mem_free(chunk
->map
, chunk
->map_alloc
* sizeof(chunk
->map
[0]));
636 * Chunk management implementation.
638 * To allow different implementations, chunk alloc/free and
639 * [de]population are implemented in a separate file which is pulled
640 * into this file and compiled together. The following functions
641 * should be implemented.
643 * pcpu_populate_chunk - populate the specified range of a chunk
644 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
645 * pcpu_create_chunk - create a new chunk
646 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
647 * pcpu_addr_to_page - translate address to physical address
648 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
650 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
651 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
652 static struct pcpu_chunk
*pcpu_create_chunk(void);
653 static void pcpu_destroy_chunk(struct pcpu_chunk
*chunk
);
654 static struct page
*pcpu_addr_to_page(void *addr
);
655 static int __init
pcpu_verify_alloc_info(const struct pcpu_alloc_info
*ai
);
657 #ifdef CONFIG_NEED_PER_CPU_KM
658 #include "percpu-km.c"
660 #include "percpu-vm.c"
664 * pcpu_chunk_addr_search - determine chunk containing specified address
665 * @addr: address for which the chunk needs to be determined.
668 * The address of the found chunk.
670 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
672 /* is it in the first chunk? */
673 if (pcpu_addr_in_first_chunk(addr
)) {
674 /* is it in the reserved area? */
675 if (pcpu_addr_in_reserved_chunk(addr
))
676 return pcpu_reserved_chunk
;
677 return pcpu_first_chunk
;
681 * The address is relative to unit0 which might be unused and
682 * thus unmapped. Offset the address to the unit space of the
683 * current processor before looking it up in the vmalloc
684 * space. Note that any possible cpu id can be used here, so
685 * there's no need to worry about preemption or cpu hotplug.
687 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
688 return pcpu_get_page_chunk(pcpu_addr_to_page(addr
));
692 * pcpu_alloc - the percpu allocator
693 * @size: size of area to allocate in bytes
694 * @align: alignment of area (max PAGE_SIZE)
695 * @reserved: allocate from the reserved chunk if available
697 * Allocate percpu area of @size bytes aligned at @align.
700 * Does GFP_KERNEL allocation.
703 * Percpu pointer to the allocated area on success, NULL on failure.
705 static void __percpu
*pcpu_alloc(size_t size
, size_t align
, bool reserved
)
707 static int warn_limit
= 10;
708 struct pcpu_chunk
*chunk
;
710 int slot
, off
, new_alloc
;
713 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
)) {
714 WARN(true, "illegal size (%zu) or align (%zu) for "
715 "percpu allocation\n", size
, align
);
719 mutex_lock(&pcpu_alloc_mutex
);
720 spin_lock_irqsave(&pcpu_lock
, flags
);
722 /* serve reserved allocations from the reserved chunk if available */
723 if (reserved
&& pcpu_reserved_chunk
) {
724 chunk
= pcpu_reserved_chunk
;
726 if (size
> chunk
->contig_hint
) {
727 err
= "alloc from reserved chunk failed";
731 while ((new_alloc
= pcpu_need_to_extend(chunk
))) {
732 spin_unlock_irqrestore(&pcpu_lock
, flags
);
733 if (pcpu_extend_area_map(chunk
, new_alloc
) < 0) {
734 err
= "failed to extend area map of reserved chunk";
735 goto fail_unlock_mutex
;
737 spin_lock_irqsave(&pcpu_lock
, flags
);
740 off
= pcpu_alloc_area(chunk
, size
, align
);
744 err
= "alloc from reserved chunk failed";
749 /* search through normal chunks */
750 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
751 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
752 if (size
> chunk
->contig_hint
)
755 new_alloc
= pcpu_need_to_extend(chunk
);
757 spin_unlock_irqrestore(&pcpu_lock
, flags
);
758 if (pcpu_extend_area_map(chunk
,
760 err
= "failed to extend area map";
761 goto fail_unlock_mutex
;
763 spin_lock_irqsave(&pcpu_lock
, flags
);
765 * pcpu_lock has been dropped, need to
766 * restart cpu_slot list walking.
771 off
= pcpu_alloc_area(chunk
, size
, align
);
777 /* hmmm... no space left, create a new chunk */
778 spin_unlock_irqrestore(&pcpu_lock
, flags
);
780 chunk
= pcpu_create_chunk();
782 err
= "failed to allocate new chunk";
783 goto fail_unlock_mutex
;
786 spin_lock_irqsave(&pcpu_lock
, flags
);
787 pcpu_chunk_relocate(chunk
, -1);
791 spin_unlock_irqrestore(&pcpu_lock
, flags
);
793 /* populate, map and clear the area */
794 if (pcpu_populate_chunk(chunk
, off
, size
)) {
795 spin_lock_irqsave(&pcpu_lock
, flags
);
796 pcpu_free_area(chunk
, off
);
797 err
= "failed to populate";
801 mutex_unlock(&pcpu_alloc_mutex
);
803 /* return address relative to base address */
804 return __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
807 spin_unlock_irqrestore(&pcpu_lock
, flags
);
809 mutex_unlock(&pcpu_alloc_mutex
);
811 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
812 "%s\n", size
, align
, err
);
815 pr_info("PERCPU: limit reached, disable warning\n");
821 * __alloc_percpu - allocate dynamic percpu area
822 * @size: size of area to allocate in bytes
823 * @align: alignment of area (max PAGE_SIZE)
825 * Allocate zero-filled percpu area of @size bytes aligned at @align.
826 * Might sleep. Might trigger writeouts.
829 * Does GFP_KERNEL allocation.
832 * Percpu pointer to the allocated area on success, NULL on failure.
834 void __percpu
*__alloc_percpu(size_t size
, size_t align
)
836 return pcpu_alloc(size
, align
, false);
838 EXPORT_SYMBOL_GPL(__alloc_percpu
);
841 * __alloc_reserved_percpu - allocate reserved percpu area
842 * @size: size of area to allocate in bytes
843 * @align: alignment of area (max PAGE_SIZE)
845 * Allocate zero-filled percpu area of @size bytes aligned at @align
846 * from reserved percpu area if arch has set it up; otherwise,
847 * allocation is served from the same dynamic area. Might sleep.
848 * Might trigger writeouts.
851 * Does GFP_KERNEL allocation.
854 * Percpu pointer to the allocated area on success, NULL on failure.
856 void __percpu
*__alloc_reserved_percpu(size_t size
, size_t align
)
858 return pcpu_alloc(size
, align
, true);
862 * pcpu_reclaim - reclaim fully free chunks, workqueue function
865 * Reclaim all fully free chunks except for the first one.
870 static void pcpu_reclaim(struct work_struct
*work
)
873 struct list_head
*head
= &pcpu_slot
[pcpu_nr_slots
- 1];
874 struct pcpu_chunk
*chunk
, *next
;
876 mutex_lock(&pcpu_alloc_mutex
);
877 spin_lock_irq(&pcpu_lock
);
879 list_for_each_entry_safe(chunk
, next
, head
, list
) {
880 WARN_ON(chunk
->immutable
);
882 /* spare the first one */
883 if (chunk
== list_first_entry(head
, struct pcpu_chunk
, list
))
886 list_move(&chunk
->list
, &todo
);
889 spin_unlock_irq(&pcpu_lock
);
891 list_for_each_entry_safe(chunk
, next
, &todo
, list
) {
892 pcpu_depopulate_chunk(chunk
, 0, pcpu_unit_size
);
893 pcpu_destroy_chunk(chunk
);
896 mutex_unlock(&pcpu_alloc_mutex
);
900 * free_percpu - free percpu area
901 * @ptr: pointer to area to free
903 * Free percpu area @ptr.
906 * Can be called from atomic context.
908 void free_percpu(void __percpu
*ptr
)
911 struct pcpu_chunk
*chunk
;
918 addr
= __pcpu_ptr_to_addr(ptr
);
920 spin_lock_irqsave(&pcpu_lock
, flags
);
922 chunk
= pcpu_chunk_addr_search(addr
);
923 off
= addr
- chunk
->base_addr
;
925 pcpu_free_area(chunk
, off
);
927 /* if there are more than one fully free chunks, wake up grim reaper */
928 if (chunk
->free_size
== pcpu_unit_size
) {
929 struct pcpu_chunk
*pos
;
931 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
933 schedule_work(&pcpu_reclaim_work
);
938 spin_unlock_irqrestore(&pcpu_lock
, flags
);
940 EXPORT_SYMBOL_GPL(free_percpu
);
943 * is_kernel_percpu_address - test whether address is from static percpu area
944 * @addr: address to test
946 * Test whether @addr belongs to in-kernel static percpu area. Module
947 * static percpu areas are not considered. For those, use
948 * is_module_percpu_address().
951 * %true if @addr is from in-kernel static percpu area, %false otherwise.
953 bool is_kernel_percpu_address(unsigned long addr
)
956 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
957 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
960 for_each_possible_cpu(cpu
) {
961 void *start
= per_cpu_ptr(base
, cpu
);
963 if ((void *)addr
>= start
&& (void *)addr
< start
+ static_size
)
967 /* on UP, can't distinguish from other static vars, always false */
972 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
973 * @addr: the address to be converted to physical address
975 * Given @addr which is dereferenceable address obtained via one of
976 * percpu access macros, this function translates it into its physical
977 * address. The caller is responsible for ensuring @addr stays valid
978 * until this function finishes.
981 * The physical address for @addr.
983 phys_addr_t
per_cpu_ptr_to_phys(void *addr
)
985 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
986 bool in_first_chunk
= false;
987 unsigned long first_start
, first_end
;
991 * The following test on first_start/end isn't strictly
992 * necessary but will speed up lookups of addresses which
993 * aren't in the first chunk.
995 first_start
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_first_unit_cpu
, 0);
996 first_end
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_last_unit_cpu
,
998 if ((unsigned long)addr
>= first_start
&&
999 (unsigned long)addr
< first_end
) {
1000 for_each_possible_cpu(cpu
) {
1001 void *start
= per_cpu_ptr(base
, cpu
);
1003 if (addr
>= start
&& addr
< start
+ pcpu_unit_size
) {
1004 in_first_chunk
= true;
1010 if (in_first_chunk
) {
1011 if (!is_vmalloc_addr(addr
))
1014 return page_to_phys(vmalloc_to_page(addr
));
1016 return page_to_phys(pcpu_addr_to_page(addr
));
1020 * pcpu_alloc_alloc_info - allocate percpu allocation info
1021 * @nr_groups: the number of groups
1022 * @nr_units: the number of units
1024 * Allocate ai which is large enough for @nr_groups groups containing
1025 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1026 * cpu_map array which is long enough for @nr_units and filled with
1027 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1028 * pointer of other groups.
1031 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1034 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
1037 struct pcpu_alloc_info
*ai
;
1038 size_t base_size
, ai_size
;
1042 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
1043 __alignof__(ai
->groups
[0].cpu_map
[0]));
1044 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
1046 ptr
= alloc_bootmem_nopanic(PFN_ALIGN(ai_size
));
1052 ai
->groups
[0].cpu_map
= ptr
;
1054 for (unit
= 0; unit
< nr_units
; unit
++)
1055 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
1057 ai
->nr_groups
= nr_groups
;
1058 ai
->__ai_size
= PFN_ALIGN(ai_size
);
1064 * pcpu_free_alloc_info - free percpu allocation info
1065 * @ai: pcpu_alloc_info to free
1067 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1069 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
1071 free_bootmem(__pa(ai
), ai
->__ai_size
);
1075 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1077 * @ai: allocation info to dump
1079 * Print out information about @ai using loglevel @lvl.
1081 static void pcpu_dump_alloc_info(const char *lvl
,
1082 const struct pcpu_alloc_info
*ai
)
1084 int group_width
= 1, cpu_width
= 1, width
;
1085 char empty_str
[] = "--------";
1086 int alloc
= 0, alloc_end
= 0;
1088 int upa
, apl
; /* units per alloc, allocs per line */
1094 v
= num_possible_cpus();
1097 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
1099 upa
= ai
->alloc_size
/ ai
->unit_size
;
1100 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
1101 apl
= rounddown_pow_of_two(max(60 / width
, 1));
1103 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1104 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
1105 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
1107 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1108 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1109 int unit
= 0, unit_end
= 0;
1111 BUG_ON(gi
->nr_units
% upa
);
1112 for (alloc_end
+= gi
->nr_units
/ upa
;
1113 alloc
< alloc_end
; alloc
++) {
1114 if (!(alloc
% apl
)) {
1116 printk("%spcpu-alloc: ", lvl
);
1118 printk("[%0*d] ", group_width
, group
);
1120 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
1121 if (gi
->cpu_map
[unit
] != NR_CPUS
)
1122 printk("%0*d ", cpu_width
,
1125 printk("%s ", empty_str
);
1132 * pcpu_setup_first_chunk - initialize the first percpu chunk
1133 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1134 * @base_addr: mapped address
1136 * Initialize the first percpu chunk which contains the kernel static
1137 * perpcu area. This function is to be called from arch percpu area
1140 * @ai contains all information necessary to initialize the first
1141 * chunk and prime the dynamic percpu allocator.
1143 * @ai->static_size is the size of static percpu area.
1145 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1146 * reserve after the static area in the first chunk. This reserves
1147 * the first chunk such that it's available only through reserved
1148 * percpu allocation. This is primarily used to serve module percpu
1149 * static areas on architectures where the addressing model has
1150 * limited offset range for symbol relocations to guarantee module
1151 * percpu symbols fall inside the relocatable range.
1153 * @ai->dyn_size determines the number of bytes available for dynamic
1154 * allocation in the first chunk. The area between @ai->static_size +
1155 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1157 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1158 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1161 * @ai->atom_size is the allocation atom size and used as alignment
1164 * @ai->alloc_size is the allocation size and always multiple of
1165 * @ai->atom_size. This is larger than @ai->atom_size if
1166 * @ai->unit_size is larger than @ai->atom_size.
1168 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1169 * percpu areas. Units which should be colocated are put into the
1170 * same group. Dynamic VM areas will be allocated according to these
1171 * groupings. If @ai->nr_groups is zero, a single group containing
1172 * all units is assumed.
1174 * The caller should have mapped the first chunk at @base_addr and
1175 * copied static data to each unit.
1177 * If the first chunk ends up with both reserved and dynamic areas, it
1178 * is served by two chunks - one to serve the core static and reserved
1179 * areas and the other for the dynamic area. They share the same vm
1180 * and page map but uses different area allocation map to stay away
1181 * from each other. The latter chunk is circulated in the chunk slots
1182 * and available for dynamic allocation like any other chunks.
1185 * 0 on success, -errno on failure.
1187 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
1190 static char cpus_buf
[4096] __initdata
;
1191 static int smap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1192 static int dmap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1193 size_t dyn_size
= ai
->dyn_size
;
1194 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ dyn_size
;
1195 struct pcpu_chunk
*schunk
, *dchunk
= NULL
;
1196 unsigned long *group_offsets
;
1197 size_t *group_sizes
;
1198 unsigned long *unit_off
;
1203 cpumask_scnprintf(cpus_buf
, sizeof(cpus_buf
), cpu_possible_mask
);
1205 #define PCPU_SETUP_BUG_ON(cond) do { \
1206 if (unlikely(cond)) { \
1207 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1208 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1209 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1215 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
1217 PCPU_SETUP_BUG_ON(!ai
->static_size
);
1219 PCPU_SETUP_BUG_ON(!base_addr
);
1220 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
1221 PCPU_SETUP_BUG_ON(ai
->unit_size
& ~PAGE_MASK
);
1222 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
1223 PCPU_SETUP_BUG_ON(ai
->dyn_size
< PERCPU_DYNAMIC_EARLY_SIZE
);
1224 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai
) < 0);
1226 /* process group information and build config tables accordingly */
1227 group_offsets
= alloc_bootmem(ai
->nr_groups
* sizeof(group_offsets
[0]));
1228 group_sizes
= alloc_bootmem(ai
->nr_groups
* sizeof(group_sizes
[0]));
1229 unit_map
= alloc_bootmem(nr_cpu_ids
* sizeof(unit_map
[0]));
1230 unit_off
= alloc_bootmem(nr_cpu_ids
* sizeof(unit_off
[0]));
1232 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
1233 unit_map
[cpu
] = UINT_MAX
;
1234 pcpu_first_unit_cpu
= NR_CPUS
;
1236 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
1237 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1239 group_offsets
[group
] = gi
->base_offset
;
1240 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
1242 for (i
= 0; i
< gi
->nr_units
; i
++) {
1243 cpu
= gi
->cpu_map
[i
];
1247 PCPU_SETUP_BUG_ON(cpu
> nr_cpu_ids
);
1248 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
1249 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
1251 unit_map
[cpu
] = unit
+ i
;
1252 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
1254 if (pcpu_first_unit_cpu
== NR_CPUS
)
1255 pcpu_first_unit_cpu
= cpu
;
1256 pcpu_last_unit_cpu
= cpu
;
1259 pcpu_nr_units
= unit
;
1261 for_each_possible_cpu(cpu
)
1262 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
1264 /* we're done parsing the input, undefine BUG macro and dump config */
1265 #undef PCPU_SETUP_BUG_ON
1266 pcpu_dump_alloc_info(KERN_DEBUG
, ai
);
1268 pcpu_nr_groups
= ai
->nr_groups
;
1269 pcpu_group_offsets
= group_offsets
;
1270 pcpu_group_sizes
= group_sizes
;
1271 pcpu_unit_map
= unit_map
;
1272 pcpu_unit_offsets
= unit_off
;
1274 /* determine basic parameters */
1275 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1276 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
1277 pcpu_atom_size
= ai
->atom_size
;
1278 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
1279 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
1282 * Allocate chunk slots. The additional last slot is for
1285 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
1286 pcpu_slot
= alloc_bootmem(pcpu_nr_slots
* sizeof(pcpu_slot
[0]));
1287 for (i
= 0; i
< pcpu_nr_slots
; i
++)
1288 INIT_LIST_HEAD(&pcpu_slot
[i
]);
1291 * Initialize static chunk. If reserved_size is zero, the
1292 * static chunk covers static area + dynamic allocation area
1293 * in the first chunk. If reserved_size is not zero, it
1294 * covers static area + reserved area (mostly used for module
1295 * static percpu allocation).
1297 schunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1298 INIT_LIST_HEAD(&schunk
->list
);
1299 schunk
->base_addr
= base_addr
;
1301 schunk
->map_alloc
= ARRAY_SIZE(smap
);
1302 schunk
->immutable
= true;
1303 bitmap_fill(schunk
->populated
, pcpu_unit_pages
);
1305 if (ai
->reserved_size
) {
1306 schunk
->free_size
= ai
->reserved_size
;
1307 pcpu_reserved_chunk
= schunk
;
1308 pcpu_reserved_chunk_limit
= ai
->static_size
+ ai
->reserved_size
;
1310 schunk
->free_size
= dyn_size
;
1311 dyn_size
= 0; /* dynamic area covered */
1313 schunk
->contig_hint
= schunk
->free_size
;
1315 schunk
->map
[schunk
->map_used
++] = -ai
->static_size
;
1316 if (schunk
->free_size
)
1317 schunk
->map
[schunk
->map_used
++] = schunk
->free_size
;
1319 /* init dynamic chunk if necessary */
1321 dchunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1322 INIT_LIST_HEAD(&dchunk
->list
);
1323 dchunk
->base_addr
= base_addr
;
1325 dchunk
->map_alloc
= ARRAY_SIZE(dmap
);
1326 dchunk
->immutable
= true;
1327 bitmap_fill(dchunk
->populated
, pcpu_unit_pages
);
1329 dchunk
->contig_hint
= dchunk
->free_size
= dyn_size
;
1330 dchunk
->map
[dchunk
->map_used
++] = -pcpu_reserved_chunk_limit
;
1331 dchunk
->map
[dchunk
->map_used
++] = dchunk
->free_size
;
1334 /* link the first chunk in */
1335 pcpu_first_chunk
= dchunk
?: schunk
;
1336 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
1339 pcpu_base_addr
= base_addr
;
1345 const char *pcpu_fc_names
[PCPU_FC_NR
] __initdata
= {
1346 [PCPU_FC_AUTO
] = "auto",
1347 [PCPU_FC_EMBED
] = "embed",
1348 [PCPU_FC_PAGE
] = "page",
1351 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
1353 static int __init
percpu_alloc_setup(char *str
)
1357 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1358 else if (!strcmp(str
, "embed"))
1359 pcpu_chosen_fc
= PCPU_FC_EMBED
;
1361 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1362 else if (!strcmp(str
, "page"))
1363 pcpu_chosen_fc
= PCPU_FC_PAGE
;
1366 pr_warning("PERCPU: unknown allocator %s specified\n", str
);
1370 early_param("percpu_alloc", percpu_alloc_setup
);
1373 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1374 * Build it if needed by the arch config or the generic setup is going
1377 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1378 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1379 #define BUILD_EMBED_FIRST_CHUNK
1382 /* build pcpu_page_first_chunk() iff needed by the arch config */
1383 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1384 #define BUILD_PAGE_FIRST_CHUNK
1387 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1388 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1390 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1391 * @reserved_size: the size of reserved percpu area in bytes
1392 * @dyn_size: minimum free size for dynamic allocation in bytes
1393 * @atom_size: allocation atom size
1394 * @cpu_distance_fn: callback to determine distance between cpus, optional
1396 * This function determines grouping of units, their mappings to cpus
1397 * and other parameters considering needed percpu size, allocation
1398 * atom size and distances between CPUs.
1400 * Groups are always mutliples of atom size and CPUs which are of
1401 * LOCAL_DISTANCE both ways are grouped together and share space for
1402 * units in the same group. The returned configuration is guaranteed
1403 * to have CPUs on different nodes on different groups and >=75% usage
1404 * of allocated virtual address space.
1407 * On success, pointer to the new allocation_info is returned. On
1408 * failure, ERR_PTR value is returned.
1410 static struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
1411 size_t reserved_size
, size_t dyn_size
,
1413 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
1415 static int group_map
[NR_CPUS
] __initdata
;
1416 static int group_cnt
[NR_CPUS
] __initdata
;
1417 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1418 int nr_groups
= 1, nr_units
= 0;
1419 size_t size_sum
, min_unit_size
, alloc_size
;
1420 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
1421 int last_allocs
, group
, unit
;
1422 unsigned int cpu
, tcpu
;
1423 struct pcpu_alloc_info
*ai
;
1424 unsigned int *cpu_map
;
1426 /* this function may be called multiple times */
1427 memset(group_map
, 0, sizeof(group_map
));
1428 memset(group_cnt
, 0, sizeof(group_cnt
));
1430 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1431 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
1432 max_t(size_t, dyn_size
, PERCPU_DYNAMIC_EARLY_SIZE
));
1433 dyn_size
= size_sum
- static_size
- reserved_size
;
1436 * Determine min_unit_size, alloc_size and max_upa such that
1437 * alloc_size is multiple of atom_size and is the smallest
1438 * which can accomodate 4k aligned segments which are equal to
1439 * or larger than min_unit_size.
1441 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1443 alloc_size
= roundup(min_unit_size
, atom_size
);
1444 upa
= alloc_size
/ min_unit_size
;
1445 while (alloc_size
% upa
|| ((alloc_size
/ upa
) & ~PAGE_MASK
))
1449 /* group cpus according to their proximity */
1450 for_each_possible_cpu(cpu
) {
1453 for_each_possible_cpu(tcpu
) {
1456 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
1457 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
1458 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
1460 nr_groups
= max(nr_groups
, group
+ 1);
1464 group_map
[cpu
] = group
;
1469 * Expand unit size until address space usage goes over 75%
1470 * and then as much as possible without using more address
1473 last_allocs
= INT_MAX
;
1474 for (upa
= max_upa
; upa
; upa
--) {
1475 int allocs
= 0, wasted
= 0;
1477 if (alloc_size
% upa
|| ((alloc_size
/ upa
) & ~PAGE_MASK
))
1480 for (group
= 0; group
< nr_groups
; group
++) {
1481 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
1482 allocs
+= this_allocs
;
1483 wasted
+= this_allocs
* upa
- group_cnt
[group
];
1487 * Don't accept if wastage is over 1/3. The
1488 * greater-than comparison ensures upa==1 always
1489 * passes the following check.
1491 if (wasted
> num_possible_cpus() / 3)
1494 /* and then don't consume more memory */
1495 if (allocs
> last_allocs
)
1497 last_allocs
= allocs
;
1502 /* allocate and fill alloc_info */
1503 for (group
= 0; group
< nr_groups
; group
++)
1504 nr_units
+= roundup(group_cnt
[group
], upa
);
1506 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
1508 return ERR_PTR(-ENOMEM
);
1509 cpu_map
= ai
->groups
[0].cpu_map
;
1511 for (group
= 0; group
< nr_groups
; group
++) {
1512 ai
->groups
[group
].cpu_map
= cpu_map
;
1513 cpu_map
+= roundup(group_cnt
[group
], upa
);
1516 ai
->static_size
= static_size
;
1517 ai
->reserved_size
= reserved_size
;
1518 ai
->dyn_size
= dyn_size
;
1519 ai
->unit_size
= alloc_size
/ upa
;
1520 ai
->atom_size
= atom_size
;
1521 ai
->alloc_size
= alloc_size
;
1523 for (group
= 0, unit
= 0; group_cnt
[group
]; group
++) {
1524 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1527 * Initialize base_offset as if all groups are located
1528 * back-to-back. The caller should update this to
1529 * reflect actual allocation.
1531 gi
->base_offset
= unit
* ai
->unit_size
;
1533 for_each_possible_cpu(cpu
)
1534 if (group_map
[cpu
] == group
)
1535 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
1536 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
1537 unit
+= gi
->nr_units
;
1539 BUG_ON(unit
!= nr_units
);
1543 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1545 #if defined(BUILD_EMBED_FIRST_CHUNK)
1547 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1548 * @reserved_size: the size of reserved percpu area in bytes
1549 * @dyn_size: minimum free size for dynamic allocation in bytes
1550 * @atom_size: allocation atom size
1551 * @cpu_distance_fn: callback to determine distance between cpus, optional
1552 * @alloc_fn: function to allocate percpu page
1553 * @free_fn: funtion to free percpu page
1555 * This is a helper to ease setting up embedded first percpu chunk and
1556 * can be called where pcpu_setup_first_chunk() is expected.
1558 * If this function is used to setup the first chunk, it is allocated
1559 * by calling @alloc_fn and used as-is without being mapped into
1560 * vmalloc area. Allocations are always whole multiples of @atom_size
1561 * aligned to @atom_size.
1563 * This enables the first chunk to piggy back on the linear physical
1564 * mapping which often uses larger page size. Please note that this
1565 * can result in very sparse cpu->unit mapping on NUMA machines thus
1566 * requiring large vmalloc address space. Don't use this allocator if
1567 * vmalloc space is not orders of magnitude larger than distances
1568 * between node memory addresses (ie. 32bit NUMA machines).
1570 * @dyn_size specifies the minimum dynamic area size.
1572 * If the needed size is smaller than the minimum or specified unit
1573 * size, the leftover is returned using @free_fn.
1576 * 0 on success, -errno on failure.
1578 int __init
pcpu_embed_first_chunk(size_t reserved_size
, size_t dyn_size
,
1580 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
1581 pcpu_fc_alloc_fn_t alloc_fn
,
1582 pcpu_fc_free_fn_t free_fn
)
1584 void *base
= (void *)ULONG_MAX
;
1585 void **areas
= NULL
;
1586 struct pcpu_alloc_info
*ai
;
1587 size_t size_sum
, areas_size
, max_distance
;
1590 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
1595 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
1596 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
1598 areas
= alloc_bootmem_nopanic(areas_size
);
1604 /* allocate, copy and determine base address */
1605 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1606 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1607 unsigned int cpu
= NR_CPUS
;
1610 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
1611 cpu
= gi
->cpu_map
[i
];
1612 BUG_ON(cpu
== NR_CPUS
);
1614 /* allocate space for the whole group */
1615 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
1618 goto out_free_areas
;
1622 base
= min(ptr
, base
);
1624 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
1625 if (gi
->cpu_map
[i
] == NR_CPUS
) {
1626 /* unused unit, free whole */
1627 free_fn(ptr
, ai
->unit_size
);
1630 /* copy and return the unused part */
1631 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
1632 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
1636 /* base address is now known, determine group base offsets */
1638 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1639 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
1640 max_distance
= max_t(size_t, max_distance
,
1641 ai
->groups
[group
].base_offset
);
1643 max_distance
+= ai
->unit_size
;
1645 /* warn if maximum distance is further than 75% of vmalloc space */
1646 if (max_distance
> (VMALLOC_END
- VMALLOC_START
) * 3 / 4) {
1647 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1649 max_distance
, VMALLOC_END
- VMALLOC_START
);
1650 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1651 /* and fail if we have fallback */
1657 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1658 PFN_DOWN(size_sum
), base
, ai
->static_size
, ai
->reserved_size
,
1659 ai
->dyn_size
, ai
->unit_size
);
1661 rc
= pcpu_setup_first_chunk(ai
, base
);
1665 for (group
= 0; group
< ai
->nr_groups
; group
++)
1666 free_fn(areas
[group
],
1667 ai
->groups
[group
].nr_units
* ai
->unit_size
);
1669 pcpu_free_alloc_info(ai
);
1671 free_bootmem(__pa(areas
), areas_size
);
1674 #endif /* BUILD_EMBED_FIRST_CHUNK */
1676 #ifdef BUILD_PAGE_FIRST_CHUNK
1678 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1679 * @reserved_size: the size of reserved percpu area in bytes
1680 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1681 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1682 * @populate_pte_fn: function to populate pte
1684 * This is a helper to ease setting up page-remapped first percpu
1685 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1687 * This is the basic allocator. Static percpu area is allocated
1688 * page-by-page into vmalloc area.
1691 * 0 on success, -errno on failure.
1693 int __init
pcpu_page_first_chunk(size_t reserved_size
,
1694 pcpu_fc_alloc_fn_t alloc_fn
,
1695 pcpu_fc_free_fn_t free_fn
,
1696 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
1698 static struct vm_struct vm
;
1699 struct pcpu_alloc_info
*ai
;
1703 struct page
**pages
;
1706 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
1708 ai
= pcpu_build_alloc_info(reserved_size
, 0, PAGE_SIZE
, NULL
);
1711 BUG_ON(ai
->nr_groups
!= 1);
1712 BUG_ON(ai
->groups
[0].nr_units
!= num_possible_cpus());
1714 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1716 /* unaligned allocations can't be freed, round up to page size */
1717 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
1719 pages
= alloc_bootmem(pages_size
);
1721 /* allocate pages */
1723 for (unit
= 0; unit
< num_possible_cpus(); unit
++)
1724 for (i
= 0; i
< unit_pages
; i
++) {
1725 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
1728 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
1730 pr_warning("PERCPU: failed to allocate %s page "
1731 "for cpu%u\n", psize_str
, cpu
);
1734 pages
[j
++] = virt_to_page(ptr
);
1737 /* allocate vm area, map the pages and copy static data */
1738 vm
.flags
= VM_ALLOC
;
1739 vm
.size
= num_possible_cpus() * ai
->unit_size
;
1740 vm_area_register_early(&vm
, PAGE_SIZE
);
1742 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
1743 unsigned long unit_addr
=
1744 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
1746 for (i
= 0; i
< unit_pages
; i
++)
1747 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
1749 /* pte already populated, the following shouldn't fail */
1750 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
1753 panic("failed to map percpu area, err=%d\n", rc
);
1756 * FIXME: Archs with virtual cache should flush local
1757 * cache for the linear mapping here - something
1758 * equivalent to flush_cache_vmap() on the local cpu.
1759 * flush_cache_vmap() can't be used as most supporting
1760 * data structures are not set up yet.
1763 /* copy static data */
1764 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
1767 /* we're ready, commit */
1768 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1769 unit_pages
, psize_str
, vm
.addr
, ai
->static_size
,
1770 ai
->reserved_size
, ai
->dyn_size
);
1772 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
1777 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
1780 free_bootmem(__pa(pages
), pages_size
);
1781 pcpu_free_alloc_info(ai
);
1784 #endif /* BUILD_PAGE_FIRST_CHUNK */
1786 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1788 * Generic SMP percpu area setup.
1790 * The embedding helper is used because its behavior closely resembles
1791 * the original non-dynamic generic percpu area setup. This is
1792 * important because many archs have addressing restrictions and might
1793 * fail if the percpu area is located far away from the previous
1794 * location. As an added bonus, in non-NUMA cases, embedding is
1795 * generally a good idea TLB-wise because percpu area can piggy back
1796 * on the physical linear memory mapping which uses large page
1797 * mappings on applicable archs.
1799 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
1800 EXPORT_SYMBOL(__per_cpu_offset
);
1802 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
1805 return __alloc_bootmem_nopanic(size
, align
, __pa(MAX_DMA_ADDRESS
));
1808 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
1810 free_bootmem(__pa(ptr
), size
);
1813 void __init
setup_per_cpu_areas(void)
1815 unsigned long delta
;
1820 * Always reserve area for module percpu variables. That's
1821 * what the legacy allocator did.
1823 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
1824 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
1825 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
1827 panic("Failed to initialize percpu areas.");
1829 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
1830 for_each_possible_cpu(cpu
)
1831 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
1833 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1835 #else /* CONFIG_SMP */
1838 * UP percpu area setup.
1840 * UP always uses km-based percpu allocator with identity mapping.
1841 * Static percpu variables are indistinguishable from the usual static
1842 * variables and don't require any special preparation.
1844 void __init
setup_per_cpu_areas(void)
1846 const size_t unit_size
=
1847 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE
,
1848 PERCPU_DYNAMIC_RESERVE
));
1849 struct pcpu_alloc_info
*ai
;
1852 ai
= pcpu_alloc_alloc_info(1, 1);
1853 fc
= __alloc_bootmem(unit_size
, PAGE_SIZE
, __pa(MAX_DMA_ADDRESS
));
1855 panic("Failed to allocate memory for percpu areas.");
1857 ai
->dyn_size
= unit_size
;
1858 ai
->unit_size
= unit_size
;
1859 ai
->atom_size
= unit_size
;
1860 ai
->alloc_size
= unit_size
;
1861 ai
->groups
[0].nr_units
= 1;
1862 ai
->groups
[0].cpu_map
[0] = 0;
1864 if (pcpu_setup_first_chunk(ai
, fc
) < 0)
1865 panic("Failed to initialize percpu areas.");
1868 #endif /* CONFIG_SMP */
1871 * First and reserved chunks are initialized with temporary allocation
1872 * map in initdata so that they can be used before slab is online.
1873 * This function is called after slab is brought up and replaces those
1874 * with properly allocated maps.
1876 void __init
percpu_init_late(void)
1878 struct pcpu_chunk
*target_chunks
[] =
1879 { pcpu_first_chunk
, pcpu_reserved_chunk
, NULL
};
1880 struct pcpu_chunk
*chunk
;
1881 unsigned long flags
;
1884 for (i
= 0; (chunk
= target_chunks
[i
]); i
++) {
1886 const size_t size
= PERCPU_DYNAMIC_EARLY_SLOTS
* sizeof(map
[0]);
1888 BUILD_BUG_ON(size
> PAGE_SIZE
);
1890 map
= pcpu_mem_alloc(size
);
1893 spin_lock_irqsave(&pcpu_lock
, flags
);
1894 memcpy(map
, chunk
->map
, size
);
1896 spin_unlock_irqrestore(&pcpu_lock
, flags
);