2 * linux/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 in vmalloc area. Each
11 * chunk is consisted of boot-time determined number of units and the
12 * first 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. ie. in
20 * ------------------- ------------------- ------------
21 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
22 * ------------------- ...... ------------------- .... ------------
24 * Allocation is done in offset-size areas of single unit space. Ie,
25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
26 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
27 * cpus. On NUMA, the mapping can be non-linear and even sparse.
28 * Percpu access can be done by configuring percpu base registers
29 * according to cpu to unit mapping and pcpu_unit_size.
31 * There are usually many small percpu allocations many of them being
32 * as small as 4 bytes. The allocator organizes chunks into lists
33 * according to free size and tries to allocate from the fullest one.
34 * Each chunk keeps the maximum contiguous area size hint which is
35 * guaranteed to be eqaul to or larger than the maximum contiguous
36 * area in the chunk. This helps the allocator not to iterate the
37 * chunk maps unnecessarily.
39 * Allocation state in each chunk is kept using an array of integers
40 * on chunk->map. A positive value in the map represents a free
41 * region and negative allocated. Allocation inside a chunk is done
42 * by scanning this map sequentially and serving the first matching
43 * entry. This is mostly copied from the percpu_modalloc() allocator.
44 * Chunks can be determined from the address using the index field
45 * in the page struct. The index field contains a pointer to the chunk.
47 * To use this allocator, arch code should do the followings.
49 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
50 * regular address to percpu pointer and back if they need to be
51 * different from the default
53 * - use pcpu_setup_first_chunk() during percpu area initialization to
54 * setup the first chunk containing the kernel static percpu area
57 #include <linux/bitmap.h>
58 #include <linux/bootmem.h>
59 #include <linux/err.h>
60 #include <linux/list.h>
61 #include <linux/log2.h>
63 #include <linux/module.h>
64 #include <linux/mutex.h>
65 #include <linux/percpu.h>
66 #include <linux/pfn.h>
67 #include <linux/slab.h>
68 #include <linux/spinlock.h>
69 #include <linux/vmalloc.h>
70 #include <linux/workqueue.h>
72 #include <asm/cacheflush.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
77 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78 #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)
95 struct list_head list
; /* linked to pcpu_slot lists */
96 int free_size
; /* free bytes in the chunk */
97 int contig_hint
; /* max contiguous size hint */
98 void *base_addr
; /* base address of this chunk */
99 int map_used
; /* # of map entries used */
100 int map_alloc
; /* # of map entries allocated */
101 int *map
; /* allocation map */
102 struct vm_struct
**vms
; /* mapped vmalloc regions */
103 bool immutable
; /* no [de]population allowed */
104 unsigned long populated
[]; /* populated bitmap */
107 static int pcpu_unit_pages __read_mostly
;
108 static int pcpu_unit_size __read_mostly
;
109 static int pcpu_nr_units __read_mostly
;
110 static int pcpu_atom_size __read_mostly
;
111 static int pcpu_nr_slots __read_mostly
;
112 static size_t pcpu_chunk_struct_size __read_mostly
;
114 /* cpus with the lowest and highest unit numbers */
115 static unsigned int pcpu_first_unit_cpu __read_mostly
;
116 static unsigned int pcpu_last_unit_cpu __read_mostly
;
118 /* the address of the first chunk which starts with the kernel static area */
119 void *pcpu_base_addr __read_mostly
;
120 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
122 static const int *pcpu_unit_map __read_mostly
; /* cpu -> unit */
123 const unsigned long *pcpu_unit_offsets __read_mostly
; /* cpu -> unit offset */
125 /* group information, used for vm allocation */
126 static int pcpu_nr_groups __read_mostly
;
127 static const unsigned long *pcpu_group_offsets __read_mostly
;
128 static const size_t *pcpu_group_sizes __read_mostly
;
131 * The first chunk which always exists. Note that unlike other
132 * chunks, this one can be allocated and mapped in several different
133 * ways and thus often doesn't live in the vmalloc area.
135 static struct pcpu_chunk
*pcpu_first_chunk
;
138 * Optional reserved chunk. This chunk reserves part of the first
139 * chunk and serves it for reserved allocations. The amount of
140 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
141 * area doesn't exist, the following variables contain NULL and 0
144 static struct pcpu_chunk
*pcpu_reserved_chunk
;
145 static int pcpu_reserved_chunk_limit
;
148 * Synchronization rules.
150 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
151 * protects allocation/reclaim paths, chunks, populated bitmap and
152 * vmalloc mapping. The latter is a spinlock and protects the index
153 * data structures - chunk slots, chunks and area maps in chunks.
155 * During allocation, pcpu_alloc_mutex is kept locked all the time and
156 * pcpu_lock is grabbed and released as necessary. All actual memory
157 * allocations are done using GFP_KERNEL with pcpu_lock released. In
158 * general, percpu memory can't be allocated with irq off but
159 * irqsave/restore are still used in alloc path so that it can be used
160 * from early init path - sched_init() specifically.
162 * Free path accesses and alters only the index data structures, so it
163 * can be safely called from atomic context. When memory needs to be
164 * returned to the system, free path schedules reclaim_work which
165 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
166 * reclaimed, release both locks and frees the chunks. Note that it's
167 * necessary to grab both locks to remove a chunk from circulation as
168 * allocation path might be referencing the chunk with only
169 * pcpu_alloc_mutex locked.
171 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* protects whole alloc and reclaim */
172 static DEFINE_SPINLOCK(pcpu_lock
); /* protects index data structures */
174 static struct list_head
*pcpu_slot __read_mostly
; /* chunk list slots */
176 /* reclaim work to release fully free chunks, scheduled from free path */
177 static void pcpu_reclaim(struct work_struct
*work
);
178 static DECLARE_WORK(pcpu_reclaim_work
, pcpu_reclaim
);
180 static bool pcpu_addr_in_first_chunk(void *addr
)
182 void *first_start
= pcpu_first_chunk
->base_addr
;
184 return addr
>= first_start
&& addr
< first_start
+ pcpu_unit_size
;
187 static bool pcpu_addr_in_reserved_chunk(void *addr
)
189 void *first_start
= pcpu_first_chunk
->base_addr
;
191 return addr
>= first_start
&&
192 addr
< first_start
+ pcpu_reserved_chunk_limit
;
195 static int __pcpu_size_to_slot(int size
)
197 int highbit
= fls(size
); /* size is in bytes */
198 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
201 static int pcpu_size_to_slot(int size
)
203 if (size
== pcpu_unit_size
)
204 return pcpu_nr_slots
- 1;
205 return __pcpu_size_to_slot(size
);
208 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
210 if (chunk
->free_size
< sizeof(int) || chunk
->contig_hint
< sizeof(int))
213 return pcpu_size_to_slot(chunk
->free_size
);
216 static int pcpu_page_idx(unsigned int cpu
, int page_idx
)
218 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
221 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
222 unsigned int cpu
, int page_idx
)
224 return (unsigned long)chunk
->base_addr
+ pcpu_unit_offsets
[cpu
] +
225 (page_idx
<< PAGE_SHIFT
);
228 static struct page
*pcpu_chunk_page(struct pcpu_chunk
*chunk
,
229 unsigned int cpu
, int page_idx
)
231 /* must not be used on pre-mapped chunk */
232 WARN_ON(chunk
->immutable
);
234 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk
, cpu
, page_idx
));
237 /* set the pointer to a chunk in a page struct */
238 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
240 page
->index
= (unsigned long)pcpu
;
243 /* obtain pointer to a chunk from a page struct */
244 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
246 return (struct pcpu_chunk
*)page
->index
;
249 static void pcpu_next_unpop(struct pcpu_chunk
*chunk
, int *rs
, int *re
, int end
)
251 *rs
= find_next_zero_bit(chunk
->populated
, end
, *rs
);
252 *re
= find_next_bit(chunk
->populated
, end
, *rs
+ 1);
255 static void pcpu_next_pop(struct pcpu_chunk
*chunk
, int *rs
, int *re
, int end
)
257 *rs
= find_next_bit(chunk
->populated
, end
, *rs
);
258 *re
= find_next_zero_bit(chunk
->populated
, end
, *rs
+ 1);
262 * (Un)populated page region iterators. Iterate over (un)populated
263 * page regions betwen @start and @end in @chunk. @rs and @re should
264 * be integer variables and will be set to start and end page index of
265 * the current region.
267 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
268 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
270 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
272 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
275 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
278 * pcpu_mem_alloc - allocate memory
279 * @size: bytes to allocate
281 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
282 * kzalloc() is used; otherwise, vmalloc() is used. The returned
283 * memory is always zeroed.
286 * Does GFP_KERNEL allocation.
289 * Pointer to the allocated area on success, NULL on failure.
291 static void *pcpu_mem_alloc(size_t size
)
293 if (size
<= PAGE_SIZE
)
294 return kzalloc(size
, GFP_KERNEL
);
296 void *ptr
= vmalloc(size
);
298 memset(ptr
, 0, size
);
304 * pcpu_mem_free - free memory
305 * @ptr: memory to free
306 * @size: size of the area
308 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
310 static void pcpu_mem_free(void *ptr
, size_t size
)
312 if (size
<= PAGE_SIZE
)
319 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
320 * @chunk: chunk of interest
321 * @oslot: the previous slot it was on
323 * This function is called after an allocation or free changed @chunk.
324 * New slot according to the changed state is determined and @chunk is
325 * moved to the slot. Note that the reserved chunk is never put on
331 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
333 int nslot
= pcpu_chunk_slot(chunk
);
335 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
337 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
339 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
344 * pcpu_chunk_addr_search - determine chunk containing specified address
345 * @addr: address for which the chunk needs to be determined.
348 * The address of the found chunk.
350 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
352 /* is it in the first chunk? */
353 if (pcpu_addr_in_first_chunk(addr
)) {
354 /* is it in the reserved area? */
355 if (pcpu_addr_in_reserved_chunk(addr
))
356 return pcpu_reserved_chunk
;
357 return pcpu_first_chunk
;
361 * The address is relative to unit0 which might be unused and
362 * thus unmapped. Offset the address to the unit space of the
363 * current processor before looking it up in the vmalloc
364 * space. Note that any possible cpu id can be used here, so
365 * there's no need to worry about preemption or cpu hotplug.
367 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
368 return pcpu_get_page_chunk(vmalloc_to_page(addr
));
372 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
373 * @chunk: chunk of interest
375 * Determine whether area map of @chunk needs to be extended to
376 * accomodate a new allocation.
382 * New target map allocation length if extension is necessary, 0
385 static int pcpu_need_to_extend(struct pcpu_chunk
*chunk
)
389 if (chunk
->map_alloc
>= chunk
->map_used
+ 2)
392 new_alloc
= PCPU_DFL_MAP_ALLOC
;
393 while (new_alloc
< chunk
->map_used
+ 2)
400 * pcpu_extend_area_map - extend area map of a chunk
401 * @chunk: chunk of interest
402 * @new_alloc: new target allocation length of the area map
404 * Extend area map of @chunk to have @new_alloc entries.
407 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
410 * 0 on success, -errno on failure.
412 static int pcpu_extend_area_map(struct pcpu_chunk
*chunk
, int new_alloc
)
414 int *old
= NULL
, *new = NULL
;
415 size_t old_size
= 0, new_size
= new_alloc
* sizeof(new[0]);
418 new = pcpu_mem_alloc(new_size
);
422 /* acquire pcpu_lock and switch to new area map */
423 spin_lock_irqsave(&pcpu_lock
, flags
);
425 if (new_alloc
<= chunk
->map_alloc
)
428 old_size
= chunk
->map_alloc
* sizeof(chunk
->map
[0]);
429 memcpy(new, chunk
->map
, old_size
);
432 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
433 * one of the first chunks and still using static map.
435 if (chunk
->map_alloc
>= PCPU_DFL_MAP_ALLOC
)
438 chunk
->map_alloc
= new_alloc
;
443 spin_unlock_irqrestore(&pcpu_lock
, flags
);
446 * pcpu_mem_free() might end up calling vfree() which uses
447 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
449 pcpu_mem_free(old
, old_size
);
450 pcpu_mem_free(new, new_size
);
456 * pcpu_split_block - split a map block
457 * @chunk: chunk of interest
458 * @i: index of map block to split
459 * @head: head size in bytes (can be 0)
460 * @tail: tail size in bytes (can be 0)
462 * Split the @i'th map block into two or three blocks. If @head is
463 * non-zero, @head bytes block is inserted before block @i moving it
464 * to @i+1 and reducing its size by @head bytes.
466 * If @tail is non-zero, the target block, which can be @i or @i+1
467 * depending on @head, is reduced by @tail bytes and @tail byte block
468 * is inserted after the target block.
470 * @chunk->map must have enough free slots to accomodate the split.
475 static void pcpu_split_block(struct pcpu_chunk
*chunk
, int i
,
478 int nr_extra
= !!head
+ !!tail
;
480 BUG_ON(chunk
->map_alloc
< chunk
->map_used
+ nr_extra
);
482 /* insert new subblocks */
483 memmove(&chunk
->map
[i
+ nr_extra
], &chunk
->map
[i
],
484 sizeof(chunk
->map
[0]) * (chunk
->map_used
- i
));
485 chunk
->map_used
+= nr_extra
;
488 chunk
->map
[i
+ 1] = chunk
->map
[i
] - head
;
489 chunk
->map
[i
++] = head
;
492 chunk
->map
[i
++] -= tail
;
493 chunk
->map
[i
] = tail
;
498 * pcpu_alloc_area - allocate area from a pcpu_chunk
499 * @chunk: chunk of interest
500 * @size: wanted size in bytes
501 * @align: wanted align
503 * Try to allocate @size bytes area aligned at @align from @chunk.
504 * Note that this function only allocates the offset. It doesn't
505 * populate or map the area.
507 * @chunk->map must have at least two free slots.
513 * Allocated offset in @chunk on success, -1 if no matching area is
516 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int size
, int align
)
518 int oslot
= pcpu_chunk_slot(chunk
);
522 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++])) {
523 bool is_last
= i
+ 1 == chunk
->map_used
;
526 /* extra for alignment requirement */
527 head
= ALIGN(off
, align
) - off
;
528 BUG_ON(i
== 0 && head
!= 0);
530 if (chunk
->map
[i
] < 0)
532 if (chunk
->map
[i
] < head
+ size
) {
533 max_contig
= max(chunk
->map
[i
], max_contig
);
538 * If head is small or the previous block is free,
539 * merge'em. Note that 'small' is defined as smaller
540 * than sizeof(int), which is very small but isn't too
541 * uncommon for percpu allocations.
543 if (head
&& (head
< sizeof(int) || chunk
->map
[i
- 1] > 0)) {
544 if (chunk
->map
[i
- 1] > 0)
545 chunk
->map
[i
- 1] += head
;
547 chunk
->map
[i
- 1] -= head
;
548 chunk
->free_size
-= head
;
550 chunk
->map
[i
] -= head
;
555 /* if tail is small, just keep it around */
556 tail
= chunk
->map
[i
] - head
- size
;
557 if (tail
< sizeof(int))
560 /* split if warranted */
562 pcpu_split_block(chunk
, i
, head
, tail
);
566 max_contig
= max(chunk
->map
[i
- 1], max_contig
);
569 max_contig
= max(chunk
->map
[i
+ 1], max_contig
);
572 /* update hint and mark allocated */
574 chunk
->contig_hint
= max_contig
; /* fully scanned */
576 chunk
->contig_hint
= max(chunk
->contig_hint
,
579 chunk
->free_size
-= chunk
->map
[i
];
580 chunk
->map
[i
] = -chunk
->map
[i
];
582 pcpu_chunk_relocate(chunk
, oslot
);
586 chunk
->contig_hint
= max_contig
; /* fully scanned */
587 pcpu_chunk_relocate(chunk
, oslot
);
589 /* tell the upper layer that this chunk has no matching area */
594 * pcpu_free_area - free area to a pcpu_chunk
595 * @chunk: chunk of interest
596 * @freeme: offset of area to free
598 * Free area starting from @freeme to @chunk. Note that this function
599 * only modifies the allocation map. It doesn't depopulate or unmap
605 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int freeme
)
607 int oslot
= pcpu_chunk_slot(chunk
);
610 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++]))
613 BUG_ON(off
!= freeme
);
614 BUG_ON(chunk
->map
[i
] > 0);
616 chunk
->map
[i
] = -chunk
->map
[i
];
617 chunk
->free_size
+= chunk
->map
[i
];
619 /* merge with previous? */
620 if (i
> 0 && chunk
->map
[i
- 1] >= 0) {
621 chunk
->map
[i
- 1] += chunk
->map
[i
];
623 memmove(&chunk
->map
[i
], &chunk
->map
[i
+ 1],
624 (chunk
->map_used
- i
) * sizeof(chunk
->map
[0]));
627 /* merge with next? */
628 if (i
+ 1 < chunk
->map_used
&& chunk
->map
[i
+ 1] >= 0) {
629 chunk
->map
[i
] += chunk
->map
[i
+ 1];
631 memmove(&chunk
->map
[i
+ 1], &chunk
->map
[i
+ 2],
632 (chunk
->map_used
- (i
+ 1)) * sizeof(chunk
->map
[0]));
635 chunk
->contig_hint
= max(chunk
->map
[i
], chunk
->contig_hint
);
636 pcpu_chunk_relocate(chunk
, oslot
);
639 static struct pcpu_chunk
*pcpu_alloc_chunk(void)
641 struct pcpu_chunk
*chunk
;
643 chunk
= kzalloc(pcpu_chunk_struct_size
, GFP_KERNEL
);
647 chunk
->map
= pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC
* sizeof(chunk
->map
[0]));
653 chunk
->map_alloc
= PCPU_DFL_MAP_ALLOC
;
654 chunk
->map
[chunk
->map_used
++] = pcpu_unit_size
;
656 INIT_LIST_HEAD(&chunk
->list
);
657 chunk
->free_size
= pcpu_unit_size
;
658 chunk
->contig_hint
= pcpu_unit_size
;
663 static void pcpu_free_chunk(struct pcpu_chunk
*chunk
)
667 pcpu_mem_free(chunk
->map
, chunk
->map_alloc
* sizeof(chunk
->map
[0]));
672 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
673 * @chunk: chunk of interest
674 * @bitmapp: output parameter for bitmap
675 * @may_alloc: may allocate the array
677 * Returns pointer to array of pointers to struct page and bitmap,
678 * both of which can be indexed with pcpu_page_idx(). The returned
679 * array is cleared to zero and *@bitmapp is copied from
680 * @chunk->populated. Note that there is only one array and bitmap
681 * and access exclusion is the caller's responsibility.
684 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
685 * Otherwise, don't care.
688 * Pointer to temp pages array on success, NULL on failure.
690 static struct page
**pcpu_get_pages_and_bitmap(struct pcpu_chunk
*chunk
,
691 unsigned long **bitmapp
,
694 static struct page
**pages
;
695 static unsigned long *bitmap
;
696 size_t pages_size
= pcpu_nr_units
* pcpu_unit_pages
* sizeof(pages
[0]);
697 size_t bitmap_size
= BITS_TO_LONGS(pcpu_unit_pages
) *
698 sizeof(unsigned long);
700 if (!pages
|| !bitmap
) {
701 if (may_alloc
&& !pages
)
702 pages
= pcpu_mem_alloc(pages_size
);
703 if (may_alloc
&& !bitmap
)
704 bitmap
= pcpu_mem_alloc(bitmap_size
);
705 if (!pages
|| !bitmap
)
709 memset(pages
, 0, pages_size
);
710 bitmap_copy(bitmap
, chunk
->populated
, pcpu_unit_pages
);
717 * pcpu_free_pages - free pages which were allocated for @chunk
718 * @chunk: chunk pages were allocated for
719 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
720 * @populated: populated bitmap
721 * @page_start: page index of the first page to be freed
722 * @page_end: page index of the last page to be freed + 1
724 * Free pages [@page_start and @page_end) in @pages for all units.
725 * The pages were allocated for @chunk.
727 static void pcpu_free_pages(struct pcpu_chunk
*chunk
,
728 struct page
**pages
, unsigned long *populated
,
729 int page_start
, int page_end
)
734 for_each_possible_cpu(cpu
) {
735 for (i
= page_start
; i
< page_end
; i
++) {
736 struct page
*page
= pages
[pcpu_page_idx(cpu
, i
)];
745 * pcpu_alloc_pages - allocates pages for @chunk
746 * @chunk: target chunk
747 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
748 * @populated: populated bitmap
749 * @page_start: page index of the first page to be allocated
750 * @page_end: page index of the last page to be allocated + 1
752 * Allocate pages [@page_start,@page_end) into @pages for all units.
753 * The allocation is for @chunk. Percpu core doesn't care about the
754 * content of @pages and will pass it verbatim to pcpu_map_pages().
756 static int pcpu_alloc_pages(struct pcpu_chunk
*chunk
,
757 struct page
**pages
, unsigned long *populated
,
758 int page_start
, int page_end
)
760 const gfp_t gfp
= GFP_KERNEL
| __GFP_HIGHMEM
| __GFP_COLD
;
764 for_each_possible_cpu(cpu
) {
765 for (i
= page_start
; i
< page_end
; i
++) {
766 struct page
**pagep
= &pages
[pcpu_page_idx(cpu
, i
)];
768 *pagep
= alloc_pages_node(cpu_to_node(cpu
), gfp
, 0);
770 pcpu_free_pages(chunk
, pages
, populated
,
771 page_start
, page_end
);
780 * pcpu_pre_unmap_flush - flush cache prior to unmapping
781 * @chunk: chunk the regions to be flushed belongs to
782 * @page_start: page index of the first page to be flushed
783 * @page_end: page index of the last page to be flushed + 1
785 * Pages in [@page_start,@page_end) of @chunk are about to be
786 * unmapped. Flush cache. As each flushing trial can be very
787 * expensive, issue flush on the whole region at once rather than
788 * doing it for each cpu. This could be an overkill but is more
791 static void pcpu_pre_unmap_flush(struct pcpu_chunk
*chunk
,
792 int page_start
, int page_end
)
795 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
796 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
799 static void __pcpu_unmap_pages(unsigned long addr
, int nr_pages
)
801 unmap_kernel_range_noflush(addr
, nr_pages
<< PAGE_SHIFT
);
805 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
806 * @chunk: chunk of interest
807 * @pages: pages array which can be used to pass information to free
808 * @populated: populated bitmap
809 * @page_start: page index of the first page to unmap
810 * @page_end: page index of the last page to unmap + 1
812 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
813 * Corresponding elements in @pages were cleared by the caller and can
814 * be used to carry information to pcpu_free_pages() which will be
815 * called after all unmaps are finished. The caller should call
816 * proper pre/post flush functions.
818 static void pcpu_unmap_pages(struct pcpu_chunk
*chunk
,
819 struct page
**pages
, unsigned long *populated
,
820 int page_start
, int page_end
)
825 for_each_possible_cpu(cpu
) {
826 for (i
= page_start
; i
< page_end
; i
++) {
829 page
= pcpu_chunk_page(chunk
, cpu
, i
);
831 pages
[pcpu_page_idx(cpu
, i
)] = page
;
833 __pcpu_unmap_pages(pcpu_chunk_addr(chunk
, cpu
, page_start
),
834 page_end
- page_start
);
837 for (i
= page_start
; i
< page_end
; i
++)
838 __clear_bit(i
, populated
);
842 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
843 * @chunk: pcpu_chunk the regions to be flushed belong to
844 * @page_start: page index of the first page to be flushed
845 * @page_end: page index of the last page to be flushed + 1
847 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
848 * TLB for the regions. This can be skipped if the area is to be
849 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
851 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
852 * for the whole region.
854 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk
*chunk
,
855 int page_start
, int page_end
)
857 flush_tlb_kernel_range(
858 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
859 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
862 static int __pcpu_map_pages(unsigned long addr
, struct page
**pages
,
865 return map_kernel_range_noflush(addr
, nr_pages
<< PAGE_SHIFT
,
870 * pcpu_map_pages - map pages into a pcpu_chunk
871 * @chunk: chunk of interest
872 * @pages: pages array containing pages to be mapped
873 * @populated: populated bitmap
874 * @page_start: page index of the first page to map
875 * @page_end: page index of the last page to map + 1
877 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
878 * caller is responsible for calling pcpu_post_map_flush() after all
879 * mappings are complete.
881 * This function is responsible for setting corresponding bits in
882 * @chunk->populated bitmap and whatever is necessary for reverse
883 * lookup (addr -> chunk).
885 static int pcpu_map_pages(struct pcpu_chunk
*chunk
,
886 struct page
**pages
, unsigned long *populated
,
887 int page_start
, int page_end
)
889 unsigned int cpu
, tcpu
;
892 for_each_possible_cpu(cpu
) {
893 err
= __pcpu_map_pages(pcpu_chunk_addr(chunk
, cpu
, page_start
),
894 &pages
[pcpu_page_idx(cpu
, page_start
)],
895 page_end
- page_start
);
900 /* mapping successful, link chunk and mark populated */
901 for (i
= page_start
; i
< page_end
; i
++) {
902 for_each_possible_cpu(cpu
)
903 pcpu_set_page_chunk(pages
[pcpu_page_idx(cpu
, i
)],
905 __set_bit(i
, populated
);
911 for_each_possible_cpu(tcpu
) {
914 __pcpu_unmap_pages(pcpu_chunk_addr(chunk
, tcpu
, page_start
),
915 page_end
- page_start
);
921 * pcpu_post_map_flush - flush cache after mapping
922 * @chunk: pcpu_chunk the regions to be flushed belong to
923 * @page_start: page index of the first page to be flushed
924 * @page_end: page index of the last page to be flushed + 1
926 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
929 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
930 * for the whole region.
932 static void pcpu_post_map_flush(struct pcpu_chunk
*chunk
,
933 int page_start
, int page_end
)
936 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
937 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
941 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
942 * @chunk: chunk to depopulate
943 * @off: offset to the area to depopulate
944 * @size: size of the area to depopulate in bytes
945 * @flush: whether to flush cache and tlb or not
947 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
948 * from @chunk. If @flush is true, vcache is flushed before unmapping
954 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
)
956 int page_start
= PFN_DOWN(off
);
957 int page_end
= PFN_UP(off
+ size
);
959 unsigned long *populated
;
962 /* quick path, check whether it's empty already */
964 pcpu_next_unpop(chunk
, &rs
, &re
, page_end
);
965 if (rs
== page_start
&& re
== page_end
)
968 /* immutable chunks can't be depopulated */
969 WARN_ON(chunk
->immutable
);
972 * If control reaches here, there must have been at least one
973 * successful population attempt so the temp pages array must
976 pages
= pcpu_get_pages_and_bitmap(chunk
, &populated
, false);
980 pcpu_pre_unmap_flush(chunk
, page_start
, page_end
);
982 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
)
983 pcpu_unmap_pages(chunk
, pages
, populated
, rs
, re
);
985 /* no need to flush tlb, vmalloc will handle it lazily */
987 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
)
988 pcpu_free_pages(chunk
, pages
, populated
, rs
, re
);
990 /* commit new bitmap */
991 bitmap_copy(chunk
->populated
, populated
, pcpu_unit_pages
);
995 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
996 * @chunk: chunk of interest
997 * @off: offset to the area to populate
998 * @size: size of the area to populate in bytes
1000 * For each cpu, populate and map pages [@page_start,@page_end) into
1001 * @chunk. The area is cleared on return.
1004 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
1006 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
)
1008 int page_start
= PFN_DOWN(off
);
1009 int page_end
= PFN_UP(off
+ size
);
1010 int free_end
= page_start
, unmap_end
= page_start
;
1011 struct page
**pages
;
1012 unsigned long *populated
;
1016 /* quick path, check whether all pages are already there */
1018 pcpu_next_pop(chunk
, &rs
, &re
, page_end
);
1019 if (rs
== page_start
&& re
== page_end
)
1022 /* need to allocate and map pages, this chunk can't be immutable */
1023 WARN_ON(chunk
->immutable
);
1025 pages
= pcpu_get_pages_and_bitmap(chunk
, &populated
, true);
1030 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
1031 rc
= pcpu_alloc_pages(chunk
, pages
, populated
, rs
, re
);
1037 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
1038 rc
= pcpu_map_pages(chunk
, pages
, populated
, rs
, re
);
1043 pcpu_post_map_flush(chunk
, page_start
, page_end
);
1045 /* commit new bitmap */
1046 bitmap_copy(chunk
->populated
, populated
, pcpu_unit_pages
);
1048 for_each_possible_cpu(cpu
)
1049 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
1053 pcpu_pre_unmap_flush(chunk
, page_start
, unmap_end
);
1054 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, unmap_end
)
1055 pcpu_unmap_pages(chunk
, pages
, populated
, rs
, re
);
1056 pcpu_post_unmap_tlb_flush(chunk
, page_start
, unmap_end
);
1058 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, free_end
)
1059 pcpu_free_pages(chunk
, pages
, populated
, rs
, re
);
1063 static void pcpu_destroy_chunk(struct pcpu_chunk
*chunk
)
1065 if (chunk
&& chunk
->vms
)
1066 pcpu_free_vm_areas(chunk
->vms
, pcpu_nr_groups
);
1067 pcpu_free_chunk(chunk
);
1070 static struct pcpu_chunk
*pcpu_create_chunk(void)
1072 struct pcpu_chunk
*chunk
;
1073 struct vm_struct
**vms
;
1075 chunk
= pcpu_alloc_chunk();
1079 vms
= pcpu_get_vm_areas(pcpu_group_offsets
, pcpu_group_sizes
,
1080 pcpu_nr_groups
, pcpu_atom_size
, GFP_KERNEL
);
1082 pcpu_free_chunk(chunk
);
1087 chunk
->base_addr
= vms
[0]->addr
- pcpu_group_offsets
[0];
1092 * pcpu_alloc - the percpu allocator
1093 * @size: size of area to allocate in bytes
1094 * @align: alignment of area (max PAGE_SIZE)
1095 * @reserved: allocate from the reserved chunk if available
1097 * Allocate percpu area of @size bytes aligned at @align.
1100 * Does GFP_KERNEL allocation.
1103 * Percpu pointer to the allocated area on success, NULL on failure.
1105 static void __percpu
*pcpu_alloc(size_t size
, size_t align
, bool reserved
)
1107 static int warn_limit
= 10;
1108 struct pcpu_chunk
*chunk
;
1110 int slot
, off
, new_alloc
;
1111 unsigned long flags
;
1113 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
)) {
1114 WARN(true, "illegal size (%zu) or align (%zu) for "
1115 "percpu allocation\n", size
, align
);
1119 mutex_lock(&pcpu_alloc_mutex
);
1120 spin_lock_irqsave(&pcpu_lock
, flags
);
1122 /* serve reserved allocations from the reserved chunk if available */
1123 if (reserved
&& pcpu_reserved_chunk
) {
1124 chunk
= pcpu_reserved_chunk
;
1126 if (size
> chunk
->contig_hint
) {
1127 err
= "alloc from reserved chunk failed";
1131 while ((new_alloc
= pcpu_need_to_extend(chunk
))) {
1132 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1133 if (pcpu_extend_area_map(chunk
, new_alloc
) < 0) {
1134 err
= "failed to extend area map of reserved chunk";
1135 goto fail_unlock_mutex
;
1137 spin_lock_irqsave(&pcpu_lock
, flags
);
1140 off
= pcpu_alloc_area(chunk
, size
, align
);
1144 err
= "alloc from reserved chunk failed";
1149 /* search through normal chunks */
1150 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
1151 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1152 if (size
> chunk
->contig_hint
)
1155 new_alloc
= pcpu_need_to_extend(chunk
);
1157 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1158 if (pcpu_extend_area_map(chunk
,
1160 err
= "failed to extend area map";
1161 goto fail_unlock_mutex
;
1163 spin_lock_irqsave(&pcpu_lock
, flags
);
1165 * pcpu_lock has been dropped, need to
1166 * restart cpu_slot list walking.
1171 off
= pcpu_alloc_area(chunk
, size
, align
);
1177 /* hmmm... no space left, create a new chunk */
1178 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1180 chunk
= pcpu_create_chunk();
1182 err
= "failed to allocate new chunk";
1183 goto fail_unlock_mutex
;
1186 spin_lock_irqsave(&pcpu_lock
, flags
);
1187 pcpu_chunk_relocate(chunk
, -1);
1191 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1193 /* populate, map and clear the area */
1194 if (pcpu_populate_chunk(chunk
, off
, size
)) {
1195 spin_lock_irqsave(&pcpu_lock
, flags
);
1196 pcpu_free_area(chunk
, off
);
1197 err
= "failed to populate";
1201 mutex_unlock(&pcpu_alloc_mutex
);
1203 /* return address relative to base address */
1204 return __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
1207 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1209 mutex_unlock(&pcpu_alloc_mutex
);
1211 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1212 "%s\n", size
, align
, err
);
1215 pr_info("PERCPU: limit reached, disable warning\n");
1221 * __alloc_percpu - allocate dynamic percpu area
1222 * @size: size of area to allocate in bytes
1223 * @align: alignment of area (max PAGE_SIZE)
1225 * Allocate percpu area of @size bytes aligned at @align. Might
1226 * sleep. Might trigger writeouts.
1229 * Does GFP_KERNEL allocation.
1232 * Percpu pointer to the allocated area on success, NULL on failure.
1234 void __percpu
*__alloc_percpu(size_t size
, size_t align
)
1236 return pcpu_alloc(size
, align
, false);
1238 EXPORT_SYMBOL_GPL(__alloc_percpu
);
1241 * __alloc_reserved_percpu - allocate reserved percpu area
1242 * @size: size of area to allocate in bytes
1243 * @align: alignment of area (max PAGE_SIZE)
1245 * Allocate percpu area of @size bytes aligned at @align from reserved
1246 * percpu area if arch has set it up; otherwise, allocation is served
1247 * from the same dynamic area. Might sleep. Might trigger writeouts.
1250 * Does GFP_KERNEL allocation.
1253 * Percpu pointer to the allocated area on success, NULL on failure.
1255 void __percpu
*__alloc_reserved_percpu(size_t size
, size_t align
)
1257 return pcpu_alloc(size
, align
, true);
1261 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1264 * Reclaim all fully free chunks except for the first one.
1267 * workqueue context.
1269 static void pcpu_reclaim(struct work_struct
*work
)
1272 struct list_head
*head
= &pcpu_slot
[pcpu_nr_slots
- 1];
1273 struct pcpu_chunk
*chunk
, *next
;
1275 mutex_lock(&pcpu_alloc_mutex
);
1276 spin_lock_irq(&pcpu_lock
);
1278 list_for_each_entry_safe(chunk
, next
, head
, list
) {
1279 WARN_ON(chunk
->immutable
);
1281 /* spare the first one */
1282 if (chunk
== list_first_entry(head
, struct pcpu_chunk
, list
))
1285 list_move(&chunk
->list
, &todo
);
1288 spin_unlock_irq(&pcpu_lock
);
1290 list_for_each_entry_safe(chunk
, next
, &todo
, list
) {
1291 pcpu_depopulate_chunk(chunk
, 0, pcpu_unit_size
);
1292 pcpu_destroy_chunk(chunk
);
1295 mutex_unlock(&pcpu_alloc_mutex
);
1299 * free_percpu - free percpu area
1300 * @ptr: pointer to area to free
1302 * Free percpu area @ptr.
1305 * Can be called from atomic context.
1307 void free_percpu(void __percpu
*ptr
)
1310 struct pcpu_chunk
*chunk
;
1311 unsigned long flags
;
1317 addr
= __pcpu_ptr_to_addr(ptr
);
1319 spin_lock_irqsave(&pcpu_lock
, flags
);
1321 chunk
= pcpu_chunk_addr_search(addr
);
1322 off
= addr
- chunk
->base_addr
;
1324 pcpu_free_area(chunk
, off
);
1326 /* if there are more than one fully free chunks, wake up grim reaper */
1327 if (chunk
->free_size
== pcpu_unit_size
) {
1328 struct pcpu_chunk
*pos
;
1330 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
1332 schedule_work(&pcpu_reclaim_work
);
1337 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1339 EXPORT_SYMBOL_GPL(free_percpu
);
1342 * is_kernel_percpu_address - test whether address is from static percpu area
1343 * @addr: address to test
1345 * Test whether @addr belongs to in-kernel static percpu area. Module
1346 * static percpu areas are not considered. For those, use
1347 * is_module_percpu_address().
1350 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1352 bool is_kernel_percpu_address(unsigned long addr
)
1354 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1355 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1358 for_each_possible_cpu(cpu
) {
1359 void *start
= per_cpu_ptr(base
, cpu
);
1361 if ((void *)addr
>= start
&& (void *)addr
< start
+ static_size
)
1368 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1369 * @addr: the address to be converted to physical address
1371 * Given @addr which is dereferenceable address obtained via one of
1372 * percpu access macros, this function translates it into its physical
1373 * address. The caller is responsible for ensuring @addr stays valid
1374 * until this function finishes.
1377 * The physical address for @addr.
1379 phys_addr_t
per_cpu_ptr_to_phys(void *addr
)
1381 if (pcpu_addr_in_first_chunk(addr
)) {
1382 if ((unsigned long)addr
< VMALLOC_START
||
1383 (unsigned long)addr
>= VMALLOC_END
)
1386 return page_to_phys(vmalloc_to_page(addr
));
1388 return page_to_phys(vmalloc_to_page(addr
));
1391 static inline size_t pcpu_calc_fc_sizes(size_t static_size
,
1392 size_t reserved_size
,
1397 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
1398 (*dyn_sizep
>= 0 ? *dyn_sizep
: 0));
1399 if (*dyn_sizep
!= 0)
1400 *dyn_sizep
= size_sum
- static_size
- reserved_size
;
1406 * pcpu_alloc_alloc_info - allocate percpu allocation info
1407 * @nr_groups: the number of groups
1408 * @nr_units: the number of units
1410 * Allocate ai which is large enough for @nr_groups groups containing
1411 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1412 * cpu_map array which is long enough for @nr_units and filled with
1413 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1414 * pointer of other groups.
1417 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1420 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
1423 struct pcpu_alloc_info
*ai
;
1424 size_t base_size
, ai_size
;
1428 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
1429 __alignof__(ai
->groups
[0].cpu_map
[0]));
1430 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
1432 ptr
= alloc_bootmem_nopanic(PFN_ALIGN(ai_size
));
1438 ai
->groups
[0].cpu_map
= ptr
;
1440 for (unit
= 0; unit
< nr_units
; unit
++)
1441 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
1443 ai
->nr_groups
= nr_groups
;
1444 ai
->__ai_size
= PFN_ALIGN(ai_size
);
1450 * pcpu_free_alloc_info - free percpu allocation info
1451 * @ai: pcpu_alloc_info to free
1453 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1455 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
1457 free_bootmem(__pa(ai
), ai
->__ai_size
);
1461 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1462 * @reserved_size: the size of reserved percpu area in bytes
1463 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1464 * @atom_size: allocation atom size
1465 * @cpu_distance_fn: callback to determine distance between cpus, optional
1467 * This function determines grouping of units, their mappings to cpus
1468 * and other parameters considering needed percpu size, allocation
1469 * atom size and distances between CPUs.
1471 * Groups are always mutliples of atom size and CPUs which are of
1472 * LOCAL_DISTANCE both ways are grouped together and share space for
1473 * units in the same group. The returned configuration is guaranteed
1474 * to have CPUs on different nodes on different groups and >=75% usage
1475 * of allocated virtual address space.
1478 * On success, pointer to the new allocation_info is returned. On
1479 * failure, ERR_PTR value is returned.
1481 struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
1482 size_t reserved_size
, ssize_t dyn_size
,
1484 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
1486 static int group_map
[NR_CPUS
] __initdata
;
1487 static int group_cnt
[NR_CPUS
] __initdata
;
1488 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1489 int group_cnt_max
= 0, nr_groups
= 1, nr_units
= 0;
1490 size_t size_sum
, min_unit_size
, alloc_size
;
1491 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
1492 int last_allocs
, group
, unit
;
1493 unsigned int cpu
, tcpu
;
1494 struct pcpu_alloc_info
*ai
;
1495 unsigned int *cpu_map
;
1497 /* this function may be called multiple times */
1498 memset(group_map
, 0, sizeof(group_map
));
1499 memset(group_cnt
, 0, sizeof(group_map
));
1502 * Determine min_unit_size, alloc_size and max_upa such that
1503 * alloc_size is multiple of atom_size and is the smallest
1504 * which can accomodate 4k aligned segments which are equal to
1505 * or larger than min_unit_size.
1507 size_sum
= pcpu_calc_fc_sizes(static_size
, reserved_size
, &dyn_size
);
1508 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1510 alloc_size
= roundup(min_unit_size
, atom_size
);
1511 upa
= alloc_size
/ min_unit_size
;
1512 while (alloc_size
% upa
|| ((alloc_size
/ upa
) & ~PAGE_MASK
))
1516 /* group cpus according to their proximity */
1517 for_each_possible_cpu(cpu
) {
1520 for_each_possible_cpu(tcpu
) {
1523 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
1524 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
1525 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
1527 nr_groups
= max(nr_groups
, group
+ 1);
1531 group_map
[cpu
] = group
;
1533 group_cnt_max
= max(group_cnt_max
, group_cnt
[group
]);
1537 * Expand unit size until address space usage goes over 75%
1538 * and then as much as possible without using more address
1541 last_allocs
= INT_MAX
;
1542 for (upa
= max_upa
; upa
; upa
--) {
1543 int allocs
= 0, wasted
= 0;
1545 if (alloc_size
% upa
|| ((alloc_size
/ upa
) & ~PAGE_MASK
))
1548 for (group
= 0; group
< nr_groups
; group
++) {
1549 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
1550 allocs
+= this_allocs
;
1551 wasted
+= this_allocs
* upa
- group_cnt
[group
];
1555 * Don't accept if wastage is over 25%. The
1556 * greater-than comparison ensures upa==1 always
1557 * passes the following check.
1559 if (wasted
> num_possible_cpus() / 3)
1562 /* and then don't consume more memory */
1563 if (allocs
> last_allocs
)
1565 last_allocs
= allocs
;
1570 /* allocate and fill alloc_info */
1571 for (group
= 0; group
< nr_groups
; group
++)
1572 nr_units
+= roundup(group_cnt
[group
], upa
);
1574 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
1576 return ERR_PTR(-ENOMEM
);
1577 cpu_map
= ai
->groups
[0].cpu_map
;
1579 for (group
= 0; group
< nr_groups
; group
++) {
1580 ai
->groups
[group
].cpu_map
= cpu_map
;
1581 cpu_map
+= roundup(group_cnt
[group
], upa
);
1584 ai
->static_size
= static_size
;
1585 ai
->reserved_size
= reserved_size
;
1586 ai
->dyn_size
= dyn_size
;
1587 ai
->unit_size
= alloc_size
/ upa
;
1588 ai
->atom_size
= atom_size
;
1589 ai
->alloc_size
= alloc_size
;
1591 for (group
= 0, unit
= 0; group_cnt
[group
]; group
++) {
1592 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1595 * Initialize base_offset as if all groups are located
1596 * back-to-back. The caller should update this to
1597 * reflect actual allocation.
1599 gi
->base_offset
= unit
* ai
->unit_size
;
1601 for_each_possible_cpu(cpu
)
1602 if (group_map
[cpu
] == group
)
1603 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
1604 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
1605 unit
+= gi
->nr_units
;
1607 BUG_ON(unit
!= nr_units
);
1613 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1615 * @ai: allocation info to dump
1617 * Print out information about @ai using loglevel @lvl.
1619 static void pcpu_dump_alloc_info(const char *lvl
,
1620 const struct pcpu_alloc_info
*ai
)
1622 int group_width
= 1, cpu_width
= 1, width
;
1623 char empty_str
[] = "--------";
1624 int alloc
= 0, alloc_end
= 0;
1626 int upa
, apl
; /* units per alloc, allocs per line */
1632 v
= num_possible_cpus();
1635 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
1637 upa
= ai
->alloc_size
/ ai
->unit_size
;
1638 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
1639 apl
= rounddown_pow_of_two(max(60 / width
, 1));
1641 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1642 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
1643 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
1645 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1646 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1647 int unit
= 0, unit_end
= 0;
1649 BUG_ON(gi
->nr_units
% upa
);
1650 for (alloc_end
+= gi
->nr_units
/ upa
;
1651 alloc
< alloc_end
; alloc
++) {
1652 if (!(alloc
% apl
)) {
1654 printk("%spcpu-alloc: ", lvl
);
1656 printk("[%0*d] ", group_width
, group
);
1658 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
1659 if (gi
->cpu_map
[unit
] != NR_CPUS
)
1660 printk("%0*d ", cpu_width
,
1663 printk("%s ", empty_str
);
1670 * pcpu_setup_first_chunk - initialize the first percpu chunk
1671 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1672 * @base_addr: mapped address
1674 * Initialize the first percpu chunk which contains the kernel static
1675 * perpcu area. This function is to be called from arch percpu area
1678 * @ai contains all information necessary to initialize the first
1679 * chunk and prime the dynamic percpu allocator.
1681 * @ai->static_size is the size of static percpu area.
1683 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1684 * reserve after the static area in the first chunk. This reserves
1685 * the first chunk such that it's available only through reserved
1686 * percpu allocation. This is primarily used to serve module percpu
1687 * static areas on architectures where the addressing model has
1688 * limited offset range for symbol relocations to guarantee module
1689 * percpu symbols fall inside the relocatable range.
1691 * @ai->dyn_size determines the number of bytes available for dynamic
1692 * allocation in the first chunk. The area between @ai->static_size +
1693 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1695 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1696 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1699 * @ai->atom_size is the allocation atom size and used as alignment
1702 * @ai->alloc_size is the allocation size and always multiple of
1703 * @ai->atom_size. This is larger than @ai->atom_size if
1704 * @ai->unit_size is larger than @ai->atom_size.
1706 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1707 * percpu areas. Units which should be colocated are put into the
1708 * same group. Dynamic VM areas will be allocated according to these
1709 * groupings. If @ai->nr_groups is zero, a single group containing
1710 * all units is assumed.
1712 * The caller should have mapped the first chunk at @base_addr and
1713 * copied static data to each unit.
1715 * If the first chunk ends up with both reserved and dynamic areas, it
1716 * is served by two chunks - one to serve the core static and reserved
1717 * areas and the other for the dynamic area. They share the same vm
1718 * and page map but uses different area allocation map to stay away
1719 * from each other. The latter chunk is circulated in the chunk slots
1720 * and available for dynamic allocation like any other chunks.
1723 * 0 on success, -errno on failure.
1725 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
1728 static char cpus_buf
[4096] __initdata
;
1729 static int smap
[2], dmap
[2];
1730 size_t dyn_size
= ai
->dyn_size
;
1731 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ dyn_size
;
1732 struct pcpu_chunk
*schunk
, *dchunk
= NULL
;
1733 unsigned long *group_offsets
;
1734 size_t *group_sizes
;
1735 unsigned long *unit_off
;
1740 cpumask_scnprintf(cpus_buf
, sizeof(cpus_buf
), cpu_possible_mask
);
1742 #define PCPU_SETUP_BUG_ON(cond) do { \
1743 if (unlikely(cond)) { \
1744 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1745 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1746 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1752 BUILD_BUG_ON(ARRAY_SIZE(smap
) >= PCPU_DFL_MAP_ALLOC
||
1753 ARRAY_SIZE(dmap
) >= PCPU_DFL_MAP_ALLOC
);
1754 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
1755 PCPU_SETUP_BUG_ON(!ai
->static_size
);
1756 PCPU_SETUP_BUG_ON(!base_addr
);
1757 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
1758 PCPU_SETUP_BUG_ON(ai
->unit_size
& ~PAGE_MASK
);
1759 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
1761 /* process group information and build config tables accordingly */
1762 group_offsets
= alloc_bootmem(ai
->nr_groups
* sizeof(group_offsets
[0]));
1763 group_sizes
= alloc_bootmem(ai
->nr_groups
* sizeof(group_sizes
[0]));
1764 unit_map
= alloc_bootmem(nr_cpu_ids
* sizeof(unit_map
[0]));
1765 unit_off
= alloc_bootmem(nr_cpu_ids
* sizeof(unit_off
[0]));
1767 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
1768 unit_map
[cpu
] = UINT_MAX
;
1769 pcpu_first_unit_cpu
= NR_CPUS
;
1771 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
1772 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1774 group_offsets
[group
] = gi
->base_offset
;
1775 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
1777 for (i
= 0; i
< gi
->nr_units
; i
++) {
1778 cpu
= gi
->cpu_map
[i
];
1782 PCPU_SETUP_BUG_ON(cpu
> nr_cpu_ids
);
1783 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
1784 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
1786 unit_map
[cpu
] = unit
+ i
;
1787 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
1789 if (pcpu_first_unit_cpu
== NR_CPUS
)
1790 pcpu_first_unit_cpu
= cpu
;
1793 pcpu_last_unit_cpu
= cpu
;
1794 pcpu_nr_units
= unit
;
1796 for_each_possible_cpu(cpu
)
1797 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
1799 /* we're done parsing the input, undefine BUG macro and dump config */
1800 #undef PCPU_SETUP_BUG_ON
1801 pcpu_dump_alloc_info(KERN_INFO
, ai
);
1803 pcpu_nr_groups
= ai
->nr_groups
;
1804 pcpu_group_offsets
= group_offsets
;
1805 pcpu_group_sizes
= group_sizes
;
1806 pcpu_unit_map
= unit_map
;
1807 pcpu_unit_offsets
= unit_off
;
1809 /* determine basic parameters */
1810 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1811 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
1812 pcpu_atom_size
= ai
->atom_size
;
1813 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
1814 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
1817 * Allocate chunk slots. The additional last slot is for
1820 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
1821 pcpu_slot
= alloc_bootmem(pcpu_nr_slots
* sizeof(pcpu_slot
[0]));
1822 for (i
= 0; i
< pcpu_nr_slots
; i
++)
1823 INIT_LIST_HEAD(&pcpu_slot
[i
]);
1826 * Initialize static chunk. If reserved_size is zero, the
1827 * static chunk covers static area + dynamic allocation area
1828 * in the first chunk. If reserved_size is not zero, it
1829 * covers static area + reserved area (mostly used for module
1830 * static percpu allocation).
1832 schunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1833 INIT_LIST_HEAD(&schunk
->list
);
1834 schunk
->base_addr
= base_addr
;
1836 schunk
->map_alloc
= ARRAY_SIZE(smap
);
1837 schunk
->immutable
= true;
1838 bitmap_fill(schunk
->populated
, pcpu_unit_pages
);
1840 if (ai
->reserved_size
) {
1841 schunk
->free_size
= ai
->reserved_size
;
1842 pcpu_reserved_chunk
= schunk
;
1843 pcpu_reserved_chunk_limit
= ai
->static_size
+ ai
->reserved_size
;
1845 schunk
->free_size
= dyn_size
;
1846 dyn_size
= 0; /* dynamic area covered */
1848 schunk
->contig_hint
= schunk
->free_size
;
1850 schunk
->map
[schunk
->map_used
++] = -ai
->static_size
;
1851 if (schunk
->free_size
)
1852 schunk
->map
[schunk
->map_used
++] = schunk
->free_size
;
1854 /* init dynamic chunk if necessary */
1856 dchunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1857 INIT_LIST_HEAD(&dchunk
->list
);
1858 dchunk
->base_addr
= base_addr
;
1860 dchunk
->map_alloc
= ARRAY_SIZE(dmap
);
1861 dchunk
->immutable
= true;
1862 bitmap_fill(dchunk
->populated
, pcpu_unit_pages
);
1864 dchunk
->contig_hint
= dchunk
->free_size
= dyn_size
;
1865 dchunk
->map
[dchunk
->map_used
++] = -pcpu_reserved_chunk_limit
;
1866 dchunk
->map
[dchunk
->map_used
++] = dchunk
->free_size
;
1869 /* link the first chunk in */
1870 pcpu_first_chunk
= dchunk
?: schunk
;
1871 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
1874 pcpu_base_addr
= base_addr
;
1878 const char *pcpu_fc_names
[PCPU_FC_NR
] __initdata
= {
1879 [PCPU_FC_AUTO
] = "auto",
1880 [PCPU_FC_EMBED
] = "embed",
1881 [PCPU_FC_PAGE
] = "page",
1884 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
1886 static int __init
percpu_alloc_setup(char *str
)
1890 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1891 else if (!strcmp(str
, "embed"))
1892 pcpu_chosen_fc
= PCPU_FC_EMBED
;
1894 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1895 else if (!strcmp(str
, "page"))
1896 pcpu_chosen_fc
= PCPU_FC_PAGE
;
1899 pr_warning("PERCPU: unknown allocator %s specified\n", str
);
1903 early_param("percpu_alloc", percpu_alloc_setup
);
1905 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1906 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1908 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1909 * @reserved_size: the size of reserved percpu area in bytes
1910 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1911 * @atom_size: allocation atom size
1912 * @cpu_distance_fn: callback to determine distance between cpus, optional
1913 * @alloc_fn: function to allocate percpu page
1914 * @free_fn: funtion to free percpu page
1916 * This is a helper to ease setting up embedded first percpu chunk and
1917 * can be called where pcpu_setup_first_chunk() is expected.
1919 * If this function is used to setup the first chunk, it is allocated
1920 * by calling @alloc_fn and used as-is without being mapped into
1921 * vmalloc area. Allocations are always whole multiples of @atom_size
1922 * aligned to @atom_size.
1924 * This enables the first chunk to piggy back on the linear physical
1925 * mapping which often uses larger page size. Please note that this
1926 * can result in very sparse cpu->unit mapping on NUMA machines thus
1927 * requiring large vmalloc address space. Don't use this allocator if
1928 * vmalloc space is not orders of magnitude larger than distances
1929 * between node memory addresses (ie. 32bit NUMA machines).
1931 * When @dyn_size is positive, dynamic area might be larger than
1932 * specified to fill page alignment. When @dyn_size is auto,
1933 * @dyn_size is just big enough to fill page alignment after static
1934 * and reserved areas.
1936 * If the needed size is smaller than the minimum or specified unit
1937 * size, the leftover is returned using @free_fn.
1940 * 0 on success, -errno on failure.
1942 int __init
pcpu_embed_first_chunk(size_t reserved_size
, ssize_t dyn_size
,
1944 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
1945 pcpu_fc_alloc_fn_t alloc_fn
,
1946 pcpu_fc_free_fn_t free_fn
)
1948 void *base
= (void *)ULONG_MAX
;
1949 void **areas
= NULL
;
1950 struct pcpu_alloc_info
*ai
;
1951 size_t size_sum
, areas_size
, max_distance
;
1954 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
1959 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
1960 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
1962 areas
= alloc_bootmem_nopanic(areas_size
);
1968 /* allocate, copy and determine base address */
1969 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1970 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1971 unsigned int cpu
= NR_CPUS
;
1974 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
1975 cpu
= gi
->cpu_map
[i
];
1976 BUG_ON(cpu
== NR_CPUS
);
1978 /* allocate space for the whole group */
1979 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
1982 goto out_free_areas
;
1986 base
= min(ptr
, base
);
1988 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
1989 if (gi
->cpu_map
[i
] == NR_CPUS
) {
1990 /* unused unit, free whole */
1991 free_fn(ptr
, ai
->unit_size
);
1994 /* copy and return the unused part */
1995 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
1996 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
2000 /* base address is now known, determine group base offsets */
2002 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2003 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
2004 max_distance
= max_t(size_t, max_distance
,
2005 ai
->groups
[group
].base_offset
);
2007 max_distance
+= ai
->unit_size
;
2009 /* warn if maximum distance is further than 75% of vmalloc space */
2010 if (max_distance
> (VMALLOC_END
- VMALLOC_START
) * 3 / 4) {
2011 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
2013 max_distance
, VMALLOC_END
- VMALLOC_START
);
2014 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2015 /* and fail if we have fallback */
2021 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2022 PFN_DOWN(size_sum
), base
, ai
->static_size
, ai
->reserved_size
,
2023 ai
->dyn_size
, ai
->unit_size
);
2025 rc
= pcpu_setup_first_chunk(ai
, base
);
2029 for (group
= 0; group
< ai
->nr_groups
; group
++)
2030 free_fn(areas
[group
],
2031 ai
->groups
[group
].nr_units
* ai
->unit_size
);
2033 pcpu_free_alloc_info(ai
);
2035 free_bootmem(__pa(areas
), areas_size
);
2038 #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
2039 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
2041 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2043 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2044 * @reserved_size: the size of reserved percpu area in bytes
2045 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2046 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
2047 * @populate_pte_fn: function to populate pte
2049 * This is a helper to ease setting up page-remapped first percpu
2050 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2052 * This is the basic allocator. Static percpu area is allocated
2053 * page-by-page into vmalloc area.
2056 * 0 on success, -errno on failure.
2058 int __init
pcpu_page_first_chunk(size_t reserved_size
,
2059 pcpu_fc_alloc_fn_t alloc_fn
,
2060 pcpu_fc_free_fn_t free_fn
,
2061 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
2063 static struct vm_struct vm
;
2064 struct pcpu_alloc_info
*ai
;
2068 struct page
**pages
;
2071 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
2073 ai
= pcpu_build_alloc_info(reserved_size
, -1, PAGE_SIZE
, NULL
);
2076 BUG_ON(ai
->nr_groups
!= 1);
2077 BUG_ON(ai
->groups
[0].nr_units
!= num_possible_cpus());
2079 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2081 /* unaligned allocations can't be freed, round up to page size */
2082 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
2084 pages
= alloc_bootmem(pages_size
);
2086 /* allocate pages */
2088 for (unit
= 0; unit
< num_possible_cpus(); unit
++)
2089 for (i
= 0; i
< unit_pages
; i
++) {
2090 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
2093 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
2095 pr_warning("PERCPU: failed to allocate %s page "
2096 "for cpu%u\n", psize_str
, cpu
);
2099 pages
[j
++] = virt_to_page(ptr
);
2102 /* allocate vm area, map the pages and copy static data */
2103 vm
.flags
= VM_ALLOC
;
2104 vm
.size
= num_possible_cpus() * ai
->unit_size
;
2105 vm_area_register_early(&vm
, PAGE_SIZE
);
2107 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2108 unsigned long unit_addr
=
2109 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
2111 for (i
= 0; i
< unit_pages
; i
++)
2112 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
2114 /* pte already populated, the following shouldn't fail */
2115 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
2118 panic("failed to map percpu area, err=%d\n", rc
);
2121 * FIXME: Archs with virtual cache should flush local
2122 * cache for the linear mapping here - something
2123 * equivalent to flush_cache_vmap() on the local cpu.
2124 * flush_cache_vmap() can't be used as most supporting
2125 * data structures are not set up yet.
2128 /* copy static data */
2129 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
2132 /* we're ready, commit */
2133 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
2134 unit_pages
, psize_str
, vm
.addr
, ai
->static_size
,
2135 ai
->reserved_size
, ai
->dyn_size
);
2137 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
2142 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
2145 free_bootmem(__pa(pages
), pages_size
);
2146 pcpu_free_alloc_info(ai
);
2149 #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
2152 * Generic percpu area setup.
2154 * The embedding helper is used because its behavior closely resembles
2155 * the original non-dynamic generic percpu area setup. This is
2156 * important because many archs have addressing restrictions and might
2157 * fail if the percpu area is located far away from the previous
2158 * location. As an added bonus, in non-NUMA cases, embedding is
2159 * generally a good idea TLB-wise because percpu area can piggy back
2160 * on the physical linear memory mapping which uses large page
2161 * mappings on applicable archs.
2163 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2164 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
2165 EXPORT_SYMBOL(__per_cpu_offset
);
2167 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
2170 return __alloc_bootmem_nopanic(size
, align
, __pa(MAX_DMA_ADDRESS
));
2173 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
2175 free_bootmem(__pa(ptr
), size
);
2178 void __init
setup_per_cpu_areas(void)
2180 unsigned long delta
;
2185 * Always reserve area for module percpu variables. That's
2186 * what the legacy allocator did.
2188 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
2189 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
2190 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
2192 panic("Failed to initialized percpu areas.");
2194 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
2195 for_each_possible_cpu(cpu
)
2196 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
2198 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */