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>
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 */
79 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
80 #ifndef __addr_to_pcpu_ptr
81 #define __addr_to_pcpu_ptr(addr) \
82 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \
83 + (unsigned long)__per_cpu_start)
85 #ifndef __pcpu_ptr_to_addr
86 #define __pcpu_ptr_to_addr(ptr) \
87 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \
88 - (unsigned long)__per_cpu_start)
92 struct list_head list
; /* linked to pcpu_slot lists */
93 int free_size
; /* free bytes in the chunk */
94 int contig_hint
; /* max contiguous size hint */
95 void *base_addr
; /* base address of this chunk */
96 int map_used
; /* # of map entries used */
97 int map_alloc
; /* # of map entries allocated */
98 int *map
; /* allocation map */
99 struct vm_struct
**vms
; /* mapped vmalloc regions */
100 bool immutable
; /* no [de]population allowed */
101 unsigned long populated
[]; /* populated bitmap */
104 static int pcpu_unit_pages __read_mostly
;
105 static int pcpu_unit_size __read_mostly
;
106 static int pcpu_nr_units __read_mostly
;
107 static int pcpu_atom_size __read_mostly
;
108 static int pcpu_nr_slots __read_mostly
;
109 static size_t pcpu_chunk_struct_size __read_mostly
;
111 /* cpus with the lowest and highest unit numbers */
112 static unsigned int pcpu_first_unit_cpu __read_mostly
;
113 static unsigned int pcpu_last_unit_cpu __read_mostly
;
115 /* the address of the first chunk which starts with the kernel static area */
116 void *pcpu_base_addr __read_mostly
;
117 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
119 static const int *pcpu_unit_map __read_mostly
; /* cpu -> unit */
120 const unsigned long *pcpu_unit_offsets __read_mostly
; /* cpu -> unit offset */
122 /* group information, used for vm allocation */
123 static int pcpu_nr_groups __read_mostly
;
124 static const unsigned long *pcpu_group_offsets __read_mostly
;
125 static const size_t *pcpu_group_sizes __read_mostly
;
128 * The first chunk which always exists. Note that unlike other
129 * chunks, this one can be allocated and mapped in several different
130 * ways and thus often doesn't live in the vmalloc area.
132 static struct pcpu_chunk
*pcpu_first_chunk
;
135 * Optional reserved chunk. This chunk reserves part of the first
136 * chunk and serves it for reserved allocations. The amount of
137 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
138 * area doesn't exist, the following variables contain NULL and 0
141 static struct pcpu_chunk
*pcpu_reserved_chunk
;
142 static int pcpu_reserved_chunk_limit
;
145 * Synchronization rules.
147 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
148 * protects allocation/reclaim paths, chunks, populated bitmap and
149 * vmalloc mapping. The latter is a spinlock and protects the index
150 * data structures - chunk slots, chunks and area maps in chunks.
152 * During allocation, pcpu_alloc_mutex is kept locked all the time and
153 * pcpu_lock is grabbed and released as necessary. All actual memory
154 * allocations are done using GFP_KERNEL with pcpu_lock released.
156 * Free path accesses and alters only the index data structures, so it
157 * can be safely called from atomic context. When memory needs to be
158 * returned to the system, free path schedules reclaim_work which
159 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
160 * reclaimed, release both locks and frees the chunks. Note that it's
161 * necessary to grab both locks to remove a chunk from circulation as
162 * allocation path might be referencing the chunk with only
163 * pcpu_alloc_mutex locked.
165 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* protects whole alloc and reclaim */
166 static DEFINE_SPINLOCK(pcpu_lock
); /* protects index data structures */
168 static struct list_head
*pcpu_slot __read_mostly
; /* chunk list slots */
170 /* reclaim work to release fully free chunks, scheduled from free path */
171 static void pcpu_reclaim(struct work_struct
*work
);
172 static DECLARE_WORK(pcpu_reclaim_work
, pcpu_reclaim
);
174 static int __pcpu_size_to_slot(int size
)
176 int highbit
= fls(size
); /* size is in bytes */
177 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
180 static int pcpu_size_to_slot(int size
)
182 if (size
== pcpu_unit_size
)
183 return pcpu_nr_slots
- 1;
184 return __pcpu_size_to_slot(size
);
187 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
189 if (chunk
->free_size
< sizeof(int) || chunk
->contig_hint
< sizeof(int))
192 return pcpu_size_to_slot(chunk
->free_size
);
195 static int pcpu_page_idx(unsigned int cpu
, int page_idx
)
197 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
200 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
201 unsigned int cpu
, int page_idx
)
203 return (unsigned long)chunk
->base_addr
+ pcpu_unit_offsets
[cpu
] +
204 (page_idx
<< PAGE_SHIFT
);
207 static struct page
*pcpu_chunk_page(struct pcpu_chunk
*chunk
,
208 unsigned int cpu
, int page_idx
)
210 /* must not be used on pre-mapped chunk */
211 WARN_ON(chunk
->immutable
);
213 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk
, cpu
, page_idx
));
216 /* set the pointer to a chunk in a page struct */
217 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
219 page
->index
= (unsigned long)pcpu
;
222 /* obtain pointer to a chunk from a page struct */
223 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
225 return (struct pcpu_chunk
*)page
->index
;
228 static void pcpu_next_unpop(struct pcpu_chunk
*chunk
, int *rs
, int *re
, int end
)
230 *rs
= find_next_zero_bit(chunk
->populated
, end
, *rs
);
231 *re
= find_next_bit(chunk
->populated
, end
, *rs
+ 1);
234 static void pcpu_next_pop(struct pcpu_chunk
*chunk
, int *rs
, int *re
, int end
)
236 *rs
= find_next_bit(chunk
->populated
, end
, *rs
);
237 *re
= find_next_zero_bit(chunk
->populated
, end
, *rs
+ 1);
241 * (Un)populated page region iterators. Iterate over (un)populated
242 * page regions betwen @start and @end in @chunk. @rs and @re should
243 * be integer variables and will be set to start and end page index of
244 * the current region.
246 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
247 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
249 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
251 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
252 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
254 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
257 * pcpu_mem_alloc - allocate memory
258 * @size: bytes to allocate
260 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
261 * kzalloc() is used; otherwise, vmalloc() is used. The returned
262 * memory is always zeroed.
265 * Does GFP_KERNEL allocation.
268 * Pointer to the allocated area on success, NULL on failure.
270 static void *pcpu_mem_alloc(size_t size
)
272 if (size
<= PAGE_SIZE
)
273 return kzalloc(size
, GFP_KERNEL
);
275 void *ptr
= vmalloc(size
);
277 memset(ptr
, 0, size
);
283 * pcpu_mem_free - free memory
284 * @ptr: memory to free
285 * @size: size of the area
287 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
289 static void pcpu_mem_free(void *ptr
, size_t size
)
291 if (size
<= PAGE_SIZE
)
298 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
299 * @chunk: chunk of interest
300 * @oslot: the previous slot it was on
302 * This function is called after an allocation or free changed @chunk.
303 * New slot according to the changed state is determined and @chunk is
304 * moved to the slot. Note that the reserved chunk is never put on
310 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
312 int nslot
= pcpu_chunk_slot(chunk
);
314 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
316 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
318 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
323 * pcpu_chunk_addr_search - determine chunk containing specified address
324 * @addr: address for which the chunk needs to be determined.
327 * The address of the found chunk.
329 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
331 void *first_start
= pcpu_first_chunk
->base_addr
;
333 /* is it in the first chunk? */
334 if (addr
>= first_start
&& addr
< first_start
+ pcpu_unit_size
) {
335 /* is it in the reserved area? */
336 if (addr
< first_start
+ pcpu_reserved_chunk_limit
)
337 return pcpu_reserved_chunk
;
338 return pcpu_first_chunk
;
342 * The address is relative to unit0 which might be unused and
343 * thus unmapped. Offset the address to the unit space of the
344 * current processor before looking it up in the vmalloc
345 * space. Note that any possible cpu id can be used here, so
346 * there's no need to worry about preemption or cpu hotplug.
348 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
349 return pcpu_get_page_chunk(vmalloc_to_page(addr
));
353 * pcpu_extend_area_map - extend area map for allocation
354 * @chunk: target chunk
356 * Extend area map of @chunk so that it can accomodate an allocation.
357 * A single allocation can split an area into three areas, so this
358 * function makes sure that @chunk->map has at least two extra slots.
361 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
362 * if area map is extended.
365 * 0 if noop, 1 if successfully extended, -errno on failure.
367 static int pcpu_extend_area_map(struct pcpu_chunk
*chunk
)
374 if (chunk
->map_alloc
>= chunk
->map_used
+ 2)
377 spin_unlock_irq(&pcpu_lock
);
379 new_alloc
= PCPU_DFL_MAP_ALLOC
;
380 while (new_alloc
< chunk
->map_used
+ 2)
383 new = pcpu_mem_alloc(new_alloc
* sizeof(new[0]));
385 spin_lock_irq(&pcpu_lock
);
390 * Acquire pcpu_lock and switch to new area map. Only free
391 * could have happened inbetween, so map_used couldn't have
394 spin_lock_irq(&pcpu_lock
);
395 BUG_ON(new_alloc
< chunk
->map_used
+ 2);
397 size
= chunk
->map_alloc
* sizeof(chunk
->map
[0]);
398 memcpy(new, chunk
->map
, size
);
401 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
402 * one of the first chunks and still using static map.
404 if (chunk
->map_alloc
>= PCPU_DFL_MAP_ALLOC
)
405 pcpu_mem_free(chunk
->map
, size
);
407 chunk
->map_alloc
= new_alloc
;
413 * pcpu_split_block - split a map block
414 * @chunk: chunk of interest
415 * @i: index of map block to split
416 * @head: head size in bytes (can be 0)
417 * @tail: tail size in bytes (can be 0)
419 * Split the @i'th map block into two or three blocks. If @head is
420 * non-zero, @head bytes block is inserted before block @i moving it
421 * to @i+1 and reducing its size by @head bytes.
423 * If @tail is non-zero, the target block, which can be @i or @i+1
424 * depending on @head, is reduced by @tail bytes and @tail byte block
425 * is inserted after the target block.
427 * @chunk->map must have enough free slots to accomodate the split.
432 static void pcpu_split_block(struct pcpu_chunk
*chunk
, int i
,
435 int nr_extra
= !!head
+ !!tail
;
437 BUG_ON(chunk
->map_alloc
< chunk
->map_used
+ nr_extra
);
439 /* insert new subblocks */
440 memmove(&chunk
->map
[i
+ nr_extra
], &chunk
->map
[i
],
441 sizeof(chunk
->map
[0]) * (chunk
->map_used
- i
));
442 chunk
->map_used
+= nr_extra
;
445 chunk
->map
[i
+ 1] = chunk
->map
[i
] - head
;
446 chunk
->map
[i
++] = head
;
449 chunk
->map
[i
++] -= tail
;
450 chunk
->map
[i
] = tail
;
455 * pcpu_alloc_area - allocate area from a pcpu_chunk
456 * @chunk: chunk of interest
457 * @size: wanted size in bytes
458 * @align: wanted align
460 * Try to allocate @size bytes area aligned at @align from @chunk.
461 * Note that this function only allocates the offset. It doesn't
462 * populate or map the area.
464 * @chunk->map must have at least two free slots.
470 * Allocated offset in @chunk on success, -1 if no matching area is
473 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int size
, int align
)
475 int oslot
= pcpu_chunk_slot(chunk
);
479 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++])) {
480 bool is_last
= i
+ 1 == chunk
->map_used
;
483 /* extra for alignment requirement */
484 head
= ALIGN(off
, align
) - off
;
485 BUG_ON(i
== 0 && head
!= 0);
487 if (chunk
->map
[i
] < 0)
489 if (chunk
->map
[i
] < head
+ size
) {
490 max_contig
= max(chunk
->map
[i
], max_contig
);
495 * If head is small or the previous block is free,
496 * merge'em. Note that 'small' is defined as smaller
497 * than sizeof(int), which is very small but isn't too
498 * uncommon for percpu allocations.
500 if (head
&& (head
< sizeof(int) || chunk
->map
[i
- 1] > 0)) {
501 if (chunk
->map
[i
- 1] > 0)
502 chunk
->map
[i
- 1] += head
;
504 chunk
->map
[i
- 1] -= head
;
505 chunk
->free_size
-= head
;
507 chunk
->map
[i
] -= head
;
512 /* if tail is small, just keep it around */
513 tail
= chunk
->map
[i
] - head
- size
;
514 if (tail
< sizeof(int))
517 /* split if warranted */
519 pcpu_split_block(chunk
, i
, head
, tail
);
523 max_contig
= max(chunk
->map
[i
- 1], max_contig
);
526 max_contig
= max(chunk
->map
[i
+ 1], max_contig
);
529 /* update hint and mark allocated */
531 chunk
->contig_hint
= max_contig
; /* fully scanned */
533 chunk
->contig_hint
= max(chunk
->contig_hint
,
536 chunk
->free_size
-= chunk
->map
[i
];
537 chunk
->map
[i
] = -chunk
->map
[i
];
539 pcpu_chunk_relocate(chunk
, oslot
);
543 chunk
->contig_hint
= max_contig
; /* fully scanned */
544 pcpu_chunk_relocate(chunk
, oslot
);
546 /* tell the upper layer that this chunk has no matching area */
551 * pcpu_free_area - free area to a pcpu_chunk
552 * @chunk: chunk of interest
553 * @freeme: offset of area to free
555 * Free area starting from @freeme to @chunk. Note that this function
556 * only modifies the allocation map. It doesn't depopulate or unmap
562 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int freeme
)
564 int oslot
= pcpu_chunk_slot(chunk
);
567 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++]))
570 BUG_ON(off
!= freeme
);
571 BUG_ON(chunk
->map
[i
] > 0);
573 chunk
->map
[i
] = -chunk
->map
[i
];
574 chunk
->free_size
+= chunk
->map
[i
];
576 /* merge with previous? */
577 if (i
> 0 && chunk
->map
[i
- 1] >= 0) {
578 chunk
->map
[i
- 1] += chunk
->map
[i
];
580 memmove(&chunk
->map
[i
], &chunk
->map
[i
+ 1],
581 (chunk
->map_used
- i
) * sizeof(chunk
->map
[0]));
584 /* merge with next? */
585 if (i
+ 1 < chunk
->map_used
&& chunk
->map
[i
+ 1] >= 0) {
586 chunk
->map
[i
] += chunk
->map
[i
+ 1];
588 memmove(&chunk
->map
[i
+ 1], &chunk
->map
[i
+ 2],
589 (chunk
->map_used
- (i
+ 1)) * sizeof(chunk
->map
[0]));
592 chunk
->contig_hint
= max(chunk
->map
[i
], chunk
->contig_hint
);
593 pcpu_chunk_relocate(chunk
, oslot
);
597 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
598 * @chunk: chunk of interest
599 * @bitmapp: output parameter for bitmap
600 * @may_alloc: may allocate the array
602 * Returns pointer to array of pointers to struct page and bitmap,
603 * both of which can be indexed with pcpu_page_idx(). The returned
604 * array is cleared to zero and *@bitmapp is copied from
605 * @chunk->populated. Note that there is only one array and bitmap
606 * and access exclusion is the caller's responsibility.
609 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
610 * Otherwise, don't care.
613 * Pointer to temp pages array on success, NULL on failure.
615 static struct page
**pcpu_get_pages_and_bitmap(struct pcpu_chunk
*chunk
,
616 unsigned long **bitmapp
,
619 static struct page
**pages
;
620 static unsigned long *bitmap
;
621 size_t pages_size
= pcpu_nr_units
* pcpu_unit_pages
* sizeof(pages
[0]);
622 size_t bitmap_size
= BITS_TO_LONGS(pcpu_unit_pages
) *
623 sizeof(unsigned long);
625 if (!pages
|| !bitmap
) {
626 if (may_alloc
&& !pages
)
627 pages
= pcpu_mem_alloc(pages_size
);
628 if (may_alloc
&& !bitmap
)
629 bitmap
= pcpu_mem_alloc(bitmap_size
);
630 if (!pages
|| !bitmap
)
634 memset(pages
, 0, pages_size
);
635 bitmap_copy(bitmap
, chunk
->populated
, pcpu_unit_pages
);
642 * pcpu_free_pages - free pages which were allocated for @chunk
643 * @chunk: chunk pages were allocated for
644 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
645 * @populated: populated bitmap
646 * @page_start: page index of the first page to be freed
647 * @page_end: page index of the last page to be freed + 1
649 * Free pages [@page_start and @page_end) in @pages for all units.
650 * The pages were allocated for @chunk.
652 static void pcpu_free_pages(struct pcpu_chunk
*chunk
,
653 struct page
**pages
, unsigned long *populated
,
654 int page_start
, int page_end
)
659 for_each_possible_cpu(cpu
) {
660 for (i
= page_start
; i
< page_end
; i
++) {
661 struct page
*page
= pages
[pcpu_page_idx(cpu
, i
)];
670 * pcpu_alloc_pages - allocates pages for @chunk
671 * @chunk: target chunk
672 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
673 * @populated: populated bitmap
674 * @page_start: page index of the first page to be allocated
675 * @page_end: page index of the last page to be allocated + 1
677 * Allocate pages [@page_start,@page_end) into @pages for all units.
678 * The allocation is for @chunk. Percpu core doesn't care about the
679 * content of @pages and will pass it verbatim to pcpu_map_pages().
681 static int pcpu_alloc_pages(struct pcpu_chunk
*chunk
,
682 struct page
**pages
, unsigned long *populated
,
683 int page_start
, int page_end
)
685 const gfp_t gfp
= GFP_KERNEL
| __GFP_HIGHMEM
| __GFP_COLD
;
689 for_each_possible_cpu(cpu
) {
690 for (i
= page_start
; i
< page_end
; i
++) {
691 struct page
**pagep
= &pages
[pcpu_page_idx(cpu
, i
)];
693 *pagep
= alloc_pages_node(cpu_to_node(cpu
), gfp
, 0);
695 pcpu_free_pages(chunk
, pages
, populated
,
696 page_start
, page_end
);
705 * pcpu_pre_unmap_flush - flush cache prior to unmapping
706 * @chunk: chunk the regions to be flushed belongs to
707 * @page_start: page index of the first page to be flushed
708 * @page_end: page index of the last page to be flushed + 1
710 * Pages in [@page_start,@page_end) of @chunk are about to be
711 * unmapped. Flush cache. As each flushing trial can be very
712 * expensive, issue flush on the whole region at once rather than
713 * doing it for each cpu. This could be an overkill but is more
716 static void pcpu_pre_unmap_flush(struct pcpu_chunk
*chunk
,
717 int page_start
, int page_end
)
720 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
721 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
724 static void __pcpu_unmap_pages(unsigned long addr
, int nr_pages
)
726 unmap_kernel_range_noflush(addr
, nr_pages
<< PAGE_SHIFT
);
730 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
731 * @chunk: chunk of interest
732 * @pages: pages array which can be used to pass information to free
733 * @populated: populated bitmap
734 * @page_start: page index of the first page to unmap
735 * @page_end: page index of the last page to unmap + 1
737 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
738 * Corresponding elements in @pages were cleared by the caller and can
739 * be used to carry information to pcpu_free_pages() which will be
740 * called after all unmaps are finished. The caller should call
741 * proper pre/post flush functions.
743 static void pcpu_unmap_pages(struct pcpu_chunk
*chunk
,
744 struct page
**pages
, unsigned long *populated
,
745 int page_start
, int page_end
)
750 for_each_possible_cpu(cpu
) {
751 for (i
= page_start
; i
< page_end
; i
++) {
754 page
= pcpu_chunk_page(chunk
, cpu
, i
);
756 pages
[pcpu_page_idx(cpu
, i
)] = page
;
758 __pcpu_unmap_pages(pcpu_chunk_addr(chunk
, cpu
, page_start
),
759 page_end
- page_start
);
762 for (i
= page_start
; i
< page_end
; i
++)
763 __clear_bit(i
, populated
);
767 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
768 * @chunk: pcpu_chunk the regions to be flushed belong to
769 * @page_start: page index of the first page to be flushed
770 * @page_end: page index of the last page to be flushed + 1
772 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
773 * TLB for the regions. This can be skipped if the area is to be
774 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
776 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
777 * for the whole region.
779 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk
*chunk
,
780 int page_start
, int page_end
)
782 flush_tlb_kernel_range(
783 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
784 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
787 static int __pcpu_map_pages(unsigned long addr
, struct page
**pages
,
790 return map_kernel_range_noflush(addr
, nr_pages
<< PAGE_SHIFT
,
795 * pcpu_map_pages - map pages into a pcpu_chunk
796 * @chunk: chunk of interest
797 * @pages: pages array containing pages to be mapped
798 * @populated: populated bitmap
799 * @page_start: page index of the first page to map
800 * @page_end: page index of the last page to map + 1
802 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
803 * caller is responsible for calling pcpu_post_map_flush() after all
804 * mappings are complete.
806 * This function is responsible for setting corresponding bits in
807 * @chunk->populated bitmap and whatever is necessary for reverse
808 * lookup (addr -> chunk).
810 static int pcpu_map_pages(struct pcpu_chunk
*chunk
,
811 struct page
**pages
, unsigned long *populated
,
812 int page_start
, int page_end
)
814 unsigned int cpu
, tcpu
;
817 for_each_possible_cpu(cpu
) {
818 err
= __pcpu_map_pages(pcpu_chunk_addr(chunk
, cpu
, page_start
),
819 &pages
[pcpu_page_idx(cpu
, page_start
)],
820 page_end
- page_start
);
825 /* mapping successful, link chunk and mark populated */
826 for (i
= page_start
; i
< page_end
; i
++) {
827 for_each_possible_cpu(cpu
)
828 pcpu_set_page_chunk(pages
[pcpu_page_idx(cpu
, i
)],
830 __set_bit(i
, populated
);
836 for_each_possible_cpu(tcpu
) {
839 __pcpu_unmap_pages(pcpu_chunk_addr(chunk
, tcpu
, page_start
),
840 page_end
- page_start
);
846 * pcpu_post_map_flush - flush cache after mapping
847 * @chunk: pcpu_chunk the regions to be flushed belong to
848 * @page_start: page index of the first page to be flushed
849 * @page_end: page index of the last page to be flushed + 1
851 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
854 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
855 * for the whole region.
857 static void pcpu_post_map_flush(struct pcpu_chunk
*chunk
,
858 int page_start
, int page_end
)
861 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
862 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
866 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
867 * @chunk: chunk to depopulate
868 * @off: offset to the area to depopulate
869 * @size: size of the area to depopulate in bytes
870 * @flush: whether to flush cache and tlb or not
872 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
873 * from @chunk. If @flush is true, vcache is flushed before unmapping
879 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
)
881 int page_start
= PFN_DOWN(off
);
882 int page_end
= PFN_UP(off
+ size
);
884 unsigned long *populated
;
887 /* quick path, check whether it's empty already */
888 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
889 if (rs
== page_start
&& re
== page_end
)
894 /* immutable chunks can't be depopulated */
895 WARN_ON(chunk
->immutable
);
898 * If control reaches here, there must have been at least one
899 * successful population attempt so the temp pages array must
902 pages
= pcpu_get_pages_and_bitmap(chunk
, &populated
, false);
906 pcpu_pre_unmap_flush(chunk
, page_start
, page_end
);
908 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
)
909 pcpu_unmap_pages(chunk
, pages
, populated
, rs
, re
);
911 /* no need to flush tlb, vmalloc will handle it lazily */
913 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
)
914 pcpu_free_pages(chunk
, pages
, populated
, rs
, re
);
916 /* commit new bitmap */
917 bitmap_copy(chunk
->populated
, populated
, pcpu_unit_pages
);
921 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
922 * @chunk: chunk of interest
923 * @off: offset to the area to populate
924 * @size: size of the area to populate in bytes
926 * For each cpu, populate and map pages [@page_start,@page_end) into
927 * @chunk. The area is cleared on return.
930 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
932 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
)
934 int page_start
= PFN_DOWN(off
);
935 int page_end
= PFN_UP(off
+ size
);
936 int free_end
= page_start
, unmap_end
= page_start
;
938 unsigned long *populated
;
942 /* quick path, check whether all pages are already there */
943 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
) {
944 if (rs
== page_start
&& re
== page_end
)
949 /* need to allocate and map pages, this chunk can't be immutable */
950 WARN_ON(chunk
->immutable
);
952 pages
= pcpu_get_pages_and_bitmap(chunk
, &populated
, true);
957 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
958 rc
= pcpu_alloc_pages(chunk
, pages
, populated
, rs
, re
);
964 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
965 rc
= pcpu_map_pages(chunk
, pages
, populated
, rs
, re
);
970 pcpu_post_map_flush(chunk
, page_start
, page_end
);
972 /* commit new bitmap */
973 bitmap_copy(chunk
->populated
, populated
, pcpu_unit_pages
);
975 for_each_possible_cpu(cpu
)
976 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
980 pcpu_pre_unmap_flush(chunk
, page_start
, unmap_end
);
981 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, unmap_end
)
982 pcpu_unmap_pages(chunk
, pages
, populated
, rs
, re
);
983 pcpu_post_unmap_tlb_flush(chunk
, page_start
, unmap_end
);
985 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, free_end
)
986 pcpu_free_pages(chunk
, pages
, populated
, rs
, re
);
990 static void free_pcpu_chunk(struct pcpu_chunk
*chunk
)
995 pcpu_free_vm_areas(chunk
->vms
, pcpu_nr_groups
);
996 pcpu_mem_free(chunk
->map
, chunk
->map_alloc
* sizeof(chunk
->map
[0]));
1000 static struct pcpu_chunk
*alloc_pcpu_chunk(void)
1002 struct pcpu_chunk
*chunk
;
1004 chunk
= kzalloc(pcpu_chunk_struct_size
, GFP_KERNEL
);
1008 chunk
->map
= pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC
* sizeof(chunk
->map
[0]));
1009 chunk
->map_alloc
= PCPU_DFL_MAP_ALLOC
;
1010 chunk
->map
[chunk
->map_used
++] = pcpu_unit_size
;
1012 chunk
->vms
= pcpu_get_vm_areas(pcpu_group_offsets
, pcpu_group_sizes
,
1013 pcpu_nr_groups
, pcpu_atom_size
,
1016 free_pcpu_chunk(chunk
);
1020 INIT_LIST_HEAD(&chunk
->list
);
1021 chunk
->free_size
= pcpu_unit_size
;
1022 chunk
->contig_hint
= pcpu_unit_size
;
1023 chunk
->base_addr
= chunk
->vms
[0]->addr
- pcpu_group_offsets
[0];
1029 * pcpu_alloc - the percpu allocator
1030 * @size: size of area to allocate in bytes
1031 * @align: alignment of area (max PAGE_SIZE)
1032 * @reserved: allocate from the reserved chunk if available
1034 * Allocate percpu area of @size bytes aligned at @align.
1037 * Does GFP_KERNEL allocation.
1040 * Percpu pointer to the allocated area on success, NULL on failure.
1042 static void *pcpu_alloc(size_t size
, size_t align
, bool reserved
)
1044 static int warn_limit
= 10;
1045 struct pcpu_chunk
*chunk
;
1049 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
)) {
1050 WARN(true, "illegal size (%zu) or align (%zu) for "
1051 "percpu allocation\n", size
, align
);
1055 mutex_lock(&pcpu_alloc_mutex
);
1056 spin_lock_irq(&pcpu_lock
);
1058 /* serve reserved allocations from the reserved chunk if available */
1059 if (reserved
&& pcpu_reserved_chunk
) {
1060 chunk
= pcpu_reserved_chunk
;
1061 if (size
> chunk
->contig_hint
||
1062 pcpu_extend_area_map(chunk
) < 0) {
1063 err
= "failed to extend area map of reserved chunk";
1066 off
= pcpu_alloc_area(chunk
, size
, align
);
1069 err
= "alloc from reserved chunk failed";
1074 /* search through normal chunks */
1075 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
1076 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1077 if (size
> chunk
->contig_hint
)
1080 switch (pcpu_extend_area_map(chunk
)) {
1084 goto restart
; /* pcpu_lock dropped, restart */
1086 err
= "failed to extend area map";
1090 off
= pcpu_alloc_area(chunk
, size
, align
);
1096 /* hmmm... no space left, create a new chunk */
1097 spin_unlock_irq(&pcpu_lock
);
1099 chunk
= alloc_pcpu_chunk();
1101 err
= "failed to allocate new chunk";
1102 goto fail_unlock_mutex
;
1105 spin_lock_irq(&pcpu_lock
);
1106 pcpu_chunk_relocate(chunk
, -1);
1110 spin_unlock_irq(&pcpu_lock
);
1112 /* populate, map and clear the area */
1113 if (pcpu_populate_chunk(chunk
, off
, size
)) {
1114 spin_lock_irq(&pcpu_lock
);
1115 pcpu_free_area(chunk
, off
);
1116 err
= "failed to populate";
1120 mutex_unlock(&pcpu_alloc_mutex
);
1122 /* return address relative to base address */
1123 return __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
1126 spin_unlock_irq(&pcpu_lock
);
1128 mutex_unlock(&pcpu_alloc_mutex
);
1130 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1131 "%s\n", size
, align
, err
);
1134 pr_info("PERCPU: limit reached, disable warning\n");
1140 * __alloc_percpu - allocate dynamic percpu area
1141 * @size: size of area to allocate in bytes
1142 * @align: alignment of area (max PAGE_SIZE)
1144 * Allocate percpu area of @size bytes aligned at @align. Might
1145 * sleep. Might trigger writeouts.
1148 * Does GFP_KERNEL allocation.
1151 * Percpu pointer to the allocated area on success, NULL on failure.
1153 void *__alloc_percpu(size_t size
, size_t align
)
1155 return pcpu_alloc(size
, align
, false);
1157 EXPORT_SYMBOL_GPL(__alloc_percpu
);
1160 * __alloc_reserved_percpu - allocate reserved percpu area
1161 * @size: size of area to allocate in bytes
1162 * @align: alignment of area (max PAGE_SIZE)
1164 * Allocate percpu area of @size bytes aligned at @align from reserved
1165 * percpu area if arch has set it up; otherwise, allocation is served
1166 * from the same dynamic area. Might sleep. Might trigger writeouts.
1169 * Does GFP_KERNEL allocation.
1172 * Percpu pointer to the allocated area on success, NULL on failure.
1174 void *__alloc_reserved_percpu(size_t size
, size_t align
)
1176 return pcpu_alloc(size
, align
, true);
1180 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1183 * Reclaim all fully free chunks except for the first one.
1186 * workqueue context.
1188 static void pcpu_reclaim(struct work_struct
*work
)
1191 struct list_head
*head
= &pcpu_slot
[pcpu_nr_slots
- 1];
1192 struct pcpu_chunk
*chunk
, *next
;
1194 mutex_lock(&pcpu_alloc_mutex
);
1195 spin_lock_irq(&pcpu_lock
);
1197 list_for_each_entry_safe(chunk
, next
, head
, list
) {
1198 WARN_ON(chunk
->immutable
);
1200 /* spare the first one */
1201 if (chunk
== list_first_entry(head
, struct pcpu_chunk
, list
))
1204 list_move(&chunk
->list
, &todo
);
1207 spin_unlock_irq(&pcpu_lock
);
1209 list_for_each_entry_safe(chunk
, next
, &todo
, list
) {
1210 pcpu_depopulate_chunk(chunk
, 0, pcpu_unit_size
);
1211 free_pcpu_chunk(chunk
);
1214 mutex_unlock(&pcpu_alloc_mutex
);
1218 * free_percpu - free percpu area
1219 * @ptr: pointer to area to free
1221 * Free percpu area @ptr.
1224 * Can be called from atomic context.
1226 void free_percpu(void *ptr
)
1229 struct pcpu_chunk
*chunk
;
1230 unsigned long flags
;
1236 addr
= __pcpu_ptr_to_addr(ptr
);
1238 spin_lock_irqsave(&pcpu_lock
, flags
);
1240 chunk
= pcpu_chunk_addr_search(addr
);
1241 off
= addr
- chunk
->base_addr
;
1243 pcpu_free_area(chunk
, off
);
1245 /* if there are more than one fully free chunks, wake up grim reaper */
1246 if (chunk
->free_size
== pcpu_unit_size
) {
1247 struct pcpu_chunk
*pos
;
1249 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
1251 schedule_work(&pcpu_reclaim_work
);
1256 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1258 EXPORT_SYMBOL_GPL(free_percpu
);
1260 static inline size_t pcpu_calc_fc_sizes(size_t static_size
,
1261 size_t reserved_size
,
1266 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
1267 (*dyn_sizep
>= 0 ? *dyn_sizep
: 0));
1268 if (*dyn_sizep
!= 0)
1269 *dyn_sizep
= size_sum
- static_size
- reserved_size
;
1275 * pcpu_alloc_alloc_info - allocate percpu allocation info
1276 * @nr_groups: the number of groups
1277 * @nr_units: the number of units
1279 * Allocate ai which is large enough for @nr_groups groups containing
1280 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1281 * cpu_map array which is long enough for @nr_units and filled with
1282 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1283 * pointer of other groups.
1286 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1289 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
1292 struct pcpu_alloc_info
*ai
;
1293 size_t base_size
, ai_size
;
1297 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
1298 __alignof__(ai
->groups
[0].cpu_map
[0]));
1299 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
1301 ptr
= alloc_bootmem_nopanic(PFN_ALIGN(ai_size
));
1307 ai
->groups
[0].cpu_map
= ptr
;
1309 for (unit
= 0; unit
< nr_units
; unit
++)
1310 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
1312 ai
->nr_groups
= nr_groups
;
1313 ai
->__ai_size
= PFN_ALIGN(ai_size
);
1319 * pcpu_free_alloc_info - free percpu allocation info
1320 * @ai: pcpu_alloc_info to free
1322 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1324 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
1326 free_bootmem(__pa(ai
), ai
->__ai_size
);
1330 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1331 * @reserved_size: the size of reserved percpu area in bytes
1332 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1333 * @atom_size: allocation atom size
1334 * @cpu_distance_fn: callback to determine distance between cpus, optional
1336 * This function determines grouping of units, their mappings to cpus
1337 * and other parameters considering needed percpu size, allocation
1338 * atom size and distances between CPUs.
1340 * Groups are always mutliples of atom size and CPUs which are of
1341 * LOCAL_DISTANCE both ways are grouped together and share space for
1342 * units in the same group. The returned configuration is guaranteed
1343 * to have CPUs on different nodes on different groups and >=75% usage
1344 * of allocated virtual address space.
1347 * On success, pointer to the new allocation_info is returned. On
1348 * failure, ERR_PTR value is returned.
1350 struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
1351 size_t reserved_size
, ssize_t dyn_size
,
1353 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
1355 static int group_map
[NR_CPUS
] __initdata
;
1356 static int group_cnt
[NR_CPUS
] __initdata
;
1357 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1358 int group_cnt_max
= 0, nr_groups
= 1, nr_units
= 0;
1359 size_t size_sum
, min_unit_size
, alloc_size
;
1360 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
1361 int last_allocs
, group
, unit
;
1362 unsigned int cpu
, tcpu
;
1363 struct pcpu_alloc_info
*ai
;
1364 unsigned int *cpu_map
;
1366 /* this function may be called multiple times */
1367 memset(group_map
, 0, sizeof(group_map
));
1368 memset(group_cnt
, 0, sizeof(group_map
));
1371 * Determine min_unit_size, alloc_size and max_upa such that
1372 * alloc_size is multiple of atom_size and is the smallest
1373 * which can accomodate 4k aligned segments which are equal to
1374 * or larger than min_unit_size.
1376 size_sum
= pcpu_calc_fc_sizes(static_size
, reserved_size
, &dyn_size
);
1377 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1379 alloc_size
= roundup(min_unit_size
, atom_size
);
1380 upa
= alloc_size
/ min_unit_size
;
1381 while (alloc_size
% upa
|| ((alloc_size
/ upa
) & ~PAGE_MASK
))
1385 /* group cpus according to their proximity */
1386 for_each_possible_cpu(cpu
) {
1389 for_each_possible_cpu(tcpu
) {
1392 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
1393 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
1394 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
1396 nr_groups
= max(nr_groups
, group
+ 1);
1400 group_map
[cpu
] = group
;
1402 group_cnt_max
= max(group_cnt_max
, group_cnt
[group
]);
1406 * Expand unit size until address space usage goes over 75%
1407 * and then as much as possible without using more address
1410 last_allocs
= INT_MAX
;
1411 for (upa
= max_upa
; upa
; upa
--) {
1412 int allocs
= 0, wasted
= 0;
1414 if (alloc_size
% upa
|| ((alloc_size
/ upa
) & ~PAGE_MASK
))
1417 for (group
= 0; group
< nr_groups
; group
++) {
1418 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
1419 allocs
+= this_allocs
;
1420 wasted
+= this_allocs
* upa
- group_cnt
[group
];
1424 * Don't accept if wastage is over 25%. The
1425 * greater-than comparison ensures upa==1 always
1426 * passes the following check.
1428 if (wasted
> num_possible_cpus() / 3)
1431 /* and then don't consume more memory */
1432 if (allocs
> last_allocs
)
1434 last_allocs
= allocs
;
1439 /* allocate and fill alloc_info */
1440 for (group
= 0; group
< nr_groups
; group
++)
1441 nr_units
+= roundup(group_cnt
[group
], upa
);
1443 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
1445 return ERR_PTR(-ENOMEM
);
1446 cpu_map
= ai
->groups
[0].cpu_map
;
1448 for (group
= 0; group
< nr_groups
; group
++) {
1449 ai
->groups
[group
].cpu_map
= cpu_map
;
1450 cpu_map
+= roundup(group_cnt
[group
], upa
);
1453 ai
->static_size
= static_size
;
1454 ai
->reserved_size
= reserved_size
;
1455 ai
->dyn_size
= dyn_size
;
1456 ai
->unit_size
= alloc_size
/ upa
;
1457 ai
->atom_size
= atom_size
;
1458 ai
->alloc_size
= alloc_size
;
1460 for (group
= 0, unit
= 0; group_cnt
[group
]; group
++) {
1461 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1464 * Initialize base_offset as if all groups are located
1465 * back-to-back. The caller should update this to
1466 * reflect actual allocation.
1468 gi
->base_offset
= unit
* ai
->unit_size
;
1470 for_each_possible_cpu(cpu
)
1471 if (group_map
[cpu
] == group
)
1472 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
1473 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
1474 unit
+= gi
->nr_units
;
1476 BUG_ON(unit
!= nr_units
);
1482 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1484 * @ai: allocation info to dump
1486 * Print out information about @ai using loglevel @lvl.
1488 static void pcpu_dump_alloc_info(const char *lvl
,
1489 const struct pcpu_alloc_info
*ai
)
1491 int group_width
= 1, cpu_width
= 1, width
;
1492 char empty_str
[] = "--------";
1493 int alloc
= 0, alloc_end
= 0;
1495 int upa
, apl
; /* units per alloc, allocs per line */
1501 v
= num_possible_cpus();
1504 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
1506 upa
= ai
->alloc_size
/ ai
->unit_size
;
1507 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
1508 apl
= rounddown_pow_of_two(max(60 / width
, 1));
1510 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1511 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
1512 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
1514 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1515 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1516 int unit
= 0, unit_end
= 0;
1518 BUG_ON(gi
->nr_units
% upa
);
1519 for (alloc_end
+= gi
->nr_units
/ upa
;
1520 alloc
< alloc_end
; alloc
++) {
1521 if (!(alloc
% apl
)) {
1523 printk("%spcpu-alloc: ", lvl
);
1525 printk("[%0*d] ", group_width
, group
);
1527 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
1528 if (gi
->cpu_map
[unit
] != NR_CPUS
)
1529 printk("%0*d ", cpu_width
,
1532 printk("%s ", empty_str
);
1539 * pcpu_setup_first_chunk - initialize the first percpu chunk
1540 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1541 * @base_addr: mapped address
1543 * Initialize the first percpu chunk which contains the kernel static
1544 * perpcu area. This function is to be called from arch percpu area
1547 * @ai contains all information necessary to initialize the first
1548 * chunk and prime the dynamic percpu allocator.
1550 * @ai->static_size is the size of static percpu area.
1552 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1553 * reserve after the static area in the first chunk. This reserves
1554 * the first chunk such that it's available only through reserved
1555 * percpu allocation. This is primarily used to serve module percpu
1556 * static areas on architectures where the addressing model has
1557 * limited offset range for symbol relocations to guarantee module
1558 * percpu symbols fall inside the relocatable range.
1560 * @ai->dyn_size determines the number of bytes available for dynamic
1561 * allocation in the first chunk. The area between @ai->static_size +
1562 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1564 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1565 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1568 * @ai->atom_size is the allocation atom size and used as alignment
1571 * @ai->alloc_size is the allocation size and always multiple of
1572 * @ai->atom_size. This is larger than @ai->atom_size if
1573 * @ai->unit_size is larger than @ai->atom_size.
1575 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1576 * percpu areas. Units which should be colocated are put into the
1577 * same group. Dynamic VM areas will be allocated according to these
1578 * groupings. If @ai->nr_groups is zero, a single group containing
1579 * all units is assumed.
1581 * The caller should have mapped the first chunk at @base_addr and
1582 * copied static data to each unit.
1584 * If the first chunk ends up with both reserved and dynamic areas, it
1585 * is served by two chunks - one to serve the core static and reserved
1586 * areas and the other for the dynamic area. They share the same vm
1587 * and page map but uses different area allocation map to stay away
1588 * from each other. The latter chunk is circulated in the chunk slots
1589 * and available for dynamic allocation like any other chunks.
1592 * 0 on success, -errno on failure.
1594 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
1597 static char cpus_buf
[4096] __initdata
;
1598 static int smap
[2], dmap
[2];
1599 size_t dyn_size
= ai
->dyn_size
;
1600 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ dyn_size
;
1601 struct pcpu_chunk
*schunk
, *dchunk
= NULL
;
1602 unsigned long *group_offsets
;
1603 size_t *group_sizes
;
1604 unsigned long *unit_off
;
1609 cpumask_scnprintf(cpus_buf
, sizeof(cpus_buf
), cpu_possible_mask
);
1611 #define PCPU_SETUP_BUG_ON(cond) do { \
1612 if (unlikely(cond)) { \
1613 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1614 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1615 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1621 BUILD_BUG_ON(ARRAY_SIZE(smap
) >= PCPU_DFL_MAP_ALLOC
||
1622 ARRAY_SIZE(dmap
) >= PCPU_DFL_MAP_ALLOC
);
1623 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
1624 PCPU_SETUP_BUG_ON(!ai
->static_size
);
1625 PCPU_SETUP_BUG_ON(!base_addr
);
1626 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
1627 PCPU_SETUP_BUG_ON(ai
->unit_size
& ~PAGE_MASK
);
1628 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
1630 /* process group information and build config tables accordingly */
1631 group_offsets
= alloc_bootmem(ai
->nr_groups
* sizeof(group_offsets
[0]));
1632 group_sizes
= alloc_bootmem(ai
->nr_groups
* sizeof(group_sizes
[0]));
1633 unit_map
= alloc_bootmem(nr_cpu_ids
* sizeof(unit_map
[0]));
1634 unit_off
= alloc_bootmem(nr_cpu_ids
* sizeof(unit_off
[0]));
1636 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
1637 unit_map
[cpu
] = UINT_MAX
;
1638 pcpu_first_unit_cpu
= NR_CPUS
;
1640 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
1641 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1643 group_offsets
[group
] = gi
->base_offset
;
1644 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
1646 for (i
= 0; i
< gi
->nr_units
; i
++) {
1647 cpu
= gi
->cpu_map
[i
];
1651 PCPU_SETUP_BUG_ON(cpu
> nr_cpu_ids
);
1652 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
1653 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
1655 unit_map
[cpu
] = unit
+ i
;
1656 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
1658 if (pcpu_first_unit_cpu
== NR_CPUS
)
1659 pcpu_first_unit_cpu
= cpu
;
1662 pcpu_last_unit_cpu
= cpu
;
1663 pcpu_nr_units
= unit
;
1665 for_each_possible_cpu(cpu
)
1666 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
1668 /* we're done parsing the input, undefine BUG macro and dump config */
1669 #undef PCPU_SETUP_BUG_ON
1670 pcpu_dump_alloc_info(KERN_INFO
, ai
);
1672 pcpu_nr_groups
= ai
->nr_groups
;
1673 pcpu_group_offsets
= group_offsets
;
1674 pcpu_group_sizes
= group_sizes
;
1675 pcpu_unit_map
= unit_map
;
1676 pcpu_unit_offsets
= unit_off
;
1678 /* determine basic parameters */
1679 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1680 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
1681 pcpu_atom_size
= ai
->atom_size
;
1682 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
1683 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
1686 * Allocate chunk slots. The additional last slot is for
1689 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
1690 pcpu_slot
= alloc_bootmem(pcpu_nr_slots
* sizeof(pcpu_slot
[0]));
1691 for (i
= 0; i
< pcpu_nr_slots
; i
++)
1692 INIT_LIST_HEAD(&pcpu_slot
[i
]);
1695 * Initialize static chunk. If reserved_size is zero, the
1696 * static chunk covers static area + dynamic allocation area
1697 * in the first chunk. If reserved_size is not zero, it
1698 * covers static area + reserved area (mostly used for module
1699 * static percpu allocation).
1701 schunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1702 INIT_LIST_HEAD(&schunk
->list
);
1703 schunk
->base_addr
= base_addr
;
1705 schunk
->map_alloc
= ARRAY_SIZE(smap
);
1706 schunk
->immutable
= true;
1707 bitmap_fill(schunk
->populated
, pcpu_unit_pages
);
1709 if (ai
->reserved_size
) {
1710 schunk
->free_size
= ai
->reserved_size
;
1711 pcpu_reserved_chunk
= schunk
;
1712 pcpu_reserved_chunk_limit
= ai
->static_size
+ ai
->reserved_size
;
1714 schunk
->free_size
= dyn_size
;
1715 dyn_size
= 0; /* dynamic area covered */
1717 schunk
->contig_hint
= schunk
->free_size
;
1719 schunk
->map
[schunk
->map_used
++] = -ai
->static_size
;
1720 if (schunk
->free_size
)
1721 schunk
->map
[schunk
->map_used
++] = schunk
->free_size
;
1723 /* init dynamic chunk if necessary */
1725 dchunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1726 INIT_LIST_HEAD(&dchunk
->list
);
1727 dchunk
->base_addr
= base_addr
;
1729 dchunk
->map_alloc
= ARRAY_SIZE(dmap
);
1730 dchunk
->immutable
= true;
1731 bitmap_fill(dchunk
->populated
, pcpu_unit_pages
);
1733 dchunk
->contig_hint
= dchunk
->free_size
= dyn_size
;
1734 dchunk
->map
[dchunk
->map_used
++] = -pcpu_reserved_chunk_limit
;
1735 dchunk
->map
[dchunk
->map_used
++] = dchunk
->free_size
;
1738 /* link the first chunk in */
1739 pcpu_first_chunk
= dchunk
?: schunk
;
1740 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
1743 pcpu_base_addr
= base_addr
;
1747 const char *pcpu_fc_names
[PCPU_FC_NR
] __initdata
= {
1748 [PCPU_FC_AUTO
] = "auto",
1749 [PCPU_FC_EMBED
] = "embed",
1750 [PCPU_FC_PAGE
] = "page",
1753 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
1755 static int __init
percpu_alloc_setup(char *str
)
1759 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1760 else if (!strcmp(str
, "embed"))
1761 pcpu_chosen_fc
= PCPU_FC_EMBED
;
1763 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1764 else if (!strcmp(str
, "page"))
1765 pcpu_chosen_fc
= PCPU_FC_PAGE
;
1768 pr_warning("PERCPU: unknown allocator %s specified\n", str
);
1772 early_param("percpu_alloc", percpu_alloc_setup
);
1774 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1775 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1777 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1778 * @reserved_size: the size of reserved percpu area in bytes
1779 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1780 * @atom_size: allocation atom size
1781 * @cpu_distance_fn: callback to determine distance between cpus, optional
1782 * @alloc_fn: function to allocate percpu page
1783 * @free_fn: funtion to free percpu page
1785 * This is a helper to ease setting up embedded first percpu chunk and
1786 * can be called where pcpu_setup_first_chunk() is expected.
1788 * If this function is used to setup the first chunk, it is allocated
1789 * by calling @alloc_fn and used as-is without being mapped into
1790 * vmalloc area. Allocations are always whole multiples of @atom_size
1791 * aligned to @atom_size.
1793 * This enables the first chunk to piggy back on the linear physical
1794 * mapping which often uses larger page size. Please note that this
1795 * can result in very sparse cpu->unit mapping on NUMA machines thus
1796 * requiring large vmalloc address space. Don't use this allocator if
1797 * vmalloc space is not orders of magnitude larger than distances
1798 * between node memory addresses (ie. 32bit NUMA machines).
1800 * When @dyn_size is positive, dynamic area might be larger than
1801 * specified to fill page alignment. When @dyn_size is auto,
1802 * @dyn_size is just big enough to fill page alignment after static
1803 * and reserved areas.
1805 * If the needed size is smaller than the minimum or specified unit
1806 * size, the leftover is returned using @free_fn.
1809 * 0 on success, -errno on failure.
1811 int __init
pcpu_embed_first_chunk(size_t reserved_size
, ssize_t dyn_size
,
1813 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
1814 pcpu_fc_alloc_fn_t alloc_fn
,
1815 pcpu_fc_free_fn_t free_fn
)
1817 void *base
= (void *)ULONG_MAX
;
1818 void **areas
= NULL
;
1819 struct pcpu_alloc_info
*ai
;
1820 size_t size_sum
, areas_size
, max_distance
;
1823 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
1828 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
1829 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
1831 areas
= alloc_bootmem_nopanic(areas_size
);
1837 /* allocate, copy and determine base address */
1838 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1839 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1840 unsigned int cpu
= NR_CPUS
;
1843 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
1844 cpu
= gi
->cpu_map
[i
];
1845 BUG_ON(cpu
== NR_CPUS
);
1847 /* allocate space for the whole group */
1848 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
1851 goto out_free_areas
;
1855 base
= min(ptr
, base
);
1857 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
1858 if (gi
->cpu_map
[i
] == NR_CPUS
) {
1859 /* unused unit, free whole */
1860 free_fn(ptr
, ai
->unit_size
);
1863 /* copy and return the unused part */
1864 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
1865 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
1869 /* base address is now known, determine group base offsets */
1871 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1872 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
1873 max_distance
= max_t(size_t, max_distance
,
1874 ai
->groups
[group
].base_offset
);
1876 max_distance
+= ai
->unit_size
;
1878 /* warn if maximum distance is further than 75% of vmalloc space */
1879 if (max_distance
> (VMALLOC_END
- VMALLOC_START
) * 3 / 4) {
1880 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1882 max_distance
, VMALLOC_END
- VMALLOC_START
);
1883 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1884 /* and fail if we have fallback */
1890 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1891 PFN_DOWN(size_sum
), base
, ai
->static_size
, ai
->reserved_size
,
1892 ai
->dyn_size
, ai
->unit_size
);
1894 rc
= pcpu_setup_first_chunk(ai
, base
);
1898 for (group
= 0; group
< ai
->nr_groups
; group
++)
1899 free_fn(areas
[group
],
1900 ai
->groups
[group
].nr_units
* ai
->unit_size
);
1902 pcpu_free_alloc_info(ai
);
1904 free_bootmem(__pa(areas
), areas_size
);
1907 #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1908 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
1910 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1912 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1913 * @reserved_size: the size of reserved percpu area in bytes
1914 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1915 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1916 * @populate_pte_fn: function to populate pte
1918 * This is a helper to ease setting up page-remapped first percpu
1919 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1921 * This is the basic allocator. Static percpu area is allocated
1922 * page-by-page into vmalloc area.
1925 * 0 on success, -errno on failure.
1927 int __init
pcpu_page_first_chunk(size_t reserved_size
,
1928 pcpu_fc_alloc_fn_t alloc_fn
,
1929 pcpu_fc_free_fn_t free_fn
,
1930 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
1932 static struct vm_struct vm
;
1933 struct pcpu_alloc_info
*ai
;
1937 struct page
**pages
;
1940 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
1942 ai
= pcpu_build_alloc_info(reserved_size
, -1, PAGE_SIZE
, NULL
);
1945 BUG_ON(ai
->nr_groups
!= 1);
1946 BUG_ON(ai
->groups
[0].nr_units
!= num_possible_cpus());
1948 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1950 /* unaligned allocations can't be freed, round up to page size */
1951 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
1953 pages
= alloc_bootmem(pages_size
);
1955 /* allocate pages */
1957 for (unit
= 0; unit
< num_possible_cpus(); unit
++)
1958 for (i
= 0; i
< unit_pages
; i
++) {
1959 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
1962 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
1964 pr_warning("PERCPU: failed to allocate %s page "
1965 "for cpu%u\n", psize_str
, cpu
);
1968 pages
[j
++] = virt_to_page(ptr
);
1971 /* allocate vm area, map the pages and copy static data */
1972 vm
.flags
= VM_ALLOC
;
1973 vm
.size
= num_possible_cpus() * ai
->unit_size
;
1974 vm_area_register_early(&vm
, PAGE_SIZE
);
1976 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
1977 unsigned long unit_addr
=
1978 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
1980 for (i
= 0; i
< unit_pages
; i
++)
1981 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
1983 /* pte already populated, the following shouldn't fail */
1984 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
1987 panic("failed to map percpu area, err=%d\n", rc
);
1990 * FIXME: Archs with virtual cache should flush local
1991 * cache for the linear mapping here - something
1992 * equivalent to flush_cache_vmap() on the local cpu.
1993 * flush_cache_vmap() can't be used as most supporting
1994 * data structures are not set up yet.
1997 /* copy static data */
1998 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
2001 /* we're ready, commit */
2002 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
2003 unit_pages
, psize_str
, vm
.addr
, ai
->static_size
,
2004 ai
->reserved_size
, ai
->dyn_size
);
2006 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
2011 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
2014 free_bootmem(__pa(pages
), pages_size
);
2015 pcpu_free_alloc_info(ai
);
2018 #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
2021 * Generic percpu area setup.
2023 * The embedding helper is used because its behavior closely resembles
2024 * the original non-dynamic generic percpu area setup. This is
2025 * important because many archs have addressing restrictions and might
2026 * fail if the percpu area is located far away from the previous
2027 * location. As an added bonus, in non-NUMA cases, embedding is
2028 * generally a good idea TLB-wise because percpu area can piggy back
2029 * on the physical linear memory mapping which uses large page
2030 * mappings on applicable archs.
2032 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2033 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
2034 EXPORT_SYMBOL(__per_cpu_offset
);
2036 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
2039 return __alloc_bootmem_nopanic(size
, align
, __pa(MAX_DMA_ADDRESS
));
2042 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
2044 free_bootmem(__pa(ptr
), size
);
2047 void __init
setup_per_cpu_areas(void)
2049 unsigned long delta
;
2054 * Always reserve area for module percpu variables. That's
2055 * what the legacy allocator did.
2057 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
2058 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
2059 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
2061 panic("Failed to initialized percpu areas.");
2063 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
2064 for_each_possible_cpu(cpu
)
2065 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
2067 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */