2 * Copyright (C) 2003 Sistina Software
3 * Copyright (C) 2006 Red Hat GmbH
5 * This file is released under the GPL.
10 #include <linux/device-mapper.h>
12 #include <linux/bio.h>
13 #include <linux/mempool.h>
14 #include <linux/module.h>
15 #include <linux/sched.h>
16 #include <linux/slab.h>
17 #include <linux/dm-io.h>
19 #define DM_MSG_PREFIX "io"
21 #define DM_IO_MAX_REGIONS BITS_PER_LONG
31 * Aligning 'struct io' reduces the number of bits required to store
32 * its address. Refer to store_io_and_region_in_bio() below.
35 unsigned long error_bits
;
37 struct task_struct
*sleeper
;
38 struct dm_io_client
*client
;
39 io_notify_fn callback
;
41 void *vma_invalidate_address
;
42 unsigned long vma_invalidate_size
;
43 } __attribute__((aligned(DM_IO_MAX_REGIONS
)));
45 static struct kmem_cache
*_dm_io_cache
;
48 * Create a client with mempool and bioset.
50 struct dm_io_client
*dm_io_client_create(void)
52 struct dm_io_client
*client
;
54 client
= kmalloc(sizeof(*client
), GFP_KERNEL
);
56 return ERR_PTR(-ENOMEM
);
58 client
->pool
= mempool_create_slab_pool(MIN_IOS
, _dm_io_cache
);
62 client
->bios
= bioset_create(MIN_BIOS
, 0);
70 mempool_destroy(client
->pool
);
72 return ERR_PTR(-ENOMEM
);
74 EXPORT_SYMBOL(dm_io_client_create
);
76 void dm_io_client_destroy(struct dm_io_client
*client
)
78 mempool_destroy(client
->pool
);
79 bioset_free(client
->bios
);
82 EXPORT_SYMBOL(dm_io_client_destroy
);
84 /*-----------------------------------------------------------------
85 * We need to keep track of which region a bio is doing io for.
86 * To avoid a memory allocation to store just 5 or 6 bits, we
87 * ensure the 'struct io' pointer is aligned so enough low bits are
88 * always zero and then combine it with the region number directly in
90 *---------------------------------------------------------------*/
91 static void store_io_and_region_in_bio(struct bio
*bio
, struct io
*io
,
94 if (unlikely(!IS_ALIGNED((unsigned long)io
, DM_IO_MAX_REGIONS
))) {
95 DMCRIT("Unaligned struct io pointer %p", io
);
99 bio
->bi_private
= (void *)((unsigned long)io
| region
);
102 static void retrieve_io_and_region_from_bio(struct bio
*bio
, struct io
**io
,
105 unsigned long val
= (unsigned long)bio
->bi_private
;
107 *io
= (void *)(val
& -(unsigned long)DM_IO_MAX_REGIONS
);
108 *region
= val
& (DM_IO_MAX_REGIONS
- 1);
111 /*-----------------------------------------------------------------
112 * We need an io object to keep track of the number of bios that
113 * have been dispatched for a particular io.
114 *---------------------------------------------------------------*/
115 static void dec_count(struct io
*io
, unsigned int region
, int error
)
118 set_bit(region
, &io
->error_bits
);
120 if (atomic_dec_and_test(&io
->count
)) {
121 if (io
->vma_invalidate_size
)
122 invalidate_kernel_vmap_range(io
->vma_invalidate_address
,
123 io
->vma_invalidate_size
);
126 wake_up_process(io
->sleeper
);
129 unsigned long r
= io
->error_bits
;
130 io_notify_fn fn
= io
->callback
;
131 void *context
= io
->context
;
133 mempool_free(io
, io
->client
->pool
);
139 static void endio(struct bio
*bio
, int error
)
144 if (error
&& bio_data_dir(bio
) == READ
)
148 * The bio destructor in bio_put() may use the io object.
150 retrieve_io_and_region_from_bio(bio
, &io
, ®ion
);
154 dec_count(io
, region
, error
);
157 /*-----------------------------------------------------------------
158 * These little objects provide an abstraction for getting a new
159 * destination page for io.
160 *---------------------------------------------------------------*/
162 void (*get_page
)(struct dpages
*dp
,
163 struct page
**p
, unsigned long *len
, unsigned *offset
);
164 void (*next_page
)(struct dpages
*dp
);
169 void *vma_invalidate_address
;
170 unsigned long vma_invalidate_size
;
174 * Functions for getting the pages from a list.
176 static void list_get_page(struct dpages
*dp
,
177 struct page
**p
, unsigned long *len
, unsigned *offset
)
179 unsigned o
= dp
->context_u
;
180 struct page_list
*pl
= (struct page_list
*) dp
->context_ptr
;
183 *len
= PAGE_SIZE
- o
;
187 static void list_next_page(struct dpages
*dp
)
189 struct page_list
*pl
= (struct page_list
*) dp
->context_ptr
;
190 dp
->context_ptr
= pl
->next
;
194 static void list_dp_init(struct dpages
*dp
, struct page_list
*pl
, unsigned offset
)
196 dp
->get_page
= list_get_page
;
197 dp
->next_page
= list_next_page
;
198 dp
->context_u
= offset
;
199 dp
->context_ptr
= pl
;
203 * Functions for getting the pages from a bvec.
205 static void bvec_get_page(struct dpages
*dp
,
206 struct page
**p
, unsigned long *len
, unsigned *offset
)
208 struct bio_vec
*bvec
= (struct bio_vec
*) dp
->context_ptr
;
211 *offset
= bvec
->bv_offset
;
214 static void bvec_next_page(struct dpages
*dp
)
216 struct bio_vec
*bvec
= (struct bio_vec
*) dp
->context_ptr
;
217 dp
->context_ptr
= bvec
+ 1;
220 static void bvec_dp_init(struct dpages
*dp
, struct bio_vec
*bvec
)
222 dp
->get_page
= bvec_get_page
;
223 dp
->next_page
= bvec_next_page
;
224 dp
->context_ptr
= bvec
;
228 * Functions for getting the pages from a VMA.
230 static void vm_get_page(struct dpages
*dp
,
231 struct page
**p
, unsigned long *len
, unsigned *offset
)
233 *p
= vmalloc_to_page(dp
->context_ptr
);
234 *offset
= dp
->context_u
;
235 *len
= PAGE_SIZE
- dp
->context_u
;
238 static void vm_next_page(struct dpages
*dp
)
240 dp
->context_ptr
+= PAGE_SIZE
- dp
->context_u
;
244 static void vm_dp_init(struct dpages
*dp
, void *data
)
246 dp
->get_page
= vm_get_page
;
247 dp
->next_page
= vm_next_page
;
248 dp
->context_u
= ((unsigned long) data
) & (PAGE_SIZE
- 1);
249 dp
->context_ptr
= data
;
253 * Functions for getting the pages from kernel memory.
255 static void km_get_page(struct dpages
*dp
, struct page
**p
, unsigned long *len
,
258 *p
= virt_to_page(dp
->context_ptr
);
259 *offset
= dp
->context_u
;
260 *len
= PAGE_SIZE
- dp
->context_u
;
263 static void km_next_page(struct dpages
*dp
)
265 dp
->context_ptr
+= PAGE_SIZE
- dp
->context_u
;
269 static void km_dp_init(struct dpages
*dp
, void *data
)
271 dp
->get_page
= km_get_page
;
272 dp
->next_page
= km_next_page
;
273 dp
->context_u
= ((unsigned long) data
) & (PAGE_SIZE
- 1);
274 dp
->context_ptr
= data
;
277 /*-----------------------------------------------------------------
278 * IO routines that accept a list of pages.
279 *---------------------------------------------------------------*/
280 static void do_region(int rw
, unsigned region
, struct dm_io_region
*where
,
281 struct dpages
*dp
, struct io
*io
)
288 sector_t remaining
= where
->count
;
289 struct request_queue
*q
= bdev_get_queue(where
->bdev
);
290 unsigned short logical_block_size
= queue_logical_block_size(q
);
291 sector_t num_sectors
;
294 * where->count may be zero if rw holds a flush and we need to
295 * send a zero-sized flush.
299 * Allocate a suitably sized-bio.
301 if ((rw
& REQ_DISCARD
) || (rw
& REQ_WRITE_SAME
))
304 num_bvecs
= min_t(int, bio_get_nr_vecs(where
->bdev
),
305 dm_sector_div_up(remaining
, (PAGE_SIZE
>> SECTOR_SHIFT
)));
307 bio
= bio_alloc_bioset(GFP_NOIO
, num_bvecs
, io
->client
->bios
);
308 bio
->bi_sector
= where
->sector
+ (where
->count
- remaining
);
309 bio
->bi_bdev
= where
->bdev
;
310 bio
->bi_end_io
= endio
;
311 store_io_and_region_in_bio(bio
, io
, region
);
313 if (rw
& REQ_DISCARD
) {
314 num_sectors
= min_t(sector_t
, q
->limits
.max_discard_sectors
, remaining
);
315 bio
->bi_size
= num_sectors
<< SECTOR_SHIFT
;
316 remaining
-= num_sectors
;
317 } else if (rw
& REQ_WRITE_SAME
) {
319 * WRITE SAME only uses a single page.
321 dp
->get_page(dp
, &page
, &len
, &offset
);
322 bio_add_page(bio
, page
, logical_block_size
, offset
);
323 num_sectors
= min_t(sector_t
, q
->limits
.max_write_same_sectors
, remaining
);
324 bio
->bi_size
= num_sectors
<< SECTOR_SHIFT
;
327 remaining
-= num_sectors
;
329 } else while (remaining
) {
331 * Try and add as many pages as possible.
333 dp
->get_page(dp
, &page
, &len
, &offset
);
334 len
= min(len
, to_bytes(remaining
));
335 if (!bio_add_page(bio
, page
, len
, offset
))
339 remaining
-= to_sector(len
);
343 atomic_inc(&io
->count
);
348 static void dispatch_io(int rw
, unsigned int num_regions
,
349 struct dm_io_region
*where
, struct dpages
*dp
,
350 struct io
*io
, int sync
)
353 struct dpages old_pages
= *dp
;
355 BUG_ON(num_regions
> DM_IO_MAX_REGIONS
);
361 * For multiple regions we need to be careful to rewind
362 * the dp object for each call to do_region.
364 for (i
= 0; i
< num_regions
; i
++) {
366 if (where
[i
].count
|| (rw
& REQ_FLUSH
))
367 do_region(rw
, i
, where
+ i
, dp
, io
);
371 * Drop the extra reference that we were holding to avoid
372 * the io being completed too early.
377 static int sync_io(struct dm_io_client
*client
, unsigned int num_regions
,
378 struct dm_io_region
*where
, int rw
, struct dpages
*dp
,
379 unsigned long *error_bits
)
382 * gcc <= 4.3 can't do the alignment for stack variables, so we must
383 * align it on our own.
384 * volatile prevents the optimizer from removing or reusing
385 * "io_" field from the stack frame (allowed in ANSI C).
387 volatile char io_
[sizeof(struct io
) + __alignof__(struct io
) - 1];
388 struct io
*io
= (struct io
*)PTR_ALIGN(&io_
, __alignof__(struct io
));
390 if (num_regions
> 1 && (rw
& RW_MASK
) != WRITE
) {
396 atomic_set(&io
->count
, 1); /* see dispatch_io() */
397 io
->sleeper
= current
;
400 io
->vma_invalidate_address
= dp
->vma_invalidate_address
;
401 io
->vma_invalidate_size
= dp
->vma_invalidate_size
;
403 dispatch_io(rw
, num_regions
, where
, dp
, io
, 1);
406 set_current_state(TASK_UNINTERRUPTIBLE
);
408 if (!atomic_read(&io
->count
))
413 set_current_state(TASK_RUNNING
);
416 *error_bits
= io
->error_bits
;
418 return io
->error_bits
? -EIO
: 0;
421 static int async_io(struct dm_io_client
*client
, unsigned int num_regions
,
422 struct dm_io_region
*where
, int rw
, struct dpages
*dp
,
423 io_notify_fn fn
, void *context
)
427 if (num_regions
> 1 && (rw
& RW_MASK
) != WRITE
) {
433 io
= mempool_alloc(client
->pool
, GFP_NOIO
);
435 atomic_set(&io
->count
, 1); /* see dispatch_io() */
439 io
->context
= context
;
441 io
->vma_invalidate_address
= dp
->vma_invalidate_address
;
442 io
->vma_invalidate_size
= dp
->vma_invalidate_size
;
444 dispatch_io(rw
, num_regions
, where
, dp
, io
, 0);
448 static int dp_init(struct dm_io_request
*io_req
, struct dpages
*dp
,
451 /* Set up dpages based on memory type */
453 dp
->vma_invalidate_address
= NULL
;
454 dp
->vma_invalidate_size
= 0;
456 switch (io_req
->mem
.type
) {
457 case DM_IO_PAGE_LIST
:
458 list_dp_init(dp
, io_req
->mem
.ptr
.pl
, io_req
->mem
.offset
);
462 bvec_dp_init(dp
, io_req
->mem
.ptr
.bvec
);
466 flush_kernel_vmap_range(io_req
->mem
.ptr
.vma
, size
);
467 if ((io_req
->bi_rw
& RW_MASK
) == READ
) {
468 dp
->vma_invalidate_address
= io_req
->mem
.ptr
.vma
;
469 dp
->vma_invalidate_size
= size
;
471 vm_dp_init(dp
, io_req
->mem
.ptr
.vma
);
475 km_dp_init(dp
, io_req
->mem
.ptr
.addr
);
486 * New collapsed (a)synchronous interface.
488 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
489 * the queue with blk_unplug() some time later or set REQ_SYNC in
490 io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
491 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
493 int dm_io(struct dm_io_request
*io_req
, unsigned num_regions
,
494 struct dm_io_region
*where
, unsigned long *sync_error_bits
)
499 r
= dp_init(io_req
, &dp
, (unsigned long)where
->count
<< SECTOR_SHIFT
);
503 if (!io_req
->notify
.fn
)
504 return sync_io(io_req
->client
, num_regions
, where
,
505 io_req
->bi_rw
, &dp
, sync_error_bits
);
507 return async_io(io_req
->client
, num_regions
, where
, io_req
->bi_rw
,
508 &dp
, io_req
->notify
.fn
, io_req
->notify
.context
);
510 EXPORT_SYMBOL(dm_io
);
512 int __init
dm_io_init(void)
514 _dm_io_cache
= KMEM_CACHE(io
, 0);
521 void dm_io_exit(void)
523 kmem_cache_destroy(_dm_io_cache
);