Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Copyright (C) 2003 Sistina Software | |
891ce207 | 3 | * Copyright (C) 2006 Red Hat GmbH |
1da177e4 LT |
4 | * |
5 | * This file is released under the GPL. | |
6 | */ | |
7 | ||
586e80e6 | 8 | #include <linux/device-mapper.h> |
1da177e4 LT |
9 | |
10 | #include <linux/bio.h> | |
11 | #include <linux/mempool.h> | |
12 | #include <linux/module.h> | |
13 | #include <linux/sched.h> | |
14 | #include <linux/slab.h> | |
a765e20e | 15 | #include <linux/dm-io.h> |
1da177e4 | 16 | |
891ce207 HM |
17 | struct dm_io_client { |
18 | mempool_t *pool; | |
19 | struct bio_set *bios; | |
20 | }; | |
21 | ||
1da177e4 LT |
22 | /* FIXME: can we shrink this ? */ |
23 | struct io { | |
e01fd7ee | 24 | unsigned long error_bits; |
1da177e4 LT |
25 | atomic_t count; |
26 | struct task_struct *sleeper; | |
891ce207 | 27 | struct dm_io_client *client; |
1da177e4 LT |
28 | io_notify_fn callback; |
29 | void *context; | |
30 | }; | |
31 | ||
32 | /* | |
33 | * io contexts are only dynamically allocated for asynchronous | |
34 | * io. Since async io is likely to be the majority of io we'll | |
891ce207 | 35 | * have the same number of io contexts as bios! (FIXME: must reduce this). |
1da177e4 | 36 | */ |
891ce207 | 37 | |
1da177e4 LT |
38 | static unsigned int pages_to_ios(unsigned int pages) |
39 | { | |
40 | return 4 * pages; /* too many ? */ | |
41 | } | |
42 | ||
c8b03afe HM |
43 | /* |
44 | * Create a client with mempool and bioset. | |
45 | */ | |
46 | struct dm_io_client *dm_io_client_create(unsigned num_pages) | |
47 | { | |
48 | unsigned ios = pages_to_ios(num_pages); | |
49 | struct dm_io_client *client; | |
50 | ||
51 | client = kmalloc(sizeof(*client), GFP_KERNEL); | |
52 | if (!client) | |
53 | return ERR_PTR(-ENOMEM); | |
54 | ||
55 | client->pool = mempool_create_kmalloc_pool(ios, sizeof(struct io)); | |
56 | if (!client->pool) | |
57 | goto bad; | |
58 | ||
bb799ca0 | 59 | client->bios = bioset_create(16, 0); |
c8b03afe HM |
60 | if (!client->bios) |
61 | goto bad; | |
62 | ||
63 | return client; | |
64 | ||
65 | bad: | |
66 | if (client->pool) | |
67 | mempool_destroy(client->pool); | |
68 | kfree(client); | |
69 | return ERR_PTR(-ENOMEM); | |
70 | } | |
71 | EXPORT_SYMBOL(dm_io_client_create); | |
72 | ||
73 | int dm_io_client_resize(unsigned num_pages, struct dm_io_client *client) | |
74 | { | |
75 | return mempool_resize(client->pool, pages_to_ios(num_pages), | |
76 | GFP_KERNEL); | |
77 | } | |
78 | EXPORT_SYMBOL(dm_io_client_resize); | |
79 | ||
80 | void dm_io_client_destroy(struct dm_io_client *client) | |
81 | { | |
82 | mempool_destroy(client->pool); | |
83 | bioset_free(client->bios); | |
84 | kfree(client); | |
85 | } | |
86 | EXPORT_SYMBOL(dm_io_client_destroy); | |
87 | ||
1da177e4 LT |
88 | /*----------------------------------------------------------------- |
89 | * We need to keep track of which region a bio is doing io for. | |
90 | * In order to save a memory allocation we store this the last | |
91 | * bvec which we know is unused (blech). | |
92 | * XXX This is ugly and can OOPS with some configs... find another way. | |
93 | *---------------------------------------------------------------*/ | |
94 | static inline void bio_set_region(struct bio *bio, unsigned region) | |
95 | { | |
f00b16ad | 96 | bio->bi_io_vec[bio->bi_max_vecs].bv_len = region; |
1da177e4 LT |
97 | } |
98 | ||
99 | static inline unsigned bio_get_region(struct bio *bio) | |
100 | { | |
f00b16ad | 101 | return bio->bi_io_vec[bio->bi_max_vecs].bv_len; |
1da177e4 LT |
102 | } |
103 | ||
104 | /*----------------------------------------------------------------- | |
105 | * We need an io object to keep track of the number of bios that | |
106 | * have been dispatched for a particular io. | |
107 | *---------------------------------------------------------------*/ | |
108 | static void dec_count(struct io *io, unsigned int region, int error) | |
109 | { | |
110 | if (error) | |
e01fd7ee | 111 | set_bit(region, &io->error_bits); |
1da177e4 LT |
112 | |
113 | if (atomic_dec_and_test(&io->count)) { | |
114 | if (io->sleeper) | |
115 | wake_up_process(io->sleeper); | |
116 | ||
117 | else { | |
e01fd7ee | 118 | unsigned long r = io->error_bits; |
1da177e4 LT |
119 | io_notify_fn fn = io->callback; |
120 | void *context = io->context; | |
121 | ||
bf17ce3a | 122 | mempool_free(io, io->client->pool); |
1da177e4 LT |
123 | fn(r, context); |
124 | } | |
125 | } | |
126 | } | |
127 | ||
6712ecf8 | 128 | static void endio(struct bio *bio, int error) |
1da177e4 | 129 | { |
c897feb3 HM |
130 | struct io *io; |
131 | unsigned region; | |
1da177e4 | 132 | |
1da177e4 LT |
133 | if (error && bio_data_dir(bio) == READ) |
134 | zero_fill_bio(bio); | |
135 | ||
c897feb3 HM |
136 | /* |
137 | * The bio destructor in bio_put() may use the io object. | |
138 | */ | |
139 | io = bio->bi_private; | |
140 | region = bio_get_region(bio); | |
141 | ||
f00b16ad | 142 | bio->bi_max_vecs++; |
1da177e4 LT |
143 | bio_put(bio); |
144 | ||
c897feb3 | 145 | dec_count(io, region, error); |
1da177e4 LT |
146 | } |
147 | ||
148 | /*----------------------------------------------------------------- | |
149 | * These little objects provide an abstraction for getting a new | |
150 | * destination page for io. | |
151 | *---------------------------------------------------------------*/ | |
152 | struct dpages { | |
153 | void (*get_page)(struct dpages *dp, | |
154 | struct page **p, unsigned long *len, unsigned *offset); | |
155 | void (*next_page)(struct dpages *dp); | |
156 | ||
157 | unsigned context_u; | |
158 | void *context_ptr; | |
159 | }; | |
160 | ||
161 | /* | |
162 | * Functions for getting the pages from a list. | |
163 | */ | |
164 | static void list_get_page(struct dpages *dp, | |
165 | struct page **p, unsigned long *len, unsigned *offset) | |
166 | { | |
167 | unsigned o = dp->context_u; | |
168 | struct page_list *pl = (struct page_list *) dp->context_ptr; | |
169 | ||
170 | *p = pl->page; | |
171 | *len = PAGE_SIZE - o; | |
172 | *offset = o; | |
173 | } | |
174 | ||
175 | static void list_next_page(struct dpages *dp) | |
176 | { | |
177 | struct page_list *pl = (struct page_list *) dp->context_ptr; | |
178 | dp->context_ptr = pl->next; | |
179 | dp->context_u = 0; | |
180 | } | |
181 | ||
182 | static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset) | |
183 | { | |
184 | dp->get_page = list_get_page; | |
185 | dp->next_page = list_next_page; | |
186 | dp->context_u = offset; | |
187 | dp->context_ptr = pl; | |
188 | } | |
189 | ||
190 | /* | |
191 | * Functions for getting the pages from a bvec. | |
192 | */ | |
193 | static void bvec_get_page(struct dpages *dp, | |
194 | struct page **p, unsigned long *len, unsigned *offset) | |
195 | { | |
196 | struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr; | |
197 | *p = bvec->bv_page; | |
198 | *len = bvec->bv_len; | |
199 | *offset = bvec->bv_offset; | |
200 | } | |
201 | ||
202 | static void bvec_next_page(struct dpages *dp) | |
203 | { | |
204 | struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr; | |
205 | dp->context_ptr = bvec + 1; | |
206 | } | |
207 | ||
208 | static void bvec_dp_init(struct dpages *dp, struct bio_vec *bvec) | |
209 | { | |
210 | dp->get_page = bvec_get_page; | |
211 | dp->next_page = bvec_next_page; | |
212 | dp->context_ptr = bvec; | |
213 | } | |
214 | ||
c8b03afe HM |
215 | /* |
216 | * Functions for getting the pages from a VMA. | |
217 | */ | |
1da177e4 LT |
218 | static void vm_get_page(struct dpages *dp, |
219 | struct page **p, unsigned long *len, unsigned *offset) | |
220 | { | |
221 | *p = vmalloc_to_page(dp->context_ptr); | |
222 | *offset = dp->context_u; | |
223 | *len = PAGE_SIZE - dp->context_u; | |
224 | } | |
225 | ||
226 | static void vm_next_page(struct dpages *dp) | |
227 | { | |
228 | dp->context_ptr += PAGE_SIZE - dp->context_u; | |
229 | dp->context_u = 0; | |
230 | } | |
231 | ||
232 | static void vm_dp_init(struct dpages *dp, void *data) | |
233 | { | |
234 | dp->get_page = vm_get_page; | |
235 | dp->next_page = vm_next_page; | |
236 | dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1); | |
237 | dp->context_ptr = data; | |
238 | } | |
239 | ||
3676347a PO |
240 | static void dm_bio_destructor(struct bio *bio) |
241 | { | |
891ce207 HM |
242 | struct io *io = bio->bi_private; |
243 | ||
bf17ce3a | 244 | bio_free(bio, io->client->bios); |
3676347a PO |
245 | } |
246 | ||
c8b03afe HM |
247 | /* |
248 | * Functions for getting the pages from kernel memory. | |
249 | */ | |
250 | static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len, | |
251 | unsigned *offset) | |
252 | { | |
253 | *p = virt_to_page(dp->context_ptr); | |
254 | *offset = dp->context_u; | |
255 | *len = PAGE_SIZE - dp->context_u; | |
256 | } | |
257 | ||
258 | static void km_next_page(struct dpages *dp) | |
259 | { | |
260 | dp->context_ptr += PAGE_SIZE - dp->context_u; | |
261 | dp->context_u = 0; | |
262 | } | |
263 | ||
264 | static void km_dp_init(struct dpages *dp, void *data) | |
265 | { | |
266 | dp->get_page = km_get_page; | |
267 | dp->next_page = km_next_page; | |
268 | dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1); | |
269 | dp->context_ptr = data; | |
270 | } | |
271 | ||
1da177e4 LT |
272 | /*----------------------------------------------------------------- |
273 | * IO routines that accept a list of pages. | |
274 | *---------------------------------------------------------------*/ | |
22a1ceb1 | 275 | static void do_region(int rw, unsigned region, struct dm_io_region *where, |
1da177e4 LT |
276 | struct dpages *dp, struct io *io) |
277 | { | |
278 | struct bio *bio; | |
279 | struct page *page; | |
280 | unsigned long len; | |
281 | unsigned offset; | |
282 | unsigned num_bvecs; | |
283 | sector_t remaining = where->count; | |
284 | ||
285 | while (remaining) { | |
286 | /* | |
f00b16ad HM |
287 | * Allocate a suitably sized-bio: we add an extra |
288 | * bvec for bio_get/set_region() and decrement bi_max_vecs | |
289 | * to hide it from bio_add_page(). | |
1da177e4 | 290 | */ |
596f138e JN |
291 | num_bvecs = dm_sector_div_up(remaining, |
292 | (PAGE_SIZE >> SECTOR_SHIFT)); | |
293 | num_bvecs = 1 + min_t(int, bio_get_nr_vecs(where->bdev), | |
294 | num_bvecs); | |
bf17ce3a | 295 | bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios); |
1da177e4 LT |
296 | bio->bi_sector = where->sector + (where->count - remaining); |
297 | bio->bi_bdev = where->bdev; | |
298 | bio->bi_end_io = endio; | |
299 | bio->bi_private = io; | |
3676347a | 300 | bio->bi_destructor = dm_bio_destructor; |
f00b16ad | 301 | bio->bi_max_vecs--; |
1da177e4 LT |
302 | bio_set_region(bio, region); |
303 | ||
304 | /* | |
305 | * Try and add as many pages as possible. | |
306 | */ | |
307 | while (remaining) { | |
308 | dp->get_page(dp, &page, &len, &offset); | |
309 | len = min(len, to_bytes(remaining)); | |
310 | if (!bio_add_page(bio, page, len, offset)) | |
311 | break; | |
312 | ||
313 | offset = 0; | |
314 | remaining -= to_sector(len); | |
315 | dp->next_page(dp); | |
316 | } | |
317 | ||
318 | atomic_inc(&io->count); | |
319 | submit_bio(rw, bio); | |
320 | } | |
321 | } | |
322 | ||
323 | static void dispatch_io(int rw, unsigned int num_regions, | |
22a1ceb1 | 324 | struct dm_io_region *where, struct dpages *dp, |
1da177e4 LT |
325 | struct io *io, int sync) |
326 | { | |
327 | int i; | |
328 | struct dpages old_pages = *dp; | |
329 | ||
330 | if (sync) | |
331 | rw |= (1 << BIO_RW_SYNC); | |
332 | ||
333 | /* | |
334 | * For multiple regions we need to be careful to rewind | |
335 | * the dp object for each call to do_region. | |
336 | */ | |
337 | for (i = 0; i < num_regions; i++) { | |
338 | *dp = old_pages; | |
339 | if (where[i].count) | |
340 | do_region(rw, i, where + i, dp, io); | |
341 | } | |
342 | ||
343 | /* | |
f00b16ad | 344 | * Drop the extra reference that we were holding to avoid |
1da177e4 LT |
345 | * the io being completed too early. |
346 | */ | |
347 | dec_count(io, 0, 0); | |
348 | } | |
349 | ||
891ce207 | 350 | static int sync_io(struct dm_io_client *client, unsigned int num_regions, |
22a1ceb1 | 351 | struct dm_io_region *where, int rw, struct dpages *dp, |
891ce207 | 352 | unsigned long *error_bits) |
1da177e4 LT |
353 | { |
354 | struct io io; | |
355 | ||
7ff14a36 | 356 | if (num_regions > 1 && (rw & RW_MASK) != WRITE) { |
1da177e4 LT |
357 | WARN_ON(1); |
358 | return -EIO; | |
359 | } | |
360 | ||
e01fd7ee | 361 | io.error_bits = 0; |
1da177e4 LT |
362 | atomic_set(&io.count, 1); /* see dispatch_io() */ |
363 | io.sleeper = current; | |
891ce207 | 364 | io.client = client; |
1da177e4 LT |
365 | |
366 | dispatch_io(rw, num_regions, where, dp, &io, 1); | |
367 | ||
368 | while (1) { | |
369 | set_current_state(TASK_UNINTERRUPTIBLE); | |
370 | ||
371 | if (!atomic_read(&io.count) || signal_pending(current)) | |
372 | break; | |
373 | ||
374 | io_schedule(); | |
375 | } | |
376 | set_current_state(TASK_RUNNING); | |
377 | ||
378 | if (atomic_read(&io.count)) | |
379 | return -EINTR; | |
380 | ||
891ce207 | 381 | if (error_bits) |
e01fd7ee | 382 | *error_bits = io.error_bits; |
891ce207 | 383 | |
e01fd7ee | 384 | return io.error_bits ? -EIO : 0; |
1da177e4 LT |
385 | } |
386 | ||
891ce207 | 387 | static int async_io(struct dm_io_client *client, unsigned int num_regions, |
22a1ceb1 | 388 | struct dm_io_region *where, int rw, struct dpages *dp, |
891ce207 | 389 | io_notify_fn fn, void *context) |
1da177e4 LT |
390 | { |
391 | struct io *io; | |
392 | ||
7ff14a36 | 393 | if (num_regions > 1 && (rw & RW_MASK) != WRITE) { |
1da177e4 LT |
394 | WARN_ON(1); |
395 | fn(1, context); | |
396 | return -EIO; | |
397 | } | |
398 | ||
bf17ce3a | 399 | io = mempool_alloc(client->pool, GFP_NOIO); |
e01fd7ee | 400 | io->error_bits = 0; |
1da177e4 LT |
401 | atomic_set(&io->count, 1); /* see dispatch_io() */ |
402 | io->sleeper = NULL; | |
891ce207 | 403 | io->client = client; |
1da177e4 LT |
404 | io->callback = fn; |
405 | io->context = context; | |
406 | ||
407 | dispatch_io(rw, num_regions, where, dp, io, 0); | |
408 | return 0; | |
409 | } | |
410 | ||
c8b03afe HM |
411 | static int dp_init(struct dm_io_request *io_req, struct dpages *dp) |
412 | { | |
413 | /* Set up dpages based on memory type */ | |
414 | switch (io_req->mem.type) { | |
415 | case DM_IO_PAGE_LIST: | |
416 | list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset); | |
417 | break; | |
418 | ||
419 | case DM_IO_BVEC: | |
420 | bvec_dp_init(dp, io_req->mem.ptr.bvec); | |
421 | break; | |
422 | ||
423 | case DM_IO_VMA: | |
424 | vm_dp_init(dp, io_req->mem.ptr.vma); | |
425 | break; | |
426 | ||
427 | case DM_IO_KMEM: | |
428 | km_dp_init(dp, io_req->mem.ptr.addr); | |
429 | break; | |
430 | ||
431 | default: | |
432 | return -EINVAL; | |
433 | } | |
434 | ||
435 | return 0; | |
436 | } | |
437 | ||
438 | /* | |
7ff14a36 MP |
439 | * New collapsed (a)synchronous interface. |
440 | * | |
441 | * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug | |
442 | * the queue with blk_unplug() some time later or set the BIO_RW_SYNC bit in | |
443 | * io_req->bi_rw. If you fail to do one of these, the IO will be submitted to | |
444 | * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c. | |
c8b03afe HM |
445 | */ |
446 | int dm_io(struct dm_io_request *io_req, unsigned num_regions, | |
22a1ceb1 | 447 | struct dm_io_region *where, unsigned long *sync_error_bits) |
c8b03afe HM |
448 | { |
449 | int r; | |
450 | struct dpages dp; | |
451 | ||
452 | r = dp_init(io_req, &dp); | |
453 | if (r) | |
454 | return r; | |
455 | ||
456 | if (!io_req->notify.fn) | |
457 | return sync_io(io_req->client, num_regions, where, | |
458 | io_req->bi_rw, &dp, sync_error_bits); | |
459 | ||
460 | return async_io(io_req->client, num_regions, where, io_req->bi_rw, | |
461 | &dp, io_req->notify.fn, io_req->notify.context); | |
462 | } | |
463 | EXPORT_SYMBOL(dm_io); |