Commit | Line | Data |
---|---|---|
86db1e29 JA |
1 | /* |
2 | * Functions related to barrier IO handling | |
3 | */ | |
4 | #include <linux/kernel.h> | |
5 | #include <linux/module.h> | |
6 | #include <linux/bio.h> | |
7 | #include <linux/blkdev.h> | |
8 | ||
9 | #include "blk.h" | |
10 | ||
11 | /** | |
12 | * blk_queue_ordered - does this queue support ordered writes | |
13 | * @q: the request queue | |
14 | * @ordered: one of QUEUE_ORDERED_* | |
15 | * @prepare_flush_fn: rq setup helper for cache flush ordered writes | |
16 | * | |
17 | * Description: | |
18 | * For journalled file systems, doing ordered writes on a commit | |
19 | * block instead of explicitly doing wait_on_buffer (which is bad | |
20 | * for performance) can be a big win. Block drivers supporting this | |
21 | * feature should call this function and indicate so. | |
22 | * | |
23 | **/ | |
24 | int blk_queue_ordered(struct request_queue *q, unsigned ordered, | |
25 | prepare_flush_fn *prepare_flush_fn) | |
26 | { | |
313e4299 TH |
27 | if (!prepare_flush_fn && (ordered & (QUEUE_ORDERED_DO_PREFLUSH | |
28 | QUEUE_ORDERED_DO_POSTFLUSH))) { | |
24c03d47 | 29 | printk(KERN_ERR "%s: prepare_flush_fn required\n", __func__); |
86db1e29 JA |
30 | return -EINVAL; |
31 | } | |
32 | ||
33 | if (ordered != QUEUE_ORDERED_NONE && | |
34 | ordered != QUEUE_ORDERED_DRAIN && | |
35 | ordered != QUEUE_ORDERED_DRAIN_FLUSH && | |
36 | ordered != QUEUE_ORDERED_DRAIN_FUA && | |
37 | ordered != QUEUE_ORDERED_TAG && | |
38 | ordered != QUEUE_ORDERED_TAG_FLUSH && | |
39 | ordered != QUEUE_ORDERED_TAG_FUA) { | |
40 | printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered); | |
41 | return -EINVAL; | |
42 | } | |
43 | ||
44 | q->ordered = ordered; | |
45 | q->next_ordered = ordered; | |
46 | q->prepare_flush_fn = prepare_flush_fn; | |
47 | ||
48 | return 0; | |
49 | } | |
86db1e29 JA |
50 | EXPORT_SYMBOL(blk_queue_ordered); |
51 | ||
52 | /* | |
53 | * Cache flushing for ordered writes handling | |
54 | */ | |
6f6a036e | 55 | unsigned blk_ordered_cur_seq(struct request_queue *q) |
86db1e29 JA |
56 | { |
57 | if (!q->ordseq) | |
58 | return 0; | |
59 | return 1 << ffz(q->ordseq); | |
60 | } | |
61 | ||
62 | unsigned blk_ordered_req_seq(struct request *rq) | |
63 | { | |
64 | struct request_queue *q = rq->q; | |
65 | ||
66 | BUG_ON(q->ordseq == 0); | |
67 | ||
68 | if (rq == &q->pre_flush_rq) | |
69 | return QUEUE_ORDSEQ_PREFLUSH; | |
70 | if (rq == &q->bar_rq) | |
71 | return QUEUE_ORDSEQ_BAR; | |
72 | if (rq == &q->post_flush_rq) | |
73 | return QUEUE_ORDSEQ_POSTFLUSH; | |
74 | ||
75 | /* | |
76 | * !fs requests don't need to follow barrier ordering. Always | |
77 | * put them at the front. This fixes the following deadlock. | |
78 | * | |
79 | * http://thread.gmane.org/gmane.linux.kernel/537473 | |
80 | */ | |
81 | if (!blk_fs_request(rq)) | |
82 | return QUEUE_ORDSEQ_DRAIN; | |
83 | ||
84 | if ((rq->cmd_flags & REQ_ORDERED_COLOR) == | |
85 | (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR)) | |
86 | return QUEUE_ORDSEQ_DRAIN; | |
87 | else | |
88 | return QUEUE_ORDSEQ_DONE; | |
89 | } | |
90 | ||
8f11b3e9 | 91 | bool blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error) |
86db1e29 JA |
92 | { |
93 | struct request *rq; | |
94 | ||
95 | if (error && !q->orderr) | |
96 | q->orderr = error; | |
97 | ||
98 | BUG_ON(q->ordseq & seq); | |
99 | q->ordseq |= seq; | |
100 | ||
101 | if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE) | |
8f11b3e9 | 102 | return false; |
86db1e29 JA |
103 | |
104 | /* | |
105 | * Okay, sequence complete. | |
106 | */ | |
107 | q->ordseq = 0; | |
108 | rq = q->orig_bar_rq; | |
40cbbb78 | 109 | __blk_end_request_all(rq, q->orderr); |
8f11b3e9 | 110 | return true; |
86db1e29 JA |
111 | } |
112 | ||
113 | static void pre_flush_end_io(struct request *rq, int error) | |
114 | { | |
115 | elv_completed_request(rq->q, rq); | |
116 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error); | |
117 | } | |
118 | ||
119 | static void bar_end_io(struct request *rq, int error) | |
120 | { | |
121 | elv_completed_request(rq->q, rq); | |
122 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error); | |
123 | } | |
124 | ||
125 | static void post_flush_end_io(struct request *rq, int error) | |
126 | { | |
127 | elv_completed_request(rq->q, rq); | |
128 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error); | |
129 | } | |
130 | ||
131 | static void queue_flush(struct request_queue *q, unsigned which) | |
132 | { | |
133 | struct request *rq; | |
134 | rq_end_io_fn *end_io; | |
135 | ||
313e4299 | 136 | if (which == QUEUE_ORDERED_DO_PREFLUSH) { |
86db1e29 JA |
137 | rq = &q->pre_flush_rq; |
138 | end_io = pre_flush_end_io; | |
139 | } else { | |
140 | rq = &q->post_flush_rq; | |
141 | end_io = post_flush_end_io; | |
142 | } | |
143 | ||
2a4aa30c | 144 | blk_rq_init(q, rq); |
1afb20f3 | 145 | rq->cmd_flags = REQ_HARDBARRIER; |
86db1e29 JA |
146 | rq->rq_disk = q->bar_rq.rq_disk; |
147 | rq->end_io = end_io; | |
148 | q->prepare_flush_fn(q, rq); | |
149 | ||
150 | elv_insert(q, rq, ELEVATOR_INSERT_FRONT); | |
151 | } | |
152 | ||
8f11b3e9 | 153 | static inline bool start_ordered(struct request_queue *q, struct request **rqp) |
86db1e29 | 154 | { |
8f11b3e9 TH |
155 | struct request *rq = *rqp; |
156 | unsigned skip = 0; | |
157 | ||
86db1e29 JA |
158 | q->orderr = 0; |
159 | q->ordered = q->next_ordered; | |
160 | q->ordseq |= QUEUE_ORDSEQ_STARTED; | |
161 | ||
58eea927 TH |
162 | /* |
163 | * For an empty barrier, there's no actual BAR request, which | |
164 | * in turn makes POSTFLUSH unnecessary. Mask them off. | |
165 | */ | |
5b93629b | 166 | if (!blk_rq_sectors(rq)) { |
58eea927 TH |
167 | q->ordered &= ~(QUEUE_ORDERED_DO_BAR | |
168 | QUEUE_ORDERED_DO_POSTFLUSH); | |
a185eb4b TH |
169 | /* |
170 | * Empty barrier on a write-through device w/ ordered | |
171 | * tag has no command to issue and without any command | |
172 | * to issue, ordering by tag can't be used. Drain | |
173 | * instead. | |
174 | */ | |
175 | if ((q->ordered & QUEUE_ORDERED_BY_TAG) && | |
176 | !(q->ordered & QUEUE_ORDERED_DO_PREFLUSH)) { | |
177 | q->ordered &= ~QUEUE_ORDERED_BY_TAG; | |
178 | q->ordered |= QUEUE_ORDERED_BY_DRAIN; | |
179 | } | |
180 | } | |
58eea927 | 181 | |
f671620e | 182 | /* stash away the original request */ |
9934c8c0 | 183 | blk_dequeue_request(rq); |
86db1e29 | 184 | q->orig_bar_rq = rq; |
f671620e | 185 | rq = NULL; |
86db1e29 JA |
186 | |
187 | /* | |
188 | * Queue ordered sequence. As we stack them at the head, we | |
189 | * need to queue in reverse order. Note that we rely on that | |
190 | * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs | |
58eea927 | 191 | * request gets inbetween ordered sequence. |
86db1e29 | 192 | */ |
58eea927 | 193 | if (q->ordered & QUEUE_ORDERED_DO_POSTFLUSH) { |
313e4299 | 194 | queue_flush(q, QUEUE_ORDERED_DO_POSTFLUSH); |
f671620e TH |
195 | rq = &q->post_flush_rq; |
196 | } else | |
8f11b3e9 | 197 | skip |= QUEUE_ORDSEQ_POSTFLUSH; |
86db1e29 | 198 | |
f671620e TH |
199 | if (q->ordered & QUEUE_ORDERED_DO_BAR) { |
200 | rq = &q->bar_rq; | |
201 | ||
202 | /* initialize proxy request and queue it */ | |
203 | blk_rq_init(q, rq); | |
204 | if (bio_data_dir(q->orig_bar_rq->bio) == WRITE) | |
205 | rq->cmd_flags |= REQ_RW; | |
206 | if (q->ordered & QUEUE_ORDERED_DO_FUA) | |
207 | rq->cmd_flags |= REQ_FUA; | |
208 | init_request_from_bio(rq, q->orig_bar_rq->bio); | |
209 | rq->end_io = bar_end_io; | |
210 | ||
211 | elv_insert(q, rq, ELEVATOR_INSERT_FRONT); | |
212 | } else | |
8f11b3e9 | 213 | skip |= QUEUE_ORDSEQ_BAR; |
86db1e29 | 214 | |
313e4299 TH |
215 | if (q->ordered & QUEUE_ORDERED_DO_PREFLUSH) { |
216 | queue_flush(q, QUEUE_ORDERED_DO_PREFLUSH); | |
86db1e29 JA |
217 | rq = &q->pre_flush_rq; |
218 | } else | |
8f11b3e9 | 219 | skip |= QUEUE_ORDSEQ_PREFLUSH; |
86db1e29 | 220 | |
0a7ae2ff | 221 | if ((q->ordered & QUEUE_ORDERED_BY_DRAIN) && queue_in_flight(q)) |
86db1e29 | 222 | rq = NULL; |
f671620e | 223 | else |
8f11b3e9 | 224 | skip |= QUEUE_ORDSEQ_DRAIN; |
86db1e29 | 225 | |
8f11b3e9 TH |
226 | *rqp = rq; |
227 | ||
228 | /* | |
229 | * Complete skipped sequences. If whole sequence is complete, | |
230 | * return false to tell elevator that this request is gone. | |
231 | */ | |
232 | return !blk_ordered_complete_seq(q, skip, 0); | |
86db1e29 JA |
233 | } |
234 | ||
8f11b3e9 | 235 | bool blk_do_ordered(struct request_queue *q, struct request **rqp) |
86db1e29 JA |
236 | { |
237 | struct request *rq = *rqp; | |
238 | const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq); | |
239 | ||
240 | if (!q->ordseq) { | |
241 | if (!is_barrier) | |
8f11b3e9 | 242 | return true; |
86db1e29 | 243 | |
8f11b3e9 TH |
244 | if (q->next_ordered != QUEUE_ORDERED_NONE) |
245 | return start_ordered(q, rqp); | |
246 | else { | |
86db1e29 | 247 | /* |
a7384677 TH |
248 | * Queue ordering not supported. Terminate |
249 | * with prejudice. | |
86db1e29 | 250 | */ |
9934c8c0 | 251 | blk_dequeue_request(rq); |
40cbbb78 | 252 | __blk_end_request_all(rq, -EOPNOTSUPP); |
86db1e29 | 253 | *rqp = NULL; |
8f11b3e9 | 254 | return false; |
86db1e29 JA |
255 | } |
256 | } | |
257 | ||
258 | /* | |
259 | * Ordered sequence in progress | |
260 | */ | |
261 | ||
262 | /* Special requests are not subject to ordering rules. */ | |
263 | if (!blk_fs_request(rq) && | |
264 | rq != &q->pre_flush_rq && rq != &q->post_flush_rq) | |
8f11b3e9 | 265 | return true; |
86db1e29 | 266 | |
313e4299 | 267 | if (q->ordered & QUEUE_ORDERED_BY_TAG) { |
86db1e29 JA |
268 | /* Ordered by tag. Blocking the next barrier is enough. */ |
269 | if (is_barrier && rq != &q->bar_rq) | |
270 | *rqp = NULL; | |
271 | } else { | |
272 | /* Ordered by draining. Wait for turn. */ | |
273 | WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q)); | |
274 | if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q)) | |
275 | *rqp = NULL; | |
276 | } | |
277 | ||
8f11b3e9 | 278 | return true; |
86db1e29 JA |
279 | } |
280 | ||
281 | static void bio_end_empty_barrier(struct bio *bio, int err) | |
282 | { | |
cc66b451 JA |
283 | if (err) { |
284 | if (err == -EOPNOTSUPP) | |
285 | set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); | |
86db1e29 | 286 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
cc66b451 | 287 | } |
86db1e29 JA |
288 | |
289 | complete(bio->bi_private); | |
290 | } | |
291 | ||
292 | /** | |
293 | * blkdev_issue_flush - queue a flush | |
294 | * @bdev: blockdev to issue flush for | |
295 | * @error_sector: error sector | |
296 | * | |
297 | * Description: | |
298 | * Issue a flush for the block device in question. Caller can supply | |
299 | * room for storing the error offset in case of a flush error, if they | |
dbdac9b7 | 300 | * wish to. |
86db1e29 JA |
301 | */ |
302 | int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector) | |
303 | { | |
304 | DECLARE_COMPLETION_ONSTACK(wait); | |
305 | struct request_queue *q; | |
306 | struct bio *bio; | |
307 | int ret; | |
308 | ||
309 | if (bdev->bd_disk == NULL) | |
310 | return -ENXIO; | |
311 | ||
312 | q = bdev_get_queue(bdev); | |
313 | if (!q) | |
314 | return -ENXIO; | |
315 | ||
316 | bio = bio_alloc(GFP_KERNEL, 0); | |
86db1e29 JA |
317 | bio->bi_end_io = bio_end_empty_barrier; |
318 | bio->bi_private = &wait; | |
319 | bio->bi_bdev = bdev; | |
2ebca85a | 320 | submit_bio(WRITE_BARRIER, bio); |
86db1e29 JA |
321 | |
322 | wait_for_completion(&wait); | |
323 | ||
324 | /* | |
325 | * The driver must store the error location in ->bi_sector, if | |
326 | * it supports it. For non-stacked drivers, this should be copied | |
83096ebf | 327 | * from blk_rq_pos(rq). |
86db1e29 JA |
328 | */ |
329 | if (error_sector) | |
330 | *error_sector = bio->bi_sector; | |
331 | ||
332 | ret = 0; | |
cc66b451 JA |
333 | if (bio_flagged(bio, BIO_EOPNOTSUPP)) |
334 | ret = -EOPNOTSUPP; | |
335 | else if (!bio_flagged(bio, BIO_UPTODATE)) | |
86db1e29 JA |
336 | ret = -EIO; |
337 | ||
338 | bio_put(bio); | |
339 | return ret; | |
340 | } | |
86db1e29 | 341 | EXPORT_SYMBOL(blkdev_issue_flush); |
fb2dce86 DW |
342 | |
343 | static void blkdev_discard_end_io(struct bio *bio, int err) | |
344 | { | |
345 | if (err) { | |
346 | if (err == -EOPNOTSUPP) | |
347 | set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); | |
348 | clear_bit(BIO_UPTODATE, &bio->bi_flags); | |
349 | } | |
350 | ||
746cd1e7 CH |
351 | if (bio->bi_private) |
352 | complete(bio->bi_private); | |
c15227de | 353 | __free_page(bio_page(bio)); |
746cd1e7 | 354 | |
fb2dce86 DW |
355 | bio_put(bio); |
356 | } | |
357 | ||
358 | /** | |
359 | * blkdev_issue_discard - queue a discard | |
360 | * @bdev: blockdev to issue discard for | |
361 | * @sector: start sector | |
362 | * @nr_sects: number of sectors to discard | |
3e6053d7 | 363 | * @gfp_mask: memory allocation flags (for bio_alloc) |
746cd1e7 | 364 | * @flags: DISCARD_FL_* flags to control behaviour |
fb2dce86 DW |
365 | * |
366 | * Description: | |
746cd1e7 | 367 | * Issue a discard request for the sectors in question. |
fb2dce86 | 368 | */ |
746cd1e7 CH |
369 | int blkdev_issue_discard(struct block_device *bdev, sector_t sector, |
370 | sector_t nr_sects, gfp_t gfp_mask, int flags) | |
fb2dce86 | 371 | { |
746cd1e7 CH |
372 | DECLARE_COMPLETION_ONSTACK(wait); |
373 | struct request_queue *q = bdev_get_queue(bdev); | |
374 | int type = flags & DISCARD_FL_BARRIER ? | |
375 | DISCARD_BARRIER : DISCARD_NOBARRIER; | |
c15227de CH |
376 | struct bio *bio; |
377 | struct page *page; | |
fb2dce86 DW |
378 | int ret = 0; |
379 | ||
fb2dce86 DW |
380 | if (!q) |
381 | return -ENXIO; | |
382 | ||
c15227de | 383 | if (!blk_queue_discard(q)) |
fb2dce86 DW |
384 | return -EOPNOTSUPP; |
385 | ||
386 | while (nr_sects && !ret) { | |
c15227de | 387 | unsigned int sector_size = q->limits.logical_block_size; |
fb2dce86 | 388 | |
c15227de CH |
389 | bio = bio_alloc(gfp_mask, 1); |
390 | if (!bio) | |
391 | goto out; | |
392 | bio->bi_sector = sector; | |
fb2dce86 DW |
393 | bio->bi_end_io = blkdev_discard_end_io; |
394 | bio->bi_bdev = bdev; | |
746cd1e7 CH |
395 | if (flags & DISCARD_FL_WAIT) |
396 | bio->bi_private = &wait; | |
fb2dce86 | 397 | |
c15227de CH |
398 | /* |
399 | * Add a zeroed one-sector payload as that's what | |
400 | * our current implementations need. If we'll ever need | |
401 | * more the interface will need revisiting. | |
402 | */ | |
403 | page = alloc_page(GFP_KERNEL | __GFP_ZERO); | |
404 | if (!page) | |
405 | goto out_free_bio; | |
406 | if (bio_add_pc_page(q, bio, page, sector_size, 0) < sector_size) | |
407 | goto out_free_page; | |
fb2dce86 | 408 | |
c15227de CH |
409 | /* |
410 | * And override the bio size - the way discard works we | |
411 | * touch many more blocks on disk than the actual payload | |
412 | * length. | |
413 | */ | |
ae03bf63 MP |
414 | if (nr_sects > queue_max_hw_sectors(q)) { |
415 | bio->bi_size = queue_max_hw_sectors(q) << 9; | |
416 | nr_sects -= queue_max_hw_sectors(q); | |
417 | sector += queue_max_hw_sectors(q); | |
fb2dce86 DW |
418 | } else { |
419 | bio->bi_size = nr_sects << 9; | |
420 | nr_sects = 0; | |
421 | } | |
746cd1e7 | 422 | |
fb2dce86 | 423 | bio_get(bio); |
746cd1e7 CH |
424 | submit_bio(type, bio); |
425 | ||
426 | if (flags & DISCARD_FL_WAIT) | |
427 | wait_for_completion(&wait); | |
fb2dce86 | 428 | |
fb2dce86 DW |
429 | if (bio_flagged(bio, BIO_EOPNOTSUPP)) |
430 | ret = -EOPNOTSUPP; | |
431 | else if (!bio_flagged(bio, BIO_UPTODATE)) | |
432 | ret = -EIO; | |
433 | bio_put(bio); | |
434 | } | |
435 | return ret; | |
c15227de CH |
436 | out_free_page: |
437 | __free_page(page); | |
438 | out_free_bio: | |
439 | bio_put(bio); | |
440 | out: | |
441 | return -ENOMEM; | |
fb2dce86 DW |
442 | } |
443 | EXPORT_SYMBOL(blkdev_issue_discard); |