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; | |
109 | ||
110 | if (__blk_end_request(rq, q->orderr, blk_rq_bytes(rq))) | |
111 | BUG(); | |
8f11b3e9 TH |
112 | |
113 | return true; | |
86db1e29 JA |
114 | } |
115 | ||
116 | static void pre_flush_end_io(struct request *rq, int error) | |
117 | { | |
118 | elv_completed_request(rq->q, rq); | |
119 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error); | |
120 | } | |
121 | ||
122 | static void bar_end_io(struct request *rq, int error) | |
123 | { | |
124 | elv_completed_request(rq->q, rq); | |
125 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error); | |
126 | } | |
127 | ||
128 | static void post_flush_end_io(struct request *rq, int error) | |
129 | { | |
130 | elv_completed_request(rq->q, rq); | |
131 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error); | |
132 | } | |
133 | ||
134 | static void queue_flush(struct request_queue *q, unsigned which) | |
135 | { | |
136 | struct request *rq; | |
137 | rq_end_io_fn *end_io; | |
138 | ||
313e4299 | 139 | if (which == QUEUE_ORDERED_DO_PREFLUSH) { |
86db1e29 JA |
140 | rq = &q->pre_flush_rq; |
141 | end_io = pre_flush_end_io; | |
142 | } else { | |
143 | rq = &q->post_flush_rq; | |
144 | end_io = post_flush_end_io; | |
145 | } | |
146 | ||
2a4aa30c | 147 | blk_rq_init(q, rq); |
1afb20f3 | 148 | rq->cmd_flags = REQ_HARDBARRIER; |
86db1e29 JA |
149 | rq->rq_disk = q->bar_rq.rq_disk; |
150 | rq->end_io = end_io; | |
151 | q->prepare_flush_fn(q, rq); | |
152 | ||
153 | elv_insert(q, rq, ELEVATOR_INSERT_FRONT); | |
154 | } | |
155 | ||
8f11b3e9 | 156 | static inline bool start_ordered(struct request_queue *q, struct request **rqp) |
86db1e29 | 157 | { |
8f11b3e9 TH |
158 | struct request *rq = *rqp; |
159 | unsigned skip = 0; | |
160 | ||
86db1e29 JA |
161 | q->orderr = 0; |
162 | q->ordered = q->next_ordered; | |
163 | q->ordseq |= QUEUE_ORDSEQ_STARTED; | |
164 | ||
58eea927 TH |
165 | /* |
166 | * For an empty barrier, there's no actual BAR request, which | |
167 | * in turn makes POSTFLUSH unnecessary. Mask them off. | |
168 | */ | |
a185eb4b | 169 | if (!rq->hard_nr_sectors) { |
58eea927 TH |
170 | q->ordered &= ~(QUEUE_ORDERED_DO_BAR | |
171 | QUEUE_ORDERED_DO_POSTFLUSH); | |
a185eb4b TH |
172 | /* |
173 | * Empty barrier on a write-through device w/ ordered | |
174 | * tag has no command to issue and without any command | |
175 | * to issue, ordering by tag can't be used. Drain | |
176 | * instead. | |
177 | */ | |
178 | if ((q->ordered & QUEUE_ORDERED_BY_TAG) && | |
179 | !(q->ordered & QUEUE_ORDERED_DO_PREFLUSH)) { | |
180 | q->ordered &= ~QUEUE_ORDERED_BY_TAG; | |
181 | q->ordered |= QUEUE_ORDERED_BY_DRAIN; | |
182 | } | |
183 | } | |
58eea927 | 184 | |
f671620e | 185 | /* stash away the original request */ |
53a08807 | 186 | elv_dequeue_request(q, rq); |
86db1e29 | 187 | q->orig_bar_rq = rq; |
f671620e | 188 | rq = NULL; |
86db1e29 JA |
189 | |
190 | /* | |
191 | * Queue ordered sequence. As we stack them at the head, we | |
192 | * need to queue in reverse order. Note that we rely on that | |
193 | * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs | |
58eea927 | 194 | * request gets inbetween ordered sequence. |
86db1e29 | 195 | */ |
58eea927 | 196 | if (q->ordered & QUEUE_ORDERED_DO_POSTFLUSH) { |
313e4299 | 197 | queue_flush(q, QUEUE_ORDERED_DO_POSTFLUSH); |
f671620e TH |
198 | rq = &q->post_flush_rq; |
199 | } else | |
8f11b3e9 | 200 | skip |= QUEUE_ORDSEQ_POSTFLUSH; |
86db1e29 | 201 | |
f671620e TH |
202 | if (q->ordered & QUEUE_ORDERED_DO_BAR) { |
203 | rq = &q->bar_rq; | |
204 | ||
205 | /* initialize proxy request and queue it */ | |
206 | blk_rq_init(q, rq); | |
207 | if (bio_data_dir(q->orig_bar_rq->bio) == WRITE) | |
208 | rq->cmd_flags |= REQ_RW; | |
209 | if (q->ordered & QUEUE_ORDERED_DO_FUA) | |
210 | rq->cmd_flags |= REQ_FUA; | |
211 | init_request_from_bio(rq, q->orig_bar_rq->bio); | |
212 | rq->end_io = bar_end_io; | |
213 | ||
214 | elv_insert(q, rq, ELEVATOR_INSERT_FRONT); | |
215 | } else | |
8f11b3e9 | 216 | skip |= QUEUE_ORDSEQ_BAR; |
86db1e29 | 217 | |
313e4299 TH |
218 | if (q->ordered & QUEUE_ORDERED_DO_PREFLUSH) { |
219 | queue_flush(q, QUEUE_ORDERED_DO_PREFLUSH); | |
86db1e29 JA |
220 | rq = &q->pre_flush_rq; |
221 | } else | |
8f11b3e9 | 222 | skip |= QUEUE_ORDSEQ_PREFLUSH; |
86db1e29 | 223 | |
f671620e | 224 | if ((q->ordered & QUEUE_ORDERED_BY_DRAIN) && q->in_flight) |
86db1e29 | 225 | rq = NULL; |
f671620e | 226 | else |
8f11b3e9 | 227 | skip |= QUEUE_ORDSEQ_DRAIN; |
86db1e29 | 228 | |
8f11b3e9 TH |
229 | *rqp = rq; |
230 | ||
231 | /* | |
232 | * Complete skipped sequences. If whole sequence is complete, | |
233 | * return false to tell elevator that this request is gone. | |
234 | */ | |
235 | return !blk_ordered_complete_seq(q, skip, 0); | |
86db1e29 JA |
236 | } |
237 | ||
8f11b3e9 | 238 | bool blk_do_ordered(struct request_queue *q, struct request **rqp) |
86db1e29 JA |
239 | { |
240 | struct request *rq = *rqp; | |
241 | const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq); | |
242 | ||
243 | if (!q->ordseq) { | |
244 | if (!is_barrier) | |
8f11b3e9 | 245 | return true; |
86db1e29 | 246 | |
8f11b3e9 TH |
247 | if (q->next_ordered != QUEUE_ORDERED_NONE) |
248 | return start_ordered(q, rqp); | |
249 | else { | |
86db1e29 | 250 | /* |
a7384677 TH |
251 | * Queue ordering not supported. Terminate |
252 | * with prejudice. | |
86db1e29 | 253 | */ |
53a08807 | 254 | elv_dequeue_request(q, rq); |
86db1e29 JA |
255 | if (__blk_end_request(rq, -EOPNOTSUPP, |
256 | blk_rq_bytes(rq))) | |
257 | BUG(); | |
258 | *rqp = NULL; | |
8f11b3e9 | 259 | return false; |
86db1e29 JA |
260 | } |
261 | } | |
262 | ||
263 | /* | |
264 | * Ordered sequence in progress | |
265 | */ | |
266 | ||
267 | /* Special requests are not subject to ordering rules. */ | |
268 | if (!blk_fs_request(rq) && | |
269 | rq != &q->pre_flush_rq && rq != &q->post_flush_rq) | |
8f11b3e9 | 270 | return true; |
86db1e29 | 271 | |
313e4299 | 272 | if (q->ordered & QUEUE_ORDERED_BY_TAG) { |
86db1e29 JA |
273 | /* Ordered by tag. Blocking the next barrier is enough. */ |
274 | if (is_barrier && rq != &q->bar_rq) | |
275 | *rqp = NULL; | |
276 | } else { | |
277 | /* Ordered by draining. Wait for turn. */ | |
278 | WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q)); | |
279 | if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q)) | |
280 | *rqp = NULL; | |
281 | } | |
282 | ||
8f11b3e9 | 283 | return true; |
86db1e29 JA |
284 | } |
285 | ||
286 | static void bio_end_empty_barrier(struct bio *bio, int err) | |
287 | { | |
cc66b451 JA |
288 | if (err) { |
289 | if (err == -EOPNOTSUPP) | |
290 | set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); | |
86db1e29 | 291 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
cc66b451 | 292 | } |
86db1e29 JA |
293 | |
294 | complete(bio->bi_private); | |
295 | } | |
296 | ||
297 | /** | |
298 | * blkdev_issue_flush - queue a flush | |
299 | * @bdev: blockdev to issue flush for | |
300 | * @error_sector: error sector | |
301 | * | |
302 | * Description: | |
303 | * Issue a flush for the block device in question. Caller can supply | |
304 | * room for storing the error offset in case of a flush error, if they | |
dbdac9b7 | 305 | * wish to. |
86db1e29 JA |
306 | */ |
307 | int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector) | |
308 | { | |
309 | DECLARE_COMPLETION_ONSTACK(wait); | |
310 | struct request_queue *q; | |
311 | struct bio *bio; | |
312 | int ret; | |
313 | ||
314 | if (bdev->bd_disk == NULL) | |
315 | return -ENXIO; | |
316 | ||
317 | q = bdev_get_queue(bdev); | |
318 | if (!q) | |
319 | return -ENXIO; | |
320 | ||
321 | bio = bio_alloc(GFP_KERNEL, 0); | |
322 | if (!bio) | |
323 | return -ENOMEM; | |
324 | ||
325 | bio->bi_end_io = bio_end_empty_barrier; | |
326 | bio->bi_private = &wait; | |
327 | bio->bi_bdev = bdev; | |
2ebca85a | 328 | submit_bio(WRITE_BARRIER, bio); |
86db1e29 JA |
329 | |
330 | wait_for_completion(&wait); | |
331 | ||
332 | /* | |
333 | * The driver must store the error location in ->bi_sector, if | |
334 | * it supports it. For non-stacked drivers, this should be copied | |
335 | * from rq->sector. | |
336 | */ | |
337 | if (error_sector) | |
338 | *error_sector = bio->bi_sector; | |
339 | ||
340 | ret = 0; | |
cc66b451 JA |
341 | if (bio_flagged(bio, BIO_EOPNOTSUPP)) |
342 | ret = -EOPNOTSUPP; | |
343 | else if (!bio_flagged(bio, BIO_UPTODATE)) | |
86db1e29 JA |
344 | ret = -EIO; |
345 | ||
346 | bio_put(bio); | |
347 | return ret; | |
348 | } | |
86db1e29 | 349 | EXPORT_SYMBOL(blkdev_issue_flush); |
fb2dce86 DW |
350 | |
351 | static void blkdev_discard_end_io(struct bio *bio, int err) | |
352 | { | |
353 | if (err) { | |
354 | if (err == -EOPNOTSUPP) | |
355 | set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); | |
356 | clear_bit(BIO_UPTODATE, &bio->bi_flags); | |
357 | } | |
358 | ||
359 | bio_put(bio); | |
360 | } | |
361 | ||
362 | /** | |
363 | * blkdev_issue_discard - queue a discard | |
364 | * @bdev: blockdev to issue discard for | |
365 | * @sector: start sector | |
366 | * @nr_sects: number of sectors to discard | |
3e6053d7 | 367 | * @gfp_mask: memory allocation flags (for bio_alloc) |
fb2dce86 DW |
368 | * |
369 | * Description: | |
370 | * Issue a discard request for the sectors in question. Does not wait. | |
371 | */ | |
3e6053d7 HD |
372 | int blkdev_issue_discard(struct block_device *bdev, |
373 | sector_t sector, sector_t nr_sects, gfp_t gfp_mask) | |
fb2dce86 DW |
374 | { |
375 | struct request_queue *q; | |
376 | struct bio *bio; | |
377 | int ret = 0; | |
378 | ||
379 | if (bdev->bd_disk == NULL) | |
380 | return -ENXIO; | |
381 | ||
382 | q = bdev_get_queue(bdev); | |
383 | if (!q) | |
384 | return -ENXIO; | |
385 | ||
386 | if (!q->prepare_discard_fn) | |
387 | return -EOPNOTSUPP; | |
388 | ||
389 | while (nr_sects && !ret) { | |
3e6053d7 | 390 | bio = bio_alloc(gfp_mask, 0); |
fb2dce86 DW |
391 | if (!bio) |
392 | return -ENOMEM; | |
393 | ||
394 | bio->bi_end_io = blkdev_discard_end_io; | |
395 | bio->bi_bdev = bdev; | |
396 | ||
397 | bio->bi_sector = sector; | |
398 | ||
399 | if (nr_sects > q->max_hw_sectors) { | |
400 | bio->bi_size = q->max_hw_sectors << 9; | |
401 | nr_sects -= q->max_hw_sectors; | |
402 | sector += q->max_hw_sectors; | |
403 | } else { | |
404 | bio->bi_size = nr_sects << 9; | |
405 | nr_sects = 0; | |
406 | } | |
407 | bio_get(bio); | |
e17fc0a1 | 408 | submit_bio(DISCARD_BARRIER, bio); |
fb2dce86 DW |
409 | |
410 | /* Check if it failed immediately */ | |
411 | if (bio_flagged(bio, BIO_EOPNOTSUPP)) | |
412 | ret = -EOPNOTSUPP; | |
413 | else if (!bio_flagged(bio, BIO_UPTODATE)) | |
414 | ret = -EIO; | |
415 | bio_put(bio); | |
416 | } | |
417 | return ret; | |
418 | } | |
419 | EXPORT_SYMBOL(blkdev_issue_discard); |