blob: f7dae57e6cabdcfaaed15762c4d2b7b9889e6c77 [file] [log] [blame]
Jens Axboe86db1e22008-01-29 14:53:40 +01001/*
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 **/
24int blk_queue_ordered(struct request_queue *q, unsigned ordered,
25 prepare_flush_fn *prepare_flush_fn)
26{
Tejun Heo313e4292008-11-28 13:32:02 +090027 if (!prepare_flush_fn && (ordered & (QUEUE_ORDERED_DO_PREFLUSH |
28 QUEUE_ORDERED_DO_POSTFLUSH))) {
Harvey Harrison24c03d42008-05-01 04:35:17 -070029 printk(KERN_ERR "%s: prepare_flush_fn required\n", __func__);
Jens Axboe86db1e22008-01-29 14:53:40 +010030 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}
Jens Axboe86db1e22008-01-29 14:53:40 +010050EXPORT_SYMBOL(blk_queue_ordered);
51
52/*
53 * Cache flushing for ordered writes handling
54 */
Adrian Bunk6f6a0362008-04-29 09:49:06 +020055unsigned blk_ordered_cur_seq(struct request_queue *q)
Jens Axboe86db1e22008-01-29 14:53:40 +010056{
57 if (!q->ordseq)
58 return 0;
59 return 1 << ffz(q->ordseq);
60}
61
62unsigned 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
Tejun Heo8f11b3e2008-11-28 13:32:05 +090091bool blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +010092{
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)
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900102 return false;
Jens Axboe86db1e22008-01-29 14:53:40 +0100103
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();
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900112
113 return true;
Jens Axboe86db1e22008-01-29 14:53:40 +0100114}
115
116static 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
122static 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
128static 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
134static void queue_flush(struct request_queue *q, unsigned which)
135{
136 struct request *rq;
137 rq_end_io_fn *end_io;
138
Tejun Heo313e4292008-11-28 13:32:02 +0900139 if (which == QUEUE_ORDERED_DO_PREFLUSH) {
Jens Axboe86db1e22008-01-29 14:53:40 +0100140 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
FUJITA Tomonori2a4aa302008-04-29 09:54:36 +0200147 blk_rq_init(q, rq);
FUJITA Tomonori1afb20f2008-04-25 12:26:28 +0200148 rq->cmd_flags = REQ_HARDBARRIER;
Jens Axboe86db1e22008-01-29 14:53:40 +0100149 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
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900156static inline bool start_ordered(struct request_queue *q, struct request **rqp)
Jens Axboe86db1e22008-01-29 14:53:40 +0100157{
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900158 struct request *rq = *rqp;
159 unsigned skip = 0;
160
Jens Axboe86db1e22008-01-29 14:53:40 +0100161 q->orderr = 0;
162 q->ordered = q->next_ordered;
163 q->ordseq |= QUEUE_ORDSEQ_STARTED;
164
Tejun Heo58eea922008-11-28 13:32:06 +0900165 /*
166 * For an empty barrier, there's no actual BAR request, which
167 * in turn makes POSTFLUSH unnecessary. Mask them off.
168 */
Tejun Heoa185eb42008-11-28 13:32:07 +0900169 if (!rq->hard_nr_sectors) {
Tejun Heo58eea922008-11-28 13:32:06 +0900170 q->ordered &= ~(QUEUE_ORDERED_DO_BAR |
171 QUEUE_ORDERED_DO_POSTFLUSH);
Tejun Heoa185eb42008-11-28 13:32:07 +0900172 /*
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 }
Tejun Heo58eea922008-11-28 13:32:06 +0900184
Tejun Heof6716202008-11-28 13:32:04 +0900185 /* stash away the original request */
Tejun Heo53a08802008-12-03 12:41:26 +0100186 elv_dequeue_request(q, rq);
Jens Axboe86db1e22008-01-29 14:53:40 +0100187 q->orig_bar_rq = rq;
Tejun Heof6716202008-11-28 13:32:04 +0900188 rq = NULL;
Jens Axboe86db1e22008-01-29 14:53:40 +0100189
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
Tejun Heo58eea922008-11-28 13:32:06 +0900194 * request gets inbetween ordered sequence.
Jens Axboe86db1e22008-01-29 14:53:40 +0100195 */
Tejun Heo58eea922008-11-28 13:32:06 +0900196 if (q->ordered & QUEUE_ORDERED_DO_POSTFLUSH) {
Tejun Heo313e4292008-11-28 13:32:02 +0900197 queue_flush(q, QUEUE_ORDERED_DO_POSTFLUSH);
Tejun Heof6716202008-11-28 13:32:04 +0900198 rq = &q->post_flush_rq;
199 } else
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900200 skip |= QUEUE_ORDSEQ_POSTFLUSH;
Jens Axboe86db1e22008-01-29 14:53:40 +0100201
Tejun Heof6716202008-11-28 13:32:04 +0900202 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
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900216 skip |= QUEUE_ORDSEQ_BAR;
Jens Axboe86db1e22008-01-29 14:53:40 +0100217
Tejun Heo313e4292008-11-28 13:32:02 +0900218 if (q->ordered & QUEUE_ORDERED_DO_PREFLUSH) {
219 queue_flush(q, QUEUE_ORDERED_DO_PREFLUSH);
Jens Axboe86db1e22008-01-29 14:53:40 +0100220 rq = &q->pre_flush_rq;
221 } else
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900222 skip |= QUEUE_ORDSEQ_PREFLUSH;
Jens Axboe86db1e22008-01-29 14:53:40 +0100223
Tejun Heof6716202008-11-28 13:32:04 +0900224 if ((q->ordered & QUEUE_ORDERED_BY_DRAIN) && q->in_flight)
Jens Axboe86db1e22008-01-29 14:53:40 +0100225 rq = NULL;
Tejun Heof6716202008-11-28 13:32:04 +0900226 else
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900227 skip |= QUEUE_ORDSEQ_DRAIN;
Jens Axboe86db1e22008-01-29 14:53:40 +0100228
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900229 *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);
Jens Axboe86db1e22008-01-29 14:53:40 +0100236}
237
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900238bool blk_do_ordered(struct request_queue *q, struct request **rqp)
Jens Axboe86db1e22008-01-29 14:53:40 +0100239{
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)
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900245 return true;
Jens Axboe86db1e22008-01-29 14:53:40 +0100246
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900247 if (q->next_ordered != QUEUE_ORDERED_NONE)
248 return start_ordered(q, rqp);
249 else {
Jens Axboe86db1e22008-01-29 14:53:40 +0100250 /*
Tejun Heoa7384672008-11-28 13:32:03 +0900251 * Queue ordering not supported. Terminate
252 * with prejudice.
Jens Axboe86db1e22008-01-29 14:53:40 +0100253 */
Tejun Heo53a08802008-12-03 12:41:26 +0100254 elv_dequeue_request(q, rq);
Jens Axboe86db1e22008-01-29 14:53:40 +0100255 if (__blk_end_request(rq, -EOPNOTSUPP,
256 blk_rq_bytes(rq)))
257 BUG();
258 *rqp = NULL;
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900259 return false;
Jens Axboe86db1e22008-01-29 14:53:40 +0100260 }
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)
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900270 return true;
Jens Axboe86db1e22008-01-29 14:53:40 +0100271
Tejun Heo313e4292008-11-28 13:32:02 +0900272 if (q->ordered & QUEUE_ORDERED_BY_TAG) {
Jens Axboe86db1e22008-01-29 14:53:40 +0100273 /* 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
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900283 return true;
Jens Axboe86db1e22008-01-29 14:53:40 +0100284}
285
286static void bio_end_empty_barrier(struct bio *bio, int err)
287{
Jens Axboecc66b452008-03-04 11:47:46 +0100288 if (err) {
289 if (err == -EOPNOTSUPP)
290 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
Jens Axboe86db1e22008-01-29 14:53:40 +0100291 clear_bit(BIO_UPTODATE, &bio->bi_flags);
Jens Axboecc66b452008-03-04 11:47:46 +0100292 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100293
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
Theodore Ts'odbdac9b2009-01-13 15:27:32 +0100305 * wish to.
Jens Axboe86db1e22008-01-29 14:53:40 +0100306 */
307int 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;
OGAWA Hirofumi2ebca852008-08-11 17:07:08 +0100328 submit_bio(WRITE_BARRIER, bio);
Jens Axboe86db1e22008-01-29 14:53:40 +0100329
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;
Jens Axboecc66b452008-03-04 11:47:46 +0100341 if (bio_flagged(bio, BIO_EOPNOTSUPP))
342 ret = -EOPNOTSUPP;
343 else if (!bio_flagged(bio, BIO_UPTODATE))
Jens Axboe86db1e22008-01-29 14:53:40 +0100344 ret = -EIO;
345
346 bio_put(bio);
347 return ret;
348}
Jens Axboe86db1e22008-01-29 14:53:40 +0100349EXPORT_SYMBOL(blkdev_issue_flush);
David Woodhousefb2dce82008-08-05 18:01:53 +0100350
351static 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
Hugh Dickins3e6053d2008-09-11 10:57:55 +0200367 * @gfp_mask: memory allocation flags (for bio_alloc)
David Woodhousefb2dce82008-08-05 18:01:53 +0100368 *
369 * Description:
370 * Issue a discard request for the sectors in question. Does not wait.
371 */
Hugh Dickins3e6053d2008-09-11 10:57:55 +0200372int blkdev_issue_discard(struct block_device *bdev,
373 sector_t sector, sector_t nr_sects, gfp_t gfp_mask)
David Woodhousefb2dce82008-08-05 18:01:53 +0100374{
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) {
Hugh Dickins3e6053d2008-09-11 10:57:55 +0200390 bio = bio_alloc(gfp_mask, 0);
David Woodhousefb2dce82008-08-05 18:01:53 +0100391 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);
David Woodhousee17fc0a2008-08-09 16:42:20 +0100408 submit_bio(DISCARD_BARRIER, bio);
David Woodhousefb2dce82008-08-05 18:01:53 +0100409
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}
419EXPORT_SYMBOL(blkdev_issue_discard);