273121c0eb805c57ce0138ffc2dfec0b2dc67a28
[linux-2.6.git] / block / blk-barrier.c
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 {
27         if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
28             prepare_flush_fn == NULL) {
29                 printk(KERN_ERR "%s: prepare_flush_fn required\n", __func__);
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 }
50 EXPORT_SYMBOL(blk_queue_ordered);
51
52 /*
53  * Cache flushing for ordered writes handling
54  */
55 unsigned blk_ordered_cur_seq(struct request_queue *q)
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
91 void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
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)
102                 return;
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();
112 }
113
114 static void pre_flush_end_io(struct request *rq, int error)
115 {
116         elv_completed_request(rq->q, rq);
117         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
118 }
119
120 static void bar_end_io(struct request *rq, int error)
121 {
122         elv_completed_request(rq->q, rq);
123         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
124 }
125
126 static void post_flush_end_io(struct request *rq, int error)
127 {
128         elv_completed_request(rq->q, rq);
129         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
130 }
131
132 static void queue_flush(struct request_queue *q, unsigned which)
133 {
134         struct request *rq;
135         rq_end_io_fn *end_io;
136
137         if (which == QUEUE_ORDERED_PREFLUSH) {
138                 rq = &q->pre_flush_rq;
139                 end_io = pre_flush_end_io;
140         } else {
141                 rq = &q->post_flush_rq;
142                 end_io = post_flush_end_io;
143         }
144
145         blk_rq_init(q, rq);
146         rq->cmd_flags = REQ_HARDBARRIER;
147         rq->rq_disk = q->bar_rq.rq_disk;
148         rq->end_io = end_io;
149         q->prepare_flush_fn(q, rq);
150
151         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
152 }
153
154 static inline struct request *start_ordered(struct request_queue *q,
155                                             struct request *rq)
156 {
157         q->orderr = 0;
158         q->ordered = q->next_ordered;
159         q->ordseq |= QUEUE_ORDSEQ_STARTED;
160
161         /*
162          * Prep proxy barrier request.
163          */
164         blkdev_dequeue_request(rq);
165         q->orig_bar_rq = rq;
166         rq = &q->bar_rq;
167         blk_rq_init(q, rq);
168         if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
169                 rq->cmd_flags |= REQ_RW;
170         if (q->ordered & QUEUE_ORDERED_FUA)
171                 rq->cmd_flags |= REQ_FUA;
172         init_request_from_bio(rq, q->orig_bar_rq->bio);
173         rq->end_io = bar_end_io;
174
175         /*
176          * Queue ordered sequence.  As we stack them at the head, we
177          * need to queue in reverse order.  Note that we rely on that
178          * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
179          * request gets inbetween ordered sequence. If this request is
180          * an empty barrier, we don't need to do a postflush ever since
181          * there will be no data written between the pre and post flush.
182          * Hence a single flush will suffice.
183          */
184         if ((q->ordered & QUEUE_ORDERED_POSTFLUSH) && !blk_empty_barrier(rq))
185                 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
186         else
187                 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
188
189         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
190
191         if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
192                 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
193                 rq = &q->pre_flush_rq;
194         } else
195                 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
196
197         if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
198                 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
199         else
200                 rq = NULL;
201
202         return rq;
203 }
204
205 int blk_do_ordered(struct request_queue *q, struct request **rqp)
206 {
207         struct request *rq = *rqp;
208         const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
209
210         if (!q->ordseq) {
211                 if (!is_barrier)
212                         return 1;
213
214                 if (q->next_ordered != QUEUE_ORDERED_NONE) {
215                         *rqp = start_ordered(q, rq);
216                         return 1;
217                 } else {
218                         /*
219                          * This can happen when the queue switches to
220                          * ORDERED_NONE while this request is on it.
221                          */
222                         blkdev_dequeue_request(rq);
223                         if (__blk_end_request(rq, -EOPNOTSUPP,
224                                               blk_rq_bytes(rq)))
225                                 BUG();
226                         *rqp = NULL;
227                         return 0;
228                 }
229         }
230
231         /*
232          * Ordered sequence in progress
233          */
234
235         /* Special requests are not subject to ordering rules. */
236         if (!blk_fs_request(rq) &&
237             rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
238                 return 1;
239
240         if (q->ordered & QUEUE_ORDERED_TAG) {
241                 /* Ordered by tag.  Blocking the next barrier is enough. */
242                 if (is_barrier && rq != &q->bar_rq)
243                         *rqp = NULL;
244         } else {
245                 /* Ordered by draining.  Wait for turn. */
246                 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
247                 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
248                         *rqp = NULL;
249         }
250
251         return 1;
252 }
253
254 static void bio_end_empty_barrier(struct bio *bio, int err)
255 {
256         if (err) {
257                 if (err == -EOPNOTSUPP)
258                         set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
259                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
260         }
261
262         complete(bio->bi_private);
263 }
264
265 /**
266  * blkdev_issue_flush - queue a flush
267  * @bdev:       blockdev to issue flush for
268  * @error_sector:       error sector
269  *
270  * Description:
271  *    Issue a flush for the block device in question. Caller can supply
272  *    room for storing the error offset in case of a flush error, if they
273  *    wish to.  Caller must run wait_for_completion() on its own.
274  */
275 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
276 {
277         DECLARE_COMPLETION_ONSTACK(wait);
278         struct request_queue *q;
279         struct bio *bio;
280         int ret;
281
282         if (bdev->bd_disk == NULL)
283                 return -ENXIO;
284
285         q = bdev_get_queue(bdev);
286         if (!q)
287                 return -ENXIO;
288
289         bio = bio_alloc(GFP_KERNEL, 0);
290         if (!bio)
291                 return -ENOMEM;
292
293         bio->bi_end_io = bio_end_empty_barrier;
294         bio->bi_private = &wait;
295         bio->bi_bdev = bdev;
296         submit_bio(1 << BIO_RW_BARRIER, bio);
297
298         wait_for_completion(&wait);
299
300         /*
301          * The driver must store the error location in ->bi_sector, if
302          * it supports it. For non-stacked drivers, this should be copied
303          * from rq->sector.
304          */
305         if (error_sector)
306                 *error_sector = bio->bi_sector;
307
308         ret = 0;
309         if (bio_flagged(bio, BIO_EOPNOTSUPP))
310                 ret = -EOPNOTSUPP;
311         else if (!bio_flagged(bio, BIO_UPTODATE))
312                 ret = -EIO;
313
314         bio_put(bio);
315         return ret;
316 }
317 EXPORT_SYMBOL(blkdev_issue_flush);
318
319 static void blkdev_discard_end_io(struct bio *bio, int err)
320 {
321         if (err) {
322                 if (err == -EOPNOTSUPP)
323                         set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
324                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
325         }
326
327         bio_put(bio);
328 }
329
330 /**
331  * blkdev_issue_discard - queue a discard
332  * @bdev:       blockdev to issue discard for
333  * @sector:     start sector
334  * @nr_sects:   number of sectors to discard
335  *
336  * Description:
337  *    Issue a discard request for the sectors in question. Does not wait.
338  */
339 int blkdev_issue_discard(struct block_device *bdev, sector_t sector,
340                          unsigned nr_sects)
341 {
342         struct request_queue *q;
343         struct bio *bio;
344         int ret = 0;
345
346         if (bdev->bd_disk == NULL)
347                 return -ENXIO;
348
349         q = bdev_get_queue(bdev);
350         if (!q)
351                 return -ENXIO;
352
353         if (!q->prepare_discard_fn)
354                 return -EOPNOTSUPP;
355
356         while (nr_sects && !ret) {
357                 bio = bio_alloc(GFP_KERNEL, 0);
358                 if (!bio)
359                         return -ENOMEM;
360
361                 bio->bi_end_io = blkdev_discard_end_io;
362                 bio->bi_bdev = bdev;
363
364                 bio->bi_sector = sector;
365
366                 if (nr_sects > q->max_hw_sectors) {
367                         bio->bi_size = q->max_hw_sectors << 9;
368                         nr_sects -= q->max_hw_sectors;
369                         sector += q->max_hw_sectors;
370                 } else {
371                         bio->bi_size = nr_sects << 9;
372                         nr_sects = 0;
373                 }
374                 bio_get(bio);
375                 submit_bio(WRITE_DISCARD, bio);
376
377                 /* Check if it failed immediately */
378                 if (bio_flagged(bio, BIO_EOPNOTSUPP))
379                         ret = -EOPNOTSUPP;
380                 else if (!bio_flagged(bio, BIO_UPTODATE))
381                         ret = -EIO;
382                 bio_put(bio);
383         }
384         return ret;
385 }
386 EXPORT_SYMBOL(blkdev_issue_discard);