2 * Functions related to barrier IO handling
4 #include <linux/kernel.h>
5 #include <linux/module.h>
7 #include <linux/blkdev.h>
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
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.
24 int blk_queue_ordered(struct request_queue *q, unsigned ordered,
25 prepare_flush_fn *prepare_flush_fn)
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__);
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);
45 q->next_ordered = ordered;
46 q->prepare_flush_fn = prepare_flush_fn;
50 EXPORT_SYMBOL(blk_queue_ordered);
53 * Cache flushing for ordered writes handling
55 unsigned blk_ordered_cur_seq(struct request_queue *q)
59 return 1 << ffz(q->ordseq);
62 unsigned blk_ordered_req_seq(struct request *rq)
64 struct request_queue *q = rq->q;
66 BUG_ON(q->ordseq == 0);
68 if (rq == &q->pre_flush_rq)
69 return QUEUE_ORDSEQ_PREFLUSH;
71 return QUEUE_ORDSEQ_BAR;
72 if (rq == &q->post_flush_rq)
73 return QUEUE_ORDSEQ_POSTFLUSH;
76 * !fs requests don't need to follow barrier ordering. Always
77 * put them at the front. This fixes the following deadlock.
79 * http://thread.gmane.org/gmane.linux.kernel/537473
81 if (!blk_fs_request(rq))
82 return QUEUE_ORDSEQ_DRAIN;
84 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
85 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
86 return QUEUE_ORDSEQ_DRAIN;
88 return QUEUE_ORDSEQ_DONE;
91 void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
95 if (error && !q->orderr)
98 BUG_ON(q->ordseq & seq);
101 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
105 * Okay, sequence complete.
110 if (__blk_end_request(rq, q->orderr, blk_rq_bytes(rq)))
114 static void pre_flush_end_io(struct request *rq, int error)
116 elv_completed_request(rq->q, rq);
117 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
120 static void bar_end_io(struct request *rq, int error)
122 elv_completed_request(rq->q, rq);
123 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
126 static void post_flush_end_io(struct request *rq, int error)
128 elv_completed_request(rq->q, rq);
129 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
132 static void queue_flush(struct request_queue *q, unsigned which)
135 rq_end_io_fn *end_io;
137 if (which == QUEUE_ORDERED_PREFLUSH) {
138 rq = &q->pre_flush_rq;
139 end_io = pre_flush_end_io;
141 rq = &q->post_flush_rq;
142 end_io = post_flush_end_io;
146 rq->cmd_flags = REQ_HARDBARRIER;
147 rq->rq_disk = q->bar_rq.rq_disk;
149 q->prepare_flush_fn(q, rq);
151 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
154 static inline struct request *start_ordered(struct request_queue *q,
158 q->ordered = q->next_ordered;
159 q->ordseq |= QUEUE_ORDSEQ_STARTED;
162 * Prep proxy barrier request.
164 blkdev_dequeue_request(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;
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.
184 if ((q->ordered & QUEUE_ORDERED_POSTFLUSH) && !blk_empty_barrier(rq))
185 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
187 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
189 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
191 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
192 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
193 rq = &q->pre_flush_rq;
195 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
197 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
198 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
205 int blk_do_ordered(struct request_queue *q, struct request **rqp)
207 struct request *rq = *rqp;
208 const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
214 if (q->next_ordered != QUEUE_ORDERED_NONE) {
215 *rqp = start_ordered(q, rq);
219 * This can happen when the queue switches to
220 * ORDERED_NONE while this request is on it.
222 blkdev_dequeue_request(rq);
223 if (__blk_end_request(rq, -EOPNOTSUPP,
232 * Ordered sequence in progress
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)
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)
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))
254 static void bio_end_empty_barrier(struct bio *bio, int err)
257 if (err == -EOPNOTSUPP)
258 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
259 clear_bit(BIO_UPTODATE, &bio->bi_flags);
262 complete(bio->bi_private);
266 * blkdev_issue_flush - queue a flush
267 * @bdev: blockdev to issue flush for
268 * @error_sector: error sector
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.
275 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
277 DECLARE_COMPLETION_ONSTACK(wait);
278 struct request_queue *q;
282 if (bdev->bd_disk == NULL)
285 q = bdev_get_queue(bdev);
289 bio = bio_alloc(GFP_KERNEL, 0);
293 bio->bi_end_io = bio_end_empty_barrier;
294 bio->bi_private = &wait;
296 submit_bio(WRITE_BARRIER, bio);
298 wait_for_completion(&wait);
301 * The driver must store the error location in ->bi_sector, if
302 * it supports it. For non-stacked drivers, this should be copied
306 *error_sector = bio->bi_sector;
309 if (bio_flagged(bio, BIO_EOPNOTSUPP))
311 else if (!bio_flagged(bio, BIO_UPTODATE))
317 EXPORT_SYMBOL(blkdev_issue_flush);
319 static void blkdev_discard_end_io(struct bio *bio, int err)
322 if (err == -EOPNOTSUPP)
323 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
324 clear_bit(BIO_UPTODATE, &bio->bi_flags);
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
337 * Issue a discard request for the sectors in question. Does not wait.
339 int blkdev_issue_discard(struct block_device *bdev, sector_t sector,
342 struct request_queue *q;
346 if (bdev->bd_disk == NULL)
349 q = bdev_get_queue(bdev);
353 if (!q->prepare_discard_fn)
356 while (nr_sects && !ret) {
357 bio = bio_alloc(GFP_KERNEL, 0);
361 bio->bi_end_io = blkdev_discard_end_io;
364 bio->bi_sector = sector;
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;
371 bio->bi_size = nr_sects << 9;
375 submit_bio(DISCARD_BARRIER, bio);
377 /* Check if it failed immediately */
378 if (bio_flagged(bio, BIO_EOPNOTSUPP))
380 else if (!bio_flagged(bio, BIO_UPTODATE))
386 EXPORT_SYMBOL(blkdev_issue_discard);