blob: 5f74fec327d53e35651d4616098a99e6e146630b [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{
27 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
28 prepare_flush_fn == NULL) {
29 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
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
51EXPORT_SYMBOL(blk_queue_ordered);
52
53/*
54 * Cache flushing for ordered writes handling
55 */
56inline unsigned blk_ordered_cur_seq(struct request_queue *q)
57{
58 if (!q->ordseq)
59 return 0;
60 return 1 << ffz(q->ordseq);
61}
62
63unsigned blk_ordered_req_seq(struct request *rq)
64{
65 struct request_queue *q = rq->q;
66
67 BUG_ON(q->ordseq == 0);
68
69 if (rq == &q->pre_flush_rq)
70 return QUEUE_ORDSEQ_PREFLUSH;
71 if (rq == &q->bar_rq)
72 return QUEUE_ORDSEQ_BAR;
73 if (rq == &q->post_flush_rq)
74 return QUEUE_ORDSEQ_POSTFLUSH;
75
76 /*
77 * !fs requests don't need to follow barrier ordering. Always
78 * put them at the front. This fixes the following deadlock.
79 *
80 * http://thread.gmane.org/gmane.linux.kernel/537473
81 */
82 if (!blk_fs_request(rq))
83 return QUEUE_ORDSEQ_DRAIN;
84
85 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
86 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
87 return QUEUE_ORDSEQ_DRAIN;
88 else
89 return QUEUE_ORDSEQ_DONE;
90}
91
92void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
93{
94 struct request *rq;
95
96 if (error && !q->orderr)
97 q->orderr = error;
98
99 BUG_ON(q->ordseq & seq);
100 q->ordseq |= seq;
101
102 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
103 return;
104
105 /*
106 * Okay, sequence complete.
107 */
108 q->ordseq = 0;
109 rq = q->orig_bar_rq;
110
111 if (__blk_end_request(rq, q->orderr, blk_rq_bytes(rq)))
112 BUG();
113}
114
115static void pre_flush_end_io(struct request *rq, int error)
116{
117 elv_completed_request(rq->q, rq);
118 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
119}
120
121static void bar_end_io(struct request *rq, int error)
122{
123 elv_completed_request(rq->q, rq);
124 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
125}
126
127static void post_flush_end_io(struct request *rq, int error)
128{
129 elv_completed_request(rq->q, rq);
130 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
131}
132
133static void queue_flush(struct request_queue *q, unsigned which)
134{
135 struct request *rq;
136 rq_end_io_fn *end_io;
137
138 if (which == QUEUE_ORDERED_PREFLUSH) {
139 rq = &q->pre_flush_rq;
140 end_io = pre_flush_end_io;
141 } else {
142 rq = &q->post_flush_rq;
143 end_io = post_flush_end_io;
144 }
145
146 rq->cmd_flags = REQ_HARDBARRIER;
147 rq_init(q, rq);
148 rq->elevator_private = NULL;
149 rq->elevator_private2 = NULL;
150 rq->rq_disk = q->bar_rq.rq_disk;
151 rq->end_io = end_io;
152 q->prepare_flush_fn(q, rq);
153
154 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
155}
156
157static inline struct request *start_ordered(struct request_queue *q,
158 struct request *rq)
159{
160 q->orderr = 0;
161 q->ordered = q->next_ordered;
162 q->ordseq |= QUEUE_ORDSEQ_STARTED;
163
164 /*
165 * Prep proxy barrier request.
166 */
167 blkdev_dequeue_request(rq);
168 q->orig_bar_rq = rq;
169 rq = &q->bar_rq;
170 rq->cmd_flags = 0;
171 rq_init(q, rq);
172 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
173 rq->cmd_flags |= REQ_RW;
174 if (q->ordered & QUEUE_ORDERED_FUA)
175 rq->cmd_flags |= REQ_FUA;
176 rq->elevator_private = NULL;
177 rq->elevator_private2 = NULL;
178 init_request_from_bio(rq, q->orig_bar_rq->bio);
179 rq->end_io = bar_end_io;
180
181 /*
182 * Queue ordered sequence. As we stack them at the head, we
183 * need to queue in reverse order. Note that we rely on that
184 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
185 * request gets inbetween ordered sequence. If this request is
186 * an empty barrier, we don't need to do a postflush ever since
187 * there will be no data written between the pre and post flush.
188 * Hence a single flush will suffice.
189 */
190 if ((q->ordered & QUEUE_ORDERED_POSTFLUSH) && !blk_empty_barrier(rq))
191 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
192 else
193 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
194
195 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
196
197 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
198 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
199 rq = &q->pre_flush_rq;
200 } else
201 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
202
203 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
204 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
205 else
206 rq = NULL;
207
208 return rq;
209}
210
211int blk_do_ordered(struct request_queue *q, struct request **rqp)
212{
213 struct request *rq = *rqp;
214 const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
215
216 if (!q->ordseq) {
217 if (!is_barrier)
218 return 1;
219
220 if (q->next_ordered != QUEUE_ORDERED_NONE) {
221 *rqp = start_ordered(q, rq);
222 return 1;
223 } else {
224 /*
225 * This can happen when the queue switches to
226 * ORDERED_NONE while this request is on it.
227 */
228 blkdev_dequeue_request(rq);
229 if (__blk_end_request(rq, -EOPNOTSUPP,
230 blk_rq_bytes(rq)))
231 BUG();
232 *rqp = NULL;
233 return 0;
234 }
235 }
236
237 /*
238 * Ordered sequence in progress
239 */
240
241 /* Special requests are not subject to ordering rules. */
242 if (!blk_fs_request(rq) &&
243 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
244 return 1;
245
246 if (q->ordered & QUEUE_ORDERED_TAG) {
247 /* Ordered by tag. Blocking the next barrier is enough. */
248 if (is_barrier && rq != &q->bar_rq)
249 *rqp = NULL;
250 } else {
251 /* Ordered by draining. Wait for turn. */
252 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
253 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
254 *rqp = NULL;
255 }
256
257 return 1;
258}
259
260static void bio_end_empty_barrier(struct bio *bio, int err)
261{
262 if (err)
263 clear_bit(BIO_UPTODATE, &bio->bi_flags);
264
265 complete(bio->bi_private);
266}
267
268/**
269 * blkdev_issue_flush - queue a flush
270 * @bdev: blockdev to issue flush for
271 * @error_sector: error sector
272 *
273 * Description:
274 * Issue a flush for the block device in question. Caller can supply
275 * room for storing the error offset in case of a flush error, if they
276 * wish to. Caller must run wait_for_completion() on its own.
277 */
278int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
279{
280 DECLARE_COMPLETION_ONSTACK(wait);
281 struct request_queue *q;
282 struct bio *bio;
283 int ret;
284
285 if (bdev->bd_disk == NULL)
286 return -ENXIO;
287
288 q = bdev_get_queue(bdev);
289 if (!q)
290 return -ENXIO;
291
292 bio = bio_alloc(GFP_KERNEL, 0);
293 if (!bio)
294 return -ENOMEM;
295
296 bio->bi_end_io = bio_end_empty_barrier;
297 bio->bi_private = &wait;
298 bio->bi_bdev = bdev;
299 submit_bio(1 << BIO_RW_BARRIER, bio);
300
301 wait_for_completion(&wait);
302
303 /*
304 * The driver must store the error location in ->bi_sector, if
305 * it supports it. For non-stacked drivers, this should be copied
306 * from rq->sector.
307 */
308 if (error_sector)
309 *error_sector = bio->bi_sector;
310
311 ret = 0;
312 if (!bio_flagged(bio, BIO_UPTODATE))
313 ret = -EIO;
314
315 bio_put(bio);
316 return ret;
317}
318
319EXPORT_SYMBOL(blkdev_issue_flush);