blk-flush.c revision 337238be1bf52e1242f940fc6fe83fb395e55057 
1/*
2 * Functions to sequence FLUSH and FUA writes.
3 */
4#include <linux/kernel.h>
5#include <linux/module.h>
6#include <linux/bio.h>
7#include <linux/blkdev.h>
8#include <linux/gfp.h>
9
10#include "blk.h"
11
12/* FLUSH/FUA sequences */
13enum {
14	QUEUE_FSEQ_STARTED	= (1 << 0), /* flushing in progress */
15	QUEUE_FSEQ_PREFLUSH	= (1 << 1), /* pre-flushing in progress */
16	QUEUE_FSEQ_DATA		= (1 << 2), /* data write in progress */
17	QUEUE_FSEQ_POSTFLUSH	= (1 << 3), /* post-flushing in progress */
18	QUEUE_FSEQ_DONE		= (1 << 4),
19};
20
21static struct request *queue_next_fseq(struct request_queue *q);
22
23unsigned blk_flush_cur_seq(struct request_queue *q)
24{
25	if (!q->flush_seq)
26		return 0;
27	return 1 << ffz(q->flush_seq);
28}
29
30static struct request *blk_flush_complete_seq(struct request_queue *q,
31					      unsigned seq, int error)
32{
33	struct request *next_rq = NULL;
34
35	if (error && !q->flush_err)
36		q->flush_err = error;
37
38	BUG_ON(q->flush_seq & seq);
39	q->flush_seq |= seq;
40
41	if (blk_flush_cur_seq(q) != QUEUE_FSEQ_DONE) {
42		/* not complete yet, queue the next flush sequence */
43		next_rq = queue_next_fseq(q);
44	} else {
45		/* complete this flush request */
46		__blk_end_request_all(q->orig_flush_rq, q->flush_err);
47		q->orig_flush_rq = NULL;
48		q->flush_seq = 0;
49
50		/* dispatch the next flush if there's one */
51		if (!list_empty(&q->pending_flushes)) {
52			next_rq = list_entry_rq(q->pending_flushes.next);
53			list_move(&next_rq->queuelist, &q->queue_head);
54		}
55	}
56	return next_rq;
57}
58
59static void pre_flush_end_io(struct request *rq, int error)
60{
61	elv_completed_request(rq->q, rq);
62	blk_flush_complete_seq(rq->q, QUEUE_FSEQ_PREFLUSH, error);
63}
64
65static void flush_data_end_io(struct request *rq, int error)
66{
67	elv_completed_request(rq->q, rq);
68	blk_flush_complete_seq(rq->q, QUEUE_FSEQ_DATA, error);
69}
70
71static void post_flush_end_io(struct request *rq, int error)
72{
73	elv_completed_request(rq->q, rq);
74	blk_flush_complete_seq(rq->q, QUEUE_FSEQ_POSTFLUSH, error);
75}
76
77static void init_flush_request(struct request *rq, struct gendisk *disk)
78{
79	rq->cmd_type = REQ_TYPE_FS;
80	rq->cmd_flags = WRITE_FLUSH;
81	rq->rq_disk = disk;
82}
83
84static struct request *queue_next_fseq(struct request_queue *q)
85{
86	struct request *orig_rq = q->orig_flush_rq;
87	struct request *rq = &q->flush_rq;
88
89	blk_rq_init(q, rq);
90
91	switch (blk_flush_cur_seq(q)) {
92	case QUEUE_FSEQ_PREFLUSH:
93		init_flush_request(rq, orig_rq->rq_disk);
94		rq->end_io = pre_flush_end_io;
95		break;
96	case QUEUE_FSEQ_DATA:
97		init_request_from_bio(rq, orig_rq->bio);
98		rq->cmd_flags &= ~(REQ_FLUSH | REQ_FUA);
99		rq->cmd_flags |= orig_rq->cmd_flags & (REQ_FLUSH | REQ_FUA);
100		rq->end_io = flush_data_end_io;
101		break;
102	case QUEUE_FSEQ_POSTFLUSH:
103		init_flush_request(rq, orig_rq->rq_disk);
104		rq->end_io = post_flush_end_io;
105		break;
106	default:
107		BUG();
108	}
109
110	elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
111	return rq;
112}
113
114struct request *blk_do_flush(struct request_queue *q, struct request *rq)
115{
116	unsigned int fflags = q->flush_flags; /* may change, cache it */
117	bool has_flush = fflags & REQ_FLUSH, has_fua = fflags & REQ_FUA;
118	bool do_preflush = has_flush && (rq->cmd_flags & REQ_FLUSH);
119	bool do_postflush = has_flush && !has_fua && (rq->cmd_flags & REQ_FUA);
120	unsigned skip = 0;
121
122	/*
123	 * Special case.  If there's data but flush is not necessary,
124	 * the request can be issued directly.
125	 *
126	 * Flush w/o data should be able to be issued directly too but
127	 * currently some drivers assume that rq->bio contains
128	 * non-zero data if it isn't NULL and empty FLUSH requests
129	 * getting here usually have bio's without data.
130	 */
131	if (blk_rq_sectors(rq) && !do_preflush && !do_postflush) {
132		rq->cmd_flags &= ~REQ_FLUSH;
133		if (!has_fua)
134			rq->cmd_flags &= ~REQ_FUA;
135		return rq;
136	}
137
138	/*
139	 * Sequenced flushes can't be processed in parallel.  If
140	 * another one is already in progress, queue for later
141	 * processing.
142	 */
143	if (q->flush_seq) {
144		list_move_tail(&rq->queuelist, &q->pending_flushes);
145		return NULL;
146	}
147
148	/*
149	 * Start a new flush sequence
150	 */
151	q->flush_err = 0;
152	q->flush_seq |= QUEUE_FSEQ_STARTED;
153
154	/* adjust FLUSH/FUA of the original request and stash it away */
155	rq->cmd_flags &= ~REQ_FLUSH;
156	if (!has_fua)
157		rq->cmd_flags &= ~REQ_FUA;
158	blk_dequeue_request(rq);
159	q->orig_flush_rq = rq;
160
161	/* skip unneded sequences and return the first one */
162	if (!do_preflush)
163		skip |= QUEUE_FSEQ_PREFLUSH;
164	if (!blk_rq_sectors(rq))
165		skip |= QUEUE_FSEQ_DATA;
166	if (!do_postflush)
167		skip |= QUEUE_FSEQ_POSTFLUSH;
168	return blk_flush_complete_seq(q, skip, 0);
169}
170
171static void bio_end_empty_barrier(struct bio *bio, int err)
172{
173	if (err) {
174		if (err == -EOPNOTSUPP)
175			set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
176		clear_bit(BIO_UPTODATE, &bio->bi_flags);
177	}
178	if (bio->bi_private)
179		complete(bio->bi_private);
180	bio_put(bio);
181}
182
183/**
184 * blkdev_issue_flush - queue a flush
185 * @bdev:	blockdev to issue flush for
186 * @gfp_mask:	memory allocation flags (for bio_alloc)
187 * @error_sector:	error sector
188 * @flags:	BLKDEV_IFL_* flags to control behaviour
189 *
190 * Description:
191 *    Issue a flush for the block device in question. Caller can supply
192 *    room for storing the error offset in case of a flush error, if they
193 *    wish to. If WAIT flag is not passed then caller may check only what
194 *    request was pushed in some internal queue for later handling.
195 */
196int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
197		sector_t *error_sector, unsigned long flags)
198{
199	DECLARE_COMPLETION_ONSTACK(wait);
200	struct request_queue *q;
201	struct bio *bio;
202	int ret = 0;
203
204	if (bdev->bd_disk == NULL)
205		return -ENXIO;
206
207	q = bdev_get_queue(bdev);
208	if (!q)
209		return -ENXIO;
210
211	/*
212	 * some block devices may not have their queue correctly set up here
213	 * (e.g. loop device without a backing file) and so issuing a flush
214	 * here will panic. Ensure there is a request function before issuing
215	 * the barrier.
216	 */
217	if (!q->make_request_fn)
218		return -ENXIO;
219
220	bio = bio_alloc(gfp_mask, 0);
221	bio->bi_end_io = bio_end_empty_barrier;
222	bio->bi_bdev = bdev;
223	if (test_bit(BLKDEV_WAIT, &flags))
224		bio->bi_private = &wait;
225
226	bio_get(bio);
227	submit_bio(WRITE_BARRIER, bio);
228	if (test_bit(BLKDEV_WAIT, &flags)) {
229		wait_for_completion(&wait);
230		/*
231		 * The driver must store the error location in ->bi_sector, if
232		 * it supports it. For non-stacked drivers, this should be
233		 * copied from blk_rq_pos(rq).
234		 */
235		if (error_sector)
236			*error_sector = bio->bi_sector;
237	}
238
239	if (bio_flagged(bio, BIO_EOPNOTSUPP))
240		ret = -EOPNOTSUPP;
241	else if (!bio_flagged(bio, BIO_UPTODATE))
242		ret = -EIO;
243
244	bio_put(bio);
245	return ret;
246}
247EXPORT_SYMBOL(blkdev_issue_flush);
248