blk-flush.c revision 4fed947cb311e5aa51781d316cefca836352f6ce 
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 queue_flush(struct request_queue *q, struct request *rq,
78			rq_end_io_fn *end_io)
79{
80	blk_rq_init(q, rq);
81	rq->cmd_type = REQ_TYPE_FS;
82	rq->cmd_flags = REQ_FLUSH;
83	rq->rq_disk = q->orig_flush_rq->rq_disk;
84	rq->end_io = end_io;
85
86	elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
87}
88
89static struct request *queue_next_fseq(struct request_queue *q)
90{
91	struct request *orig_rq = q->orig_flush_rq;
92	struct request *rq = &q->flush_rq;
93
94	switch (blk_flush_cur_seq(q)) {
95	case QUEUE_FSEQ_PREFLUSH:
96		queue_flush(q, rq, pre_flush_end_io);
97		break;
98
99	case QUEUE_FSEQ_DATA:
100		/* initialize proxy request, inherit FLUSH/FUA and queue it */
101		blk_rq_init(q, rq);
102		init_request_from_bio(rq, orig_rq->bio);
103		rq->cmd_flags &= ~(REQ_FLUSH | REQ_FUA);
104		rq->cmd_flags |= orig_rq->cmd_flags & (REQ_FLUSH | REQ_FUA);
105		rq->end_io = flush_data_end_io;
106
107		elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
108		break;
109
110	case QUEUE_FSEQ_POSTFLUSH:
111		queue_flush(q, rq, post_flush_end_io);
112		break;
113
114	default:
115		BUG();
116	}
117	return rq;
118}
119
120struct request *blk_do_flush(struct request_queue *q, struct request *rq)
121{
122	unsigned int fflags = q->flush_flags; /* may change, cache it */
123	bool has_flush = fflags & REQ_FLUSH, has_fua = fflags & REQ_FUA;
124	bool do_preflush = has_flush && (rq->cmd_flags & REQ_FLUSH);
125	bool do_postflush = has_flush && !has_fua && (rq->cmd_flags & REQ_FUA);
126	unsigned skip = 0;
127
128	/*
129	 * Special case.  If there's data but flush is not necessary,
130	 * the request can be issued directly.
131	 *
132	 * Flush w/o data should be able to be issued directly too but
133	 * currently some drivers assume that rq->bio contains
134	 * non-zero data if it isn't NULL and empty FLUSH requests
135	 * getting here usually have bio's without data.
136	 */
137	if (blk_rq_sectors(rq) && !do_preflush && !do_postflush) {
138		rq->cmd_flags &= ~REQ_FLUSH;
139		if (!has_fua)
140			rq->cmd_flags &= ~REQ_FUA;
141		return rq;
142	}
143
144	/*
145	 * Sequenced flushes can't be processed in parallel.  If
146	 * another one is already in progress, queue for later
147	 * processing.
148	 */
149	if (q->flush_seq) {
150		list_move_tail(&rq->queuelist, &q->pending_flushes);
151		return NULL;
152	}
153
154	/*
155	 * Start a new flush sequence
156	 */
157	q->flush_err = 0;
158	q->flush_seq |= QUEUE_FSEQ_STARTED;
159
160	/* adjust FLUSH/FUA of the original request and stash it away */
161	rq->cmd_flags &= ~REQ_FLUSH;
162	if (!has_fua)
163		rq->cmd_flags &= ~REQ_FUA;
164	blk_dequeue_request(rq);
165	q->orig_flush_rq = rq;
166
167	/* skip unneded sequences and return the first one */
168	if (!do_preflush)
169		skip |= QUEUE_FSEQ_PREFLUSH;
170	if (!blk_rq_sectors(rq))
171		skip |= QUEUE_FSEQ_DATA;
172	if (!do_postflush)
173		skip |= QUEUE_FSEQ_POSTFLUSH;
174	return blk_flush_complete_seq(q, skip, 0);
175}
176
177static void bio_end_empty_barrier(struct bio *bio, int err)
178{
179	if (err) {
180		if (err == -EOPNOTSUPP)
181			set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
182		clear_bit(BIO_UPTODATE, &bio->bi_flags);
183	}
184	if (bio->bi_private)
185		complete(bio->bi_private);
186	bio_put(bio);
187}
188
189/**
190 * blkdev_issue_flush - queue a flush
191 * @bdev:	blockdev to issue flush for
192 * @gfp_mask:	memory allocation flags (for bio_alloc)
193 * @error_sector:	error sector
194 * @flags:	BLKDEV_IFL_* flags to control behaviour
195 *
196 * Description:
197 *    Issue a flush for the block device in question. Caller can supply
198 *    room for storing the error offset in case of a flush error, if they
199 *    wish to. If WAIT flag is not passed then caller may check only what
200 *    request was pushed in some internal queue for later handling.
201 */
202int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
203		sector_t *error_sector, unsigned long flags)
204{
205	DECLARE_COMPLETION_ONSTACK(wait);
206	struct request_queue *q;
207	struct bio *bio;
208	int ret = 0;
209
210	if (bdev->bd_disk == NULL)
211		return -ENXIO;
212
213	q = bdev_get_queue(bdev);
214	if (!q)
215		return -ENXIO;
216
217	/*
218	 * some block devices may not have their queue correctly set up here
219	 * (e.g. loop device without a backing file) and so issuing a flush
220	 * here will panic. Ensure there is a request function before issuing
221	 * the barrier.
222	 */
223	if (!q->make_request_fn)
224		return -ENXIO;
225
226	bio = bio_alloc(gfp_mask, 0);
227	bio->bi_end_io = bio_end_empty_barrier;
228	bio->bi_bdev = bdev;
229	if (test_bit(BLKDEV_WAIT, &flags))
230		bio->bi_private = &wait;
231
232	bio_get(bio);
233	submit_bio(WRITE_BARRIER, bio);
234	if (test_bit(BLKDEV_WAIT, &flags)) {
235		wait_for_completion(&wait);
236		/*
237		 * The driver must store the error location in ->bi_sector, if
238		 * it supports it. For non-stacked drivers, this should be
239		 * copied from blk_rq_pos(rq).
240		 */
241		if (error_sector)
242			*error_sector = bio->bi_sector;
243	}
244
245	if (bio_flagged(bio, BIO_EOPNOTSUPP))
246		ret = -EOPNOTSUPP;
247	else if (!bio_flagged(bio, BIO_UPTODATE))
248		ret = -EIO;
249
250	bio_put(bio);
251	return ret;
252}
253EXPORT_SYMBOL(blkdev_issue_flush);
254