2 * SPDX-License-Identifier: Beerware
4 * ----------------------------------------------------------------------------
5 * "THE BEER-WARE LICENSE" (Revision 42):
6 * <phk@FreeBSD.ORG> wrote this file. As long as you retain this notice you
7 * can do whatever you want with this stuff. If we meet some day, and you think
8 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
9 * ----------------------------------------------------------------------------
11 * The bioq_disksort() (and the specification of the bioq API)
12 * have been written by Luigi Rizzo and Fabio Checconi under the same
16 #include <sys/cdefs.h>
17 __FBSDID("$FreeBSD$");
21 #include <sys/param.h>
22 #include <sys/systm.h>
26 #include <sys/sysctl.h>
27 #include <geom/geom_disk.h>
29 static int bioq_batchsize = 0;
30 SYSCTL_INT(_debug, OID_AUTO, bioq_batchsize, CTLFLAG_RW,
31 &bioq_batchsize, 0, "BIOQ batch size");
34 * Disk error is the preface to plaintive error messages
35 * about failing disk transfers. It prints messages of the form
36 * "hp0g: BLABLABLA cmd=read fsbn 12345 of 12344-12347"
37 * blkdone should be -1 if the position of the error is unknown.
38 * The message is printed with printf.
41 disk_err(struct bio *bp, const char *what, int blkdone, int nl)
45 if (bp->bio_dev != NULL)
46 printf("%s: %s ", devtoname(bp->bio_dev), what);
47 else if (bp->bio_disk != NULL)
49 bp->bio_disk->d_name, bp->bio_disk->d_unit, what);
51 printf("disk??: %s ", what);
53 case BIO_READ: printf("cmd=read "); break;
54 case BIO_WRITE: printf("cmd=write "); break;
55 case BIO_DELETE: printf("cmd=delete "); break;
56 case BIO_GETATTR: printf("cmd=getattr "); break;
57 case BIO_FLUSH: printf("cmd=flush "); break;
58 default: printf("cmd=%x ", bp->bio_cmd); break;
61 if (bp->bio_bcount <= DEV_BSIZE) {
62 printf("fsbn %jd%s", (intmax_t)sn, nl ? "\n" : "");
67 printf("fsbn %jd of ", (intmax_t)sn);
69 printf("%jd-%jd", (intmax_t)bp->bio_pblkno,
70 (intmax_t)(bp->bio_pblkno + (bp->bio_bcount - 1) / DEV_BSIZE));
76 * BIO queue implementation
78 * Please read carefully the description below before making any change
79 * to the code, or you might change the behaviour of the data structure
80 * in undesirable ways.
82 * A bioq stores disk I/O request (bio), normally sorted according to
83 * the distance of the requested position (bio->bio_offset) from the
84 * current head position (bioq->last_offset) in the scan direction, i.e.
86 * (uoff_t)(bio_offset - last_offset)
88 * Note that the cast to unsigned (uoff_t) is fundamental to insure
89 * that the distance is computed in the scan direction.
91 * The main methods for manipulating the bioq are:
93 * bioq_disksort() performs an ordered insertion;
95 * bioq_first() return the head of the queue, without removing;
97 * bioq_takefirst() return and remove the head of the queue,
98 * updating the 'current head position' as
99 * bioq->last_offset = bio->bio_offset + bio->bio_length;
101 * When updating the 'current head position', we assume that the result of
102 * bioq_takefirst() is dispatched to the device, so bioq->last_offset
103 * represents the head position once the request is complete.
105 * If the bioq is manipulated using only the above calls, it starts
106 * with a sorted sequence of requests with bio_offset >= last_offset,
107 * possibly followed by another sorted sequence of requests with
108 * 0 <= bio_offset < bioq->last_offset
110 * NOTE: historical behaviour was to ignore bio->bio_length in the
111 * update, but its use tracks the head position in a better way.
112 * Historical behaviour was also to update the head position when
113 * the request under service is complete, rather than when the
114 * request is extracted from the queue. However, the current API
115 * has no method to update the head position; secondly, once
116 * a request has been submitted to the disk, we have no idea of
117 * the actual head position, so the final one is our best guess.
119 * --- Direct queue manipulation ---
121 * A bioq uses an underlying TAILQ to store requests, so we also
122 * export methods to manipulate the TAILQ, in particular:
124 * bioq_insert_tail() insert an entry at the end.
125 * It also creates a 'barrier' so all subsequent
126 * insertions through bioq_disksort() will end up
129 * bioq_insert_head() insert an entry at the head, update
130 * bioq->last_offset = bio->bio_offset so that
131 * all subsequent insertions through bioq_disksort()
132 * will end up after this entry;
134 * bioq_remove() remove a generic element from the queue, act as
135 * bioq_takefirst() if invoked on the head of the queue.
137 * The semantic of these methods is the same as the operations
138 * on the underlying TAILQ, but with additional guarantees on
139 * subsequent bioq_disksort() calls. E.g. bioq_insert_tail()
140 * can be useful for making sure that all previous ops are flushed
141 * to disk before continuing.
143 * Updating bioq->last_offset on a bioq_insert_head() guarantees
144 * that the bio inserted with the last bioq_insert_head() will stay
145 * at the head of the queue even after subsequent bioq_disksort().
147 * Note that when the direct queue manipulation functions are used,
148 * the queue may contain multiple inversion points (i.e. more than
149 * two sorted sequences of requests).
154 bioq_init(struct bio_queue_head *head)
157 TAILQ_INIT(&head->queue);
158 head->last_offset = 0;
159 head->insert_point = NULL;
165 bioq_remove(struct bio_queue_head *head, struct bio *bp)
168 if (head->insert_point == NULL) {
169 if (bp == TAILQ_FIRST(&head->queue))
170 head->last_offset = bp->bio_offset + bp->bio_length;
171 } else if (bp == head->insert_point)
172 head->insert_point = NULL;
174 TAILQ_REMOVE(&head->queue, bp, bio_queue);
179 bioq_flush(struct bio_queue_head *head, struct devstat *stp, int error)
183 while ((bp = bioq_takefirst(head)) != NULL)
184 biofinish(bp, stp, error);
188 bioq_insert_head(struct bio_queue_head *head, struct bio *bp)
191 if (head->insert_point == NULL)
192 head->last_offset = bp->bio_offset;
193 TAILQ_INSERT_HEAD(&head->queue, bp, bio_queue);
199 bioq_insert_tail(struct bio_queue_head *head, struct bio *bp)
202 TAILQ_INSERT_TAIL(&head->queue, bp, bio_queue);
204 head->insert_point = bp;
205 head->last_offset = bp->bio_offset;
209 bioq_first(struct bio_queue_head *head)
212 return (TAILQ_FIRST(&head->queue));
216 bioq_takefirst(struct bio_queue_head *head)
220 bp = TAILQ_FIRST(&head->queue);
222 bioq_remove(head, bp);
227 * Compute the sorting key. The cast to unsigned is
228 * fundamental for correctness, see the description
229 * near the beginning of the file.
232 bioq_bio_key(struct bio_queue_head *head, struct bio *bp)
235 return ((uoff_t)(bp->bio_offset - head->last_offset));
239 * Seek sort for disks.
241 * Sort all requests in a single queue while keeping
242 * track of the current position of the disk with last_offset.
243 * See above for details.
246 bioq_disksort(struct bio_queue_head *head, struct bio *bp)
248 struct bio *cur, *prev;
251 if ((bp->bio_flags & BIO_ORDERED) != 0) {
253 * Ordered transactions can only be dispatched
254 * after any currently queued transactions. They
255 * also have barrier semantics - no transactions
256 * queued in the future can pass them.
258 bioq_insert_tail(head, bp);
263 * We should only sort requests of types that have concept of offset.
264 * Other types, such as BIO_FLUSH or BIO_ZONE, may imply some degree
265 * of ordering even if strict ordering is not requested explicitly.
267 if (bp->bio_cmd != BIO_READ && bp->bio_cmd != BIO_WRITE &&
268 bp->bio_cmd != BIO_DELETE) {
269 bioq_insert_tail(head, bp);
273 if (bioq_batchsize > 0 && head->batched > bioq_batchsize) {
274 bioq_insert_tail(head, bp);
279 key = bioq_bio_key(head, bp);
280 cur = TAILQ_FIRST(&head->queue);
282 if (head->insert_point) {
283 prev = head->insert_point;
284 cur = TAILQ_NEXT(head->insert_point, bio_queue);
287 while (cur != NULL && key >= bioq_bio_key(head, cur)) {
289 cur = TAILQ_NEXT(cur, bio_queue);
293 TAILQ_INSERT_HEAD(&head->queue, bp, bio_queue);
295 TAILQ_INSERT_AFTER(&head->queue, prev, bp, bio_queue);