d/freebsd-skq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

When read requests are sent from a filesystem running above g_journal,

Browse Source 
the g_journal level needs to check whether it is holding a newer
copy of the block than that which exists on the disk. If so, it
needs to return its copy. If not, it should pass the request down
to the disk to fulfill. It currently considers six queues:

0) delayed queue,
1) unsent (current queue),
2) in-flight to the journal (flush queue),
3) active journal (active queue),
4) inactive journal (inactive queue), and
5) inflight to the disk (copy queue).

Checking on two of these queues is unnecessary:

0) The delayed requests should not be used for reads because they
   have not yet been entered into the journal, so their value should
   reflect the disk contents, not the future contents that are not
   yet committed.

2) Because all the bio's in the flush queue are also found on the
   active queue, there is no need to inspect the flush queue for
   reads since they will be found when searching the active queue.

Submitted by: Dr. Andreas Longwitz <longwitz@incore.de>
Discussed with: kib
MFC after: 1 week
This commit is contained in:
mckusick 2017-08-13 18:09:22 +00:00
parent 9732028639
commit 12015c2c0a
1 changed files with 5 additions and 53 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
sys/geom/journal/g_journal.c
View File

@ -1514,49 +1514,10 @@ g_journal_read_find(struct bio *head, int sorted, struct bio *pbp, off_t ostart,
}
/*
* Try to find requested data in cache.
*/
static struct bio *
g_journal_read_queue_find(struct bio_queue *head, struct bio *pbp, off_t ostart,
off_t oend)
{
off_t cstart, cend;
struct bio *bp;
TAILQ_FOREACH(bp, head, bio_queue) {
cstart = MAX(ostart, bp->bio_offset);
cend = MIN(oend, bp->bio_offset + bp->bio_length);
if (cend <= ostart)
continue;
else if (cstart >= oend)
continue;
KASSERT(bp->bio_data != NULL,
("%s: bio_data == NULL", __func__));
GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend,
bp);
bcopy(bp->bio_data + cstart - bp->bio_offset,
pbp->bio_data + cstart - pbp->bio_offset, cend - cstart);
pbp->bio_completed += cend - cstart;
if (pbp->bio_completed == pbp->bio_length) {
/*
* Cool, the whole request was in cache, deliver happy
* message.
*/
g_io_deliver(pbp, 0);
return (pbp);
}
break;
}
return (bp);
}
/*
* This function is used for colecting data on read.
* This function is used for collecting data on read.
* The complexity is because parts of the data can be stored in four different
* places:
* - in delayed requests
* - in memory - the data not yet send to the active journal provider
* - in requests which are going to be sent to the active journal
* - in the active journal
* - in the inactive journal
* - in the data provider
@ -1574,20 +1535,14 @@ g_journal_read(struct g_journal_softc *sc, struct bio *pbp, off_t ostart,
cstart = cend = -1;
bp = NULL;
head = NULL;
for (i = 0; i <= 5; i++) {
for (i = 1; i <= 5; i++) {
switch (i) {
case 0: /* Delayed requests. */
head = NULL;
sorted = 0;
break;
case 1: /* Not-yet-send data. */
head = sc->sc_current_queue;
sorted = 1;
break;
case 2: /* In-flight to the active journal. */
head = sc->sc_flush_queue;
sorted = 0;
break;
case 2: /* Skip flush queue as they are also in active queue */
continue;
case 3: /* Active journal. */
head = sc->sc_active.jj_queue;
sorted = 1;
@ -1606,10 +1561,7 @@ g_journal_read(struct g_journal_softc *sc, struct bio *pbp, off_t ostart,
default:
panic("gjournal %s: i=%d", __func__, i);
}
if (i == 0)
bp = g_journal_read_queue_find(&sc->sc_delayed_queue.queue, pbp, ostart, oend);
else
bp = g_journal_read_find(head, sorted, pbp, ostart, oend);
bp = g_journal_read_find(head, sorted, pbp, ostart, oend);
if (bp == pbp) { /* Got the whole request. */
GJ_DEBUG(2, "Got the whole request from %u.", i);
return;