svn commit: r218052 - projects/graid/head/sys/geom/raid
Warner Losh
imp at FreeBSD.org
Sat Jan 29 00:21:40 UTC 2011
Author: imp
Date: Sat Jan 29 00:21:40 2011
New Revision: 218052
URL: http://svn.freebsd.org/changeset/base/218052
Log:
Stop leaking BIOs on rebuild. We used to leak 20k of them on a 20GB
rebuild. Now we leak none.
Also, if g_clone_bp() fails during a rebuild, just stop rebuilding for
a while. Maybe the next step here would be actually reusing the BPs
since we don't really need to clone them for a rebuild (we do for
resync however).
# But we do leak them when we do actual I/O to the disk still :(
Modified:
projects/graid/head/sys/geom/raid/tr_raid1.c
Modified: projects/graid/head/sys/geom/raid/tr_raid1.c
==============================================================================
--- projects/graid/head/sys/geom/raid/tr_raid1.c Fri Jan 28 23:44:57 2011 (r218051)
+++ projects/graid/head/sys/geom/raid/tr_raid1.c Sat Jan 29 00:21:40 2011 (r218052)
@@ -179,9 +179,6 @@ g_raid_tr_raid1_rebuild_some(struct g_ra
g_raid_tr_raid1_rebuild_abort(tr, sd->sd_volume);
return;
}
- trs->trso_flags |= TR_RAID1_F_DOING_SOME;
- trs->trso_recover_slabs = SD_REBUILD_CLUSTER_IDLE;
- trs->trso_fair_io = SD_REBUILD_FAIR_IO;
bp = &trs->trso_bio;
memset(bp, 0, sizeof(*bp));
bp->bio_offset = sd->sd_rebuild_pos;
@@ -190,8 +187,13 @@ g_raid_tr_raid1_rebuild_some(struct g_ra
bp->bio_data = trs->trso_buffer;
bp->bio_cmd = BIO_READ;
bp2 = g_clone_bio(bp);
+ if (bp2 == NULL) /* We'll try again later */
+ return;
bp2->bio_cflags = G_RAID_BIO_FLAG_SYNC;
bp2->bio_caller1 = good_sd;
+ trs->trso_recover_slabs = SD_REBUILD_CLUSTER_IDLE;
+ trs->trso_fair_io = SD_REBUILD_FAIR_IO;
+ trs->trso_flags |= TR_RAID1_F_DOING_SOME;
g_raid_lock_range(sd->sd_volume, /* Lock callback starts I/O */
bp2->bio_offset, bp2->bio_length, bp2);
}
@@ -609,11 +611,23 @@ g_raid_tr_iodone_raid1(struct g_raid_tr_
return;
}
cbp = g_clone_bio(pbp);
+ if (cbp == NULL) {
+ /*
+ * By flagging that we're not doing anything,
+ * we'll pick up the rebuild at a later point
+ * either by timeout or when we steal a small
+ * part of the active I/O.
+ */
+ g_destroy_bio(bp); /* reuse? */
+ trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
+ return;
+ }
cbp->bio_cmd = BIO_WRITE;
cbp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
cbp->bio_offset = bp->bio_offset;
cbp->bio_length = bp->bio_length;
G_RAID_LOGREQ(4, bp, "Queueing reguild write.");
+ g_destroy_bio(bp); /* reuse? */
g_raid_subdisk_iostart(trs->trso_failed_sd, cbp);
} else {
/*
@@ -630,8 +644,10 @@ g_raid_tr_iodone_raid1(struct g_raid_tr_
"rebuild write done. Error %d", bp->bio_error);
if (bp->bio_error != 0) {
g_raid_tr_raid1_rebuild_abort(tr, vol);
+ g_destroy_bio(bp); /* reuse? */
return;
}
+ g_destroy_bio(bp); /* reuse? */
/* XXX A lot of the following is needed when we kick of the work -- refactor */
nsd = trs->trso_failed_sd;
g_raid_unlock_range(sd->sd_volume,
@@ -657,6 +673,16 @@ g_raid_tr_iodone_raid1(struct g_raid_tr_
}
pbp->bio_offset = nsd->sd_rebuild_pos;
cbp = g_clone_bio(pbp);
+ if (cbp == NULL) {
+ /*
+ * By flagging that we're not doing anything,
+ * we'll pick up the rebuild at a later point
+ * either by timeout or when we steal a small
+ * part of the active I/O.
+ */
+ trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
+ return;
+ }
cbp->bio_cmd = BIO_READ;
cbp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
cbp->bio_offset = nsd->sd_rebuild_pos;
More information about the svn-src-projects
mailing list