On Wed, Sep 06, 2017 at 07:10:15PM +0200, Patrick Wildt wrote:
> On Wed, Sep 06, 2017 at 05:44:21PM +0200, Patrick Wildt wrote:
> > Hi,
> >
> > I'm testing some SoftRAID in a Mirroring setup and stumbled upon
> > something. When I set a disk offline, zero out the disk and the
> > metadata blocks, reboot, and then try to assemble it, it will
> > fail. Rightfully so, since the metadata information is completely
> > gone from one disk. That's not a reassemble, that's a rebuild.
> >
> > Now in the case that I want to actually force a whole new RAID
> > it looks like I have to pass -C force, to clear the metadata and
> > force creation of a new RAID. Unfortunately this does not work,
> > because sr_meta_native_attach(), which is called by
> > sr_meta_native_attach(), wants to make sure that all devices are
> > either non-SoftRAID or SoftRAID. Now in my case the disks are
> > kinda both. One zeroed, one not. In that case sr && not_sr is
> > true and it will exit with "not all chunks are of the native
> > metadata format".
> >
> > Is there a reason that sr_meta_attach() has to be called before
> > clearing the metadata on force?
> >
> > Untested diff below.
> >
> > Patrick
>
> Ah, well, sr_meta_clear() memsets sd->sd_meta, which is only later
> allocated during sr_meta_native_attach(), and sr_meta_read() of
> course also makes use of it. Maybe sr_meta_native_attach() can
> be more lax.
Simply ignoring the "incorrect" metadata format when we are forcing
assembly (and thus clearing the metadata) makes it work.
Opinions?
Patrick
diff --git a/sys/dev/softraid.c b/sys/dev/softraid.c
index 8b75f238072..b46cc9e27a4 100644
--- a/sys/dev/softraid.c
+++ b/sys/dev/softraid.c
@@ -1658,7 +1658,7 @@ sr_meta_native_attach(struct sr_discipline *sd, int force)
not_sr++;
}
- if (sr && not_sr) {
+ if (sr && not_sr && !force) {
sr_error(sc, "not all chunks are of the native metadata "
"format");
goto bad;
