> [...] because in g_eli_read_metadata(), it doesn't check the return
> from eli_metadata_decode(), so it doesn't notice the EINVAL and
> happily reports back success without ever having touched the md :(
Fixing that is enough to prevent the panic at any rate. At least in
the normal case; if it randomly gets good magic, weird stuff might
still happen. But that lack of checking is a Real Bug by itself
anyway, so merits a fix.
Index: g_eli.c
===================================================================
--- g_eli.c (revision 280158)
+++ g_eli.c (working copy)
@@ -633,7 +633,9 @@
g_topology_lock();
if (buf == NULL)
goto end;
- eli_metadata_decode(buf, md);
+ error = eli_metadata_decode(buf, md);
+ if (error != 0)
+ goto end;
end:
if (buf != NULL)
g_free(buf);
(yes, the goto is redundant as all heck there. I decided to put it in
for symmetry with the other checking, as well as for a JIC, if e.g.
some code later got inserted after the decode before the end: label)
So for the moment, I've reinstated the conditional in
geli-delete.5.patch (which includes the above, though it's worthy of
commission by itself anyway). Technically it shouldn't be necessary
with that, since it'll fail on the reading anyway, but
geli: Cannot change configuration of onetime provider ada0s1.nop.
is friendlier and more informative than
geli: Cannot read metadata from ada0s1.nop (error=22).
and it also provides a convenient point to leave a comment about why
it doesn't [yet] work, until it does.
--
Matthew Fuller (MF4839) | [email protected]
Systems/Network Administrator | http://www.over-yonder.net/~fullermd/
On the Internet, nobody can hear you scream.
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-geom
To unsubscribe, send any mail to "[email protected]"
