Re: ceph: don't unregister an MDS session before removing its caps

Hi Max,

Deferring unregister until after remove_session_caps() is the right fix
for the BUG_ON. One gap though: check_new_map() wasn't updated for the
new invariant. A CLOSED session now stays in mdsc->sessions[], but that
loop still treats a registered session as live — on a rank failover it
can overwrite CLOSED with RESTARTING and call send_mds_reconnect() on a
dying object, or in the export-target path treat the occupied slot as
"no session" and fail to open a replacement.

The fix: in check_new_map(), skip CLOSED sessions the same way you skip
NULL slots.

Could you respin?

--
Alex Markuze