A single osd connection fault (e.g. tcp disconnect) wasn't
reopening the connection, which causes all current and future
requests for that osd to hang.
Signed-off-by: Sage Weil <sage@newdream.net>