Ceph » rbd

A Linux I/O request on an rbd image is broken into one or
more rbd requests, one request directed to each osd object
backing the image.

Each of these rbd requests receives a response, and its
result code is recorded in a table (called a "collection").

I'm not exactly sure what happens from the request initiator's
perspective when an error occurs, but it looks to me like
the aggregate result may not reflect the fact that an error
was reported when one of the rbd requests completed.

I'll have to spend a little more time on this to make sure
it's right, but for now wanted to document the problem.

I am going to fix this in my new request management code
(http://tracker.newdream.net/issues/3741) but the problem
exists in the current code as well.

I'm not sure how important it is to fix the old code, but
if it needs fixing I'd rather fix the old code first.