"Peter T. Breuer" wrote:

"A month of sundays ago ptb wrote:"

quoted

quoted

Actually, I think you want to make sure, in the async case, that the
write to the _primary_ disk completes, as opposed to just the first
write...I think you could get the I/O completing out of order, right?

Hmm .. that's a point. It's not a question of out of order, so much as
we want to give the maximum chance of getting one successful write and

Ooops .. I looked up the wrong bit of code in my reply. I looked at the
final end_req, instead of the mirrored i/o end_req. Yes, don't worry,
it's already done right. The ack happens early only if there is success
to report ("uptodate" is set)..

Right, but what if the data is successfully written to the secondary and
then ACKed back to userland and then the primary server fails before the
data gets written locally (I know this is unlikely, but possible,
nonetheless)? The primary will now have out-of-date data that userland
thinks is committed...

--
Paul