Joe Thornber wrote:
As a result of Jens' global unplug work we've had to revist all the uses of blk_run_queues(), and be specific about which queues we need to run. In a lot of cases I've just knocked them out completely where they appear to be hangovers from 2.4 weirdness. If you start getting lumpy io with multipath, snapshots or mirroring this is probably the reason - please yell.
Incase people do test multipath, I wanted to say I fixed - but broke in the process - multipath failover. I fixed failfast in the block layer, but that SCSI code path was also broken and the patch did not get merged yet. You need the patch here http://marc.theaimsgroup.com/?l=linux-scsi&m=107904932710899&w=2 if you are using a SCSI driver, or you will get some weird IO patterns when doing failover.
Sorry, Mike