John Baldwin wrote: > > My hunch, which is why I suggested decreasing the number of > > tags seen by the driver, is that the tagged queues are over > > used, and this locks the disk up. My best guess is an off-by-one > > or an exceptional condition handler that was not an issue until > > recently, because of a FreeBSD interrupt architecture change > > having nothing to do with the driver itself (i.e. the reason it > > only happens under load, and didn't happen under the same load, > > before). > > Terry, we've had threaded interrupt handlers for over a year and a half > now. If the had really broken things in this basic a fashion we wouldn't > have made it this far with running systems. Your hypothesis about > something busted in the tagged queueing code seems sound but blaiming > this on interrupt threads doesn't make much sense to me.
The problems don't show up, except under extreme loads, with particular drives. Therefore, it is still my hunch. ;^). Dropping the queue depth to 8 from 16 to attempt to verify my hunch won't hurt anything, and may find the problem. It could still be an off-by-one error in Soren's code, as well (but I don't think it is). -- Terry To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message