>>>>> "Ewan" == Ewan D Milne <emilne@xxxxxxxxxx> writes: Ewan, Ewan> So, this is good, the current implementation has a flaw in that Ewan> under certain conditions, a device will get offlined immediately, Ewan> (i.e. if there are a few medium access commands pending, and they Ewan> all timeout), which isn't what was intended. Yeah. That was OK for my use case. I was trying to prevent the server from going into a tail spin. There was no chance of recovering the disk. But ideally we'd be offlining based on how many times we retry the same medium access command. Ewan> as separate medium access timeouts, but I think the original Ewan> intent of Martin's change wasn't to operate on such a short Ewan> time-scale, am I right, Martin? On the device that begat my original patch, SPC command responses were handled by the SAS controller firmware on behalf of all discovered devices. Regardless of whether said drives were still alive or not. Medium Access commands, however, would always get passed on to the physical drive for processing. So when a drive went pining for the fjords, TUR would always succeed whereas reads and writes would time out. -- Martin K. Petersen Oracle Linux Engineering
