On Tue, 2012-05-22 at 17:59 +0900, Jun'ichi Nomura wrote: > When a scsi host is kept busy by a dm-mpath device, other dm-mpath device > on the same host could be starved. For example: > http://www.redhat.com/archives/dm-devel/2012-May/msg00123.html > > It happens because dm-mpath delays request submission when the underlying > scsi host is busy even if sdev is not busy. > > For case like this, it is better to send the request down and let scsi do > appropriate starvation control over the shared resource. > > Though it might seem odd to change scsi's definition of a LLD being "busy", > it is reasonable because scsi_lld_busy (and blk_lld_busy) was introduced > to provide a hint for request-based stacking driver (i.e. dm-multipath) > and dm-multipath is the only user of this function. This explanation is rather dense. Isn't a more understandable explanation: block congestion control doesn't have any concept of fairness across multiple queues. This means that if SCSI reports the host as busy in the queue congestion control it can result in an unfair starvation situation in dm-mp if there are multiple multipath devices on the same host. The fix for this is to report only the sdev busy state (and ignore the host busy state) in the block congestion control call back. The host is still congested, but the SCSI subsystem will sort out the congestion in a fair way because it knows the relation between the queues and the host. And please put a comment in the code as well otherwise someone will eventually send a "fix" for this because we're not paying attention to host busy (and I'll have forgotten about the issue by then and might apply it). A final note is that this is more a band aid than a fix because this is still a congestion situation dm-mp should be aware of. James -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
