On Mon, 2019-11-25 at 13:42 -0700, Drew Hastings wrote: > On Mon, Nov 25, 2019 at 12:48 PM Martin Wilck <mwilck@xxxxxxx> wrote: > > I think you are seeing this FIXME: > > > > https://elixir.bootlin.com/linux/v4.19.79/source/drivers/md/dm-mpath.c#L612 > > > > For your testing, perhaps you just remove that if(!pgpath) > > condition. > > > > Regards, > > Martin > > That's correct, thanks. It works as expected after removing that > condition. > > With some limited testing using fio, I don't see any obvious negative > impacts of allowing choose_pgpath to run each bio mapping. I also > don't see any significant increase in CPU usage or IO latency during > benchmarking. > > The FIXME comment acknowledges that this sort of defeats the > balancing function of multipath, so I'm assuming this was removed for > a good reason. Does anyone know why this was done? In my particular > use case, I benefit from balancing the paths for throughput, so it's > useful to me. This originates from the former use of dm-multipath for NVMe devices, the now obsolete "nvme" queue mode, and from the attempt to separate dm-multipath from SCSI and its device handler model. See 8d47e65948dd ("dm mpath: remove unnecessary NVMe branching in favor of scsi_dh checks"). It seems indeed strange to skip choosing the path in the absence of the SCSI device handler, which is (with a grain of salt) responsible for switching *path groups*, not paths inside a group. It's a corner case, I don't think many real-world multipath setups deploy bio-based dm- mpath without a device handler. So, I suggest you submit a patch and discuss this with Mike :-) Regards, Martin -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
