On 4/20/2021 5:25 PM, Arun Easi wrote:
Hi Daniel,
On Tue, 20 Apr 2021, 11:28am, Daniel Wagner wrote:
----------------------------------------------------------------------
Hi Roman,
On Tue, Apr 20, 2021 at 08:35:10PM +0300, Roman Bolshakov wrote:
+ James S.
Daniel, WRT to your patch I don't think we should add one more approach
to set dev_loss_tmo via kernel module parameter as NVMe adopters are
going to be even more confused about the parameter. Just imagine
knowledge bases populated with all sorts of the workarounds, that apply
to kernel version x, y, z, etc :)
Totally agree. I consider this patch just a hack and way to get the
discussion going, hence the RFC :) Well, maybe we are going to add it
downstream in our kernels until we have a better way for setting the
dev_loss_tmo.
As explained the debugfs interface is not working (okay, that's
something which could be fixed) and it has the big problem that it is
not under control by udevd. Not sure if we with some new udev rules the
debugfs could automatically discovered or not.
Curious, which udev script does this today for FC SCSI?
In theory, the exsting fc nvmediscovery udev event has enough information
to find out the right qla2xxx debugfs node and set dev_loss_tmo.
What exists for FCP/SCSI is quite clear and reasonable. I don't know why
FC-NVMe rports should be way too different.
The lpfc driver does expose the FCP/SCSI and the FC-NVMe rports nicely
via the fc_remote_ports and this is what I would like to have from the
qla2xxx driver as well. qla2xxx exposes the FCP/SCSI rports but not the
FC-NVMe rports.
Given that FC NVME does not have sysfs hierarchy like FC SCSI, I see
utility in making FC-NVME ports available via fc_remote_ports. If, though,
a FC target port is dual protocol aware this would leave with only one
knob to control both.
I think, going with fc_remote_ports is better than introducing one more
way (like this patch) to set this.
Regards,
-Arun
Thanks Arun,
I think we're all better off if the qla exports the nvme nodes via the
scsi-side fc_remote_ports. In the end - we will commonize a fc
transport that then sits above scsi and nvme and will definitely be
compatible with what's there. The registration with scsi was rather
straight-forward for lpfc, so I assume it will be for qla as well and
the devloss interface, although kludegy to have the driver propagate the
scsi callback to nvme, also isn't much.
I also don't think we want to keep creating new mgmt points. it's
already ugly enough.
-- james
