[dm-devel] [PATCH v5 3/3] dm mpath: add CONFIG_DM_MULTIPATH_SG_IO - failover for SG_IO

[Martin Wilck]

On Do, 2021-07-01 at 09:56 +0200, Christoph Hellwig wrote:
> On Mon, Jun 28, 2021 at 05:15:58PM +0200, mwilck at suse.com wrote:
> > The qemu "pr-helper" was specifically invented for it. I
> > believe that this is the most important real-world scenario for
> > sending
> > SG_IO ioctls to device-mapper devices.
> 
> pr-helper obviously does not SG_IO on dm-multipath as that simply
> does
> not work.
> 
> More importantly - if you want to use persistent reservations use the
> kernel ioctls for that.  These work on SCSI, NVMe and device mapper
> without any extra magic.

This set is not about persistent reservations. Sorry if my mentioning
pr-helper made this impression. It was only meant to emphasize the fact
that qemu SCSI passthrough using SG_IO is an important use case.

> Failing over SG_IO does not make sense.  It is an interface
> specically
> designed to leave all error handling to the userspace program using
> it,
> and we should not change that for one specific error case.  If you
> want the kernel to handle errors for you, use the proper interfaces.
> In this case this is the persistent reservation ioctls.  If they miss
> some features that qemu needs we should add those.

I respectfully disagree. Users of dm-multipath devices expect that IO
succeeds as long as there's at least one healthy path. This is a
fundamental property of multipath devices. Whether IO is sent
"normally" or via SG_IO doesn't make a difference wrt this expectation.

The fact that qemu implements SCSI passthrough this way shows that this
is not just a naïve user expectation, but is shared by experienced
developers as well. AFAICS, we can't simply move the path error
detection and retry logic into user space qemu, because user space
doesn't have information about the status of the multipath map; not to
mention that doing this would be highly inefficient.

I agree that in principle, SG_IO error handling should be left to user
space. But in this specific case, it makes sense to handle just the
"path error vs. target error" distinction in the kernel. All else is of
course still handled by user space.

Regards,
Martin