Re: [libvirt] [PATCH v2 1/2] vfio/mdev: add version attribute for mdev device

On Fri, 10 May 2019 10:36:09 +0100 "Dr. David Alan Gilbert" <dgilbert(a)redhat.com&gt; wrote: > * Cornelia Huck (cohuck(a)redhat.com) wrote: > > On Thu, 9 May 2019 17:48:26 +0100 > > "Dr. David Alan Gilbert" <dgilbert(a)redhat.com&gt; wrote: > > > > > * Cornelia Huck (cohuck(a)redhat.com) wrote: > > > > On Thu, 9 May 2019 16:48:57 +0100 > > > > "Dr. David Alan Gilbert" <dgilbert(a)redhat.com&gt; wrote: > > > > > > > > > * Cornelia Huck (cohuck(a)redhat.com) wrote: > > > > > > On Tue, 7 May 2019 15:18:26 -0600 > > > > > > Alex Williamson <alex.williamson(a)redhat.com&gt; wrote: > > > > > > > > > > > > > On Sun, 5 May 2019 21:49:04 -0400 > > > > > > > Yan Zhao <yan.y.zhao(a)intel.com&gt; wrote: > > > > > > > > > > > > > > + Errno: > > > > > > > > + If vendor driver wants to claim a mdev device incompatible to all other mdev > > > > > > > > + devices, it should not register version attribute for this mdev device. But if > > > > > > > > + a vendor driver has already registered version attribute and it wants to claim > > > > > > > > + a mdev device incompatible to all other mdev devices, it needs to return > > > > > > > > + -ENODEV on access to this mdev device's version attribute. > > > > > > > > + If a mdev device is only incompatible to certain mdev devices, write of > > > > > > > > + incompatible mdev devices's version strings to its version attribute should > > > > > > > > + return -EINVAL; > > > > > > > > > > > > > > I think it's best not to define the specific errno returned for a > > > > > > > specific situation, let the vendor driver decide, userspace simply > > > > > > > needs to know that an errno on read indicates the device does not > > > > > > > support migration version comparison and that an errno on write > > > > > > > indicates the devices are incompatible or the target doesn't support > > > > > > > migration versions. > > > > > > > > > > > > I think I have to disagree here: It's probably valuable to have an > > > > > > agreed error for 'cannot migrate at all' vs 'cannot migrate between > > > > > > those two particular devices'. Userspace might want to do different > > > > > > things (e.g. trying with different device pairs). > > > > > > > > > > Trying to stuff these things down an errno seems a bad idea; we can't > > > > > get much information that way. > > > > > > > > So, what would be a reasonable approach? Userspace should first read > > > > the version attributes on both devices (to find out whether migration > > > > is supported at all), and only then figure out via writing whether they > > > > are compatible? > > > > > > > > (Or just go ahead and try, if it does not care about the reason.) > > > > > > Well, I'm OK with something like writing to test whether it's > > > compatible, it's just we need a better way of saying 'no'. > > > I'm not sure if that involves reading back from somewhere after > > > the write or what. > > > > Hm, so I basically see two ways of doing that: > > - standardize on some error codes... problem: error codes can be hard > > to fit to reasons > > - make the error available in some attribute that can be read > > > > I'm not sure how we can serialize the readback with the last write, > > though (this looks inherently racy). > > > > How important is detailed error reporting here? > > I think we need something, otherwise we're just going to get vague > user reports of 'but my VM doesn't migrate'; I'd like the error to be > good enough to point most users to something they can understand > (e.g. wrong card family/too old a driver etc). Ok, that sounds like a reasonable point. Not that I have a better idea how to achieve that, though... we could also log a more verbose error message to the kernel log, but that's not necessarily where a user will look first. Ideally, we'd want to have the user space program setting up things querying the general compatibility for migration (so that it becomes their problem on how to alert the user to problems :), but I'm not sure how to eliminate the race between asking the vendor driver for compatibility and getting the result of that operation. Unless we introduce an interface that can retrieve _all_ results together with the written value? Or is that not going to be much of a problem in practice?