Openstack developers reported that newly-created mdevs were not recognized by libvirt until after a libvirt daemon restart. The source of the problem appears to be that when libvirt gets the udev 'add' event, the sysfs tree for that device might not be ready and so libvirt waits 100ms for it to appear (max 100 waits of 1ms each). But in the OpenStack environment, the sysfs tree for new mediated devices was taking closer to 250ms to appear and therefore libvirt gave up waiting and didn't add these new devices to its list of nodedevs. By changing the wait time to 1 second (max 100 waits of 10ms each), this should provide enough time to enable these deployments to recognize newly-created mediated devices, but it shouldn't increase the delay for more traditional deployments too much. Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2109450 Signed-off-by: Jonathon Jongsma <jjongsma(a)redhat.com&gt; --- Alternatively, we could switch to triggering off of the udev 'bind' event rather than the 'add' event, but I wasn't able to convince myself that this would result in 100% compatible behavior, so this felt like the safest solution. If others can convince me that switching to 'bind' is safe, I can re-submit this patch.

On 8/24/22 2:09 AM, Erik Skultety wrote: > On Tue, Aug 23, 2022 at 12:43:03PM -0500, Jonathon Jongsma wrote: > > Openstack developers reported that newly-created mdevs were not > > recognized by libvirt until after a libvirt daemon restart. The source > > of the problem appears to be that when libvirt gets the udev 'add' > > event, the sysfs tree for that device might not be ready and so libvirt > > waits 100ms for it to appear (max 100 waits of 1ms each). But in the > > OpenStack environment, the sysfs tree for new mediated devices was > > taking closer to 250ms to appear and therefore libvirt gave up waiting > > and didn't add these new devices to its list of nodedevs. > > > > By changing the wait time to 1 second (max 100 waits of 10ms each), this > > should provide enough time to enable these deployments to recognize > > newly-created mediated devices, but it shouldn't increase the delay for > > more traditional deployments too much. > > > > Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2109450 > > > > Signed-off-by: Jonathon Jongsma <jjongsma(a)redhat.com&gt; > > --- > > > > Alternatively, we could switch to triggering off of the udev 'bind' event > > rather than the 'add' event, but I wasn't able to convince myself that this > > would result in 100% compatible behavior, so this felt like the safest > > solution. If others can convince me that switching to 'bind' is safe, I can > > re-submit this patch. > > Is there a guarantee that the filesystem tree is ready by the time the event > arrives? I remember back in the day when I implemented this, this was even > discussed on the kernel list and the outcome was that each application needs to > sort this out on its own hinting that at least at that time there wasn't > any other way to do this reliably? Has something changed in the meantime? > > Erik > I'm afraid I don't actually know if anything has changed in the kernel in this area. That's basically the reason that I proposed the approach that I did. But I do know that in the bug referenced, the 'bind' event comes about 250ms later than the 'add' event. I'm not sure if the filesystem tree is necessarily ready on 'bind', but the fact that it is 250ms later means that, at minimum, there's a significantly better chance that it is ready by that point than at the time of 'add'.

On Wed, Aug 24, 2022 at 10:56:22AM -0500, Jonathon Jongsma wrote: > On 8/24/22 2:09 AM, Erik Skultety wrote: > > On Tue, Aug 23, 2022 at 12:43:03PM -0500, Jonathon Jongsma wrote: > > > Openstack developers reported that newly-created mdevs were not > > > recognized by libvirt until after a libvirt daemon restart. The source > > > of the problem appears to be that when libvirt gets the udev 'add' > > > event, the sysfs tree for that device might not be ready and so libvirt > > > waits 100ms for it to appear (max 100 waits of 1ms each). But in the > > > OpenStack environment, the sysfs tree for new mediated devices was > > > taking closer to 250ms to appear and therefore libvirt gave up waiting > > > and didn't add these new devices to its list of nodedevs. > > > > > > By changing the wait time to 1 second (max 100 waits of 10ms each), this > > > should provide enough time to enable these deployments to recognize > > > newly-created mediated devices, but it shouldn't increase the delay for > > > more traditional deployments too much. > > > > > > Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2109450 > > > > > > Signed-off-by: Jonathon Jongsma <jjongsma(a)redhat.com&gt; > > > --- > > > > > > Alternatively, we could switch to triggering off of the udev 'bind' event > > > rather than the 'add' event, but I wasn't able to convince myself that this > > > would result in 100% compatible behavior, so this felt like the safest > > > solution. If others can convince me that switching to 'bind' is safe, I can > > > re-submit this patch. > > > > Is there a guarantee that the filesystem tree is ready by the time the event > > arrives? I remember back in the day when I implemented this, this was even > > discussed on the kernel list and the outcome was that each application needs to > > sort this out on its own hinting that at least at that time there wasn't > > any other way to do this reliably? Has something changed in the meantime? > > > > Erik > > > > I'm afraid I don't actually know if anything has changed in the kernel in > this area. That's basically the reason that I proposed the approach that I > did. But I do know that in the bug referenced, the 'bind' event comes about > 250ms later than the 'add' event. I'm not sure if the filesystem tree is > necessarily ready on 'bind', but the fact that it is 250ms later means that, > at minimum, there's a significantly better chance that it is ready by that > point than at the time of 'add'. Looking at the commit message for 'bind' events https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit... [quote] Also, many drivers create additional driver-specific device attributes when binding to the device, to provide userspace with additional controls. The new events allow userspace to adjust these driver-specific attributes without worrying that they are not there yet. [/quote] My reading of this is that we should NOT assume sysfs files exist until we've seen the 'bind' event. IOW, the reason we have this problem with sysfs is precisely because we hooked into 'add' instead of 'bind'. The sleep papers over the mistake, while using 'bind' instead would address the root cause.

From the quote you posted I don't see how we're guaranteed that the sysfs tree

NB, if we use 'bind' we need to double check all our supported platforms are targetting kerenl >= 4.14 already.

On Thu, Aug 25, 2022 at 09:06:27AM +0200, Erik Skultety wrote: > On Wed, Aug 24, 2022 at 06:14:51PM +0100, Daniel P. Berrangé wrote: >> On Wed, Aug 24, 2022 at 10:56:22AM -0500, Jonathon Jongsma wrote: >>> On 8/24/22 2:09 AM, Erik Skultety wrote: >>>> On Tue, Aug 23, 2022 at 12:43:03PM -0500, Jonathon Jongsma wrote: >>>>> Openstack developers reported that newly-created mdevs were not >>>>> recognized by libvirt until after a libvirt daemon restart. The source >>>>> of the problem appears to be that when libvirt gets the udev 'add' >>>>> event, the sysfs tree for that device might not be ready and so libvirt >>>>> waits 100ms for it to appear (max 100 waits of 1ms each). But in the >>>>> OpenStack environment, the sysfs tree for new mediated devices was >>>>> taking closer to 250ms to appear and therefore libvirt gave up waiting >>>>> and didn't add these new devices to its list of nodedevs. >>>>> >>>>> By changing the wait time to 1 second (max 100 waits of 10ms each), this >>>>> should provide enough time to enable these deployments to recognize >>>>> newly-created mediated devices, but it shouldn't increase the delay for >>>>> more traditional deployments too much. >>>>> >>>>> Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2109450 >>>>> >>>>> Signed-off-by: Jonathon Jongsma <jjongsma(a)redhat.com&gt; >>>>> --- >>>>> >>>>> Alternatively, we could switch to triggering off of the udev 'bind' event >>>>> rather than the 'add' event, but I wasn't able to convince myself that this >>>>> would result in 100% compatible behavior, so this felt like the safest >>>>> solution. If others can convince me that switching to 'bind' is safe, I can >>>>> re-submit this patch. >>>> >>>> Is there a guarantee that the filesystem tree is ready by the time the event >>>> arrives? I remember back in the day when I implemented this, this was even >>>> discussed on the kernel list and the outcome was that each application needs to >>>> sort this out on its own hinting that at least at that time there wasn't >>>> any other way to do this reliably? Has something changed in the meantime? >>>> >>>> Erik >>>> >>> >>> I'm afraid I don't actually know if anything has changed in the kernel in >>> this area. That's basically the reason that I proposed the approach that I >>> did. But I do know that in the bug referenced, the 'bind' event comes about >>> 250ms later than the 'add' event. I'm not sure if the filesystem tree is >>> necessarily ready on 'bind', but the fact that it is 250ms later means that, >>> at minimum, there's a significantly better chance that it is ready by that >>> point than at the time of 'add'. >> >> Looking at the commit message for 'bind' events >> >> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit... >> >> [quote] >> Also, many drivers create additional driver-specific device attributes >> when binding to the device, to provide userspace with additional controls. >> The new events allow userspace to adjust these driver-specific attributes >> without worrying that they are not there yet. >> [/quote] >> >> My reading of this is that we should NOT assume sysfs files exist >> until we've seen the 'bind' event. >> >> IOW, the reason we have this problem with sysfs is precisely because >> we hooked into 'add' instead of 'bind'. The sleep papers over the >> mistake, while using 'bind' instead would address the root cause. > > From the quote you posted I don't see how we're guaranteed that the sysfs tree > is ready and it looks like each driver has too much room for interpreting > 'bind' whichever way they want. That said, I agree that switching to listening > to 'bind' instead of 'add' is an improvement, however we still may need to have > a timeout in place anyway as I still don't see any guarantees from the NVIDIA > driver that the sysfs tree will be ready by then. Right, lets do both, so might as well commit this timeout change immediately before release since it should be safe. With regards, Daniel