<div dir="ltr"><div dir="ltr"> </div> <div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Wed, Jun 19, 2019 at 11:57 AM Cornelia Huck <<a href="mailto:cohuck@redhat.com">cohuck@redhat.com</a>> wrote: </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">On Wed, 19 Jun 2019 11:04:15 +0200 Sylvain Bauza <<a href="mailto:sbauza@redhat.com" target="_blank">sbauza@redhat.com</a>> wrote: > On Wed, Jun 19, 2019 at 12:27 AM Alex Williamson <<a href="mailto:alex.williamson@redhat.com" target="_blank">alex.williamson@redhat.com</a>> > wrote: > > > On Tue, 18 Jun 2019 14:48:11 +0200 > > Sylvain Bauza <<a href="mailto:sbauza@redhat.com" target="_blank">sbauza@redhat.com</a>> wrote: > > > > > On Tue, Jun 18, 2019 at 1:01 PM Cornelia Huck <<a href="mailto:cohuck@redhat.com" target="_blank">cohuck@redhat.com</a>> wrote: > > > > I think we need to reach consensus about the actual scope of the > > > > mdevctl tool. > > > > > > > > > > > Thanks Cornelia, my thoughts: > > > > > > - Is it supposed to be responsible for managing *all* mdev devices in > > > > the system, or is it more supposed to be a convenience helper for > > > > users/software wanting to manage mdevs? > > > > > > > > > > The latter. If an operator (or some software) wants to create mdevs by > > not > > > using mdevctl (and rather directly calling the sysfs), I think it's OK. > > > That said, mdevs created by mdevctl would be supported by systemctl, > > while > > > the others not but I think it's okay. > > > > I agree (sort of), and I'm hearing that we should drop any sort of > > automatic persistence of mdevs created outside of mdevctl. The problem > > comes when we try to draw the line between unmanaged and manged > > devices. For instance, if we have a command to list mdevs it would > > feel incomplete if it didn't list all mdevs both those managed by > > mdevctl and those created elsewhere. For managed devices, I expect > > we'll also have commands that allow the mode of the device to be > > switched between transient, saved, and persistent. Should a user then > > be allowed to promote an unmanaged device to one of these modes via the > > same command? Should they be allowed to stop an unmanaged device > > through driverctl? Through systemctl? These all seem like reasonable > > things to do, so what then is the difference between transient and > > unmanaged mdev and is mdevctl therefore managing all mdevs, not just > > those it has created? > > > > > Well, IMHO, mdevs created by mdevctl could all be persisted or transient > just by adding an option when calling mdevctl, like : > "mdevctl create-mdev [--transient] <uuid> <pci_id> <type>" where default > would be persisting the mdev. This sounds useful; the caller can avoid fiddling with sysfs entries directly, while not committing to a permanent configuration. > > For mdevs *not* created by mdevctl, then a usecase could be "I'd like to > ask mdevctl to manage mdevs I already created" and if so, a mdevctl command > like : > "mdevctl manage-mdev [--transient] <mdev_uuid>" What kind of 'managing' would this actually enable? If we rely on mdevctl working with sysfs directly for transient devices, I can't really think of anything... </blockquote><div> </div><div>Just for getting the list of mdevs and see whether they are persistent.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> > > Of course, that would mean that when you list mdevs by "mdev list-all" you > wouldn't get mdevs managed by mdevctl. > Thoughts ? > > > - Do we want mdevctl to manage config files for individual mdevs, or > > > > are they supposed to be in a common format that can also be managed > > > > by e.g. libvirt? > > > > > > > > > > Unless I misunderstand, I think mdevctl just helps to create mdevs for > > > being used by guests created either by libvirt or QEMU or even others. > > > How a guest would allocate a mdev (ie. saying "I'll use this specific > > mdev > > > UUID") is IMHO not something for mdevctl. > > > > Right, mdevctl isn't concerned with how a specific mdev is used, but I > > think what Connie is after is more the proposal from Daniel where > > libvirt can essentially manage mdevctl config files itself and then > > only invoke mdevctl for the dirty work of creating and deleting > > devices. In fact, assuming systemd, libvirt could avoid direct > > interaction with mdevctl entirely, instead using systemctl device units > > to start and stop the mdevs. Maybe where that proposal takes a turn is > > when we again consider non-systemd hosts, where maybe mdevctl needs to > > write out an init script per mdev and libvirt injecting itself into > > manipulation of the config files would either need to perform the same > > or fall back to mdevctl. Unfortunately there seems to be an ultimatum > > to either condone external config file manipulation or expand the scope > > of the project into becoming a library. > > > > > Well, like I said, I think it's maybe another user case : just using > libvirt when you want a guest having vGPUs and then libvirt would create > mdevs (so users wouln't need to know at that). > That said, for the moment, I think we don't really need it so maybe a new > RFE once we at least have mdevctl packaged and supported by RHEL ? If we allow config file handling directly, libvirt could start using it even without mdevctl present? (Not sure if that makes sense.) </blockquote><div> </div><div>Well, sure.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> > > > > - Should mdevctl be a stand-alone tool, provide library functions, or > > > > both? Related: should it keep any internal state that is not written > > > > to disk? (I think that also plays into the transient vs. persistent > > > > question.) > > > > I don't think we want an mdevctld, if that's what you mean by internal Yeah, mdevctld--. > > state not written to disk. I think we ideally want all state in the > > mdev config files or discerned through sysfs. How we handle > > non-systemd hosts may throw a wrench in that though since currently the > > systemd integration relies on a template to support arbitrary mdevs and > > I'm not sure how to replicate that in other init services. If we need > > to dynamically manage per mdev init files in addition to config files, > > we're not so self contained. > > > > > FWIW, I'd love using mdevctl for OpenStack (Nova) just at least for > > > creating persisted mdevs (ie. mdevs that would be recreated after > > rebooting > > > using systemctl). That's the real use case I need. > > > Whether libvirt would internally support mdevctl would be nice but that's > > > not really something Nova needs, so I leave others providing their own > > > thoughts. > > > > > > > > > My personal opinion is that mdevctl should be able to tolerate mdevs > > > > being configured by other means, but probably should not try to impose > > > > its own configuration if it detects that (unless explicitly asked to do > > > > so). Not sure how feasible that goal is. > > > > > > > > That's what I misunderstand : in order to have a guest using a vGPU, > > you > > > need to do two things : > > > 1/ create the mdev > > > 2/ allocate this created dev to a specific guest config > > > > > > Of course, we could imagine a way to have both steps to be done directly > > by > > > libvirt, but from my opinion, mdevctl is really helping 1/ and not 2/. > > > > Yep, we also don't want to presume libvirt is the only consumer here. > > mdevctl should also support other VM management tools, users who write > > their own management scripts, and even non-VM related use cases. > > > > > Oh yes, please don't premuse mdevctl is only needed by libvirt. > FWIW, once mdevctl is supported by RHEL, I'd love to use it for OpenStack > Nova at least because I want to persist the mdevs. > At the moment, Nova just creates mdevs directly by sysfs and look the > existing onces up by sysfs, but upstream community in Nova thinks the > mission statement is not about managing mdevs so we don't really want to > add in Nova some kind of DB persistence just for mdevs. So, Nova would basically poke mdevctl, but not interact with the config files directly? Or am I misunderstanding? </blockquote><div> </div><div>Correct, instead of doing something like <a href="https://github.com/openstack/nova/blob/master/nova/privsep/libvirt.py#L207-L216">https://github.com/openstack/nova/blob/master/nova/privsep/libvirt.py#L207-L216</a></div><div>That said, Nova could do like libvirt and create a config file, for sure.</div><div> </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> > > > > A well-defined config file format is probably a win, even if it only > > > > ends up being used by mdevctl itself. > > > > Yes, regardless of whether others touch them, conversion scripts on > > upgrade should be avoided. Do we need something beyond a key=value > > file? So far we're only storing the mdev type and startup mode, but > > vfio-ap clearly needs more, apparently key=value1,value2,... type > > representation. Still, I think I'd prefer simple over jumping to xml > > or json or yaml. Thanks, > > > > > Heh, in OpenStack discussing about a file format is possibly one of the > longest arguments we already have, so I leave others to say their own > opinions but FWIW, as we use Python we tend to prefer YAML files. I don't > care about the format tho, just take the most convenient for libvirt I'd > say. Aww, and here I was looking forward to a nice file format discussion... More seriously, as I said in my other reply, .ini style would be good, but using JSON probably gives us more flexibility in the long run. </blockquote></div></div>