[libvirt] Introduce vGPU mdev framework to libvirt

Alex Williamson alex.williamson at redhat.com
Wed Feb 15 16:55:28 UTC 2017


On Wed, 15 Feb 2017 14:43:22 +0100
Erik Skultety <eskultet at redhat.com> wrote:

> On Wed, Feb 15, 2017 at 09:50:03AM +0100, Martin Polednik wrote:
> > On 14/02/17 09:58 -0700, Alex Williamson wrote:  
> > > On Tue, 14 Feb 2017 16:50:14 +0100
> > > Martin Polednik <mpolednik at redhat.com> wrote:
> > >   
> > > > On 07/02/17 12:29 -0700, Alex Williamson wrote:  
> > > > >On Tue, 7 Feb 2017 17:26:51 +0100
> > > > >Erik Skultety <eskultet at redhat.com> wrote:
> > > > >  
> > > > >> On Mon, Feb 06, 2017 at 09:33:14AM -0700, Alex Williamson wrote:  
> > > > >> > On Mon,  6 Feb 2017 13:19:42 +0100
> > > > >> > Erik Skultety <eskultet at redhat.com> wrote:
> > > > >> >  
> > > > >> > > Finally. It's here. This is the initial suggestion on how libvirt might
> > > > >> > > interract with the mdev framework, currently only focussing on the non-managed
> > > > >> > > devices, i.e. those pre-created by the user, since that will be revisited once
> > > > >> > > we all settled on how the XML should look like, given we might not want to use
> > > > >> > > the sysfs path directly as an attribute in the domain XML. My proposal on the
> > > > >> > > XML is the following:
> > > > >> > >
> > > > >> > > <hostdev mode='subsystem' type='mdev'>
> > > > >> > >     <source>
> > > > >> > >         <!-- this is the host's physical device address -->
> > > > >> > >         <address domain='0x0000' bus='0x00' slot='0x00' function='0x00'>
> > > > >> > >         <uuid>vGPU_UUID<uuid>
> > > > >> > >     <source>
> > > > >> > >     <!-- target PCI address can be omitted to assign it automatically -->
> > > > >> > > </hostdev>
> > > > >> > >
> > > > >> > > So the mediated device is identified by the physical parent device visible on
> > > > >> > > the host and a UUID which allows us to construct the sysfs path by ourselves,
> > > > >> > > which we then put on the QEMU's command line.  
> > > > >> >
> > > > >> > Based on your test code, I think you're creating something like this:
> > > > >> >
> > > > >> > -device vfio-pci,sysfsdev=/sys/class/mdev_bus/0000:00:03.0/53764d0e-85a0-42b4-af5c-2046b460b1dc
> > > > >> >
> > > > >> > That would explain the need for the parent device address, but that's
> > > > >> > an entirely self inflicted requirement.  For a managed="no" scenarios,
> > > > >> > we shouldn't need the parent, we can get to the mdev device
> > > > >> > via /sys/bus/mdev/devices/53764d0e-85a0-42b4-af5c-2046b460b1dc.  So it  
> > > > >>
> > > > >> True, for managed="no" would this path be a nice optimization.
> > > > >>  
> > > > >> > seems that the UUID should be the only required source element for
> > > > >> > managed="no".
> > > > >> >
> > > > >> > For managed="yes", it seems like the parent device is still an optional  
> > > > >>
> > > > >> The reason I went with the parent address element (and purposely neglecting the
> > > > >> sample mtty driver) was that I assumed any modern mdev capable HW would be
> > > > >> accessible through the PCI bus on the host. Also I wanted to explicitly hint
> > > > >> libvirt as much as possible which parent device a vGPU device instance should
> > > > >> be created on in case there are more than one of them, rather then scanning
> > > > >> sysfs for a suitable parent which actually supports the given vGPU type.
> > > > >>  
> > > > >> > field.  The most important thing that libvirt needs to know when
> > > > >> > creating a mdev device for a VM is the mdev type name.  The parent
> > > > >> > device should be an optional field to help higher level management
> > > > >> > tools deal with placement of the device for locality or load balancing.
> > > > >> > Also, we can't assume that the parent device is a PCI device, the
> > > > >> > sample mtty driver already breaks this assumption.  
> > > > >>
> > > > >> Since we need to assume non-PCI devices and we still need to enable management
> > > > >> to hint libvirt about the parent to utilize load balancing and stuff, I've come
> > > > >> up with the following adjustments/ideas on how to reflect that in the XML:
> > > > >> - still use the address element but use it with the 'type' attribute [1] (still
> > > > >>   breaks the sample mtty driver though) while making the element truly optional
> > > > >>   if I'm going to be outvoted in favor of scanning the directory for a suitable
> > > > >>   parent device on our own, rather than requiring the user to provide that
> > > > >>
> > > > >> - providing either an attribute or a standalone element for the parent device
> > > > >>   name, like a string version of the PCI address or whatever form the parent
> > > > >>   device comes in (doesn't break the mtty driver but I don't quite like this)
> > > > >>
> > > > >> - providing a path element/attribute to sysfs pointing to the parent device
> > > > >>   which I'm afraid is what Daniel is not in favor of libvirt doing
> > > > >>
> > > > >> So, this is what I've so far come up with in terms of hinting libvirt about the
> > > > >> parent device, do you have any input on this, maybe some more ideas on how we
> > > > >> should identify the parent device?  
> > > > >
> > > > >IMO, if we cannot account for the mtty sample driver, we're doing it
> > > > >wrong.  I suppose we can leave it unspecified how one selects a parent
> > > > >device for the mtty driver, but it should be possible to expand the
> > > > >syntax to include it.  So I think that means that when the parent
> > > > >address is provided, the parent address type needs to be specified as
> > > > >PCI.  So...
> > > > >
> > > > > <hostdev mode='subsystem' type='mdev'>
> > > > >
> > > > >This needs to encompass the device API or else the optional VM address
> > > > >cannot be resolved.  Perhaps model='vfio-pci' here?  Seems similar to
> > > > >how we specify the device type for PCI controllers where we have
> > > > >multiple options:
> > > > >
> > > > > <hostdev mode='subsystem' type='mdev' model='vfio-pci'>
> > > > >
> > > > >   <source>
> > > > >
> > > > >For managed='no', I don't see that anything other than the mdev UUID is
> > > > >useful.
> > > > >
> > > > >     <uuid>MDEV_UUID</uuid>
> > > > >
> > > > >If libvirt gets into the business of creating mdev devices and we call
> > > > >that managed='yes', then the mdev type to create is required.  I don't
> > > > >know whether there's anything similar we can steal syntax from:
> > > > >
> > > > >     <type>"nvidia-11"</type>
> > > > >
> > > > >That's pretty horrible, needs some xml guru love.
> > > > >
> > > > >We need to provide for specifying a parent, but we can't assume the  
> > > > 
> > > > From higher level perspective, I believe it would be "good
> > > > enough" for most of the cases to only specify the type. Libvirt will
> > > > anyway have to be able to enumerate the devices for listAllDevices
> > > > afaik.
> > > > 
> > > > My wish would be specifying
> > > > <hostdev mode='subsystem' type='mdev'>
> > > >     <type>nvidia-11</type>
> > > > </hostdev>
> > > > unless the user has specific requests or some other decision (mmio
> > > > numa placement) takes place.  
> > > 
> > > Yes, the <type> is the minimum information necessary for libvirt to
> > > create the mdev device itself.  A <source> section could add optional
> > > placement information.  Note though that without an nvidia-11 type
> > > device on the system to query, the xml doesn't tell us what sort of
> > > device this creates in the VM.  We could assume that it's vfio-pci, but
> > > designing in an assumption isn't a great idea.  So, as above, some
> > > mechanism to make the xml self contained, such as specifying the model
> > > as vfio-pci, helps avoid that assumption and allows us to know the
> > > format for expressing the VM <address>  
> > 
> > As long as libvirt provides means to determine the model via device
> > listing (listAllDevices), OK.
> >   
> 
> Yes, libvirt will provide means expose this information.
> 
> > > > We would additionally need (allocated instances/max instances of that
> > > > type) in listAllDevices to account for the specific assignment
> > > > possibility.  
> > > 
> > > mdev devices support an available_instances per mdev type that is
> > > dynamically updated as devices are created.  The interaction of
> > > available_instances between different types is going to require some
> > > heuristics to understand.  Some vendors may not support heterogeneous
> > > types, others may pull from a common pool of resources, where each type
> > > may consume resources from that pool at different rates.  
> > 
> > Given common pool semantics, will we be able to calculate how many of
> > each type will be available in the pool if we were to instantiate
> > certain type? Example:
> > 
> > available types:
> > type_a: 4 devices (each consumes 1 "slot")
> > type_b: 1 device  (each consumes 4 "slots")
> > total "slots": 4
> >   
> 
> Well, if we could assume that the number of instances for a specific type would
> always be a power of 2 and the resources are distributed in that manner, then
> it's simple, you're allocating a resources that a more resource-demanding type
> would need to instantiate a single device, so you'll end up with one less
> device for each more resource-demanding type recursively. However, that is a
> strong assumption to make, so I'm not sure, it's possible that available
> instances, which only updates once you instantiated a specific type, is the
> only thing we should rely on.

Agree, and vendors can change how they manage this at any time.  For
instance if I boot one version of the kernel, i915 gives me:

i915-GVTg_V4_1
i915-GVTg_V4_2
i915-GVTg_V4_4

If I boot another, I get:

i915-GVTg_V4_1
i915-GVTg_V4_2
i915-GVTg_V4_5
i915-GVTg_V4_7

Now we don't have evenly divisible numbers.  If I create a type _1
device, available_instances still says I can create one type _5 or _7.
It's perhaps best for libvirt to just look at the current state and not
try to predict the future.

> > we know that creating type_a device prevents any
> > more type_b devices to be created.
> > 
> > Does NVIDIA or AMD use the common pool?

AMD isn't a player here yet, Intel and NVIDIA have vGPUs, IBM has a
model under development for S390 channel I/O.  The only thing you can
rely on is available_instances per mdev type at a given point of time.
How available_instances chances when we start creating devices is
vendor specific and may change at any time.

> > > > I'm not sure what the decision was wrt type naming, can 2 different
> > > > cards have similarly named type with different meaning?  
> > > 
> > > We don't deal in similarities, each type ID is unique and it's up to
> > > the mdev vendor driver to make sure that an "nvidia-11" on and M60 card
> > > is software equivalent to an "nvidia-11" on an M10 card.  If they're
> > > not equivalent, the type ID will be different.  Something we may want
> > > to consider eventually is whether we want/need to deal with
> > > compatibility strings.  For instance, NVIDIA seems to be tying the type
> > > ID strongly to specific implementations, an nvidia-11 may only be
> > > available on an M60 card.  An M10 card may offer an nvidia-21 type with
> > > similar capabilities.  There may be a need to express an mdev device as
> > > compatible with various type IDs for hardware availability, at the risk
> > > of exposing slight variations to the VM.  This could also make
> > > placement easier for vendor drivers that only support homogeneous mdev
> > > devices, "I prefer an mdev ID of type 'nvidia-11', but will accept one
> > > of type 'nvidia-12,nvidia-21'".  Thanks,  
> > 
> > I like the idea of libvirt being able to select one of specified
> > types, we have to bear in mind that it'll slightly complicate the XML:
> > 
> >    <mdev_types>
> >        <type>nvidia-11</type>
> >        <type>nvidia-21</type>
> >    </mdev_types>  
> 
> ^^ are you referring to nodedev XML or domain XML, because in case of a domain,
> there should be only one type per <hostdev type='mdev'>. There is also the
> ongoing question what's the best way to approach creation of mdev with libvirt
> and we have to be very careful with that so it won't bite us back in the
> future.
> However, for 7.4 the priority is to accept a pre-created device and to provide
> means in the nodedev driver to list all existing mdev devices and their
> corresponding parent devices.

Sorry if I confused the topic with some sort of compatibility listing.
I agree that for a <hostdev> there needs to be a single type if
libvirt is to create an instance of that type.  Any notion of
compatible or secondary acceptable types is lower priority than the
necessary basic behaviors, I'm just trying to plan ahead for later
extensions that might be useful.  Perhaps any notion of compatibility
lives in user define lists above libvirt.

> > That luckily shouldn't be problem for libvirt or management software.
> > On the other hand, the type equivalence will require some kind of
> > labeling on the management side -- user defines "mygpu" as "vgpu with
> > type nvidia-11 or nvidia-21" unless libvirt commits to a maintaining a
> > database with capability-equivalent types for devices (which, given
> > the generic-ness of the mdev, doesn't seem like a good idea).
> >   
> 
> Libvirt definitely shouldn't be handling type compatibility-related issues.
> As Alex pointed out, this should be vendor driver's responsibility. There's
> also Intel's KVMGT which has a different approach to it's type IDs. IIUC they
> based their type IDs on the fraction of actual resources used, i.e. type _1
> consumes the whole HW _2 consumes half, etc. but this is a question for Alex as
> he's been playing with it for some time. Anyhow, from my understanding Intel's
> types look more generic, thus more compatible with different HW revisions, if
> so, then in that case by dealing with the type compatibility, libvirt would be
> tailoring its logic to a specific vendor's use whereas I think libvirt
> should only focus on interacting with the mdev framework using the data it's
> got from the user. IOW new mdev-capable HW will be coming out which would in
> turn just bring more types to deal with. If the vendor driver won't be willing
> to accept any other type than just the set it's exporting, then I think the
> management may want to try to compensate for this with the information it can
> query from libvirt.

AIUI, looking at the example Intel mdev types I list above, that "V4"
indicates a Broadwell class GPU.  So don't be fooled into thinking
Intel is making some sort of generic device that it can represent on
any platform.  I'd expect a Skylake system to export similar types, but
with a "V5" component to the name.  NVIDIA naming is just more opaque,
possibly changing not only across generations but across
implementations.  I fully agree that it's the vendor's responsibility
to maintain that a given type is compatible wherever it is exposed and
libvirt's first priority is to focus on specifying a single type in the
xml and working towards instantiating an mdev of that type.  Libvirt
should never assume that anything other than that single, exact type is
compatible or sufficient for the VM.

I did plant the seed above about whether user defined compatibility
lists might be useful, it seems like something we should keep in mind,
but at a lower priority than any sort of initial support.  As Erik
suggested in a separate discussion, perhaps any notion of user defined
compatibility happens at a management layer above libvirt.  Migration
also needs to be considered when we think about compatible devices.
Compatibility likely only refers to the point at which we instantiate
the VM, if we were ever to support migration of an mdev device, the
target and source would need to be identical.  Yet more reasons for
libvirt to leave compatibility to higher layers of management tools.
Thanks,

Alex




More information about the libvir-list mailing list