On Tue, 20 Sep 2016 02:05:52 +0530 Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > 'fb_length': > Read-only file. Mandatory. > Returns <number>{K,M,G}, size of framebuffer. This can't be mandatory, it's only relevant to vGPU devices, vGPUs are just one user of mediated devices. [...] > 'params' > Write-Only file. Optional. > String input. Libvirt would pass the string given in XML file to > this file and then create mdev device. Set empty string to clear params. > For example, set parameter 'frame_rate_limiter=0' to disable frame rate > limiter for performance benchmarking, then create device of type 11. The > device created would have that parameter set by vendor driver. Might as well just call it "magic", there's absolutely no ability to introspect what parameters are allowed or even valid here. Can all parameters be changed dynamically? Do parameter changes apply to existing devices or only future devices? This is a poorly specified interface. I'd prefer this be done via module options on the vendor driver.

> The parent device would look like: > > <device> > <name>pci_0000_86_00_0</name> > <capability type='pci'> > <domain>0</domain> > <bus>134</bus> > <slot>0</slot> > <function>0</function> This is the parent device address?

> 2. Create/destroy mediated device > > With above example, vGPU device XML would look like: > > <device> > <name>my-vgpu</name> > <parent>pci_0000_86_00_0</parent> > <capability type='mdev'> > <type id='11'/> > <group>1</group> > <params>'frame_rate_limiter=0'</params> > </capability> > </device> > > 'type id' is mandatory. I was under the impression that the vendor supplied "name" was the one unique identifier. We're certainly not going to create a registrar to hand out type ids to each vendor, so what makes type id unique?

> * Clear params, if set earlier: > > echo "" > /sys/../0000\:86\:00.0/11/params So params only applies to the devices created while those params are set? This is inherently racy.

On Mon, 19 Sep 2016 23:50:56 +0200 Paolo Bonzini <pbonzini(a)redhat.com&gt; wrote: > On 19/09/2016 23:36, Alex Williamson wrote: >> On Tue, 20 Sep 2016 02:05:52 +0530 >> Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: >>> 'fb_length': >>> Read-only file. Mandatory. >>> Returns <number>{K,M,G}, size of framebuffer. >> >> This can't be mandatory, it's only relevant to vGPU devices, vGPUs are >> just one user of mediated devices. [...] >> >>> 'params' >>> Write-Only file. Optional. >>> String input. Libvirt would pass the string given in XML file to >>> this file and then create mdev device. Set empty string to clear params. >>> For example, set parameter 'frame_rate_limiter=0' to disable frame rate >>> limiter for performance benchmarking, then create device of type 11. The >>> device created would have that parameter set by vendor driver. >> >> Might as well just call it "magic", there's absolutely no ability to >> introspect what parameters are allowed or even valid here. Can all >> parameters be changed dynamically? Do parameter changes apply to >> existing devices or only future devices? This is a poorly specified >> interface. I'd prefer this be done via module options on the vendor >> driver. > > Or additional files in the mdev's sysfs directory? Sure, the vendor driver could certainly support that, it'd be vendor specific of course.

As I've said in my earlier reply - libvirt will *NOT* support passing arbitrary vendor specific parameters as a blob via the XML. Everything that appears in the XML must be *fully* specified and explicitly represented in the XML as a distinct attribute or element.

> > As I've said in my earlier reply - libvirt will *NOT* support passing > > arbitrary vendor specific parameters as a blob via the XML. Everything > > that appears in the XML must be *fully* specified and explicitly > > represented in the XML as a distinct attribute or element. > > Are generic key/value attributes (e.g. a <attribute> element) acceptable? Only if libvirt has a known list of valid attribute key names upfront. We don't want to just blindly expose arbitary vendor specific keys exposed by the kernel. Libvirt's job is to ensure the XML representation is vendor portable

On Tue, Sep 20, 2016 at 05:05:43PM +0200, Paolo Bonzini wrote: > > > On 20/09/2016 16:58, Daniel P. Berrange wrote: >>>> As I've said in my earlier reply - libvirt will *NOT* support passing >>>> arbitrary vendor specific parameters as a blob via the XML. Everything >>>> that appears in the XML must be *fully* specified and explicitly >>>> represented in the XML as a distinct attribute or element. >>> >>> Are generic key/value attributes (e.g. a <attribute> element) acceptable? >> >> Only if libvirt has a known list of valid attribute key names upfront. >> We don't want to just blindly expose arbitary vendor specific keys exposed >> by the kernel. Libvirt's job is to ensure the XML representation is vendor >> portable >

> Ok, then I guess vendor-specific mdev parameters are out completely. Or > could they be added under a separate namespace, like QEMU passthrough? The QEMU XML namespace that allows passthrough is intended either as a short term hack to let people experiment with new QEMU features, before they get explicitly represented in the main XML namespace, or in a few cases, for things we never want to support officially. Any VM which uses the separate namespace is tainted, which means if theres a bug report we'll require the reported to remove whatever config caused the tainting and then reproduce the problem. If the vendor specific mdev parameters are things that apps will often need to be able to set, then allowing it via a separate namespace is just providing a backdoor that everyone needs to use, defeating the point of using a separate namespace. We don't want to knowingly design a system which defacto requires the app to use the separate namespace most/all of the time. Regards, Daniel

On Tue, Sep 20, 2016 at 10:01:18PM +0530, Kirti Wankhede wrote: > > > On 9/20/2016 8:44 PM, Daniel P. Berrange wrote: >> On Tue, Sep 20, 2016 at 05:05:43PM +0200, Paolo Bonzini wrote: >>> >>> >>> On 20/09/2016 16:58, Daniel P. Berrange wrote: >>>>>> As I've said in my earlier reply - libvirt will *NOT* support passing >>>>>> arbitrary vendor specific parameters as a blob via the XML. Everything >>>>>> that appears in the XML must be *fully* specified and explicitly >>>>>> represented in the XML as a distinct attribute or element. >>>>> >>>>> Are generic key/value attributes (e.g. a <attribute> element) acceptable? >>>> >>>> Only if libvirt has a known list of valid attribute key names upfront. >>>> We don't want to just blindly expose arbitary vendor specific keys exposed >>>> by the kernel. Libvirt's job is to ensure the XML representation is vendor >>>> portable >>> > > In key/value attributes (taking example from proposed xml file) > > <attribute name='resolution'>2560x1600</attribute> > > 'Key' (i.e. 'resolution') should be known upfront, not the value, right? Yes, the actual value is not important - only its structured. ie, libvirt would check that it is in the format '$WIDTHx$HEIGHT' and reject it if not.

On 9/20/2016 8:44 PM, Daniel P. Berrange wrote: > On Tue, Sep 20, 2016 at 05:05:43PM +0200, Paolo Bonzini wrote: >> >> >> On 20/09/2016 16:58, Daniel P. Berrange wrote: >>>>> As I've said in my earlier reply - libvirt will *NOT* support passing >>>>> arbitrary vendor specific parameters as a blob via the XML. Everything >>>>> that appears in the XML must be *fully* specified and explicitly >>>>> represented in the XML as a distinct attribute or element. >>>> >>>> Are generic key/value attributes (e.g. a <attribute> element) acceptable? >>> >>> Only if libvirt has a known list of valid attribute key names upfront. >>> We don't want to just blindly expose arbitary vendor specific keys exposed >>> by the kernel. Libvirt's job is to ensure the XML representation is vendor >>> portable >> In key/value attributes (taking example from proposed xml file) <attribute name='resolution'>2560x1600</attribute> 'Key' (i.e. 'resolution') should be known upfront, not the value, right?

On 20/09/2016 17:14, Daniel P. Berrange wrote: > Any VM which > uses the separate namespace is tainted, which means if theres a bug > report we'll require the reported to remove whatever config caused > the tainting and then reproduce the problem. > > If the vendor specific mdev parameters are things that apps will often > need to be able to set, then allowing it via a separate namespace is > just providing a backdoor that everyone needs to use, defeating the > point of using a separate namespace. No, they don't need to be often set, and tainting the domain is a good idea.

From: Alex Williamson [mailto:alex.williamson@redhat.com] Sent: Tuesday, September 20, 2016 5:36 AM > In the above example directory '11' represents a type id of mdev device. > 'name', 'fb_length', 'resolution', 'heads', 'max_instance' and > 'requires_group' would be Read-Only files that vendor would provide to > describe about that type. > > 'create': > Write-only file. Mandatory. > Accepts string to create mediated device. > > 'name': > Read-Only file. Mandatory. > Returns string, the name of that type id. > > 'fb_length': > Read-only file. Mandatory. > Returns <number>{K,M,G}, size of framebuffer. This can't be mandatory, it's only relevant to vGPU devices, vGPUs are just one user of mediated devices.

> > 'max_instance': > Read-Only file. Mandatory. > Returns integer. Returns maximum mdev device could be created > at the moment when this file is read. This count would be updated by > vendor driver. Before creating mdev device of this type, check if > max_instance is > 0. Didn't we discuss this being being something like "available_instances" with a "devices" sub-directory linking current devices of that type?

> 'params' > Write-Only file. Optional. > String input. Libvirt would pass the string given in XML file to > this file and then create mdev device. Set empty string to clear params. > For example, set parameter 'frame_rate_limiter=0' to disable frame rate > limiter for performance benchmarking, then create device of type 11. The > device created would have that parameter set by vendor driver. Might as well just call it "magic", there's absolutely no ability to introspect what parameters are allowed or even valid here. Can all parameters be changed dynamically? Do parameter changes apply to existing devices or only future devices? This is a poorly specified interface. I'd prefer this be done via module options on the vendor driver.

> > 2. Create/destroy mediated device > > With above example, vGPU device XML would look like: > > <device> > <name>my-vgpu</name> > <parent>pci_0000_86_00_0</parent> > <capability type='mdev'> > <type id='11'/> > <group>1</group> > <params>'frame_rate_limiter=0'</params> > </capability> > </device> > > 'type id' is mandatory. I was under the impression that the vendor supplied "name" was the one unique identifier. We're certainly not going to create a registrar to hand out type ids to each vendor, so what makes type id unique? I have a hard time believing that a given vendor can even allocate unique type ids for their own devices. Unique type id across vendors is not practical. So which attribute are we actually using to identify which type of mdev device we need and why would we ever specify additional attributes like fb_length? Doesn't the vendor guarantee that "GRID M60-0B" has a fixed setup of those attributes?

> 'params' is optional field. User should set this field if extra > parameters need to be set for a particular vGPU device. Libvirt don't > need to parse these params. These are meant for vendor driver. ie. magic black hole. nope.

On 9/20/2016 3:06 AM, Alex Williamson wrote: > On Tue, 20 Sep 2016 02:05:52 +0530 > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > >> Hi libvirt experts, >> >> Thanks for valuable input on v1 version of RFC. >> >> Quick brief, VFIO based mediated device framework provides a way to >> virtualize their devices without SR-IOV, like NVIDIA vGPU, Intel KVMGT >> and IBM's channel IO. This framework reuses VFIO APIs for all the >> functionalities for mediated devices which are currently being used for >> pass through devices. This framework introduces a set of new sysfs files >> for device creation and its life cycle management. >> >> Here is the summary of discussion on v1: >> 1. Discover mediated device: >> As part of physical device initialization process, vendor driver will >> register their physical devices, which will be used to create virtual >> device (mediated device, aka mdev) to the mediated framework. >> >> Vendor driver should specify mdev_supported_types in directory format. >> This format is class based, for example, display class directory format >> should be as below. We need to define such set for each class of devices >> which would be supported by mediated device framework. >> >> --- mdev_destroy > > I thought we replaced mdev_destroy with a remove attribute for each > mdev device. > >> --- mdev_supported_types >> |-- 11 >> | |-- create >> | |-- name >> | |-- fb_length >> | |-- resolution >> | |-- heads >> | |-- max_instances >> | |-- params >> | |-- requires_group >> |-- 12 >> | |-- create >> | |-- name >> | |-- fb_length >> | |-- resolution >> | |-- heads >> | |-- max_instances >> | |-- params >> | |-- requires_group >> |-- 13 >> |-- create >> |-- name >> |-- fb_length >> |-- resolution >> |-- heads >> |-- max_instances >> |-- params >> |-- requires_group >> >> >> In the above example directory '11' represents a type id of mdev device. >> 'name', 'fb_length', 'resolution', 'heads', 'max_instance' and >> 'requires_group' would be Read-Only files that vendor would provide to >> describe about that type. >> >> 'create': >> Write-only file. Mandatory. >> Accepts string to create mediated device. >> >> 'name': >> Read-Only file. Mandatory. >> Returns string, the name of that type id. >> >> 'fb_length': >> Read-only file. Mandatory. >> Returns <number>{K,M,G}, size of framebuffer. > > This can't be mandatory, it's only relevant to vGPU devices, vGPUs are > just one user of mediated devices. > As per our discussion in BOF, we would define class and its attributes. As I mentioned above these are attributes of "display" class.

>> >> 'resolution': >> Read-Only file. Mandatory. >> Returns 'hres x vres' format. Maximum supported resolution. > > Same. > >> >> 'heads': >> Read-Only file. Mandatory. >> Returns integer. Number of maximum heads supported. > > Same. > >> >> 'max_instance': >> Read-Only file. Mandatory. >> Returns integer. Returns maximum mdev device could be created >> at the moment when this file is read. This count would be updated by >> vendor driver. Before creating mdev device of this type, check if >> max_instance is > 0. > > Didn't we discuss this being being something like "available_instances" > with a "devices" sub-directory linking current devices of that type? > I'm ok with 'available_instances' as name of this file. Yes, mdev device directory will have links to devices of that type. I think that is part of mdev core module discussion. I'm trying to list sysfs files for libvirt interface here to focus more on libvirt interface. Hence didn't add that. I'll make sure to add all such details in future. >> 'params' >> Write-Only file. Optional. >> String input. Libvirt would pass the string given in XML file to >> this file and then create mdev device. Set empty string to clear params. >> For example, set parameter 'frame_rate_limiter=0' to disable frame rate >> limiter for performance benchmarking, then create device of type 11. The >> device created would have that parameter set by vendor driver. > > Might as well just call it "magic", there's absolutely no ability to > introspect what parameters are allowed or even valid here. Libvirt don't need to introspect these parameters. These are meant for vendor driver. Let vendor driver interpret these parameters and take action based on that.

> Can all parameters be changed dynamically? Parameters here would be extra parameters which are not listed above in mandatory list. If such parameters are required to set, these parameters should be set per mdev device before VM is booted. > Do parameter changes apply to existing devices or only future devices? No, it should be applied at the time when mdev device is created or after mdev device is created but before VM is booted. It will not be applicable to existing devices. > This is a poorly specified > interface. I'd prefer this be done via module options on the vendor > driver. > Module option is not suitable here, in that case such parameters would be applied to all mdev device created for vendor driver and that is not what we want to.

>> 'requires_group' >> Read-Only file. Optional. >> This should be provided by vendor driver if vendor driver need to >> group mdev devices in one domain so that vendor driver can use 'first >> open' to commit resources of all mdev devices associated to that domain >> and 'last close' to free those. >> >> The parent device would look like: >> >> <device> >> <name>pci_0000_86_00_0</name> >> <capability type='pci'> >> <domain>0</domain> >> <bus>134</bus> >> <slot>0</slot> >> <function>0</function> > > This is the parent device address? (I think we'd re-use the > specification for assigned devices) Is this optional? Couldn't > libvirt choose to pick any parent device supplying the specified type > if not supplied? > >> <capability type='mdev'> >> <!-- one type element per sysfs directory --> >> <type id='11'> >> <!-- one element per sysfs file roughly --> >> <name>GRID M60-0B</name> >> <attribute name='fb_length'>512M</attribute> >> <attribute name='resolution'>2560x1600</attribute> >> <attribute name='heads'>2</attribute> >> <attribute name='max_instances'>16</attribute> >> <attribute name='requires_group'>1</attribute> >> </type> >> </capability> >> <product id='...'>GRID M60</product> >> <vendor id='0x10de'>NVIDIA</vendor> > > What are these specifying? > >> </capability> >> </device> >> >> 2. Create/destroy mediated device >> >> With above example, vGPU device XML would look like: >> >> <device> >> <name>my-vgpu</name> >> <parent>pci_0000_86_00_0</parent> >> <capability type='mdev'> >> <type id='11'/> >> <group>1</group> >> <params>'frame_rate_limiter=0'</params> >> </capability> >> </device> >> >> 'type id' is mandatory. > > I was under the impression that the vendor supplied "name" was the one > unique identifier. We're certainly not going to create a registrar to > hand out type ids to each vendor, so what makes type id unique? Type id is unique, 'name' is the name (string) of device that would be displayed in guest OS for that type of mdev device.

> I have > a hard time believing that a given vendor can even allocate unique type > ids for their own devices. Unique type id across vendors is not > practical. So which attribute are we actually using to identify which > type of mdev device we need and why would we ever specify additional > attributes like fb_length? Doesn't the vendor guarantee that "GRID > M60-0B" has a fixed setup of those attributes? > Specifying attributes here is not our requirement. Yes we have fixed set of attributes for "GRID-M60-0B" and on. We are defining the attributed here for "display" class for all other vendor of gpu can use. >> 'group' is optional. It should be a unique number in the system among >> all the groups created for mdev devices. Its usage is: >> - not needed if single vGPU device is being assigned to a domain. >> - only need to be set if multiple vGPUs need to be assigned to a >> domain and vendor driver have 'requires_group' file in type id directory. >> - if type id directory include 'requires_group' and user tries to >> assign multiple vGPUs to a domain without having <group> field in XML, >> it will create single vGPU. > > We never finished our discussion of how this gets implemented or > whether a group applies only to devices from the same vendor, same > device, how heterogeneous groups are handled, etc. > We agreed on above points that I mentioned here and we wish to know what libvirt folks think, right? Since there were no inputs on that thread I thought I should summarize what had been discussed and get libvirt experts thoughts on this.

>> 'params' is optional field. User should set this field if extra >> parameters need to be set for a particular vGPU device. Libvirt don't >> need to parse these params. These are meant for vendor driver. > > ie. magic black hole. nope. > >> Libvirt need to follow the sequence to create device: >> * Read /sys/../0000\:86\:00.0/11/max_instances. If it is greater than 0, >> then only proceed else fail. >> >> * Set extra params if 'params' field exist in device XML and 'params' >> file exist in type id directory >> >> echo "frame_rate_limiter=0" > /sys/../0000\:86\:00.0/11/params >> >> * Autogenerate UUID >> * Create device: >> >> echo "$UUID:<group>" > /sys/../0000\:86\:00.0/11/create >> >> where <group> is optional. Group should be unique number among all >> the groups created for mdev devices. > > Is it an integer, a UUID, a string? How is the group for an mdev > device discovered? Can it be changed later? > Group number could be integer, that's where we need input from libvirt folks. It should be provided at the time of mdev device creation. It can't be changed after that. If user want to change group number destroy/remove mdev device and create it again with proper group number.

>> * Clear params, if set earlier: >> >> echo "" > /sys/../0000\:86\:00.0/11/params > > So params only applies to the devices created while those params are > set? This is inherently racy. > >> * To destroy device: >> >> echo $UUID > /sys/../0000\:86\:00.0/mdev_destroy > > echo 1 > /sys/path/to/mdev/device/remove > >> 3. Start/stop mediated device >> >> No change or requirement for libvirt as this will be handled by open() >> and close() callbacks to vendor driver. In case of multiple devices and >> 'requires_group' set, this will be handled in 'first open()' and 'last >> close()' on device in that group. >> >> 4. Launch QEMU/VM >> >> Pass the mdev sysfs path to QEMU as vfio-pci device. >> For above vGPU device example: >> >> -device vfio-pci,sysfsdev=/sys/bus/mdev/devices/$UUID >> >> 5. QEMU/VM Shutdown sequence >> >> No change or requirement for libvirt. >> >> 6. VM Reset >> >> No change or requirement for libvirt as this will be handled via VFIO >> reset API and QEMU process will keep running as before. >> >> 7. Hot-plug >> >> It is same syntax to create a virtual device for hot-plug. > > How do groups work with hotplug? Can a device be creating into an > existing, running group? Can a device be removed from an existing, > running group? Thanks, > Group is for vendor driver to decide when to commit resources for a group of devices to support multiple devices in a domain. It will be upto vendor driver how it manage it.

On 9/20/2016 8:13 PM, Alex Williamson wrote: > On Tue, 20 Sep 2016 19:51:58 +0530 > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > >> On 9/20/2016 3:06 AM, Alex Williamson wrote: >>> On Tue, 20 Sep 2016 02:05:52 +0530 >>> Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: >>> >>>> Hi libvirt experts, >>>> >>>> >>>> 'create': >>>> Write-only file. Mandatory. >>>> Accepts string to create mediated device. >>>> >>>> 'name': >>>> Read-Only file. Mandatory. >>>> Returns string, the name of that type id. >>>> >>>> 'fb_length': >>>> Read-only file. Mandatory. >>>> Returns <number>{K,M,G}, size of framebuffer. >>> >>> This can't be mandatory, it's only relevant to vGPU devices, vGPUs are >>> just one user of mediated devices. >>> >> >> As per our discussion in BOF, we would define class and its attributes. >> As I mentioned above these are attributes of "display" class. > > Just as I didn't know here, how does libvirt know the class of a given > type id? > Yes, we need some way to identify the class as Daniel mentioned in his suggestion. Add a file 'class' in mdev_supported_types that would return the class. >>> Can all parameters be changed dynamically? >> >> Parameters here would be extra parameters which are not listed above in >> mandatory list. If such parameters are required to set, these parameters >> should be set per mdev device before VM is booted. >> >>> Do parameter changes apply to existing devices or only future devices? >> >> No, it should be applied at the time when mdev device is created or >> after mdev device is created but before VM is booted. It will not be >> applicable to existing devices. >> >>> This is a poorly specified >>> interface. I'd prefer this be done via module options on the vendor >>> driver. >>> >> >> Module option is not suitable here, in that case such parameters would >> be applied to all mdev device created for vendor driver and that is not >> what we want to. > > Then use attributes on the mdev device itself, this is not an > acceptable interface. > >>>> >>>> With above example, vGPU device XML would look like: >>>> >>>> <device> >>>> <name>my-vgpu</name> >>>> <parent>pci_0000_86_00_0</parent> >>>> <capability type='mdev'> >>>> <type id='11'/> >>>> <group>1</group> >>>> <params>'frame_rate_limiter=0'</params> >>>> </capability> >>>> </device> >>>> >>>> 'type id' is mandatory. >>> >>> I was under the impression that the vendor supplied "name" was the one >>> unique identifier. We're certainly not going to create a registrar to >>> hand out type ids to each vendor, so what makes type id unique? >> >> Type id is unique, 'name' is the name (string) of device that would be >> displayed in guest OS for that type of mdev device. > > We were told at the BoF that name was the unique string which would be > consistent and you again indicate below that "GRID-M60-0B" has a fixed > set of attributes, so why do we care what type id is associated with > that name? > >>> I have >>> a hard time believing that a given vendor can even allocate unique type >>> ids for their own devices. Unique type id across vendors is not >>> practical. So which attribute are we actually using to identify which >>> type of mdev device we need and why would we ever specify additional >>> attributes like fb_length? Doesn't the vendor guarantee that "GRID >>> M60-0B" has a fixed setup of those attributes? >>> >> >> Specifying attributes here is not our requirement. Yes we have fixed set >> of attributes for "GRID-M60-0B" and on. >> We are defining the attributed here for "display" class for all other >> vendor of gpu can use. >> >> >>>> 'group' is optional. It should be a unique number in the system among >>>> all the groups created for mdev devices. Its usage is: >>>> - not needed if single vGPU device is being assigned to a domain. >>>> - only need to be set if multiple vGPUs need to be assigned to a >>>> domain and vendor driver have 'requires_group' file in type id directory. >>>> - if type id directory include 'requires_group' and user tries to >>>> assign multiple vGPUs to a domain without having <group> field in XML, >>>> it will create single vGPU. >>> >>> We never finished our discussion of how this gets implemented or >>> whether a group applies only to devices from the same vendor, same >>> device, how heterogeneous groups are handled, etc. >>> >> >> We agreed on above points that I mentioned here and we wish to know what >> libvirt folks think, right? >> Since there were no inputs on that thread I thought I should summarize >> what had been discussed and get libvirt experts thoughts on this. > > No, that was just an idea that was never full defined. We had that > idea, the IOMMU group idea, the invalid container idea... This is > still an open question in the design. > We do want to close down on the design as soon as possible. I had tried to address open questions in earlier discussion also. Let me address your concerns you have: > whether a group applies only to devices from the same vendor, Yes.

> same device, Group can be of mdev devices from different supported types and from different physical device but should be from same vendor.

>how heterogeneous groups are handled, 'heterogeneous' does that mean device grouped together with different vendors?

>>>> 'params' is optional field. User should set this field if extra >>>> parameters need to be set for a particular vGPU device. Libvirt don't >>>> need to parse these params. These are meant for vendor driver. >>> >>> ie. magic black hole. nope. >>> >>>> Libvirt need to follow the sequence to create device: >>>> * Read /sys/../0000\:86\:00.0/11/max_instances. If it is greater than 0, >>>> then only proceed else fail. >>>> >>>> * Set extra params if 'params' field exist in device XML and 'params' >>>> file exist in type id directory >>>> >>>> echo "frame_rate_limiter=0" > /sys/../0000\:86\:00.0/11/params >>>> >>>> * Autogenerate UUID >>>> * Create device: >>>> >>>> echo "$UUID:<group>" > /sys/../0000\:86\:00.0/11/create >>>> >>>> where <group> is optional. Group should be unique number among all >>>> the groups created for mdev devices. >>> >>> Is it an integer, a UUID, a string? How is the group for an mdev >>> device discovered? Can it be changed later? >>> >> >> Group number could be integer, that's where we need input from libvirt >> folks. It should be provided at the time of mdev device creation. It >> can't be changed after that. If user want to change group number >> destroy/remove mdev device and create it again with proper group number. > > My concern is that a type id seems arbitrary but we're specifying that > it be unique. We already have something unique, the name. So why try > to make the type id unique as well? A vendor can accidentally create > their vendor driver so that a given name means something very > specific. On the other hand they need to be extremely deliberate to > coordinate that a type id means a unique thing across all their product > lines. > Let me clarify, type id should be unique in the list of mdev_supported_types. You can't have 2 directories in with same name.

I think here 'name' is the problem and it is not used in xml file. Lets not make it mandatory. But vendor driver can have it which 3rd party can use to read. Lets remove 'id' from type id in XML if that is the concern. Supported types is going to be defined by vendor driver, so let vendor driver decide what to use for directory name and same should be used in device xml file, it could be '11' or "GRID M60-0B": <device> <name>my-vgpu</name> <parent>pci_0000_86_00_0</parent> <capability type='mdev'> <type='11'/> ... </capability> </device> >>>> >>>> 7. Hot-plug >>>> >>>> It is same syntax to create a virtual device for hot-plug. >>> >>> How do groups work with hotplug? Can a device be creating into an >>> existing, running group? Can a device be removed from an existing, >>> running group? Thanks, >>> >> >> Group is for vendor driver to decide when to commit resources for a >> group of devices to support multiple devices in a domain. It will be >> upto vendor driver how it manage it. > > Then it's not sufficiently specified for libvirt to use. Thanks, > Can you clarify what if required for libvirt to use?

On 9/20/2016 10:20 PM, Alex Williamson wrote: > On Tue, 20 Sep 2016 21:53:16 +0530 > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > >> On 9/20/2016 8:13 PM, Alex Williamson wrote: >>> On Tue, 20 Sep 2016 19:51:58 +0530 >>> Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: >>> >>>> On 9/20/2016 3:06 AM, Alex Williamson wrote: >>>>> On Tue, 20 Sep 2016 02:05:52 +0530 >>>>> Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: >>>>> >>>>>> Hi libvirt experts, >>>>>> >>>>>> >>>>>> 'create': >>>>>> Write-only file. Mandatory. >>>>>> Accepts string to create mediated device. >>>>>> >>>>>> 'name': >>>>>> Read-Only file. Mandatory. >>>>>> Returns string, the name of that type id. >>>>>> >>>>>> 'fb_length': >>>>>> Read-only file. Mandatory. >>>>>> Returns <number>{K,M,G}, size of framebuffer. >>>>> >>>>> This can't be mandatory, it's only relevant to vGPU devices, vGPUs are >>>>> just one user of mediated devices. >>>>> >>>> >>>> As per our discussion in BOF, we would define class and its attributes. >>>> As I mentioned above these are attributes of "display" class. >>> >>> Just as I didn't know here, how does libvirt know the class of a given >>> type id? >>> >> >> Yes, we need some way to identify the class as Daniel mentioned in his >> suggestion. Add a file 'class' in mdev_supported_types that would return >> the class. >> >> >>>>> Can all parameters be changed dynamically? >>>> >>>> Parameters here would be extra parameters which are not listed above in >>>> mandatory list. If such parameters are required to set, these parameters >>>> should be set per mdev device before VM is booted. >>>> >>>>> Do parameter changes apply to existing devices or only future devices? >>>> >>>> No, it should be applied at the time when mdev device is created or >>>> after mdev device is created but before VM is booted. It will not be >>>> applicable to existing devices. >>>> >>>>> This is a poorly specified >>>>> interface. I'd prefer this be done via module options on the vendor >>>>> driver. >>>>> >>>> >>>> Module option is not suitable here, in that case such parameters would >>>> be applied to all mdev device created for vendor driver and that is not >>>> what we want to. >>> >>> Then use attributes on the mdev device itself, this is not an >>> acceptable interface. >>> >>>>>> >>>>>> With above example, vGPU device XML would look like: >>>>>> >>>>>> <device> >>>>>> <name>my-vgpu</name> >>>>>> <parent>pci_0000_86_00_0</parent> >>>>>> <capability type='mdev'> >>>>>> <type id='11'/> >>>>>> <group>1</group> >>>>>> <params>'frame_rate_limiter=0'</params> >>>>>> </capability> >>>>>> </device> >>>>>> >>>>>> 'type id' is mandatory. >>>>> >>>>> I was under the impression that the vendor supplied "name" was the one >>>>> unique identifier. We're certainly not going to create a registrar to >>>>> hand out type ids to each vendor, so what makes type id unique? >>>> >>>> Type id is unique, 'name' is the name (string) of device that would be >>>> displayed in guest OS for that type of mdev device. >>> >>> We were told at the BoF that name was the unique string which would be >>> consistent and you again indicate below that "GRID-M60-0B" has a fixed >>> set of attributes, so why do we care what type id is associated with >>> that name? >>> >>>>> I have >>>>> a hard time believing that a given vendor can even allocate unique type >>>>> ids for their own devices. Unique type id across vendors is not >>>>> practical. So which attribute are we actually using to identify which >>>>> type of mdev device we need and why would we ever specify additional >>>>> attributes like fb_length? Doesn't the vendor guarantee that "GRID >>>>> M60-0B" has a fixed setup of those attributes? >>>>> >>>> >>>> Specifying attributes here is not our requirement. Yes we have fixed set >>>> of attributes for "GRID-M60-0B" and on. >>>> We are defining the attributed here for "display" class for all other >>>> vendor of gpu can use. >>>> >>>> >>>>>> 'group' is optional. It should be a unique number in the system among >>>>>> all the groups created for mdev devices. Its usage is: >>>>>> - not needed if single vGPU device is being assigned to a domain. >>>>>> - only need to be set if multiple vGPUs need to be assigned to a >>>>>> domain and vendor driver have 'requires_group' file in type id directory. >>>>>> - if type id directory include 'requires_group' and user tries to >>>>>> assign multiple vGPUs to a domain without having <group> field in XML, >>>>>> it will create single vGPU. >>>>> >>>>> We never finished our discussion of how this gets implemented or >>>>> whether a group applies only to devices from the same vendor, same >>>>> device, how heterogeneous groups are handled, etc. >>>>> >>>> >>>> We agreed on above points that I mentioned here and we wish to know what >>>> libvirt folks think, right? >>>> Since there were no inputs on that thread I thought I should summarize >>>> what had been discussed and get libvirt experts thoughts on this. >>> >>> No, that was just an idea that was never full defined. We had that >>> idea, the IOMMU group idea, the invalid container idea... This is >>> still an open question in the design. >>> >> >> We do want to close down on the design as soon as possible. >> I had tried to address open questions in earlier discussion also. Let me >> address your concerns you have: >> >>> whether a group applies only to devices from the same vendor, >> >> Yes. > > So should we call this a vendor_group? How does libvirt learn the > vendor? We can't assume that the parent device will always be a PCI > device. In sysfs, devices, even other than PCI devices keeps link to its owner driver. We can make use of that to find the owner driver. > Maybe it's actually a driver_group and libvirt should put any > mdev devices with the same ->parent->driver into the same group. > Yes. > If these groups are managed via the vendor driver, we don't really need > a global namespace of group ids, isn't it per vendor driver? Maybe the > group should be a uuid rather than an integer to effectively remove the > collision/uniqueness problem. > Ok, UUID is also fine. It should be unique. >>> same device, >> >> Group can be of mdev devices from different supported types and from >> different physical device but should be from same vendor. > > I think we're just reinforcing that this is a driver-group. Another > vendor might require a device-group, where devices must all be attached > to the same parent device. "group" is too generic. > >>> how heterogeneous groups are handled, >> >> 'heterogeneous' does that mean device grouped together with different >> vendors? > > Different vendors or different type-ids. I think you've clarified both > of these, but it certainly wasn't specified previously. > >>>>>> 'params' is optional field. User should set this field if extra >>>>>> parameters need to be set for a particular vGPU device. Libvirt don't >>>>>> need to parse these params. These are meant for vendor driver. >>>>> >>>>> ie. magic black hole. nope. >>>>> >>>>>> Libvirt need to follow the sequence to create device: >>>>>> * Read /sys/../0000\:86\:00.0/11/max_instances. If it is greater than 0, >>>>>> then only proceed else fail. >>>>>> >>>>>> * Set extra params if 'params' field exist in device XML and 'params' >>>>>> file exist in type id directory >>>>>> >>>>>> echo "frame_rate_limiter=0" > /sys/../0000\:86\:00.0/11/params >>>>>> >>>>>> * Autogenerate UUID >>>>>> * Create device: >>>>>> >>>>>> echo "$UUID:<group>" > /sys/../0000\:86\:00.0/11/create >>>>>> >>>>>> where <group> is optional. Group should be unique number among all >>>>>> the groups created for mdev devices. >>>>> >>>>> Is it an integer, a UUID, a string? How is the group for an mdev >>>>> device discovered? Can it be changed later? >>>>> >>>> >>>> Group number could be integer, that's where we need input from libvirt >>>> folks. It should be provided at the time of mdev device creation. It >>>> can't be changed after that. If user want to change group number >>>> destroy/remove mdev device and create it again with proper group number. >>> >>> My concern is that a type id seems arbitrary but we're specifying that >>> it be unique. We already have something unique, the name. So why try >>> to make the type id unique as well? A vendor can accidentally create >>> their vendor driver so that a given name means something very >>> specific. On the other hand they need to be extremely deliberate to >>> coordinate that a type id means a unique thing across all their product >>> lines. >>> >> >> Let me clarify, type id should be unique in the list of >> mdev_supported_types. You can't have 2 directories in with same name. > > Of course, but does that mean it's only unique to the machine I'm > currently running on? Let's say I have a Tesla P100 on my system and > type-id 11 is named "GRID-M60-0B". At some point in the future I > replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that > new card still going to be a "GRID-M60-0B"? If not then we've based > our XML on the wrong attribute. If the new device does not support > "GRID-M60-0B" then we should generate an error, not simply initialize > whatever type-id 11 happens to be on this new card. > If there are 2 M60 in the system then you would find '11' type directory in mdev_supported_types of both M60. If you have P100, '11' type would not be there in its mdev_supported_types, it will have different types. For example, if you replace M60 with P100, but XML is not updated. XML have type '11'. When libvirt would try to create mdev device, libvirt would have to find 'create' file in sysfs in following directory format: --- mdev_supported_types |-- 11 | |-- create but now for P100, '11' directory is not there, so libvirt should throw error on not able to find '11' directory.

> Again, we've defined that name is unique, so why should type-id be > anything other than arbitrary? It relates only to the current instance > on the machine. > >> I think here 'name' is the problem and it is not used in xml file. >> Lets not make it mandatory. But vendor driver can have it which 3rd >> party can use to read. >> Lets remove 'id' from type id in XML if that is the concern. Supported >> types is going to be defined by vendor driver, so let vendor driver >> decide what to use for directory name and same should be used in device >> xml file, it could be '11' or "GRID M60-0B": >> >> <device> >> <name>my-vgpu</name> >> <parent>pci_0000_86_00_0</parent> >> <capability type='mdev'> >> <type='11'/> >> ... >> </capability> >> </device> >> >> >>>>>> >>>>>> 7. Hot-plug >>>>>> >>>>>> It is same syntax to create a virtual device for hot-plug. >>>>> >>>>> How do groups work with hotplug? Can a device be creating into an >>>>> existing, running group? Can a device be removed from an existing, >>>>> running group? Thanks, >>>>> >>>> >>>> Group is for vendor driver to decide when to commit resources for a >>>> group of devices to support multiple devices in a domain. It will be >>>> upto vendor driver how it manage it. >>> >>> Then it's not sufficiently specified for libvirt to use. Thanks, >>> >> >> Can you clarify what if required for libvirt to use? > > Everything we're discussing above. We can't simply say that group > management is defined by the vendor driver, that means that libvirt > needs to know how to uniquely behave for each vendor driver. We want > to create a consistent framework where libvirt doesn't care what the > vendor driver is. Therefore we need to make group manipulation fully > specified, regardless of the vendor driver. Thanks, > There is no communication between different vendor drivers. So if libvirt creates a group with one device from one vendor and one device from another vendor, both vendor driver would think they have a single device and accordingly commit resources for that single device from its own driver. Ideally nothing would fail. But that is not the main aim of grouping, right? To make this independent of vendor driver, we had initial proposal of having same UUID and different instance numbers. But that was also not acceptable.

>>>>> My concern is that a type id seems arbitrary but we're specifying that >>>>> it be unique. We already have something unique, the name. So why try >>>>> to make the type id unique as well? A vendor can accidentally create >>>>> their vendor driver so that a given name means something very >>>>> specific. On the other hand they need to be extremely deliberate to >>>>> coordinate that a type id means a unique thing across all their product >>>>> lines. >>>>> >>>> >>>> Let me clarify, type id should be unique in the list of >>>> mdev_supported_types. You can't have 2 directories in with same name. >>> >>> Of course, but does that mean it's only unique to the machine I'm >>> currently running on? Let's say I have a Tesla P100 on my system and >>> type-id 11 is named "GRID-M60-0B". At some point in the future I >>> replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that >>> new card still going to be a "GRID-M60-0B"? If not then we've based >>> our XML on the wrong attribute. If the new device does not support >>> "GRID-M60-0B" then we should generate an error, not simply initialize >>> whatever type-id 11 happens to be on this new card. >>> >> >> If there are 2 M60 in the system then you would find '11' type directory >> in mdev_supported_types of both M60. If you have P100, '11' type would >> not be there in its mdev_supported_types, it will have different types. >> >> For example, if you replace M60 with P100, but XML is not updated. XML >> have type '11'. When libvirt would try to create mdev device, libvirt >> would have to find 'create' file in sysfs in following directory format: >> >> --- mdev_supported_types >> |-- 11 >> | |-- create >> >> but now for P100, '11' directory is not there, so libvirt should throw >> error on not able to find '11' directory. > > This really seems like an accident waiting to happen. What happens > when the user replaces their M60 with an Intel XYZ device that happens > to expose a type 11 mdev class gpu device? How is libvirt supposed to > know that the XML used to refer to a GRID-M60-0B and now it's an > INTEL-IGD-XYZ? Doesn't basing the XML entry on the name and removing > yet another arbitrary requirement that we have some sort of globally > unique type-id database make a lot of sense? The same issue applies > for simple debug-ability, if I'm reviewing the XML for a domain and the > name is the primary index for the mdev device, I know what it is. > Seeing type-id='11' is meaningless. > Let me clarify again, type '11' is a string that vendor driver would define (see my previous reply below) it could be "11" or "GRID-M60-0B". If 2 vendors used same string we can't control that. right? >>>> Lets remove 'id' from type id in XML if that is the concern. Supported >>>> types is going to be defined by vendor driver, so let vendor driver >>>> decide what to use for directory name and same should be used in device >>>> xml file, it could be '11' or "GRID M60-0B": >>>> >>>> <device> >>>> <name>my-vgpu</name> >>>> <parent>pci_0000_86_00_0</parent> >>>> <capability type='mdev'> >>>> <type='11'/> >>>> ... >>>> </capability> >>>> </device>

On Thu, Sep 22, 2016 at 08:19:21AM -0600, Alex Williamson wrote: > On Thu, 22 Sep 2016 09:41:20 +0530 > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > > >>>>> My concern is that a type id seems arbitrary but we're specifying that > > >>>>> it be unique. We already have something unique, the name. So why try > > >>>>> to make the type id unique as well? A vendor can accidentally create > > >>>>> their vendor driver so that a given name means something very > > >>>>> specific. On the other hand they need to be extremely deliberate to > > >>>>> coordinate that a type id means a unique thing across all their product > > >>>>> lines. > > >>>>> > > >>>> > > >>>> Let me clarify, type id should be unique in the list of > > >>>> mdev_supported_types. You can't have 2 directories in with same name. > > >>> > > >>> Of course, but does that mean it's only unique to the machine I'm > > >>> currently running on? Let's say I have a Tesla P100 on my system and > > >>> type-id 11 is named "GRID-M60-0B". At some point in the future I > > >>> replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that > > >>> new card still going to be a "GRID-M60-0B"? If not then we've based > > >>> our XML on the wrong attribute. If the new device does not support > > >>> "GRID-M60-0B" then we should generate an error, not simply initialize > > >>> whatever type-id 11 happens to be on this new card. > > >>> > > >> > > >> If there are 2 M60 in the system then you would find '11' type directory > > >> in mdev_supported_types of both M60. If you have P100, '11' type would > > >> not be there in its mdev_supported_types, it will have different types. > > >> > > >> For example, if you replace M60 with P100, but XML is not updated. XML > > >> have type '11'. When libvirt would try to create mdev device, libvirt > > >> would have to find 'create' file in sysfs in following directory format: > > >> > > >> --- mdev_supported_types > > >> |-- 11 > > >> | |-- create > > >> > > >> but now for P100, '11' directory is not there, so libvirt should throw > > >> error on not able to find '11' directory. > > > > > > This really seems like an accident waiting to happen. What happens > > > when the user replaces their M60 with an Intel XYZ device that happens > > > to expose a type 11 mdev class gpu device? How is libvirt supposed to > > > know that the XML used to refer to a GRID-M60-0B and now it's an > > > INTEL-IGD-XYZ? Doesn't basing the XML entry on the name and removing > > > yet another arbitrary requirement that we have some sort of globally > > > unique type-id database make a lot of sense? The same issue applies > > > for simple debug-ability, if I'm reviewing the XML for a domain and the > > > name is the primary index for the mdev device, I know what it is. > > > Seeing type-id='11' is meaningless. > > > > > > > Let me clarify again, type '11' is a string that vendor driver would > > define (see my previous reply below) it could be "11" or "GRID-M60-0B". > > If 2 vendors used same string we can't control that. right? > > > > > > >>>> Lets remove 'id' from type id in XML if that is the concern. Supported > > >>>> types is going to be defined by vendor driver, so let vendor driver > > >>>> decide what to use for directory name and same should be used in device > > >>>> xml file, it could be '11' or "GRID M60-0B": > > >>>> > > >>>> <device> > > >>>> <name>my-vgpu</name> > > >>>> <parent>pci_0000_86_00_0</parent> > > >>>> <capability type='mdev'> > > >>>> <type='11'/> > > >>>> ... > > >>>> </capability> > > >>>> </device> > > Then let's get rid of the 'name' attribute and let the sysfs directory > simply be the name. Then we can get rid of 'type' altogether so we > don't have this '11' vs 'GRID-M60-0B' issue. Thanks, That sounds nice to me - we don't need two unique identifiers if one will do.

Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|

> From: Neo Jia [mailto:cjia@nvidia.com] > Sent: Thursday, September 29, 2016 3:23 AM > > On Thu, Sep 22, 2016 at 03:26:38PM +0100, Daniel P. Berrange wrote: > > On Thu, Sep 22, 2016 at 08:19:21AM -0600, Alex Williamson wrote: > > > On Thu, 22 Sep 2016 09:41:20 +0530 > > > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > > > > > > >>>>> My concern is that a type id seems arbitrary but we're specifying that > > > > >>>>> it be unique. We already have something unique, the name. So why try > > > > >>>>> to make the type id unique as well? A vendor can accidentally create > > > > >>>>> their vendor driver so that a given name means something very > > > > >>>>> specific. On the other hand they need to be extremely deliberate to > > > > >>>>> coordinate that a type id means a unique thing across all their product > > > > >>>>> lines. > > > > >>>>> > > > > >>>> > > > > >>>> Let me clarify, type id should be unique in the list of > > > > >>>> mdev_supported_types. You can't have 2 directories in with same name. > > > > >>> > > > > >>> Of course, but does that mean it's only unique to the machine I'm > > > > >>> currently running on? Let's say I have a Tesla P100 on my system and > > > > >>> type-id 11 is named "GRID-M60-0B". At some point in the future I > > > > >>> replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that > > > > >>> new card still going to be a "GRID-M60-0B"? If not then we've based > > > > >>> our XML on the wrong attribute. If the new device does not support > > > > >>> "GRID-M60-0B" then we should generate an error, not simply initialize > > > > >>> whatever type-id 11 happens to be on this new card. > > > > >>> > > > > >> > > > > >> If there are 2 M60 in the system then you would find '11' type directory > > > > >> in mdev_supported_types of both M60. If you have P100, '11' type would > > > > >> not be there in its mdev_supported_types, it will have different types. > > > > >> > > > > >> For example, if you replace M60 with P100, but XML is not updated. XML > > > > >> have type '11'. When libvirt would try to create mdev device, libvirt > > > > >> would have to find 'create' file in sysfs in following directory format: > > > > >> > > > > >> --- mdev_supported_types > > > > >> |-- 11 > > > > >> | |-- create > > > > >> > > > > >> but now for P100, '11' directory is not there, so libvirt should throw > > > > >> error on not able to find '11' directory. > > > > > > > > > > This really seems like an accident waiting to happen. What happens > > > > > when the user replaces their M60 with an Intel XYZ device that happens > > > > > to expose a type 11 mdev class gpu device? How is libvirt supposed to > > > > > know that the XML used to refer to a GRID-M60-0B and now it's an > > > > > INTEL-IGD-XYZ? Doesn't basing the XML entry on the name and removing > > > > > yet another arbitrary requirement that we have some sort of globally > > > > > unique type-id database make a lot of sense? The same issue applies > > > > > for simple debug-ability, if I'm reviewing the XML for a domain and the > > > > > name is the primary index for the mdev device, I know what it is. > > > > > Seeing type-id='11' is meaningless. > > > > > > > > > > > > > Let me clarify again, type '11' is a string that vendor driver would > > > > define (see my previous reply below) it could be "11" or "GRID-M60-0B". > > > > If 2 vendors used same string we can't control that. right? > > > > > > > > > > > > >>>> Lets remove 'id' from type id in XML if that is the concern. Supported > > > > >>>> types is going to be defined by vendor driver, so let vendor driver > > > > >>>> decide what to use for directory name and same should be used in device > > > > >>>> xml file, it could be '11' or "GRID M60-0B": > > > > >>>> > > > > >>>> <device> > > > > >>>> <name>my-vgpu</name> > > > > >>>> <parent>pci_0000_86_00_0</parent> > > > > >>>> <capability type='mdev'> > > > > >>>> <type='11'/> > > > > >>>> ... > > > > >>>> </capability> > > > > >>>> </device> > > > > > > Then let's get rid of the 'name' attribute and let the sysfs directory > > > simply be the name. Then we can get rid of 'type' altogether so we > > > don't have this '11' vs 'GRID-M60-0B' issue. Thanks, > > > > That sounds nice to me - we don't need two unique identifiers if > > one will do. > > Hi Alex and Daniel, > > I just had some internal discussions here within NVIDIA and found out that > actually the name/label potentially might not be unique and the "id" will be. > So I think we still would like to keep both so the id is the programmatic id > and the name/label is a human readable string for it, which might get changed to > be non-unique by outside of engineering. > > Sorry for the change. > > Thanks, > Neo > A curious question. How do we expect such a descriptive name/label used by upper-level stack (e.g. openstack)? Should openstack define a vGPU flavor just using ID (GRID-type11) or using both ID/name (GRID-type11- M60-0B) for end customer to choose? If it's only for human information, does it make sense e.g. providing only unique ID in sysfs while relying on vendor specific documentation to describe what the ID actually means?

Thanks, Kevin

On 09/28/2016 03:59 PM, Neo Jia wrote: > On Wed, Sep 28, 2016 at 07:45:38PM +0000, Tian, Kevin wrote: > > > From: Neo Jia [mailto:cjia@nvidia.com] > > > Sent: Thursday, September 29, 2016 3:23 AM > > > > > > On Thu, Sep 22, 2016 at 03:26:38PM +0100, Daniel P. Berrange wrote: > > > > On Thu, Sep 22, 2016 at 08:19:21AM -0600, Alex Williamson wrote: > > > > > On Thu, 22 Sep 2016 09:41:20 +0530 > > > > > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > > > > > > > > > > > > > > > My concern is that a type id seems arbitrary but we're specifying that > > > > > > > > > > > it be unique. We already have something unique, the name. So why try > > > > > > > > > > > to make the type id unique as well? A vendor can accidentally create > > > > > > > > > > > their vendor driver so that a given name means something very > > > > > > > > > > > specific. On the other hand they need to be extremely deliberate to > > > > > > > > > > > coordinate that a type id means a unique thing across all their product > > > > > > > > > > > lines. > > > > > > > > > > > > > > > > > > > > > Let me clarify, type id should be unique in the list of > > > > > > > > > > mdev_supported_types. You can't have 2 directories in with same name. > > > > > > > > > Of course, but does that mean it's only unique to the machine I'm > > > > > > > > > currently running on? Let's say I have a Tesla P100 on my system and > > > > > > > > > type-id 11 is named "GRID-M60-0B". At some point in the future I > > > > > > > > > replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that > > > > > > > > > new card still going to be a "GRID-M60-0B"? If not then we've based > > > > > > > > > our XML on the wrong attribute. If the new device does not support > > > > > > > > > "GRID-M60-0B" then we should generate an error, not simply initialize > > > > > > > > > whatever type-id 11 happens to be on this new card. > > > > > > > > > > > > > > > > > If there are 2 M60 in the system then you would find '11' type directory > > > > > > > > in mdev_supported_types of both M60. If you have P100, '11' type would > > > > > > > > not be there in its mdev_supported_types, it will have different types. > > > > > > > > > > > > > > > > For example, if you replace M60 with P100, but XML is not updated. XML > > > > > > > > have type '11'. When libvirt would try to create mdev device, libvirt > > > > > > > > would have to find 'create' file in sysfs in following directory format: > > > > > > > > > > > > > > > > --- mdev_supported_types > > > > > > > > |-- 11 > > > > > > > > | |-- create > > > > > > > > > > > > > > > > but now for P100, '11' directory is not there, so libvirt should throw > > > > > > > > error on not able to find '11' directory. > > > > > > > This really seems like an accident waiting to happen. What happens > > > > > > > when the user replaces their M60 with an Intel XYZ device that happens > > > > > > > to expose a type 11 mdev class gpu device? How is libvirt supposed to > > > > > > > know that the XML used to refer to a GRID-M60-0B and now it's an > > > > > > > INTEL-IGD-XYZ? Doesn't basing the XML entry on the name and removing > > > > > > > yet another arbitrary requirement that we have some sort of globally > > > > > > > unique type-id database make a lot of sense? The same issue applies > > > > > > > for simple debug-ability, if I'm reviewing the XML for a domain and the > > > > > > > name is the primary index for the mdev device, I know what it is. > > > > > > > Seeing type-id='11' is meaningless. > > > > > > > > > > > > > Let me clarify again, type '11' is a string that vendor driver would > > > > > > define (see my previous reply below) it could be "11" or "GRID-M60-0B". > > > > > > If 2 vendors used same string we can't control that. right? > > > > > > > > > > > > > > > > > > > > > > Lets remove 'id' from type id in XML if that is the concern. Supported > > > > > > > > > > types is going to be defined by vendor driver, so let vendor driver > > > > > > > > > > decide what to use for directory name and same should be used in device > > > > > > > > > > xml file, it could be '11' or "GRID M60-0B": > > > > > > > > > > > > > > > > > > > > <device> > > > > > > > > > > <name>my-vgpu</name> > > > > > > > > > > <parent>pci_0000_86_00_0</parent> > > > > > > > > > > <capability type='mdev'> > > > > > > > > > > <type='11'/> > > > > > > > > > > ... > > > > > > > > > > </capability> > > > > > > > > > > </device> > > > > > Then let's get rid of the 'name' attribute and let the sysfs directory > > > > > simply be the name. Then we can get rid of 'type' altogether so we > > > > > don't have this '11' vs 'GRID-M60-0B' issue. Thanks, > > > > That sounds nice to me - we don't need two unique identifiers if > > > > one will do. > > > Hi Alex and Daniel, > > > > > > I just had some internal discussions here within NVIDIA and found out that > > > actually the name/label potentially might not be unique and the "id" will be. My comment below follows the above statement^^^^. > > > So I think we still would like to keep both so the id is the programmatic id > > > and the name/label is a human readable string for it, which might get changed to > > > be non-unique by outside of engineering. > > > > > > Sorry for the change. > > > > > > Thanks, > > > Neo > > > > > A curious question. How do we expect such a descriptive name/label used > > by upper-level stack (e.g. openstack)? Should openstack define a vGPU > > flavor just using ID (GRID-type11) or using both ID/name (GRID-type11- > > M60-0B) for end customer to choose? If it's only for human information, > > does it make sense e.g. providing only unique ID in sysfs while relying on > > vendor specific documentation to describe what the ID actually means? > Hi Kevin, > > The id is not visible to the upper-level stack, only the name / label will be > shown to the end customer to choose, such as "GRID-M60-0B", as we might expose > the same virtual device (name/label) with some internal difference which will > be tracked by the different unique id. If the upper layer will only see the descriptive name/label, then that label must be unique. It's not acceptable for a particular key used by management software to sometimes lead to one flavor of device and sometimes another (no matter how small the differences may be). So if only the ID is unique, then the ID is what must be used in any configuration at any level.

> > I think having the ability to allow libvirt or upper-level stack to display a > human readable string for a given type of vgpu will make the user life easier. > > Thanks, > Neo > > > Thanks, > > Kevin > > > > > -- > libvir-list mailing list > libvir-list(a)redhat.com > https://www.redhat.com/mailman/listinfo/libvir-list >

On Thu, Sep 22, 2016 at 03:26:38PM +0100, Daniel P. Berrange wrote: > On Thu, Sep 22, 2016 at 08:19:21AM -0600, Alex Williamson wrote: > > On Thu, 22 Sep 2016 09:41:20 +0530 > > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > > > > >>>>> My concern is that a type id seems arbitrary but we're specifying that > > > >>>>> it be unique. We already have something unique, the name. So why try > > > >>>>> to make the type id unique as well? A vendor can accidentally create > > > >>>>> their vendor driver so that a given name means something very > > > >>>>> specific. On the other hand they need to be extremely deliberate to > > > >>>>> coordinate that a type id means a unique thing across all their product > > > >>>>> lines. > > > >>>>> > > > >>>> > > > >>>> Let me clarify, type id should be unique in the list of > > > >>>> mdev_supported_types. You can't have 2 directories in with same name. > > > >>> > > > >>> Of course, but does that mean it's only unique to the machine I'm > > > >>> currently running on? Let's say I have a Tesla P100 on my system and > > > >>> type-id 11 is named "GRID-M60-0B". At some point in the future I > > > >>> replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that > > > >>> new card still going to be a "GRID-M60-0B"? If not then we've based > > > >>> our XML on the wrong attribute. If the new device does not support > > > >>> "GRID-M60-0B" then we should generate an error, not simply initialize > > > >>> whatever type-id 11 happens to be on this new card. > > > >>> > > > >> > > > >> If there are 2 M60 in the system then you would find '11' type directory > > > >> in mdev_supported_types of both M60. If you have P100, '11' type would > > > >> not be there in its mdev_supported_types, it will have different types. > > > >> > > > >> For example, if you replace M60 with P100, but XML is not updated. XML > > > >> have type '11'. When libvirt would try to create mdev device, libvirt > > > >> would have to find 'create' file in sysfs in following directory format: > > > >> > > > >> --- mdev_supported_types > > > >> |-- 11 > > > >> | |-- create > > > >> > > > >> but now for P100, '11' directory is not there, so libvirt should throw > > > >> error on not able to find '11' directory. > > > > > > > > This really seems like an accident waiting to happen. What happens > > > > when the user replaces their M60 with an Intel XYZ device that happens > > > > to expose a type 11 mdev class gpu device? How is libvirt supposed to > > > > know that the XML used to refer to a GRID-M60-0B and now it's an > > > > INTEL-IGD-XYZ? Doesn't basing the XML entry on the name and removing > > > > yet another arbitrary requirement that we have some sort of globally > > > > unique type-id database make a lot of sense? The same issue applies > > > > for simple debug-ability, if I'm reviewing the XML for a domain and the > > > > name is the primary index for the mdev device, I know what it is. > > > > Seeing type-id='11' is meaningless. > > > > > > > > > > Let me clarify again, type '11' is a string that vendor driver would > > > define (see my previous reply below) it could be "11" or "GRID-M60-0B". > > > If 2 vendors used same string we can't control that. right? > > > > > > > > > >>>> Lets remove 'id' from type id in XML if that is the concern. Supported > > > >>>> types is going to be defined by vendor driver, so let vendor driver > > > >>>> decide what to use for directory name and same should be used in device > > > >>>> xml file, it could be '11' or "GRID M60-0B": > > > >>>> > > > >>>> <device> > > > >>>> <name>my-vgpu</name> > > > >>>> <parent>pci_0000_86_00_0</parent> > > > >>>> <capability type='mdev'> > > > >>>> <type='11'/> > > > >>>> ... > > > >>>> </capability> > > > >>>> </device> > > > > Then let's get rid of the 'name' attribute and let the sysfs directory > > simply be the name. Then we can get rid of 'type' altogether so we > > don't have this '11' vs 'GRID-M60-0B' issue. Thanks, > > That sounds nice to me - we don't need two unique identifiers if > one will do. Hi Alex and Daniel, I just had some internal discussions here within NVIDIA and found out that actually the name/label potentially might not be unique and the "id" will be. So I think we still would like to keep both so the id is the programmatic id and the name/label is a human readable string for it, which might get changed to be non-unique by outside of engineering.

On Wed, Sep 28, 2016 at 01:55:47PM -0600, Alex Williamson wrote: > On Wed, 28 Sep 2016 12:22:35 -0700 > Neo Jia <cjia(a)nvidia.com&gt; wrote: > > > On Thu, Sep 22, 2016 at 03:26:38PM +0100, Daniel P. Berrange wrote: > > > On Thu, Sep 22, 2016 at 08:19:21AM -0600, Alex Williamson wrote: > > > > On Thu, 22 Sep 2016 09:41:20 +0530 > > > > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > > > > > > > > >>>>> My concern is that a type id seems arbitrary but we're specifying that > > > > > >>>>> it be unique. We already have something unique, the name. So why try > > > > > >>>>> to make the type id unique as well? A vendor can accidentally create > > > > > >>>>> their vendor driver so that a given name means something very > > > > > >>>>> specific. On the other hand they need to be extremely deliberate to > > > > > >>>>> coordinate that a type id means a unique thing across all their product > > > > > >>>>> lines. > > > > > >>>>> > > > > > >>>> > > > > > >>>> Let me clarify, type id should be unique in the list of > > > > > >>>> mdev_supported_types. You can't have 2 directories in with same name. > > > > > >>> > > > > > >>> Of course, but does that mean it's only unique to the machine I'm > > > > > >>> currently running on? Let's say I have a Tesla P100 on my system and > > > > > >>> type-id 11 is named "GRID-M60-0B". At some point in the future I > > > > > >>> replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that > > > > > >>> new card still going to be a "GRID-M60-0B"? If not then we've based > > > > > >>> our XML on the wrong attribute. If the new device does not support > > > > > >>> "GRID-M60-0B" then we should generate an error, not simply initialize > > > > > >>> whatever type-id 11 happens to be on this new card. > > > > > >>> > > > > > >> > > > > > >> If there are 2 M60 in the system then you would find '11' type directory > > > > > >> in mdev_supported_types of both M60. If you have P100, '11' type would > > > > > >> not be there in its mdev_supported_types, it will have different types. > > > > > >> > > > > > >> For example, if you replace M60 with P100, but XML is not updated. XML > > > > > >> have type '11'. When libvirt would try to create mdev device, libvirt > > > > > >> would have to find 'create' file in sysfs in following directory format: > > > > > >> > > > > > >> --- mdev_supported_types > > > > > >> |-- 11 > > > > > >> | |-- create > > > > > >> > > > > > >> but now for P100, '11' directory is not there, so libvirt should throw > > > > > >> error on not able to find '11' directory. > > > > > > > > > > > > This really seems like an accident waiting to happen. What happens > > > > > > when the user replaces their M60 with an Intel XYZ device that happens > > > > > > to expose a type 11 mdev class gpu device? How is libvirt supposed to > > > > > > know that the XML used to refer to a GRID-M60-0B and now it's an > > > > > > INTEL-IGD-XYZ? Doesn't basing the XML entry on the name and removing > > > > > > yet another arbitrary requirement that we have some sort of globally > > > > > > unique type-id database make a lot of sense? The same issue applies > > > > > > for simple debug-ability, if I'm reviewing the XML for a domain and the > > > > > > name is the primary index for the mdev device, I know what it is. > > > > > > Seeing type-id='11' is meaningless. > > > > > > > > > > > > > > > > Let me clarify again, type '11' is a string that vendor driver would > > > > > define (see my previous reply below) it could be "11" or "GRID-M60-0B". > > > > > If 2 vendors used same string we can't control that. right? > > > > > > > > > > > > > > > >>>> Lets remove 'id' from type id in XML if that is the concern. Supported > > > > > >>>> types is going to be defined by vendor driver, so let vendor driver > > > > > >>>> decide what to use for directory name and same should be used in device > > > > > >>>> xml file, it could be '11' or "GRID M60-0B": > > > > > >>>> > > > > > >>>> <device> > > > > > >>>> <name>my-vgpu</name> > > > > > >>>> <parent>pci_0000_86_00_0</parent> > > > > > >>>> <capability type='mdev'> > > > > > >>>> <type='11'/> > > > > > >>>> ... > > > > > >>>> </capability> > > > > > >>>> </device> > > > > > > > > Then let's get rid of the 'name' attribute and let the sysfs directory > > > > simply be the name. Then we can get rid of 'type' altogether so we > > > > don't have this '11' vs 'GRID-M60-0B' issue. Thanks, > > > > > > That sounds nice to me - we don't need two unique identifiers if > > > one will do. > > > > Hi Alex and Daniel, > > > > I just had some internal discussions here within NVIDIA and found out that > > actually the name/label potentially might not be unique and the "id" will be. > > So I think we still would like to keep both so the id is the programmatic id > > and the name/label is a human readable string for it, which might get changed to > > be non-unique by outside of engineering. > > I think your discovery only means that for your vendor driver, the name > will be "11" (as a string). Perhaps you'd like some sort of vendor > provided description within each type, but I am not in favor of having > an arbitrary integer value imply something specific within the sysfs > interface. IOW, the NVIDIA vendor driver should be able to create: > > 11 > ├── create > ├── description > ├── etc > └── resolution > > While Intel might create: > > Skylake-vGPU > ├── create > ├── description > ├── etc > └── resolution > > Maybe "description" is optional for vendors that use useful names? > Thanks, I think we should be able to have a unique vendor type string instead of an arbitrary integer value there as long as we are allowed to have a description field that can be used to show to the end user as "name / label".

On Wed, Sep 28, 2016 at 12:22:35PM -0700, Neo Jia wrote: > On Thu, Sep 22, 2016 at 03:26:38PM +0100, Daniel P. Berrange wrote: > > On Thu, Sep 22, 2016 at 08:19:21AM -0600, Alex Williamson wrote: > > > On Thu, 22 Sep 2016 09:41:20 +0530 > > > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > > > > > > >>>>> My concern is that a type id seems arbitrary but we're specifying that > > > > >>>>> it be unique. We already have something unique, the name. So why try > > > > >>>>> to make the type id unique as well? A vendor can accidentally create > > > > >>>>> their vendor driver so that a given name means something very > > > > >>>>> specific. On the other hand they need to be extremely deliberate to > > > > >>>>> coordinate that a type id means a unique thing across all their product > > > > >>>>> lines. > > > > >>>>> > > > > >>>> > > > > >>>> Let me clarify, type id should be unique in the list of > > > > >>>> mdev_supported_types. You can't have 2 directories in with same name. > > > > >>> > > > > >>> Of course, but does that mean it's only unique to the machine I'm > > > > >>> currently running on? Let's say I have a Tesla P100 on my system and > > > > >>> type-id 11 is named "GRID-M60-0B". At some point in the future I > > > > >>> replace the Tesla P100 with a Q1000 (made up). Is type-id 11 on that > > > > >>> new card still going to be a "GRID-M60-0B"? If not then we've based > > > > >>> our XML on the wrong attribute. If the new device does not support > > > > >>> "GRID-M60-0B" then we should generate an error, not simply initialize > > > > >>> whatever type-id 11 happens to be on this new card. > > > > >>> > > > > >> > > > > >> If there are 2 M60 in the system then you would find '11' type directory > > > > >> in mdev_supported_types of both M60. If you have P100, '11' type would > > > > >> not be there in its mdev_supported_types, it will have different types. > > > > >> > > > > >> For example, if you replace M60 with P100, but XML is not updated. XML > > > > >> have type '11'. When libvirt would try to create mdev device, libvirt > > > > >> would have to find 'create' file in sysfs in following directory format: > > > > >> > > > > >> --- mdev_supported_types > > > > >> |-- 11 > > > > >> | |-- create > > > > >> > > > > >> but now for P100, '11' directory is not there, so libvirt should throw > > > > >> error on not able to find '11' directory. > > > > > > > > > > This really seems like an accident waiting to happen. What happens > > > > > when the user replaces their M60 with an Intel XYZ device that happens > > > > > to expose a type 11 mdev class gpu device? How is libvirt supposed to > > > > > know that the XML used to refer to a GRID-M60-0B and now it's an > > > > > INTEL-IGD-XYZ? Doesn't basing the XML entry on the name and removing > > > > > yet another arbitrary requirement that we have some sort of globally > > > > > unique type-id database make a lot of sense? The same issue applies > > > > > for simple debug-ability, if I'm reviewing the XML for a domain and the > > > > > name is the primary index for the mdev device, I know what it is. > > > > > Seeing type-id='11' is meaningless. > > > > > > > > > > > > > Let me clarify again, type '11' is a string that vendor driver would > > > > define (see my previous reply below) it could be "11" or "GRID-M60-0B". > > > > If 2 vendors used same string we can't control that. right? > > > > > > > > > > > > >>>> Lets remove 'id' from type id in XML if that is the concern. Supported > > > > >>>> types is going to be defined by vendor driver, so let vendor driver > > > > >>>> decide what to use for directory name and same should be used in device > > > > >>>> xml file, it could be '11' or "GRID M60-0B": > > > > >>>> > > > > >>>> <device> > > > > >>>> <name>my-vgpu</name> > > > > >>>> <parent>pci_0000_86_00_0</parent> > > > > >>>> <capability type='mdev'> > > > > >>>> <type='11'/> > > > > >>>> ... > > > > >>>> </capability> > > > > >>>> </device> > > > > > > Then let's get rid of the 'name' attribute and let the sysfs directory > > > simply be the name. Then we can get rid of 'type' altogether so we > > > don't have this '11' vs 'GRID-M60-0B' issue. Thanks, > > > > That sounds nice to me - we don't need two unique identifiers if > > one will do. > > Hi Alex and Daniel, > > I just had some internal discussions here within NVIDIA and found out that > actually the name/label potentially might not be unique and the "id" will be. > So I think we still would like to keep both so the id is the programmatic id > and the name/label is a human readable string for it, which might get changed to > be non-unique by outside of engineering. The name is what the end user will used to decide which type of device to create. If the name is not unique, how is the user supposed to decide which of the duplicates to use. Allowing duplicate names just sounds broken to me.

Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://entangle-photo.org -o- http://search.cpan.org/~danberr/ :|

From: Kirti Wankhede [mailto:kwankhede@nvidia.com] Sent: Wednesday, September 21, 2016 12:23 AM On 9/20/2016 8:13 PM, Alex Williamson wrote: > On Tue, 20 Sep 2016 19:51:58 +0530 > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > >> On 9/20/2016 3:06 AM, Alex Williamson wrote: >>> On Tue, 20 Sep 2016 02:05:52 +0530 >>> Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: >>> >>>> Hi libvirt experts, >>>> >>>> >>>> 'create': >>>> Write-only file. Mandatory. >>>> Accepts string to create mediated device. >>>> >>>> 'name': >>>> Read-Only file. Mandatory. >>>> Returns string, the name of that type id. >>>> >>>> 'fb_length': >>>> Read-only file. Mandatory. >>>> Returns <number>{K,M,G}, size of framebuffer. >>> >>> This can't be mandatory, it's only relevant to vGPU devices, vGPUs are >>> just one user of mediated devices. >>> >> >> As per our discussion in BOF, we would define class and its attributes. >> As I mentioned above these are attributes of "display" class. > > Just as I didn't know here, how does libvirt know the class of a given > type id? > Yes, we need some way to identify the class as Daniel mentioned in his suggestion. Add a file 'class' in mdev_supported_types that would return the class.

From: Alex Williamson [mailto:alex.williamson@redhat.com] Sent: Wednesday, September 21, 2016 12:43 PM On Wed, 21 Sep 2016 04:10:53 +0000 "Tian, Kevin" <kevin.tian(a)intel.com&gt; wrote: > > From: Kirti Wankhede [mailto:kwankhede@nvidia.com] > > Sent: Wednesday, September 21, 2016 12:23 AM > > > > > > On 9/20/2016 8:13 PM, Alex Williamson wrote: > > > On Tue, 20 Sep 2016 19:51:58 +0530 > > > Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > > > > >> On 9/20/2016 3:06 AM, Alex Williamson wrote: > > >>> On Tue, 20 Sep 2016 02:05:52 +0530 > > >>> Kirti Wankhede <kwankhede(a)nvidia.com&gt; wrote: > > >>> > > >>>> Hi libvirt experts, > > >>>> > > >>>> > > >>>> 'create': > > >>>> Write-only file. Mandatory. > > >>>> Accepts string to create mediated device. > > >>>> > > >>>> 'name': > > >>>> Read-Only file. Mandatory. > > >>>> Returns string, the name of that type id. > > >>>> > > >>>> 'fb_length': > > >>>> Read-only file. Mandatory. > > >>>> Returns <number>{K,M,G}, size of framebuffer. > > >>> > > >>> This can't be mandatory, it's only relevant to vGPU devices, vGPUs are > > >>> just one user of mediated devices. > > >>> > > >> > > >> As per our discussion in BOF, we would define class and its attributes. > > >> As I mentioned above these are attributes of "display" class. > > > > > > Just as I didn't know here, how does libvirt know the class of a given > > > type id? > > > > > > > Yes, we need some way to identify the class as Daniel mentioned in his > > suggestion. Add a file 'class' in mdev_supported_types that would return > > the class. > > > > > > 'display' class or 'gpu' class? display represents only part of GPU > functionalities. I assume you want to provide an unique class ID > for each type, instead of allowing multiple classes each identifying > a subset of controllable GPU resources (e.g. 'display', 'render', etc.)... Not sure where you're going with this, we're not creating a programming interface for the device, that exists through the vfio API. I believe the intention here is simply a user level classification for the purpose of being able to interpret the attributes provided in a logical way.

> for unique class ID, I once considered whether it could be directly > inherited from class of parent device, since typically a vendor driver > creates mdev devices using the same type as physical device. But later > I realized one potential value of keep a class field for mdev types, > e.g. if we want to extend mdev to FPGA which could be programmed > as different classes of functionalities. :-) And how do we generically determine the class of a parent device? We cannot assume the parent device is PCI. Thanks,

> From: Kirti Wankhede [mailto:kwankhede@nvidia.com] > Sent: Wednesday, September 21, 2016 12:23 AM >> >>>> I have >>>> a hard time believing that a given vendor can even allocate unique type >>>> ids for their own devices. Unique type id across vendors is not >>>> practical. So which attribute are we actually using to identify which >>>> type of mdev device we need and why would we ever specify additional >>>> attributes like fb_length? Doesn't the vendor guarantee that "GRID >>>> M60-0B" has a fixed setup of those attributes? >>>> >>> >>> Specifying attributes here is not our requirement. Yes we have fixed set >>> of attributes for "GRID-M60-0B" and on. >>> We are defining the attributed here for "display" class for all other >>> vendor of gpu can use. >>> Hi, Kirti, We decide to go with above type-based interface for KVMGT, with fixed setup of attributes too. If both Intel and NVIDIA solutions use such fixed manner, can we go with a proposal which doesn't include 'class' extension for now? Later if there is a concrete usage requiring such class-specific attribute setting, then it's easy to extend following discussion in this thread. I'm thinking how we can converge the discussion here into something simple enough (and extensible) to accelerate upstreaming of both Intel/NVIDIA solutions...

> From: Kirti Wankhede [mailto:kwankhede@nvidia.com] > Sent: Tuesday, September 20, 2016 10:22 PM > >> > >> 'max_instance': > >> Read-Only file. Mandatory. > >> Returns integer. Returns maximum mdev device could be created > >> at the moment when this file is read. This count would be updated by > >> vendor driver. Before creating mdev device of this type, check if > >> max_instance is > 0. > > > > Didn't we discuss this being being something like "available_instances" > > with a "devices" sub-directory linking current devices of that type? > > > > I'm ok with 'available_instances' as name of this file. > Yes, mdev device directory will have links to devices of that type. I > think that is part of mdev core module discussion. I'm trying to list > sysfs files for libvirt interface here to focus more on libvirt > interface. Hence didn't add that. I'll make sure to add all such details > in future. > > Definitely you need provide those details since they also contain libvirt interface, e.g. 'destroy' which we discussed should be in mdev directory.

Another example is class-specific attributes such as 'resolution', etc. which you listed under 'display class'. All those attributes should be moved to mdev directory. Only class ID is required under each type.

> From: Alex Williamson [mailto:alex.williamson@redhat.com] > Sent: Wednesday, September 21, 2016 12:37 PM > > On Wed, 21 Sep 2016 03:56:21 +0000 > "Tian, Kevin" <kevin.tian(a)intel.com&gt; wrote: > > > > From: Kirti Wankhede [mailto:kwankhede@nvidia.com] > > > Sent: Tuesday, September 20, 2016 10:22 PM > > > >> > > > >> 'max_instance': > > > >> Read-Only file. Mandatory. > > > >> Returns integer. Returns maximum mdev device could be created > > > >> at the moment when this file is read. This count would be updated by > > > >> vendor driver. Before creating mdev device of this type, check if > > > >> max_instance is > 0. > > > > > > > > Didn't we discuss this being being something like "available_instances" > > > > with a "devices" sub-directory linking current devices of that type? > > > > > > > > > > I'm ok with 'available_instances' as name of this file. > > > Yes, mdev device directory will have links to devices of that type. I > > > think that is part of mdev core module discussion. I'm trying to list > > > sysfs files for libvirt interface here to focus more on libvirt > > > interface. Hence didn't add that. I'll make sure to add all such details > > > in future. > > > > > > > > > > Definitely you need provide those details since they also contain > > libvirt interface, e.g. 'destroy' which we discussed should be in > > mdev directory. > > $ find /sys/devices -name remove | wc -l > 24 > $ find /sys/devices -name destroy | wc -l > 0 > > 'remove' is the sysfs deconstructor we should use. make sense. > > > Another example is class-specific attributes such > > as 'resolution', etc. which you listed under 'display class'. All those > > attributes should be moved to mdev directory. Only class ID is > > required under each type. > > Can you elaborate on what you're proposing here? If we don't have > attributes like 'resolution' under the type-id then we can't describe > to the user the features of the mdev before we create it. I don't > think anybody wants a 'create it and find out' type of interface. > Thanks, > I think such way would be racy. What about two clients creating mdev devices simultaneously? How to guarantee their configuration of class specific attributes not mutual-impacted since before creation any such configuration would be global under that type?

My feeling is that we'd better keep create simple - just write a UUID to "type-id/create". Any class-specific attribute, if we really want to support, should be able to be individually configured with required resource allocated incrementally on top of basic resources allocated at create stage. Then libvirt can set those attributes between create and open a mdev device. If this direction is accepted, then naturally such attributes should be put under mdev directory. Possibly we don't need a class description under type-id. libvirt just checks directly whether any known class represented in each mdev directory (vendor driver will expose it on demand), and then operate attributes under that class.

From: Kirti Wankhede [mailto:kwankhede@nvidia.com] Sent: Tuesday, September 20, 2016 4:36 AM Hi libvirt experts, Thanks for valuable input on v1 version of RFC. Quick brief, VFIO based mediated device framework provides a way to virtualize their devices without SR-IOV, like NVIDIA vGPU, Intel KVMGT and IBM's channel IO. This framework reuses VFIO APIs for all the functionalities for mediated devices which are currently being used for pass through devices. This framework introduces a set of new sysfs files for device creation and its life cycle management. Here is the summary of discussion on v1: 1. Discover mediated device: As part of physical device initialization process, vendor driver will register their physical devices, which will be used to create virtual device (mediated device, aka mdev) to the mediated framework. Vendor driver should specify mdev_supported_types in directory format. This format is class based, for example, display class directory format should be as below. We need to define such set for each class of devices which would be supported by mediated device framework. --- mdev_destroy --- mdev_supported_types |-- 11 | |-- create | |-- name | |-- fb_length | |-- resolution | |-- heads | |-- max_instances | |-- params | |-- requires_group |-- 12 | |-- create | |-- name | |-- fb_length | |-- resolution | |-- heads | |-- max_instances | |-- params | |-- requires_group |-- 13 |-- create |-- name |-- fb_length |-- resolution |-- heads |-- max_instances |-- params |-- requires_group

On Tue, Sep 20, 2016 at 02:05:52AM +0530, Kirti Wankhede wrote: > > Hi libvirt experts, > > Thanks for valuable input on v1 version of RFC. > > Quick brief, VFIO based mediated device framework provides a way to > virtualize their devices without SR-IOV, like NVIDIA vGPU, Intel KVMGT > and IBM's channel IO. This framework reuses VFIO APIs for all the > functionalities for mediated devices which are currently being used for > pass through devices. This framework introduces a set of new sysfs files > for device creation and its life cycle management. > > Here is the summary of discussion on v1: > 1. Discover mediated device: > As part of physical device initialization process, vendor driver will > register their physical devices, which will be used to create virtual > device (mediated device, aka mdev) to the mediated framework. > > Vendor driver should specify mdev_supported_types in directory format. > This format is class based, for example, display class directory format > should be as below. We need to define such set for each class of devices > which would be supported by mediated device framework. > > --- mdev_destroy > --- mdev_supported_types > |-- 11 > | |-- create > | |-- name > | |-- fb_length > | |-- resolution > | |-- heads > | |-- max_instances > | |-- params > | |-- requires_group > |-- 12 > | |-- create > | |-- name > | |-- fb_length > | |-- resolution > | |-- heads > | |-- max_instances > | |-- params > | |-- requires_group > |-- 13 > |-- create > |-- name > |-- fb_length > |-- resolution > |-- heads > |-- max_instances > |-- params > |-- requires_group > > > In the above example directory '11' represents a type id of mdev device. > 'name', 'fb_length', 'resolution', 'heads', 'max_instance' and > 'requires_group' would be Read-Only files that vendor would provide to > describe about that type. > > 'create': > Write-only file. Mandatory. > Accepts string to create mediated device. > > 'name': > Read-Only file. Mandatory. > Returns string, the name of that type id. Presumably this is a human-targetted title/description of the device. > > 'fb_length': > Read-only file. Mandatory. > Returns <number>{K,M,G}, size of framebuffer. > > 'resolution': > Read-Only file. Mandatory. > Returns 'hres x vres' format. Maximum supported resolution. > > 'heads': > Read-Only file. Mandatory. > Returns integer. Number of maximum heads supported. None of these should be mandatory as that makes the mdev useless for non-GPU devices. I'd expect to see a 'class' or 'type' attribute in the directory whcih tells you what kind of mdev it is. A valid 'class' value would be 'gpu'. The fb_length, resolution, and heads parameters would only be mandatory when class==gpu.

> 'max_instance': > Read-Only file. Mandatory. > Returns integer. Returns maximum mdev device could be created > at the moment when this file is read. This count would be updated by > vendor driver. Before creating mdev device of this type, check if > max_instance is > 0. > > 'params' > Write-Only file. Optional. > String input. Libvirt would pass the string given in XML file to > this file and then create mdev device. Set empty string to clear params. > For example, set parameter 'frame_rate_limiter=0' to disable frame rate > limiter for performance benchmarking, then create device of type 11. The > device created would have that parameter set by vendor driver. Nope, libvirt will explicitly *NEVER* allow arbitrary opaque passthrough of vendor specific data in this way. > The parent device would look like: > > <device> > <name>pci_0000_86_00_0</name> > <capability type='pci'> > <domain>0</domain> > <bus>134</bus> > <slot>0</slot> > <function>0</function> > <capability type='mdev'> > <!-- one type element per sysfs directory --> > <type id='11'> > <!-- one element per sysfs file roughly --> > <name>GRID M60-0B</name> > <attribute name='fb_length'>512M</attribute> > <attribute name='resolution'>2560x1600</attribute> > <attribute name='heads'>2</attribute> > <attribute name='max_instances'>16</attribute> > <attribute name='requires_group'>1</attribute> > </type> There would need to be a <class> element, eg <class>gpu</class> We would then have further elements based on the class. eg <type id='11'> <!-- one element per sysfs file roughly --> <name>GRID M60-0B</name> <fb_length>512M</fb_length> <resolution>2560x1600</resolution> <heads>2</heads> <max_instances>16</max_instances> <requires_group>1</requires_group> </type> > </capability> > <product id='...'>GRID M60</product> > <vendor id='0x10de'>NVIDIA</vendor> > </capability> > </device> > > 2. Create/destroy mediated device > > With above example, vGPU device XML would look like: > > <device> > <name>my-vgpu</name> > <parent>pci_0000_86_00_0</parent> > <capability type='mdev'> > <type id='11'/> > <group>1</group> > <params>'frame_rate_limiter=0'</params> No, we will not support <params> in this manner in libvirt. The entire purpose of libvirt is to represent data in a vendor agnostic manner and not do abitrary passthrough of vendor specific data. Simply saying this field is optional does not get around that either. > </capability> > </device> > > 'type id' is mandatory. > 'group' is optional. It should be a unique number in the system among > all the groups created for mdev devices. Its usage is: > - not needed if single vGPU device is being assigned to a domain. > - only need to be set if multiple vGPUs need to be assigned to a > domain and vendor driver have 'requires_group' file in type id directory. > - if type id directory include 'requires_group' and user tries to > assign multiple vGPUs to a domain without having <group> field in XML, > it will create single vGPU. > > 'params' is optional field. User should set this field if extra > parameters need to be set for a particular vGPU device. Libvirt don't > need to parse these params. These are meant for vendor driver. > > Libvirt need to follow the sequence to create device: > * Read /sys/../0000\:86\:00.0/11/max_instances. If it is greater than 0, > then only proceed else fail. > > * Set extra params if 'params' field exist in device XML and 'params' > file exist in type id directory > > echo "frame_rate_limiter=0" > /sys/../0000\:86\:00.0/11/params We cannot do that step. > > * Autogenerate UUID > * Create device: > > echo "$UUID:<group>" > /sys/../0000\:86\:00.0/11/create > > where <group> is optional. Group should be unique number among all > the groups created for mdev devices. > > * Clear params, if set earlier: > > echo "" > /sys/../0000\:86\:00.0/11/params > > * To destroy device: > > echo $UUID > /sys/../0000\:86\:00.0/mdev_destroy > > > 3. Start/stop mediated device > > No change or requirement for libvirt as this will be handled by open() > and close() callbacks to vendor driver. In case of multiple devices and > 'requires_group' set, this will be handled in 'first open()' and 'last > close()' on device in that group. > > 4. Launch QEMU/VM > > Pass the mdev sysfs path to QEMU as vfio-pci device. > For above vGPU device example: > > -device vfio-pci,sysfsdev=/sys/bus/mdev/devices/$UUID > > 5. QEMU/VM Shutdown sequence > > No change or requirement for libvirt. > > 6. VM Reset > > No change or requirement for libvirt as this will be handled via VFIO > reset API and QEMU process will keep running as before. > > 7. Hot-plug > > It is same syntax to create a virtual device for hot-plug. Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|

From: Daniel P. Berrange [mailto:berrange@redhat.com] Sent: Thursday, September 29, 2016 4:06 PM On Wed, Sep 28, 2016 at 12:48:33PM -0700, Neo Jia wrote: > On Tue, Sep 20, 2016 at 10:47:53AM +0100, Daniel P. Berrange wrote: > > On Tue, Sep 20, 2016 at 02:05:52AM +0530, Kirti Wankhede wrote: > > > > > > Hi libvirt experts, > > > > > > Thanks for valuable input on v1 version of RFC. > > > > > > Quick brief, VFIO based mediated device framework provides a way to > > > virtualize their devices without SR-IOV, like NVIDIA vGPU, Intel KVMGT > > > and IBM's channel IO. This framework reuses VFIO APIs for all the > > > functionalities for mediated devices which are currently being used for > > > pass through devices. This framework introduces a set of new sysfs files > > > for device creation and its life cycle management. > > > > > > Here is the summary of discussion on v1: > > > 1. Discover mediated device: > > > As part of physical device initialization process, vendor driver will > > > register their physical devices, which will be used to create virtual > > > device (mediated device, aka mdev) to the mediated framework. > > > > > > Vendor driver should specify mdev_supported_types in directory format. > > > This format is class based, for example, display class directory format > > > should be as below. We need to define such set for each class of devices > > > which would be supported by mediated device framework. > > > > > > --- mdev_destroy > > > --- mdev_supported_types > > > |-- 11 > > > | |-- create > > > | |-- name > > > | |-- fb_length > > > | |-- resolution > > > | |-- heads > > > | |-- max_instances > > > | |-- params > > > | |-- requires_group > > > |-- 12 > > > | |-- create > > > | |-- name > > > | |-- fb_length > > > | |-- resolution > > > | |-- heads > > > | |-- max_instances > > > | |-- params > > > | |-- requires_group > > > |-- 13 > > > |-- create > > > |-- name > > > |-- fb_length > > > |-- resolution > > > |-- heads > > > |-- max_instances > > > |-- params > > > |-- requires_group > > > > > > > > > In the above example directory '11' represents a type id of mdev device. > > > 'name', 'fb_length', 'resolution', 'heads', 'max_instance' and > > > 'requires_group' would be Read-Only files that vendor would provide to > > > describe about that type. > > > > > > 'create': > > > Write-only file. Mandatory. > > > Accepts string to create mediated device. > > > > > > 'name': > > > Read-Only file. Mandatory. > > > Returns string, the name of that type id. > > > > Presumably this is a human-targetted title/description of > > the device. > > > > > > > > 'fb_length': > > > Read-only file. Mandatory. > > > Returns <number>{K,M,G}, size of framebuffer. > > > > > > 'resolution': > > > Read-Only file. Mandatory. > > > Returns 'hres x vres' format. Maximum supported resolution. > > > > > > 'heads': > > > Read-Only file. Mandatory. > > > Returns integer. Number of maximum heads supported. > > > > None of these should be mandatory as that makes the mdev > > useless for non-GPU devices. > > > > I'd expect to see a 'class' or 'type' attribute in the > > directory whcih tells you what kind of mdev it is. A > > valid 'class' value would be 'gpu'. The fb_length, > > resolution, and heads parameters would only be mandatory > > when class==gpu. > > > > Hi Daniel, > > Here you are proposing to add a class named "gpu", which will make all those gpu > related attributes mandatory, which libvirt can allow user to better > parse/present a particular mdev configuration? > > I am just wondering if there is another option that we just make all those > attributes that a mdev device can have as optional but still meaningful to > libvirt, so libvirt can still parse / recognize them as an class "mdev". 'mdev' isn't a class - mdev is the name of the kernel module. The class refers to the broad capability of the device. class would be things like "gpu", "nic", "fpga" or other such things. The point of the class is to identify which other attributes will be considered mandatory.

From: Daniel P. Berrange [mailto:berrange@redhat.com] Sent: Thursday, September 29, 2016 10:39 PM On Thu, Sep 29, 2016 at 02:35:48PM +0000, Tian, Kevin wrote: > > From: Daniel P. Berrange [mailto:berrange@redhat.com] > > Sent: Thursday, September 29, 2016 4:06 PM > > > > On Wed, Sep 28, 2016 at 12:48:33PM -0700, Neo Jia wrote: > > > On Tue, Sep 20, 2016 at 10:47:53AM +0100, Daniel P. Berrange wrote: > > > > On Tue, Sep 20, 2016 at 02:05:52AM +0530, Kirti Wankhede wrote: > > > > > > > > > > Hi libvirt experts, > > > > > > > > > > Thanks for valuable input on v1 version of RFC. > > > > > > > > > > Quick brief, VFIO based mediated device framework provides a way to > > > > > virtualize their devices without SR-IOV, like NVIDIA vGPU, Intel KVMGT > > > > > and IBM's channel IO. This framework reuses VFIO APIs for all the > > > > > functionalities for mediated devices which are currently being used for > > > > > pass through devices. This framework introduces a set of new sysfs files > > > > > for device creation and its life cycle management. > > > > > > > > > > Here is the summary of discussion on v1: > > > > > 1. Discover mediated device: > > > > > As part of physical device initialization process, vendor driver will > > > > > register their physical devices, which will be used to create virtual > > > > > device (mediated device, aka mdev) to the mediated framework. > > > > > > > > > > Vendor driver should specify mdev_supported_types in directory format. > > > > > This format is class based, for example, display class directory format > > > > > should be as below. We need to define such set for each class of devices > > > > > which would be supported by mediated device framework. > > > > > > > > > > --- mdev_destroy > > > > > --- mdev_supported_types > > > > > |-- 11 > > > > > | |-- create > > > > > | |-- name > > > > > | |-- fb_length > > > > > | |-- resolution > > > > > | |-- heads > > > > > | |-- max_instances > > > > > | |-- params > > > > > | |-- requires_group > > > > > |-- 12 > > > > > | |-- create > > > > > | |-- name > > > > > | |-- fb_length > > > > > | |-- resolution > > > > > | |-- heads > > > > > | |-- max_instances > > > > > | |-- params > > > > > | |-- requires_group > > > > > |-- 13 > > > > > |-- create > > > > > |-- name > > > > > |-- fb_length > > > > > |-- resolution > > > > > |-- heads > > > > > |-- max_instances > > > > > |-- params > > > > > |-- requires_group > > > > > > > > > > > > > > > In the above example directory '11' represents a type id of mdev device. > > > > > 'name', 'fb_length', 'resolution', 'heads', 'max_instance' and > > > > > 'requires_group' would be Read-Only files that vendor would provide to > > > > > describe about that type. > > > > > > > > > > 'create': > > > > > Write-only file. Mandatory. > > > > > Accepts string to create mediated device. > > > > > > > > > > 'name': > > > > > Read-Only file. Mandatory. > > > > > Returns string, the name of that type id. > > > > > > > > Presumably this is a human-targetted title/description of > > > > the device. > > > > > > > > > > > > > > 'fb_length': > > > > > Read-only file. Mandatory. > > > > > Returns <number>{K,M,G}, size of framebuffer. > > > > > > > > > > 'resolution': > > > > > Read-Only file. Mandatory. > > > > > Returns 'hres x vres' format. Maximum supported resolution. > > > > > > > > > > 'heads': > > > > > Read-Only file. Mandatory. > > > > > Returns integer. Number of maximum heads supported. > > > > > > > > None of these should be mandatory as that makes the mdev > > > > useless for non-GPU devices. > > > > > > > > I'd expect to see a 'class' or 'type' attribute in the > > > > directory whcih tells you what kind of mdev it is. A > > > > valid 'class' value would be 'gpu'. The fb_length, > > > > resolution, and heads parameters would only be mandatory > > > > when class==gpu. > > > > > > > > > > Hi Daniel, > > > > > > Here you are proposing to add a class named "gpu", which will make all those gpu > > > related attributes mandatory, which libvirt can allow user to better > > > parse/present a particular mdev configuration? > > > > > > I am just wondering if there is another option that we just make all those > > > attributes that a mdev device can have as optional but still meaningful to > > > libvirt, so libvirt can still parse / recognize them as an class "mdev". > > > > 'mdev' isn't a class - mdev is the name of the kernel module. The class > > refers to the broad capability of the device. class would be things > > like "gpu", "nic", "fpga" or other such things. The point of the class > > is to identify which other attributes will be considered mandatory. > > > > > > Thanks Daniel. This class definition makes sense to me. > > However I'm not sure whether we should define such common mandatory attributes > of a 'gpu' class now. Intel will go with a 2's power sharing of type definition... actual > type name to be finalized, but an example looks like below: > > [GVTG-SKL-x2]: available instances (2) > [GVTG-SKL-x4]: available instances (4) > [GVTG-SKL-x8]: available instances (8) > ... > > User can create different types of vGPUs simultaneously. A GVTG-SKL-x2 type > vGPU will get half of the physical GPU resource, while a GVTG-SKL-x4 type will > get a quarter. However it's unclear to me how we want to enumerate those > resources into resolution or heads. I feel it'd be more reasonable for us to push > initial libvirt mdev support w/o vgpu specific class definition, until we see > a clear value of doing so (at that time we then follow Daniel's guideline to define > mandatory attributes common to all GPU vendors). Libvirt won't report arbitrary vendor define attributes. So if we are not going to define a gpu class & associated attributes, then there will be no reporting of the 'heads', 'resolution', 'fb_length' data described above.

We can have 'class' as optional attribute. So Intel don't have to provide 'class' attribute and they don't have to specify mandatory attributes of that class. We would provide 'class' attribute and provide mandatory attributes.

Ping.. Pulling the questions at the top. >> Will libvirt report 'description' RO attribute, its output would be >> string, so that user could be able to see the configuration of that >> profile? >> > > Daniel, > Waiting for your input on this. > >> We can have 'class' as optional attribute. So Intel don't have to >> provide 'class' attribute and they don't have to specify mandatory >> attributes of that class. We would provide 'class' attribute and provide >> mandatory attributes.