[Cc +libvirt folks] On Thu, 14 Apr 2022 03:46:52 -0700 Yi Liu <yi.l.liu(a)intel.com&gt; wrote: > With the introduction of iommufd[1], the linux kernel provides a generic > interface for userspace drivers to propagate their DMA mappings to kernel > for assigned devices. This series does the porting of the VFIO devices > onto the /dev/iommu uapi and let it coexist with the legacy implementation. > Other devices like vpda, vfio mdev and etc. are not considered yet.

> The selection of the backend is made on a device basis using the new > iommufd option (on/off/auto). By default the iommufd backend is selected > if supported by the host and by QEMU (iommufd KConfig). This option is > currently available only for the vfio-pci device. For other types of > devices, it does not yet exist and the legacy BE is chosen by default. I've discussed this a bit with Eric, but let me propose a different command line interface. Libvirt generally likes to pass file descriptors to QEMU rather than grant it access to those files directly. This was problematic with vfio-pci because libvirt can't easily know when QEMU will want to grab another /dev/vfio/vfio container. Therefore we abandoned this approach and instead libvirt grants file permissions. However, with iommufd there's no reason that QEMU ever needs more than a single instance of /dev/iommufd and we're using per device vfio file descriptors, so it seems like a good time to revisit this.

The interface I was considering would be to add an iommufd object to QEMU, so we might have a: -device iommufd[,fd=#][,id=foo] For non-libivrt usage this would have the ability to open /dev/iommufd itself if an fd is not provided. This object could be shared with other iommufd users in the VM and maybe we'd allow multiple instances for more esoteric use cases. [NB, maybe this should be a -object rather than -device since the iommufd is not a guest visible device?]

The vfio-pci device might then become: -device vfio-pci[,host=DDDD:BB:DD.f][,sysfsdev=/sys/path/to/device][,fd=#][,iommufd=foo] So essentially we can specify the device via host, sysfsdev, or passing an fd to the vfio device file. When an iommufd object is specified, "foo" in the example above, each of those options would use the vfio-device access mechanism, essentially the same as iommufd=on in your example. With the fd passing option, an iommufd object would be required and necessarily use device level access. In your example, the iommufd=auto seems especially troublesome for libvirt because QEMU is going to have different locked memory requirements based on whether we're using type1 or iommufd, where the latter resolves the duplicate accounting issues. libvirt needs to know deterministically which backed is being used, which this proposal seems to provide, while at the same time bringing us more in line with fd passing. Thoughts? Thanks,

On Fri, Apr 22, 2022 at 04:09:43PM -0600, Alex Williamson wrote: > [Cc +libvirt folks] > > On Thu, 14 Apr 2022 03:46:52 -0700 > Yi Liu <yi.l.liu(a)intel.com&gt; wrote: > > > With the introduction of iommufd[1], the linux kernel provides a generic > > interface for userspace drivers to propagate their DMA mappings to kernel > > for assigned devices. This series does the porting of the VFIO devices > > onto the /dev/iommu uapi and let it coexist with the legacy implementation. > > Other devices like vpda, vfio mdev and etc. are not considered yet. snip > > The selection of the backend is made on a device basis using the new > > iommufd option (on/off/auto). By default the iommufd backend is selected > > if supported by the host and by QEMU (iommufd KConfig). This option is > > currently available only for the vfio-pci device. For other types of > > devices, it does not yet exist and the legacy BE is chosen by default. > > I've discussed this a bit with Eric, but let me propose a different > command line interface. Libvirt generally likes to pass file > descriptors to QEMU rather than grant it access to those files > directly. This was problematic with vfio-pci because libvirt can't > easily know when QEMU will want to grab another /dev/vfio/vfio > container. Therefore we abandoned this approach and instead libvirt > grants file permissions. > > However, with iommufd there's no reason that QEMU ever needs more than > a single instance of /dev/iommufd and we're using per device vfio file > descriptors, so it seems like a good time to revisit this. I assume access to '/dev/iommufd' gives the process somewhat elevated privileges, such that you don't want to unconditionally give QEMU access to this device ?

> The interface I was considering would be to add an iommufd object to > QEMU, so we might have a: > > -device iommufd[,fd=#][,id=foo] > > For non-libivrt usage this would have the ability to open /dev/iommufd > itself if an fd is not provided. This object could be shared with > other iommufd users in the VM and maybe we'd allow multiple instances > for more esoteric use cases. [NB, maybe this should be a -object rather than > -device since the iommufd is not a guest visible device?] Yes, -object would be the right answer for something that's purely a host side backend impl selector. > The vfio-pci device might then become: > > -device vfio-pci[,host=DDDD:BB:DD.f][,sysfsdev=/sys/path/to/device][,fd=#][,iommufd=foo] > > So essentially we can specify the device via host, sysfsdev, or passing > an fd to the vfio device file. When an iommufd object is specified, > "foo" in the example above, each of those options would use the > vfio-device access mechanism, essentially the same as iommufd=on in > your example. With the fd passing option, an iommufd object would be > required and necessarily use device level access. > > In your example, the iommufd=auto seems especially troublesome for > libvirt because QEMU is going to have different locked memory > requirements based on whether we're using type1 or iommufd, where the > latter resolves the duplicate accounting issues. libvirt needs to know > deterministically which backed is being used, which this proposal seems > to provide, while at the same time bringing us more in line with fd > passing. Thoughts? Thanks, Yep, I agree that libvirt needs to have more direct control over this. This is also even more important if there are notable feature differences in the 2 backends. I wonder if anyone has considered an even more distinct impl, whereby we have a completely different device type on the backend, eg -device vfio-iommu-pci[,host=DDDD:BB:DD.f][,sysfsdev=/sys/path/to/device][,fd=#][,iommufd=foo] If a vendor wants to fully remove the legacy impl, they can then use the Kconfig mechanism to disable the build of the legacy impl device, while keeping the iommu impl (or vica-verca if the new iommu impl isn't considered reliable enough for them to support yet). Libvirt would use -object iommu,id=iommu0,fd=NNN -device vfio-iommu-pci,fd=MMM,iommu=iommu0 Non-libvirt would use a simpler -device vfio-iommu-pci,host=0000:03:22.1 with QEMU auto-creating a 'iommu' object in the background. This would fit into libvirt's existing modelling better. We currently have a concept of a PCI assignment backend, which previously supported the legacy PCI assignment, vs the VFIO PCI assignment. This new iommu impl feels like a 3rd PCI assignment approach, and so fits with how we modelled it as a different device type in the past.

Based on current plan there is probably a transition window between the point where the first vfio device type (vfio-pci) gaining iommufd support and the point where all vfio types supporting iommufd.

We also need to be able to advise libvirt as to how each iommufd object or user of that object factors into the VM locked memory requirement. When used by vfio-pci, we're only mapping VM RAM, so we'd ask libvirt to set the locked memory limit to the size of VM RAM per iommufd, regardless of the number of devices using a given iommufd. However, I don't know if all users of iommufd will be exclusively mapping VM RAM. Combinations of devices where some map VM RAM and others map QEMU buffer space could still require some incremental increase per device (I'm not sure if vfio-nvme is such a device). It seems like heuristics will still be involved even after iommufd solves the per-device vfio-pci locked memory limit issue. Thanks,

On Tue, 26 Apr 2022 13:42:17 -0300 Jason Gunthorpe <jgg(a)nvidia.com&gt; wrote: > On Tue, Apr 26, 2022 at 10:21:59AM -0600, Alex Williamson wrote: > > We also need to be able to advise libvirt as to how each iommufd object > > or user of that object factors into the VM locked memory requirement. > > When used by vfio-pci, we're only mapping VM RAM, so we'd ask libvirt > > to set the locked memory limit to the size of VM RAM per iommufd, > > regardless of the number of devices using a given iommufd. However, I > > don't know if all users of iommufd will be exclusively mapping VM RAM. > > Combinations of devices where some map VM RAM and others map QEMU > > buffer space could still require some incremental increase per device > > (I'm not sure if vfio-nvme is such a device). It seems like heuristics > > will still be involved even after iommufd solves the per-device > > vfio-pci locked memory limit issue. Thanks, > > If the model is to pass the FD, how about we put a limit on the FD > itself instead of abusing the locked memory limit? > > We could have a no-way-out ioctl that directly limits the # of PFNs > covered by iopt_pages inside an iommufd. FD passing would likely only be the standard for libvirt invoked VMs. The QEMU vfio-pci device would still parse a host= or sysfsdev= option when invoked by mortals and associate to use the legacy vfio group interface or the new vfio device interface based on whether an iommufd is specified.

Does that rule out your suggestion? I don't know, please reveal more about the mechanics of putting a limit on the FD itself and this no-way-out ioctl. The latter name suggests to me that I should also note that we need to support memory hotplug with these devices. Thanks,

> From: Alex Williamson <alex.williamson(a)redhat.com&gt; > Sent: Wednesday, April 27, 2022 12:22 AM > > > > > > My expectation would be that libvirt uses: > > > > > > -object iommufd,id=iommufd0,fd=NNN > > > -device vfio-pci,fd=MMM,iommufd=iommufd0 > > > > > > Whereas simple QEMU command line would be: > > > > > > -object iommufd,id=iommufd0 > > > -device vfio-pci,iommufd=iommufd0,host=0000:02:00.0 > > > > > > The iommufd object would open /dev/iommufd itself. Creating an > > > implicit iommufd object is someone problematic because one of the > > > things I forgot to highlight in my previous description is that the > > > iommufd object is meant to be shared across not only various vfio > > > devices (platform, ccw, ap, nvme, etc), but also across subsystems, ex. > > > vdpa. > > > > Out of curiosity - in concept one iommufd is sufficient to support all > > ioas requirements across subsystems while having multiple iommufd's > > instead lose the benefit of centralized accounting. The latter will also > > cause some trouble when we start virtualizing ENQCMD which requires > > VM-wide PASID virtualization thus further needs to share that > > information across iommufd's. Not unsolvable but really no gain by > > adding such complexity. So I'm curious whether Qemu provide > > a way to restrict that certain object type can only have one instance > > to discourage such multi-iommufd attempt? > > I don't see any reason for QEMU to restrict iommufd objects. The QEMU > philosophy seems to be to let users create whatever configuration they > want. For libvirt though, the assumption would be that a single > iommufd object can be used across subsystems, so libvirt would never > automatically create multiple objects. I like the flexibility what the objection approach gives in your proposal. But with the said complexity in mind (with no foreseen benefit), I wonder

whether an alternative approach which treats iommufd as a global property instead of an object is acceptable in Qemu, i.e.: -iommufd on/off -device vfio-pci,iommufd,[fd=MMM/host=0000:02:00.0] All devices with iommufd specified then implicitly share a single iommufd object within Qemu.

This still allows vfio devices to be specified via fd but just requires Libvirt to grant file permission on /dev/iommu. Is it a worthwhile tradeoff to be considered or just not a typical way in Qemu philosophy e.g. any object associated with a device must be explicitly specified?

From: Daniel P. Berrangé <berrange(a)redhat.com&gt; Sent: Friday, April 29, 2022 12:20 AM On Thu, Apr 28, 2022 at 08:24:48AM -0600, Alex Williamson wrote: > On Thu, 28 Apr 2022 03:21:45 +0000 > "Tian, Kevin" <kevin.tian(a)intel.com&gt; wrote: > > > > From: Alex Williamson <alex.williamson(a)redhat.com&gt; > > > Sent: Wednesday, April 27, 2022 12:22 AM > > > > > > > > > > My expectation would be that libvirt uses: > > > > > > > > > > -object iommufd,id=iommufd0,fd=NNN > > > > > -device vfio-pci,fd=MMM,iommufd=iommufd0 > > > > > > > > > > Whereas simple QEMU command line would be: > > > > > > > > > > -object iommufd,id=iommufd0 > > > > > -device vfio-pci,iommufd=iommufd0,host=0000:02:00.0 > > > > > > > > > > The iommufd object would open /dev/iommufd itself. Creating an > > > > > implicit iommufd object is someone problematic because one of the > > > > > things I forgot to highlight in my previous description is that the > > > > > iommufd object is meant to be shared across not only various vfio > > > > > devices (platform, ccw, ap, nvme, etc), but also across subsystems, ex. > > > > > vdpa. > > > > > > > > Out of curiosity - in concept one iommufd is sufficient to support all > > > > ioas requirements across subsystems while having multiple iommufd's > > > > instead lose the benefit of centralized accounting. The latter will also > > > > cause some trouble when we start virtualizing ENQCMD which requires > > > > VM-wide PASID virtualization thus further needs to share that > > > > information across iommufd's. Not unsolvable but really no gain by > > > > adding such complexity. So I'm curious whether Qemu provide > > > > a way to restrict that certain object type can only have one instance > > > > to discourage such multi-iommufd attempt? > > > > > > I don't see any reason for QEMU to restrict iommufd objects. The QEMU > > > philosophy seems to be to let users create whatever configuration they > > > want. For libvirt though, the assumption would be that a single > > > iommufd object can be used across subsystems, so libvirt would never > > > automatically create multiple objects. > > > > I like the flexibility what the objection approach gives in your proposal. > > But with the said complexity in mind (with no foreseen benefit), I wonder > > What's the actual complexity? Front-end/backend splits are very common > in QEMU. We're making the object connection via name, why is it > significantly more complicated to allow multiple iommufd objects? On > the contrary, it seems to me that we'd need to go out of our way to add > code to block multiple iommufd objects.

> > > whether an alternative approach which treats iommufd as a global > > property instead of an object is acceptable in Qemu, i.e.: > > > > -iommufd on/off > > -device vfio-pci,iommufd,[fd=MMM/host=0000:02:00.0] > > > > All devices with iommufd specified then implicitly share a single iommufd > > object within Qemu. > > QEMU requires key-value pairs AFAIK, so the above doesn't work, then > we're just back to the iommufd=on/off. > > > This still allows vfio devices to be specified via fd but just requires Libvirt > > to grant file permission on /dev/iommu. Is it a worthwhile tradeoff to be > > considered or just not a typical way in Qemu philosophy e.g. any object > > associated with a device must be explicitly specified? > > Avoiding QEMU opening files was a significant focus of my alternate > proposal. Also note that we must be able to support hotplug, so we > need to be able to dynamically add and remove the iommufd object, I > don't see that a global property allows for that. Implicit > associations of devices to shared resources doesn't seem particularly > desirable to me. Thanks, Adding new global properties/options is rather an anti-pattern for QEMU these days. Using -object is the right approach. If you only want to allow for one of them, just document this requirement. We've got other objects which are singletons like all the confidential guest classes for each arch.

Hi Alex, On 4/23/22 12:09 AM, Alex Williamson wrote: > [Cc +libvirt folks] > > On Thu, 14 Apr 2022 03:46:52 -0700 > Yi Liu <yi.l.liu(a)intel.com&gt; wrote: > >> With the introduction of iommufd[1], the linux kernel provides a generic >> interface for userspace drivers to propagate their DMA mappings to kernel >> for assigned devices. This series does the porting of the VFIO devices >> onto the /dev/iommu uapi and let it coexist with the legacy implementation. >> Other devices like vpda, vfio mdev and etc. are not considered yet. >> >> For vfio devices, the new interface is tied with device fd and iommufd >> as the iommufd solution is device-centric. This is different from legacy >> vfio which is group-centric. To support both interfaces in QEMU, this >> series introduces the iommu backend concept in the form of different >> container classes. The existing vfio container is named legacy container >> (equivalent with legacy iommu backend in this series), while the new >> iommufd based container is named as iommufd container (may also be mentioned >> as iommufd backend in this series). The two backend types have their own >> way to setup secure context and dma management interface. Below diagram >> shows how it looks like with both BEs. >> >> VFIO AddressSpace/Memory >> +-------+ +----------+ +-----+ +-----+ >> | pci | | platform | | ap | | ccw | >> +---+---+ +----+-----+ +--+--+ +--+--+ +----------------------+ >> | | | | | AddressSpace | >> | | | | +------------+---------+ >> +---V-----------V-----------V--------V----+ / >> | VFIOAddressSpace | <------------+ >> | | | MemoryListener >> | VFIOContainer list | >> +-------+----------------------------+----+ >> | | >> | | >> +-------V------+ +--------V----------+ >> | iommufd | | vfio legacy | >> | container | | container | >> +-------+------+ +--------+----------+ >> | | >> | /dev/iommu | /dev/vfio/vfio >> | /dev/vfio/devices/vfioX | /dev/vfio/$group_id >> Userspace | | >> ===========+============================+================================ >> Kernel | device fd | >> +---------------+ | group/container fd >> | (BIND_IOMMUFD | | (SET_CONTAINER/SET_IOMMU) >> | ATTACH_IOAS) | | device fd >> | | | >> | +-------V------------V-----------------+ >> iommufd | | vfio | >> (map/unmap | +---------+--------------------+-------+ >> ioas_copy) | | | map/unmap >> | | | >> +------V------+ +-----V------+ +------V--------+ >> | iommfd core | | device | | vfio iommu | >> +-------------+ +------------+ +---------------+ >> >> [Secure Context setup] >> - iommufd BE: uses device fd and iommufd to setup secure context >> (bind_iommufd, attach_ioas) >> - vfio legacy BE: uses group fd and container fd to setup secure context >> (set_container, set_iommu) >> [Device access] >> - iommufd BE: device fd is opened through /dev/vfio/devices/vfioX >> - vfio legacy BE: device fd is retrieved from group fd ioctl >> [DMA Mapping flow] >> - VFIOAddressSpace receives MemoryRegion add/del via MemoryListener >> - VFIO populates DMA map/unmap via the container BEs >> *) iommufd BE: uses iommufd >> *) vfio legacy BE: uses container fd >> >> This series qomifies the VFIOContainer object which acts as a base class >> for a container. This base class is derived into the legacy VFIO container >> and the new iommufd based container. The base class implements generic code >> such as code related to memory_listener and address space management whereas >> the derived class implements callbacks that depend on the kernel user space >> being used. >> >> The selection of the backend is made on a device basis using the new >> iommufd option (on/off/auto). By default the iommufd backend is selected >> if supported by the host and by QEMU (iommufd KConfig). This option is >> currently available only for the vfio-pci device. For other types of >> devices, it does not yet exist and the legacy BE is chosen by default. > I've discussed this a bit with Eric, but let me propose a different > command line interface. Libvirt generally likes to pass file > descriptors to QEMU rather than grant it access to those files > directly. This was problematic with vfio-pci because libvirt can't > easily know when QEMU will want to grab another /dev/vfio/vfio > container. Therefore we abandoned this approach and instead libvirt > grants file permissions. > > However, with iommufd there's no reason that QEMU ever needs more than > a single instance of /dev/iommufd and we're using per device vfio file > descriptors, so it seems like a good time to revisit this. > > The interface I was considering would be to add an iommufd object to > QEMU, so we might have a: > > -device iommufd[,fd=#][,id=foo] > > For non-libivrt usage this would have the ability to open /dev/iommufd > itself if an fd is not provided. This object could be shared with > other iommufd users in the VM and maybe we'd allow multiple instances > for more esoteric use cases. [NB, maybe this should be a -object rather than > -device since the iommufd is not a guest visible device?] > > The vfio-pci device might then become: > > -device vfio-pci[,host=DDDD:BB:DD.f][,sysfsdev=/sys/path/to/device][,fd=#][,iommufd=foo] > > So essentially we can specify the device via host, sysfsdev, or passing > an fd to the vfio device file. When an iommufd object is specified, > "foo" in the example above, each of those options would use the > vfio-device access mechanism, essentially the same as iommufd=on in > your example. With the fd passing option, an iommufd object would be > required and necessarily use device level access. What is the use case you foresee for the "fd=#" option?

> > In your example, the iommufd=auto seems especially troublesome for > libvirt because QEMU is going to have different locked memory > requirements based on whether we're using type1 or iommufd, where the > latter resolves the duplicate accounting issues. libvirt needs to know > deterministically which backed is being used, which this proposal seems > to provide, while at the same time bringing us more in line with fd > passing. Thoughts? Thanks, I like your proposal (based on the -object iommufd). The only thing that may be missing I think is for a qemu end-user who actually does not care about the iommu backend being used but just wishes to use the most recent available one it adds some extra complexity. But this is not the most important use case ;)