On Fri, Jun 01, 2018 at 06:43:41PM +0300, Michael S. Tsirkin wrote: > On Thu, May 31, 2018 at 08:28:32PM -0500, Venu Busireddy wrote: > > > > I looked at the discussion in the threads [1] and [2], where it was > > suggested placing the passthrough device behind one bridge, and the virtio > > device behind another bridge, and storing in those bridges' configuration > > space some unique identifier that can be used to pair the two devices. > > > > After some discussions with Si-Wei Liu and others, we believe that the > > following scheme may be a viable approach. > > Thanks for writing this up! I mostly agree which isn't surprising since > I proposed this scheme originally ;) > https://patchwork.ozlabs.org/patch/901596/#1901665 > > Some minor comments though: > > > Please take a look at this > > proposal and provide your thoughts. > > > > 1. Enhance the QEMU CLI to include a "group_id" option to the bridge > > devices for Q35 as well as i440FX models. I have already made changes > > for the Q35 model (ioh3420 bridge). > > > > 2. When the guest is created, the operator creates two bridge devices > > (for example, using '-device ioh3420,group_id="string"'), and specifies > > a unique identifier string for both bridges. This identifier can be > > the UUID generated by 'uuidgen' command. > > I suggest we limit this to a bridge with QEMU vendor/device ID. > > We have pci-bridge and pci-bridge-seat. > > This could be similar: pci-bridge-group. > Less of a chance there's a conflict. > > Accordingly bridge is the group, not the id :) > > > 3. QEMU places this unique identifier in the PCI configuration space of > > the bridge as Vendor-Specific capability (0x09). > > Why not use the standard UUID capability? Accordingly name property > uuid? > > > The "Vendor" here is > > not to be confused with a specific organization. Instead, the vendor > > of the bridge is QEMU (with vendor ID 0x8086 and device ID 0x3420). > > So that has a small chance of conflicting with a vendor > specific capability used by some driver of this bridge. > We are better off with an actual vendor/device ID IMHO. > > > 4. The operator places the passthrough device behind one of the bridges, > > and the virtio device behind the other bridge. > > > > 5. Patch 4 in patch series [3] should be modified to use the unique > > identifier string stored in the bridges' configuration space instead > > of the MAC address for pairing the devices. > > > > If it is desirable to create only one bridge instead of two (to conserve > > the number of devices in the system), then the passthrough device can be > > attached to that single bridge (with the identifier), and the identifier > > for the virtio device can be stored in the virtio device's configuration > > space itself. To do that, we need to update the virtio specification, > > I think we can use the standard PCI UUID. No need to update the spec > for this.
Thinking some more we might want to say in the spec that PCI UUID is used for pairing so I guess yes, we do need to update the spec after all. > > and I have sent a proposal [4] to the OASIS team to update the virtio > > specification. If that proposal is accepted, then we can modify QEMU to > > use the virtio device's configuration space instead of the second bridge > > to store the unique identifier. > > > > Thank you for sparing the time. > > > > Venu > > > > [1] https://www.spinics.net/lists/linux-virtualization/msg33518.html > > [2] https://www.spinics.net/lists/netdev/msg499011.html > > [3] https://patchwork.ozlabs.org/cover/920005/ > > [4] https://lists.oasis-open.org/archives/virtio-dev/201805/msg00118.html --------------------------------------------------------------------- To unsubscribe, e-mail: virtio-dev-unsubscr...@lists.oasis-open.org For additional commands, e-mail: virtio-dev-h...@lists.oasis-open.org
