[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]
Subject: Re: [virtio-dev] [PATCH v3 0/4] introduce virtio vhost-user backend device type
On Mon, Aug 05, 2019 at 01:32:14AM +0300, Nikos Dragazis wrote: > On 31/7/19 11:19 Î.Î., Michael S. Tsirkin wrote: > > On Wed, Jul 31, 2019 at 05:49:57PM +0300, Nikos Dragazis wrote: > >> On 21/6/19 11:43 Î.Î., Michael S. Tsirkin wrote: > >>> On Wed, Jun 19, 2019 at 03:54:30PM +0100, Stefan Hajnoczi wrote: > >>>> On Sat, May 11, 2019 at 09:47:36PM +0300, Nikos Dragazis wrote: > >>>>> Hi everyone, > >>>>> > >>>>> this PATCH presents an updated version of the RFC virtio device spec for > >>>>> the virtio-vhost-user device. The initial RFC implementation can be > >>>>> found here: [1]. > >>>>> > >>>>> This PATCH is split into four parts: > >>>>> > >>>>> 1. the first commit is just a refactored version of the initial RFC > >>>>> implementation [1]. I just moved the device spec into a separate .tex > >>>>> file and changed the device id from #24 to #28 since the ids #24-#27 are > >>>>> reserved. > >>>>> > >>>>> 2. the second commit fixes some minor issues with the device spec. This > >>>>> is quite straightforward. > >>>>> > >>>>> 3. the third commit enhances the notification capability with some > >>>>> device/driver requirements. This makes sense because the notification > >>>>> capability behaves similarly to the MSI-X capability, thereby having > >>>>> some requirements that need to be mentioned. > >>>>> > >>>>> 4. the fourth commit synchronizes the shared memory capability with a > >>>>> recent patch [2] that attempts to standardize a standalone > >>>>> SHARED_MEMORY_CFG virtio capability. > >>>>> > >>>>> v3 changes: > >>>>> * Device Requirements for the Notification Capability: point out the > >>>>> difference between the MSI-X Table Size stored in the Message Control > >>>>> register of the MSI-X capability structure and the actual MSI-X Table > >>>>> Size > >>>>> > >>>>> v2 changes: > >>>>> * Change device id from #25 to #28 > >>>>> > >>>>> Looking forward to your comments. > >>>> I'm happy with v3. > >>>> > >>>> Any other comments or shall we proceed to a vote? > >>>> > >>>> Stefan > >>> I think we need to wait for shared memory part to be finalized, right? > >>> > >> Michael, > >> > >> given that David's patchset for the shared memory regions has been > >> approved, I think we can start discussing on the spec for the > >> virtio-vhost-user device. A link to the latest version is here: [1]. > >> > >> I really think that this device is useful and should be part of the > >> VIRTIO specification. I will submit a fifth version of the patchset soon > >> with some minor changes that seem reasonable to me. > >> > >> Best regards, > >> Nikos > >> > >> [1] https://lists.oasis-open.org/archives/virtio-dev/201906/msg00036.html > > One part I dislike there is "Additional Device Resources over PCI". > > That mostly seems to deal with allocating interrupts. > > Actually, it deals with allocating notification addresses (doorbells) > and device interrupts for the vhost-user virtqueues. The > virtio-vhost-user device must be able to handle both its own RX/TX > virtqueues and the vhost-user virtqueues. So, the device must offer > separate notification addresses and interrupt vectors for the vhost-user > virtqueues. We are standardizing these resources with the > VIRTIO_PCI_CAP_DOORBELL_CFG and VIRTIO_PCI_CAP_NOTIFICATION_CFG > configuration structures. > > > Can't we (ab)use the virtio pci registers for this? > > Just extend VQ number to VQ/notification number. > > > > I am not sure I get your point. But, yes, I think we could use the > existing registers (queue_select, queue_msix_vector, queue_notify_off) > in the common configuration structure for both the deviceâs virtqueues > and the vhost-user virtqueues. So, for example, we could use indexes 0, > 1 for the deviceâs RX/TX virtqueues and index i+2 for the i-th > vhost-user virtqueue. > > However, I don't see why mixing up the configuration of the RX/TX and > the vhost-user virtqueues is a better solution. The reason is that this will automatically buy you support in all transports: PCI/CCW/MMIO ... > Wouldnât it be more > clean if we had separate configuration structures for the device's RX/TX > virtqueues and the vhost-user virtqueues? Flip this on its head and you will see that it buys you nothing except cosmetics to separate them, and costs you portability across transports. > -- > Nikos
[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]