On 12/13/17 10:29, Paolo Bonzini wrote: > On 13/12/2017 09:35, Laszlo Ersek wrote: >> I consider the lack of a "VIRTIO_SCSI_F_MQ" feature bit an issue with >> the virtio specification (and consequently with vhost-scsi), not with >> the guest driver(s). > > VIRTIO_SCSI_F_MQ does not exist because virtio-scsi has _always_ > supported multiqueue and has always had a "num_queues" field in the > configuration space. For virtio-net, VIRTIO_NET_F_MQ does not say > "the device or driver knows about multiqueue", it says "the device or > driver wants to read max_virtqueue_pairs" from configuration space. > It's perfectly fine for a device to propose VIRTIO_NET_F_MQ and set > max_virtqueue_pairs=1, or for a driver to negotiate VIRTIO_NET_F_MQ > and then skip initialization of some virtqueues. > > This also means that Maxime's patch to DPDK is also not enough. :) > Virtio-net actually does have a configuration mechanism for > multiqueue, namely the VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET command; the > driver sends that command specifying the number of the transmit and > receive queues to use. However, in my understanding, that command is > only needed for the device to configure receive flow steering, so > virtio-scsi doesn't need that either. > >> Perhaps you can update vhost-scsi similarly to the last patch of >> Maxime's v4 series, even without "VIRTIO_SCSI_F_MQ" -- in the >> SET_FEATURES request handler, just destroy the unused virtqueues that >> have not been configured by the guest driver until that time? > > Yes, this is the right solution. We can assume that if the descriptor > address is equal to zero, the queue is not in use. This is not in the > spec as far as I can see, but it is QEMU's assumption. I will send a > patch to the virtio specification.
Hmmm, I'm not so sure about this, on a second thought. Reviewing the OVMF code, I see that I added a comment (to all of the virtio drivers actually): > // > // In virtio-1.0, feature negotiation is expected to complete before queue > // discovery, and the device can also reject the selected set of features. > // I added this because of the following sections in the 1.0 spec: - 3.1.1 Driver Requirements: Device Initialization - 3.1.2 Legacy Interface: Device Initialization In particular 3.1.2 writes, "The result was [...] steps 4, 7 and 8 were conflated.". (When I added virtio-1.0 support to OVMF, I paid attention to conform to the new ordering for modern transports, and to keep the ordering unchanged otherwise.) I think this is a problem then; if a 1.0 driver is required to finish feature negotiation (steps 4-6) before configuring the queues (step 7), then the host side cannot derive any clues from the state of the queues when the guest completes step 5 (= set FEATURES_OK). Am I wrong? ... On the other hand, when the driver sets DRIVER_OK (step 8), then the host *can* derive clues from the state of the queues; I think. Thanks Laszlo _______________________________________________ edk2-devel mailing list [email protected] https://lists.01.org/mailman/listinfo/edk2-devel
