Matias Ezequiel Vara Larsen
2023-Oct-17 15:19 UTC
virtio-sound: release control request clarification
Hello, This email is to clarify the VirtIO specification regarding the RELEASE control request. Section 5.14.6.6.5.1 [1] states the following device requirements for the RELEASE control request: 1. The device MUST complete all pending I/O messages for the specified stream ID. 2. The device MUST NOT complete the control request while there are pending I/O messages for the specified stream ID. The 1) requirement does not indicate what "complete" means. Does it mean that the pending I/O messages in the tx queue shall be outputted in the host, i.e., consumed by the audio backend? Or, completion means simply to put the requests in the used-ring without consuming them? Regarding 2), I interpret it as "the device shall wait until all I/O messages are proceeded to complete the RELEASE control request". Currently, the kernel driver seems not expecting such a delay when the RELEASE command is sent. If I understand correctly, the kernel driver first sends the RELEASE command and waits a fixed amount of time until the device can process it. Then, the driver waits a fixed amount of time until all pending IO messages are completed. If the device follows the specification and waits until all messages IO are completed to issue the completion of the RELEASE command, the kernel driver may timeout. The time to complete N IO messages in the TX queue could be proportional with the number of pending messages. In our device implementation [2], RELEASE is handled as follows: - Drop all messages in the TX queue without outputting in the host. - Complete the RELEASE control request. This seems to be working, however, I can observe that sometimes there are still requests in the TX queue when we get RELEASE. Those requests are never reproduced in the host. My questions are: - In the specification, should we modify it to clarify that all pending IO messages in the device are discarded during RELEASE, that is, not output to the host, but signaled to the guest as completed? - According to the specification, should the driver wait in RELEASE an amount of time proportional to the number of periods yet to be reproduced? Thanks, Matias. [1] https://docs.oasis-open.org/virtio/virtio/v1.2/csd01/virtio-v1.2-csd01.html [2] https://github.com/rust-vmm/vhost-device/tree/main/staging/vhost-device-sound
Manos Pitsidianakis
2023-Oct-17 15:29 UTC
virtio-sound: release control request clarification
On Tue, 17 Oct 2023 at 18:19, Matias Ezequiel Vara Larsen <mvaralar at redhat.com> wrote:> > Hello, > > This email is to clarify the VirtIO specification regarding the RELEASE > control request. Section 5.14.6.6.5.1 [1] states the following device > requirements for the RELEASE control request: > 1. The device MUST complete all pending I/O messages for the specified > stream ID. > 2. The device MUST NOT complete the control request while there are > pending I/O messages for the specified stream ID. > > The 1) requirement does not indicate what "complete" means. Does it mean > that the pending I/O messages in the tx queue shall be outputted in the > host, i.e., consumed by the audio backend? Or, completion means simply > to put the requests in the used-ring without consuming them?It means the latter. At no point is the host's consumption of audio data mentioned except for implicit or explicit period notifications.> > Regarding 2), I interpret it as "the device shall wait until all I/O > messages are proceeded to complete the RELEASE control request".Possible state transitions to RELEASE state are from PREPARE and STOP, which neither are associated with active I/O in the streams. The correct interpretation is "Do not reply to the control request if you have pending I/O messages for this stream ID".> Currently, the kernel driver seems not expecting such a delay when the > RELEASE command is sent. If I understand correctly, the kernel driver > first sends the RELEASE command and waits a fixed amount of time until > the device can process it. Then, the driver waits a fixed amount of time > until all pending IO messages are completed. If the device follows the > specification and waits until all messages IO are completed to issue the > completion of the RELEASE command, the kernel driver may timeout. The > time to complete N IO messages in the TX queue could be proportional > with the number of pending messages. > > In our device implementation [2], RELEASE is handled as follows: > - Drop all messages in the TX queue without outputting in the host. > - Complete the RELEASE control request. > > This seems to be working, however, I can observe that sometimes there > are still requests in the TX queue when we get RELEASE. Those requests > are never reproduced in the host.My guess is this is because of the guest alsa doing prebuffering, not that the host is supposed to handle those I/O messages. -- Manos
Hi Matias, On 18.10.2023 00:19, Matias Ezequiel Vara Larsen wrote:> Hello, > > This email is to clarify the VirtIO specification regarding the RELEASE > control request. Section 5.14.6.6.5.1 [1] states the following device > requirements for the RELEASE control request: > 1. The device MUST complete all pending I/O messages for the specified > stream ID. > 2. The device MUST NOT complete the control request while there are > pending I/O messages for the specified stream ID. > > The 1) requirement does not indicate what "complete" means. Does it mean > that the pending I/O messages in the tx queue shall be outputted in the > host, i.e., consumed by the audio backend? Or, completion means simply > to put the requests in the used-ring without consuming them?Here "to complete" means moving the buffers to the used list in vring. Technically, the specification only requires that the device "return" all referenced DMA memory to the guest before completing the RELEASE control request. What the device actually does with these I/O messages is implementation dependent and is not within the scope of the specification. Thus...> Regarding 2), I interpret it as "the device shall wait until all I/O > messages are proceeded to complete the RELEASE control request"....you can do this way if you really need to.> Currently, the kernel driver seems not expecting such a delay when the > RELEASE command is sent. If I understand correctly, the kernel driver > first sends the RELEASE command and waits a fixed amount of time until > the device can process it. Then, the driver waits a fixed amount of time > until all pending IO messages are completed. If the device follows the > specification and waits until all messages IO are completed to issue the > completion of the RELEASE command, the kernel driver may timeout. The > time to complete N IO messages in the TX queue could be proportional > with the number of pending messages.The default timeout for control requests in the ALSA driver is 1 second. In theory, this time should be enough to completely reproduce/fill the 500ms buffer, and complete all requests, including the RELEASE control request. If the device fails to do this, then most likely there are some problems with the implementation.> In our device implementation [2], RELEASE is handled as follows: > - Drop all messages in the TX queue without outputting in the host. > - Complete the RELEASE control request. > > This seems to be working, however, I can observe that sometimes there > are still requests in the TX queue when we get RELEASE. Those requests > are never reproduced in the host. > > My questions are: > - In the specification, should we modify it to clarify that all pending > IO messages in the device are discarded during RELEASE, that is, not > output to the host, but signaled to the guest as completed?No, we shouldn't. See comment above.> - According to the specification, should the driver wait in RELEASE an > amount of time proportional to the number of periods yet to be > reproduced?This is purely a matter of driver implementation. It is possible to implement the driver without timeouts, but this would be a bad idea. Because bugs in the device could lead to an infinite wait in the kernel. Best regards,> > Thanks, Matias. > > [1] > https://ddec1-0-en-ctp.trendmicro.com:443/wis/clicktime/v1/query?url=https%3a%2f%2fdocs.oasis%2dopen.org%2fvirtio%2fvirtio%2fv1.2%2fcsd01%2fvirtio%2dv1.2%2dcsd01.html&umid=31e1136e-6322-4698-9f1d-d631ac36403e&auth=53c7c7de28b92dfd96e93d9dd61a23e634d2fbec-586f0596c89224a3bc9e20df81eaea8933bb129a > [2] > https://github.com/rust-vmm/vhost-device/tree/main/staging/vhost-device-sound-- Anton Yakovlev Senior Software Engineer OpenSynergy GmbH Rotherstr. 20, 10245 Berlin