Jorge Moreira <[email protected]> pointed out that the ring state machine is underspecified. In the discussion that followed, we discovered that the spec says one thing and implementations do something else. This patch updates the spec to reflect how things are actually implemented across widely-used front-ends and back-ends including QEMU, crosvm, rust-vmm, and DPDK. Do this while taking care not to make any other existing implementations non-compliant by changing the sepc.
The spec says rings are started when a kick is received but the implementations actually start rings when VHOST_USER_SET_VRING_KICK is received. Reconcile this as follows: - Clarify that a ring can be stopped and then started again. The back-end must resume processing available requests when the ring is restarted. - Update the spec to say rings are started when VHOST_USER_SET_VRING_KICK is received. - Ensure compatibility by saying front-ends SHOULD inject a kick in case the back-end strictly implemented the old spec. - Avoid future back-end dependencies on injected kicks by saying that back-ends SHOULD NOT expect a kick to start rings. This way implementors have clarity on how things work while still allowing compatibility for existing implementations. Reported-by: Jorge Moreira <[email protected]> Cc: "Michael S . Tsirkin" <[email protected]> Cc: Stefano Garzarella <[email protected]> Signed-off-by: Stefan Hajnoczi <[email protected]> --- docs/interop/vhost-user.rst | 26 ++++++++++++++++++++------ 1 file changed, 20 insertions(+), 6 deletions(-) diff --git a/docs/interop/vhost-user.rst b/docs/interop/vhost-user.rst index 137c9f3669..6ced95e080 100644 --- a/docs/interop/vhost-user.rst +++ b/docs/interop/vhost-user.rst @@ -462,12 +462,26 @@ Rings have two independent states: started/stopped, and enabled/disabled. * started and enabled: The back-end must process the ring normally, i.e. process all requests and execute them. -Each ring is initialized in a stopped and disabled state. The back-end -must start a ring upon receiving a kick (that is, detecting that file -descriptor is readable) on the descriptor specified by -``VHOST_USER_SET_VRING_KICK`` or receiving the in-band message -``VHOST_USER_VRING_KICK`` if negotiated, and stop a ring upon receiving -``VHOST_USER_GET_VRING_BASE``. +Each ring is initialized in a stopped and disabled state. Rings are started +with ``VHOST_USER_SET_VRING_KICK`` (or ``VHOST_USER_VRING_KICK`` if +``VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS`` is negotiated) and stopped with +``VHOST_USER_GET_VRING_BASE``. A stopped ring enters the started state again +with ``VHOST_USER_SET_VRING_KICK`` (or ``VHOST_USER_VRING_KICK`` if +``VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS`` is negotiated) and the back-end +resumes processing requests. + +Note that previous versions of this specification stated that rings start when +the back-end receives a kick (that is, detecting that file descriptor is +readable) on the descriptor specified by ``VHOST_USER_SET_VRING_KICK`` or +receiving the in-band message ``VHOST_USER_VRING_KICK`` if negotiated. +Widely-used front-ends and back-ends did not implement this behavior and it +complicates poll mode back-ends that do not rely on the kick file descriptor. + +For compatibility with back-ends that implemented the start on kick behavior, +front-ends SHOULD inject a kick after ``VHOST_USER_SET_VRING_KICK``. This +ensures that the back-end processes any available requests in the ring. +Back-ends SHOULD NOT rely on receiving a kick after +``VHOST_USER_SET_VRING_KICK``. Rings can be enabled or disabled by ``VHOST_USER_SET_VRING_ENABLE``. -- 2.54.0
