On 07/02/2018 20:06, Roman Kagan wrote: >>> + struct hyperv_message *msg; >>> + /* >>> + * the state of the message staged in .msg: >>> + * 0 - the staging area is not in use (after init or message >>> + * successfully delivered to guest) >>> + * -EBUSY - the staging area is being used in vcpu thread >>> + * -EAGAIN - delivery attempt failed due to slot being busy, retry >>> + * -EXXXX - error >>> + */ >>> + int msg_status; >>> + >> Access to these fields needs to be protected by a mutex (the refcount is >> okay because it is only released in the bottom half). > Hmm, I'll double-check, but the original idea was that this stuff is > only used/ref-d/unref-d in the main thread so no mutex was necessary.
"Passing the buck" from iothread to vCPU and vice versa should work; async_run_on_cpu and aio_bh_schedule_oneshot introduce the necessary ordering. However you could still have concurrent access to the state. The mutex is the safest option, but please document whatever you come up with. Thanks, Paolo