google / alioth / 36653607736

38%

push

github

fix(virtio): reserve an io_uring entry for every enabled queue

All queues of a device share one io_uring. To keep a busy queue from
starving the others, submit_buffers() lets each queue keep
QUEUE_RESERVE_SIZE requests in flight even when no shared entry is
left, and a stalled queue is retried when one of its own requests
completes.

However, the reserve was only taken out of the shared budget for
queues with a queue notifier polled by the ring. Queues notified
through WakeEvent, e.g. when StartParam has no notifiers, could still
use their reserve, so up to one request per queue could be in flight
beyond the budget.

Take the reserve for every enabled queue before the device is
activated, and only charge the notifier polls against the shared
budget.

Assisted-by: Antigravity:Claude-Opus-5
Signed-off-by: Changyuan Lyu <changyuanl@google.com>

785 of 1596 branches covered (49.19%)

Branch coverage included in aggregate %.

0 of 3 new or added lines in 1 file covered. (0.0%)

5684 of 15492 relevant lines covered (36.69%)

922.19 hits per line