On 03/04/2015 12:38, Fam Zheng wrote: > Stopping the vm will drive the block job all the way to the end, because the > sleep duration is too short, which means the block_job_sleep_ns in the block > jobs are unhelpful. That is because the timer will fire too soon, even before > the aio_poll in bdrv_drain_all returns. > > Lengthen the sleep and add a test case to catch this issue in the future. > > It's not perfect, because the aio_poll returning point could still be far > enough that we wake up the job earlier, but this patch is already making it > better in common cases - setting up a timer with timeout=0 was definitely too > short anyway. > > A complete solution would be adding a "sleep until next iteration" timer/BH > API, but I'm not sure that is worth the complexity.
Would it work if vm_stop pauses block jobs before drain, and restarts the paused ones afterwards? Paolo