On Thu, Jul 09, 2015 at 11:47:56AM +0800, Fam Zheng wrote:
> block_job_sleep_ns is called by block job coroutines to yield the
> execution to VCPU threads and monitor etc. It is pointless to sleep for
> 0 or a few nanoseconds, because that equals to a "yield + enter" with no
> intermission in between (the timer fires immediately in the same
> iteration of event loop), which means other code still doesn't get a
> fair share of main loop / BQL.
> 
> Introduce block_job_relax_cpu which will at least for
> BLOCK_JOB_RELAX_CPU_NS. Existing block_job_sleep_ns(job, 0) callers can
> be replaced by this later.
> 
> Reported-by: Alexandre DERUMIER <aderumier@odiso.com>
> Signed-off-by: Fam Zheng <famz@redhat.com>
> ---
>  include/block/blockjob.h | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
> 
> diff --git a/include/block/blockjob.h b/include/block/blockjob.h
> index 57d8ef1..53ac4f4 100644
> --- a/include/block/blockjob.h
> +++ b/include/block/blockjob.h
> @@ -157,6 +157,22 @@ void *block_job_create(const BlockJobDriver *driver, BlockDriverState *bs,
>   */
>  void block_job_sleep_ns(BlockJob *job, QEMUClockType type, int64_t ns);
>  
> +#define BLOCK_JOB_RELAX_CPU_NS 10000000L

By the way, why did you choose 10 milliseconds?  That is quite long.

If this function is called once per 10 ms disk I/O operations then we
lose 50% utilization.  1 ms or less would be reasonable.

> +
> +/**
> + * block_job_relax_cpu:
> + * @job: The job that calls the function.
> + *
> + * Sleep a little to avoid intensive cpu time occupation. Block jobs should
> + * call this or block_job_sleep_ns (for more precision, but note that 0 ns is
> + * usually not enought) periodically, otherwise the QMP and VCPU could starve

s/enought/enough/

> + * on CPU and/or BQL.
> + */
> +static inline void block_job_relax_cpu(BlockJob *job)

coroutine_fn is missing.