Re: [Qemu-devel] [Qemu-block] [PATCH] blockjob: kick jobs on set-speed

[John Snow <jsnow@redhat.com>]

On 12/13/2017 05:48 AM, Stefan Hajnoczi wrote:
> On Tue, Dec 12, 2017 at 01:22:28PM -0500, John Snow wrote:
>>
>>
>> On 12/11/2017 07:08 PM, Jeff Cody wrote:
>>> On Mon, Dec 11, 2017 at 06:46:09PM -0500, John Snow wrote:
>>>> If users set an unreasonably low speed (like one byte per second), the
>>>> calculated delay may exceed many hours. While we like to punish users
>>>> for asking for stupid things, we do also like to allow users to correct
>>>> their wicked ways.
>>>>
>>>> When a user provides a new speed, kick the job to allow it to recalculate
>>>> its delay.
>>>>
>>>> Signed-off-by: John Snow <jsnow@redhat.com>
>>>> ---
>>>>  blockjob.c | 5 +++++
>>>>  1 file changed, 5 insertions(+)
>>>>
>>>> diff --git a/blockjob.c b/blockjob.c
>>>> index 715c2c2680..43f01ad190 100644
>>>> --- a/blockjob.c
>>>> +++ b/blockjob.c
>>>> @@ -483,6 +483,7 @@ static void block_job_completed_txn_success(BlockJob *job)
>>>>  void block_job_set_speed(BlockJob *job, int64_t speed, Error **errp)
>>>>  {
>>>>      Error *local_err = NULL;
>>>> +    int64_t old_speed = job->speed;
>>>>  
>>>>      if (!job->driver->set_speed) {
>>>>          error_setg(errp, QERR_UNSUPPORTED);
>>>> @@ -495,6 +496,10 @@ void block_job_set_speed(BlockJob *job, int64_t speed, Error **errp)
>>>>      }
>>>>  
>>>>      job->speed = speed;
>>>> +    /* Kick the job to recompute its delay */
>>>> +    if ((speed > old_speed) && timer_pending(&job->sleep_timer)) {
>>>
>>> job->sleep_timer is protected by block_job_mutex (via
>>> block_job_lock/unlock); is it safe for us to check it here outside the
>>> mutex?
>>>
>>
>> My hunch is that in this specific case that it is; but only because of
>> assumptions about holding the aio_context and the QEMU global mutex here.
>>
>>> But in any case, I think we could get rid of the timer_pending check, and
>>> just always kick the job if we have a speed increase.  block_job_enter()
>>> should do the right thing (mutex protected check on job->busy and
>>> job->sleep_timer).
>>>
>>
>> I could lock it for inarguable correctness; I just didn't want to kick a
>> job that didn't actually require any kicking to limit any potential
>> problems from that interaction.
>>
>> (I'm fond of the extra conditional because I feel like it makes the
>> intent of the kick explicit.)
>>
>> I can remove it.
> 
> Removing the conditional would introduce a bug.  block_job_enter() will
> unpause the job.
> 

It will almost certainly pause again immediately, but maybe not before
work is performed.