Re: [RFC][PATCH v2 1/7] taskstats: Add new taskstats command TASKSTATS_CMD_ATTR_PIDS

[Michael Holzheu <holzheu@linux.vnet.ibm.com>]

Hello Peter,

On Mon, 2010-11-15 at 18:21 +0100, Peter Zijlstra wrote:
> On Mon, 2010-11-15 at 18:09 +0100, Michael Holzheu wrote:
> 
> > > That you should not use sched_clock(),
> > 
> > What should we use instead?
> 
> Depends on what you want, look at kernel/sched_clock.c
> 
> > > What does last departed mean? That is what timeline are you counting in?
> > > Do you want time as tasks see it, or time as your wallclock sees it?
> > 
> > "last_depart" should be the time stamp, where the task has left a CPU
> > the last time.
> > 
> > We assume that we can compare "last_depart" with "time_ns" in the
> > taskstats structure,
> 
> I think you assume I actually know anything about taskstat :-), its the
> thing I always say =n to in my config file and have so far happily
> ignored all code of.
> 
> >  if we use task_rq(t)->clock for last_depart and
> > sched_clock() for stats->time_ns.
> 
> Then you're up shit creek because rq->clock doesn't necessarily have any
> correlation to sched_clock().
> 
> >  We also assume that we get wallclock
> > intervals in nanoseconds, if we look at two sched_clock() timestamps.
> 
> Invalid assumption.

Ok, thanks. So sched_clock() seems to be a bad idea for our purposes.

An alternative approach could be to have a global counter for the task
snapshots, which is increased each time a snapshot is created for
userspace. In addition to that we had to add a snapshot counter field to
the task_struct that is set to the current value of the global counter
each time a task leaves a CPU. Then userspace could ask for all tasks
that have been active after snapshot number x. In the response userspace
gets all tasks that have a snapshot number bigger than x together with
the new snapshot number y that can be used for the next query.

Still it would be useful to add a timestamp of the creation of the
taskstats data in the response to userspace for calculating the interval
time between two snapshots. Would the usage of ktime_get() be valid for
that purpose?

Michael