Re: [PATCH] proc/schedstat: Expose /proc/<pid>/schedstat if delay accounting is enabled

[Ingo Molnar <mingo@kernel.org>]

* Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com> wrote:

> On 2015/06/02 09:58AM, Ingo Molnar wrote:
> > 
> > * Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com> wrote:
> > 
> > > On 2015/05/29 11:54AM, Ingo Molnar wrote:
> > > > 
> > > > > On a related note, even though sum_exec_runtime is available 
> > > > > unconditionally, I dump all zeroes in my patch if 
> > > > > !sched_info_on() to make it clear that some of the fields are 
> > > > > not available. Is this ok or should be display sum_exec_runtime 
> > > > > regardless of sched_info?
> > > > 
> > > > So I'd suggest printing -1 for non-available fields, that should be unambigous 
> > > > enough and makes it also possible to write out 0 in some cases.
> > > 
> > > Per Documentation/scheduler/sched-stats.txt (and the linked latency.c there), 
> > > user-space seems to be expecting unsigned values here. Would displaying -1 here 
> > > be ok?
> > 
> > Probably not (the code is silly, why doesn't it split up the string and use 
> > atol()?) - hopefully real user-space is better? Can you try some real, packaged up 
> > tools that read schedstats, to see whether they work with -1?
> 
> Hi Ingo,
> Sorry for the delay - I had been off on vacation.
> 
> I see that quite a few packages are using /proc/<pid>/schedstat - pcp, systemd, 
> dstat, android, among others. While most of these seem to be splitting up the 
> fields properly, they are using a variant of strtoull(), which returns 
> ULLONG_MAX for -1, and none of these check for that condition. If any of the 
> tools use the value read to report total execution time or run delay, it will be 
> incorrect.
> 
> At this point, I feel it is better to display all the three fields in schedstat 
> only if sched_info_on() is true, as explained above. What do you suggest?

Ok, agreed - thanks for the analysis.

Mind (re-)submitting the patches accordingly?

Thanks,

	Ingo