[PATCH v4 6/6] pseries/sysfs: Minimise IPI noise while reading [idle_][s]purr

[Nathan Lynch]

Gautham R Shenoy <ego at linux.vnet.ibm.com> writes:
> On Thu, Apr 02, 2020 at 01:04:34PM +0530, Naveen N. Rao wrote:
>> >>
>> >>I wonder if we should introduce a sysctl interface to control thresholding.
>> >>It can default to 0, which disables thresholding so that the existing
>> >>behavior continues. Applications (lparstat) can optionally set it to suit
>> >>their use.
>> >
>> >We would be introducing 3 new sysfs interfaces that way instead of
>> >two.
>> >
>> >/sys/devices/system/cpu/purr_spurr_staleness
>> >/sys/devices/system/cpu/cpuX/idle_purr
>> >/sys/devices/system/cpu/cpuX/idle_spurr
>> >
>> >I don't have a problem with this. Nathan, Michael, thoughts on this?

No, I don't think this warrants a tunable when the issue it's intended
to address is still a bit speculative at this point. (Also, note that
this would be a system-wide value, but you could have multiple
concurrent users of the interface with different needs.)


>> >The alternative is to have a procfs interface, something like
>> >/proc/powerpc/resource_util_stats
>> >
>> >which gives a listing similar to /proc/stat, i.e
>> >
>> >      CPUX  <purr>  <idle_purr>  <spurr>  <idle_spurr>
>> >
>> >Even in this case, the values can be obtained in one-shot with a
>> >single IPI and be printed in the row corresponding to the CPU.
>> 
>> Right -- and that would be optimal requiring a single system call, at the
>> cost of using a legacy interface.
>> 
>> The other option would be to drop this patch and to just go with patches 1-5
>> introducing the new sysfs interfaces for idle_[s]purr. It isn't entirely
>> clear how often this would be used, or its actual impact. We can perhaps
>> consider this optimization if and when this causes problems...
>
> I am ok with that. We can revisit the problem if IPI noise becomes
> noticable. However, if Nathan or Michael feel that this problem is
> better solved now, than leaving it for the future, we will have to
> take a call on what the interface is going to be.

While I maintain some concern about the overhead on larger LPARs (150us
per CPU works out to ~0.15s total to serially sample 1024 CPUs, ~0.3s
for 2048 and so on), I am OK with the straightforward addition of the
attributes without any batching or sampling thresholds behind the scenes
for now. I appreciate your consideration of the issue.

If this turns out to be too inefficient then I think we should consider
a non-sysfs mechanism such as chardev+ioctl.