Re: [PATCH v5] x86/resctrl: Add event choices for mba_MBps

[Tony Luck]

On Mon, Dec 04, 2023 at 12:03:23PM -0800, Reinette Chatre wrote:
> Hi Tony,
> 
> On 12/4/2023 11:45 AM, Luck, Tony wrote:
> >> Yes. I saw the thread. Even then I feel having two similar options can
> >> cause confusion. I feel it is enough just to solve the original problem.
> >> Giving more options to a corner cases is a overkill in my opinion.
> > 
> > The "original" problem was systems without "local" bandwidth event. I
> > wanted to give a way for users of mba_MBps to still have some way to
> > use it (assuming that "total" bandwidth event was present).
> > 
> > Reinette suggested that some people might want to use "total", even
> > on systems that support "local". I firmly agree with that.  It is easy to
> > construct scenarios where most bandwidth is to a remote node. using
> > "local" event will do nothing to throttle in these case. I'm not at all sure
> > why "local" event was picked. There's nothing in the LKML threads to
> > give clues.
> > 
> > I proposed a mount option "total" as a modifier to be used in conjunction
> > with "mba_MBps". Reinette said it was too generic. Her suggestion was
> > to add "mba_MBps_total" to be used instead of "mba_MBps".
> 
> No, it cannot be used instead of "mba_MBps". My intention was for it to be
> in addition to existing "mba_MBps" since taking "mba_MBps" away would be
> considered breaking user space ABI.

I was unclear. The mba_MBps option must remain as legacy ABI. My
"instead of" was intended to convey that a user wanting total bandwidth
would use:

# mount -t resctrl -o mba_MBps_total resctrl /sys/fs/resctrl

rather than the new option being a modifier and requiring both
the legacy option and the modifier like this:

# mount -t resctrl -o mba_MBps,mba_MBps_total resctrl /sys/fs/resctrl

which seems overly verbose.

> 
> Even so ...
> 
> > 
> > Is that where I should have gone, instead of "mba_MBps={local|total}"?
> 
> While I did propose "mba_MBps_total" (in addition to "mba_MBps") I do
> recognize your comment that a new key of mba_MBps_event does give more
> flexibility if different events become available in future. Emphasis is
> on "different" since I do not believe the parsing can support multiple
> events and thus mba_MBps_event cannot be treated as a general bucket for all
> mba_sc options, just different events guiding the feedback loop.
> 
> "mba_MBps" must be kept and having it continue to use local bw as default,
> but total bw on systems that do not support local bw seems appropriate,
> (which is what this patch does).

So we defintely have:

"mba_MBps" - defaults to local, on systems without local may switch to
total if that is available. Should this switch get a pr_info()? Or just happen
silently (as I've done in patches so far).

and need to come to agreement on which of these to implement:

A) "mba_MBps_total" - forces use of total. Fails the mount if total is not
   available.

B) "mba_MBps={local|total)" forces use of chosen event, fails if event
is unavailable.

C) Something else.

D) Don't provide any way to force use of total event.

> 
> Reinette

-Tony