Re: [PATCH v2 09/18] x86/resctrl: Allow resctrl_arch_rmid_read() to sleep

[Peter Newman]

Hi James,

On Mon, Mar 6, 2023 at 12:34 PM James Morse <james.morse@xxxxxxx> wrote:
> On 23/01/2023 15:33, Peter Newman wrote:
> > On Fri, Jan 13, 2023 at 6:56 PM James Morse <james.morse@xxxxxxx> wrote:
> >> MPAM's cache occupancy counters can take a little while to settle once
> >> the monitor has been configured. The maximum settling time is described
> >> to the driver via a firmware table. The value could be large enough
> >> that it makes sense to sleep.
> >
> > Would it be easier to return an error when reading the occupancy count
> > too soon after configuration? On Intel it is already normal for counter
> > reads to fail on newly-allocated RMIDs.
>
> For x86, you have as many counters as there are RMIDs, so there is no issue just accessing
> the counter.

I should have said AMD instead of Intel, because their implementations
have far fewer counters than RMIDs.

>
> With MPAM there may be as few as 1 monitor for the CSU (cache storage utilisation)
> counter, which needs to be multiplexed between different PARTID to find the cache
> occupancy (This works for CSU because its a stable count, it doesn't work for the
> bandwidth monitors)
> On such a platform the monitor needs to be allocated and programmed before it reads a
> value for a particular PARTID/CLOSID. If you had two threads trying to read the same
> counter, they could interleave perfectly to prevent either thread managing to read a value.
> The 'not ready' time is advertised in a firmware table, and the driver will wait at most
> that long before giving up and returning an error.

Likewise, on AMD, a repeating sequence of tasks which are LRU in terms
of counter -> RMID allocation could prevent RMID event reads from ever
returning a value.

The main difference I see with MPAM is that software allocates the
counters instead of hardware, but the overall behavior sounds the same.

The part I object to is introducing the wait to the counter read because
existing software already expects an immediate error when reading a
counter too soon. To produce accurate data, these readings are usually
read at intervals of multiple seconds.

Instead, when configuring a counter, could you use the firmware table
value to compute the time when the counter will next be valid and return
errors on read requests received before that?

-Peter