Hi Ben,
On 28/08/2025 17:13, Ben Horgan wrote:
> On 8/22/25 16:30, James Morse wrote:
>> When CPUs come online the original configuration should be restored.
>> Once the maximum partid is known, allocate an configuration array for
>> each component, and reprogram each RIS configuration from this.
>>
>> The MPAM spec describes how multiple controls can interact. To prevent
>> this happening by accident, always reset controls that don't have a
>> valid configuration. This allows the same helper to be used for
>> configuration and reset.
> What in particular are you worried about here?
'other' controls being left in an unknown state - meaning the one you did set, is useless.
In a sane world, the thing writing the controls would write all the supported registers.
In practice, resctrl only knows about bitmaps. The glue code could provide all the other
values, but I figured it was better for the driver to do it. I'm sure they'll add other
control types, and it would be a nuisance to update multiple callers if there is ever more
than one.
Another angle down here is mismatched components/devices mean a control type could be
hidden if its not available everywhere - so the caller may not be aware of all the
controls it was supposed to provide.
> It does seem a bit
> wasteful that to update a single control in a ris all the controls in
> that ris are updated.
I don't think anyone would ever build something that supports all these. One is most
likely, pushing to three for platforms that support CPOR and CMIN/MAX. By the time you've
taken the IPI to access a cache MSC, the cost of an additional register access is negligible.
> This is needed for reset and restore but do we
> really want if we are just changing one control, e.g. the cache portion
> bitmap.
The original config has been blown away by this point, but we do have the bitmap of what
changed. I guess this is an emergent effect of __write_config() originating from the reset
helper, and the 'empty config' being used to reset devices.
I'd like to keep it as a single function that actually touches these registers.
I'll change to generate a 'maximal' config instead of empty for reset - which pulls the
policy on those values out, and drops the '0 for reset'.
huh ... that is what the ALL_FEATURES macro you pointed out was for ...
I suspect it was the bitmaps that are larger than a u32 that made this hard.
>> diff --git a/drivers/resctrl/mpam_devices.c b/drivers/resctrl/mpam_devices.c
>> index b424af666b1e..8f6df2406c22 100644
>> --- a/drivers/resctrl/mpam_devices.c
>> +++ b/drivers/resctrl/mpam_devices.c
>> @@ -1960,6 +2094,79 @@ void mpam_enable(struct work_struct *work)
>> mpam_enable_once();
>> }
>>
>> +struct mpam_write_config_arg {
>> + struct mpam_msc_ris *ris;
>> + struct mpam_component *comp;
>> + u16 partid;
>> +};
>> +
>> +static int __write_config(void *arg)
>> +{
>> + struct mpam_write_config_arg *c = arg;
>> +
>> + mpam_reprogram_ris_partid(c->ris, c->partid, &c->comp->cfg[c->partid]);
>> +
>> + return 0;
>> +}
>> +
>> +#define maybe_update_config(cfg, feature, newcfg, member, changes) do { \
>> + if (mpam_has_feature(feature, newcfg) && \
>> + (newcfg)->member != (cfg)->member) { \
>> + (cfg)->member = (newcfg)->member; \
>> + cfg->features |= (1 << feature); \
>> + \
>> + (changes) |= (1 << feature); \
>> + } \
>> +} while (0)
>> +
>> +static mpam_features_t mpam_update_config(struct mpam_config *cfg,
>> + const struct mpam_config *newcfg)
>> +{
>> + mpam_features_t changes = 0;
>> +
>> + maybe_update_config(cfg, mpam_feat_cpor_part, newcfg, cpbm, changes);
>> + maybe_update_config(cfg, mpam_feat_mbw_part, newcfg, mbw_pbm, changes);
>> + maybe_update_config(cfg, mpam_feat_mbw_max, newcfg, mbw_max, changes);
>> +
>> + return changes;
>> +}
>> +
>> +/* TODO: split into write_config/sync_config */
>> +/* TODO: add config_dirty bitmap to drive sync_config */
> Any changes to come for these TODO comments?
No time. The dirty bitmap was to help with the problem you highlighted above. Separating
into write/sync was to make it easier to support the firmware-backed thing, which can be
a problem for another day.
I'll drop these.
Thanks,
James
