On 10.06.2023 19:46, Stephan Gerhold wrote:
> On Fri, Jun 09, 2023 at 10:19:24PM +0200, Konrad Dybcio wrote:
>> SMD RPM only provides two buckets, one each for the active-only and
>> active-sleep RPM contexts. Use the correct constant to allocate and
>> operate on them.
>>
>> Fixes: dcbce7b0a79c ("interconnect: qcom: icc-rpm: Support multiple buckets")
>> Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx>
>> Signed-off-by: Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
>> ---
>> drivers/interconnect/qcom/icc-rpm.c | 14 +++++++-------
>> 1 file changed, 7 insertions(+), 7 deletions(-)
>>
>> diff --git a/drivers/interconnect/qcom/icc-rpm.c b/drivers/interconnect/qcom/icc-rpm.c
>> index 6d40815c5401..3ac47b818afe 100644
>> --- a/drivers/interconnect/qcom/icc-rpm.c
>> +++ b/drivers/interconnect/qcom/icc-rpm.c
>> [...]
>> @@ -275,7 +275,7 @@ static int qcom_icc_bw_aggregate(struct icc_node *node, u32 tag, u32 avg_bw,
>> if (!tag)
>> tag = QCOM_ICC_TAG_ALWAYS;
>>
>> - for (i = 0; i < QCOM_ICC_NUM_BUCKETS; i++) {
>> + for (i = 0; i < QCOM_SMD_RPM_STATE_NUM; i++) {
>> if (tag & BIT(i)) {
>
> Hm, I think QCOM_ICC_NUM_BUCKETS is actually intentional here. There is
> a hint about this in the description of the commit in your Fixes line:
>
>> This patch studies the implementation from interconnect rpmh driver to
>> support multiple buckets. The rpmh driver provides three buckets for
>> AMC, WAKE, and SLEEP; this driver only needs to use WAKE and SLEEP
>> buckets, but we keep the same way with rpmh driver, this can allow us
>> to reuse the DT binding and avoid to define duplicated data structures.
>
> As far as I understand, the idea was to reuse the definitions in
> qcom,icc.h and just ignore the AMC bucket for now. AFAIU AMC (or rather
> the lack thereof) is basically caching: Sending requests without AMC bit
> set is delayed until the next rpmh_flush() call that happens when
> entering a deep idle state. It requires some work but I guess
> theoretically one could implement exactly the same for RPM.
That's trying to shove a cube into a circle-shaped hole.. AMC is a
hardware (well, firmware.. you know what I mean) feature, which SMD
RPM lacks. Plus it'd result in useless allocations.
>
> What you're actually doing here is not fixing the commit but changing
> the bindings. On MSM8909 I defined the ICC path for CPU<->RAM like this:
>
> interconnects = <&bimc MAS_APPS_PROC QCOM_ICC_TAG_ACTIVE_ONLY
> &bimc SLV_EBI QCOM_ICC_TAG_ACTIVE_ONLY>;
>
> Per definition in qcom,icc.h:
>
> QCOM_ICC_TAG_ACTIVE_ONLY = (AMC | WAKE) = (BIT(0) | BIT(1))
>
> Without your patch series this behaves correctly. It results in an
> active-only vote.
>
> The change of behavior is in PATCH 17/22 "interconnect: qcom: icc-rpm:
> Control bus rpmcc from icc". It silently switches from
> QCOM_ICC_BUCKET_WAKE (1) and QCOM_ICC_BUCKET_SLEEP (2) to
> QCOM_SMD_RPM_ACTIVE_STATE (0) and QCOM_SMD_RPM_SLEEP_STATE (1).
>
> In other words, QCOM_ICC_TAG_ACTIVE_ONLY (BIT(0) | BIT(1)) now results
> in an active+sleep vote, not an active-only one. :)
>
> There doesn't seem to be an upstream user of the ICC tags/buckets for
> icc-rpm yet so personally I would be fine with changing it. However,
> then qcom,icc.h should get a clear comment that it's rpmh-only and we
> should define a new qcom,icc-rpm.h.
Right, I'd argue the original commit was in the wrong here. It was trying
to reuse bindings which were intended for a different hw architecture (and
perhaps not described very well - there's no word of RPMh neither in the
name nor in the defines themselves). That's an abuse in my view..
I think introducing RPM-specific bindings and communicating the change
clearly is the way to go. As you've noticed, there are no users so that
should not be problematic at all.
Konrad
>
> Or perhaps we should just drop this patch and continue using
> QCOM_ICC_BUCKET_WAKE and QCOM_ICC_BUCKET_SLEEP as before?
>
> Thanks,
> Stephan
