Date:   Sat, 10 Jun 2023 21:09:18 +0200
From:   Konrad Dybcio <konrad.dybcio@...aro.org>
To:     Stephan Gerhold <stephan@...hold.net>
Cc:     Andy Gross <agross@...nel.org>,
        Bjorn Andersson <andersson@...nel.org>,
        Michael Turquette <mturquette@...libre.com>,
        Stephen Boyd <sboyd@...nel.org>,
        Georgi Djakov <djakov@...nel.org>,
        Leo Yan <leo.yan@...aro.org>,
        Evan Green <evgreen@...omium.org>,
        Marijn Suijten <marijn.suijten@...ainline.org>,
        linux-arm-msm@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-clk@...r.kernel.org, linux-pm@...r.kernel.org
Subject: Re: [PATCH v2 22/22] interconnect: qcom: icc-rpm: Fix bandwidth
 calculations



On 10.06.2023 21:06, Stephan Gerhold wrote:
> On Fri, Jun 09, 2023 at 10:19:27PM +0200, Konrad Dybcio wrote:
>> Up until now, we've been aggregating the bandwidth values and only
>> dividing them by the bus width of the source node. This was completely
>> wrong, as different nodes on a given path may (and usually do) have
>> varying bus widths.  That in turn, resulted in the calculated clock rates
>> being completely bogus - usually they ended up being much higher, as
>> NoC_A<->NoC_B links are very wide.
>>
>> Since we're not using the aggregate bandwidth value for anything other
>> than clock rate calculations, remodel qcom_icc_bus_aggregate() to
>> calculate the per-context clock rate for a given provider, taking into
>> account the bus width of every individual node.
>>
>> Fixes: 30c8fa3ec61a ("interconnect: qcom: Add MSM8916 interconnect provider driver")
>> Reported-by: Stephan Gerhold <stephan@...hold.net>
>> Signed-off-by: Konrad Dybcio <konrad.dybcio@...aro.org>
>> ---
>>  drivers/interconnect/qcom/icc-rpm.c | 59 ++++++++++++-------------------------
>>  1 file changed, 19 insertions(+), 40 deletions(-)
>>
>> diff --git a/drivers/interconnect/qcom/icc-rpm.c b/drivers/interconnect/qcom/icc-rpm.c
>> index 1508233632f6..d177a76abe2a 100644
>> --- a/drivers/interconnect/qcom/icc-rpm.c
>> +++ b/drivers/interconnect/qcom/icc-rpm.c
>> @@ -293,58 +293,44 @@ static int qcom_icc_bw_aggregate(struct icc_node *node, u32 tag, u32 avg_bw,
>>  }
>>  
>>  /**
>> - * qcom_icc_bus_aggregate - aggregate bandwidth by traversing all nodes
>> + * qcom_icc_bus_aggregate - calculate bus clock rates by traversing all nodes
>>   * @provider: generic interconnect provider
>> - * @agg_avg: an array for aggregated average bandwidth of buckets
>> - * @agg_peak: an array for aggregated peak bandwidth of buckets
>> - * @max_agg_avg: pointer to max value of aggregated average bandwidth
>> + * @agg_clk_rate: array containing the aggregated clock rates in kHz
>>   */
>> -static void qcom_icc_bus_aggregate(struct icc_provider *provider,
>> -				   u64 *agg_avg, u64 *agg_peak,
>> -				   u64 *max_agg_avg)
>> +static void qcom_icc_bus_aggregate(struct icc_provider *provider, u64 *agg_clk_rate)
>>  {
>> -	struct icc_node *node;
>> +	u64 agg_avg_rate, agg_rate;
>>  	struct qcom_icc_node *qn;
>> -	u64 sum_avg[QCOM_SMD_RPM_STATE_NUM];
>> +	struct icc_node *node;
>>  	int i;
>>  
>> -	/* Initialise aggregate values */
>> -	for (i = 0; i < QCOM_SMD_RPM_STATE_NUM; i++) {
>> -		agg_avg[i] = 0;
>> -		agg_peak[i] = 0;
>> -	}
>> -
>> -	*max_agg_avg = 0;
>> -
>>  	/*
>> -	 * Iterate nodes on the interconnect and aggregate bandwidth
>> -	 * requests for every bucket.
>> +	 * Iterate nodes on the provider, aggregate bandwidth requests for
>> +	 * every bucket and convert them into bus clock rates.
>>  	 */
>>  	list_for_each_entry(node, &provider->nodes, node_list) {
>>  		qn = node->data;
>>  		for (i = 0; i < QCOM_SMD_RPM_STATE_NUM; i++) {
>>  			if (qn->channels)
>> -				sum_avg[i] = div_u64(qn->sum_avg[i], qn->channels);
>> +				agg_avg_rate = div_u64(qn->sum_avg[i], qn->channels);
>>  			else
>> -				sum_avg[i] = qn->sum_avg[i];
>> -			agg_avg[i] += sum_avg[i];
>> -			agg_peak[i] = max_t(u64, agg_peak[i], qn->max_peak[i]);
>> +				agg_avg_rate = qn->sum_avg[i];
>> +
>> +			agg_rate = max_t(u64, agg_avg_rate, qn->max_peak[i]);
>> +			do_div(agg_rate, qn->buswidth);
>> +
>> +			agg_clk_rate[i] = max_t(u64, agg_clk_rate[i], agg_rate);
>>  		}
>>  	}
>> -
>> -	/* Find maximum values across all buckets */
>> -	for (i = 0; i < QCOM_SMD_RPM_STATE_NUM; i++)
>> -		*max_agg_avg = max_t(u64, *max_agg_avg, agg_avg[i]);
>>  }
>>  
>>  static int qcom_icc_set(struct icc_node *src, struct icc_node *dst)
>>  {
>> -	struct qcom_icc_provider *qp;
>>  	struct qcom_icc_node *src_qn = NULL, *dst_qn = NULL;
>> +	u64 agg_clk_rate[QCOM_SMD_RPM_STATE_NUM] = { 0 };
>>  	struct icc_provider *provider;
>> +	struct qcom_icc_provider *qp;
>>  	u64 active_rate, sleep_rate;
>> -	u64 agg_avg[QCOM_SMD_RPM_STATE_NUM], agg_peak[QCOM_SMD_RPM_STATE_NUM];
>> -	u64 max_agg_avg;
>>  	int ret;
>>  
>>  	src_qn = src->data;
>> @@ -353,7 +339,9 @@ static int qcom_icc_set(struct icc_node *src, struct icc_node *dst)
>>  	provider = src->provider;
>>  	qp = to_qcom_provider(provider);
>>  
>> -	qcom_icc_bus_aggregate(provider, agg_avg, agg_peak, &max_agg_avg);
>> +	qcom_icc_bus_aggregate(provider, agg_clk_rate);
>> +	active_rate = agg_clk_rate[QCOM_SMD_RPM_ACTIVE_STATE];
>> +	sleep_rate = agg_clk_rate[QCOM_SMD_RPM_SLEEP_STATE];
>>  
>>  	ret = qcom_icc_rpm_set(src_qn, src_qn->sum_avg);
>>  	if (ret)
>> @@ -369,15 +357,6 @@ static int qcom_icc_set(struct icc_node *src, struct icc_node *dst)
>>  	if (!qp->bus_clk_desc && !qp->bus_clk)
>>  		return 0;
>>  
>> -	/* Intentionally keep the rates in kHz as that's what RPM accepts */
> 
> I kind of liked this comment because otherwise it's not obvious why
> you're not converting from "ICC units" anywhere.
I figured the kerneldoc change would be enough:

@agg_clk_rate: array containing the aggregated clock rates in kHz

> 
> Anyway:
> 
> Reviewed-by: Stephan Gerhold <stephan@...hold.net>
> 
> Thanks for going through the giant maze to fix this!
Thanks for being there along the way.. This spaghetti is far too much
for a single human..

Konrad

> 
> Stephan

Powered by blists - more mailing lists