[1/4] interconnect: qcom: icc-rpm: Add AB/IB calculations coefficients
Commit Message
Presumably due to the hardware being so complex, some nodes (or busses)
have different (usually higher) requirements for bandwidth than what
the usual calculations would suggest.
Looking at the available downstream files, it seems like AB values are
adjusted per-bus and IB values are adjusted per-node.
With that in mind, introduce percentage-based coefficient struct members
and use them in the calculations.
One thing to note is that downstream does (X%)*AB and IB/(Y%) which
feels a bit backwards, especially given that the divisors for IB turn
out to always be 25, 50, 200 making this a convenient conversion to 4x,
2x, 0.5x.. This commit uses the more sane, non-inverse approach.
Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
---
drivers/interconnect/qcom/icc-rpm.c | 10 +++++++++-
drivers/interconnect/qcom/icc-rpm.h | 5 +++++
2 files changed, 14 insertions(+), 1 deletion(-)
Comments
On Wed, Jul 26, 2023 at 06:25:43PM +0200, Konrad Dybcio wrote:
> Presumably due to the hardware being so complex, some nodes (or busses)
> have different (usually higher) requirements for bandwidth than what
> the usual calculations would suggest.
>
Weird. I just hope this was never abused to workaround other broken
configuration. A nice round ib_percent = 200 has mostly the same effect as
- Doubling the requested peek bandwidth in the consumer driver (perhaps
they were too lazy to fix the driver in downstream at some point)
- Halving the node buswidth
It's probably hard to say for sure...
> Looking at the available downstream files, it seems like AB values are
> adjusted per-bus and IB values are adjusted per-node.
> With that in mind, introduce percentage-based coefficient struct members
> and use them in the calculations.
>
> One thing to note is that downstream does (X%)*AB and IB/(Y%) which
> feels a bit backwards, especially given that the divisors for IB turn
> out to always be 25, 50, 200 making this a convenient conversion to 4x,
> 2x, 0.5x.. This commit uses the more sane, non-inverse approach.
>
> Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
> ---
> drivers/interconnect/qcom/icc-rpm.c | 10 +++++++++-
> drivers/interconnect/qcom/icc-rpm.h | 5 +++++
> 2 files changed, 14 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/interconnect/qcom/icc-rpm.c b/drivers/interconnect/qcom/icc-rpm.c
> index 2c16917ba1fd..2de0e1dfe225 100644
> --- a/drivers/interconnect/qcom/icc-rpm.c
> +++ b/drivers/interconnect/qcom/icc-rpm.c
> @@ -298,9 +298,11 @@ static int qcom_icc_bw_aggregate(struct icc_node *node, u32 tag, u32 avg_bw,
> */
> static void qcom_icc_bus_aggregate(struct icc_provider *provider, u64 *agg_clk_rate)
> {
> - u64 agg_avg_rate, agg_rate;
> + struct qcom_icc_provider *qp = to_qcom_provider(provider);
> + u64 agg_avg_rate, agg_peak_rate, agg_rate;
> struct qcom_icc_node *qn;
> struct icc_node *node;
> + u16 percent;
> int i;
>
> /*
> @@ -315,6 +317,12 @@ static void qcom_icc_bus_aggregate(struct icc_provider *provider, u64 *agg_clk_r
> else
> agg_avg_rate = qn->sum_avg[i];
>
> + percent = qp->ab_percent ? qp->ab_percent : 100;
> + agg_avg_rate = mult_frac(percent, agg_avg_rate, 100);
if (qp->ab_percent)
agg_avg_rate = mult_frac(qp->ab_percent, agg_avg_rate, 100);
Would be likely more efficient (no calculation if unspecified) and not
much harder to read.
> +
> + percent = qn->ib_percent ? qn->ib_percent : 100;
> + agg_peak_rate = mult_frac(percent, qn->max_peak[i], 100);
> +
agg_peak_rate doesn't seem to be used anywhere else? 🤔
Thanks,
Stephan
On 26.07.2023 19:16, Stephan Gerhold wrote:
> On Wed, Jul 26, 2023 at 06:25:43PM +0200, Konrad Dybcio wrote:
>> Presumably due to the hardware being so complex, some nodes (or busses)
>> have different (usually higher) requirements for bandwidth than what
>> the usual calculations would suggest.
>>
>
> Weird. I just hope this was never abused to workaround other broken
> configuration. A nice round ib_percent = 200 has mostly the same effect as
>
> - Doubling the requested peek bandwidth in the consumer driver (perhaps
> they were too lazy to fix the driver in downstream at some point)
> - Halving the node buswidth
>
> It's probably hard to say for sure...
As per usual..
[...]
>>
>> /*
>> @@ -315,6 +317,12 @@ static void qcom_icc_bus_aggregate(struct icc_provider *provider, u64 *agg_clk_r
>> else
>> agg_avg_rate = qn->sum_avg[i];
>>
>> + percent = qp->ab_percent ? qp->ab_percent : 100;
>> + agg_avg_rate = mult_frac(percent, agg_avg_rate, 100);
>
> if (qp->ab_percent)
> agg_avg_rate = mult_frac(qp->ab_percent, agg_avg_rate, 100);
>
> Would be likely more efficient (no calculation if unspecified) and not
> much harder to read.
Oh right!
>
>> +
>> + percent = qn->ib_percent ? qn->ib_percent : 100;
>> + agg_peak_rate = mult_frac(percent, qn->max_peak[i], 100);
>> +
>
> agg_peak_rate doesn't seem to be used anywhere else? 🤔
Whoooooops....
Konrad
Hi Konrad,
kernel test robot noticed the following build warnings:
[auto build test WARNING on 1e25dd7772483f477f79986d956028e9f47f990a]
url: https://github.com/intel-lab-lkp/linux/commits/Konrad-Dybcio/interconnect-qcom-icc-rpm-Add-AB-IB-calculations-coefficients/20230727-002710
base: 1e25dd7772483f477f79986d956028e9f47f990a
patch link: https://lore.kernel.org/r/20230726-topic-icc_coeff-v1-1-31616960818c%40linaro.org
patch subject: [PATCH 1/4] interconnect: qcom: icc-rpm: Add AB/IB calculations coefficients
config: arm-allmodconfig (https://download.01.org/0day-ci/archive/20230729/202307291745.0JcdYvBz-lkp@intel.com/config)
compiler: arm-linux-gnueabi-gcc (GCC) 12.3.0
reproduce: (https://download.01.org/0day-ci/archive/20230729/202307291745.0JcdYvBz-lkp@intel.com/reproduce)
If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <lkp@intel.com>
| Closes: https://lore.kernel.org/oe-kbuild-all/202307291745.0JcdYvBz-lkp@intel.com/
All warnings (new ones prefixed by >>):
drivers/interconnect/qcom/icc-rpm.c: In function 'qcom_icc_bus_aggregate':
>> drivers/interconnect/qcom/icc-rpm.c:302:27: warning: variable 'agg_peak_rate' set but not used [-Wunused-but-set-variable]
302 | u64 agg_avg_rate, agg_peak_rate, agg_rate;
| ^~~~~~~~~~~~~
vim +/agg_peak_rate +302 drivers/interconnect/qcom/icc-rpm.c
293
294 /**
295 * qcom_icc_bus_aggregate - calculate bus clock rates by traversing all nodes
296 * @provider: generic interconnect provider
297 * @agg_clk_rate: array containing the aggregated clock rates in kHz
298 */
299 static void qcom_icc_bus_aggregate(struct icc_provider *provider, u64 *agg_clk_rate)
300 {
301 struct qcom_icc_provider *qp = to_qcom_provider(provider);
> 302 u64 agg_avg_rate, agg_peak_rate, agg_rate;
303 struct qcom_icc_node *qn;
304 struct icc_node *node;
305 u16 percent;
306 int i;
307
308 /*
309 * Iterate nodes on the provider, aggregate bandwidth requests for
310 * every bucket and convert them into bus clock rates.
311 */
312 list_for_each_entry(node, &provider->nodes, node_list) {
313 qn = node->data;
314 for (i = 0; i < QCOM_SMD_RPM_STATE_NUM; i++) {
315 if (qn->channels)
316 agg_avg_rate = div_u64(qn->sum_avg[i], qn->channels);
317 else
318 agg_avg_rate = qn->sum_avg[i];
319
320 percent = qp->ab_percent ? qp->ab_percent : 100;
321 agg_avg_rate = mult_frac(percent, agg_avg_rate, 100);
322
323 percent = qn->ib_percent ? qn->ib_percent : 100;
324 agg_peak_rate = mult_frac(percent, qn->max_peak[i], 100);
325
326 agg_rate = max_t(u64, agg_avg_rate, qn->max_peak[i]);
327 do_div(agg_rate, qn->buswidth);
328
329 agg_clk_rate[i] = max_t(u64, agg_clk_rate[i], agg_rate);
330 }
331 }
332 }
333
@@ -298,9 +298,11 @@ static int qcom_icc_bw_aggregate(struct icc_node *node, u32 tag, u32 avg_bw,
*/
static void qcom_icc_bus_aggregate(struct icc_provider *provider, u64 *agg_clk_rate)
{
- u64 agg_avg_rate, agg_rate;
+ struct qcom_icc_provider *qp = to_qcom_provider(provider);
+ u64 agg_avg_rate, agg_peak_rate, agg_rate;
struct qcom_icc_node *qn;
struct icc_node *node;
+ u16 percent;
int i;
/*
@@ -315,6 +317,12 @@ static void qcom_icc_bus_aggregate(struct icc_provider *provider, u64 *agg_clk_r
else
agg_avg_rate = qn->sum_avg[i];
+ percent = qp->ab_percent ? qp->ab_percent : 100;
+ agg_avg_rate = mult_frac(percent, agg_avg_rate, 100);
+
+ percent = qn->ib_percent ? qn->ib_percent : 100;
+ agg_peak_rate = mult_frac(percent, qn->max_peak[i], 100);
+
agg_rate = max_t(u64, agg_avg_rate, qn->max_peak[i]);
do_div(agg_rate, qn->buswidth);
@@ -45,6 +45,7 @@ struct rpm_clk_resource {
* @regmap: regmap for QoS registers read/write access
* @qos_offset: offset to QoS registers
* @bus_clk_rate: bus clock rate in Hz
+ * @ab_percent: a percentage-based coefficient for compensating the AB calculations
* @bus_clk_desc: a pointer to a rpm_clk_resource description of bus clocks
* @bus_clk: a pointer to a HLOS-owned bus clock
* @intf_clks: a clk_bulk_data array of interface clocks
@@ -58,6 +59,7 @@ struct qcom_icc_provider {
struct regmap *regmap;
unsigned int qos_offset;
u32 bus_clk_rate[QCOM_SMD_RPM_STATE_NUM];
+ u16 ab_percent;
const struct rpm_clk_resource *bus_clk_desc;
struct clk *bus_clk;
struct clk_bulk_data *intf_clks;
@@ -93,6 +95,7 @@ struct qcom_icc_qos {
* @num_links: the total number of @links
* @channels: number of channels at this node (e.g. DDR channels)
* @buswidth: width of the interconnect between a node and the bus (bytes)
+ * @ib_percent: a percentage-based coefficient for compensating the IB calculations
* @sum_avg: current sum aggregate value of all avg bw requests
* @max_peak: current max aggregate value of all peak bw requests
* @mas_rpm_id: RPM id for devices that are bus masters
@@ -106,6 +109,7 @@ struct qcom_icc_node {
u16 num_links;
u16 channels;
u16 buswidth;
+ u16 ib_percent;
u64 sum_avg[QCOM_SMD_RPM_STATE_NUM];
u64 max_peak[QCOM_SMD_RPM_STATE_NUM];
int mas_rpm_id;
@@ -123,6 +127,7 @@ struct qcom_icc_desc {
enum qcom_icc_type type;
const struct regmap_config *regmap_cfg;
unsigned int qos_offset;
+ u16 ab_percent;
};
/* Valid for all bus types */