Re: [PATCH v4 4/4] arm64: dts: qcom: sdm845: Add CPU BWMON

From: Rajendra Nayak
Date: Tue Jun 28 2022 - 11:20:34 EST




On 6/28/2022 7:32 PM, Krzysztof Kozlowski wrote:
On 28/06/2022 15:15, Rajendra Nayak wrote:


On 6/28/2022 4:20 PM, Krzysztof Kozlowski wrote:
On 28/06/2022 12:36, Rajendra Nayak wrote:

On 6/27/2022 6:09 PM, Krzysztof Kozlowski wrote:
On 26/06/2022 05:28, Bjorn Andersson wrote:
On Thu 23 Jun 07:58 CDT 2022, Krzysztof Kozlowski wrote:

On 23/06/2022 08:48, Rajendra Nayak wrote:
diff --git a/arch/arm64/boot/dts/qcom/sdm845.dtsi b/arch/arm64/boot/dts/qcom/sdm845.dtsi
index 83e8b63f0910..adffb9c70566 100644
--- a/arch/arm64/boot/dts/qcom/sdm845.dtsi
+++ b/arch/arm64/boot/dts/qcom/sdm845.dtsi
@@ -2026,6 +2026,60 @@ llcc: system-cache-controller@1100000 {
interrupts = <GIC_SPI 582 IRQ_TYPE_LEVEL_HIGH>;
};
+ pmu@1436400 {
+ compatible = "qcom,sdm845-cpu-bwmon";
+ reg = <0 0x01436400 0 0x600>;
+
+ interrupts = <GIC_SPI 581 IRQ_TYPE_LEVEL_HIGH>;
+
+ interconnects = <&gladiator_noc MASTER_APPSS_PROC 3 &mem_noc SLAVE_EBI1 3>,
+ <&osm_l3 MASTER_OSM_L3_APPS &osm_l3 SLAVE_OSM_L3>;
+ interconnect-names = "ddr", "l3c";

Is this the pmu/bwmon instance between the cpu and caches or the one between the caches and DDR?

To my understanding this is the one between CPU and caches.

Ok, but then because the OPP table lists the DDR bw first and Cache bw second, isn't the driver
ending up comparing the bw values thrown by the pmu against the DDR bw instead of the Cache BW?

I double checked now and you're right.

Atleast with my testing on sc7280 I found this to mess things up and I always was ending up at
higher OPPs even while the system was completely idle. Comparing the values against the Cache bw
fixed it.(sc7280 also has a bwmon4 instance between the cpu and caches and a bwmon5 between the cache
and DDR)

In my case it exposes different issue - under performance. Somehow the
bwmon does not report bandwidth high enough to vote for high bandwidth.

After removing the DDR interconnect and bandwidth OPP values I have for:
sysbench --threads=8 --time=60 --memory-total-size=20T --test=memory
--memory-block-size=4M run

1. Vanilla: 29768 MB/s
2. Vanilla without CPU votes: 8728 MB/s
3. Previous bwmon (voting too high): 32007 MB/s
4. Fixed bwmon 24911 MB/s
Bwmon does not vote for maximum L3 speed:
bwmon report 9408 MB/s (thresholds set: <9216000 15052801>
)
osm l3 aggregate 14355 MBps -> 897 MHz, level 7, bw 14355 MBps

Maybe that's just problem with missing governor which would vote for
bandwidth rounding up or anticipating higher needs.

Depending on which one it is, shouldn;t we just be scaling either one and not both the interconnect paths?

The interconnects are the same as ones used for CPU nodes, therefore if
we want to scale both when scaling CPU, then we also want to scale both
when seeing traffic between CPU and cache.

Well, they were both associated with the CPU node because with no other input to decide on _when_
to scale the caches and DDR, we just put a mapping table which simply mapped a CPU freq to a L3 _and_
DDR freq. So with just one input (CPU freq) we decided on what should be both the L3 freq and DDR freq.

Now with 2 pmu's, we have 2 inputs, so we can individually scale the L3 based on the cache PMU
counters and DDR based on the DDR PMU counters, no?

Since you said you have plans to add the other pmu support as well (bwmon5 between the cache and DDR)
how else would you have the OPP table associated with that pmu instance? Would you again have both the
L3 and DDR scale based on the inputs from that bwmon too?

Good point, thanks for sharing. I think you're right. I'll keep only the
l3c interconnect path.


If I understand correctly, <&osm_l3 MASTER_OSM_L3_APPS &osm_l3
SLAVE_OSM_L3> relates to the L3 cache speed, which sits inside the CPU
subsystem. As such traffic hitting this cache will not show up in either
bwmon instance.

The path <&gladiator_noc MASTER_APPSS_PROC 3 &mem_noc SLAVE_EBI1 3>
affects the DDR frequency. So the traffic measured by the cpu-bwmon
would be the CPU subsystems traffic that missed the L1/L2/L3 caches and
hits the memory bus towards DDR.

That seems right, looking some more into the downstream code and register definitions
I see the 2 bwmon instances actually lie on the path outside CPU SS towards DDR,
first one (bwmon4) is between the CPUSS and LLCC (system cache) and the second one
(bwmon5) between LLCC and DDR. So we should use the counters from bwmon4 to
scale the CPU-LLCC path (and not L3), on sc7280 that would mean splitting the
<&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3> into
<&gem_noc MASTER_APPSS_PROC 3 &gem_noc SLAVE_LLCC 3> (voting based on the bwmon4 inputs)

For sdm845 SLAVE_LLCC is in mem_noc, so I guess mc_virt on sc7280?

thats correct,


and <&mc_virt MASTER_LLCC 3 &mc_virt SLAVE_EBI1 3> (voting based on the bwmon5 inputs)
and similar for sdm845 too.

L3 should perhaps still be voted based on the cpu freq as done today.

This would mean that original bandwidth values (800 - 7216 MB/s) were
correct. However we have still your observation that bwmon kicks in very
fast and my measurements that sampled bwmon data shows bandwidth ~20000
MB/s.

Right, thats because the bandwidth supported between the cpu<->llcc path is much higher
than the DDR frequencies. For instance on sc7280, I see (2288 - 15258 MB/s) for LLCC while
the DDR max is 8532 MB/s.

OK, that sounds right.

Another point is that I did not find actual scaling of throughput via
that interconnect path:
<&gladiator_noc MASTER_APPSS_PROC 3 &mem_noc SLAVE_LLCC 3>

Shouldn't this be <&gladiator_noc MASTER_APPSS_PROC 3 &gladiator_noc SLAVE_LLCC 3> on sdm845?


so I cannot test impact of bwmon that way.

Best regards,
Krzysztof