| Message ID | 1619792901-32701-1-git-send-email-sibis@codeaurora.org |
|---|---|
| Headers | show |
| Series | DDR/L3 Scaling support on SC7280 SoCs | expand |
On Fri, Apr 30, 2021 at 07:58:21PM +0530, Sibi Sankar wrote: > Add OPP tables required to scale DDR/L3 per freq-domain on SC7280 SoCs. > > Signed-off-by: Sibi Sankar <sibis@codeaurora.org> > --- > arch/arm64/boot/dts/qcom/sc7280.dtsi | 135 +++++++++++++++++++++++++++++++++++ > 1 file changed, 135 insertions(+) > > diff --git a/arch/arm64/boot/dts/qcom/sc7280.dtsi b/arch/arm64/boot/dts/qcom/sc7280.dtsi > index 0bb835aeae33..90220cecb368 100644 > --- a/arch/arm64/boot/dts/qcom/sc7280.dtsi > +++ b/arch/arm64/boot/dts/qcom/sc7280.dtsi > @@ -7,6 +7,7 @@ > > #include <dt-bindings/clock/qcom,gcc-sc7280.h> > #include <dt-bindings/clock/qcom,rpmh.h> > +#include <dt-bindings/interconnect/qcom,osm-l3.h> > #include <dt-bindings/interconnect/qcom,sc7280.h> > #include <dt-bindings/interrupt-controller/arm-gic.h> > #include <dt-bindings/mailbox/qcom-ipcc.h> > @@ -71,6 +72,9 @@ > &LITTLE_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_0>; > + operating-points-v2 = <&cpu0_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; This patch seems to depend on the 'Add SC7280 interconnect provider driver' series (https://patchwork.kernel.org/project/linux-arm-msm/list/?series=473747) and 'Add L3 provider support for SC7280' (https://patchwork.kernel.org/project/linux-arm-msm/list/?series=468285), none of them has landed yet. The dependencies should be mentioned in the commit notes (under '---'). > qcom,freq-domain = <&cpufreq_hw 0>; > L2_0: l2-cache { > compatible = "cache"; > @@ -90,6 +94,9 @@ > &LITTLE_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_100>; > + operating-points-v2 = <&cpu0_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; > qcom,freq-domain = <&cpufreq_hw 0>; > L2_100: l2-cache { > compatible = "cache"; > @@ -106,6 +113,9 @@ > &LITTLE_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_200>; > + operating-points-v2 = <&cpu0_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; > qcom,freq-domain = <&cpufreq_hw 0>; > L2_200: l2-cache { > compatible = "cache"; > @@ -122,6 +132,9 @@ > &LITTLE_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_300>; > + operating-points-v2 = <&cpu0_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; > qcom,freq-domain = <&cpufreq_hw 0>; > L2_300: l2-cache { > compatible = "cache"; > @@ -138,6 +151,9 @@ > &BIG_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_400>; > + operating-points-v2 = <&cpu4_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; > qcom,freq-domain = <&cpufreq_hw 1>; > L2_400: l2-cache { > compatible = "cache"; > @@ -154,6 +170,9 @@ > &BIG_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_500>; > + operating-points-v2 = <&cpu4_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; > qcom,freq-domain = <&cpufreq_hw 1>; > L2_500: l2-cache { > compatible = "cache"; > @@ -170,6 +189,9 @@ > &BIG_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_600>; > + operating-points-v2 = <&cpu4_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; > qcom,freq-domain = <&cpufreq_hw 1>; > L2_600: l2-cache { > compatible = "cache"; > @@ -186,6 +208,9 @@ > &BIG_CPU_SLEEP_1 > &CLUSTER_SLEEP_0>; > next-level-cache = <&L2_700>; > + operating-points-v2 = <&cpu4_opp_table>; > + interconnects = <&gem_noc MASTER_APPSS_PROC 3 &mc_virt SLAVE_EBI1 3>, > + <&epss_l3 MASTER_EPSS_L3_APPS &epss_l3 SLAVE_EPSS_L3_SHARED>; > qcom,freq-domain = <&cpufreq_hw 2>; > L2_700: l2-cache { > compatible = "cache"; > @@ -248,6 +273,116 @@ > }; > }; > > + cpu0_opp_table: cpu0_opp_table { the node name should use dashes as separators instead of underscores, i.e. it should be 'cpu0-opp-table'. > + compatible = "operating-points-v2"; > + opp-shared; > + > + cpu0_opp1: opp-300000000 { > + opp-hz = /bits/ 64 <300000000>; > + opp-peak-kBps = <800000 9600000>; > + }; > + > + cpu0_opp2: opp-691200000 { > + opp-hz = /bits/ 64 <691200000>; > + opp-peak-kBps = <800000 17817600>; > + }; > + > + cpu0_opp3: opp-806400000 { > + opp-hz = /bits/ 64 <806400000>; > + opp-peak-kBps = <800000 20889600>; > + }; > + > + cpu0_opp4: opp-940800000 { > + opp-hz = /bits/ 64 <940800000>; > + opp-peak-kBps = <1804000 24576000>; > + }; > + > + cpu0_opp5: opp-1152000000 { > + opp-hz = /bits/ 64 <1152000000>; > + opp-peak-kBps = <2188000 27033600>; > + }; > + > + cpu0_opp6: opp-1324800000 { > + opp-hz = /bits/ 64 <1324800000>; > + opp-peak-kBps = <2188000 33792000>; > + }; > + > + cpu0_opp7: opp-1516800000 { > + opp-hz = /bits/ 64 <1516800000>; > + opp-peak-kBps = <3072000 38092800>; > + }; > + > + cpu0_opp8: opp-1651200000 { > + opp-hz = /bits/ 64 <1651200000>; > + opp-peak-kBps = <3072000 41779200>; > + }; > + > + cpu0_opp9: opp-1804800000 { > + opp-hz = /bits/ 64 <1804800000>; > + opp-peak-kBps = <4068000 48537600>; > + }; > + > + cpu0_opp10: opp-1958400000 { > + opp-hz = /bits/ 64 <1958400000>; > + opp-peak-kBps = <4068000 48537600>; > + }; > + }; > + > + cpu4_opp_table: cpu4_opp_table { node name should be 'cpu4-opp-table' > + compatible = "operating-points-v2"; > + opp-shared; > + > + cpu4_opp1: opp-691200000 { > + opp-hz = /bits/ 64 <691200000>; > + opp-peak-kBps = <1804000 9600000>; > + }; > + > + cpu4_opp2: opp-940800000 { > + opp-hz = /bits/ 64 <940800000>; > + opp-peak-kBps = <2188000 17817600>; > + }; > + > + cpu4_opp3: opp-1228800000 { > + opp-hz = /bits/ 64 <1228800000>; > + opp-peak-kBps = <4068000 24576000>; > + }; > + > + cpu4_opp4: opp-1344000000 { > + opp-hz = /bits/ 64 <1344000000>; > + opp-peak-kBps = <4068000 24576000>; > + }; > + > + cpu4_opp5: opp-1516800000 { > + opp-hz = /bits/ 64 <1516800000>; > + opp-peak-kBps = <4068000 24576000>; > + }; > + > + cpu4_opp6: opp-1651200000 { > + opp-hz = /bits/ 64 <1651200000>; > + opp-peak-kBps = <6220000 38092800>; > + }; > + > + cpu4_opp7: opp-1900800000 { > + opp-hz = /bits/ 64 <1900800000>; > + opp-peak-kBps = <6220000 44851200>; > + }; > + > + cpu4_opp8: opp-2054400000 { > + opp-hz = /bits/ 64 <2054400000>; > + opp-peak-kBps = <6220000 44851200>; > + }; > + > + cpu4_opp9: opp-2131200000 { > + opp-hz = /bits/ 64 <2131200000>; > + opp-peak-kBps = <6220000 44851200>; > + }; > + > + cpu4_opp10: opp-2400000000 { > + opp-hz = /bits/ 64 <2400000000>; > + opp-peak-kBps = <6832000 48537600>; > + }; > + }; > + > memory@80000000 { > device_type = "memory"; > /* We expect the bootloader to fill in the size */
Hi Sibi, On Tue, May 04, 2021 at 11:55:10PM +0530, Sibi Sankar wrote: > Hey Sudeep, > > Thanks for the review! > > On 2021-05-04 20:12, Sudeep Holla wrote: [...] > > > > NACK, this breaks if there is a mismatch from what is read from the > > hardware and what is presented in this table above. Either add it from the > > some bootloader or other boot code to this table reading from the > > hardware/firmware or find a way to link them without this. > > > > Sorry I had warned long back about this when such links were discussed > > as part of interconnect binding. > > Not sure why this warrants a NACK, as this was consensus for mapping cpu > freq to DDR/L3 bandwidth votes. (We use the same solution on SDM845 and > SC7180). The opp tables are optional and when specified puts in votes for > DDR/L3. In the future the table can be safely dropped when more useful > devfreq governors are upstreamed. > cpufreq: qcom: Don't add frequencies without an OPP (You can always add commit sha to make it easy to search) But I am not sure how this is related to the above commit anyways. > > I guess your main concern for breakage is ^^ commit? The original design is > to list a super set of frequencies supported by all variants of the SoC > along with the required DDR/L3 bandwidth values. When we run into > non-documented frequency we just wouldn't put in bw votes for it which > should be fine since the entire opp_table is optional. If this is the reason > for the NACK I can try get it reverted with Matthias's ack. No my main concern is this platform uses "qcom-cpufreq-hw" driver and the fact that the OPPs are retrieved from the hardware lookup table invalidates whatever we have in DT. In short it will be junk and becomes obsolete. So what I suggested before is still valid. You simply can't have static OPP tables in the DT for this platform. Do get some boot code to fetch the same from the h/w LUT and patch to the DT or figure out any other way to manage dynamically. So NACK still stands for static addition of OPPs to the DT as in this patch. -- Regards, Sudeep
On 2021-05-05 14:19, Sudeep Holla wrote: > Hi Sibi, > > On Tue, May 04, 2021 at 11:55:10PM +0530, Sibi Sankar wrote: >> Hey Sudeep, >> >> Thanks for the review! >> >> On 2021-05-04 20:12, Sudeep Holla wrote: > > [...] > >> > >> > NACK, this breaks if there is a mismatch from what is read from the >> > hardware and what is presented in this table above. Either add it from the >> > some bootloader or other boot code to this table reading from the >> > hardware/firmware or find a way to link them without this. >> > >> > Sorry I had warned long back about this when such links were discussed >> > as part of interconnect binding. >> >> Not sure why this warrants a NACK, as this was consensus for mapping >> cpu >> freq to DDR/L3 bandwidth votes. (We use the same solution on SDM845 >> and >> SC7180). The opp tables are optional and when specified puts in votes >> for >> DDR/L3. In the future the table can be safely dropped when more useful >> devfreq governors are upstreamed. >> cpufreq: qcom: Don't add frequencies without an OPP > > (You can always add commit sha to make it easy to search) > > But I am not sure how this is related to the above commit anyways. > >> >> I guess your main concern for breakage is ^^ commit? The original >> design is >> to list a super set of frequencies supported by all variants of the >> SoC >> along with the required DDR/L3 bandwidth values. When we run into >> non-documented frequency we just wouldn't put in bw votes for it which >> should be fine since the entire opp_table is optional. If this is the >> reason >> for the NACK I can try get it reverted with Matthias's ack. > > No my main concern is this platform uses "qcom-cpufreq-hw" driver and > the > fact that the OPPs are retrieved from the hardware lookup table > invalidates > whatever we have in DT. In short it will be junk and becomes obsolete. The table provides mapping to bandwidths which aren't available in the firmware though. In short we do have to store the mapping somewhere i.e. a mapping that lists all possible frequencies to its bandwidth requirements needs to be present and using a opp table with the interconnect bw bindings was the consensus reached. Given that a duplicate mapping that lists all possible frequencies to bw is inevitable and Qualcomm has a way of listing all the supported frequencies for the SoC, I feel that dt breakage in the future should be a non-concern. Not sure why you call it junk since it solves the perf/power requirements on SDM845/SC7180 SoCs. When it becomes obsolete it would mean that they are better devfreq governors available upstream and that's a good reason for the opp tables to go away. > So what I suggested before is still valid. You simply can't have static > OPP tables in the DT for this platform. Do get some boot code to fetch > the > same from the h/w LUT and patch to the DT or figure out any other way > to > manage dynamically. moving the logic to boot loader doesn't magically fix your concerns though (since it would also need a superset of available frequencies). It will suffer from the same problems with an additional dependency on firmware propagation in case of breakages which is something you can avoid for the simple cpu based scaling solution. > > So NACK still stands for static addition of OPPs to the DT as in this > patch. I'll let Viresh take the call since this solution is already used on older SoCs. > > -- > Regards, > Sudeep -- Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.
On Wed, May 05, 2021 at 03:39:17PM +0530, Sibi Sankar wrote: > On 2021-05-05 14:19, Sudeep Holla wrote: [...] > > But I am not sure how this is related to the above commit anyways. > > > > > > > > I guess your main concern for breakage is ^^ commit? The original > > > design is > > > to list a super set of frequencies supported by all variants of the > > > SoC > > > along with the required DDR/L3 bandwidth values. When we run into > > > non-documented frequency we just wouldn't put in bw votes for it which > > > should be fine since the entire opp_table is optional. If this is > > > the reason > > > for the NACK I can try get it reverted with Matthias's ack. > > > > No my main concern is this platform uses "qcom-cpufreq-hw" driver and > > the > > fact that the OPPs are retrieved from the hardware lookup table > > invalidates > > whatever we have in DT. In short it will be junk and becomes obsolete. > > The table provides mapping to bandwidths which aren't available in the > firmware though. In short we do have to store the mapping somewhere i.e. a > mapping that lists all possible frequencies to its bandwidth requirements > needs to be present and using a opp table with the interconnect bw bindings > was the consensus reached. I understand and that's exactly what I had pointed out earlier when I mentioned that I had raised this concern previously. > Given that a duplicate mapping that lists all possible frequencies to bw is > inevitable I disagree, it was made inevitable by not listening to all feedback, sorry. > and Qualcomm has a way of listing all the supported frequencies for the SoC, > I feel that dt breakage in the future should be a non-concern. I don't completely understand this TBH. Also my main worry is as we move more towards abstract scale and/or index based, any addition or deletion of OPPs results in change in the index or scale. It may be dealt on absolute scale today everywhere and not a problem *today*, but it will break IMO. > Not sure why you call it junk since it solves the perf/power requirements on > SDM845/SC7180 SoCs. When it becomes obsolete it would mean that they are > better devfreq governors available upstream and that's a good reason for the > opp tables to go away. > Nope, I meant the firmware updates the OPP table underneath for whatever valid reasons it may have. > > So what I suggested before is still valid. You simply can't have static > > OPP tables in the DT for this platform. Do get some boot code to fetch > > the same from the h/w LUT and patch to the DT or figure out any other way > > to manage dynamically. > > moving the logic to boot loader doesn't magically fix your concerns though > (since it would also need a superset of available frequencies). OK that's interesting. I wanted to fetch the exact list from the hardware every time. > It will suffer from the same problems with an additional dependency on > firmware propagation in case of breakages which is something you can avoid > for the simple cpu based scaling solution. IMO the cross dependency is one part of problem and fetching the exact list of OPPs available for the CPU is another. I meant to fix the latter in boot code. The former is something I assume DT bindings deals with. > > > > So NACK still stands for static addition of OPPs to the DT as in this > > patch. > > I'll let Viresh take the call since this solution is already used on older SoCs. Sure, definitely I am just expressing my concern with NACK and I don't have the final say 😁. -- Regards, Sudeep
On Wed, May 05, 2021 at 05:07:24PM +0530, Viresh Kumar wrote: > On 05-05-21, 09:49, Sudeep Holla wrote: > > No my main concern is this platform uses "qcom-cpufreq-hw" driver and the > > fact that the OPPs are retrieved from the hardware lookup table invalidates > > whatever we have in DT. > > Not exactly. > > It disables them all, and then call dev_pm_opp_adjust_voltage() and > enable them again. This is how it started initially. Though the driver > also works if the DT doesn't have the table, in that case it calls > dev_pm_opp_add() for all the OPPs. > Ah OK, if it is handled in the driver, I will shut up then 😄. I did a quick look at it but couldn't understand the connection, so I started and continued the discussion. Thanks for the confirmation. I am fine if it is handled. -- Regards, Sudeep