cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variant
Subsystems (2)
PATCH 1/1 cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variant
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.
The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 ++++++ 1 file changed, 6 insertions(+)
drivers/cpufreq/armada-37xx-cpufreq.c +6 -0
@@ -104,7 +104,13 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { +#if 0 +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +#endif {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } },
Signature · 1 line
2.31.1
Quoted 1 line · On Wed, 30 Jun 2021 15:59:42 +0200 Marek Behún <[email protected]> wrote:
Signed-off-by Marek Behún <[email protected]>
OMG, this should be
Marek
Andrew Lunn
maintainerQuoted 29 lines · Marek Behún
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 ++++++ 1 file changed, 6 insertions(+)drivers/cpufreq/armada-37xx-cpufreq.c +0
-0
drivers/cpufreq/armada-37xx-cpufreq.c struct armada_37xx_dvfs { · line 113
Quoted 10 lines
drivers/cpufreq/armada-37xx-cpufreq.c +6
-0@@ -104,7 +104,13 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { +#if 0 +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +#endif
I suspect you will get some drive by patches from bot handlers removing the #if 0, since such code is not liked within the kernel, and bot handlers blindly do what the bot tells them.
So i would suggest you avoid #if 0, and move the .cpu_req_max entry into the comment.
Andrew
PATCH v2 1/1 cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variant
In reply to Marek BehúnThe 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.
The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.
Ken Ma / Victor Gu, you have worked on commit
https://github.com/MarvellEmbeddedProcessors/linux-marvell/commit/d6719fdc2b3cac58064f41b531f86993c919aa9a
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base CPU frequency and instead adds code that computes the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).
Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/cpufreq/armada-37xx-cpufreq.c +5 -1
@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} }, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } },
Signature · 1 line
2.31.1
Viresh Kumar
maintainerQuoted 28 lines · Marek Behún
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://github.com/MarvellEmbeddedProcessors/linux-marvell/commit/d6719fdc2b3cac58064f41b531f86993c919aa9a
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base CPU frequency and instead adds code that computes the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).
Quoted 6 lines
Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
I am not picking it up for 5.14-rc1 to make sure others get a chance to provide reviews.
Signature · 1 line
viresh
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 55 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://github.com/MarvellEmbeddedProcessors/linux-marvell/commit/d6719fdc2b3cac58064f41b531f86993c919aa9a
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base CPU frequency and instead adds code that computes the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +5
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} }, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } },Signature · 1 line
2.31.1
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 58 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 55 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://github.com/MarvellEmbeddedProcessors/linux-marvell/commit/d6719fdc2b3cac58064f41b531f86993c919aa9a
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base CPU frequency and instead adds code that computes the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +5
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} }, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } },Signature · 1 line
2.31.1
Ping! Gentle reminder for Marvell people.
Quoted 49 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 45 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 42 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://github.com/MarvellEmbeddedProcessors/linux-marvell/commit/d6719fdc2b3cac58064f41b531f86993c919aa9a
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base CPU frequency and instead adds code that computes the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +0
-0@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };
Quoted 13 lines
Quoted 5 lines
Quoted 5 lines
drivers/cpufreq/armada-37xx-cpufreq.c +3
-1static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. Quoted 8 lines
Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +2
-0+*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} }, */ {.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} }, {.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} }, {.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5, 6} },Signature · 1 line
2.31.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 65 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 62 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 58 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 55 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://github.com/MarvellEmbeddedProcessors/linux-marvell/commit/d6719fdc2b3cac58064f41b531f86993c919aa9a
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base CPU frequency and instead adds code that computes the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +5
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} }, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } },Signature · 1 line
2.31.1
Viresh Kumar
maintainerQuoted 3 lines · Pali Rohár
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
We can not hang forever for something that breaks stuff. Applied this for 5.14 now.
Signature · 1 line
viresh
+ Elad and Wojciech from Marvell
Could you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://lore.kernel.org/linux-pm/20210809040224.j2rvopmmqda3utc5@vireshk-i7/
Quoted 70 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 65 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 62 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 58 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 55 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.
We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://github.com/MarvellEmbeddedProcessors/linux-marvell/commit/d6719fdc2b3cac58064f41b531f86993c919aa9a
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base CPU frequency and instead adds code that computes the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +5
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} }, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } },Signature · 1 line
2.31.1
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz) requires sudden changes of VDD supply, and it requires time to stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state.
"
I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variant
External Email
----------------------------------------------------------------------
+ Elad and Wojciech from Marvell
Could you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_linux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=cXiCZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=
Quoted 79 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 73 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 70 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 66 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 63 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with
DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is
used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719fdc2b3 cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtf Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKyKOOM3X dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b9cDKem t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base
CPU frequency and instead adds code that computes the voltages
from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-2@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} +}, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } }, --2.31.1
Hello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=61c40f35f5cd6f67ccbd7319a1722eb78c815989
There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 113 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz) requires sudden changes of VDD supply, and it requires time to stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state.
"I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantExternal Email
----------------------------------------------------------------------
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_linux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=cXiCZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 79 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 73 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 70 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 66 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 63 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with
DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling is
used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719fdc2b3 cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtf Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKyKOOM3X dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b9cDKem t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base
CPU frequency and instead adds code that computes the voltages
from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-2@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, 6} +}, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 } }, --2.31.1
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variant
Hello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5cd6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJIPPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iym5bjoM6l5zLrbh_GVs&e=
There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 131 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantExternal Email
----------------------------------------------------------------------
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ&c=n KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m= 5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with
DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling
is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKyKOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b9cDK
em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base
CPU frequency and instead adds code that computes the
voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, +6} }, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 }},
--
2.31.1
Hello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 139 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5cd6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJIPPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iym5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 116 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantExternal Email
----------------------------------------------------------------------
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ&c=n KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m= 5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 69 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 63 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 60 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 56 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 53 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with
DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling
is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKyKOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b9cDK
em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base
CPU frequency and instead adds code that computes the
voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +0
-0@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };
Quoted 15 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
drivers/cpufreq/armada-37xx-cpufreq.c +5
-1static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +1
-0+6} }, */ {.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} }, {.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} }, {.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5, 6}},
--
2.31.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
A message that is not in the archive (sent off-list or not yet received): BN9PR18MB42518C761E574D862D30CDA7DB9A9@BN9PR18MB4251.namprd18.prod.outlook.com
Quoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random
and they are always different, like a real voltage issue:
https://gist.github.com/robimarko/113216f566ccf159dfd33933889da042
https://gist.github.com/robimarko/990d757870d44a3c5acdfeb957547705
Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://gist.github.com/robimarko/1a81b0c6e93735b75ff4461d405c8033
I am still digging to print the voltage changes as _set_opp_voltage is not being used.
Regards,
Robert
Quoted 169 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]> נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 154 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5cd6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJIPPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iym5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 131 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantExternal Email
----------------------------------------------------------------------
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ&c=n KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m= 5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable with
DVFS: when the SOC boots, the WTMI firmware sets clocks and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when scaling
is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKyKOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b9cDK
em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz base
CPU frequency and instead adds code that computes the
voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some 1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?
--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 ,6 } }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4, +6} }, */ {.cpu_freq_max =1000 *1000 *1000 , .divider = {1 ,2 ,4 ,5 } }, {.cpu_freq_max =800 *1000 *1000 , .divider = {1 ,2 ,3 ,4 } }, {.cpu_freq_max =600 *1000 *1000 , .divider = {2 ,4 ,5 ,6 }},
--
2.31.1
Signature · 7 lines
Robert Marko Staff Embedded Linux Engineer Sartura Ltd. Lendavska ulica 16a 10000 Zagreb, Croatia Email: [email protected] Web: www.sartura.hr
Hi,
Unless the logs are misleading, then I see here:
cpu cpu0: _set_opp: switching OPP: Freq 200000000 -> 1200000000 Hz, Level 0 -> 0, Bw 0 -> 0
Which violates the errata.
If there is an interim step in between, I think it should be printed out in the debug so we can clearly understand what is the interim frequency setting between 200 and 1200 MHz.
Elad.
-----Original Message-----
From: Robert Marko <[email protected]>
Sent: Tuesday, August 2, 2022 7:42 PM
To: Elad Nachman <[email protected]>
Cc: Pali Rohár <[email protected]>; Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>; Gérald Kerma <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variant
Quoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random and they are always different, like a real voltage issue:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_113216f566ccf159dfd33933889da042&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=jvmR3Myk443DelvNZv1OkhmpqnMp9Y8mvzzYz2g13rM&e=
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_990d757870d44a3c5acdfeb957547705&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=XrMFeJpEGO5A4rIKjkHLNc4MHzPGOBKeOktDWCbQMAc&e=
Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_1a81b0c6e93735b75ff4461d405c8033&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=02ljqhQAdZki-JwDYNPKaStmzSkhuitBRP6R17iOZqA&e=
I am still digging to print the voltage changes as _set_opp_voltage is not being used.
Regards,
Robert
Quoted 198 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]>
נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 172 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_
pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5c d6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eT eNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJI PPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iy m5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 136 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for
1.2 GHz variantExternal Email
------------------------------------------------------------------
----
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ
&c=n
KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrw
Q&m=
5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s
=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable
with
DVFS: when the SOC boots, the WTMI firmware sets clocks
and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when
scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support forArmada
37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719
fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyP
az7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKy
KOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b
9cDK em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz
base CPU frequency and instead adds code that computes
the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some
1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 , 6} }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC +is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2,
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +1
-0+4, 6} }, */ {.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} }, {.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} }, {.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5,6} },
--
2.31.1
--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: https://urldefense.proofpoint.com/v2/url?u=http-3A__www.sartura.hr&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=_aBokTETNVzTrHqewupr4PeLusBNf7LGrTmjI2hppFk&e=
Quoted 9 lines · Elad Nachman
Hi,
Unless the logs are misleading, then I see here:
cpu cpu0:
_set_opp: switching OPP: Freq 200000000 -> 1200000000 Hz, Level 0 -> 0, Bw 0 -> 0Which violates the errata.
If there is an interim step in between, I think it should be printed out in the debug so we can clearly understand what is the interim frequency setting between 200 and 1200 MHz.
This is printed directly by the _set_opp from the cpufreq core, so it should be accurate.
Pali, am I doing this correctly or I need to print from the A3K cpufreq or clk drivers?
Regards,
Robert
Quoted 254 lines
Elad.
-----Original Message-----
From: Robert Marko <[email protected]>
Sent: Tuesday, August 2, 2022 7:42 PM
To: Elad Nachman <[email protected]>
Cc: Pali Rohár <[email protected]>; Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>; Gérald Kerma <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantQuoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random and they are always different, like a real voltage issue:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_113216f566ccf159dfd33933889da042&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=jvmR3Myk443DelvNZv1OkhmpqnMp9Y8mvzzYz2g13rM&e=
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_990d757870d44a3c5acdfeb957547705&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=XrMFeJpEGO5A4rIKjkHLNc4MHzPGOBKeOktDWCbQMAc&e=Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_1a81b0c6e93735b75ff4461d405c8033&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=02ljqhQAdZki-JwDYNPKaStmzSkhuitBRP6R17iOZqA&e=I am still digging to print the voltage changes as
_set_opp_voltageis not being used.Regards,
RobertQuoted 198 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]>
נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 172 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_
pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5c d6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eT eNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJI PPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iy m5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 136 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for
1.2 GHz variantExternal Email
------------------------------------------------------------------
----
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ
&c=n
KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrw
Q&m=
5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s
=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable
with
DVFS: when the SOC boots, the WTMI firmware sets clocks
and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when
scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support forArmada
37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719
fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyP
az7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKy
KOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b
9cDK em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz
base CPU frequency and instead adds code that computes
the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some
1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 , 6} }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC +is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +1
-0+4, 6} }, */ {.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} }, {.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} }, {.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5,6} },
--
2.31.1--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: https://urldefense.proofpoint.com/v2/url?u=http-3A__www.sartura.hr&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=_aBokTETNVzTrHqewupr4PeLusBNf7LGrTmjI2hppFk&e=
Signature · 7 lines
Robert Marko Staff Embedded Linux Engineer Sartura Ltd. Lendavska ulica 16a 10000 Zagreb, Croatia Email: [email protected] Web: www.sartura.hr
Quoted 8 lines · On Tuesday 02 August 2022 18:56:07 Robert Marko wrote:
Quoted 7 lines · Elad Nachman
Hi,
Unless the logs are misleading, then I see here:
cpu cpu0:
_set_opp: switching OPP: Freq 200000000 -> 1200000000 Hz, Level 0 -> 0, Bw 0 -> 0
Quoted 7 lines
Quoted 2 lines
Which violates the errata.
If there is an interim step in between, I think it should be printed out in the debug so we can clearly understand what is the interim frequency setting between 200 and 1200 MHz.This is printed directly by the
_set_oppfrom the cpufreq core, so it should be accurate.
Pali, am I doing this correctly or I need to print from the A3K cpufreq or clk drivers?
Hello! You need to print it from a3k clk driver. cpufreq core just ask driver to switch speed from 200000000 to 1200000000 and clk driver then change it with its own workaround function.
The real change of Level is done at these places:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n548
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n592
Check places where is done write operation to register ARMADA_37XX_NB_CPU_LOAD.
Quoted 267 lines
Regards,
RobertQuoted 254 lines
Elad.
-----Original Message-----
From: Robert Marko <[email protected]>
Sent: Tuesday, August 2, 2022 7:42 PM
To: Elad Nachman <[email protected]>
Cc: Pali Rohár <[email protected]>; Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>; Gérald Kerma <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantQuoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random and they are always different, like a real voltage issue:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_113216f566ccf159dfd33933889da042&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=jvmR3Myk443DelvNZv1OkhmpqnMp9Y8mvzzYz2g13rM&e=
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_990d757870d44a3c5acdfeb957547705&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=XrMFeJpEGO5A4rIKjkHLNc4MHzPGOBKeOktDWCbQMAc&e=Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_1a81b0c6e93735b75ff4461d405c8033&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=02ljqhQAdZki-JwDYNPKaStmzSkhuitBRP6R17iOZqA&e=I am still digging to print the voltage changes as
_set_opp_voltageis not being used.Regards,
RobertQuoted 198 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]>
נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 172 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_
pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5c d6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eT eNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJI PPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iy m5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 136 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for
1.2 GHz variantExternal Email
------------------------------------------------------------------
----
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ
&c=n
KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrw
Q&m=
5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s
=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable
with
DVFS: when the SOC boots, the WTMI firmware sets clocks
and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when
scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support forArmada
37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719
fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyP
az7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKy
KOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b
9cDK em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz
base CPU frequency and instead adds code that computes
the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some
1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 , 6} }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC +is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +1
-0+4, 6} }, */ {.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} }, {.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} }, {.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5,6} },
--
2.31.1--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: https://urldefense.proofpoint.com/v2/url?u=http-3A__www.sartura.hr&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=_aBokTETNVzTrHqewupr4PeLusBNf7LGrTmjI2hppFk&e=Signature · 7 lines
Robert Marko Staff Embedded Linux Engineer Sartura Ltd. Lendavska ulica 16a 10000 Zagreb, Croatia Email: [email protected] Web: www.sartura.hr
Quoted 22 lines · Pali Rohár
Quoted 8 lines · On Tuesday 02 August 2022 18:56:07 Robert Marko wrote:
Quoted 7 lines · Elad Nachman
Hi,
Unless the logs are misleading, then I see here:
cpu cpu0:
_set_opp: switching OPP: Freq 200000000 -> 1200000000 Hz, Level 0 -> 0, Bw 0 -> 0Quoted 7 lines
Quoted 2 lines
Which violates the errata.
If there is an interim step in between, I think it should be printed out in the debug so we can clearly understand what is the interim frequency setting between 200 and 1200 MHz.This is printed directly by the
_set_oppfrom the cpufreq core, so it should be accurate.
Pali, am I doing this correctly or I need to print from the A3K cpufreq or clk drivers?Hello! You need to print it from a3k clk driver. cpufreq core just ask driver to switch speed from 200000000 to 1200000000 and clk driver then change it with its own workaround function.
Quoted 6 lines
The real change of Level is done at these places:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n548
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n592Check places where is done write operation to register
ARMADA_37XX_NB_CPU_LOAD.
Ok, finally got time to try it.
I am now printing from the clk driver instead, hopefully in the right places:
https://gist.github.com/robimarko/d297c81f70ef9620c830435bad8a6a8d
Trying to enlarge the wait to 100ms does not help.
Regards,
Robert
Quoted 268 lines
Quoted 268 lines
Regards,
RobertQuoted 246 lines
Elad.
-----Original Message-----
From: Robert Marko <[email protected]>
Sent: Tuesday, August 2, 2022 7:42 PM
To: Elad Nachman <[email protected]>
Cc: Pali Rohár <[email protected]>; Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>; Gérald Kerma <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantQuoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random and they are always different, like a real voltage issue:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_113216f566ccf159dfd33933889da042&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=jvmR3Myk443DelvNZv1OkhmpqnMp9Y8mvzzYz2g13rM&e=
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_990d757870d44a3c5acdfeb957547705&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=XrMFeJpEGO5A4rIKjkHLNc4MHzPGOBKeOktDWCbQMAc&e=Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_1a81b0c6e93735b75ff4461d405c8033&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=02ljqhQAdZki-JwDYNPKaStmzSkhuitBRP6R17iOZqA&e=I am still digging to print the voltage changes as
_set_opp_voltageis not being used.Regards,
RobertQuoted 198 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]>
נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 172 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_
pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5c d6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eT eNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJI PPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iy m5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 136 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for
1.2 GHz variantExternal Email
------------------------------------------------------------------
----
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ
&c=n
KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrw
Q&m=
5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s
=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable
with
DVFS: when the SOC boots, the WTMI firmware sets clocks
and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when
scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support forArmada
37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719
fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyP
az7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKy
KOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b
9cDK em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz
base CPU frequency and instead adds code that computes
the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some
1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 , 6} }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC +is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +1
-0+4, 6} }, */ {.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} }, {.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} }, {.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5,6} },
--
2.31.1Quoted 8 lines
--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: https://urldefense.proofpoint.com/v2/url?u=http-3A__www.sartura.hr&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=_aBokTETNVzTrHqewupr4PeLusBNf7LGrTmjI2hppFk&e=--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: www.sartura.hr
Signature · 7 lines
Robert Marko Staff Embedded Linux Engineer Sartura Ltd. Lendavska ulica 16a 10000 Zagreb, Croatia Email: [email protected] Web: www.sartura.hr
Quoted 27 lines · On Wednesday 17 August 2022 11:40:32 Robert Marko wrote:
Quoted 26 lines · Pali Rohár
Quoted 8 lines · On Tuesday 02 August 2022 18:56:07 Robert Marko wrote:
Quoted 7 lines · Elad Nachman
Hi,
Unless the logs are misleading, then I see here:
cpu cpu0:
_set_opp: switching OPP: Freq 200000000 -> 1200000000 Hz, Level 0 -> 0, Bw 0 -> 0Quoted 7 lines
Quoted 2 lines
Which violates the errata.
If there is an interim step in between, I think it should be printed out in the debug so we can clearly understand what is the interim frequency setting between 200 and 1200 MHz.This is printed directly by the
_set_oppfrom the cpufreq core, so it should be accurate.
Pali, am I doing this correctly or I need to print from the A3K cpufreq or clk drivers?Hello! You need to print it from a3k clk driver. cpufreq core just ask driver to switch speed from 200000000 to 1200000000 and clk driver then change it with its own workaround function.
The real change of Level is done at these places:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n548
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n592
Quoted 8 lines
Quoted 2 lines
Check places where is done write operation to register
ARMADA_37XX_NB_CPU_LOAD.Ok, finally got time to try it.
I am now printing from the clk driver instead, hopefully in the right places:
https://gist.github.com/robimarko/d297c81f70ef9620c830435bad8a6a8dTrying to enlarge the wait to 100ms does not help.
Could you provide also diff which you applied to driver?
Quoted 281 lines
Regards,
RobertQuoted 268 lines
Quoted 268 lines
Regards,
RobertQuoted 246 lines
Elad.
-----Original Message-----
From: Robert Marko <[email protected]>
Sent: Tuesday, August 2, 2022 7:42 PM
To: Elad Nachman <[email protected]>
Cc: Pali Rohár <[email protected]>; Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>; Gérald Kerma <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantQuoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random and they are always different, like a real voltage issue:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_113216f566ccf159dfd33933889da042&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=jvmR3Myk443DelvNZv1OkhmpqnMp9Y8mvzzYz2g13rM&e=
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_990d757870d44a3c5acdfeb957547705&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=XrMFeJpEGO5A4rIKjkHLNc4MHzPGOBKeOktDWCbQMAc&e=Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_1a81b0c6e93735b75ff4461d405c8033&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=02ljqhQAdZki-JwDYNPKaStmzSkhuitBRP6R17iOZqA&e=I am still digging to print the voltage changes as
_set_opp_voltageis not being used.Regards,
RobertQuoted 198 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]>
נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 172 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_
pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5c d6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eT eNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJI PPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iy m5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 136 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for
1.2 GHz variantExternal Email
------------------------------------------------------------------
----
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ
&c=n
KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrw
Q&m=
5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s
=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable
with
DVFS: when the SOC boots, the WTMI firmware sets clocks
and AVS values that work correctly with 1.2 GHz CPU frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided by the vendor in the OTP is simply incorrect when
scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid the cpufreq driver on 1.2 GHz variant.
Signed-off-by Marek Behún <[email protected]>Fixes 92ce45fb875d ("cpufreq: Add DVFS support forArmada
37xx")
---
If someone from Marvell could look into this, it would be great since basically 1.2 GHz variant cannot scale, which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719
fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyP
az7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKy
KOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b
9cDK em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz
base CPU frequency and instead adds code that computes
the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for some
1.2 GHz boards? Do we need to force the L0 voltage to 1202 mV if it is lower, or something?--- drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)drivers/cpufreq/armada-37xx-cpufreq.c +6
-1@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = { - {.cpu_freq_max =1200 *1000 *1000 , .divider = {1 ,2 ,4 , 6} }, +/* +* The cpufreq scaling for 1.2 GHz variant of the SOC +is currently +* unstable because we do not know how to configure it properly. +*/ +/* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +1
-0+4, 6} }, */ {.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} }, {.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} }, {.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5,6} },
--
2.31.1Quoted 8 lines
--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: https://urldefense.proofpoint.com/v2/url?u=http-3A__www.sartura.hr&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=_aBokTETNVzTrHqewupr4PeLusBNf7LGrTmjI2hppFk&e=--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: www.sartura.hrSignature · 7 lines
Robert Marko Staff Embedded Linux Engineer Sartura Ltd. Lendavska ulica 16a 10000 Zagreb, Croatia Email: [email protected] Web: www.sartura.hr
Quoted 32 lines · Pali Rohár
Quoted 30 lines · On Wednesday 17 August 2022 11:40:32 Robert Marko wrote:
Quoted 28 lines · Pali Rohár
Quoted 8 lines · On Tuesday 02 August 2022 18:56:07 Robert Marko wrote:
Quoted 7 lines · Elad Nachman
Hi,
Unless the logs are misleading, then I see here:
cpu cpu0:
_set_opp: switching OPP: Freq 200000000 -> 1200000000 Hz, Level 0 -> 0, Bw 0 -> 0Quoted 7 lines
Quoted 2 lines
Which violates the errata.
If there is an interim step in between, I think it should be printed out in the debug so we can clearly understand what is the interim frequency setting between 200 and 1200 MHz.This is printed directly by the
_set_oppfrom the cpufreq core, so it should be accurate.
Pali, am I doing this correctly or I need to print from the A3K cpufreq or clk drivers?Hello! You need to print it from a3k clk driver. cpufreq core just ask driver to switch speed from 200000000 to 1200000000 and clk driver then change it with its own workaround function.
The real change of Level is done at these places:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n548
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n592Check places where is done write operation to register
ARMADA_37XX_NB_CPU_LOAD.
Quoted 7 lines
Quoted 5 lines
Ok, finally got time to try it.
I am now printing from the clk driver instead, hopefully in the right places:
https://gist.github.com/robimarko/d297c81f70ef9620c830435bad8a6a8dTrying to enlarge the wait to 100ms does not help.
Could you provide also diff which you applied to driver?
Sure, here it is:
https://gist.github.com/robimarko/a2b8942b5f22b107c62fba9695220881
Regards,
Robert
Quoted 282 lines
Quoted 282 lines
Regards,
RobertQuoted 260 lines
Quoted 260 lines
Regards,
RobertQuoted 254 lines
Elad.
-----Original Message-----
From: Robert Marko <[email protected]>
Sent: Tuesday, August 2, 2022 7:42 PM
To: Elad Nachman <[email protected]>
Cc: Pali Rohár <[email protected]>; Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>; Gérald Kerma <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantQuoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random and they are always different, like a real voltage issue:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_113216f566ccf159dfd33933889da042&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=jvmR3Myk443DelvNZv1OkhmpqnMp9Y8mvzzYz2g13rM&e=
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_990d757870d44a3c5acdfeb957547705&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=XrMFeJpEGO5A4rIKjkHLNc4MHzPGOBKeOktDWCbQMAc&e=Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_1a81b0c6e93735b75ff4461d405c8033&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=02ljqhQAdZki-JwDYNPKaStmzSkhuitBRP6R17iOZqA&e=I am still digging to print the voltage changes as
_set_opp_voltageis not being used.Regards,
RobertQuoted 198 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]>
נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 172 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_
pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5c d6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eT eNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJI PPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iy m5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 136 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for
1.2 GHz variantExternal Email
------------------------------------------------------------------
----
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ
&c=n
KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrw
Q&m=
5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s
=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
Quoted 68 lines · On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable
with
DVFS: when the SOC boots, the WTMI firmware sets clocks
and AVS values that work correctly with 1.2 GHz CPU
frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided
by the vendor in the OTP is simply incorrect when
scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid
the cpufreq driver on 1.2 GHz variant.Signed-off-by: Marek Behún <[email protected]>
Fixes: 92ce45fb875d ("cpufreq: Add DVFS support for
Armada
37xx")
---
If someone from Marvell could look into this, it would
be great since basically 1.2 GHz variant cannot scale,
which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719
fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyP
az7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKy
KOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b
9cDK em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz
base CPU frequency and instead adds code that computes
the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for
some
1.2 GHz boards? Do we need to force the L0 voltage to
1202 mV if it is lower, or something?
---
drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)diff --git a/drivers/cpufreq/armada-37xx-cpufreq.c
b/drivers/cpufreq/armada-37xx-cpufreq.c
index 3fc98a3ffd91..c10fc33b29b1 100644
--- a/drivers/cpufreq/armada-37xx-cpufreq.c
+++ b/drivers/cpufreq/armada-37xx-cpufreq.c
@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = {
- {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4,
6} },
+ /*
+ * The cpufreq scaling for 1.2 GHz variant of the SOC
+is currently
+ * unstable because we do not know how to configure it properly.
+ */
+ /* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2,Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
+4, 6} }, */
{.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} },
{.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} },
{.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5,
6} },
--
2.31.1--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: https://urldefense.proofpoint.com/v2/url?u=http-3A__www.sartura.hr&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=_aBokTETNVzTrHqewupr4PeLusBNf7LGrTmjI2hppFk&e=Quoted 8 lines
Quoted 8 lines
--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: www.sartura.hr--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: www.sartura.hr
Signature · 7 lines
Robert Marko Staff Embedded Linux Engineer Sartura Ltd. Lendavska ulica 16a 10000 Zagreb, Croatia Email: [email protected] Web: www.sartura.hr
Quoted 37 lines · On Thursday 18 August 2022 10:14:26 Robert Marko wrote:
Quoted 36 lines · Pali Rohár
Quoted 34 lines · On Wednesday 17 August 2022 11:40:32 Robert Marko wrote:
Quoted 22 lines · Pali Rohár
Quoted 8 lines · On Tuesday 02 August 2022 18:56:07 Robert Marko wrote:
Quoted 7 lines · Elad Nachman
Hi,
Unless the logs are misleading, then I see here:
cpu cpu0:
_set_opp: switching OPP: Freq 200000000 -> 1200000000 Hz, Level 0 -> 0, Bw 0 -> 0Quoted 7 lines
Quoted 2 lines
Which violates the errata.
If there is an interim step in between, I think it should be printed out in the debug so we can clearly understand what is the interim frequency setting between 200 and 1200 MHz.This is printed directly by the
_set_oppfrom the cpufreq core, so it should be accurate.
Pali, am I doing this correctly or I need to print from the A3K cpufreq or clk drivers?Hello! You need to print it from a3k clk driver. cpufreq core just ask driver to switch speed from 200000000 to 1200000000 and clk driver then change it with its own workaround function.
Quoted 6 lines
The real change of Level is done at these places:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n548
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/clk/mvebu/armada-37xx-periph.c?h=v5.19#n592Check places where is done write operation to register
ARMADA_37XX_NB_CPU_LOAD.Ok, finally got time to try it.
I am now printing from the clk driver instead, hopefully in the right places:
https://gist.github.com/robimarko/d297c81f70ef9620c830435bad8a6a8d
Quoted 6 lines
Quoted 3 lines
Quoted 1 line
Trying to enlarge the wait to 100ms does not help.
Could you provide also diff which you applied to driver?
Sure, here it is:
https://gist.github.com/robimarko/a2b8942b5f22b107c62fba9695220881
You should print debug logs when _calling_ regmap_update_bits(). And current log is very strange, you are printing message line "%pCn requested rate %lu load_level %u" for iteration in for-loop, also when "val != div". So log contains lot of incorrect lines.
Quoted 295 lines
Regards,
RobertQuoted 282 lines
Quoted 282 lines
Regards,
RobertQuoted 260 lines
Quoted 260 lines
Regards,
RobertQuoted 254 lines
Elad.
-----Original Message-----
From: Robert Marko <[email protected]>
Sent: Tuesday, August 2, 2022 7:42 PM
To: Elad Nachman <[email protected]>
Cc: Pali Rohár <[email protected]>; Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>; Gérald Kerma <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantQuoted 5 lines · Elad Nachman
Hi Pali,
Could you please provide the crash dump / call trace?
Quoted 7 lines
Also, if you can please annotate with printk the exact voltage/frequency changes taken by the driver, up to the point of the crash?
This will help understand the sequence of events leading to the crash.
Thanks,
Elad.
Hi Elad,
Here are 2 bootlogs, but I dont think they are of any use as the traces are rather random and they are always different, like a real voltage issue:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_113216f566ccf159dfd33933889da042&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=jvmR3Myk443DelvNZv1OkhmpqnMp9Y8mvzzYz2g13rM&e=
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_990d757870d44a3c5acdfeb957547705&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=XrMFeJpEGO5A4rIKjkHLNc4MHzPGOBKeOktDWCbQMAc&e=Here is a bootleg with the frequency changes, OPP points that are set by the CPUFreq driver are also here:
https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_robimarko_1a81b0c6e93735b75ff4461d405c8033&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=02ljqhQAdZki-JwDYNPKaStmzSkhuitBRP6R17iOZqA&e=I am still digging to print the voltage changes as
_set_opp_voltageis not being used.Regards,
RobertQuoted 198 lines
________________________________
מאת: Pali Rohár <[email protected]>
נשלח: יום שני 01 אוגוסט 2022 20:56
אל: Elad Nachman <[email protected]>
עותק: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected] <[email protected]>; Sebastian Hesselbarth <[email protected]>; [email protected] <[email protected]>; nnet <[email protected]>; Gérald Kerma <[email protected]>
נושא: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad!
Robert (in CC) tested this proposed change. But increasing delay to 100ms does not help. CPU still crashes early during boot.
Quoted 172 lines · On Monday 01 August 2022 14:15:27 Elad Nachman wrote:
Hi,
As first step, please try to increase the delay to 100ms, see if it helps.
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 5:13 PM
To: Elad Nachman <[email protected]>
Cc: Wojciech Bartczak <[email protected]>; Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: Re: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for 1.2 GHz variantHello Elad and thank you for response!
This errata is already implemented in the kernel for a longer time by Gregory's commit:
https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_
pub_scm_linux_kernel_git_stable_linux.git_commit_-3Fid-3D61c40f35f5c d6f67ccbd7319a1722eb78c815989&d=DwIDaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eT eNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=-E-AwB9STVx8xgapaCNSpDJI PPnkrzrWkZX0uFz2bfNGFnckZelT_XaovUUPrNIg&s=4EUcdDWB_gqnEV8nREQi9E_iy m5bjoM6l5zLrbh_GVs&e=There is also 20ms delay after L2/L3 to L1 state switch.
Any idea what could be wrong here? Or is something more than above commit needed to correctly implement that errata?
Quoted 136 lines · On Monday 01 August 2022 14:01:07 Elad Nachman wrote:
Hi Pali,
There is an errata for that.
"
Switching from L2/L3 state (200/300 MHz) to L0 state (1200 MHz)
requires sudden changes of VDD supply, and it requires time to
stabilize the VDD supply. The solution is to use gradual switching from L2/L3 to L1 and then L1 to L0 state. "I would also add additional delay for the VDD supply stabilization.
FYI,
Elad.
-----Original Message-----
From: Pali Rohár <[email protected]>
Sent: Monday, August 1, 2022 3:36 PM
To: Elad Nachman <[email protected]>; Wojciech Bartczak <[email protected]>
Cc: Marek Behún <[email protected]>; Viresh Kumar <[email protected]>; Gregory CLEMENT <[email protected]>; Robert Marko <[email protected]>; Tomasz Maciej Nowak <[email protected]>; Anders Trier Olesen <[email protected]>; Philip Soares <[email protected]>; [email protected]; Sebastian Hesselbarth <[email protected]>; [email protected]; nnet <[email protected]>
Subject: [EXT] Re: [PATCH v2] cpufreq: armada-37xx: forbid cpufreq for
1.2 GHz variantExternal Email
------------------------------------------------------------------
----
+ Elad and Wojciech from MarvellCould you please look at this issue and/or forward it to relevant Marvell team?
Maintainer Viresh already wrote that we cannot hang forever for Marvell and patch which disables support for 1.2 GHz was merged:
https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_l
inux-2Dpm_20210809040224.j2rvopmmqda3utc5-40vireshk-2Di7_&d=DwIDaQ
&c=n
KjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrw
Q&m=
5nMMKyKOOM3XdMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s
=cXi CZByknfz1rOIgJl4fJHl1KLLRq2shHul2-VPpYP0&e=Quoted 84 lines · On Sunday 08 August 2021 21:30:26 Pali Rohár wrote:
Gentle reminder. This is really serious issue. Could you please look at it?
Adding more MarvellEmbeddedProcessors people to the loop: Evan, Benjamin an Igal
Quoted 78 lines · On Thursday 15 July 2021 21:33:21 Pali Rohár wrote:
Ping! Gentle reminder for Marvell people.
Quoted 75 lines · On Thursday 08 July 2021 16:34:51 Pali Rohár wrote:
Konstantin, Nadav, Ken, Victor, Jason: This issue is pretty serious, CPU on 1.2GHz A3720 is crashing. Could you please look at it?
Quoted 71 lines · On Friday 02 July 2021 18:30:35 Pali Rohár wrote:
+Jason from GlobalScale as this issue affects GlobalScale Espressobin Ultra and V7 1.2 GHz boards.
On Thursday 01 July 2021 00:56:01 Marek Behún wrote:
The 1.2 GHz variant of the Armada 3720 SOC is unstable
with
DVFS: when the SOC boots, the WTMI firmware sets clocks
and AVS values that work correctly with 1.2 GHz CPU
frequency, but random crashes occur once cpufreq driver starts scaling.We do not know currently what is the reason:
- it may be that the voltage value for L0 for 1.2 GHz variant provided
by the vendor in the OTP is simply incorrect when
scaling is used,
- it may be that some delay is needed somewhere,
- it may be something else.The most sane solution now seems to be to simply forbid
the cpufreq driver on 1.2 GHz variant.Signed-off-by: Marek Behún <[email protected]>
Fixes: 92ce45fb875d ("cpufreq: Add DVFS support for
Armada
37xx")
---
If someone from Marvell could look into this, it would
be great since basically 1.2 GHz variant cannot scale,
which is a feature that was claimed to be supported by the SOC.Ken Ma / Victor Gu, you have worked on commit
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.
co
m_MarvellEmbeddedProcessors_linux-2Dmarvell_commit_d6719
fdc2
b3
cac58064f41b531f86993c919aa9a&d=DwIDaQ&c=nKjWec2b6R0mOyP
az7x
tf
Q&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=5nMMKy
KOOM
3X
dMe_PerZRx8L7-D7MkWhCl7GxpXTPiotVf1TR4j8v3bpjQmRKCLC&s=b
9cDK em t70OiTJF6KXj0ySzbxpsB_nuteXJE87via80&e=
in linux-marvell.
Your patch takes away the 1202 mV constant for 1.2 GHz
base CPU frequency and instead adds code that computes
the voltages from the voltage found in L0 AVS register (which is filled in by WTMI firmware).Do you know why the code does not work correctly for
some
1.2 GHz boards? Do we need to force the L0 voltage to
1202 mV if it is lower, or something?
---
drivers/cpufreq/armada-37xx-cpufreq.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)diff --git a/drivers/cpufreq/armada-37xx-cpufreq.c
b/drivers/cpufreq/armada-37xx-cpufreq.c
index 3fc98a3ffd91..c10fc33b29b1 100644
--- a/drivers/cpufreq/armada-37xx-cpufreq.c
+++ b/drivers/cpufreq/armada-37xx-cpufreq.c
@@ -104,7 +104,11 @@ struct armada_37xx_dvfs { };static struct armada_37xx_dvfs armada_37xx_dvfs[] = {
- {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2, 4,
6} },
+ /*
+ * The cpufreq scaling for 1.2 GHz variant of the SOC
+is currently
+ * unstable because we do not know how to configure it properly.
+ */
+ /* {.cpu_freq_max = 1200*1000*1000, .divider = {1, 2,Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
+4, 6} }, */
{.cpu_freq_max = 1000*1000*1000, .divider = {1, 2, 4, 5} },
{.cpu_freq_max = 800*1000*1000, .divider = {1, 2, 3, 4} },
{.cpu_freq_max = 600*1000*1000, .divider = {2, 4, 5,
6} },
--
2.31.1--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: https://urldefense.proofpoint.com/v2/url?u=http-3A__www.sartura.hr&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=eTeNTLEK5-TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=u39n7XPBdQVaoaviM32QcFaiO0KDs3BVzkeF-4zrqPKElNH3igH9KqEKfxSKLz-H&s=_aBokTETNVzTrHqewupr4PeLusBNf7LGrTmjI2hppFk&e=Quoted 8 lines
Quoted 8 lines
--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: www.sartura.hr--
Robert Marko
Staff Embedded Linux Engineer
Sartura Ltd.
Lendavska ulica 16a
10000 Zagreb, Croatia
Email: [email protected]
Web: www.sartura.hrSignature · 7 lines
Robert Marko Staff Embedded Linux Engineer Sartura Ltd. Lendavska ulica 16a 10000 Zagreb, Croatia Email: [email protected] Web: www.sartura.hr