| Message ID | 1626429658-18961-3-git-send-email-rnayak@codeaurora.org |
|---|---|
| State | Superseded |
| Headers | show |
| Series | None | expand |
On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: > Quoting Bjorn Andersson (2021-07-16 13:18:56) > > On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: > > > > > qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) > > > Though qup-i2c does not support DVFS, it still needs to vote for a > > > performance state on 'CX' to satisfy the 19.2 Mhz clock frequency > > > requirement. > > > > > > > Sounds good, but... > > > > > Use 'required-opps' to pass this information from > > > device tree, and also add the power-domains property to specify > > > the CX power-domain. > > > > > > > ..is the required-opps really needed with my rpmhpd patch in place? > > > > Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for > CX. On e.g. sm8250 the first available non-zero corner presented in cmd-db is low_svs. And if this (which?) clock requires a higher corner than the lowest possible in order to tick at this "lowest" frequency, I'm certainly interested in some more details. Regards, Bjorn
On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: > Quoting Bjorn Andersson (2021-07-16 13:52:12) > > On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: > > > > > Quoting Bjorn Andersson (2021-07-16 13:18:56) > > > > On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) > > > > > Though qup-i2c does not support DVFS, it still needs to vote for a > > > > > performance state on 'CX' to satisfy the 19.2 Mhz clock frequency > > > > > requirement. > > > > > > > > > > > > > Sounds good, but... > > > > > > > > > Use 'required-opps' to pass this information from > > > > > device tree, and also add the power-domains property to specify > > > > > the CX power-domain. > > > > > > > > > > > > > ..is the required-opps really needed with my rpmhpd patch in place? > > > > > > > > > > Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for > > > CX. > > > > On e.g. sm8250 the first available non-zero corner presented in cmd-db > > is low_svs. > > Indeed. On sc7180 it's not the first non-zero corner. I suppose > retention for CX isn't actually used when the SoC is awake so your > rpmhpd patch is putting in a vote for something that doesn't do anything > at runtime for CX? I imagine that rpmh only sets the aggregate corner to > retention when the whole SoC is suspended/sleeping, otherwise things > wouldn't go very well. Similarly, min_svs may be VDD minimization? If > so, those first two states are basically states that shouldn't be used > at runtime, almost like sleep states. > But if that's the case, I don't think it's appropriate for the "enabled state" of the domain to use any of those corners. As this means that anyone who needs any of the rpmhpd domains active also needs to specify required-opps, which wouldn't be needed for any other power domain provider. And more importantly it means that a device sitting in a GDSC, which would be parented by a rpmhpd domain has no way to specify the GDSC and trickle the minimum-vote up to the rpmhpd domain. (And I know that we don't describe the parentship of the GDSCs today, but this patch tells me that it's around the corner - for more than MMCX) Regards, Bjorn > > > > And if this (which?) clock requires a higher corner than the lowest > > possible in order to tick at this "lowest" frequency, I'm certainly > > interested in some more details. > >
On 7/17/2021 3:29 AM, Bjorn Andersson wrote: > On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: > >> Quoting Bjorn Andersson (2021-07-16 13:52:12) >>> On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: >>> >>>> Quoting Bjorn Andersson (2021-07-16 13:18:56) >>>>> On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: >>>>> >>>>>> qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) >>>>>> Though qup-i2c does not support DVFS, it still needs to vote for a >>>>>> performance state on 'CX' to satisfy the 19.2 Mhz clock frequency >>>>>> requirement. >>>>>> >>>>> >>>>> Sounds good, but... >>>>> >>>>>> Use 'required-opps' to pass this information from >>>>>> device tree, and also add the power-domains property to specify >>>>>> the CX power-domain. >>>>>> >>>>> >>>>> ..is the required-opps really needed with my rpmhpd patch in place? >>>>> >>>> >>>> Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for >>>> CX. >>> >>> On e.g. sm8250 the first available non-zero corner presented in cmd-db >>> is low_svs. what rail is this? the mmcx? Perhaps it does not support RET. cx usually supports both collapse state and RET. >> >> Indeed. On sc7180 it's not the first non-zero corner. I suppose >> retention for CX isn't actually used when the SoC is awake so your >> rpmhpd patch is putting in a vote for something that doesn't do anything >> at runtime for CX? I imagine that rpmh only sets the aggregate corner to >> retention when the whole SoC is suspended/sleeping, otherwise things >> wouldn't go very well. Similarly, min_svs may be VDD minimization? If >> so, those first two states are basically states that shouldn't be used >> at runtime, almost like sleep states. >> > > But if that's the case, I don't think it's appropriate for the "enabled > state" of the domain to use any of those corners. I rechecked the downstream kernels where all this voting happens from within the clock drivers, and I do see votes to min_svs for some clocks, but Stephen is right that RET is not something that's voted on while in active state. But always going with something just above the ret level while active will also not work for all devices, for instance for i2c on 7180, it needs a cx vote of low svs while the rail (cx) does support something lower than that which is min svs. (why can't it just work with min svs?, I don't know, these values and recommendations come in from the voltage plans published by HW teams for every SoC and we just end up using them in SW, perhaps something to dig further and understand which I will try and do but these are the values in voltage plans and downstream kernels which work for now) > > As this means that anyone who needs any of the rpmhpd domains active > also needs to specify required-opps, which wouldn't be needed for any > other power domain provider. > > And more importantly it means that a device sitting in a GDSC, which > would be parented by a rpmhpd domain has no way to specify the GDSC and > trickle the minimum-vote up to the rpmhpd domain. (And I know that we > don't describe the parentship of the GDSCs today, but this patch > tells me that it's around the corner - for more than MMCX) > > Regards, > Bjorn > >>> >>> And if this (which?) clock requires a higher corner than the lowest >>> possible in order to tick at this "lowest" frequency, I'm certainly >>> interested in some more details. >>>
On 7/20/2021 12:49 AM, Bjorn Andersson wrote: > On Mon 19 Jul 04:37 CDT 2021, Rajendra Nayak wrote: > >> >> >> On 7/17/2021 3:29 AM, Bjorn Andersson wrote: >>> On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: >>> >>>> Quoting Bjorn Andersson (2021-07-16 13:52:12) >>>>> On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: >>>>> >>>>>> Quoting Bjorn Andersson (2021-07-16 13:18:56) >>>>>>> On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: >>>>>>> >>>>>>>> qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) >>>>>>>> Though qup-i2c does not support DVFS, it still needs to vote for a >>>>>>>> performance state on 'CX' to satisfy the 19.2 Mhz clock frequency >>>>>>>> requirement. >>>>>>>> >>>>>>> >>>>>>> Sounds good, but... >>>>>>> >>>>>>>> Use 'required-opps' to pass this information from >>>>>>>> device tree, and also add the power-domains property to specify >>>>>>>> the CX power-domain. >>>>>>>> >>>>>>> >>>>>>> ..is the required-opps really needed with my rpmhpd patch in place? >>>>>>> >>>>>> >>>>>> Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for >>>>>> CX. >>>>> >>>>> On e.g. sm8250 the first available non-zero corner presented in cmd-db >>>>> is low_svs. >> >> what rail is this? the mmcx? Perhaps it does not support RET. >> cx usually supports both collapse state and RET. >> > > That was the one I was specifically looking at for the MDSS_GDSC->MMCX > issue, so it's likely I didn't look elsewhere. > >>>> >>>> Indeed. On sc7180 it's not the first non-zero corner. I suppose >>>> retention for CX isn't actually used when the SoC is awake so your >>>> rpmhpd patch is putting in a vote for something that doesn't do anything >>>> at runtime for CX? I imagine that rpmh only sets the aggregate corner to >>>> retention when the whole SoC is suspended/sleeping, otherwise things >>>> wouldn't go very well. Similarly, min_svs may be VDD minimization? If >>>> so, those first two states are basically states that shouldn't be used >>>> at runtime, almost like sleep states. >>>> >>> >>> But if that's the case, I don't think it's appropriate for the "enabled >>> state" of the domain to use any of those corners. >> >> I rechecked the downstream kernels where all this voting happens from within >> the clock drivers, and I do see votes to min_svs for some clocks, but Stephen is >> right that RET is not something that's voted on while in active state. >> >> But always going with something just above the ret level while active will also >> not work for all devices, for instance for i2c on 7180, it needs a cx vote of >> low svs while the rail (cx) does support something lower than that which is min svs. >> (why can't it just work with min svs?, I don't know, these values and recommendations >> come in from the voltage plans published by HW teams for every SoC and we just end up >> using them in SW, perhaps something to dig further and understand which I will try and >> do but these are the values in voltage plans and downstream kernels which work for now) >> > > So to some degree this invalidates my argumentation about the > enabled_corner in rpmhpd, given that "enabled" means a different corner > for each rail - not just the one with lowest non-zero value. Right, it might work in some cases but might not work for all. > > So perhaps instead of introducing the enabled_corner we need to > introduce your patch and slap a WARN_ON(corner == 0) in > rpmhpd_power_on() - to ensure that all clients that uses a rpmhpd domain > actually do vote for a high enough corner? So this would mean the expectation is that the clients set the perf state/corner before they call power_on? I don;t think that's the case today with most clients, infact its the opposite, we power on first and then make a call to set the perf state of the domain. > > Regards, > Bjorn > >>> >>> As this means that anyone who needs any of the rpmhpd domains active >>> also needs to specify required-opps, which wouldn't be needed for any >>> other power domain provider. >>> >>> And more importantly it means that a device sitting in a GDSC, which >>> would be parented by a rpmhpd domain has no way to specify the GDSC and >>> trickle the minimum-vote up to the rpmhpd domain. (And I know that we >>> don't describe the parentship of the GDSCs today, but this patch >>> tells me that it's around the corner - for more than MMCX) >>> >>> Regards, >>> Bjorn >>> >>>>> >>>>> And if this (which?) clock requires a higher corner than the lowest >>>>> possible in order to tick at this "lowest" frequency, I'm certainly >>>>> interested in some more details. >>>>> >> >> -- >> QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member >> of Code Aurora Forum, hosted by The Linux Foundation -- QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation
On Mon 19 Jul 23:29 CDT 2021, Rajendra Nayak wrote: > > > On 7/20/2021 12:49 AM, Bjorn Andersson wrote: > > On Mon 19 Jul 04:37 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > On 7/17/2021 3:29 AM, Bjorn Andersson wrote: > > > > On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: > > > > > > > > > Quoting Bjorn Andersson (2021-07-16 13:52:12) > > > > > > On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: > > > > > > > > > > > > > Quoting Bjorn Andersson (2021-07-16 13:18:56) > > > > > > > > On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > > > > > > > qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) > > > > > > > > > Though qup-i2c does not support DVFS, it still needs to vote for a > > > > > > > > > performance state on 'CX' to satisfy the 19.2 Mhz clock frequency > > > > > > > > > requirement. > > > > > > > > > > > > > > > > > > > > > > > > > Sounds good, but... > > > > > > > > > > > > > > > > > Use 'required-opps' to pass this information from > > > > > > > > > device tree, and also add the power-domains property to specify > > > > > > > > > the CX power-domain. > > > > > > > > > > > > > > > > > > > > > > > > > ..is the required-opps really needed with my rpmhpd patch in place? > > > > > > > > > > > > > > > > > > > > > > Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for > > > > > > > CX. > > > > > > > > > > > > On e.g. sm8250 the first available non-zero corner presented in cmd-db > > > > > > is low_svs. > > > > > > what rail is this? the mmcx? Perhaps it does not support RET. > > > cx usually supports both collapse state and RET. > > > > > > > That was the one I was specifically looking at for the MDSS_GDSC->MMCX > > issue, so it's likely I didn't look elsewhere. > > > > > > > > > > > > Indeed. On sc7180 it's not the first non-zero corner. I suppose > > > > > retention for CX isn't actually used when the SoC is awake so your > > > > > rpmhpd patch is putting in a vote for something that doesn't do anything > > > > > at runtime for CX? I imagine that rpmh only sets the aggregate corner to > > > > > retention when the whole SoC is suspended/sleeping, otherwise things > > > > > wouldn't go very well. Similarly, min_svs may be VDD minimization? If > > > > > so, those first two states are basically states that shouldn't be used > > > > > at runtime, almost like sleep states. > > > > > > > > > > > > > But if that's the case, I don't think it's appropriate for the "enabled > > > > state" of the domain to use any of those corners. > > > > > > I rechecked the downstream kernels where all this voting happens from within > > > the clock drivers, and I do see votes to min_svs for some clocks, but Stephen is > > > right that RET is not something that's voted on while in active state. > > > > > > But always going with something just above the ret level while active will also > > > not work for all devices, for instance for i2c on 7180, it needs a cx vote of > > > low svs while the rail (cx) does support something lower than that which is min svs. > > > (why can't it just work with min svs?, I don't know, these values and recommendations > > > come in from the voltage plans published by HW teams for every SoC and we just end up > > > using them in SW, perhaps something to dig further and understand which I will try and > > > do but these are the values in voltage plans and downstream kernels which work for now) > > > > > > > So to some degree this invalidates my argumentation about the > > enabled_corner in rpmhpd, given that "enabled" means a different corner > > for each rail - not just the one with lowest non-zero value. > > Right, it might work in some cases but might not work for all. > Which makes it way less desirable. The enable state for rpmhpd power domains doesn't meet my expectations for how a power domain should behave, but we should at least be consistent across all consumers of it then... But the original issue remains, that when a device is powered by MDSS_GDSC, which is a subdomain of MMCX we still need to ensure that "on" for MMCX is actually "on" - which just happens to be the first non-0 corner. But I presume we will end up having to do the same with &gcc's GDSCs, which are subdomains of CX and MX where this isn't true. > > > > So perhaps instead of introducing the enabled_corner we need to > > introduce your patch and slap a WARN_ON(corner == 0) in > > rpmhpd_power_on() - to ensure that all clients that uses a rpmhpd domain > > actually do vote for a high enough corner? > > So this would mean the expectation is that the clients set the perf state/corner > before they call power_on? I don;t think that's the case today with most clients, > infact its the opposite, we power on first and then make a call to set the perf > state of the domain. > You're right, it's pretty much always the opposite, given that genpd will always enable the domain during attach. Regards, Bjorn > > > > Regards, > > Bjorn > > > > > > > > > > As this means that anyone who needs any of the rpmhpd domains active > > > > also needs to specify required-opps, which wouldn't be needed for any > > > > other power domain provider. > > > > > > > > And more importantly it means that a device sitting in a GDSC, which > > > > would be parented by a rpmhpd domain has no way to specify the GDSC and > > > > trickle the minimum-vote up to the rpmhpd domain. (And I know that we > > > > don't describe the parentship of the GDSCs today, but this patch > > > > tells me that it's around the corner - for more than MMCX) > > > > > > > > Regards, > > > > Bjorn > > > > > > > > > > > > > > > > And if this (which?) clock requires a higher corner than the lowest > > > > > > possible in order to tick at this "lowest" frequency, I'm certainly > > > > > > interested in some more details. > > > > > > > > > > > > -- > > > QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member > > > of Code Aurora Forum, hosted by The Linux Foundation > > -- > QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member > of Code Aurora Forum, hosted by The Linux Foundation
On 7/25/2021 10:31 PM, Bjorn Andersson wrote: > On Mon 19 Jul 23:29 CDT 2021, Rajendra Nayak wrote: > >> >> >> On 7/20/2021 12:49 AM, Bjorn Andersson wrote: >>> On Mon 19 Jul 04:37 CDT 2021, Rajendra Nayak wrote: >>> >>>> >>>> >>>> On 7/17/2021 3:29 AM, Bjorn Andersson wrote: >>>>> On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: >>>>> >>>>>> Quoting Bjorn Andersson (2021-07-16 13:52:12) >>>>>>> On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: >>>>>>> >>>>>>>> Quoting Bjorn Andersson (2021-07-16 13:18:56) >>>>>>>>> On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: >>>>>>>>> >>>>>>>>>> qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) >>>>>>>>>> Though qup-i2c does not support DVFS, it still needs to vote for a >>>>>>>>>> performance state on 'CX' to satisfy the 19.2 Mhz clock frequency >>>>>>>>>> requirement. >>>>>>>>>> >>>>>>>>> >>>>>>>>> Sounds good, but... >>>>>>>>> >>>>>>>>>> Use 'required-opps' to pass this information from >>>>>>>>>> device tree, and also add the power-domains property to specify >>>>>>>>>> the CX power-domain. >>>>>>>>>> >>>>>>>>> >>>>>>>>> ..is the required-opps really needed with my rpmhpd patch in place? >>>>>>>>> >>>>>>>> >>>>>>>> Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for >>>>>>>> CX. >>>>>>> >>>>>>> On e.g. sm8250 the first available non-zero corner presented in cmd-db >>>>>>> is low_svs. >>>> >>>> what rail is this? the mmcx? Perhaps it does not support RET. >>>> cx usually supports both collapse state and RET. >>>> >>> >>> That was the one I was specifically looking at for the MDSS_GDSC->MMCX >>> issue, so it's likely I didn't look elsewhere. >>> >>>>>> >>>>>> Indeed. On sc7180 it's not the first non-zero corner. I suppose >>>>>> retention for CX isn't actually used when the SoC is awake so your >>>>>> rpmhpd patch is putting in a vote for something that doesn't do anything >>>>>> at runtime for CX? I imagine that rpmh only sets the aggregate corner to >>>>>> retention when the whole SoC is suspended/sleeping, otherwise things >>>>>> wouldn't go very well. Similarly, min_svs may be VDD minimization? If >>>>>> so, those first two states are basically states that shouldn't be used >>>>>> at runtime, almost like sleep states. >>>>>> >>>>> >>>>> But if that's the case, I don't think it's appropriate for the "enabled >>>>> state" of the domain to use any of those corners. >>>> >>>> I rechecked the downstream kernels where all this voting happens from within >>>> the clock drivers, and I do see votes to min_svs for some clocks, but Stephen is >>>> right that RET is not something that's voted on while in active state. >>>> >>>> But always going with something just above the ret level while active will also >>>> not work for all devices, for instance for i2c on 7180, it needs a cx vote of >>>> low svs while the rail (cx) does support something lower than that which is min svs. >>>> (why can't it just work with min svs?, I don't know, these values and recommendations >>>> come in from the voltage plans published by HW teams for every SoC and we just end up >>>> using them in SW, perhaps something to dig further and understand which I will try and >>>> do but these are the values in voltage plans and downstream kernels which work for now) >>>> >>> >>> So to some degree this invalidates my argumentation about the >>> enabled_corner in rpmhpd, given that "enabled" means a different corner >>> for each rail - not just the one with lowest non-zero value. >> >> Right, it might work in some cases but might not work for all. >> > > Which makes it way less desirable. > > The enable state for rpmhpd power domains doesn't meet my expectations > for how a power domain should behave, Right and that's perhaps because these are not the usual power-domains, which have one "on/active" state and one or more "off/inactive" states (off/ret/clock-stop) Rpmhpd has multiple "on/active" states, and whats "on/active" for one consumer might not be "on/active" for another, so this information is hard to be managed at a generic level and these requests in some way or the other need to come in explicitly from the resp. consumers. > but we should at least be > consistent across all consumers of it then... > > > But the original issue remains, that when a device is powered by > MDSS_GDSC, which is a subdomain of MMCX we still need to ensure that > "on" for MMCX is actually "on" - which just happens to be the first > non-0 corner. > > But I presume we will end up having to do the same with &gcc's GDSCs, > which are subdomains of CX and MX where this isn't true. > >>> >>> So perhaps instead of introducing the enabled_corner we need to >>> introduce your patch and slap a WARN_ON(corner == 0) in >>> rpmhpd_power_on() - to ensure that all clients that uses a rpmhpd domain >>> actually do vote for a high enough corner? >> >> So this would mean the expectation is that the clients set the perf state/corner >> before they call power_on? I don;t think that's the case today with most clients, >> infact its the opposite, we power on first and then make a call to set the perf >> state of the domain. >> > > You're right, it's pretty much always the opposite, given that genpd > will always enable the domain during attach. > > Regards, > Bjorn > >>> >>> Regards, >>> Bjorn >>> >>>>> >>>>> As this means that anyone who needs any of the rpmhpd domains active >>>>> also needs to specify required-opps, which wouldn't be needed for any >>>>> other power domain provider. >>>>> >>>>> And more importantly it means that a device sitting in a GDSC, which >>>>> would be parented by a rpmhpd domain has no way to specify the GDSC and >>>>> trickle the minimum-vote up to the rpmhpd domain. (And I know that we >>>>> don't describe the parentship of the GDSCs today, but this patch >>>>> tells me that it's around the corner - for more than MMCX) >>>>> >>>>> Regards, >>>>> Bjorn >>>>> >>>>>>> >>>>>>> And if this (which?) clock requires a higher corner than the lowest >>>>>>> possible in order to tick at this "lowest" frequency, I'm certainly >>>>>>> interested in some more details. >>>>>>> >>>> >>>> -- >>>> QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member >>>> of Code Aurora Forum, hosted by The Linux Foundation >> >> -- >> QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member >> of Code Aurora Forum, hosted by The Linux Foundation
On Tue 27 Jul 02:35 CDT 2021, Rajendra Nayak wrote: > > On 7/25/2021 10:31 PM, Bjorn Andersson wrote: > > On Mon 19 Jul 23:29 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > On 7/20/2021 12:49 AM, Bjorn Andersson wrote: > > > > On Mon 19 Jul 04:37 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > > > > > > > > > On 7/17/2021 3:29 AM, Bjorn Andersson wrote: > > > > > > On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: > > > > > > > > > > > > > Quoting Bjorn Andersson (2021-07-16 13:52:12) > > > > > > > > On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: > > > > > > > > > > > > > > > > > Quoting Bjorn Andersson (2021-07-16 13:18:56) > > > > > > > > > > On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > > > > > > > > > > > qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) > > > > > > > > > > > Though qup-i2c does not support DVFS, it still needs to vote for a > > > > > > > > > > > performance state on 'CX' to satisfy the 19.2 Mhz clock frequency > > > > > > > > > > > requirement. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Sounds good, but... > > > > > > > > > > > > > > > > > > > > > Use 'required-opps' to pass this information from > > > > > > > > > > > device tree, and also add the power-domains property to specify > > > > > > > > > > > the CX power-domain. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > ..is the required-opps really needed with my rpmhpd patch in place? > > > > > > > > > > > > > > > > > > > > > > > > > > > > Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for > > > > > > > > > CX. > > > > > > > > > > > > > > > > On e.g. sm8250 the first available non-zero corner presented in cmd-db > > > > > > > > is low_svs. > > > > > > > > > > what rail is this? the mmcx? Perhaps it does not support RET. > > > > > cx usually supports both collapse state and RET. > > > > > > > > > > > > > That was the one I was specifically looking at for the MDSS_GDSC->MMCX > > > > issue, so it's likely I didn't look elsewhere. > > > > > > > > > > > > > > > > > > Indeed. On sc7180 it's not the first non-zero corner. I suppose > > > > > > > retention for CX isn't actually used when the SoC is awake so your > > > > > > > rpmhpd patch is putting in a vote for something that doesn't do anything > > > > > > > at runtime for CX? I imagine that rpmh only sets the aggregate corner to > > > > > > > retention when the whole SoC is suspended/sleeping, otherwise things > > > > > > > wouldn't go very well. Similarly, min_svs may be VDD minimization? If > > > > > > > so, those first two states are basically states that shouldn't be used > > > > > > > at runtime, almost like sleep states. > > > > > > > > > > > > > > > > > > > But if that's the case, I don't think it's appropriate for the "enabled > > > > > > state" of the domain to use any of those corners. > > > > > > > > > > I rechecked the downstream kernels where all this voting happens from within > > > > > the clock drivers, and I do see votes to min_svs for some clocks, but Stephen is > > > > > right that RET is not something that's voted on while in active state. > > > > > > > > > > But always going with something just above the ret level while active will also > > > > > not work for all devices, for instance for i2c on 7180, it needs a cx vote of > > > > > low svs while the rail (cx) does support something lower than that which is min svs. > > > > > (why can't it just work with min svs?, I don't know, these values and recommendations > > > > > come in from the voltage plans published by HW teams for every SoC and we just end up > > > > > using them in SW, perhaps something to dig further and understand which I will try and > > > > > do but these are the values in voltage plans and downstream kernels which work for now) > > > > > > > > > > > > > So to some degree this invalidates my argumentation about the > > > > enabled_corner in rpmhpd, given that "enabled" means a different corner > > > > for each rail - not just the one with lowest non-zero value. > > > > > > Right, it might work in some cases but might not work for all. > > > > > > > Which makes it way less desirable. > > > > The enable state for rpmhpd power domains doesn't meet my expectations > > for how a power domain should behave, > > Right and that's perhaps because these are not the usual power-domains, > which have one "on/active" state and one or more "off/inactive" states (off/ret/clock-stop) > Rpmhpd has multiple "on/active" states, and whats "on/active" for one consumer > might not be "on/active" for another, so this information is hard to be managed > at a generic level and these requests in some way or the other need to come > in explicitly from the resp. consumers. > I think it's fine if we just acknowledge that this is how the rpmhpd domains works. But I am worried about how we're going to handle the case where the consumer is indirectly referencing one of these power-domains using a subdomain (gdsc). And the open question is if a solution to that problem will solve this problem as well, or if we need to have this and some mechanism to describe the "on state" for the parent of a subdomain. Regards, Bjorn
On 28/07/2021 06:46, Bjorn Andersson wrote: > On Tue 27 Jul 02:35 CDT 2021, Rajendra Nayak wrote: > >> >> On 7/25/2021 10:31 PM, Bjorn Andersson wrote: >>> On Mon 19 Jul 23:29 CDT 2021, Rajendra Nayak wrote: >>> >>>> >>>> >>>> On 7/20/2021 12:49 AM, Bjorn Andersson wrote: >>>>> On Mon 19 Jul 04:37 CDT 2021, Rajendra Nayak wrote: >>>>> >>>>>> >>>>>> >>>>>> On 7/17/2021 3:29 AM, Bjorn Andersson wrote: >>>>>>> On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: >>>>>>> >>>>>>>> Quoting Bjorn Andersson (2021-07-16 13:52:12) >>>>>>>>> On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: >>>>>>>>> >>>>>>>>>> Quoting Bjorn Andersson (2021-07-16 13:18:56) >>>>>>>>>>> On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: >>>>>>>>>>> >>>>>>>>>>>> qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) >>>>>>>>>>>> Though qup-i2c does not support DVFS, it still needs to vote for a >>>>>>>>>>>> performance state on 'CX' to satisfy the 19.2 Mhz clock frequency >>>>>>>>>>>> requirement. >>>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> Sounds good, but... >>>>>>>>>>> >>>>>>>>>>>> Use 'required-opps' to pass this information from >>>>>>>>>>>> device tree, and also add the power-domains property to specify >>>>>>>>>>>> the CX power-domain. >>>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> ..is the required-opps really needed with my rpmhpd patch in place? >>>>>>>>>>> >>>>>>>>>> >>>>>>>>>> Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for >>>>>>>>>> CX. >>>>>>>>> >>>>>>>>> On e.g. sm8250 the first available non-zero corner presented in cmd-db >>>>>>>>> is low_svs. >>>>>> >>>>>> what rail is this? the mmcx? Perhaps it does not support RET. >>>>>> cx usually supports both collapse state and RET. >>>>>> >>>>> >>>>> That was the one I was specifically looking at for the MDSS_GDSC->MMCX >>>>> issue, so it's likely I didn't look elsewhere. >>>>> >>>>>>>> >>>>>>>> Indeed. On sc7180 it's not the first non-zero corner. I suppose >>>>>>>> retention for CX isn't actually used when the SoC is awake so your >>>>>>>> rpmhpd patch is putting in a vote for something that doesn't do anything >>>>>>>> at runtime for CX? I imagine that rpmh only sets the aggregate corner to >>>>>>>> retention when the whole SoC is suspended/sleeping, otherwise things >>>>>>>> wouldn't go very well. Similarly, min_svs may be VDD minimization? If >>>>>>>> so, those first two states are basically states that shouldn't be used >>>>>>>> at runtime, almost like sleep states. >>>>>>>> >>>>>>> >>>>>>> But if that's the case, I don't think it's appropriate for the "enabled >>>>>>> state" of the domain to use any of those corners. >>>>>> >>>>>> I rechecked the downstream kernels where all this voting happens from within >>>>>> the clock drivers, and I do see votes to min_svs for some clocks, but Stephen is >>>>>> right that RET is not something that's voted on while in active state. >>>>>> >>>>>> But always going with something just above the ret level while active will also >>>>>> not work for all devices, for instance for i2c on 7180, it needs a cx vote of >>>>>> low svs while the rail (cx) does support something lower than that which is min svs. >>>>>> (why can't it just work with min svs?, I don't know, these values and recommendations >>>>>> come in from the voltage plans published by HW teams for every SoC and we just end up >>>>>> using them in SW, perhaps something to dig further and understand which I will try and >>>>>> do but these are the values in voltage plans and downstream kernels which work for now) >>>>>> >>>>> >>>>> So to some degree this invalidates my argumentation about the >>>>> enabled_corner in rpmhpd, given that "enabled" means a different corner >>>>> for each rail - not just the one with lowest non-zero value. >>>> >>>> Right, it might work in some cases but might not work for all. >>>> >>> >>> Which makes it way less desirable. >>> >>> The enable state for rpmhpd power domains doesn't meet my expectations >>> for how a power domain should behave, >> >> Right and that's perhaps because these are not the usual power-domains, >> which have one "on/active" state and one or more "off/inactive" states (off/ret/clock-stop) >> Rpmhpd has multiple "on/active" states, and whats "on/active" for one consumer >> might not be "on/active" for another, so this information is hard to be managed >> at a generic level and these requests in some way or the other need to come >> in explicitly from the resp. consumers. >> > > I think it's fine if we just acknowledge that this is how the rpmhpd > domains works. > > But I am worried about how we're going to handle the case where the > consumer is indirectly referencing one of these power-domains using a > subdomain (gdsc). With the proper subdomain relationship in place and with Ulf's patches, this seems to be handled correctly. gdsc sets proper level for the parent power domain, which gets voted and unvoted by the core pm code when gdsc's power domain is powered on or off. > And the open question is if a solution to that problem will solve this > problem as well, or if we need to have this and some mechanism to > describe the "on state" for the parent of a subdomain.
On Wed 28 Jul 07:01 PDT 2021, Dmitry Baryshkov wrote: > On 28/07/2021 06:46, Bjorn Andersson wrote: > > On Tue 27 Jul 02:35 CDT 2021, Rajendra Nayak wrote: > > > > > > > > On 7/25/2021 10:31 PM, Bjorn Andersson wrote: > > > > On Mon 19 Jul 23:29 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > > > > > > > > > On 7/20/2021 12:49 AM, Bjorn Andersson wrote: > > > > > > On Mon 19 Jul 04:37 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > > On 7/17/2021 3:29 AM, Bjorn Andersson wrote: > > > > > > > > On Fri 16 Jul 16:49 CDT 2021, Stephen Boyd wrote: > > > > > > > > > > > > > > > > > Quoting Bjorn Andersson (2021-07-16 13:52:12) > > > > > > > > > > On Fri 16 Jul 15:21 CDT 2021, Stephen Boyd wrote: > > > > > > > > > > > > > > > > > > > > > Quoting Bjorn Andersson (2021-07-16 13:18:56) > > > > > > > > > > > > On Fri 16 Jul 05:00 CDT 2021, Rajendra Nayak wrote: > > > > > > > > > > > > > > > > > > > > > > > > > qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) > > > > > > > > > > > > > Though qup-i2c does not support DVFS, it still needs to vote for a > > > > > > > > > > > > > performance state on 'CX' to satisfy the 19.2 Mhz clock frequency > > > > > > > > > > > > > requirement. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Sounds good, but... > > > > > > > > > > > > > > > > > > > > > > > > > Use 'required-opps' to pass this information from > > > > > > > > > > > > > device tree, and also add the power-domains property to specify > > > > > > > > > > > > > the CX power-domain. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > ..is the required-opps really needed with my rpmhpd patch in place? > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Yes? Because rpmhpd_opp_low_svs is not the lowest performance state for > > > > > > > > > > > CX. > > > > > > > > > > > > > > > > > > > > On e.g. sm8250 the first available non-zero corner presented in cmd-db > > > > > > > > > > is low_svs. > > > > > > > > > > > > > > what rail is this? the mmcx? Perhaps it does not support RET. > > > > > > > cx usually supports both collapse state and RET. > > > > > > > > > > > > > > > > > > > That was the one I was specifically looking at for the MDSS_GDSC->MMCX > > > > > > issue, so it's likely I didn't look elsewhere. > > > > > > > > > > > > > > > > > > > > > > > > Indeed. On sc7180 it's not the first non-zero corner. I suppose > > > > > > > > > retention for CX isn't actually used when the SoC is awake so your > > > > > > > > > rpmhpd patch is putting in a vote for something that doesn't do anything > > > > > > > > > at runtime for CX? I imagine that rpmh only sets the aggregate corner to > > > > > > > > > retention when the whole SoC is suspended/sleeping, otherwise things > > > > > > > > > wouldn't go very well. Similarly, min_svs may be VDD minimization? If > > > > > > > > > so, those first two states are basically states that shouldn't be used > > > > > > > > > at runtime, almost like sleep states. > > > > > > > > > > > > > > > > > > > > > > > > > But if that's the case, I don't think it's appropriate for the "enabled > > > > > > > > state" of the domain to use any of those corners. > > > > > > > > > > > > > > I rechecked the downstream kernels where all this voting happens from within > > > > > > > the clock drivers, and I do see votes to min_svs for some clocks, but Stephen is > > > > > > > right that RET is not something that's voted on while in active state. > > > > > > > > > > > > > > But always going with something just above the ret level while active will also > > > > > > > not work for all devices, for instance for i2c on 7180, it needs a cx vote of > > > > > > > low svs while the rail (cx) does support something lower than that which is min svs. > > > > > > > (why can't it just work with min svs?, I don't know, these values and recommendations > > > > > > > come in from the voltage plans published by HW teams for every SoC and we just end up > > > > > > > using them in SW, perhaps something to dig further and understand which I will try and > > > > > > > do but these are the values in voltage plans and downstream kernels which work for now) > > > > > > > > > > > > > > > > > > > So to some degree this invalidates my argumentation about the > > > > > > enabled_corner in rpmhpd, given that "enabled" means a different corner > > > > > > for each rail - not just the one with lowest non-zero value. > > > > > > > > > > Right, it might work in some cases but might not work for all. > > > > > > > > > > > > > Which makes it way less desirable. > > > > > > > > The enable state for rpmhpd power domains doesn't meet my expectations > > > > for how a power domain should behave, > > > > > > Right and that's perhaps because these are not the usual power-domains, > > > which have one "on/active" state and one or more "off/inactive" states (off/ret/clock-stop) > > > Rpmhpd has multiple "on/active" states, and whats "on/active" for one consumer > > > might not be "on/active" for another, so this information is hard to be managed > > > at a generic level and these requests in some way or the other need to come > > > in explicitly from the resp. consumers. > > > > > > > I think it's fine if we just acknowledge that this is how the rpmhpd > > domains works. > > > > But I am worried about how we're going to handle the case where the > > consumer is indirectly referencing one of these power-domains using a > > subdomain (gdsc). > > With the proper subdomain relationship in place and with Ulf's patches, this > seems to be handled correctly. gdsc sets proper level for the parent power > domain, which gets voted and unvoted by the core pm code when gdsc's power > domain is powered on or off. > Right, but this works only in our case because "on" for MMCX happens to be the first non-zero corner. What this patch points out is that for some of the other power domains my patch in the rpmhpd driver isn't sufficient - and presumably wouldn't work for other gdsc's (that are parented by CX or MX). Regards, Bjorn > > And the open question is if a solution to that problem will solve this > > problem as well, or if we need to have this and some mechanism to > > describe the "on state" for the parent of a subdomain. > > > -- > With best wishes > Dmitry
diff --git a/arch/arm64/boot/dts/qcom/sc7180.dtsi b/arch/arm64/boot/dts/qcom/sc7180.dtsi index a5d58eb..cd30185 100644 --- a/arch/arm64/boot/dts/qcom/sc7180.dtsi +++ b/arch/arm64/boot/dts/qcom/sc7180.dtsi @@ -785,8 +785,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_0 0>, <&aggre1_noc MASTER_QUP_0 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi0: spi@880000 { @@ -837,8 +839,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_0 0>, <&aggre1_noc MASTER_QUP_0 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi1: spi@884000 { @@ -889,8 +893,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_0 0>, <&aggre1_noc MASTER_QUP_0 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; uart2: serial@888000 { @@ -923,8 +929,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_0 0>, <&aggre1_noc MASTER_QUP_0 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi3: spi@88c000 { @@ -975,8 +983,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_0 0>, <&aggre1_noc MASTER_QUP_0 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; uart4: serial@890000 { @@ -1009,8 +1019,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_0 0>, <&aggre1_noc MASTER_QUP_0 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi5: spi@894000 { @@ -1074,8 +1086,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_1 0>, <&aggre2_noc MASTER_QUP_1 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi6: spi@a80000 { @@ -1126,8 +1140,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_1 0>, <&aggre2_noc MASTER_QUP_1 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; uart7: serial@a84000 { @@ -1160,8 +1176,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_1 0>, <&aggre2_noc MASTER_QUP_1 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi8: spi@a88000 { @@ -1212,8 +1230,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_1 0>, <&aggre2_noc MASTER_QUP_1 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; uart9: serial@a8c000 { @@ -1246,8 +1266,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_1 0>, <&aggre2_noc MASTER_QUP_1 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi10: spi@a90000 { @@ -1298,8 +1320,10 @@ <&gem_noc MASTER_APPSS_PROC 0 &config_noc SLAVE_QUP_1 0>, <&aggre2_noc MASTER_QUP_1 0 &mc_virt SLAVE_EBI1 0>; interconnect-names = "qup-core", "qup-config", "qup-memory"; + power-domains = <&rpmhpd SC7180_CX>; + required-opps = <&rpmhpd_opp_low_svs>; status = "disabled"; }; spi11: spi@a94000 {
qup-i2c devices on sc7180 are clocked with a fixed clock (19.2 MHz) Though qup-i2c does not support DVFS, it still needs to vote for a performance state on 'CX' to satisfy the 19.2 Mhz clock frequency requirement. Use 'required-opps' to pass this information from device tree, and also add the power-domains property to specify the CX power-domain. Signed-off-by: Rajendra Nayak <rnayak@codeaurora.org> --- arch/arm64/boot/dts/qcom/sc7180.dtsi | 24 ++++++++++++++++++++++++ 1 file changed, 24 insertions(+)