| Message ID | 20200407063345.4484-3-Jason@zx2c4.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/3] x86/mce/therm_throt: remove unused platform_thermal_notify function pointer | expand |
On Tue, 2020-04-07 at 00:33 -0600, Jason A. Donenfeld wrote: > The thermal IRQ handler uses 1.21% CPU on my system when it's hot > from > compiling things. Indeed looking at /proc/interrupts reveals quite a > lot I am curious why you are hitting threshold frequently? What is rdmsr 0x1a2 > of events coming in. Beyond logging them, the existing drivers on the > system don't appear to do very much that I'm interested in. So, add a > way to disable this entirely so that I can regain precious CPU > cycles. It is showing amount of time system is running in a constrained environment. Lots of real time and HPC folks really care about this. Thanks, Srinivas > > Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> > --- > arch/x86/Kconfig | 4 ++++ > arch/x86/kernel/cpu/mce/intel.c | 3 ++- > 2 files changed, 6 insertions(+), 1 deletion(-) > > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig > index 39e7444353af..3125a11932f2 100644 > --- a/arch/x86/Kconfig > +++ b/arch/x86/Kconfig > @@ -1157,7 +1157,11 @@ config X86_MCE_INJECT > > config X86_THERMAL_VECTOR > def_bool y > + prompt "Machine check thermal vector" > depends on X86_MCE_INTEL > + ---help--- > + Provide support for capturing thermal events, logging them, > and > + passing them off to other drivers. > > config X86_MCE_THERMAL_VERBOSE > bool "Verbose logging for thermal events" > diff --git a/arch/x86/kernel/cpu/mce/intel.c > b/arch/x86/kernel/cpu/mce/intel.c > index f996ffb887bc..d14f1922fb49 100644 > --- a/arch/x86/kernel/cpu/mce/intel.c > +++ b/arch/x86/kernel/cpu/mce/intel.c > @@ -511,7 +511,8 @@ static void intel_ppin_init(struct cpuinfo_x86 > *c) > > void mce_intel_feature_init(struct cpuinfo_x86 *c) > { > - intel_init_thermal(c); > + if (IS_ENABLED(CONFIG_X86_THERMAL_VECTOR)) > + intel_init_thermal(c); > intel_init_cmci(); > intel_init_lmce(); > intel_ppin_init(c);
On Mon, Apr 13, 2020 at 9:38 PM Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com> wrote: > > On Tue, 2020-04-07 at 00:33 -0600, Jason A. Donenfeld wrote: > > The thermal IRQ handler uses 1.21% CPU on my system when it's hot > > from > > compiling things. Indeed looking at /proc/interrupts reveals quite a > > lot > I am curious why you are hitting threshold frequently? > What is rdmsr 0x1a2 5640000 > > of events coming in. Beyond logging them, the existing drivers on the > > system don't appear to do very much that I'm interested in. So, add a > > way to disable this entirely so that I can regain precious CPU > > cycles. > It is showing amount of time system is running in a constrained > environment. Lots of real time and HPC folks really care about this. Which is why this patch adds an option, not a full removal or something. Real time and HPC people can keep their expensive interrupt. Other people with different varieties of system can disable it.
On Mon, 2020-04-13 at 22:21 -0600, Jason A. Donenfeld wrote: > On Mon, Apr 13, 2020 at 9:38 PM Srinivas Pandruvada > <srinivas.pandruvada@linux.intel.com> wrote: > > On Tue, 2020-04-07 at 00:33 -0600, Jason A. Donenfeld wrote: > > > The thermal IRQ handler uses 1.21% CPU on my system when it's hot > > > from > > > compiling things. Indeed looking at /proc/interrupts reveals > > > quite a > > > lot > > I am curious why you are hitting threshold frequently? > > What is rdmsr 0x1a2 > > 5640000 You are getting too many interrupts at 95C. You should look at your cooling system. > > > > of events coming in. Beyond logging them, the existing drivers on > > > the > > > system don't appear to do very much that I'm interested in. So, > > > add a > > > way to disable this entirely so that I can regain precious CPU > > > cycles. > > It is showing amount of time system is running in a constrained > > environment. Lots of real time and HPC folks really care about > > this. > > Which is why this patch adds an option, not a full removal or > something. Real time and HPC people can keep their expensive > interrupt. Other people with different varieties of system > disable > it. Generally compile time flag is not desirable. If it is what required then we should have boot time flag something in lines of existing "int_pln_enable" option. Thanks, Srinivas
On Tue, Apr 14, 2020 at 8:45 AM Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com> wrote: > > On Mon, 2020-04-13 at 22:21 -0600, Jason A. Donenfeld wrote: > > On Mon, Apr 13, 2020 at 9:38 PM Srinivas Pandruvada > > <srinivas.pandruvada@linux.intel.com> wrote: > > > On Tue, 2020-04-07 at 00:33 -0600, Jason A. Donenfeld wrote: > > > > The thermal IRQ handler uses 1.21% CPU on my system when it's hot > > > > from > > > > compiling things. Indeed looking at /proc/interrupts reveals > > > > quite a > > > > lot > > > I am curious why you are hitting threshold frequently? > > > What is rdmsr 0x1a2 > > > > 5640000 > You are getting too many interrupts at 95C. You should look at your > cooling system. > > > > > > > of events coming in. Beyond logging them, the existing drivers on > > > > the > > > > system don't appear to do very much that I'm interested in. So, > > > > add a > > > > way to disable this entirely so that I can regain precious CPU > > > > cycles. > > > It is showing amount of time system is running in a constrained > > > environment. Lots of real time and HPC folks really care about > > > this. > > > > Which is why this patch adds an option, not a full removal or > > something. Real time and HPC people can keep their expensive > > interrupt. Other people with different varieties of system > > disable > > it. > Generally compile time flag is not desirable. If it is what required > then we should have boot time flag something in lines of existing > "int_pln_enable" option. Generally it is desirable, and extremely common too. This thermal code -- which mostly functions to print some messages into kmsg -- is very verbose. This is not something I want to compile into smaller systems. This is the reason why kconfig has options in the first place. I'm not sure yet-another boottime flag makes sense for this.
On Tue, 2020-04-14 at 13:41 -0600, Jason A. Donenfeld wrote: > On Tue, Apr 14, 2020 at 8:45 AM Srinivas Pandruvada > <srinivas.pandruvada@linux.intel.com> wrote: > > On Mon, 2020-04-13 at 22:21 -0600, Jason A. Donenfeld wrote: > > > On Mon, Apr 13, 2020 at 9:38 PM Srinivas Pandruvada > > > <srinivas.pandruvada@linux.intel.com> wrote: > > > > On Tue, 2020-04-07 at 00:33 -0600, Jason A. Donenfeld wrote: > > > > > The thermal IRQ handler uses 1.21% CPU on my system when it's > > > > > hot > > > > > from > > > > > compiling things. Indeed looking at /proc/interrupts reveals > > > > > quite a > > > > > lot > > > > I am curious why you are hitting threshold frequently? > > > > What is rdmsr 0x1a2 > > > > > > 5640000 > > You are getting too many interrupts at 95C. You should look at your > > cooling system. > > > > > > > of events coming in. Beyond logging them, the existing > > > > > drivers on > > > > > the > > > > > system don't appear to do very much that I'm interested in. > > > > > So, > > > > > add a > > > > > way to disable this entirely so that I can regain precious > > > > > CPU > > > > > cycles. > > > > It is showing amount of time system is running in a constrained > > > > environment. Lots of real time and HPC folks really care about > > > > this. > > > > > > Which is why this patch adds an option, not a full removal or > > > something. Real time and HPC people can keep their expensive > > > interrupt. Other people with different varieties of system > > > disable > > > it. > > Generally compile time flag is not desirable. If it is what > > required > > then we should have boot time flag something in lines of existing > > "int_pln_enable" option. > > Generally it is desirable, and extremely common too. This thermal > code > -- which mostly functions to print some messages into kmsg -- is very > verbose. This is not something I want to compile into smaller > systems. > This is the reason why kconfig has options in the first place. I'm > not > sure yet-another boottime flag makes sense for this. Can you send log which is still showing verbose prints with the latest kernel? I can see interrupts will still fire. If it is, then temperature trend is still above 95C and cooling systems is not in control. In another window, print in loop (with sleep 1) /sys/class/thermal/thermal_zone*/temp for the zone for which "type == x86_pkg_temp"
On Tue, Apr 14, 2020 at 1:58 PM Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com> wrote: > > On Tue, 2020-04-14 at 13:41 -0600, Jason A. Donenfeld wrote: > > On Tue, Apr 14, 2020 at 8:45 AM Srinivas Pandruvada > > <srinivas.pandruvada@linux.intel.com> wrote: > > > On Mon, 2020-04-13 at 22:21 -0600, Jason A. Donenfeld wrote: > > > > On Mon, Apr 13, 2020 at 9:38 PM Srinivas Pandruvada > > > > <srinivas.pandruvada@linux.intel.com> wrote: > > > > > On Tue, 2020-04-07 at 00:33 -0600, Jason A. Donenfeld wrote: > > > > > > The thermal IRQ handler uses 1.21% CPU on my system when it's > > > > > > hot > > > > > > from > > > > > > compiling things. Indeed looking at /proc/interrupts reveals > > > > > > quite a > > > > > > lot > > > > > I am curious why you are hitting threshold frequently? > > > > > What is rdmsr 0x1a2 > > > > > > > > 5640000 > > > You are getting too many interrupts at 95C. You should look at your > > > cooling system. > > > > > > > > > of events coming in. Beyond logging them, the existing > > > > > > drivers on > > > > > > the > > > > > > system don't appear to do very much that I'm interested in. > > > > > > So, > > > > > > add a > > > > > > way to disable this entirely so that I can regain precious > > > > > > CPU > > > > > > cycles. > > > > > It is showing amount of time system is running in a constrained > > > > > environment. Lots of real time and HPC folks really care about > > > > > this. > > > > > > > > Which is why this patch adds an option, not a full removal or > > > > something. Real time and HPC people can keep their expensive > > > > interrupt. Other people with different varieties of system > > > > disable > > > > it. > > > Generally compile time flag is not desirable. If it is what > > > required > > > then we should have boot time flag something in lines of existing > > > "int_pln_enable" option. > > > > Generally it is desirable, and extremely common too. This thermal > > code > > -- which mostly functions to print some messages into kmsg -- is very > > verbose. This is not something I want to compile into smaller > > systems. > > This is the reason why kconfig has options in the first place. I'm > > not > > sure yet-another boottime flag makes sense for this. > > Can you send log which is still showing verbose prints with the latest > kernel? I can see interrupts will still fire. > > If it is, then temperature trend is still above 95C and cooling systems > is not in control. In another window, print in loop (with sleep 1) > /sys/class/thermal/thermal_zone*/temp > for the zone for which "type == x86_pkg_temp" It sounds like you're interested in debugging the cooling system of my [brand new Thinkpad P1 gen 2] laptop. I appreciate the concern, but that's really not the purpose of this patch. The purpose of this patch is to enable disabling the sizable thermal interrupt code, that does absolutely nothing for a wide variety of systems that do not want that code in their kernel. In other words, I don't think the particular thinkpad cooling aspects have anything to do with the need to add the trivial Kconfig option to not compile this code. I realize it's your code and you might really like it, but that doesn't mean everyone wants it in their kernel image.
+ Tony. On Tue, Apr 14, 2020 at 01:41:08PM -0600, Jason A. Donenfeld wrote: > Generally it is desirable, and extremely common too. This thermal code > -- which mostly functions to print some messages into kmsg -- is very > verbose. This is not something I want to compile into smaller systems. > This is the reason why kconfig has options in the first place. I'm not > sure yet-another boottime flag makes sense for this. I don't mind making the already existing option selectable and leaving it default y, i.e., keeping the current situation by default. And people who want to disable it, can then do so. I do mind to having yet another config option though. No thanks - they're already too many. So it should be an all or nothing thing. Poking quickly at that and drivers/thermal/intel/x86_pkg_temp_thermal.c, all those things do is report trip points. therm_throt reports how long the hw throttled due to hitting a trip point, etc. IINM, of course so please correct me if I'm missing anything. But if not and this all is only for reporting and doesn't have any detrimental effects on the hardware when missing from the system, then I guess we could make CONFIG_X86_THERMAL_VECTOR user-selectable. Thx.
On Tue, 2020-04-14 at 22:23 +0200, Borislav Petkov wrote: > + Tony. > > On Tue, Apr 14, 2020 at 01:41:08PM -0600, Jason A. Donenfeld wrote: > > Generally it is desirable, and extremely common too. This thermal > > code > > -- which mostly functions to print some messages into kmsg -- is > > very > > verbose. This is not something I want to compile into smaller > > systems. > > This is the reason why kconfig has options in the first place. I'm > > not > > sure yet-another boottime flag makes sense for this. > > I don't mind making the already existing option selectable and > leaving > it default y, i.e., keeping the current situation by default. And > people > who want to disable it, can then do so. > > I do mind to having yet another config option though. No thanks - > they're already too many. > > So it should be an all or nothing thing. > > Poking quickly at that and > drivers/thermal/intel/x86_pkg_temp_thermal.c, > all those things do is report trip points. therm_throt reports how > long > the hw throttled due to hitting a trip point, etc. > > IINM, of course so please correct me if I'm missing anything. > > But if not and this all is only for reporting and doesn't have any > detrimental effects on the hardware when missing from the system, > then I > guess we could make CONFIG_X86_THERMAL_VECTOR user-selectable. We can make user selectable These drivers are used for reporting only. User space can select a trip temperature via x86_pkg_temp and get notification via uevent to start additional cooling system (additional fans, liquid coooling etc), so that processor don't have to go self throttling mode. Self throttling depending on processor series and firmware can be very aggressive. In client systems thermald will set a temperature and starts power control once it reaches passive temperature limit. But it can function without x86_pkg_temp also, so even if user disables thermal reporting it can still function. Thanks, Srinivas > > Thx. >
On Tue, Apr 14, 2020 at 2:49 PM Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com> wrote: > > On Tue, 2020-04-14 at 22:23 +0200, Borislav Petkov wrote: > > + Tony. > > > > On Tue, Apr 14, 2020 at 01:41:08PM -0600, Jason A. Donenfeld wrote: > > > Generally it is desirable, and extremely common too. This thermal > > > code > > > -- which mostly functions to print some messages into kmsg -- is > > > very > > > verbose. This is not something I want to compile into smaller > > > systems. > > > This is the reason why kconfig has options in the first place. I'm > > > not > > > sure yet-another boottime flag makes sense for this. > > > > I don't mind making the already existing option selectable and > > leaving > > it default y, i.e., keeping the current situation by default. And > > people > > who want to disable it, can then do so. > > > > I do mind to having yet another config option though. No thanks - > > they're already too many. > > > > So it should be an all or nothing thing. > > > > Poking quickly at that and > > drivers/thermal/intel/x86_pkg_temp_thermal.c, > > all those things do is report trip points. therm_throt reports how > > long > > the hw throttled due to hitting a trip point, etc. > > > > IINM, of course so please correct me if I'm missing anything. > > > > But if not and this all is only for reporting and doesn't have any > > detrimental effects on the hardware when missing from the system, > > then I > > guess we could make CONFIG_X86_THERMAL_VECTOR user-selectable. > > We can make user selectable > > These drivers are used for reporting only. > User space can select a trip temperature via x86_pkg_temp and get > notification via uevent to start additional cooling system (additional I didn't see any uevent stuff. Is this part of out-of-tree modules or proprietary code that's hooking into those EXPORT_SYMBOL (non-GPL) exports? > fans, liquid coooling etc), so that processor don't have to go self > throttling mode. Self throttling depending on processor series and > firmware can be very aggressive. > In client systems thermald will set a temperature and starts power > control once it reaches passive temperature limit. But it can function > without x86_pkg_temp also, so even if user disables thermal reporting > it can still function. The 2/3 patch may be interesting as well to you. This removes the expensive work queue management stuff if the option is deselected, since all those workqueues do is print messages to kmsg, while retaining the rest of the infra.
On Tue, Apr 07, 2020 at 12:33:45AM -0600, Jason A. Donenfeld wrote: > The thermal IRQ handler uses 1.21% CPU on my system when it's hot from > compiling things. Indeed looking at /proc/interrupts reveals quite a lot > of events coming in. Beyond logging them, the existing drivers on the > system don't appear to do very much that I'm interested in. So, add a > way to disable this entirely so that I can regain precious CPU cycles. Why is this MCE code?!? hysterical raisins? Anyway, I wonder if this is something we should hook up to SCHED_THERMAL_PRESSURE, Rafael? > diff --git a/arch/x86/kernel/cpu/mce/intel.c b/arch/x86/kernel/cpu/mce/intel.c > index f996ffb887bc..d14f1922fb49 100644 > --- a/arch/x86/kernel/cpu/mce/intel.c > +++ b/arch/x86/kernel/cpu/mce/intel.c > @@ -511,7 +511,8 @@ static void intel_ppin_init(struct cpuinfo_x86 *c) > > void mce_intel_feature_init(struct cpuinfo_x86 *c) > { > - intel_init_thermal(c); > + if (IS_ENABLED(CONFIG_X86_THERMAL_VECTOR)) > + intel_init_thermal(c); > intel_init_cmci(); > intel_init_lmce(); > intel_ppin_init(c); > -- > 2.26.0 >
On Tue, 2020-04-14 at 23:40 +0200, Peter Zijlstra wrote: > On Tue, Apr 07, 2020 at 12:33:45AM -0600, Jason A. Donenfeld wrote: > > The thermal IRQ handler uses 1.21% CPU on my system when it's hot > > from > > compiling things. Indeed looking at /proc/interrupts reveals quite > > a lot > > of events coming in. Beyond logging them, the existing drivers on > > the > > system don't appear to do very much that I'm interested in. So, add > > a > > way to disable this entirely so that I can regain precious CPU > > cycles. > > Why is this MCE code?!? hysterical raisins? When this code was developed long time ago, it was also doing mce logging. But it is no longer doing mce logging, but code is still there. > > Anyway, I wonder if this is something we should hook up to > SCHED_THERMAL_PRESSURE, Rafael? > > > diff --git a/arch/x86/kernel/cpu/mce/intel.c > > b/arch/x86/kernel/cpu/mce/intel.c > > index f996ffb887bc..d14f1922fb49 100644 > > --- a/arch/x86/kernel/cpu/mce/intel.c > > +++ b/arch/x86/kernel/cpu/mce/intel.c > > @@ -511,7 +511,8 @@ static void intel_ppin_init(struct cpuinfo_x86 > > *c) > > > > void mce_intel_feature_init(struct cpuinfo_x86 *c) > > { > > - intel_init_thermal(c); > > + if (IS_ENABLED(CONFIG_X86_THERMAL_VECTOR)) > > + intel_init_thermal(c); > > intel_init_cmci(); > > intel_init_lmce(); > > intel_ppin_init(c); > > -- > > 2.26.0 > >
On Tue, 2020-04-14 at 15:07 -0600, Jason A. Donenfeld wrote: > On Tue, Apr 14, 2020 at 2:49 PM Srinivas Pandruvada > <srinivas.pandruvada@linux.intel.com> wrote: > > On Tue, 2020-04-14 at 22:23 +0200, Borislav Petkov wrote: > > > + Tony. > > > > > > On Tue, Apr 14, 2020 at 01:41:08PM -0600, Jason A. Donenfeld > > > wrote: > > > > Generally it is desirable, and extremely common too. This > > > > thermal > > > > code > > > > -- which mostly functions to print some messages into kmsg -- > > > > is > > > > very > > > > verbose. This is not something I want to compile into smaller > > > > systems. > > > > This is the reason why kconfig has options in the first place. > > > > I'm > > > > not > > > > sure yet-another boottime flag makes sense for this. > > > > > > I don't mind making the already existing option selectable and > > > leaving > > > it default y, i.e., keeping the current situation by default. And > > > people > > > who want to disable it, can then do so. > > > > > > I do mind to having yet another config option though. No thanks - > > > they're already too many. > > > > > > So it should be an all or nothing thing. > > > > > > Poking quickly at that and > > > drivers/thermal/intel/x86_pkg_temp_thermal.c, > > > all those things do is report trip points. therm_throt reports > > > how > > > long > > > the hw throttled due to hitting a trip point, etc. > > > > > > IINM, of course so please correct me if I'm missing anything. > > > > > > But if not and this all is only for reporting and doesn't have > > > any > > > detrimental effects on the hardware when missing from the system, > > > then I > > > guess we could make CONFIG_X86_THERMAL_VECTOR user-selectable. > > > > We can make user selectable > > > > These drivers are used for reporting only. > > User space can select a trip temperature via x86_pkg_temp and get > > notification via uevent to start additional cooling system > > (additional > > I didn't see any uevent stuff. Is this part of out-of-tree modules or > proprietary code that's hooking into those EXPORT_SYMBOL (non-GPL) > exports? This is not out of tree. This is x86_pkg_temp driver as part of thermal subsystem, and thermal_zone_device_update() user space governor issues uevent. But those are different modifiable thresholds not the high/low temperature thresholds. > > > fans, liquid coooling etc), so that processor don't have to go self > > throttling mode. Self throttling depending on processor series and > > firmware can be very aggressive. > > In client systems thermald will set a temperature and starts power > > control once it reaches passive temperature limit. But it can > > function > > without x86_pkg_temp also, so even if user disables thermal > > reporting > > it can still function. > > The 2/3 patch may be interesting as well to you. This removes the > expensive work queue management stuff if the option is deselected, > since all those workqueues do is print messages to kmsg, while > retaining the rest of the infra.
diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig index 39e7444353af..3125a11932f2 100644 --- a/arch/x86/Kconfig +++ b/arch/x86/Kconfig @@ -1157,7 +1157,11 @@ config X86_MCE_INJECT config X86_THERMAL_VECTOR def_bool y + prompt "Machine check thermal vector" depends on X86_MCE_INTEL + ---help--- + Provide support for capturing thermal events, logging them, and + passing them off to other drivers. config X86_MCE_THERMAL_VERBOSE bool "Verbose logging for thermal events" diff --git a/arch/x86/kernel/cpu/mce/intel.c b/arch/x86/kernel/cpu/mce/intel.c index f996ffb887bc..d14f1922fb49 100644 --- a/arch/x86/kernel/cpu/mce/intel.c +++ b/arch/x86/kernel/cpu/mce/intel.c @@ -511,7 +511,8 @@ static void intel_ppin_init(struct cpuinfo_x86 *c) void mce_intel_feature_init(struct cpuinfo_x86 *c) { - intel_init_thermal(c); + if (IS_ENABLED(CONFIG_X86_THERMAL_VECTOR)) + intel_init_thermal(c); intel_init_cmci(); intel_init_lmce(); intel_ppin_init(c);
The thermal IRQ handler uses 1.21% CPU on my system when it's hot from compiling things. Indeed looking at /proc/interrupts reveals quite a lot of events coming in. Beyond logging them, the existing drivers on the system don't appear to do very much that I'm interested in. So, add a way to disable this entirely so that I can regain precious CPU cycles. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> --- arch/x86/Kconfig | 4 ++++ arch/x86/kernel/cpu/mce/intel.c | 3 ++- 2 files changed, 6 insertions(+), 1 deletion(-)