x86/perf: Default freeze_on_smi on for Comet Lake and later.
Subsystems (2)
PATCH 1/1 x86/perf: Default freeze_on_smi on for Comet Lake and later.
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as /sys/devices/cpu/freeze_on_smi). While we can and will tell our users to set freeze_on_smi manually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.
In this patch I have assumed that all non-Atom Intel microarchitectures starting with Comet Lake behave like this but it would be good for someone at Intel to verify that.
[0] See the Intel white paper "Trustworthy SMM on the Intel vPro Platform" at https://bugzilla.kernel.org/attachment.cgi?id=300300, particularly the end of page 5.
--- arch/x86/events/intel/core.c | 8 ++++++++ 1 file changed, 8 insertions(+)
arch/x86/events/intel/core.c +8 -0
@@ -6094,6 +6094,11 @@ __init int intel_pmu_init(void) x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; } +if (boot_cpu_data.x86_model ==INTEL_FAM6_COMETLAKE_L || + boot_cpu_data.x86_model ==INTEL_FAM6_COMETLAKE ) { + x86_pmu.attr_freeze_on_smi =1 ; + } +pr_cont ("Skylake events, " ); name ="skylake" ;break ;
@@ -6135,6 +6140,7 @@ __init int intel_pmu_init(void) x86_pmu.num_topdown_events =4 ; x86_pmu.update_topdown_event = icl_update_topdown_event; x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; + x86_pmu.attr_freeze_on_smi =1 ;pr_cont ("Icelake events, " ); name ="icelake" ;break ;
@@ -6172,6 +6178,7 @@ __init int intel_pmu_init(void) x86_pmu.num_topdown_events =8 ; x86_pmu.update_topdown_event = icl_update_topdown_event; x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; + x86_pmu.attr_freeze_on_smi =1 ;pr_cont ("Sapphire Rapids events, " ); name ="sapphire_rapids" ;break ;
@@ -6217,6 +6224,7 @@ __init int intel_pmu_init(void)* x86_pmu.rtm_abort_event. */ x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); + x86_pmu.attr_freeze_on_smi =1 ; td_attr = adl_hybrid_events_attrs; mem_attr = adl_hybrid_mem_attrs;
Signature · 1 line
2.34.1
Peter Zijlstra
maintainerQuoted 13 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.
Quoted 4 lines
In this patch I have assumed that all non-Atom Intel microarchitectures starting with Comet Lake behave like this but it would be good for someone at Intel to verify that.
Kan, can you look at that?
Quoted 7 lines
[0] See the Intel white paper "Trustworthy SMM on the Intel vPro Platform" at https://bugzilla.kernel.org/attachment.cgi?id=300300, particularly the end of page 5.
Signed-off-by Kyle Huey <[email protected]>
Patch seems sensible enough; I'll go queue it up unless Kan comes back with anything troublesome.
Quoted 6 lines · Peter Zijlstra
Quoted 5 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
Quoted 7 lines
Quoted 7 lines
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.
+ Andi
From we heard many times from sophisticated customers, they really hate blind spots. They want to see everything. That's why we set freeze_on_smi to 0 as default. I think the patch breaks the principle.
I don't think there is a way to notify all the users that the default kernel value will be changed. (Yes, the end user can always check the /sys/devices/cpu/freeze_on_smi to get the latest value. But in practice, no one checks it unless some errors found.) I think it may bring troubles to the users if they rely on the counts in SMM.
The patch only changes the default values for some platforms, not all platforms. The default value is not consistent among platforms anymore. It can bring confusion.
All in all, we have already exposed an interface for the end-users to change the value. If some apps, e.g., rr, doesn't want the default value, I think they can always change it in the app for all platforms. We should still keep the freeze_on_smi to 0 as default, which should benefit more users.
Quoted 8 lines
Quoted 5 lines
In this patch I have assumed that all non-Atom Intel microarchitectures starting with Comet Lake behave like this but it would be good for someone at Intel to verify that.
Kan, can you look at that?
I'm asking internally.
Thanks,
Kan
Quoted 10 lines
Quoted 7 lines
[0] See the Intel white paper "Trustworthy SMM on the Intel vPro Platform" at https://bugzilla.kernel.org/attachment.cgi?id=300300, particularly the end of page 5.
Signed-off-by Kyle Huey <[email protected]>Patch seems sensible enough; I'll go queue it up unless Kan comes back with anything troublesome.
Peter Zijlstra
maintainerQuoted 17 lines · Kan
Quoted 6 lines · Peter Zijlstra
Quoted 5 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
Quoted 7 lines
Quoted 7 lines
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.
Quoted 5 lines
+ Andi
From we heard many times from sophisticated customers, they really hate blind spots. They want to see everything. That's why we set
freeze_on_smito 0 as default. I think the patch breaks the principle.
Well, USR really, as in *REALLY* should not be counting SMM. That's just plain broken.
There's maybe an argument to include it in OS, but USR is ring-3.
Quoted 18 lines · Peter Zijlstra
Quoted 17 lines · Kan
Quoted 13 lines · Peter Zijlstra
Quoted 12 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.
Quoted 9 lines
Quoted 5 lines
+ Andi
From we heard many times from sophisticated customers, they really hate
blind spots. They want to see everything. That's why we setfreeze_on_smito 0 as default. I think the patch breaks the principle.Well, USR really, as in *REALLY* should not be counting SMM. That's just plain broken.
For the USR only case, the bit could be set.
Quoted 1 line
There's maybe an argument to include it in OS, but USR is ring-3.
But we don't have an option for the USR only case. The changing will impact the SYS case as well. Personally, maybe it's better to let the user apace app control the bit as needed rather than changing the default kernel value for all cases.
Thanks,
Kan
Peter Zijlstra
maintainerQuoted 31 lines · Kan
Quoted 18 lines · Peter Zijlstra
Quoted 17 lines · Kan
Quoted 13 lines · Peter Zijlstra
Quoted 12 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.Quoted 9 lines
Quoted 5 lines
+ Andi
From we heard many times from sophisticated customers, they really hate
blind spots. They want to see everything. That's why we setfreeze_on_smito 0 as default. I think the patch breaks the principle.Well, USR really, as in *REALLY* should not be counting SMM. That's just plain broken.
Quoted 8 lines
For the USR only case, the bit could be set.
Quoted 1 line
There's maybe an argument to include it in OS, but USR is ring-3.
But we don't have an option for the USR only case. The changing will impact the SYS case as well. Personally, maybe it's better to let the user apace app control the bit as needed rather than changing the default kernel value for all cases.
The bit is system wide, you can't sanely control it per counter. The change that made USR include SMM is insane and broken.
Quoted 18 lines · Kan
Quoted 6 lines · Peter Zijlstra
Quoted 5 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
Quoted 7 lines
Quoted 7 lines
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.
Quoted 5 lines
+ Andi
From we heard many times from sophisticated customers, they really hate
blind spots. They want to see everything. That's why we setfreeze_on_smito 0 as default. I think the patch breaks the principle.
The default kernel settings for perf events prioritize preventing information leaks to less privileged code. perf_event_paranoid defaults to 2, preventing unprivileged users from observing kernel space. If "sophisticated customers" want to see everything they have already needed privileges (or an explicit opt-in through decreasing perf_event_paranoid) for some time.
The current situation on Comet Lake+ where an unprivileged user *cannot* observe kernel code due to security concerns but simultaneously *must* observe SMM code seems rather absurd.
Quoted 5 lines
I don't think there is a way to notify all the users that the default kernel value will be changed. (Yes, the end user can always check the
/sys/devices/cpu/freeze_on_smito get the latest value. But in practice, no one checks it unless some errors found.) I think it may bring troubles to the users if they rely on the counts in SMM.
Unfortunately the new hardware has already changed the behavior without notifying users, no matter what we do here.
Quoted 3 lines
The patch only changes the default values for some platforms, not all platforms. The default value is not consistent among platforms anymore. It can bring confusion.
I don't personally object to changing freeze_on_smi for all platforms :) I was merely trying to limit the changes.
Quoted 5 lines
All in all, we have already exposed an interface for the end-users to change the value. If some apps, e.g., rr, doesn't want the default value, I think they can always change it in the app for all platforms. We should still keep the
freeze_on_smito 0 as default, which should benefit more users.
I think "people who want to just do userspace profiling like they did before can just change the value" is an unsatisfying answer, especially because freeze_on_smi requires root to change.
- Kyle
Quoted 25 lines
Quoted 9 lines
Quoted 6 lines
In this patch I have assumed that all non-Atom Intel microarchitectures starting with Comet Lake behave like this but it would be good for someone at Intel to verify that.
Kan, can you look at that?
I'm asking internally.
Thanks,
KanQuoted 10 lines
Quoted 7 lines
[0] See the Intel white paper "Trustworthy SMM on the Intel vPro Platform" at https://bugzilla.kernel.org/attachment.cgi?id=300300, particularly the end of page 5.
Signed-off-by Kyle Huey <[email protected]>Patch seems sensible enough; I'll go queue it up unless Kan comes back with anything troublesome.
Quoted 32 lines · Kyle Huey
Quoted 18 lines · Kan
Quoted 6 lines · Peter Zijlstra
Quoted 5 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0] A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
Quoted 7 lines
Quoted 7 lines
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.Quoted 5 lines
+ Andi
From we heard many times from sophisticated customers, they really hate
blind spots. They want to see everything. That's why we setfreeze_on_smito 0 as default. I think the patch breaks the principle.The default kernel settings for perf events prioritize preventing information leaks to less privileged code.
perf_event_paranoiddefaults to 2, preventing unprivileged users from observing kernel space. If "sophisticated customers" want to see everything they have already needed privileges (or an explicit opt-in through decreasingperf_event_paranoid) for some time.
Quoted 4 lines
The current situation on Comet Lake+ where an unprivileged user *cannot* observe kernel code due to security concerns but simultaneously *must* observe SMM code seems rather absurd.
I see. I was thought the unprivileged user can observe the SMM code on the previous platforms. The CML+ change only makes part of the SMM code CPL0. Seems I'm wrong. The change looks like changing the previous CPL0 code to CPL3 code. If so, yes, I think we should prevent the information leaks for the unprivileged user.
Quoted 9 lines
Quoted 5 lines
I don't think there is a way to notify all the users that the default kernel value will be changed. (Yes, the end user can always check the
/sys/devices/cpu/freeze_on_smito get the latest value. But in practice, no one checks it unless some errors found.) I think it may bring troubles to the users if they rely on the counts in SMM.Unfortunately the new hardware has already changed the behavior without notifying users, no matter what we do here.
Quoted 6 lines
Quoted 3 lines
The patch only changes the default values for some platforms, not all platforms. The default value is not consistent among platforms anymore. It can bring confusion.
I don't personally object to changing
freeze_on_smifor all platforms :) I was merely trying to limit the changes.
Changing it to all platforms seems a too big hammer. I agree we should limit it to the impacted platforms.
I've contacted the author of the white paper. I was told that the change is for the client vPro platforms. They are not sure whether it impacts Server platform or Atom platforms. I'm still working on it. I will let you and Peter know once I get more information.
Thanks,
Kan
Peter Zijlstra
maintainerQuoted 5 lines · Kan
I see. I was thought the unprivileged user can observe the SMM code on the previous platforms. The CML+ change only makes part of the SMM code CPL0. Seems I'm wrong. The change looks like changing the previous CPL0 code to CPL3 code. If so, yes, I think we should prevent the information leaks for the unprivileged user.
Right.
Quoted 7 lines
Changing it to all platforms seems a too big hammer. I agree we should limit it to the impacted platforms.
I've contacted the author of the white paper. I was told that the change is for the client vPro platforms. They are not sure whether it impacts Server platform or Atom platforms. I'm still working on it. I will let you and Peter know once I get more information.
For now I've updated the patch as per the below. I'm tempted to simply apply it as is and let it be.
Having different defaults for vPro vs !vPro chips seems more confusing than not.
We should also very much get this change reverted for future chips.
arch/x86/events/intel/core.c +16 -0
@@ -6094,6 +6094,16 @@ __init int intel_pmu_init(void) x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; } +if (boot_cpu_data.x86_model ==INTEL_FAM6_COMETLAKE_L || + boot_cpu_data.x86_model ==INTEL_FAM6_COMETLAKE ) { +/* +* For some idiotic reason SMM is visible to USR +* counters. Since this is a privilege issue, default +* disable counters in SMM for these chips. +*/ + x86_pmu.attr_freeze_on_smi =1 ; + } +pr_cont ("Skylake events, " ); name ="skylake" ;break ;
@@ -6135,6 +6145,8 @@ __init int intel_pmu_init(void) x86_pmu.num_topdown_events =4 ; x86_pmu.update_topdown_event = icl_update_topdown_event; x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; +/* SMM visible in USR, see above */ + x86_pmu.attr_freeze_on_smi =1 ;pr_cont ("Icelake events, " ); name ="icelake" ;break ;
@@ -6172,6 +6184,8 @@ __init int intel_pmu_init(void) x86_pmu.num_topdown_events =8 ; x86_pmu.update_topdown_event = icl_update_topdown_event; x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; +/* SMM visible in USR, see above */ + x86_pmu.attr_freeze_on_smi =1 ;pr_cont ("Sapphire Rapids events, " ); name ="sapphire_rapids" ;break ;
@@ -6217,6 +6231,8 @@ __init int intel_pmu_init(void)* x86_pmu.rtm_abort_event. */ x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); +/* SMM visible in USR, see above */ + x86_pmu.attr_freeze_on_smi =1 ; td_attr = adl_hybrid_events_attrs; mem_attr = adl_hybrid_mem_attrs;
Quoted 17 lines · Peter Zijlstra
Quoted 5 lines · Kan
I see. I was thought the unprivileged user can observe the SMM code on the previous platforms. The CML+ change only makes part of the SMM code CPL0. Seems I'm wrong. The change looks like changing the previous CPL0 code to CPL3 code. If so, yes, I think we should prevent the information leaks for the unprivileged user.
Right.
Quoted 7 lines
Changing it to all platforms seems a too big hammer. I agree we should limit it to the impacted platforms.
I've contacted the author of the white paper. I was told that the change is for the client vPro platforms. They are not sure whether it impacts Server platform or Atom platforms. I'm still working on it. I will let you and Peter know once I get more information.
Quoted 6 lines
For now I've updated the patch as per the below. I'm tempted to simply apply it as is and let it be.
Having different defaults for vPro vs !vPro chips seems more confusing than not.
But I think it should make sense to have different defaults for client vs server, or big core vs atom.
I'm still working on it and trying to figure out the affected generations. (Hope I can find the right contacts in the next few days.)
Quoted 2 lines
We should also very much get this change reverted for future chips.
Yes. I will discuss it with the contacts once I get the name.
Thanks,
Kan
Quoted 46 lines
arch/x86/events/intel/core.c +16
-0@@ -6094,6 +6094,16 @@ __init int intel_pmu_init(void) x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; } +if (boot_cpu_data.x86_model ==INTEL_FAM6_COMETLAKE_L || + boot_cpu_data.x86_model ==INTEL_FAM6_COMETLAKE ) { +/* +* For some idiotic reason SMM is visible to USR +* counters. Since this is a privilege issue, default +* disable counters in SMM for these chips. +*/ + x86_pmu.attr_freeze_on_smi =1 ; + } +pr_cont ("Skylake events, " ); name ="skylake" ;break ;@@ -6135,6 +6145,8 @@ __init int intel_pmu_init(void) x86_pmu.num_topdown_events =4 ; x86_pmu.update_topdown_event = icl_update_topdown_event; x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; +/* SMM visible in USR, see above */ + x86_pmu.attr_freeze_on_smi =1 ;pr_cont ("Icelake events, " ); name ="icelake" ;break ;@@ -6172,6 +6184,8 @@ __init int intel_pmu_init(void) x86_pmu.num_topdown_events =8 ; x86_pmu.update_topdown_event = icl_update_topdown_event; x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; +/* SMM visible in USR, see above */ + x86_pmu.attr_freeze_on_smi =1 ;pr_cont ("Sapphire Rapids events, " ); name ="sapphire_rapids" ;break ;@@ -6217,6 +6231,8 @@ __init int intel_pmu_init(void)* x86_pmu.rtm_abort_event. */ x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); +/* SMM visible in USR, see above */ + x86_pmu.attr_freeze_on_smi =1 ; td_attr = adl_hybrid_events_attrs; mem_attr = adl_hybrid_mem_attrs;
Quoted 2 lines · Kyle Huey
Beginning in Comet Lake, Intel extended the concept of privilege rings to SMM.[0]
SMM has always has full access to all 4 rings of protection.
Blame anyone who uses the terms "Ring -1/-2", because they are horribly misleading terms and show a fundamental misunderstanding of how this works.
On entry to SMM, the processors is in Real Mode, which is CPL0. Pretty much every handler switches to Protected Mode as soon as possible (because programming for Unreal Mode is evil). UEFI systems with 64bit firmware will set up pagetables and switch into Long mode.
But from a code organisation point of view, SMM has traditionally been a RWX free-for-all with more-than-kernel privileges and all the input handling gotchas/etc. It's no wonder that SMM is a fertile source of security issues.
Quoted 3 lines
A side effect of this is that events caused by execution of code in SMM are now visible to performance counters with IA32_PERFEVTSELx.USR set.
What is new in Comet Lake is a kernel running in SMM CPL0 which sets up usermode to run the main logic.
However, if e.g. an enterprising Coreboot developer were to decide that this CPL3 SMM plan might be a good idea, older CPUs would start manifesting the same behaviour.
Quoted 7 lines
rr[1] depends on exact counts of performance events for the user space tracee, so this change in behavior is fatal for us. It is, however, easily corrected by setting IA32_DEBUGCTL.FREEZE_WHILE_SMM to 1 (visible in sysfs as
/sys/devices/cpu/freeze_on_smi). While we can and will tell our users to setfreeze_on_smimanually when appropriate, because observing events in SMM is rarely useful to anyone, we propose to change the default value of this switch.
Frankly, it is an error that FREEZE_WHILE_SMM is under the kernels control, and not SMM's control. After all, it's SMM handling all the UEFI secrets/etc.
Linux ought to set FREEZE_WHILE_SMM unilaterally, because most kernel profiling probably won't want interference from SMM. Root can always disable FREEZE_WHILE_SMM if profiling is really wanted.
I'm not sure if anything can be done on pre-FREEZE_WHILE_SMM CPUs. Nor AMD CPUs which are also gaining CPL3 SMM logic, and don't appear to have any equivalent functionality.
~Andrew
Peter Zijlstra
maintainerQuoted 11 lines · Andrew Cooper
Frankly, it is an error that
FREEZE_WHILE_SMMis under the kernels control, and not SMM's control. After all, it's SMM handling all the UEFI secrets/etc.Linux ought to set
FREEZE_WHILE_SMMunilaterally, because most kernel profiling probably won't want interference from SMM. Root can always disableFREEZE_WHILE_SMMif profiling is really wanted.I'm not sure if anything can be done on pre-FREEZE_WHILE_SMM CPUs. Nor AMD CPUs which are also gaining CPL3 SMM logic, and don't appear to have any equivalent functionality.
Which suggests something like this?
---
arch/x86/events/intel/core.c +13 -0
@@ -4703,6 +4703,19 @@ static __initconst const struct x86_pmu intel_pmu = { .lbr_read = intel_pmu_lbr_read_64, .lbr_save = intel_pmu_lbr_save, .lbr_restore = intel_pmu_lbr_restore, + +/* +* SMM has access to all 4 rings and while traditionally SMM code only +* ran in CPL0, newer firmware is starting to make use of CPL3 in SMM. +* +* Since the EVENTSEL.{USR,OS} CPL filtering makes no distinction +* between SMM or not, this results in what should be pure userspace +* counters including SMM data. +* +* This is a clear privilege issue, therefore globally disable +* counting SMM by default. +*/ + .attr_freeze_on_smi = 1, };static __init void intel_clovertown_quirk (void )
Quoted 21 lines · Peter Zijlstra
Quoted 11 lines · Andrew Cooper
Frankly, it is an error that
FREEZE_WHILE_SMMis under the kernels control, and not SMM's control. After all, it's SMM handling all the UEFI secrets/etc.Linux ought to set
FREEZE_WHILE_SMMunilaterally, because most kernel profiling probably won't want interference from SMM. Root can always disableFREEZE_WHILE_SMMif profiling is really wanted.I'm not sure if anything can be done on pre-FREEZE_WHILE_SMM CPUs. Nor AMD CPUs which are also gaining CPL3 SMM logic, and don't appear to have any equivalent functionality.
Which suggests something like this?
---
arch/x86/events/intel/core.c +0
-0
arch/x86/events/intel/core.c static __initconst const struct x86_pmu intel_pmu = { · line 4709
Quoted 7 lines
arch/x86/events/intel/core.c +4
-0@@ -4703,6 +4703,19 @@ static __initconst const struct x86_pmu intel_pmu = { .lbr_read = intel_pmu_lbr_read_64, .lbr_save = intel_pmu_lbr_save, .lbr_restore = intel_pmu_lbr_restore, + +/* +* SMM has access to all 4 rings and while traditionally SMM code only +* ran in CPL0, newer firmware is starting to make use of CPL3 in SMM.
"2021-era firmware" ?
arch/x86/events/intel/core.c static __initconst const struct x86_pmu intel_pmu = { · line 4721
Quoted 12 lines
arch/x86/events/intel/core.c +9
-0+* +* Since the EVENTSEL.{USR,OS} CPL filtering makes no distinction +* between SMM or not, this results in what should be pure userspace +* counters including SMM data. +* +* This is a clear privilege issue, therefore globally disable +* counting SMM by default. +*/ + .attr_freeze_on_smi = 1, };static __init void intel_clovertown_quirk (void )
I'd say so, yes.
Elsewhere on this thread, there has been claims of "everyone like seeing SMM samples" as an argument for the current default, but I don't buy this.
NMIs are blocked in SMM, so PMIs will be attributed to an arbitrary wrong instruction boundary, and the entire SMM Handler's worth of counter changes will disappear into thin air, appearing as over-counting as far as the profiling target is concerned.
When it comes to SMIs, the two interesting cases are:
1) Am I deterministically triggering SMIs when I oughtn't to be, and
2) If I did race with an SMI, should I discard everything and start again.
Both require paying attention to MSR_SMI_COUNT, even if the answer to 2) is probably not when freeze_on_smi is active.
~Andrew
[tip: perf/urgent] x86/perf: Default set FREEZE_ON_SMI for all
In reply to Peter ZijlstraThe following commit has been merged into the perf/urgent branch of tip:
The following commit has been merged into the perf/urgent branch of tip:
Commit-ID: a01994f5e5c79d3a35e5e8cf4252c7f2147323c3 Gitweb: https://git.kernel.org/tip/a01994f5e5c79d3a35e5e8cf4252c7f2147323c3 Author: Peter Zijlstra <[email protected]> AuthorDate: Thu, 27 Jan 2022 12:32:51 +01:00 Committer: Peter Zijlstra <[email protected]> CommitterDate: Wed, 02 Feb 2022 13:11:39 +01:00
x86/perf: Default set FREEZE_ON_SMI for all
Kyle reported that rr[0] has started to malfunction on Comet Lake and later CPUs due to EFI starting to make use of CPL3 [1] and the PMU event filtering not distinguishing between regular CPL3 and SMM CPL3.
Since this is a privilege violation, default disable SMM visibility where possible.
Administrators wanting to observe SMM cycles can easily change this using the sysfs attribute while regular users don't have access to this file.
[1] See the Intel white paper "Trustworthy SMM on the Intel vPro Platform" at https://bugzilla.kernel.org/attachment.cgi?id=300300, particularly the end of page 5.
--- arch/x86/events/intel/core.c | 13 +++++++++++++ 1 file changed, 13 insertions(+)
arch/x86/events/intel/core.c +13 -0
@@ -4703,6 +4703,19 @@ static __initconst const struct x86_pmu intel_pmu = { .lbr_read = intel_pmu_lbr_read_64, .lbr_save = intel_pmu_lbr_save, .lbr_restore = intel_pmu_lbr_restore, + +/* +* SMM has access to all 4 rings and while traditionally SMM code only +* ran in CPL0, 2021-era firmware is starting to make use of CPL3 in SMM. +* +* Since the EVENTSEL.{USR,OS} CPL filtering makes no distinction +* between SMM or not, this results in what should be pure userspace +* counters including SMM data. +* +* This is a clear privilege issue, therefore globally disable +* counting SMM by default. +*/ + .attr_freeze_on_smi = 1, };static __init void intel_clovertown_quirk (void )