Contains fixes w.r.t usage of LBR Freeze, Erratum 1452 and a bug in the
CPU offline path where PMU-related registers are reset.
Previous versions can be found at:
v2: https://lore.kernel.org/all/[email protected]/
v1: https://lore.kernel.org/all/[email protected]/
Changes in v3:
- As suggested by Boris, update the commit message of the first patch
with the reason behind making the LBR and PMC Freeze feature bit
visible in /proc/cpuinfo.
Changes in v2:
- Make the LBR and PMC Freeze feature bit visible in /proc/cpuinfo. As
suggested by Stephane, this will be useful to determine if it is
feasible to perform kernel FDO on a system.
Sandipan Das (3):
perf/x86/amd/lbr: Use freeze based on availability
perf/x86/amd/lbr: Discard erroneous branch entries
perf/x86/amd/core: Avoid register reset when CPU is dead
arch/x86/events/amd/core.c | 5 ++---
arch/x86/events/amd/lbr.c | 22 ++++++++++++++--------
arch/x86/include/asm/cpufeatures.h | 2 +-
arch/x86/kernel/cpu/scattered.c | 1 +
4 files changed, 18 insertions(+), 12 deletions(-)
--
2.34.1
Currently, it is assumed that LBR Freeze is supported on all processors
which have CPUID leaf 0x80000022[EAX] bit 1 set. This is incorrect as
the feature availability is additionally dependent on CPUID leaf
0x80000022[EAX] bit 2 being set which may not be set for all Zen 4
processors. Define a new feature bit for LBR and PMC freeze and set the
freeze enable bit (FLBRI) in DebugCtl (MSR 0x1d9) conditionally.
It should still be possible to use LBR without freeze for profile-guided
optimization of user programs by using an user-only branch filter during
profiling. When the user-only filter is enabled, branches are no longer
recorded after the transition to CPL 0 upon PMI arrival. When branch
entries are read in the PMI handler, the branch stack does not change.
E.g.
$ perf record -j any,u -e ex_ret_brn_tkn ./workload
Since the feature bit is visible under flags in /proc/cpuinfo, it can be
used to determine the feasibility of use-cases which require LBR Freeze
to be supported by the hardware such as profile-guided optimization of
kernels.
Fixes: ca5b7c0d9621 ("perf/x86/amd/lbr: Add LbrExtV2 branch record support")
Signed-off-by: Sandipan Das <[email protected]>
---
arch/x86/events/amd/core.c | 4 ++--
arch/x86/events/amd/lbr.c | 16 ++++++++++------
arch/x86/include/asm/cpufeatures.h | 2 +-
arch/x86/kernel/cpu/scattered.c | 1 +
4 files changed, 14 insertions(+), 9 deletions(-)
diff --git a/arch/x86/events/amd/core.c b/arch/x86/events/amd/core.c
index 4ee6390b45c9..ffdfaee08b08 100644
--- a/arch/x86/events/amd/core.c
+++ b/arch/x86/events/amd/core.c
@@ -905,8 +905,8 @@ static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
if (!status)
goto done;
- /* Read branch records before unfreezing */
- if (status & GLOBAL_STATUS_LBRS_FROZEN) {
+ /* Read branch records */
+ if (x86_pmu.lbr_nr) {
amd_pmu_lbr_read();
status &= ~GLOBAL_STATUS_LBRS_FROZEN;
}
diff --git a/arch/x86/events/amd/lbr.c b/arch/x86/events/amd/lbr.c
index eb31f850841a..110e34c59643 100644
--- a/arch/x86/events/amd/lbr.c
+++ b/arch/x86/events/amd/lbr.c
@@ -400,10 +400,12 @@ void amd_pmu_lbr_enable_all(void)
wrmsrl(MSR_AMD64_LBR_SELECT, lbr_select);
}
- rdmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl);
- rdmsrl(MSR_AMD_DBG_EXTN_CFG, dbg_extn_cfg);
+ if (cpu_feature_enabled(X86_FEATURE_AMD_LBR_PMC_FREEZE)) {
+ rdmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl);
+ wrmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
+ }
- wrmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
+ rdmsrl(MSR_AMD_DBG_EXTN_CFG, dbg_extn_cfg);
wrmsrl(MSR_AMD_DBG_EXTN_CFG, dbg_extn_cfg | DBG_EXTN_CFG_LBRV2EN);
}
@@ -416,10 +418,12 @@ void amd_pmu_lbr_disable_all(void)
return;
rdmsrl(MSR_AMD_DBG_EXTN_CFG, dbg_extn_cfg);
- rdmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl);
-
wrmsrl(MSR_AMD_DBG_EXTN_CFG, dbg_extn_cfg & ~DBG_EXTN_CFG_LBRV2EN);
- wrmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl & ~DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
+
+ if (cpu_feature_enabled(X86_FEATURE_AMD_LBR_PMC_FREEZE)) {
+ rdmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl);
+ wrmsrl(MSR_IA32_DEBUGCTLMSR, dbg_ctl & ~DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
+ }
}
__init int amd_pmu_lbr_init(void)
diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
index 4af140cf5719..e47ea31b019d 100644
--- a/arch/x86/include/asm/cpufeatures.h
+++ b/arch/x86/include/asm/cpufeatures.h
@@ -97,7 +97,7 @@
#define X86_FEATURE_SYSENTER32 ( 3*32+15) /* "" sysenter in IA32 userspace */
#define X86_FEATURE_REP_GOOD ( 3*32+16) /* REP microcode works well */
#define X86_FEATURE_AMD_LBR_V2 ( 3*32+17) /* AMD Last Branch Record Extension Version 2 */
-/* FREE, was #define X86_FEATURE_LFENCE_RDTSC ( 3*32+18) "" LFENCE synchronizes RDTSC */
+#define X86_FEATURE_AMD_LBR_PMC_FREEZE ( 3*32+18) /* AMD LBR and PMC Freeze */
#define X86_FEATURE_ACC_POWER ( 3*32+19) /* AMD Accumulated Power Mechanism */
#define X86_FEATURE_NOPL ( 3*32+20) /* The NOPL (0F 1F) instructions */
#define X86_FEATURE_ALWAYS ( 3*32+21) /* "" Always-present feature */
diff --git a/arch/x86/kernel/cpu/scattered.c b/arch/x86/kernel/cpu/scattered.c
index 0dad49a09b7a..a515328d9d7d 100644
--- a/arch/x86/kernel/cpu/scattered.c
+++ b/arch/x86/kernel/cpu/scattered.c
@@ -49,6 +49,7 @@ static const struct cpuid_bit cpuid_bits[] = {
{ X86_FEATURE_BMEC, CPUID_EBX, 3, 0x80000020, 0 },
{ X86_FEATURE_PERFMON_V2, CPUID_EAX, 0, 0x80000022, 0 },
{ X86_FEATURE_AMD_LBR_V2, CPUID_EAX, 1, 0x80000022, 0 },
+ { X86_FEATURE_AMD_LBR_PMC_FREEZE, CPUID_EAX, 2, 0x80000022, 0 },
{ 0, 0, 0, 0, 0 }
};
--
2.34.1
The Revision Guide for AMD Family 19h Model 10-1Fh processors, found at
the link below, declares Erratum 1452 which states that non-branch
entries may erroneously be recorded in the Last Branch Record (LBR)
stack with the valid and spec bits set. Such entries can be recognized
by inspecting bit 61 of the corresponding LastBranchStackToIp register.
This bit is currently reserved but if found to be set, the associated
branch entry should be discarded.
Link: https://bugzilla.kernel.org/attachment.cgi?id=305518
Signed-off-by: Sandipan Das <[email protected]>
---
arch/x86/events/amd/lbr.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/arch/x86/events/amd/lbr.c b/arch/x86/events/amd/lbr.c
index 110e34c59643..43bf2dbcdb82 100644
--- a/arch/x86/events/amd/lbr.c
+++ b/arch/x86/events/amd/lbr.c
@@ -173,9 +173,11 @@ void amd_pmu_lbr_read(void)
/*
* Check if a branch has been logged; if valid = 0, spec = 0
- * then no branch was recorded
+ * then no branch was recorded; if reserved = 1 then an
+ * erroneous branch was recorded (see erratum 1452)
*/
- if (!entry.to.split.valid && !entry.to.split.spec)
+ if ((!entry.to.split.valid && !entry.to.split.spec) ||
+ entry.to.split.reserved)
continue;
perf_clear_branch_entry_bitfields(br + out);
--
2.34.1
When bringing a CPU online, some of the PMC and LBR related registers
are reset. The same is done when a CPU is taken offline although that
is unnecessary. This currently happens in the "cpu_dead" callback which
is also incorrect as the callback runs on a control CPU instead of the
one that is being taken offline. This also affects hibernation and
suspend to RAM on some platforms as reported in the link below.
Link: https://lore.kernel.org/all/[email protected]/
Reported-by: Mario Limonciello <[email protected]>
Fixes: 21d59e3e2c40 ("perf/x86/amd/core: Detect PerfMonV2 support")
Signed-off-by: Sandipan Das <[email protected]>
---
arch/x86/events/amd/core.c | 1 -
1 file changed, 1 deletion(-)
diff --git a/arch/x86/events/amd/core.c b/arch/x86/events/amd/core.c
index ffdfaee08b08..63514c311f44 100644
--- a/arch/x86/events/amd/core.c
+++ b/arch/x86/events/amd/core.c
@@ -604,7 +604,6 @@ static void amd_pmu_cpu_dead(int cpu)
kfree(cpuhw->lbr_sel);
cpuhw->lbr_sel = NULL;
- amd_pmu_cpu_reset(cpu);
if (!x86_pmu.amd_nb_constraints)
return;
--
2.34.1
* Sandipan Das <[email protected]> wrote:
> Currently, it is assumed that LBR Freeze is supported on all processors
> which have CPUID leaf 0x80000022[EAX] bit 1 set. This is incorrect as
That's X86_FEATURE_AMD_LBR_V2, right? Should probably be mentioned in the
changelog.
> the feature availability is additionally dependent on CPUID leaf
> 0x80000022[EAX] bit 2 being set which may not be set for all Zen 4
> processors. Define a new feature bit for LBR and PMC freeze and set the
> freeze enable bit (FLBRI) in DebugCtl (MSR 0x1d9) conditionally.
What happens on such Zen 4 CPUs that don't support LBR Freeze? Does the CPU
just ignore it, or something worse?
> It should still be possible to use LBR without freeze for profile-guided
> optimization of user programs by using an user-only branch filter during
> profiling. When the user-only filter is enabled, branches are no longer
> recorded after the transition to CPL 0 upon PMI arrival. When branch
> entries are read in the PMI handler, the branch stack does not change.
>
> E.g.
>
> $ perf record -j any,u -e ex_ret_brn_tkn ./workload
>
> Since the feature bit is visible under flags in /proc/cpuinfo, it can be
> used to determine the feasibility of use-cases which require LBR Freeze
> to be supported by the hardware such as profile-guided optimization of
> kernels.
Sounds good to me.
> diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
> index 4af140cf5719..e47ea31b019d 100644
> --- a/arch/x86/include/asm/cpufeatures.h
> +++ b/arch/x86/include/asm/cpufeatures.h
> @@ -97,7 +97,7 @@
> #define X86_FEATURE_SYSENTER32 ( 3*32+15) /* "" sysenter in IA32 userspace */
> #define X86_FEATURE_REP_GOOD ( 3*32+16) /* REP microcode works well */
> #define X86_FEATURE_AMD_LBR_V2 ( 3*32+17) /* AMD Last Branch Record Extension Version 2 */
> -/* FREE, was #define X86_FEATURE_LFENCE_RDTSC ( 3*32+18) "" LFENCE synchronizes RDTSC */
> +#define X86_FEATURE_AMD_LBR_PMC_FREEZE ( 3*32+18) /* AMD LBR and PMC Freeze */
> #define X86_FEATURE_ACC_POWER ( 3*32+19) /* AMD Accumulated Power Mechanism */
> #define X86_FEATURE_NOPL ( 3*32+20) /* The NOPL (0F 1F) instructions */
> #define X86_FEATURE_ALWAYS ( 3*32+21) /* "" Always-present feature */
Could you please port this to the latest upstream kernel? The 3*32+18 slot
is now used for another purpose, and we need to define a new synthethic
CPUID word, word 21 if I'm counting it right.
Don't forget to increase NCAPINTS from 21 to 22, and consider the fixed
asserts in the x86_bug_flags[] definitions in <asm/cpufeature.h>, and the
asserts in <asm/disabled-features.h> and <asm/required-features.h>. This
new word should probably be added in a separate preparatory patch.
Thanks,
Ingo
On 3/13/2024 3:45 PM, Ingo Molnar wrote:
>
> * Sandipan Das <[email protected]> wrote:
>
>> Currently, it is assumed that LBR Freeze is supported on all processors
>> which have CPUID leaf 0x80000022[EAX] bit 1 set. This is incorrect as
>
> That's X86_FEATURE_AMD_LBR_V2, right? Should probably be mentioned in the
> changelog.
>
Yes. I'll add it to the changelog.
>> the feature availability is additionally dependent on CPUID leaf
>> 0x80000022[EAX] bit 2 being set which may not be set for all Zen 4
>> processors. Define a new feature bit for LBR and PMC freeze and set the
>> freeze enable bit (FLBRI) in DebugCtl (MSR 0x1d9) conditionally.
>
> What happens on such Zen 4 CPUs that don't support LBR Freeze? Does the CPU
> just ignore it, or something worse?
>
In this case, LBR ignores PMC overflows and the branch records keep getting
updated continuously as execution progresses.
>> It should still be possible to use LBR without freeze for profile-guided
>> optimization of user programs by using an user-only branch filter during
>> profiling. When the user-only filter is enabled, branches are no longer
>> recorded after the transition to CPL 0 upon PMI arrival. When branch
>> entries are read in the PMI handler, the branch stack does not change.
>>
>> E.g.
>>
>> $ perf record -j any,u -e ex_ret_brn_tkn ./workload
>>
>> Since the feature bit is visible under flags in /proc/cpuinfo, it can be
>> used to determine the feasibility of use-cases which require LBR Freeze
>> to be supported by the hardware such as profile-guided optimization of
>> kernels.
>
> Sounds good to me.
>
>> diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
>> index 4af140cf5719..e47ea31b019d 100644
>> --- a/arch/x86/include/asm/cpufeatures.h
>> +++ b/arch/x86/include/asm/cpufeatures.h
>> @@ -97,7 +97,7 @@
>> #define X86_FEATURE_SYSENTER32 ( 3*32+15) /* "" sysenter in IA32 userspace */
>> #define X86_FEATURE_REP_GOOD ( 3*32+16) /* REP microcode works well */
>> #define X86_FEATURE_AMD_LBR_V2 ( 3*32+17) /* AMD Last Branch Record Extension Version 2 */
>> -/* FREE, was #define X86_FEATURE_LFENCE_RDTSC ( 3*32+18) "" LFENCE synchronizes RDTSC */
>> +#define X86_FEATURE_AMD_LBR_PMC_FREEZE ( 3*32+18) /* AMD LBR and PMC Freeze */
>> #define X86_FEATURE_ACC_POWER ( 3*32+19) /* AMD Accumulated Power Mechanism */
>> #define X86_FEATURE_NOPL ( 3*32+20) /* The NOPL (0F 1F) instructions */
>> #define X86_FEATURE_ALWAYS ( 3*32+21) /* "" Always-present feature */
>
> Could you please port this to the latest upstream kernel? The 3*32+18 slot
> is now used for another purpose, and we need to define a new synthethic
> CPUID word, word 21 if I'm counting it right.
>
> Don't forget to increase NCAPINTS from 21 to 22, and consider the fixed
> asserts in the x86_bug_flags[] definitions in <asm/cpufeature.h>, and the
> asserts in <asm/disabled-features.h> and <asm/required-features.h>. This
> new word should probably be added in a separate preparatory patch.
>
Sure. I'll rebase and send this along with the other changes.
Thanks for the review.