tools/power turbostat: Fix RAPL summary collection on AMD processors
PATCH v4 1/1 tools/power turbostat: Fix RAPL summary collection on AMD processors
Turbostat fails to correctly collect and display RAPL summary information on Family 17h and 19h AMD processors. Running turbostat on these processors returns immediately. If turbostat is working correctly then RAPL summary data is displayed until the user provided command completes. If a command is not provided by the user then turbostat is designed to continuously display RAPL information until interrupted.
The issue is due to offset_to_idx() and idx_to_offset() missing support for AMD MSR addresses/offsets. offset_to_idx()'s switch statement is missing cases for AMD MSRs and idx_to_offset() does not include a path to return AMD MSR(s) for any idx.
The solution is add AMD MSR support to offset_to_idx() and idx_to_offset(). These functions are split-out and renamed along architecture vendor lines for supporting both AMD and Intel MSRs.
--- tools/power/x86/turbostat/turbostat.c | 61 ++++++++++++++++++++++++--- 1 file changed, 56 insertions(+), 5 deletions(-)
tools/power/x86/turbostat/turbostat.c +56 -5
@@ -291,7 +291,7 @@ struct msr_sum_array {/* The percpu MSR sum array.*/ struct msr_sum_array *per_cpu_msr_sum; -int idx_to_offset (int idx) +int idx_to_offset_intel (int idx) {int offset;
@@ -320,7 +320,7 @@ int idx_to_offset(int idx)return offset; } -int offset_to_idx (int offset) +int offset_to_idx_intel (int offset) {int idx;
@@ -349,7 +349,7 @@ int offset_to_idx(int offset)return idx; } -int idx_valid (int idx) +int idx_valid_intel (int idx) {switch (idx) {case IDX_PKG_ENERGY :
@@ -368,6 +368,51 @@ int idx_valid(int idx)return 0 ; } } + +int (*idx_to_offset)(int idx) = idx_to_offset_intel; +int (*offset_to_idx)(int offset) = offset_to_idx_intel; +int (*idx_valid)(int idx) = idx_valid_intel; + +int idx_to_offset_amd (int idx) +{ +int offset; + +switch (idx) { +case IDX_PKG_ENERGY : + offset =MSR_PKG_ENERGY_STAT ; +break ; +default : + offset =-1 ; + } + +return offset; +} + +int offset_to_idx_amd (int offset) +{ +int idx; + +switch (offset) { +case MSR_PKG_ENERGY_STAT : + idx =IDX_PKG_ENERGY ; +break ; +default : + idx =-1 ; + } + +return idx; +} + +int idx_valid_amd (int idx) +{ +switch (idx) { +case IDX_PKG_ENERGY : +return do_rapl &MSR_PKG_ENERGY_STAT ; +default : +return 0 ; + } +} +struct sys_counters {unsigned int added_thread_counters;unsigned int added_core_counters;
@@ -3249,7 +3294,7 @@ int get_msr_sum(int cpu, off_t offset, unsigned long long *msr)return 1 ; idx =offset_to_idx (offset); -if (idx <0 ) +if (idx ==-1 )return idx;/* get_msr_sum() = sum + (get_msr() - last) */ ret =get_msr (cpu, offset, &msr_cur);
@@ -3277,7 +3322,7 @@ static int update_msr_sum(struct thread_data *t, struct core_data *c, struct pkgif (!idx_valid (i))continue ; offset =idx_to_offset (i); -if (offset <0 ) +if (offset ==-1 )continue ; ret =get_msr (cpu, offset, &msr_cur);if (ret) {
@@ -5348,6 +5393,12 @@ void process_cpuid()if (!quiet)decode_misc_feature_control (); +if (authentic_amd || hygon_genuine) { + idx_to_offset = idx_to_offset_amd; + offset_to_idx = offset_to_idx_amd; + idx_valid = idx_valid_amd; + } +return ; }
Signature · 1 line
2.25.1
tools/power/x86/turbostat/turbostat.c int idx_valid(int idx) · line 410
Quoted 6 lines · Terry Bowman
tools/power/x86/turbostat/turbostat.c +5
-0+int idx_valid_amd (int idx) +{ + switch (idx) { + case IDX_PKG_ENERGY: + return do_rapl & MSR_PKG_ENERGY_STAT;
This isn't correct - MSR_PKG_ENERGY_STAT is the MSR offset, not a bit mask for the do_rapl bit field.
The presence of MSR_PKG_ENERGY_STAT (along with MSR_RAPL_PWR_UNIT and MSR_CORE_ENERGY_STAT) is indicated by the RAPL_AMD_F17H bit in do_rapl, and can be checked with:
do_rapl & RAPL_AMD_F17H
Signature · 1 line
Calvin Walton <[email protected]>
Hi Calvin,
Thanks for the feedback. I'll begin making the change and testing. I'll respond with V2 patch in this thread.
Regards,
Terry
Quoted 15 lines · Calvin Walton
tools/power/x86/turbostat/turbostat.c int idx_valid(int idx) · line 410
Quoted 5 lines · Terry Bowman
tools/power/x86/turbostat/turbostat.c +5
-0+int idx_valid_amd (int idx) +{ + switch (idx) { + case IDX_PKG_ENERGY: + return do_rapl & MSR_PKG_ENERGY_STAT;This isn't correct -
MSR_PKG_ENERGY_STATis the MSR offset, not a bit mask for thedo_raplbit field.The presence of
MSR_PKG_ENERGY_STAT(along withMSR_RAPL_PWR_UNITandMSR_CORE_ENERGY_STAT) is indicated by theRAPL_AMD_F17Hbit indo_rapl, and can be checked with:do_rapl & RAPL_AMD_F17H
Quoted 9 lines · Terry Bowman
Hi Calvin,
Thanks for the feedback. I'll begin making the change and testing. I'll
respond with V2 patch in this thread.Regards,
Terry
It looks like there might already be a patch in the pipeline for this issue; see Chen Yu's response to my patch here:
https://lkml.org/lkml/2021/4/14/1322
I'm hoping we get some clarification of the status soon.
While you're respinning your patch, there's one other issue that I noticed - all the idx/offset-related functions pass the offset value in a variable of type "int" (32bit signed integer), but the offset of the AMD MSR_CORE_ENERGY_STAT MSR is 0xC001029A, which exceeds INT_MAX. The offsets should all use "off_t" to get a 64bit type and avoid wrapping or sign extension issues.
Signature · 1 line
Calvin Walton <[email protected]>