perf pmu intel: Generalize SNC cpumask adjustment for multiple platforms
PATCH v6 1/2 perf pmu intel: Generalize SNC cpumask adjustment for multiple platforms
Prepare for supporting more Intel platforms with sub-NUMA clustering by generalizing the GNR specific logic.
---
v6:
Make string literal.
Add SPR into SNC2.
v5: lore.kernel.org/[email protected]/
Split patch.
v4: lore.kernel.org/[email protected]
Rebase.
v3: lore.kernel.org/[email protected]
Fix a typo.
v2: lore.kernel.org/[email protected]
Split EMR and GNR in the SNC2 IMC cpu map.
v1: lore.kernel.org/[email protected]
tools/perf/arch/x86/util/pmu.c | 46 +++++++++++++++++++++------------- 1 file changed, 29 insertions(+), 17 deletions(-)
tools/perf/arch/x86/util/pmu.c +29 -17
@@ -22,20 +22,30 @@#include "util/env.h" #include "util/header.h" -static bool x86__is_intel_graniterapids (void ) +#define GENUINE_INTEL_GNR "GenuineIntel-6-A[DE]" + +static bool x86__is_snc_supported (void ) { -static bool checked_if_graniterapids; -static bool is_graniterapids; +static bool checked_if_snc_supported; +static bool is_supported; + +if (!checked_if_snc_supported) { -if (!checked_if_graniterapids) { -const char *graniterapids_cpuid ="GenuineIntel-6-A[DE]" ; +/* Graniterapids supports SNC configuration. */ +static const char *const supported_cpuids[] = { +GENUINE_INTEL_GNR ,/* Graniterapids */ + };char *cpuid =get_cpuid_str ((struct perf_cpu ){0 }); - is_graniterapids = cpuid &&strcmp_cpuid_str (graniterapids_cpuid, cpuid) ==0 ; +for (size_t i =0 ; i <ARRAY_SIZE (supported_cpuids); i++) { + is_supported = cpuid &&strcmp_cpuid_str (supported_cpuids[i], cpuid) ==0 ; +if (is_supported) +break ; + }free (cpuid); - checked_if_graniterapids = true; + checked_if_snc_supported = true; } -return is_graniterapids; +return is_supported; }static struct perf_cpu_map *read_sysfs_cpu_map (const char *sysfs_path)
@@ -132,8 +142,8 @@ static int uncore_imc_snc(struct perf_pmu *pmu)// Compute the IMC SNC using lookup tables. unsigned int imc_num;int snc_nodes =snc_nodes_per_l3_cache (); -const u8 snc2_map[] = {1 ,1 ,0 ,0 ,1 ,1 ,0 ,0 }; -const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 ,1 ,1 ,0 ,0 ,2 ,2 }; +const u8 snc2_map[] = {1 ,1 ,0 ,0 }; +const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 };const u8 *snc_map;size_t snc_map_len;
@@ -156,11 +166,12 @@ static int uncore_imc_snc(struct perf_pmu *pmu)pr_warning ("Unexpected: unable to compute IMC number '%s'\n" , pmu->name);return 0 ; } -if (imc_num >= snc_map_len) { +if (imc_num >= snc_map_len *perf_cpu_map__nr (pmu->cpus)) {pr_warning ("Unexpected IMC %d for SNC%d mapping\n" , imc_num, snc_nodes);return 0 ; } -return snc_map[imc_num]; + +return snc_map[imc_num % snc_map_len]; }static int uncore_cha_imc_compute_cpu_adjust (int pmu_snc)
@@ -200,7 +211,7 @@ static int uncore_cha_imc_compute_cpu_adjust(int pmu_snc)return cpu_adjust[pmu_snc]; } -static void gnr_uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) +static void uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) {// With sub-NUMA clustering (SNC) there is a NUMA node per SNC in the // topology. For example, a two socket graniterapids machine may be set
@@ -300,11 +311,12 @@ void perf_pmu__arch_init(struct perf_pmu *pmu) pmu->mem_events = perf_mem_events_intel_aux; else pmu->mem_events = perf_mem_events_intel; - }else if (x86__is_intel_graniterapids ()) { + }else if (x86__is_snc_supported ()) {if (strstarts (pmu->name,"uncore_cha_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); -else if (strstarts (pmu->name,"uncore_imc_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); +else if (strstarts (pmu->name,"uncore_imc_" ) && + !strstarts (pmu->name,"uncore_imc_free_running" )) +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); } } }
--
2.54.0.669.g59709faab0-goog
PATCH v6 2/2 perf pmu intel: Adjust cpumasks for sub-NUMA clusters on Sapphire Rapids and Emerald Rapids
Similar to GNR [1], Sapphire Rapids and Emerald Rapids support sub-NUMA clusters as well. Adjust cpumasks using the same logic as GNR in [1].
Tested on Emerald Rapids with SNC2 enabled:
$ perf stat --per-node -e 'UNC_CHA_CLOCKTICKS,UNC_M_CLOCKTICKS' -a -- sleep 1
Performance counter stats for 'system wide':
N0 30 72125876670 UNC_CHA_CLOCKTICKS N0 4 8815163648 UNC_M_CLOCKTICKS N1 30 72124958844 UNC_CHA_CLOCKTICKS N1 4 8815014974 UNC_M_CLOCKTICKS N2 30 72121049022 UNC_CHA_CLOCKTICKS N2 4 8814592626 UNC_M_CLOCKTICKS N3 30 72117133854 UNC_CHA_CLOCKTICKS N3 4 8814012840 UNC_M_CLOCKTICKS
1.001574118 seconds time elapsed
[1] lore.kernel.org/[email protected]
--- tools/perf/arch/x86/util/pmu.c | 60 ++++++++++++++++++++++++---------- 1 file changed, 42 insertions(+), 18 deletions(-)
tools/perf/arch/x86/util/pmu.c +42 -18
@@ -22,6 +22,8 @@#include "util/env.h" #include "util/header.h" +#define GENUINE_INTEL_SPR "GenuineIntel-6-8F" +#define GENUINE_INTEL_EMR "GenuineIntel-6-CF"#define GENUINE_INTEL_GNR "GenuineIntel-6-A[DE]"static bool x86__is_snc_supported (void )
@@ -31,8 +33,10 @@ static bool x86__is_snc_supported(void)if (!checked_if_snc_supported) { -/* Graniterapids supports SNC configuration. */ +/* Sapphirerapids Emeraldrapids Graniterapids support SNC configuration. */ static const char *const supported_cpuids[] = { +GENUINE_INTEL_SPR ,/* Sapphirerapids */ +GENUINE_INTEL_EMR ,/* Emeraldrapids */ GENUINE_INTEL_GNR ,/* Graniterapids */ };char *cpuid =get_cpuid_str ((struct perf_cpu ){0 });
@@ -142,23 +146,43 @@ static int uncore_imc_snc(struct perf_pmu *pmu)// Compute the IMC SNC using lookup tables. unsigned int imc_num;int snc_nodes =snc_nodes_per_l3_cache (); -const u8 snc2_map[] = {1 ,1 ,0 ,0 }; -const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 }; -const u8 *snc_map; -size_t snc_map_len; - -switch (snc_nodes) { -case 2 : - snc_map = snc2_map; - snc_map_len =ARRAY_SIZE (snc2_map); -break ; -case 3 : - snc_map = snc3_map; - snc_map_len =ARRAY_SIZE (snc3_map); -break ; -default : -/* Error or no lookup support for SNC with >3 nodes. */ -return 0 ; +char *cpuid; +static const u8 spr_emr_snc2_map[] = {0 ,0 ,1 ,1 }; +static const u8 gnr_snc2_map[] = {1 ,1 ,0 ,0 }; +static const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 }; +static const u8 *snc_map; +static size_t snc_map_len; + +/* snc_map is not inited yet. We only look up once to avoid expensive operations. */ +if (!snc_map) { +switch (snc_nodes) { +case 2 : + cpuid =get_cpuid_str ((struct perf_cpu ){0 }); +if (cpuid) { +if (strcmp_cpuid_str (GENUINE_INTEL_SPR , cpuid) ==0 || +strcmp_cpuid_str (GENUINE_INTEL_EMR , cpuid) ==0 ) { + snc_map = spr_emr_snc2_map; + snc_map_len =ARRAY_SIZE (spr_emr_snc2_map); + }else if (strcmp_cpuid_str (GENUINE_INTEL_GNR , cpuid) ==0 ) { + snc_map = gnr_snc2_map; + snc_map_len =ARRAY_SIZE (gnr_snc2_map); + } +free (cpuid); + } +break ; +case 3 : + snc_map = snc3_map; + snc_map_len =ARRAY_SIZE (snc3_map); +break ; +default : +/* Error or no lookup support for SNC with >3 nodes. */ +return 0 ; + } + +if (!snc_map) { +pr_warning ("Unexpected: can not find snc map config" ); +return 0 ; + } }/* Compute SNC for PMU. */
--
2.54.0.669.g59709faab0-goog
[email protected]
AI reviewRe: [PATCH v6 2/2] perf pmu intel: Adjust cpumasks for sub-NUMA clusters on Sapphire Rapids and Emerald Rapids
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] Missing IMC cpumask adjustment and silent failure for SNC4 configurations on Sapphire Rapids. --
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Missing IMC cpumask adjustment and silent failure for SNC4 configurations on Sapphire Rapids. --
commit 312ea8a1894ae6a4d9318072afccaf0abbbb29bd
Author: Chun-Tse Shao <[email protected]>
perf pmu intel: Adjust cpumasks for sub-NUMA clusters on Sapphire Rapids and Emerald Rapids
This commit adjusts the cpumasks for sub-NUMA clusters on Sapphire Rapids and Emerald Rapids processors. It updates the detection logic and the mapping arrays to support SNC configurations for these architectures.
Quoted 4 lines
tools/perf/arch/x86/util/pmu.c +0
-0
[ ... ]
Quoted 4 lines
tools/perf/arch/x86/util/pmu.c +0
-0@@ -142,23 +146,43 @@ static int uncore_imc_snc(struct perf_pmu *pmu)// Compute the IMC SNC using lookup tables. unsigned int imc_num;int snc_nodes =snc_nodes_per_l3_cache ();
[ ... ]
Quoted 4 lines
tools/perf/arch/x86/util/pmu.c +4
-0+/* snc_map is not inited yet. We only look up once to avoid expensive operations. */ +if (!snc_map) { +switch (snc_nodes) { +case 2 :
[ ... ]
tools/perf/arch/x86/util/pmu.c static int uncore_imc_snc(struct perf_pmu *pmu) · line 179
Quoted 8 lines
tools/perf/arch/x86/util/pmu.c +8
-0+break ; +case 3 : + snc_map = snc3_map; + snc_map_len =ARRAY_SIZE (snc3_map); +break ; +default : +/* Error or no lookup support for SNC with >3 nodes. */ +return 0 ;
Sapphire Rapids natively supports 4-way sub-NUMA clustering, which would make snc_nodes evaluate to 4. Does this regression silently skip IMC cpumask adjustments for SNC4 configurations?
Because this default case returns 0 immediately, it entirely bypasses the newly added warning block below.
tools/perf/arch/x86/util/pmu.c static int uncore_imc_snc(struct perf_pmu *pmu) · line 185
Quoted 6 lines
tools/perf/arch/x86/util/pmu.c +6
-0+ } + +if (!snc_map) { +pr_warning ("Unexpected: can not find snc map config" ); +return 0 ; + }
Could this leave users unaware that the IMC PMUs are unmapped, even though CHA PMUs are correctly adjusted for SNC4 elsewhere?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider: - [High] The bounds check for `imc_num` incorrectly limits the maximum IMC number based on the number of…
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] The bounds check for imc_num incorrectly limits the maximum IMC number based on the number of active sockets, causing valid IMC PMUs to be rejected and leaving their cpumasks unadjusted on single-socket systems.
- [High] A divide-by-zero crash can occur when computing cha_snc due to an unhandled edge case where the number of discovered CHA PMUs is less than the calculated snc_nodes.
--
commit c91c5bd96c1936822e95d0f574b8f633d5afb4c1
Author: Chun-Tse Shao <[email protected]>
perf pmu intel: Generalize SNC cpumask adjustment for multiple platforms
Prepare for supporting more Intel platforms with sub-NUMA clustering by generalizing the GNR specific logic.
Quoted 4 lines
tools/perf/arch/x86/util/pmu.c +0
-0
[ ... ]
tools/perf/arch/x86/util/pmu.c static int uncore_imc_snc(struct perf_pmu *pmu) · line 169
Quoted 6 lines
tools/perf/arch/x86/util/pmu.c +1
-1@@ -156,11 +166,12 @@ static int uncore_imc_snc(struct perf_pmu *pmu)pr_warning ("Unexpected: unable to compute IMC number '%s'\n" , pmu->name);return 0 ; } -if (imc_num >= snc_map_len) { +if (imc_num >= snc_map_len *perf_cpu_map__nr (pmu->cpus)) {
Will this condition incorrectly limit the maximum IMC number based on the number of active sockets?
The expression perf_cpu_map__nr(pmu->cpus) represents the number of active sockets, meaning it will evaluate to 1 on a single-socket system. In an SNC3 configuration where snc_map_len is 6, this would limit imc_num to less than 6.
Could this cause single-socket systems to wrongly reject imc_num >= 6, given that the hardware might still expose 12 IMC blocks on that socket? If so, processing uncore_imc_6 through uncore_imc_11 would trigger the "Unexpected IMC" warning and skip the necessary cpumask adjustments.
(Note that subsequent patches in the series, such as "perf pmu intel:
Adjust cpumasks for sub-NUMA clusters on Sapphire Rapids and Emerald Rapids", do not appear to resolve this bounds check).
Quoted 3 lines
tools/perf/arch/x86/util/pmu.c +0
-0pr_warning ("Unexpected IMC %d for SNC%d mapping\n" , imc_num, snc_nodes);return 0 ; }
[ ... ]
tools/perf/arch/x86/util/pmu.c void perf_pmu__arch_init(struct perf_pmu *pmu) · line 316
Quoted 11 lines
tools/perf/arch/x86/util/pmu.c +2
-4@@ -300,11 +311,12 @@ void perf_pmu__arch_init(struct perf_pmu *pmu) pmu->mem_events = perf_mem_events_intel_aux; else pmu->mem_events = perf_mem_events_intel; - }else if (x86__is_intel_graniterapids ()) { + }else if (x86__is_snc_supported ()) {if (strstarts (pmu->name,"uncore_cha_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); -else if (strstarts (pmu->name,"uncore_imc_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true);
Is it possible for a divide-by-zero to occur when this generalized adjustment logic invokes uncore_cha_snc()?
If a system or container restricts access to sysfs and exposes fewer CHA devices, or if testing under a fake PMU environment, the dynamically determined num_cha could be less than the calculated snc_nodes.
In uncore_cha_snc(), the code performs:
chas_per_node = num_cha / snc_nodes;
And later evaluates:
cha_snc = cha_num / chas_per_node;
Could the initial integer division result in chas_per_node = 0, causing an unhandled crash when evaluating cha_snc?
Quoted 6 lines
tools/perf/arch/x86/util/pmu.c +3
-0+else if (strstarts (pmu->name,"uncore_imc_" ) && + !strstarts (pmu->name,"uncore_imc_free_running" )) +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); } } }
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Quoted 5 lines · Chun-Tse Shao
Prepare for supporting more Intel platforms with sub-NUMA clustering by generalizing the GNR specific logic.
Signed-off-by Chun-Tse Shao <[email protected]>---
Quoted 120 lines
v6:
Make string literal.
Add SPR into SNC2.v5: lore.kernel.org/[email protected]/
Split patch.v4: lore.kernel.org/[email protected]
Rebase.v3: lore.kernel.org/[email protected]
Fix a typo.v2: lore.kernel.org/[email protected]
Split EMR and GNR in the SNC2 IMC cpu map.v1: lore.kernel.org/[email protected]
tools/perf/arch/x86/util/pmu.c | 46 +++++++++++++++++++++------------- 1 file changed, 29 insertions(+), 17 deletions(-)tools/perf/arch/x86/util/pmu.c +29
-18@@ -22,20 +22,30 @@#include "util/env.h" #include "util/header.h" -static bool x86__is_intel_graniterapids (void ) +#define GENUINE_INTEL_GNR "GenuineIntel-6-A[DE]" + +static bool x86__is_snc_supported (void ) { -static bool checked_if_graniterapids; -static bool is_graniterapids; +static bool checked_if_snc_supported; +static bool is_supported; + +if (!checked_if_snc_supported) { -if (!checked_if_graniterapids) { -const char *graniterapids_cpuid ="GenuineIntel-6-A[DE]" ; +/* Graniterapids supports SNC configuration. */ +static const char *const supported_cpuids[] = { +GENUINE_INTEL_GNR ,/* Graniterapids */ + };char *cpuid =get_cpuid_str ((struct perf_cpu ){0 }); - is_graniterapids = cpuid &&strcmp_cpuid_str (graniterapids_cpuid, cpuid) ==0 ; +for (size_t i =0 ; i <ARRAY_SIZE (supported_cpuids); i++) { + is_supported = cpuid &&strcmp_cpuid_str (supported_cpuids[i], cpuid) ==0 ; +if (is_supported) +break ; + }free (cpuid); - checked_if_graniterapids = true; + checked_if_snc_supported = true; } -return is_graniterapids; +return is_supported; }static struct perf_cpu_map *read_sysfs_cpu_map (const char *sysfs_path)@@ -132,8 +142,8 @@ static int uncore_imc_snc(struct perf_pmu *pmu)// Compute the IMC SNC using lookup tables. unsigned int imc_num;int snc_nodes =snc_nodes_per_l3_cache (); -const u8 snc2_map[] = {1 ,1 ,0 ,0 ,1 ,1 ,0 ,0 }; -const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 ,1 ,1 ,0 ,0 ,2 ,2 }; +const u8 snc2_map[] = {1 ,1 ,0 ,0 }; +const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 };const u8 *snc_map;size_t snc_map_len;@@ -156,11 +166,12 @@ static int uncore_imc_snc(struct perf_pmu *pmu)pr_warning ("Unexpected: unable to compute IMC number '%s'\n" , pmu->name);return 0 ; } -if (imc_num >= snc_map_len) { +if (imc_num >= snc_map_len *perf_cpu_map__nr (pmu->cpus)) {pr_warning ("Unexpected IMC %d for SNC%d mapping\n" , imc_num, snc_nodes);return 0 ; } -return snc_map[imc_num]; + +return snc_map[imc_num % snc_map_len]; }static int uncore_cha_imc_compute_cpu_adjust (int pmu_snc)@@ -200,7 +211,7 @@ static int uncore_cha_imc_compute_cpu_adjust(int pmu_snc)return cpu_adjust[pmu_snc]; } -static void gnr_uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) +static void uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) {// With sub-NUMA clustering (SNC) there is a NUMA node per SNC in the // topology. For example, a two socket graniterapids machine may be set @@ -300,11 +311,12 @@ void perf_pmu__arch_init(struct perf_pmu *pmu) pmu->mem_events = perf_mem_events_intel_aux; else pmu->mem_events = perf_mem_events_intel; - }else if (x86__is_intel_graniterapids ()) { + }else if (x86__is_snc_supported ()) {if (strstarts (pmu->name,"uncore_cha_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); -else if (strstarts (pmu->name,"uncore_imc_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); +else if (strstarts (pmu->name,"uncore_imc_" ) && + !strstarts (pmu->name,"uncore_imc_free_running" )) +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); } } } --2.54.0.669.g59709faab0-goog
PATCH v7 1/2 perf pmu intel: Generalize SNC cpumask adjustment for multiple platforms
In reply to Chun-Tse ShaoPrepare for supporting more Intel platforms with sub-NUMA clustering by generalizing the GNR specific logic.
---
v7:
Fixed based on Sashiko review:
- Refactor PMU initialization to be fully thread-safe using pthread_once and pthread_mutex locks.
- Fix a critical bounds check bug on single-socket systems.
- Fix potential divide-by-zero crash in uncore_cha_snc.
- Avoid spurious warnings on non-SNC systems.
- Avoid silent and inconsistent uncore mappings on unsupported SNC configurations.
- Resolve checkpatch linter warnings.
v6: lore.kernel.org/[email protected]
Make string literal.
Add SPR into SNC2.
v5: lore.kernel.org/[email protected]
Split patch.
v4: lore.kernel.org/[email protected]
Rebase.
v3: lore.kernel.org/[email protected]
Fix a typo.
v2: lore.kernel.org/[email protected]
Split EMR and GNR in the SNC2 IMC cpu map.
v1: lore.kernel.org/[email protected]
tools/perf/arch/x86/util/pmu.c | 199 +++++++++++++++++++++------------ 1 file changed, 130 insertions(+), 69 deletions(-)
tools/perf/arch/x86/util/pmu.c +130 -69
@@ -1,4 +1,5 @@// SPDX-License-Identifier: GPL-2.0 +#include <pthread.h> #include <string.h> #include <stdio.h> #include <sys/types.h>
@@ -22,20 +23,31 @@#include "util/env.h" #include "util/header.h" -static bool x86__is_intel_graniterapids (void ) -{ -static bool checked_if_graniterapids; -static bool is_graniterapids; +#define GENUINE_INTEL_GNR "GenuineIntel-6-A[DE]" -if (!checked_if_graniterapids) { -const char *graniterapids_cpuid ="GenuineIntel-6-A[DE]" ; -char *cpuid =get_cpuid_str ((struct perf_cpu ){0 }); +static bool cached_snc_supported; +static pthread_once_t snc_support_once =PTHREAD_ONCE_INIT ; - is_graniterapids = cpuid &&strcmp_cpuid_str (graniterapids_cpuid, cpuid) ==0 ; -free (cpuid); - checked_if_graniterapids = true; +static void init_snc_support (void ) +{ +/* Graniterapids supports SNC configuration. */ +static const char *const supported_cpuids[] = { +GENUINE_INTEL_GNR ,/* Graniterapids */ + }; +char *cpuid =get_cpuid_str ((struct perf_cpu ){0 }); + +for (size_t i =0 ; i <ARRAY_SIZE (supported_cpuids); i++) { + cached_snc_supported = cpuid &&strcmp_cpuid_str (supported_cpuids[i], cpuid) ==0 ; +if (cached_snc_supported) +break ; } -return is_graniterapids; +free (cpuid); +} + +static bool x86__is_snc_supported (void ) +{ +pthread_once (&snc_support_once, init_snc_support); +return cached_snc_supported; }static struct perf_cpu_map *read_sysfs_cpu_map (const char *sysfs_path)
@@ -52,49 +64,58 @@ static struct perf_cpu_map *read_sysfs_cpu_map(const char *sysfs_path)return cpus; } -static int snc_nodes_per_l3_cache (void ) +static int cached_snc_nodes; +static pthread_once_t snc_nodes_once =PTHREAD_ONCE_INIT ; + +static void init_snc_nodes (void ) { -static bool checked_snc; -static int snc_nodes; - -if (!checked_snc) { -struct perf_cpu_map *node_cpus = -read_sysfs_cpu_map ("devices/system/node/node0/cpulist" ); -struct perf_cpu_map *cache_cpus = -read_sysfs_cpu_map ("devices/system/cpu/cpu0/cache/index3/shared_cpu_list" ); - - snc_nodes =perf_cpu_map__nr (cache_cpus) /perf_cpu_map__nr (node_cpus); -perf_cpu_map__put (cache_cpus); -perf_cpu_map__put (node_cpus); - checked_snc = true; - } -return snc_nodes; +struct perf_cpu_map *node_cpus = +read_sysfs_cpu_map ("devices/system/node/node0/cpulist" ); +struct perf_cpu_map *cache_cpus = +read_sysfs_cpu_map ("devices/system/cpu/cpu0/cache/index3/shared_cpu_list" ); + +if (node_cpus && cache_cpus) + cached_snc_nodes =perf_cpu_map__nr (cache_cpus) /perf_cpu_map__nr (node_cpus); +else + cached_snc_nodes =0 ; +perf_cpu_map__put (cache_cpus); +perf_cpu_map__put (node_cpus); } -static int num_chas (void ) +static int snc_nodes_per_l3_cache (void ) { -static bool checked_chas; -static int num_chas; +pthread_once (&snc_nodes_once, init_snc_nodes); +return cached_snc_nodes; +} -if (!checked_chas) { -int fd =perf_pmu__event_source_devices_fd (); -struct io_dir dir; -struct io_dirent64 *dent; +static int cached_num_chas; +static pthread_once_t num_chas_once =PTHREAD_ONCE_INIT ; -if (fd <0 ) -return -1 ; +static void init_num_chas (void ) +{ +int fd =perf_pmu__event_source_devices_fd (); +struct io_dir dir; +struct io_dirent64 *dent; -io_dir__init (&dir, fd); +if (fd <0 ) { + cached_num_chas =-1 ; +return ; + } -while ((dent =io_dir__readdir (&dir)) !=NULL ) { -/* Note, dent->d_type will be DT_LNK and so isn't a useful filter. */ -if (strstarts (dent->d_name,"uncore_cha_" )) - num_chas++; - } -close (fd); - checked_chas = true; +io_dir__init (&dir, fd); + +while ((dent =io_dir__readdir (&dir)) !=NULL ) { +/* Note, dent->d_type will be DT_LNK and so isn't a useful filter. */ +if (strstarts (dent->d_name,"uncore_cha_" )) + cached_num_chas++; } -return num_chas; +close (fd); +} + +static int num_chas (void ) +{ +pthread_once (&num_chas_once, init_num_chas); +return cached_num_chas; }#define MAX_SNCS 6
@@ -121,46 +142,73 @@ static int uncore_cha_snc(struct perf_pmu *pmu)return 0 ; } chas_per_node = num_cha / snc_nodes; +if (chas_per_node ==0 ) { +pr_warning ("Unexpected: chas_per_node is 0 (num_cha=%d, snc_nodes=%d)\n" , + num_cha, snc_nodes); +return 0 ; + } cha_snc = cha_num / chas_per_node;/* Range check cha_snc. for unexpected out of bounds. */ return cha_snc >=MAX_SNCS ?0 : cha_snc; } -static int uncore_imc_snc (struct perf_pmu *pmu) +static const u8 *cached_imc_snc_map; +static size_t cached_imc_snc_map_len; +static pthread_once_t imc_snc_map_once =PTHREAD_ONCE_INIT ; + +static void init_snc_map (void ) { -// Compute the IMC SNC using lookup tables. -unsigned int imc_num;int snc_nodes =snc_nodes_per_l3_cache (); -const u8 snc2_map[] = {1 ,1 ,0 ,0 ,1 ,1 ,0 ,0 }; -const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 ,1 ,1 ,0 ,0 ,2 ,2 }; -const u8 *snc_map; -size_t snc_map_len; +char *cpuid; +static const u8 gnr_snc2_map[] = {1 ,1 ,0 ,0 }; +static const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 };switch (snc_nodes) {case 2 : - snc_map = snc2_map; - snc_map_len =ARRAY_SIZE (snc2_map); + cpuid =get_cpuid_str ((struct perf_cpu ){0 }); +if (cpuid) { +if (strcmp_cpuid_str (GENUINE_INTEL_GNR , cpuid) ==0 ) { + cached_imc_snc_map = gnr_snc2_map; + cached_imc_snc_map_len =ARRAY_SIZE (gnr_snc2_map); + } +free (cpuid); + }break ;case 3 : - snc_map = snc3_map; - snc_map_len =ARRAY_SIZE (snc3_map); + cached_imc_snc_map = snc3_map; + cached_imc_snc_map_len =ARRAY_SIZE (snc3_map);break ;default :/* Error or no lookup support for SNC with >3 nodes. */ -return 0 ; +break ; } +if (!cached_imc_snc_map) +pr_warning ("Unexpected: can not find snc map config\n" ); +} + +static int uncore_imc_snc (struct perf_pmu *pmu) +{ +// Compute the IMC SNC using lookup tables. +unsigned int imc_num; +int snc_nodes =snc_nodes_per_l3_cache (); + +if (snc_nodes <=1 ) +return 0 ; + +pthread_once (&imc_snc_map_once, init_snc_map); +/* Compute SNC for PMU. */ if (sscanf (pmu->name,"uncore_imc_%u" , &imc_num) !=1 ) {pr_warning ("Unexpected: unable to compute IMC number '%s'\n" , pmu->name);return 0 ; } -if (imc_num >= snc_map_len) { -pr_warning ("Unexpected IMC %d for SNC%d mapping\n" , imc_num, snc_nodes); + +if (!cached_imc_snc_map)return 0 ; - } -return snc_map[imc_num]; + +return cached_imc_snc_map[imc_num % cached_imc_snc_map_len]; }static int uncore_cha_imc_compute_cpu_adjust (int pmu_snc)
@@ -200,7 +248,9 @@ static int uncore_cha_imc_compute_cpu_adjust(int pmu_snc)return cpu_adjust[pmu_snc]; } -static void gnr_uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) +static pthread_mutex_t pmu_adjust_mutex =PTHREAD_MUTEX_INITIALIZER ; + +static void uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) {// With sub-NUMA clustering (SNC) there is a NUMA node per SNC in the // topology. For example, a two socket graniterapids machine may be set
@@ -231,9 +281,11 @@ static void gnr_uncore_cha_imc_adjust_cpumask_for_snc(struct perf_pmu *pmu, boolreturn ; } +pthread_mutex_lock (&pmu_adjust_mutex); + pmu_snc = cha ?uncore_cha_snc (pmu) :uncore_imc_snc (pmu);if (pmu_snc ==0 ) { -// No adjustment necessary for the first SNC. +pthread_mutex_unlock (&pmu_adjust_mutex);return ; }
@@ -242,8 +294,10 @@ static void gnr_uncore_cha_imc_adjust_cpumask_for_snc(struct perf_pmu *pmu, bool// Hold onto the perf_cpu_map globally to avoid recomputation. cpu_adjust =uncore_cha_imc_compute_cpu_adjust (pmu_snc); adjusted[pmu_snc] =perf_cpu_map__empty_new (perf_cpu_map__nr (pmu->cpus)); -if (!adjusted[pmu_snc]) +if (!adjusted[pmu_snc]) { +pthread_mutex_unlock (&pmu_adjust_mutex);return ; + } }perf_cpu_map__for_each_cpu (cpu, idx, pmu->cpus) {
@@ -263,6 +317,8 @@ static void gnr_uncore_cha_imc_adjust_cpumask_for_snc(struct perf_pmu *pmu, boolperf_cpu_map__put (pmu->cpus); pmu->cpus =perf_cpu_map__get (adjusted[pmu_snc]); + +pthread_mutex_unlock (&pmu_adjust_mutex); }void perf_pmu__arch_init (struct perf_pmu *pmu)
@@ -300,11 +356,16 @@ void perf_pmu__arch_init(struct perf_pmu *pmu) pmu->mem_events = perf_mem_events_intel_aux; else pmu->mem_events = perf_mem_events_intel; - }else if (x86__is_intel_graniterapids ()) { -if (strstarts (pmu->name,"uncore_cha_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); -else if (strstarts (pmu->name,"uncore_imc_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); + }else if (x86__is_snc_supported ()) { +int snc_nodes =snc_nodes_per_l3_cache (); + +if (snc_nodes ==2 || snc_nodes ==3 ) { +if (strstarts (pmu->name,"uncore_cha_" )) +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); +else if (strstarts (pmu->name,"uncore_imc_" ) && + !strstarts (pmu->name,"uncore_imc_free_running" )) +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); + } } } }
--
2.54.0.823.g6e5bcc1fc9-goog
PATCH v7 2/2 perf pmu intel: Adjust cpumasks for sub-NUMA clusters on Sapphire Rapids and Emerald Rapids
Similar to GNR [1], Sapphire Rapids and Emerald Rapids support sub-NUMA clusters as well. Adjust cpumasks using the same logic as GNR in [1].
Tested on Emerald Rapids with SNC2 enabled:
$ perf stat --per-node -e 'UNC_CHA_CLOCKTICKS,UNC_M_CLOCKTICKS' -a -- sleep 1
Performance counter stats for 'system wide':
N0 30 72125876670 UNC_CHA_CLOCKTICKS N0 4 8815163648 UNC_M_CLOCKTICKS N1 30 72124958844 UNC_CHA_CLOCKTICKS N1 4 8815014974 UNC_M_CLOCKTICKS N2 30 72121049022 UNC_CHA_CLOCKTICKS N2 4 8814592626 UNC_M_CLOCKTICKS N3 30 72117133854 UNC_CHA_CLOCKTICKS N3 4 8814012840 UNC_M_CLOCKTICKS
1.001574118 seconds time elapsed
[1] lore.kernel.org/[email protected]
--- tools/perf/arch/x86/util/pmu.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)
tools/perf/arch/x86/util/pmu.c +11 -2
@@ -23,6 +23,8 @@#include "util/env.h" #include "util/header.h" +#define GENUINE_INTEL_SPR "GenuineIntel-6-8F" +#define GENUINE_INTEL_EMR "GenuineIntel-6-CF"#define GENUINE_INTEL_GNR "GenuineIntel-6-A[DE]"static bool cached_snc_supported;
@@ -30,8 +32,10 @@ static pthread_once_t snc_support_once = PTHREAD_ONCE_INIT;static void init_snc_support (void ) { -/* Graniterapids supports SNC configuration. */ +/* Sapphirerapids Emeraldrapids Graniterapids support SNC configuration. */ static const char *const supported_cpuids[] = { +GENUINE_INTEL_SPR ,/* Sapphirerapids */ +GENUINE_INTEL_EMR ,/* Emeraldrapids */ GENUINE_INTEL_GNR ,/* Graniterapids */ };char *cpuid =get_cpuid_str ((struct perf_cpu ){0 });
@@ -161,6 +165,7 @@ static void init_snc_map(void) {int snc_nodes =snc_nodes_per_l3_cache ();char *cpuid; +static const u8 spr_emr_snc2_map[] = {0 ,0 ,1 ,1 };static const u8 gnr_snc2_map[] = {1 ,1 ,0 ,0 };static const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 };
@@ -168,7 +173,11 @@ static void init_snc_map(void)case 2 : cpuid =get_cpuid_str ((struct perf_cpu ){0 });if (cpuid) { -if (strcmp_cpuid_str (GENUINE_INTEL_GNR , cpuid) ==0 ) { +if (strcmp_cpuid_str (GENUINE_INTEL_SPR , cpuid) ==0 || +strcmp_cpuid_str (GENUINE_INTEL_EMR , cpuid) ==0 ) { + cached_imc_snc_map = spr_emr_snc2_map; + cached_imc_snc_map_len =ARRAY_SIZE (spr_emr_snc2_map); + }else if (strcmp_cpuid_str (GENUINE_INTEL_GNR , cpuid) ==0 ) { cached_imc_snc_map = gnr_snc2_map; cached_imc_snc_map_len =ARRAY_SIZE (gnr_snc2_map); }
--
2.54.0.823.g6e5bcc1fc9-goog
Accidentally in reply to the previous version of patchset, will resend. Sorry for the interruption.
Quoted 346 lines · Chun-Tse Shao
Prepare for supporting more Intel platforms with sub-NUMA clustering by generalizing the GNR specific logic.
Reviewed-by Zide Chen <[email protected]>Signed-off-by Chun-Tse Shao <[email protected]>Assisted-by Gemini:gemini-3.1-pro-preview---
v7:
Fixed based on Sashiko review:
- Refactor PMU initialization to be fully thread-safe usingpthread_onceandpthread_mutexlocks.
- Fix a critical bounds check bug on single-socket systems.
- Fix potential divide-by-zero crash inuncore_cha_snc.
- Avoid spurious warnings on non-SNC systems.
- Avoid silent and inconsistent uncore mappings on unsupported SNC configurations.
- Resolve checkpatch linter warnings.v6: lore.kernel.org/[email protected]
Make string literal.
Add SPR into SNC2.v5: lore.kernel.org/[email protected]
Split patch.v4: lore.kernel.org/[email protected]
Rebase.v3: lore.kernel.org/[email protected]
Fix a typo.v2: lore.kernel.org/[email protected]
Split EMR and GNR in the SNC2 IMC cpu map.v1: lore.kernel.org/[email protected]
tools/perf/arch/x86/util/pmu.c | 199 +++++++++++++++++++++------------ 1 file changed, 130 insertions(+), 69 deletions(-)tools/perf/arch/x86/util/pmu.c +130
-70@@ -1,4 +1,5 @@// SPDX-License-Identifier: GPL-2.0 +#include <pthread.h> #include <string.h> #include <stdio.h> #include <sys/types.h> @@ -22,20 +23,31 @@#include "util/env.h" #include "util/header.h" -static bool x86__is_intel_graniterapids (void ) -{ -static bool checked_if_graniterapids; -static bool is_graniterapids; +#define GENUINE_INTEL_GNR "GenuineIntel-6-A[DE]" -if (!checked_if_graniterapids) { -const char *graniterapids_cpuid ="GenuineIntel-6-A[DE]" ; -char *cpuid =get_cpuid_str ((struct perf_cpu ){0 }); +static bool cached_snc_supported; +static pthread_once_t snc_support_once =PTHREAD_ONCE_INIT ; - is_graniterapids = cpuid &&strcmp_cpuid_str (graniterapids_cpuid, cpuid) ==0 ; -free (cpuid); - checked_if_graniterapids = true; +static void init_snc_support (void ) +{ +/* Graniterapids supports SNC configuration. */ +static const char *const supported_cpuids[] = { +GENUINE_INTEL_GNR ,/* Graniterapids */ + }; +char *cpuid =get_cpuid_str ((struct perf_cpu ){0 }); + +for (size_t i =0 ; i <ARRAY_SIZE (supported_cpuids); i++) { + cached_snc_supported = cpuid &&strcmp_cpuid_str (supported_cpuids[i], cpuid) ==0 ; +if (cached_snc_supported) +break ; } -return is_graniterapids; +free (cpuid); +} + +static bool x86__is_snc_supported (void ) +{ +pthread_once (&snc_support_once, init_snc_support); +return cached_snc_supported; }static struct perf_cpu_map *read_sysfs_cpu_map (const char *sysfs_path)@@ -52,49 +64,58 @@ static struct perf_cpu_map *read_sysfs_cpu_map(const char *sysfs_path)return cpus; } -static int snc_nodes_per_l3_cache (void ) +static int cached_snc_nodes; +static pthread_once_t snc_nodes_once =PTHREAD_ONCE_INIT ; + +static void init_snc_nodes (void ) { -static bool checked_snc; -static int snc_nodes; - -if (!checked_snc) { -struct perf_cpu_map *node_cpus = -read_sysfs_cpu_map ("devices/system/node/node0/cpulist" ); -struct perf_cpu_map *cache_cpus = -read_sysfs_cpu_map ("devices/system/cpu/cpu0/cache/index3/shared_cpu_list" ); - - snc_nodes =perf_cpu_map__nr (cache_cpus) /perf_cpu_map__nr (node_cpus); -perf_cpu_map__put (cache_cpus); -perf_cpu_map__put (node_cpus); - checked_snc = true; - } -return snc_nodes; +struct perf_cpu_map *node_cpus = +read_sysfs_cpu_map ("devices/system/node/node0/cpulist" ); +struct perf_cpu_map *cache_cpus = +read_sysfs_cpu_map ("devices/system/cpu/cpu0/cache/index3/shared_cpu_list" ); + +if (node_cpus && cache_cpus) + cached_snc_nodes =perf_cpu_map__nr (cache_cpus) /perf_cpu_map__nr (node_cpus); +else + cached_snc_nodes =0 ; +perf_cpu_map__put (cache_cpus); +perf_cpu_map__put (node_cpus); } -static int num_chas (void ) +static int snc_nodes_per_l3_cache (void ) { -static bool checked_chas; -static int num_chas; +pthread_once (&snc_nodes_once, init_snc_nodes); +return cached_snc_nodes; +} -if (!checked_chas) { -int fd =perf_pmu__event_source_devices_fd (); -struct io_dir dir; -struct io_dirent64 *dent; +static int cached_num_chas; +static pthread_once_t num_chas_once =PTHREAD_ONCE_INIT ; -if (fd <0 ) -return -1 ; +static void init_num_chas (void ) +{ +int fd =perf_pmu__event_source_devices_fd (); +struct io_dir dir; +struct io_dirent64 *dent; -io_dir__init (&dir, fd); +if (fd <0 ) { + cached_num_chas =-1 ; +return ; + } -while ((dent =io_dir__readdir (&dir)) !=NULL ) { -/* Note, dent->d_type will be DT_LNK and so isn't a useful filter. */ -if (strstarts (dent->d_name,"uncore_cha_" )) - num_chas++; - } -close (fd); - checked_chas = true; +io_dir__init (&dir, fd); + +while ((dent =io_dir__readdir (&dir)) !=NULL ) { +/* Note, dent->d_type will be DT_LNK and so isn't a useful filter. */ +if (strstarts (dent->d_name,"uncore_cha_" )) + cached_num_chas++; } -return num_chas; +close (fd); +} + +static int num_chas (void ) +{ +pthread_once (&num_chas_once, init_num_chas); +return cached_num_chas; }#define MAX_SNCS 6@@ -121,46 +142,73 @@ static int uncore_cha_snc(struct perf_pmu *pmu)return 0 ; } chas_per_node = num_cha / snc_nodes; +if (chas_per_node ==0 ) { +pr_warning ("Unexpected: chas_per_node is 0 (num_cha=%d, snc_nodes=%d)\n" , + num_cha, snc_nodes); +return 0 ; + } cha_snc = cha_num / chas_per_node;/* Range check cha_snc. for unexpected out of bounds. */ return cha_snc >=MAX_SNCS ?0 : cha_snc; } -static int uncore_imc_snc (struct perf_pmu *pmu) +static const u8 *cached_imc_snc_map; +static size_t cached_imc_snc_map_len; +static pthread_once_t imc_snc_map_once =PTHREAD_ONCE_INIT ; + +static void init_snc_map (void ) { -// Compute the IMC SNC using lookup tables. -unsigned int imc_num;int snc_nodes =snc_nodes_per_l3_cache (); -const u8 snc2_map[] = {1 ,1 ,0 ,0 ,1 ,1 ,0 ,0 }; -const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 ,1 ,1 ,0 ,0 ,2 ,2 }; -const u8 *snc_map; -size_t snc_map_len; +char *cpuid; +static const u8 gnr_snc2_map[] = {1 ,1 ,0 ,0 }; +static const u8 snc3_map[] = {1 ,1 ,0 ,0 ,2 ,2 };switch (snc_nodes) {case 2 : - snc_map = snc2_map; - snc_map_len =ARRAY_SIZE (snc2_map); + cpuid =get_cpuid_str ((struct perf_cpu ){0 }); +if (cpuid) { +if (strcmp_cpuid_str (GENUINE_INTEL_GNR , cpuid) ==0 ) { + cached_imc_snc_map = gnr_snc2_map; + cached_imc_snc_map_len =ARRAY_SIZE (gnr_snc2_map); + } +free (cpuid); + }break ;case 3 : - snc_map = snc3_map; - snc_map_len =ARRAY_SIZE (snc3_map); + cached_imc_snc_map = snc3_map; + cached_imc_snc_map_len =ARRAY_SIZE (snc3_map);break ;default :/* Error or no lookup support for SNC with >3 nodes. */ -return 0 ; +break ; } +if (!cached_imc_snc_map) +pr_warning ("Unexpected: can not find snc map config\n" ); +} + +static int uncore_imc_snc (struct perf_pmu *pmu) +{ +// Compute the IMC SNC using lookup tables. +unsigned int imc_num; +int snc_nodes =snc_nodes_per_l3_cache (); + +if (snc_nodes <=1 ) +return 0 ; + +pthread_once (&imc_snc_map_once, init_snc_map); +/* Compute SNC for PMU. */ if (sscanf (pmu->name,"uncore_imc_%u" , &imc_num) !=1 ) {pr_warning ("Unexpected: unable to compute IMC number '%s'\n" , pmu->name);return 0 ; } -if (imc_num >= snc_map_len) { -pr_warning ("Unexpected IMC %d for SNC%d mapping\n" , imc_num, snc_nodes); + +if (!cached_imc_snc_map)return 0 ; - } -return snc_map[imc_num]; + +return cached_imc_snc_map[imc_num % cached_imc_snc_map_len]; }static int uncore_cha_imc_compute_cpu_adjust (int pmu_snc)@@ -200,7 +248,9 @@ static int uncore_cha_imc_compute_cpu_adjust(int pmu_snc)return cpu_adjust[pmu_snc]; } -static void gnr_uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) +static pthread_mutex_t pmu_adjust_mutex =PTHREAD_MUTEX_INITIALIZER ; + +static void uncore_cha_imc_adjust_cpumask_for_snc (struct perf_pmu *pmu,bool cha) {// With sub-NUMA clustering (SNC) there is a NUMA node per SNC in the // topology. For example, a two socket graniterapids machine may be set @@ -231,9 +281,11 @@ static void gnr_uncore_cha_imc_adjust_cpumask_for_snc(struct perf_pmu *pmu, boolreturn ; } +pthread_mutex_lock (&pmu_adjust_mutex); + pmu_snc = cha ?uncore_cha_snc (pmu) :uncore_imc_snc (pmu);if (pmu_snc ==0 ) { -// No adjustment necessary for the first SNC. +pthread_mutex_unlock (&pmu_adjust_mutex);return ; }@@ -242,8 +294,10 @@ static void gnr_uncore_cha_imc_adjust_cpumask_for_snc(struct perf_pmu *pmu, bool// Hold onto the perf_cpu_map globally to avoid recomputation. cpu_adjust =uncore_cha_imc_compute_cpu_adjust (pmu_snc); adjusted[pmu_snc] =perf_cpu_map__empty_new (perf_cpu_map__nr (pmu->cpus)); -if (!adjusted[pmu_snc]) +if (!adjusted[pmu_snc]) { +pthread_mutex_unlock (&pmu_adjust_mutex);return ; + } }perf_cpu_map__for_each_cpu (cpu, idx, pmu->cpus) {@@ -263,6 +317,8 @@ static void gnr_uncore_cha_imc_adjust_cpumask_for_snc(struct perf_pmu *pmu, boolperf_cpu_map__put (pmu->cpus); pmu->cpus =perf_cpu_map__get (adjusted[pmu_snc]); + +pthread_mutex_unlock (&pmu_adjust_mutex); }void perf_pmu__arch_init (struct perf_pmu *pmu)@@ -300,11 +356,16 @@ void perf_pmu__arch_init(struct perf_pmu *pmu) pmu->mem_events = perf_mem_events_intel_aux; else pmu->mem_events = perf_mem_events_intel; - }else if (x86__is_intel_graniterapids ()) { -if (strstarts (pmu->name,"uncore_cha_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); -else if (strstarts (pmu->name,"uncore_imc_" )) -gnr_uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); + }else if (x86__is_snc_supported ()) { +int snc_nodes =snc_nodes_per_l3_cache (); + +if (snc_nodes ==2 || snc_nodes ==3 ) { +if (strstarts (pmu->name,"uncore_cha_" )) +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ true); +else if (strstarts (pmu->name,"uncore_imc_" ) && + !strstarts (pmu->name,"uncore_imc_free_running" )) +uncore_cha_imc_adjust_cpumask_for_snc (pmu,/*cha=*/ false); + } } } } --2.54.0.823.g6e5bcc1fc9-goog