tools/power turbostat: Read Package-cstates via perf
authorPatryk Wlazlyn <patryk.wlazlyn@linux.intel.com>
Wed, 8 May 2024 13:00:14 +0000 (15:00 +0200)
committerLen Brown <len.brown@intel.com>
Thu, 16 May 2024 01:50:16 +0000 (21:50 -0400)
Reading the counters via perf can be done in bulk with a single syscall,
making the counter values more accurate with respect to one another by
minimizing the time gap between individual counter reads.

Signed-off-by: Patryk Wlazlyn <patryk.wlazlyn@linux.intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>
tools/power/x86/turbostat/turbostat.c

index 9d0278d179655e2fee9605e645ebbbac7f566501..a3842e927799268b93d2303cc267fb1770aaeb91 100644 (file)
@@ -224,6 +224,28 @@ unsigned long long bic_present = BIC_USEC | BIC_TOD | BIC_sysfs | BIC_APIC | BIC
 #define BIC_NOT_PRESENT(COUNTER_BIT) (bic_present &= ~COUNTER_BIT)
 #define BIC_IS_ENABLED(COUNTER_BIT) (bic_enabled & COUNTER_BIT)
 
+/*
+ * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
+ * If you change the values, note they are used both in comparisons
+ * (>= PCL__7) and to index pkg_cstate_limit_strings[].
+ */
+#define PCLUKN 0               /* Unknown */
+#define PCLRSV 1               /* Reserved */
+#define PCL__0 2               /* PC0 */
+#define PCL__1 3               /* PC1 */
+#define PCL__2 4               /* PC2 */
+#define PCL__3 5               /* PC3 */
+#define PCL__4 6               /* PC4 */
+#define PCL__6 7               /* PC6 */
+#define PCL_6N 8               /* PC6 No Retention */
+#define PCL_6R 9               /* PC6 Retention */
+#define PCL__7 10              /* PC7 */
+#define PCL_7S 11              /* PC7 Shrink */
+#define PCL__8 12              /* PC8 */
+#define PCL__9 13              /* PC9 */
+#define PCL_10 14              /* PC10 */
+#define PCLUNL 15              /* Unlimited */
+
 struct amperf_group_fd;
 
 char *proc_stat = "/proc/stat";
@@ -1190,21 +1212,30 @@ enum ccstate_rci_index {
        CCSTATE_RCI_INDEX_C3_RESIDENCY = 1,
        CCSTATE_RCI_INDEX_C6_RESIDENCY = 2,
        CCSTATE_RCI_INDEX_C7_RESIDENCY = 3,
-       NUM_CCSTATE_COUNTERS,
+       PCSTATE_RCI_INDEX_C2_RESIDENCY = 4,
+       PCSTATE_RCI_INDEX_C3_RESIDENCY = 5,
+       PCSTATE_RCI_INDEX_C6_RESIDENCY = 6,
+       PCSTATE_RCI_INDEX_C7_RESIDENCY = 7,
+       PCSTATE_RCI_INDEX_C8_RESIDENCY = 8,
+       PCSTATE_RCI_INDEX_C9_RESIDENCY = 9,
+       PCSTATE_RCI_INDEX_C10_RESIDENCY = 10,
+       NUM_CSTATE_COUNTERS,
 };
 
 struct cstate_counter_info_t {
-       unsigned long long data[NUM_CCSTATE_COUNTERS];
-       enum cstate_source source[NUM_CCSTATE_COUNTERS];
-       unsigned long long msr[NUM_CCSTATE_COUNTERS];
-       int fd_perf;
+       unsigned long long data[NUM_CSTATE_COUNTERS];
+       enum cstate_source source[NUM_CSTATE_COUNTERS];
+       unsigned long long msr[NUM_CSTATE_COUNTERS];
+       int fd_perf_core;
+       int fd_perf_pkg;
 };
 
 struct cstate_counter_info_t *ccstate_counter_info;
 unsigned int ccstate_counter_info_size;
 
-#define CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD (1u << 0)
-#define CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY (1u << 1)
+#define CSTATE_COUNTER_FLAG_COLLECT_PER_CORE   (1u << 0)
+#define CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD ((1u << 1) | CSTATE_COUNTER_FLAG_COLLECT_PER_CORE)
+#define CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY (1u << 2)
 
 struct cstate_counter_arch_info {
        int feature_mask;       /* Mask for testing if the counter is supported on host */
@@ -1214,6 +1245,7 @@ struct cstate_counter_arch_info {
        unsigned int rci_index; /* Maps data from perf counters to global variables */
        unsigned long long bic;
        unsigned long long flags;
+       int pkg_cstate_limit;
 };
 
 static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
@@ -1225,6 +1257,7 @@ static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
         .rci_index = CCSTATE_RCI_INDEX_C1_RESIDENCY,
         .bic = BIC_CPU_c1,
         .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD,
+        .pkg_cstate_limit = 0,
          },
        {
         .feature_mask = CC3,
@@ -1233,7 +1266,8 @@ static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
         .msr = MSR_CORE_C3_RESIDENCY,
         .rci_index = CCSTATE_RCI_INDEX_C3_RESIDENCY,
         .bic = BIC_CPU_c3,
-        .flags = CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
+        .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
+        .pkg_cstate_limit = 0,
          },
        {
         .feature_mask = CC6,
@@ -1242,7 +1276,8 @@ static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
         .msr = MSR_CORE_C6_RESIDENCY,
         .rci_index = CCSTATE_RCI_INDEX_C6_RESIDENCY,
         .bic = BIC_CPU_c6,
-        .flags = CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
+        .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
+        .pkg_cstate_limit = 0,
          },
        {
         .feature_mask = CC7,
@@ -1251,7 +1286,78 @@ static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
         .msr = MSR_CORE_C7_RESIDENCY,
         .rci_index = CCSTATE_RCI_INDEX_C7_RESIDENCY,
         .bic = BIC_CPU_c7,
-        .flags = CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
+        .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
+        .pkg_cstate_limit = 0,
+         },
+       {
+        .feature_mask = PC2,
+        .perf_subsys = "cstate_pkg",
+        .perf_name = "c2-residency",
+        .msr = MSR_PKG_C2_RESIDENCY,
+        .rci_index = PCSTATE_RCI_INDEX_C2_RESIDENCY,
+        .bic = BIC_Pkgpc2,
+        .flags = 0,
+        .pkg_cstate_limit = PCL__2,
+         },
+       {
+        .feature_mask = PC3,
+        .perf_subsys = "cstate_pkg",
+        .perf_name = "c3-residency",
+        .msr = MSR_PKG_C3_RESIDENCY,
+        .rci_index = PCSTATE_RCI_INDEX_C3_RESIDENCY,
+        .bic = BIC_Pkgpc3,
+        .flags = 0,
+        .pkg_cstate_limit = PCL__3,
+         },
+       {
+        .feature_mask = PC6,
+        .perf_subsys = "cstate_pkg",
+        .perf_name = "c6-residency",
+        .msr = MSR_PKG_C6_RESIDENCY,
+        .rci_index = PCSTATE_RCI_INDEX_C6_RESIDENCY,
+        .bic = BIC_Pkgpc6,
+        .flags = 0,
+        .pkg_cstate_limit = PCL__6,
+         },
+       {
+        .feature_mask = PC7,
+        .perf_subsys = "cstate_pkg",
+        .perf_name = "c7-residency",
+        .msr = MSR_PKG_C7_RESIDENCY,
+        .rci_index = PCSTATE_RCI_INDEX_C7_RESIDENCY,
+        .bic = BIC_Pkgpc7,
+        .flags = 0,
+        .pkg_cstate_limit = PCL__7,
+         },
+       {
+        .feature_mask = PC8,
+        .perf_subsys = "cstate_pkg",
+        .perf_name = "c8-residency",
+        .msr = MSR_PKG_C8_RESIDENCY,
+        .rci_index = PCSTATE_RCI_INDEX_C8_RESIDENCY,
+        .bic = BIC_Pkgpc8,
+        .flags = 0,
+        .pkg_cstate_limit = PCL__8,
+         },
+       {
+        .feature_mask = PC9,
+        .perf_subsys = "cstate_pkg",
+        .perf_name = "c9-residency",
+        .msr = MSR_PKG_C9_RESIDENCY,
+        .rci_index = PCSTATE_RCI_INDEX_C9_RESIDENCY,
+        .bic = BIC_Pkgpc9,
+        .flags = 0,
+        .pkg_cstate_limit = PCL__9,
+         },
+       {
+        .feature_mask = PC10,
+        .perf_subsys = "cstate_pkg",
+        .perf_name = "c10-residency",
+        .msr = MSR_PKG_C10_RESIDENCY,
+        .rci_index = PCSTATE_RCI_INDEX_C10_RESIDENCY,
+        .bic = BIC_Pkgpc10,
+        .flags = 0,
+        .pkg_cstate_limit = PCL_10,
          },
 };
 
@@ -1641,15 +1747,8 @@ int get_msr_fd(int cpu)
 
 static void bic_disable_msr_access(void)
 {
-       const unsigned long bic_msrs =
-           BIC_SMI |
-           BIC_Mod_c6 |
-           BIC_CoreTmp |
-           BIC_Totl_c0 |
-           BIC_Any_c0 |
-           BIC_GFX_c0 |
-           BIC_CPUGFX |
-           BIC_Pkgpc2 | BIC_Pkgpc3 | BIC_Pkgpc6 | BIC_Pkgpc7 | BIC_Pkgpc8 | BIC_Pkgpc9 | BIC_Pkgpc10 | BIC_PkgTmp;
+       const unsigned long bic_msrs = BIC_SMI | BIC_Mod_c6 | BIC_CoreTmp |
+           BIC_Totl_c0 | BIC_Any_c0 | BIC_GFX_c0 | BIC_CPUGFX | BIC_PkgTmp;
 
        bic_enabled &= ~bic_msrs;
 
@@ -3493,7 +3592,7 @@ static size_t cstate_counter_info_count_perf(const struct cstate_counter_info_t
 {
        size_t ret = 0;
 
-       for (int i = 0; i < NUM_CCSTATE_COUNTERS; ++i)
+       for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i)
                if (cci->source[i] == CSTATE_SOURCE_PERF)
                        ++ret;
 
@@ -3598,9 +3697,16 @@ char *find_sysfs_path_by_id(struct sysfs_path *sp, int id)
        return NULL;
 }
 
-int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_data *c)
+int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_data *c, struct pkg_data *p)
 {
-       unsigned long long perf_data[NUM_CCSTATE_COUNTERS + 1];
+       /*
+        * Overcommit memory a little bit here,
+        * but skip calculating exact sizes for the buffers.
+        */
+       unsigned long long perf_data[NUM_CSTATE_COUNTERS];
+       unsigned long long perf_data_core[NUM_CSTATE_COUNTERS + 1];
+       unsigned long long perf_data_pkg[NUM_CSTATE_COUNTERS + 1];
+
        struct cstate_counter_info_t *cci;
 
        if (debug)
@@ -3609,35 +3715,72 @@ int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_dat
        assert(ccstate_counter_info);
        assert(cpu <= ccstate_counter_info_size);
 
+       memset(perf_data, 0, sizeof(perf_data));
+       memset(perf_data_core, 0, sizeof(perf_data_core));
+       memset(perf_data_pkg, 0, sizeof(perf_data_pkg));
+
        cci = &ccstate_counter_info[cpu];
 
        /*
         * If we have any perf counters to read, read them all now, in bulk
         */
-       if (cci->fd_perf != -1) {
-               const size_t num_perf_counters = cstate_counter_info_count_perf(cci);
-               const ssize_t expected_read_size =
-                       (num_perf_counters + 1) * sizeof(unsigned long long);
-               const ssize_t actual_read_size =
-                       read(cci->fd_perf, &perf_data[0], sizeof(perf_data));
+       const size_t num_perf_counters = cstate_counter_info_count_perf(cci);
+       ssize_t expected_read_size = num_perf_counters * sizeof(unsigned long long);
+       ssize_t actual_read_size_core = 0, actual_read_size_pkg = 0;
 
-               if (actual_read_size != expected_read_size)
-                       err(-1, "%s: failed to read perf_data (%zu %zu)",
-                               __func__, expected_read_size, actual_read_size);
+       if (cci->fd_perf_core != -1) {
+               /* Each descriptor read begins with number of counters read. */
+               expected_read_size += sizeof(unsigned long long);
+
+               actual_read_size_core = read(cci->fd_perf_core, &perf_data_core[0], sizeof(perf_data_core));
+
+               if (actual_read_size_core <= 0)
+                       err(-1, "%s: read perf %s: %ld", __func__, "core", actual_read_size_core);
+       }
+
+       if (cci->fd_perf_pkg != -1) {
+               /* Each descriptor read begins with number of counters read. */
+               expected_read_size += sizeof(unsigned long long);
+
+               actual_read_size_pkg = read(cci->fd_perf_pkg, &perf_data_pkg[0], sizeof(perf_data_pkg));
+
+               if (actual_read_size_pkg <= 0)
+                       err(-1, "%s: read perf %s: %ld", __func__, "pkg", actual_read_size_pkg);
        }
 
-       for (unsigned int i = 0, pi = 1; i < NUM_CCSTATE_COUNTERS; ++i) {
+       const ssize_t actual_read_size_total = actual_read_size_core + actual_read_size_pkg;
+
+       if (actual_read_size_total != expected_read_size)
+               err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size_total);
+
+       /*
+        * Copy ccstate and pcstate data into unified buffer.
+        *
+        * Skip first element from core and pkg buffers.
+        * Kernel puts there how many counters were read.
+        */
+       const size_t num_core_counters = perf_data_core[0];
+       const size_t num_pkg_counters = perf_data_pkg[0];
+
+       assert(num_perf_counters == num_core_counters + num_pkg_counters);
+
+       /* Copy ccstate perf data */
+       memcpy(&perf_data[0], &perf_data_core[1], num_core_counters * sizeof(unsigned long long));
+
+       /* Copy pcstate perf data */
+       memcpy(&perf_data[num_core_counters], &perf_data_pkg[1], num_pkg_counters * sizeof(unsigned long long));
+
+       for (unsigned int i = 0, pi = 0; i < NUM_CSTATE_COUNTERS; ++i) {
                switch (cci->source[i]) {
                case CSTATE_SOURCE_NONE:
                        break;
 
                case CSTATE_SOURCE_PERF:
                        assert(pi < ARRAY_SIZE(perf_data));
-                       assert(cci->fd_perf != -1);
+                       assert(cci->fd_perf_core != -1 || cci->fd_perf_pkg != -1);
 
                        if (debug) {
-                               fprintf(stderr, "cstate via %s %u: %llu\n",
-                                       "perf", i, perf_data[pi]);
+                               fprintf(stderr, "cstate via %s %u: %llu\n", "perf", i, perf_data[pi]);
                        }
 
                        cci->data[i] = perf_data[pi];
@@ -3651,8 +3794,7 @@ int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_dat
                                return -13 - i;
 
                        if (debug) {
-                               fprintf(stderr, "cstate via %s0x%llx %u: %llu\n",
-                                       "msr", cci->msr[i], i, cci->data[i]);
+                               fprintf(stderr, "cstate via %s0x%llx %u: %llu\n", "msr", cci->msr[i], i, cci->data[i]);
                        }
 
                        break;
@@ -3671,12 +3813,21 @@ int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_dat
                out_counter = cci->data[index];                 \
 } while (0)
 
-       BUILD_BUG_ON(NUM_CCSTATE_COUNTERS != 4);
+       BUILD_BUG_ON(NUM_CSTATE_COUNTERS != 11);
+
        PERF_COUNTER_WRITE_DATA(t->c1, CCSTATE_RCI_INDEX_C1_RESIDENCY);
        PERF_COUNTER_WRITE_DATA(c->c3, CCSTATE_RCI_INDEX_C3_RESIDENCY);
        PERF_COUNTER_WRITE_DATA(c->c6, CCSTATE_RCI_INDEX_C6_RESIDENCY);
        PERF_COUNTER_WRITE_DATA(c->c7, CCSTATE_RCI_INDEX_C7_RESIDENCY);
 
+       PERF_COUNTER_WRITE_DATA(p->pc2, PCSTATE_RCI_INDEX_C2_RESIDENCY);
+       PERF_COUNTER_WRITE_DATA(p->pc3, PCSTATE_RCI_INDEX_C3_RESIDENCY);
+       PERF_COUNTER_WRITE_DATA(p->pc6, PCSTATE_RCI_INDEX_C6_RESIDENCY);
+       PERF_COUNTER_WRITE_DATA(p->pc7, PCSTATE_RCI_INDEX_C7_RESIDENCY);
+       PERF_COUNTER_WRITE_DATA(p->pc8, PCSTATE_RCI_INDEX_C8_RESIDENCY);
+       PERF_COUNTER_WRITE_DATA(p->pc9, PCSTATE_RCI_INDEX_C9_RESIDENCY);
+       PERF_COUNTER_WRITE_DATA(p->pc10, PCSTATE_RCI_INDEX_C10_RESIDENCY);
+
 #undef PERF_COUNTER_WRITE_DATA
 
        return 0;
@@ -3738,7 +3889,7 @@ int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
                t->smi_count = msr & 0xFFFFFFFF;
        }
 
-       get_cstate_counters(cpu, t, c);
+       get_cstate_counters(cpu, t, c, p);
 
        for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
                if (get_mp(cpu, mp, &t->counter[i], mp->sp->path))
@@ -3803,34 +3954,6 @@ int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
                if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
                        return -13;
        }
-       if (DO_BIC(BIC_Pkgpc3))
-               if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
-                       return -9;
-       if (DO_BIC(BIC_Pkgpc6)) {
-               if (platform->has_msr_atom_pkg_c6_residency) {
-                       if (get_msr(cpu, MSR_ATOM_PKG_C6_RESIDENCY, &p->pc6))
-                               return -10;
-               } else {
-                       if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
-                               return -10;
-               }
-       }
-
-       if (DO_BIC(BIC_Pkgpc2))
-               if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
-                       return -11;
-       if (DO_BIC(BIC_Pkgpc7))
-               if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
-                       return -12;
-       if (DO_BIC(BIC_Pkgpc8))
-               if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
-                       return -13;
-       if (DO_BIC(BIC_Pkgpc9))
-               if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
-                       return -13;
-       if (DO_BIC(BIC_Pkgpc10))
-               if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
-                       return -13;
 
        if (DO_BIC(BIC_CPU_LPI))
                p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
@@ -3889,29 +4012,6 @@ done:
        return 0;
 }
 
-/*
- * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
- * If you change the values, note they are used both in comparisons
- * (>= PCL__7) and to index pkg_cstate_limit_strings[].
- */
-
-#define PCLUKN 0               /* Unknown */
-#define PCLRSV 1               /* Reserved */
-#define PCL__0 2               /* PC0 */
-#define PCL__1 3               /* PC1 */
-#define PCL__2 4               /* PC2 */
-#define PCL__3 5               /* PC3 */
-#define PCL__4 6               /* PC4 */
-#define PCL__6 7               /* PC6 */
-#define PCL_6N 8               /* PC6 No Retention */
-#define PCL_6R 9               /* PC6 Retention */
-#define PCL__7 10              /* PC7 */
-#define PCL_7S 11              /* PC7 Shrink */
-#define PCL__8 12              /* PC8 */
-#define PCL__9 13              /* PC9 */
-#define PCL_10 14              /* PC10 */
-#define PCLUNL 15              /* Unlimited */
-
 int pkg_cstate_limit = PCLUKN;
 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
        "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "pc10", "unlimited"
@@ -4410,8 +4510,11 @@ void free_fd_cstate(void)
        const int counter_info_num = ccstate_counter_info_size;
 
        for (int counter_id = 0; counter_id < counter_info_num; ++counter_id) {
-               if (ccstate_counter_info[counter_id].fd_perf != -1)
-                       close(ccstate_counter_info[counter_id].fd_perf);
+               if (ccstate_counter_info[counter_id].fd_perf_core != -1)
+                       close(ccstate_counter_info[counter_id].fd_perf_core);
+
+               if (ccstate_counter_info[counter_id].fd_perf_pkg != -1)
+                       close(ccstate_counter_info[counter_id].fd_perf_pkg);
        }
 
        free(ccstate_counter_info);
@@ -6914,30 +7017,43 @@ static int has_amperf_access(void)
        return 0;
 }
 
-int add_cstate_perf_counter_(int cpu, struct cstate_counter_info_t *cci,
-                            const struct cstate_counter_arch_info *cai)
+int *get_cstate_perf_group_fd(struct cstate_counter_info_t *cci, const char *group_name)
+{
+       if (strcmp(group_name, "cstate_core") == 0)
+               return &cci->fd_perf_core;
+
+       if (strcmp(group_name, "cstate_pkg") == 0)
+               return &cci->fd_perf_pkg;
+
+       return NULL;
+}
+
+int add_cstate_perf_counter_(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
 {
        if (no_perf)
                return -1;
 
+       int *pfd_group = get_cstate_perf_group_fd(cci, cai->perf_subsys);
+
+       if (pfd_group == NULL)
+               return -1;
+
        const unsigned int type = read_perf_type(cai->perf_subsys);
        const unsigned int config = read_rapl_config(cai->perf_subsys, cai->perf_name);
 
-       const int fd_counter =
-               open_perf_counter(cpu, type, config, cci->fd_perf, PERF_FORMAT_GROUP);
+       const int fd_counter = open_perf_counter(cpu, type, config, *pfd_group, PERF_FORMAT_GROUP);
 
        if (fd_counter == -1)
                return -1;
 
        /* If it's the first counter opened, make it a group descriptor */
-       if (cci->fd_perf == -1)
-               cci->fd_perf = fd_counter;
+       if (*pfd_group == -1)
+               *pfd_group = fd_counter;
 
        return fd_counter;
 }
 
-int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci,
-                           const struct cstate_counter_arch_info *cai)
+int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
 {
        int ret = add_cstate_perf_counter_(cpu, cci, cai);
 
@@ -6950,8 +7066,9 @@ int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci,
 void cstate_perf_init_(bool soft_c1)
 {
        bool has_counter;
-       bool *cores_visited;
+       bool *cores_visited = NULL, *pkg_visited = NULL;
        const int cores_visited_elems = topo.max_core_id + 1;
+       const int pkg_visited_elems = topo.max_package_id + 1;
        const int cci_num = topo.max_cpu_num + 1;
 
        ccstate_counter_info = calloc(cci_num, sizeof(*ccstate_counter_info));
@@ -6963,13 +7080,20 @@ void cstate_perf_init_(bool soft_c1)
        if (!cores_visited)
                err(1, "calloc cores_visited");
 
+       pkg_visited = calloc(pkg_visited_elems, sizeof(*pkg_visited));
+       if (!pkg_visited)
+               err(1, "calloc pkg_visited");
+
        /* Initialize cstate_counter_info_percpu */
-       for (int cpu = 0; cpu < cci_num; ++cpu)
-               ccstate_counter_info[cpu].fd_perf = -1;
+       for (int cpu = 0; cpu < cci_num; ++cpu) {
+               ccstate_counter_info[cpu].fd_perf_core = -1;
+               ccstate_counter_info[cpu].fd_perf_pkg = -1;
+       }
 
-       for (int cidx = 0; cidx < NUM_CCSTATE_COUNTERS; ++cidx) {
+       for (int cidx = 0; cidx < NUM_CSTATE_COUNTERS; ++cidx) {
                has_counter = false;
                memset(cores_visited, 0, cores_visited_elems * sizeof(*cores_visited));
+               memset(pkg_visited, 0, pkg_visited_elems * sizeof(*pkg_visited));
 
                const struct cstate_counter_arch_info *cai = &ccstate_counter_arch_infos[cidx];
 
@@ -6981,23 +7105,29 @@ void cstate_perf_init_(bool soft_c1)
                                continue;
 
                        const int core_id = cpus[cpu].physical_core_id;
+                       const int pkg_id = cpus[cpu].physical_package_id;
 
                        assert(core_id < cores_visited_elems);
+                       assert(pkg_id < pkg_visited_elems);
 
                        const bool per_thread = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD;
+                       const bool per_core = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_CORE;
 
                        if (!per_thread && cores_visited[core_id])
                                continue;
 
+                       if (!per_core && pkg_visited[pkg_id])
+                               continue;
+
                        const bool counter_needed = BIC_IS_ENABLED(cai->bic) ||
                            (soft_c1 && (cai->flags & CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY));
                        const bool counter_supported =
-                               platform->supported_cstates & cai->feature_mask;
+                           (platform->supported_cstates & cai->feature_mask) &&
+                           (pkg_cstate_limit >= cai->pkg_cstate_limit);
 
                        if (counter_needed && counter_supported) {
                                /* Use perf API for this counter */
-                               if (!no_perf && cai->perf_name
-                                   && add_cstate_perf_counter(cpu, cci, cai) != -1) {
+                               if (!no_perf && cai->perf_name && add_cstate_perf_counter(cpu, cci, cai) != -1) {
 
                                        cci->source[cai->rci_index] = CSTATE_SOURCE_PERF;
 
@@ -7011,6 +7141,7 @@ void cstate_perf_init_(bool soft_c1)
                        if (cci->source[cai->rci_index] != CSTATE_SOURCE_NONE) {
                                has_counter = true;
                                cores_visited[core_id] = true;
+                               pkg_visited[pkg_id] = true;
                        }
                }
 
@@ -7020,6 +7151,7 @@ void cstate_perf_init_(bool soft_c1)
        }
 
        free(cores_visited);
+       free(pkg_visited);
 }
 
 void cstate_perf_init(void)
@@ -7029,7 +7161,7 @@ void cstate_perf_init(void)
         * but we need APERF, MPERF too.
         */
        const bool soft_c1 = !platform->has_msr_core_c1_res && has_amperf_access()
-                            && platform->supported_cstates & CC1;
+           && platform->supported_cstates & CC1;
 
        if (soft_c1)
                BIC_PRESENT(BIC_CPU_c1);
@@ -7041,27 +7173,6 @@ void probe_cstates(void)
 {
        probe_cst_limit();
 
-       if (platform->supported_cstates & PC2 && (pkg_cstate_limit >= PCL__2))
-               BIC_PRESENT(BIC_Pkgpc2);
-
-       if (platform->supported_cstates & PC3 && (pkg_cstate_limit >= PCL__3))
-               BIC_PRESENT(BIC_Pkgpc3);
-
-       if (platform->supported_cstates & PC6 && (pkg_cstate_limit >= PCL__6))
-               BIC_PRESENT(BIC_Pkgpc6);
-
-       if (platform->supported_cstates & PC7 && (pkg_cstate_limit >= PCL__7))
-               BIC_PRESENT(BIC_Pkgpc7);
-
-       if (platform->supported_cstates & PC8 && (pkg_cstate_limit >= PCL__8))
-               BIC_PRESENT(BIC_Pkgpc8);
-
-       if (platform->supported_cstates & PC9 && (pkg_cstate_limit >= PCL__9))
-               BIC_PRESENT(BIC_Pkgpc9);
-
-       if (platform->supported_cstates & PC10 && (pkg_cstate_limit >= PCL_10))
-               BIC_PRESENT(BIC_Pkgpc10);
-
        if (platform->has_msr_module_c6_res_ms)
                BIC_PRESENT(BIC_Mod_c6);
 
@@ -7320,7 +7431,7 @@ void process_cpuid()
 
 static void counter_info_init(void)
 {
-       for (int i = 0; i < NUM_CCSTATE_COUNTERS; ++i) {
+       for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i) {
                struct cstate_counter_arch_info *const cai = &ccstate_counter_arch_infos[i];
 
                if (platform->has_msr_knl_core_c6_residency && cai->msr == MSR_CORE_C6_RESIDENCY)
@@ -7328,6 +7439,9 @@ static void counter_info_init(void)
 
                if (!platform->has_msr_core_c1_res && cai->msr == MSR_CORE_C1_RES)
                        cai->msr = 0;
+
+               if (platform->has_msr_atom_pkg_c6_residency && cai->msr == MSR_PKG_C6_RESIDENCY)
+                       cai->msr = MSR_ATOM_PKG_C6_RESIDENCY;
        }
 }