]> www.infradead.org Git - users/dwmw2/linux.git/commitdiff
tools/power turbostat: Group SMI counter with APERF and MPERF
authorPatryk Wlazlyn <patryk.wlazlyn@linux.intel.com>
Fri, 24 May 2024 12:31:52 +0000 (14:31 +0200)
committerLen Brown <len.brown@intel.com>
Tue, 9 Jul 2024 13:19:43 +0000 (09:19 -0400)
These three counters now are treated similar to other perf counters
groups. This simplifies and gets rid of a lot of special cases for APERF
and MPERF.

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 29751b41ea0dceab0a54bb1ac509ca484cec0de5..495235055fa22d9d21bde3a5722067b1e973dfa2 100644 (file)
@@ -85,7 +85,6 @@
 enum counter_scope { SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE };
 enum counter_type { COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC, COUNTER_K2M };
 enum counter_format { FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT, FORMAT_AVERAGE };
-enum amperf_source { AMPERF_SOURCE_PERF, AMPERF_SOURCE_MSR };
 enum counter_source { COUNTER_SOURCE_NONE, COUNTER_SOURCE_PERF, COUNTER_SOURCE_MSR };
 
 struct sysfs_path {
@@ -275,7 +274,6 @@ char *proc_stat = "/proc/stat";
 FILE *outf;
 int *fd_percpu;
 int *fd_instr_count_percpu;
-struct amperf_group_fd *fd_amperf_percpu;      /* File descriptors for perf group with APERF and MPERF counters. */
 struct timeval interval_tv = { 5, 0 };
 struct timespec interval_ts = { 5, 0 };
 
@@ -290,6 +288,7 @@ unsigned int summary_only;
 unsigned int list_header_only;
 unsigned int dump_only;
 unsigned int has_aperf;
+unsigned int has_aperf_access;
 unsigned int has_epb;
 unsigned int has_turbo;
 unsigned int is_hybrid;
@@ -330,7 +329,6 @@ unsigned int first_counter_read = 1;
 int ignore_stdin;
 bool no_msr;
 bool no_perf;
-enum amperf_source amperf_source;
 
 enum gfx_sysfs_idx {
        GFX_rc6,
@@ -1381,6 +1379,67 @@ static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
          },
 };
 
+/* Indexes used to map data read from perf and MSRs into global variables */
+enum msr_rci_index {
+       MSR_RCI_INDEX_APERF = 0,
+       MSR_RCI_INDEX_MPERF = 1,
+       MSR_RCI_INDEX_SMI = 2,
+       NUM_MSR_COUNTERS,
+};
+
+struct msr_counter_info_t {
+       unsigned long long data[NUM_MSR_COUNTERS];
+       enum counter_source source[NUM_MSR_COUNTERS];
+       unsigned long long msr[NUM_MSR_COUNTERS];
+       unsigned long long msr_mask[NUM_MSR_COUNTERS];
+       int fd_perf;
+};
+
+struct msr_counter_info_t *msr_counter_info;
+unsigned int msr_counter_info_size;
+
+struct msr_counter_arch_info {
+       const char *perf_subsys;
+       const char *perf_name;
+       unsigned long long msr;
+       unsigned long long msr_mask;
+       unsigned int rci_index; /* Maps data from perf counters to global variables */
+       bool needed;
+       bool present;
+};
+
+enum msr_arch_info_index {
+       MSR_ARCH_INFO_APERF_INDEX = 0,
+       MSR_ARCH_INFO_MPERF_INDEX = 1,
+       MSR_ARCH_INFO_SMI_INDEX = 2,
+};
+
+static struct msr_counter_arch_info msr_counter_arch_infos[] = {
+       [MSR_ARCH_INFO_APERF_INDEX] = {
+               .perf_subsys = "msr",
+               .perf_name = "aperf",
+               .msr = MSR_IA32_APERF,
+               .msr_mask = 0xFFFFFFFFFFFFFFFF,
+               .rci_index = MSR_RCI_INDEX_APERF,
+       },
+
+       [MSR_ARCH_INFO_MPERF_INDEX] = {
+               .perf_subsys = "msr",
+               .perf_name = "mperf",
+               .msr = MSR_IA32_MPERF,
+               .msr_mask = 0xFFFFFFFFFFFFFFFF,
+               .rci_index = MSR_RCI_INDEX_MPERF,
+       },
+
+       [MSR_ARCH_INFO_SMI_INDEX] = {
+               .perf_subsys = "msr",
+               .perf_name = "smi",
+               .msr = MSR_SMI_COUNT,
+               .msr_mask = 0xFFFFFFFF,
+               .rci_index = MSR_RCI_INDEX_SMI,
+       },
+};
+
 struct thread_data {
        struct timeval tv_begin;
        struct timeval tv_end;
@@ -1767,7 +1826,7 @@ 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 |
+       const unsigned long bic_msrs = BIC_Mod_c6 | BIC_CoreTmp |
            BIC_Totl_c0 | BIC_Any_c0 | BIC_GFX_c0 | BIC_CPUGFX | BIC_PkgTmp;
 
        bic_enabled &= ~bic_msrs;
@@ -3402,30 +3461,6 @@ static unsigned int read_perf_counter_info_n(const char *const path, const char
        return v;
 }
 
-static unsigned int read_msr_type(void)
-{
-       const char *const path = "/sys/bus/event_source/devices/msr/type";
-       const char *const format = "%u";
-
-       return read_perf_counter_info_n(path, format);
-}
-
-static unsigned int read_aperf_config(void)
-{
-       const char *const path = "/sys/bus/event_source/devices/msr/events/aperf";
-       const char *const format = "event=%x";
-
-       return read_perf_counter_info_n(path, format);
-}
-
-static unsigned int read_mperf_config(void)
-{
-       const char *const path = "/sys/bus/event_source/devices/msr/events/mperf";
-       const char *const format = "event=%x";
-
-       return read_perf_counter_info_n(path, format);
-}
-
 static unsigned int read_perf_type(const char *subsys)
 {
        const char *const path_format = "/sys/bus/event_source/devices/%s/type";
@@ -3479,124 +3514,6 @@ static double read_perf_rapl_scale(const char *subsys, const char *event_name)
        return scale;
 }
 
-static struct amperf_group_fd open_amperf_fd(int cpu)
-{
-       const unsigned int msr_type = read_msr_type();
-       const unsigned int aperf_config = read_aperf_config();
-       const unsigned int mperf_config = read_mperf_config();
-       struct amperf_group_fd fds = {.aperf = -1, .mperf = -1 };
-
-       fds.aperf = open_perf_counter(cpu, msr_type, aperf_config, -1, PERF_FORMAT_GROUP);
-       fds.mperf = open_perf_counter(cpu, msr_type, mperf_config, fds.aperf, PERF_FORMAT_GROUP);
-
-       return fds;
-}
-
-static int get_amperf_fd(int cpu)
-{
-       assert(fd_amperf_percpu);
-
-       if (fd_amperf_percpu[cpu].aperf)
-               return fd_amperf_percpu[cpu].aperf;
-
-       fd_amperf_percpu[cpu] = open_amperf_fd(cpu);
-
-       return fd_amperf_percpu[cpu].aperf;
-}
-
-/* Read APERF, MPERF and TSC using the perf API. */
-static int read_aperf_mperf_tsc_perf(struct thread_data *t, int cpu)
-{
-       union {
-               struct {
-                       unsigned long nr_entries;
-                       unsigned long aperf;
-                       unsigned long mperf;
-               };
-
-               unsigned long as_array[3];
-       } cnt;
-
-       const int fd_amperf = get_amperf_fd(cpu);
-
-       /*
-        * Read the TSC with rdtsc, because we want the absolute value and not
-        * the offset from the start of the counter.
-        */
-       t->tsc = rdtsc();
-
-       const int n = read(fd_amperf, &cnt.as_array[0], sizeof(cnt.as_array));
-
-       if (n != sizeof(cnt.as_array))
-               return -2;
-
-       t->aperf = cnt.aperf * aperf_mperf_multiplier;
-       t->mperf = cnt.mperf * aperf_mperf_multiplier;
-
-       return 0;
-}
-
-/* Read APERF, MPERF and TSC using the MSR driver and rdtsc instruction. */
-static int read_aperf_mperf_tsc_msr(struct thread_data *t, int cpu)
-{
-       unsigned long long tsc_before, tsc_between, tsc_after, aperf_time, mperf_time;
-       int aperf_mperf_retry_count = 0;
-
-       /*
-        * The TSC, APERF and MPERF must be read together for
-        * APERF/MPERF and MPERF/TSC to give accurate results.
-        *
-        * Unfortunately, APERF and MPERF are read by
-        * individual system call, so delays may occur
-        * between them.  If the time to read them
-        * varies by a large amount, we re-read them.
-        */
-
-       /*
-        * This initial dummy APERF read has been seen to
-        * reduce jitter in the subsequent reads.
-        */
-
-       if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
-               return -3;
-
-retry:
-       t->tsc = rdtsc();       /* re-read close to APERF */
-
-       tsc_before = t->tsc;
-
-       if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
-               return -3;
-
-       tsc_between = rdtsc();
-
-       if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
-               return -4;
-
-       tsc_after = rdtsc();
-
-       aperf_time = tsc_between - tsc_before;
-       mperf_time = tsc_after - tsc_between;
-
-       /*
-        * If the system call latency to read APERF and MPERF
-        * differ by more than 2x, then try again.
-        */
-       if ((aperf_time > (2 * mperf_time)) || (mperf_time > (2 * aperf_time))) {
-               aperf_mperf_retry_count++;
-               if (aperf_mperf_retry_count < 5)
-                       goto retry;
-               else
-                       warnx("cpu%d jitter %lld %lld", cpu, aperf_time, mperf_time);
-       }
-       aperf_mperf_retry_count = 0;
-
-       t->aperf = t->aperf * aperf_mperf_multiplier;
-       t->mperf = t->mperf * aperf_mperf_multiplier;
-
-       return 0;
-}
-
 size_t rapl_counter_info_count_perf(const struct rapl_counter_info_t *rci)
 {
        size_t ret = 0;
@@ -3853,6 +3770,84 @@ int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_dat
        return 0;
 }
 
+size_t msr_counter_info_count_perf(const struct msr_counter_info_t *mci)
+{
+       size_t ret = 0;
+
+       for (int i = 0; i < NUM_MSR_COUNTERS; ++i)
+               if (mci->source[i] == COUNTER_SOURCE_PERF)
+                       ++ret;
+
+       return ret;
+}
+
+int get_smi_aperf_mperf(unsigned int cpu, struct thread_data *t)
+{
+       unsigned long long perf_data[NUM_MSR_COUNTERS + 1];
+
+       struct msr_counter_info_t *mci;
+
+       if (debug)
+               fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
+
+       assert(msr_counter_info);
+       assert(cpu <= msr_counter_info_size);
+
+       mci = &msr_counter_info[cpu];
+
+       ZERO_ARRAY(perf_data);
+       ZERO_ARRAY(mci->data);
+
+       if (mci->fd_perf != -1) {
+               const size_t num_perf_counters = msr_counter_info_count_perf(mci);
+               const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
+               const ssize_t actual_read_size = read(mci->fd_perf, &perf_data[0], sizeof(perf_data));
+
+               if (actual_read_size != expected_read_size)
+                       err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size,
+                           actual_read_size);
+       }
+
+       for (unsigned int i = 0, pi = 1; i < NUM_MSR_COUNTERS; ++i) {
+               switch (mci->source[i]) {
+               case COUNTER_SOURCE_NONE:
+                       break;
+
+               case COUNTER_SOURCE_PERF:
+                       assert(pi < ARRAY_SIZE(perf_data));
+                       assert(mci->fd_perf != -1);
+
+                       if (debug)
+                               fprintf(stderr, "Reading msr counter via perf at %u: %llu\n", i, perf_data[pi]);
+
+                       mci->data[i] = perf_data[pi];
+
+                       ++pi;
+                       break;
+
+               case COUNTER_SOURCE_MSR:
+                       assert(!no_msr);
+
+                       if (get_msr(cpu, mci->msr[i], &mci->data[i]))
+                               return -2 - i;
+
+                       mci->data[i] &= mci->msr_mask[i];
+
+                       if (debug)
+                               fprintf(stderr, "Reading msr counter via msr at %u: %llu\n", i, mci->data[i]);
+
+                       break;
+               }
+       }
+
+       BUILD_BUG_ON(NUM_MSR_COUNTERS != 3);
+       t->aperf = mci->data[MSR_RCI_INDEX_APERF];
+       t->mperf = mci->data[MSR_RCI_INDEX_MPERF];
+       t->smi_count = mci->data[MSR_RCI_INDEX_SMI];
+
+       return 0;
+}
+
 /*
  * get_counters(...)
  * migrate to cpu
@@ -3878,24 +3873,7 @@ int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 
        t->tsc = rdtsc();       /* we are running on local CPU of interest */
 
-       if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) || DO_BIC(BIC_IPC)
-           || soft_c1_residency_display(BIC_Avg_MHz)) {
-               int status = -1;
-
-               assert(!no_perf || !no_msr);
-
-               switch (amperf_source) {
-               case AMPERF_SOURCE_PERF:
-                       status = read_aperf_mperf_tsc_perf(t, cpu);
-                       break;
-               case AMPERF_SOURCE_MSR:
-                       status = read_aperf_mperf_tsc_msr(t, cpu);
-                       break;
-               }
-
-               if (status != 0)
-                       return status;
-       }
+       get_smi_aperf_mperf(cpu, t);
 
        if (DO_BIC(BIC_IPC))
                if (read(get_instr_count_fd(cpu), &t->instr_count, sizeof(long long)) != sizeof(long long))
@@ -3903,11 +3881,6 @@ int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 
        if (DO_BIC(BIC_IRQ))
                t->irq_count = irqs_per_cpu[cpu];
-       if (DO_BIC(BIC_SMI)) {
-               if (get_msr(cpu, MSR_SMI_COUNT, &msr))
-                       return -5;
-               t->smi_count = msr & 0xFFFFFFFF;
-       }
 
        get_cstate_counters(cpu, t, c, p);
 
@@ -4489,25 +4462,6 @@ void free_fd_percpu(void)
        fd_percpu = NULL;
 }
 
-void free_fd_amperf_percpu(void)
-{
-       int i;
-
-       if (!fd_amperf_percpu)
-               return;
-
-       for (i = 0; i < topo.max_cpu_num + 1; ++i) {
-               if (fd_amperf_percpu[i].mperf != 0)
-                       close(fd_amperf_percpu[i].mperf);
-
-               if (fd_amperf_percpu[i].aperf != 0)
-                       close(fd_amperf_percpu[i].aperf);
-       }
-
-       free(fd_amperf_percpu);
-       fd_amperf_percpu = NULL;
-}
-
 void free_fd_instr_count_percpu(void)
 {
        if (!fd_instr_count_percpu)
@@ -4542,6 +4496,21 @@ void free_fd_cstate(void)
        ccstate_counter_info_size = 0;
 }
 
+void free_fd_msr(void)
+{
+       if (!msr_counter_info)
+               return;
+
+       for (int cpu = 0; cpu < topo.max_cpu_num; ++cpu) {
+               if (msr_counter_info[cpu].fd_perf != -1)
+                       close(msr_counter_info[cpu].fd_perf);
+       }
+
+       free(msr_counter_info);
+       msr_counter_info = NULL;
+       msr_counter_info_size = 0;
+}
+
 void free_fd_rapl_percpu(void)
 {
        if (!rapl_counter_info_perdomain)
@@ -4601,7 +4570,7 @@ void free_all_buffers(void)
 
        free_fd_percpu();
        free_fd_instr_count_percpu();
-       free_fd_amperf_percpu();
+       free_fd_msr();
        free_fd_rapl_percpu();
        free_fd_cstate();
 
@@ -4938,6 +4907,7 @@ static void update_effective_set(bool startup)
 }
 
 void linux_perf_init(void);
+void msr_perf_init(void);
 void rapl_perf_init(void);
 void cstate_perf_init(void);
 
@@ -4946,6 +4916,7 @@ void re_initialize(void)
        free_all_buffers();
        setup_all_buffers(false);
        linux_perf_init();
+       msr_perf_init();
        rapl_perf_init();
        cstate_perf_init();
        fprintf(outf, "turbostat: re-initialized with num_cpus %d, allowed_cpus %d\n", topo.num_cpus,
@@ -6799,15 +6770,6 @@ static int has_instr_count_access(void)
        return has_access;
 }
 
-bool is_aperf_access_required(void)
-{
-       return BIC_IS_ENABLED(BIC_Avg_MHz)
-           || BIC_IS_ENABLED(BIC_Busy)
-           || BIC_IS_ENABLED(BIC_Bzy_MHz)
-           || BIC_IS_ENABLED(BIC_IPC)
-           || BIC_IS_ENABLED(BIC_CPU_c1);
-}
-
 int add_rapl_perf_counter_(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
                           double *scale_, enum rapl_unit *unit_)
 {
@@ -6866,14 +6828,6 @@ void linux_perf_init(void)
                if (fd_instr_count_percpu == NULL)
                        err(-1, "calloc fd_instr_count_percpu");
        }
-
-       const bool aperf_required = is_aperf_access_required();
-
-       if (aperf_required && has_aperf && amperf_source == AMPERF_SOURCE_PERF) {
-               fd_amperf_percpu = calloc(topo.max_cpu_num + 1, sizeof(*fd_amperf_percpu));
-               if (fd_amperf_percpu == NULL)
-                       err(-1, "calloc fd_amperf_percpu");
-       }
 }
 
 void rapl_perf_init(void)
@@ -6966,75 +6920,11 @@ void rapl_perf_init(void)
        free(domain_visited);
 }
 
-static int has_amperf_access_via_msr(void)
-{
-       if (no_msr)
-               return 0;
-
-       if (probe_msr(base_cpu, MSR_IA32_APERF))
-               return 0;
-
-       if (probe_msr(base_cpu, MSR_IA32_MPERF))
-               return 0;
-
-       return 1;
-}
-
-static int has_amperf_access_via_perf(void)
-{
-       struct amperf_group_fd fds;
-
-       /*
-        * Cache the last result, so we don't warn the user multiple times
-        *
-        * Negative means cached, no access
-        * Zero means not cached
-        * Positive means cached, has access
-        */
-       static int has_access_cached;
-
-       if (no_perf)
-               return 0;
-
-       if (has_access_cached != 0)
-               return has_access_cached > 0;
-
-       fds = open_amperf_fd(base_cpu);
-       has_access_cached = (fds.aperf != -1) && (fds.mperf != -1);
-
-       if (fds.aperf == -1)
-               warnx("Failed to access %s. Some of the counters may not be available\n"
-                     "\tRun as root to enable them or use %s to disable the access explicitly",
-                     "APERF perf counter", "--no-perf");
-       else
-               close(fds.aperf);
-
-       if (fds.mperf == -1)
-               warnx("Failed to access %s. Some of the counters may not be available\n"
-                     "\tRun as root to enable them or use %s to disable the access explicitly",
-                     "MPERF perf counter", "--no-perf");
-       else
-               close(fds.mperf);
-
-       if (has_access_cached == 0)
-               has_access_cached = -1;
-
-       return has_access_cached > 0;
-}
-
-/* Check if we can access APERF and MPERF */
+/* Assumes msr_counter_info is populated */
 static int has_amperf_access(void)
 {
-       if (!is_aperf_access_required())
-               return 0;
-
-       if (!no_msr && has_amperf_access_via_msr())
-               return 1;
-
-       if (!no_perf && has_amperf_access_via_perf())
-               return 1;
-
-       return 0;
+       return msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].present &&
+           msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].present;
 }
 
 int *get_cstate_perf_group_fd(struct cstate_counter_info_t *cci, const char *group_name)
@@ -7083,6 +6973,114 @@ int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci, const st
        return ret;
 }
 
+int add_msr_perf_counter_(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
+{
+       if (no_perf)
+               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);
+
+       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;
+
+       return fd_counter;
+}
+
+int add_msr_perf_counter(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
+{
+       int ret = add_msr_perf_counter_(cpu, cci, cai);
+
+       if (debug)
+               fprintf(stderr, "%s: %s/%s: %d (cpu: %d)\n", __func__, cai->perf_subsys, cai->perf_name, ret, cpu);
+
+       return ret;
+}
+
+void msr_perf_init_(void)
+{
+       const int mci_num = topo.max_cpu_num + 1;
+
+       msr_counter_info = calloc(mci_num, sizeof(*msr_counter_info));
+       if (!msr_counter_info)
+               err(1, "calloc msr_counter_info");
+       msr_counter_info_size = mci_num;
+
+       for (int cpu = 0; cpu < mci_num; ++cpu)
+               msr_counter_info[cpu].fd_perf = -1;
+
+       for (int cidx = 0; cidx < NUM_MSR_COUNTERS; ++cidx) {
+
+               struct msr_counter_arch_info *cai = &msr_counter_arch_infos[cidx];
+
+               cai->present = false;
+
+               for (int cpu = 0; cpu < mci_num; ++cpu) {
+
+                       struct msr_counter_info_t *const cci = &msr_counter_info[cpu];
+
+                       if (cpu_is_not_allowed(cpu))
+                               continue;
+
+                       if (cai->needed) {
+                               /* Use perf API for this counter */
+                               if (!no_perf && cai->perf_name && add_msr_perf_counter(cpu, cci, cai) != -1) {
+                                       cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
+                                       cai->present = true;
+
+                                       /* User MSR for this counter */
+                               } else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
+                                       cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
+                                       cci->msr[cai->rci_index] = cai->msr;
+                                       cci->msr_mask[cai->rci_index] = cai->msr_mask;
+                                       cai->present = true;
+                               }
+                       }
+               }
+       }
+}
+
+/* Initialize data for reading perf counters from the MSR group. */
+void msr_perf_init(void)
+{
+       bool need_amperf = false, need_smi = false;
+       const bool need_soft_c1 = (!platform->has_msr_core_c1_res) && (platform->supported_cstates & CC1);
+
+       need_amperf = BIC_IS_ENABLED(BIC_Avg_MHz) || BIC_IS_ENABLED(BIC_Busy) || BIC_IS_ENABLED(BIC_Bzy_MHz)
+           || BIC_IS_ENABLED(BIC_IPC) || need_soft_c1;
+
+       if (BIC_IS_ENABLED(BIC_SMI))
+               need_smi = true;
+
+       /* Enable needed counters */
+       msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].needed = need_amperf;
+       msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].needed = need_amperf;
+       msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].needed = need_smi;
+
+       msr_perf_init_();
+
+       const bool has_amperf = has_amperf_access();
+       const bool has_smi = msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].present;
+
+       has_aperf_access = has_amperf;
+
+       if (has_amperf) {
+               BIC_PRESENT(BIC_Avg_MHz);
+               BIC_PRESENT(BIC_Busy);
+               BIC_PRESENT(BIC_Bzy_MHz);
+               BIC_PRESENT(BIC_SMI);
+       }
+
+       if (has_smi)
+               BIC_PRESENT(BIC_SMI);
+}
+
 void cstate_perf_init_(bool soft_c1)
 {
        bool has_counter;
@@ -7340,12 +7338,6 @@ void process_cpuid()
 
        __cpuid(0x6, eax, ebx, ecx, edx);
        has_aperf = ecx & (1 << 0);
-       if (has_aperf && has_amperf_access()) {
-               BIC_PRESENT(BIC_Avg_MHz);
-               BIC_PRESENT(BIC_Busy);
-               BIC_PRESENT(BIC_Bzy_MHz);
-               BIC_PRESENT(BIC_IPC);
-       }
        do_dts = eax & (1 << 0);
        if (do_dts)
                BIC_PRESENT(BIC_CoreTmp);
@@ -7462,6 +7454,11 @@ static void counter_info_init(void)
                if (platform->has_msr_atom_pkg_c6_residency && cai->msr == MSR_PKG_C6_RESIDENCY)
                        cai->msr = MSR_ATOM_PKG_C6_RESIDENCY;
        }
+
+       for (int i = 0; i < NUM_MSR_COUNTERS; ++i) {
+               msr_counter_arch_infos[i].present = false;
+               msr_counter_arch_infos[i].needed = false;
+       }
 }
 
 void probe_pm_features(void)
@@ -7837,21 +7834,6 @@ void set_base_cpu(void)
        err(-ENODEV, "No valid cpus found");
 }
 
-static void set_amperf_source(void)
-{
-       amperf_source = AMPERF_SOURCE_PERF;
-
-       const bool aperf_required = is_aperf_access_required();
-
-       if (no_perf || !aperf_required || !has_amperf_access_via_perf())
-               amperf_source = AMPERF_SOURCE_MSR;
-
-       if (quiet || !debug)
-               return;
-
-       fprintf(outf, "aperf/mperf source preference: %s\n", amperf_source == AMPERF_SOURCE_MSR ? "msr" : "perf");
-}
-
 bool has_added_counters(void)
 {
        /*
@@ -7862,54 +7844,8 @@ bool has_added_counters(void)
        return sys.added_core_counters | sys.added_thread_counters | sys.added_package_counters;
 }
 
-bool is_msr_access_required(void)
-{
-       if (no_msr)
-               return false;
-
-       if (has_added_counters())
-               return true;
-
-       return BIC_IS_ENABLED(BIC_SMI)
-           || BIC_IS_ENABLED(BIC_CPU_c1)
-           || BIC_IS_ENABLED(BIC_CPU_c3)
-           || BIC_IS_ENABLED(BIC_CPU_c6)
-           || BIC_IS_ENABLED(BIC_CPU_c7)
-           || BIC_IS_ENABLED(BIC_Mod_c6)
-           || BIC_IS_ENABLED(BIC_CoreTmp)
-           || BIC_IS_ENABLED(BIC_Totl_c0)
-           || BIC_IS_ENABLED(BIC_Any_c0)
-           || BIC_IS_ENABLED(BIC_GFX_c0)
-           || BIC_IS_ENABLED(BIC_CPUGFX)
-           || BIC_IS_ENABLED(BIC_Pkgpc3)
-           || BIC_IS_ENABLED(BIC_Pkgpc6)
-           || BIC_IS_ENABLED(BIC_Pkgpc2)
-           || BIC_IS_ENABLED(BIC_Pkgpc7)
-           || BIC_IS_ENABLED(BIC_Pkgpc8)
-           || BIC_IS_ENABLED(BIC_Pkgpc9)
-           || BIC_IS_ENABLED(BIC_Pkgpc10)
-           /* TODO: Multiplex access with perf */
-           || BIC_IS_ENABLED(BIC_CorWatt)
-           || BIC_IS_ENABLED(BIC_Cor_J)
-           || BIC_IS_ENABLED(BIC_PkgWatt)
-           || BIC_IS_ENABLED(BIC_CorWatt)
-           || BIC_IS_ENABLED(BIC_GFXWatt)
-           || BIC_IS_ENABLED(BIC_RAMWatt)
-           || BIC_IS_ENABLED(BIC_Pkg_J)
-           || BIC_IS_ENABLED(BIC_Cor_J)
-           || BIC_IS_ENABLED(BIC_GFX_J)
-           || BIC_IS_ENABLED(BIC_RAM_J)
-           || BIC_IS_ENABLED(BIC_PKG__)
-           || BIC_IS_ENABLED(BIC_RAM__)
-           || BIC_IS_ENABLED(BIC_PkgTmp)
-           || (is_aperf_access_required() && !has_amperf_access_via_perf());
-}
-
 void check_msr_access(void)
 {
-       if (!is_msr_access_required())
-               no_msr = 1;
-
        check_dev_msr();
        check_msr_permission();
 
@@ -7919,19 +7855,8 @@ void check_msr_access(void)
 
 void check_perf_access(void)
 {
-       const bool intrcount_required = BIC_IS_ENABLED(BIC_IPC);
-
-       if (no_perf || !intrcount_required || !has_instr_count_access())
+       if (no_perf || !BIC_IS_ENABLED(BIC_IPC) || !has_instr_count_access())
                bic_enabled &= ~BIC_IPC;
-
-       const bool aperf_required = is_aperf_access_required();
-
-       if (!aperf_required || !has_amperf_access()) {
-               bic_enabled &= ~BIC_Avg_MHz;
-               bic_enabled &= ~BIC_Busy;
-               bic_enabled &= ~BIC_Bzy_MHz;
-               bic_enabled &= ~BIC_IPC;
-       }
 }
 
 void turbostat_init()
@@ -7943,7 +7868,7 @@ void turbostat_init()
        process_cpuid();
        counter_info_init();
        probe_pm_features();
-       set_amperf_source();
+       msr_perf_init();
        linux_perf_init();
        rapl_perf_init();
        cstate_perf_init();
@@ -7951,8 +7876,8 @@ void turbostat_init()
        for_all_cpus(get_cpu_type, ODD_COUNTERS);
        for_all_cpus(get_cpu_type, EVEN_COUNTERS);
 
-       if (DO_BIC(BIC_IPC))
-               (void)get_instr_count_fd(base_cpu);
+       if (BIC_IS_ENABLED(BIC_IPC) && has_aperf_access && get_instr_count_fd(base_cpu) != -1)
+               BIC_PRESENT(BIC_IPC);
 
        /*
         * If TSC tweak is needed, but couldn't get it,