From c3cb2075deef7cd35401f9d2f00f3aeb23ed16d5 Mon Sep 17 00:00:00 2001 From: Ray Kinsella Date: Tue, 7 Sep 2021 16:26:58 +0100 Subject: perfmon: adding support for papi TMAM Adding support for Linux papi TMAM on Intel Snowridge. Adds the ability to indicate that a bundle should be thread or node bundle type based on available cpu features (rdpmc support). Type: feature Signed-off-by: Ray Kinsella Change-Id: Ib871b2644fdb2410fbb580e0d21c3a8e2be13aba --- src/plugins/perfmon/cli.c | 12 ++- src/plugins/perfmon/intel/bundle/topdown_metrics.c | 110 ++++++++++++--------- src/plugins/perfmon/intel/core.h | 10 +- src/plugins/perfmon/perfmon.h | 35 ++++++- 4 files changed, 113 insertions(+), 54 deletions(-) (limited to 'src/plugins') diff --git a/src/plugins/perfmon/cli.c b/src/plugins/perfmon/cli.c index 3305480031d..1535570de7e 100644 --- a/src/plugins/perfmon/cli.c +++ b/src/plugins/perfmon/cli.c @@ -128,7 +128,8 @@ show_perfmon_bundle_command_fn (vlib_main_t *vm, unformat_input_t *input, vlib_cli_output (vm, "%U\n", format_perfmon_bundle, 0, 0); for (int i = 0; i < vec_len (vb); i++) - if (!vb[i]->cpu_supports || vb[i]->cpu_supports ()) + /* bundle type will be unknown if no cpu_supports matched */ + if (vb[i]->type != PERFMON_BUNDLE_TYPE_UNKNOWN) vlib_cli_output (vm, "%U\n", format_perfmon_bundle, vb[i], verbose); vec_free (vb); @@ -312,7 +313,7 @@ show_perfmon_stats_command_fn (vlib_main_t *vm, unformat_input_t *input, { in = vec_elt_at_index (it->instances, i); r = vec_elt_at_index (readings, i); - table_format_cell (t, col, -1, "%s", in->name); + table_format_cell (t, col, -1, "%s", in->name, b->type); if (b->type == PERFMON_BUNDLE_TYPE_NODE) { perfmon_thread_runtime_t *tr; @@ -322,19 +323,20 @@ show_perfmon_stats_command_fn (vlib_main_t *vm, unformat_input_t *input, { perfmon_node_stats_t ns; table_format_cell (t, ++col, -1, "%U", format_vlib_node_name, - vm, j); + vm, j, b->type); table_set_cell_align (t, col, -1, TTAA_RIGHT); table_set_cell_fg_color (t, col, -1, TTAC_CYAN); clib_memcpy_fast (&ns, tr->node_stats + j, sizeof (ns)); for (int j = 0; j < n_row; j++) - table_format_cell (t, col, j, "%U", b->format_fn, &ns, j); + table_format_cell (t, col, j, "%U", b->format_fn, &ns, j, + b->type); } } else { for (int j = 0; j < n_row; j++) - table_format_cell (t, i, j, "%U", b->format_fn, r, j); + table_format_cell (t, i, j, "%U", b->format_fn, r, j, b->type); } col++; } diff --git a/src/plugins/perfmon/intel/bundle/topdown_metrics.c b/src/plugins/perfmon/intel/bundle/topdown_metrics.c index 386f3843bc3..6913781be90 100644 --- a/src/plugins/perfmon/intel/bundle/topdown_metrics.c +++ b/src/plugins/perfmon/intel/bundle/topdown_metrics.c @@ -27,29 +27,42 @@ typedef enum { - TOPDOWN_E_METRIC_RETIRING = 0, - TOPDOWN_E_METRIC_BAD_SPEC, - TOPDOWN_E_METRIC_FE_BOUND, - TOPDOWN_E_METRIC_BE_BOUND, -} topdown_lvl1_counters_t; + TOPDOWN_E_RETIRING = 0, + TOPDOWN_E_BAD_SPEC, + TOPDOWN_E_FE_BOUND, + TOPDOWN_E_BE_BOUND, +} topdown_lvl1_t; enum { - TOPDOWN_SLOTS = 0, - TOPDOWN_METRICS, -} topdown_lvl1_metrics_t; + TOPDOWN_E_RDPMC_SLOTS = 0, + TOPDOWN_E_RDPMC_METRICS, +}; + +typedef f64 (topdown_lvl1_parse_fn_t) (void *, topdown_lvl1_t); -static_always_inline f32 -topdown_lvl1_parse_row (perfmon_node_stats_t *ns, topdown_lvl1_counters_t e) +/* Parse thread level states from perfmon_reading */ +static_always_inline f64 +topdown_lvl1_perf_reading (void *ps, topdown_lvl1_t e) { + perfmon_reading_t *ss = (perfmon_reading_t *) ps; + + /* slots are at value[0], everthing else follows at +1 */ + return ((f64) ss->value[e + 1] / ss->value[0]) * 100; +} + +static_always_inline f64 +topdown_lvl1_rdpmc_metric (void *ps, topdown_lvl1_t e) +{ + perfmon_node_stats_t *ss = (perfmon_node_stats_t *) ps; f64 slots_t0 = - ns->t[0].value[TOPDOWN_SLOTS] * - ((f64) GET_METRIC (ns->t[0].value[TOPDOWN_METRICS], e) / 0xff); + ss->t[0].value[TOPDOWN_E_RDPMC_SLOTS] * + ((f64) GET_METRIC (ss->t[0].value[TOPDOWN_E_RDPMC_METRICS], e) / 0xff); f64 slots_t1 = - ns->t[1].value[TOPDOWN_SLOTS] * - ((f64) GET_METRIC (ns->t[1].value[TOPDOWN_METRICS], e) / 0xff); - u64 slots_delta = - ns->t[1].value[TOPDOWN_SLOTS] - ns->t[0].value[TOPDOWN_SLOTS]; + ss->t[1].value[TOPDOWN_E_RDPMC_SLOTS] * + ((f64) GET_METRIC (ss->t[1].value[TOPDOWN_E_RDPMC_METRICS], e) / 0xff); + u64 slots_delta = ss->t[1].value[TOPDOWN_E_RDPMC_SLOTS] - + ss->t[0].value[TOPDOWN_E_RDPMC_SLOTS]; slots_t1 = slots_t1 - slots_t0; @@ -59,53 +72,60 @@ topdown_lvl1_parse_row (perfmon_node_stats_t *ns, topdown_lvl1_counters_t e) static u8 * format_topdown_lvl1 (u8 *s, va_list *args) { - perfmon_node_stats_t *st = va_arg (*args, perfmon_node_stats_t *); - u64 row = va_arg (*args, int); + void *ps = va_arg (*args, void *); + u64 idx = va_arg (*args, int); + perfmon_bundle_type_t type = va_arg (*args, perfmon_bundle_type_t); + f64 sv = 0; + + topdown_lvl1_parse_fn_t *parse_fn, + *parse_fns[PERFMON_BUNDLE_TYPE_MAX] = { 0, topdown_lvl1_rdpmc_metric, + topdown_lvl1_perf_reading, 0 }; + parse_fn = parse_fns[type]; + ASSERT (parse_fn); - switch (row) + switch (idx) { case 0: - s = format (s, "%f", - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_BAD_SPEC) + - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_RETIRING)); + sv = + parse_fn (ps, TOPDOWN_E_BAD_SPEC) + parse_fn (ps, TOPDOWN_E_RETIRING); break; case 1: - s = format (s, "%f", - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_BE_BOUND) + - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_FE_BOUND)); + sv = + parse_fn (ps, TOPDOWN_E_BE_BOUND) + parse_fn (ps, TOPDOWN_E_FE_BOUND); break; - case 2: - s = format (s, "%f", - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_RETIRING)); - break; - case 3: - s = format (s, "%f", - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_BAD_SPEC)); - break; - case 4: - s = format (s, "%f", - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_FE_BOUND)); - break; - case 5: - s = format (s, "%f", - topdown_lvl1_parse_row (st, TOPDOWN_E_METRIC_BE_BOUND)); + default: + sv = parse_fn (ps, (topdown_lvl1_t) idx - 2); break; } + + s = format (s, "%f", sv); + return s; } -PERFMON_REGISTER_BUNDLE (topdown_lvl1) = { +static perfmon_cpu_supports_t topdown_lvl1_cpu_supports[] = { + /* Intel SNR supports papi/thread only */ + { clib_cpu_supports_movdiri, PERFMON_BUNDLE_TYPE_THREAD }, + /* Intel ICX supports papi/thread or rdpmc/node */ + { clib_cpu_supports_avx512_bitalg, PERFMON_BUNDLE_TYPE_NODE } +}; + +PERFMON_REGISTER_BUNDLE (topdown_lvl1_metric) = { .name = "topdown-level1", .description = "Top-down Microarchitecture Analysis Level 1", .source = "intel-core", - .type = PERFMON_BUNDLE_TYPE_NODE, .offset_type = PERFMON_OFFSET_TYPE_METRICS, .events[0] = INTEL_CORE_E_TOPDOWN_SLOTS, - .events[1] = INTEL_CORE_E_TOPDOWN_L1_METRICS, + .events[1] = INTEL_CORE_E_TOPDOWN_L1_RETIRING_METRIC, + .events[2] = INTEL_CORE_E_TOPDOWN_L1_BAD_SPEC_METRIC, + .events[3] = INTEL_CORE_E_TOPDOWN_L1_FE_BOUND_METRIC, + .events[4] = INTEL_CORE_E_TOPDOWN_L1_BE_BOUND_METRIC, + .n_events = 5, .metrics[0] = RDPMC_FIXED_SLOTS | FIXED_COUNTER_SLOTS, .metrics[1] = RDPMC_L1_METRICS | METRIC_COUNTER_TOPDOWN_L1, - .n_events = 2, - .cpu_supports = clib_cpu_supports_avx512_bitalg, + .n_metrics = 2, + .cpu_supports = topdown_lvl1_cpu_supports, + .n_cpu_supports = ARRAY_LEN (topdown_lvl1_cpu_supports), .format_fn = format_topdown_lvl1, .column_headers = PERFMON_STRINGS ("% NS", "% ST", "% NS.RT", "% NS.BS", "% ST.FE", "% ST.BE"), diff --git a/src/plugins/perfmon/intel/core.h b/src/plugins/perfmon/intel/core.h index cd5c31ba260..a6a5269d772 100644 --- a/src/plugins/perfmon/intel/core.h +++ b/src/plugins/perfmon/intel/core.h @@ -29,8 +29,14 @@ "Reference cycles when the core is not in halt state.") \ _ (0x00, 0x04, 0, 0, 0, 0x00, TOPDOWN, SLOTS, \ "TMA slots available for an unhalted logical processor.") \ - _ (0x00, 0x80, 0, 0, 0, 0x00, TOPDOWN, L1_METRICS, \ - "TMA slots metrics for an unhalted logical processor.") \ + _ (0x00, 0x80, 0, 0, 0, 0x00, TOPDOWN, L1_RETIRING_METRIC, \ + "TMA retiring slots for an unhalted logical processor.") \ + _ (0x00, 0x81, 0, 0, 0, 0x00, TOPDOWN, L1_BAD_SPEC_METRIC, \ + "TMA bad spec slots or an unhalted logical processor.") \ + _ (0x00, 0x82, 0, 0, 0, 0x00, TOPDOWN, L1_FE_BOUND_METRIC, \ + "TMA fe bound slots for an unhalted logical processor.") \ + _ (0x00, 0x83, 0, 0, 0, 0x00, TOPDOWN, L1_BE_BOUND_METRIC, \ + "TMA be bound slots for an unhalted logical processor.") \ _ (0x03, 0x02, 0, 0, 0, 0x00, LD_BLOCKS, STORE_FORWARD, \ "Loads blocked due to overlapping with a preceding store that cannot be" \ " forwarded.") \ diff --git a/src/plugins/perfmon/perfmon.h b/src/plugins/perfmon/perfmon.h index bba22cf6b1d..967b92e0e4a 100644 --- a/src/plugins/perfmon/perfmon.h +++ b/src/plugins/perfmon/perfmon.h @@ -31,6 +31,7 @@ typedef enum PERFMON_BUNDLE_TYPE_NODE, PERFMON_BUNDLE_TYPE_THREAD, PERFMON_BUNDLE_TYPE_SYSTEM, + PERFMON_BUNDLE_TYPE_MAX, } perfmon_bundle_type_t; typedef enum @@ -95,6 +96,12 @@ struct perfmon_bundle; typedef clib_error_t *(perfmon_bundle_init_fn_t) (vlib_main_t *vm, struct perfmon_bundle *); +typedef struct +{ + clib_cpu_supports_func_t cpu_supports; + perfmon_bundle_type_t bundle_type; +} perfmon_cpu_supports_t; + typedef struct perfmon_bundle { char *name; @@ -104,14 +111,18 @@ typedef struct perfmon_bundle perfmon_bundle_type_t type; perfmon_offset_type_t offset_type; u32 events[PERF_MAX_EVENTS]; - u32 metrics[PERF_MAX_EVENTS]; u32 n_events; + u32 metrics[PERF_MAX_EVENTS]; + u32 n_metrics; + + perfmon_cpu_supports_t *cpu_supports; + u32 n_cpu_supports; + perfmon_bundle_init_fn_t *init_fn; char **column_headers; format_function_t *format_fn; - clib_cpu_supports_func_t cpu_supports; /* do not set manually */ perfmon_source_t *src; @@ -168,6 +179,24 @@ typedef struct extern perfmon_main_t perfmon_main; +always_inline uword +perfmon_cpu_supported_bundle_type (perfmon_bundle_t *b) +{ + perfmon_cpu_supports_t *supports = b->cpu_supports; + uword type = 0; + + /* if nothing specific for this bundle, go with the default */ + if (!supports) + return b->type; + + /* the last specified type, will always win */ + for (int i = 0; i < b->n_cpu_supports; ++i) + if (supports[i].cpu_supports ()) + type = supports[i].bundle_type; + + return type; +} + #define PERFMON_REGISTER_SOURCE(x) \ perfmon_source_t __perfmon_source_##x; \ static void __clib_constructor __perfmon_source_registration_##x (void) \ @@ -184,6 +213,8 @@ extern perfmon_main_t perfmon_main; { \ perfmon_main_t *pm = &perfmon_main; \ __perfmon_bundle_##x.next = pm->bundles; \ + __perfmon_bundle_##x.type = \ + perfmon_cpu_supported_bundle_type (&__perfmon_bundle_##x); \ pm->bundles = &__perfmon_bundle_##x; \ } \ perfmon_bundle_t __perfmon_bundle_##x -- cgit 1.2.3-korg