1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/compiler.h>
3 #include <string.h>
4 #include <perf/cpumap.h>
5 #include <perf/evlist.h>
6 #include "metricgroup.h"
7 #include "tests.h"
8 #include "pmu-events/pmu-events.h"
9 #include "evlist.h"
10 #include "rblist.h"
11 #include "debug.h"
12 #include "expr.h"
13 #include "stat.h"
14 #include "pmus.h"
15
16 struct value {
17 const char *event;
18 u64 val;
19 };
20
find_value(const char * name,struct value * values)21 static u64 find_value(const char *name, struct value *values)
22 {
23 struct value *v = values;
24
25 while (v->event) {
26 if (!strcmp(name, v->event))
27 return v->val;
28 v++;
29 }
30 return 0;
31 }
32
load_runtime_stat(struct evlist * evlist,struct value * vals)33 static void load_runtime_stat(struct evlist *evlist, struct value *vals)
34 {
35 struct evsel *evsel;
36 u64 count;
37
38 evlist__alloc_aggr_stats(evlist, 1);
39 evlist__for_each_entry(evlist, evsel) {
40 count = find_value(evsel->name, vals);
41 evsel->supported = true;
42 evsel->stats->aggr->counts.val = count;
43 if (evsel__name_is(evsel, "duration_time"))
44 update_stats(&walltime_nsecs_stats, count);
45 }
46 }
47
compute_single(struct evlist * evlist,const char * name)48 static double compute_single(struct evlist *evlist, const char *name)
49 {
50 struct metric_expr *mexp;
51 struct metric_event *me;
52 struct evsel *evsel;
53
54 evlist__for_each_entry(evlist, evsel) {
55 me = metricgroup__lookup(&evlist->metric_events, evsel, false);
56 if (me != NULL) {
57 list_for_each_entry (mexp, &me->head, nd) {
58 if (strcmp(mexp->metric_name, name))
59 continue;
60 return test_generic_metric(mexp, 0);
61 }
62 }
63 }
64 return 0.;
65 }
66
__compute_metric(const char * name,struct value * vals,const char * name1,double * ratio1,const char * name2,double * ratio2)67 static int __compute_metric(const char *name, struct value *vals,
68 const char *name1, double *ratio1,
69 const char *name2, double *ratio2)
70 {
71 const struct pmu_metrics_table *pme_test;
72 struct perf_cpu_map *cpus;
73 struct evlist *evlist;
74 int err;
75
76 /*
77 * We need to prepare evlist for stat mode running on CPU 0
78 * because that's where all the stats are going to be created.
79 */
80 evlist = evlist__new();
81 if (!evlist)
82 return -ENOMEM;
83
84 cpus = perf_cpu_map__new("0");
85 if (!cpus) {
86 evlist__delete(evlist);
87 return -ENOMEM;
88 }
89
90 perf_evlist__set_maps(&evlist->core, cpus, NULL);
91
92 /* Parse the metric into metric_events list. */
93 pme_test = find_core_metrics_table("testarch", "testcpu");
94 err = metricgroup__parse_groups_test(evlist, pme_test, name);
95 if (err)
96 goto out;
97
98 err = evlist__alloc_stats(/*config=*/NULL, evlist, /*alloc_raw=*/false);
99 if (err)
100 goto out;
101
102 /* Load the runtime stats with given numbers for events. */
103 load_runtime_stat(evlist, vals);
104
105 /* And execute the metric */
106 if (name1 && ratio1)
107 *ratio1 = compute_single(evlist, name1);
108 if (name2 && ratio2)
109 *ratio2 = compute_single(evlist, name2);
110
111 out:
112 /* ... cleanup. */
113 evlist__free_stats(evlist);
114 perf_cpu_map__put(cpus);
115 evlist__delete(evlist);
116 return err;
117 }
118
compute_metric(const char * name,struct value * vals,double * ratio)119 static int compute_metric(const char *name, struct value *vals, double *ratio)
120 {
121 return __compute_metric(name, vals, name, ratio, NULL, NULL);
122 }
123
compute_metric_group(const char * name,struct value * vals,const char * name1,double * ratio1,const char * name2,double * ratio2)124 static int compute_metric_group(const char *name, struct value *vals,
125 const char *name1, double *ratio1,
126 const char *name2, double *ratio2)
127 {
128 return __compute_metric(name, vals, name1, ratio1, name2, ratio2);
129 }
130
test_ipc(void)131 static int test_ipc(void)
132 {
133 double ratio;
134 struct value vals[] = {
135 { .event = "inst_retired.any", .val = 300 },
136 { .event = "cpu_clk_unhalted.thread", .val = 200 },
137 { .event = NULL, },
138 };
139
140 TEST_ASSERT_VAL("failed to compute metric",
141 compute_metric("IPC", vals, &ratio) == 0);
142
143 TEST_ASSERT_VAL("IPC failed, wrong ratio",
144 ratio == 1.5);
145 return 0;
146 }
147
test_frontend(void)148 static int test_frontend(void)
149 {
150 double ratio;
151 struct value vals[] = {
152 { .event = "idq_uops_not_delivered.core", .val = 300 },
153 { .event = "cpu_clk_unhalted.thread", .val = 200 },
154 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 },
155 { .event = "cpu_clk_unhalted.ref_xclk", .val = 600 },
156 { .event = NULL, },
157 };
158
159 TEST_ASSERT_VAL("failed to compute metric",
160 compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0);
161
162 TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio",
163 ratio == 0.45);
164 return 0;
165 }
166
test_cache_miss_cycles(void)167 static int test_cache_miss_cycles(void)
168 {
169 double ratio;
170 struct value vals[] = {
171 { .event = "l1d-loads-misses", .val = 300 },
172 { .event = "l1i-loads-misses", .val = 200 },
173 { .event = "inst_retired.any", .val = 400 },
174 { .event = NULL, },
175 };
176
177 TEST_ASSERT_VAL("failed to compute metric",
178 compute_metric("cache_miss_cycles", vals, &ratio) == 0);
179
180 TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio",
181 ratio == 1.25);
182 return 0;
183 }
184
185
186 /*
187 * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi
188 * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) +
189 * l2_rqsts.pf_miss + l2_rqsts.rfo_miss
190 * DCache_L2_All = dcache_l2_all_hits + dcache_l2_all_miss
191 * DCache_L2_Hits = d_ratio(dcache_l2_all_hits, dcache_l2_all)
192 * DCache_L2_Misses = d_ratio(dcache_l2_all_miss, dcache_l2_all)
193 *
194 * l2_rqsts.demand_data_rd_hit = 100
195 * l2_rqsts.pf_hit = 200
196 * l2_rqsts.rfo_hi = 300
197 * l2_rqsts.all_demand_data_rd = 400
198 * l2_rqsts.pf_miss = 500
199 * l2_rqsts.rfo_miss = 600
200 *
201 * DCache_L2_All_Hits = 600
202 * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400
203 * DCache_L2_All = 600 + 1400 = 2000
204 * DCache_L2_Hits = 600 / 2000 = 0.3
205 * DCache_L2_Misses = 1400 / 2000 = 0.7
206 */
test_dcache_l2(void)207 static int test_dcache_l2(void)
208 {
209 double ratio;
210 struct value vals[] = {
211 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 },
212 { .event = "l2_rqsts.pf_hit", .val = 200 },
213 { .event = "l2_rqsts.rfo_hit", .val = 300 },
214 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 },
215 { .event = "l2_rqsts.pf_miss", .val = 500 },
216 { .event = "l2_rqsts.rfo_miss", .val = 600 },
217 { .event = NULL, },
218 };
219
220 TEST_ASSERT_VAL("failed to compute metric",
221 compute_metric("DCache_L2_Hits", vals, &ratio) == 0);
222
223 TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio",
224 ratio == 0.3);
225
226 TEST_ASSERT_VAL("failed to compute metric",
227 compute_metric("DCache_L2_Misses", vals, &ratio) == 0);
228
229 TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio",
230 ratio == 0.7);
231 return 0;
232 }
233
test_recursion_fail(void)234 static int test_recursion_fail(void)
235 {
236 double ratio;
237 struct value vals[] = {
238 { .event = "inst_retired.any", .val = 300 },
239 { .event = "cpu_clk_unhalted.thread", .val = 200 },
240 { .event = NULL, },
241 };
242
243 TEST_ASSERT_VAL("failed to find recursion",
244 compute_metric("M1", vals, &ratio) == -1);
245
246 TEST_ASSERT_VAL("failed to find recursion",
247 compute_metric("M3", vals, &ratio) == -1);
248 return 0;
249 }
250
test_memory_bandwidth(void)251 static int test_memory_bandwidth(void)
252 {
253 double ratio;
254 struct value vals[] = {
255 { .event = "l1d.replacement", .val = 4000000 },
256 { .event = "duration_time", .val = 200000000 },
257 { .event = NULL, },
258 };
259
260 TEST_ASSERT_VAL("failed to compute metric",
261 compute_metric("L1D_Cache_Fill_BW", vals, &ratio) == 0);
262 TEST_ASSERT_VAL("L1D_Cache_Fill_BW, wrong ratio",
263 1.28 == ratio);
264
265 return 0;
266 }
267
test_metric_group(void)268 static int test_metric_group(void)
269 {
270 double ratio1, ratio2;
271 struct value vals[] = {
272 { .event = "cpu_clk_unhalted.thread", .val = 200 },
273 { .event = "l1d-loads-misses", .val = 300 },
274 { .event = "l1i-loads-misses", .val = 200 },
275 { .event = "inst_retired.any", .val = 400 },
276 { .event = NULL, },
277 };
278
279 TEST_ASSERT_VAL("failed to find recursion",
280 compute_metric_group("group1", vals,
281 "IPC", &ratio1,
282 "cache_miss_cycles", &ratio2) == 0);
283
284 TEST_ASSERT_VAL("group IPC failed, wrong ratio",
285 ratio1 == 2.0);
286
287 TEST_ASSERT_VAL("group cache_miss_cycles failed, wrong ratio",
288 ratio2 == 1.25);
289 return 0;
290 }
291
test__parse_metric(struct test_suite * test __maybe_unused,int subtest __maybe_unused)292 static int test__parse_metric(struct test_suite *test __maybe_unused, int subtest __maybe_unused)
293 {
294 TEST_ASSERT_VAL("IPC failed", test_ipc() == 0);
295 TEST_ASSERT_VAL("frontend failed", test_frontend() == 0);
296 TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0);
297 TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0);
298 TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0);
299 TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0);
300 TEST_ASSERT_VAL("test metric group", test_metric_group() == 0);
301 return 0;
302 }
303
304 DEFINE_SUITE("Parse and process metrics", parse_metric);
305