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