projects/linux·blob master · refs

306 lines · 8.1 KiB

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
17struct value {
18 const char *event;
19 u64 val;
20};
21
22static 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
34static 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 evsel->stats->aggr->counts.ena = 1;
45 evsel->stats->aggr->counts.run = 1;
46 }
47}
48
49static double compute_single(struct evlist *evlist, const char *name)
50{
51 struct metric_expr *mexp;
52 struct metric_event *me;
53 struct evsel *evsel;
54
55 evlist__for_each_entry(evlist, evsel) {
56 me = metricgroup__lookup(evlist__metric_events(evlist), evsel, false);
57 if (me != NULL) {
58 list_for_each_entry (mexp, &me->head, nd) {
59 if (strcmp(mexp->metric_name, name))
60 continue;
61 return test_generic_metric(mexp, 0);
62 }
63 }
64 }
65 return 0.;
66}
67
68static int __compute_metric(const char *name, struct value *vals,
69 const char *name1, double *ratio1,
70 const char *name2, double *ratio2)
71{
72 const struct pmu_metrics_table *pme_test;
73 struct perf_cpu_map *cpus;
74 struct evlist *evlist;
75 int err;
76
77 /*
78 * We need to prepare evlist for stat mode running on CPU 0
79 * because that's where all the stats are going to be created.
80 */
81 evlist = evlist__new();
82 if (!evlist)
83 return -ENOMEM;
84
85 cpus = perf_cpu_map__new("0");
86 if (!cpus) {
87 evlist__put(evlist);
88 return -ENOMEM;
89 }
90
91 perf_evlist__set_maps(evlist__core(evlist), cpus, NULL);
92
93 /* Parse the metric into metric_events list. */
94 pme_test = find_core_metrics_table("testarch", "testcpu");
95 err = metricgroup__parse_groups_test(evlist, pme_test, name,
96 /*cputype_filter=*/false);
97 if (err)
98 goto out;
99
100 err = evlist__alloc_stats(/*config=*/NULL, evlist, /*alloc_raw=*/false);
101 if (err)
102 goto out;
103
104 /* Load the runtime stats with given numbers for events. */
105 load_runtime_stat(evlist, vals);
106
107 /* And execute the metric */
108 if (name1 && ratio1)
109 *ratio1 = compute_single(evlist, name1);
110 if (name2 && ratio2)
111 *ratio2 = compute_single(evlist, name2);
112
113out:
114 /* ... cleanup. */
115 evlist__free_stats(evlist);
116 perf_cpu_map__put(cpus);
117 evlist__put(evlist);
118 return err;
119}
120
121static int compute_metric(const char *name, struct value *vals, double *ratio)
122{
123 return __compute_metric(name, vals, name, ratio, NULL, NULL);
124}
125
126static int compute_metric_group(const char *name, struct value *vals,
127 const char *name1, double *ratio1,
128 const char *name2, double *ratio2)
129{
130 return __compute_metric(name, vals, name1, ratio1, name2, ratio2);
131}
132
133static int test_ipc(void)
134{
135 double ratio;
136 struct value vals[] = {
137 { .event = "inst_retired.any", .val = 300 },
138 { .event = "cpu_clk_unhalted.thread", .val = 200 },
139 { .event = NULL, },
140 };
141
142 TEST_ASSERT_VAL("failed to compute metric",
143 compute_metric("IPC", vals, &ratio) == 0);
144
145 TEST_ASSERT_VAL("IPC failed, wrong ratio",
146 ratio == 1.5);
147 return 0;
148}
149
150static int test_frontend(void)
151{
152 double ratio;
153 struct value vals[] = {
154 { .event = "idq_uops_not_delivered.core", .val = 300 },
155 { .event = "cpu_clk_unhalted.thread", .val = 200 },
156 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 },
157 { .event = "cpu_clk_unhalted.ref_xclk", .val = 600 },
158 { .event = NULL, },
159 };
160
161 TEST_ASSERT_VAL("failed to compute metric",
162 compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0);
163
164 TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio",
165 ratio == 0.45);
166 return 0;
167}
168
169static int test_cache_miss_cycles(void)
170{
171 double ratio;
172 struct value vals[] = {
173 { .event = "l1d-loads-misses", .val = 300 },
174 { .event = "l1i-loads-misses", .val = 200 },
175 { .event = "inst_retired.any", .val = 400 },
176 { .event = NULL, },
177 };
178
179 TEST_ASSERT_VAL("failed to compute metric",
180 compute_metric("cache_miss_cycles", vals, &ratio) == 0);
181
182 TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio",
183 ratio == 1.25);
184 return 0;
185}
186
187
188/*
189 * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi
190 * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) +
191 * l2_rqsts.pf_miss + l2_rqsts.rfo_miss
192 * DCache_L2_All = dcache_l2_all_hits + dcache_l2_all_miss
193 * DCache_L2_Hits = d_ratio(dcache_l2_all_hits, dcache_l2_all)
194 * DCache_L2_Misses = d_ratio(dcache_l2_all_miss, dcache_l2_all)
195 *
196 * l2_rqsts.demand_data_rd_hit = 100
197 * l2_rqsts.pf_hit = 200
198 * l2_rqsts.rfo_hi = 300
199 * l2_rqsts.all_demand_data_rd = 400
200 * l2_rqsts.pf_miss = 500
201 * l2_rqsts.rfo_miss = 600
202 *
203 * DCache_L2_All_Hits = 600
204 * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400
205 * DCache_L2_All = 600 + 1400 = 2000
206 * DCache_L2_Hits = 600 / 2000 = 0.3
207 * DCache_L2_Misses = 1400 / 2000 = 0.7
208 */
209static int test_dcache_l2(void)
210{
211 double ratio;
212 struct value vals[] = {
213 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 },
214 { .event = "l2_rqsts.pf_hit", .val = 200 },
215 { .event = "l2_rqsts.rfo_hit", .val = 300 },
216 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 },
217 { .event = "l2_rqsts.pf_miss", .val = 500 },
218 { .event = "l2_rqsts.rfo_miss", .val = 600 },
219 { .event = NULL, },
220 };
221
222 TEST_ASSERT_VAL("failed to compute metric",
223 compute_metric("DCache_L2_Hits", vals, &ratio) == 0);
224
225 TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio",
226 ratio == 0.3);
227
228 TEST_ASSERT_VAL("failed to compute metric",
229 compute_metric("DCache_L2_Misses", vals, &ratio) == 0);
230
231 TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio",
232 ratio == 0.7);
233 return 0;
234}
235
236static int test_recursion_fail(void)
237{
238 double ratio;
239 struct value vals[] = {
240 { .event = "inst_retired.any", .val = 300 },
241 { .event = "cpu_clk_unhalted.thread", .val = 200 },
242 { .event = NULL, },
243 };
244
245 TEST_ASSERT_VAL("failed to find recursion",
246 compute_metric("M1", vals, &ratio) == -1);
247
248 TEST_ASSERT_VAL("failed to find recursion",
249 compute_metric("M3", vals, &ratio) == -1);
250 return 0;
251}
252
253static int test_memory_bandwidth(void)
254{
255 double ratio;
256 struct value vals[] = {
257 { .event = "l1d.replacement", .val = 4000000 },
258 { .event = "duration_time", .val = 200000000 },
259 { .event = NULL, },
260 };
261
262 TEST_ASSERT_VAL("failed to compute metric",
263 compute_metric("L1D_Cache_Fill_BW", vals, &ratio) == 0);
264 TEST_ASSERT_VAL("L1D_Cache_Fill_BW, wrong ratio",
265 1.28 == ratio);
266
267 return 0;
268}
269
270static int test_metric_group(void)
271{
272 double ratio1, ratio2;
273 struct value vals[] = {
274 { .event = "cpu_clk_unhalted.thread", .val = 200 },
275 { .event = "l1d-loads-misses", .val = 300 },
276 { .event = "l1i-loads-misses", .val = 200 },
277 { .event = "inst_retired.any", .val = 400 },
278 { .event = NULL, },
279 };
280
281 TEST_ASSERT_VAL("failed to find recursion",
282 compute_metric_group("group1", vals,
283 "IPC", &ratio1,
284 "cache_miss_cycles", &ratio2) == 0);
285
286 TEST_ASSERT_VAL("group IPC failed, wrong ratio",
287 ratio1 == 2.0);
288
289 TEST_ASSERT_VAL("group cache_miss_cycles failed, wrong ratio",
290 ratio2 == 1.25);
291 return 0;
292}
293
294static int test__parse_metric(struct test_suite *test __maybe_unused, int subtest __maybe_unused)
295{
296 TEST_ASSERT_VAL("IPC failed", test_ipc() == 0);
297 TEST_ASSERT_VAL("frontend failed", test_frontend() == 0);
298 TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0);
299 TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0);
300 TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0);
301 TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0);
302 TEST_ASSERT_VAL("test metric group", test_metric_group() == 0);
303 return 0;
304}
305
306DEFINE_SUITE("Parse and process metrics", parse_metric);