]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - tools/perf/util/stat.c
powerpc/dma: dma_set_coherent_mask() should not be GPL only
[karo-tx-linux.git] / tools / perf / util / stat.c
1 #include <math.h>
2 #include "stat.h"
3 #include "evlist.h"
4 #include "evsel.h"
5 #include "thread_map.h"
6
7 void update_stats(struct stats *stats, u64 val)
8 {
9         double delta;
10
11         stats->n++;
12         delta = val - stats->mean;
13         stats->mean += delta / stats->n;
14         stats->M2 += delta*(val - stats->mean);
15
16         if (val > stats->max)
17                 stats->max = val;
18
19         if (val < stats->min)
20                 stats->min = val;
21 }
22
23 double avg_stats(struct stats *stats)
24 {
25         return stats->mean;
26 }
27
28 /*
29  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
30  *
31  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
32  * s^2 = -------------------------------
33  *                  n - 1
34  *
35  * http://en.wikipedia.org/wiki/Stddev
36  *
37  * The std dev of the mean is related to the std dev by:
38  *
39  *             s
40  * s_mean = -------
41  *          sqrt(n)
42  *
43  */
44 double stddev_stats(struct stats *stats)
45 {
46         double variance, variance_mean;
47
48         if (stats->n < 2)
49                 return 0.0;
50
51         variance = stats->M2 / (stats->n - 1);
52         variance_mean = variance / stats->n;
53
54         return sqrt(variance_mean);
55 }
56
57 double rel_stddev_stats(double stddev, double avg)
58 {
59         double pct = 0.0;
60
61         if (avg)
62                 pct = 100.0 * stddev/avg;
63
64         return pct;
65 }
66
67 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
68                            enum perf_stat_evsel_id id)
69 {
70         struct perf_stat *ps = evsel->priv;
71
72         return ps->id == id;
73 }
74
75 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
76 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
77         ID(NONE,                x),
78         ID(CYCLES_IN_TX,        cpu/cycles-t/),
79         ID(TRANSACTION_START,   cpu/tx-start/),
80         ID(ELISION_START,       cpu/el-start/),
81         ID(CYCLES_IN_TX_CP,     cpu/cycles-ct/),
82 };
83 #undef ID
84
85 void perf_stat_evsel_id_init(struct perf_evsel *evsel)
86 {
87         struct perf_stat *ps = evsel->priv;
88         int i;
89
90         /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
91
92         for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
93                 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
94                         ps->id = i;
95                         break;
96                 }
97         }
98 }
99
100 void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
101 {
102         int i;
103         struct perf_stat *ps = evsel->priv;
104
105         for (i = 0; i < 3; i++)
106                 init_stats(&ps->res_stats[i]);
107
108         perf_stat_evsel_id_init(evsel);
109 }
110
111 int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
112 {
113         evsel->priv = zalloc(sizeof(struct perf_stat));
114         if (evsel->priv == NULL)
115                 return -ENOMEM;
116         perf_evsel__reset_stat_priv(evsel);
117         return 0;
118 }
119
120 void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
121 {
122         zfree(&evsel->priv);
123 }
124
125 int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
126                                       int ncpus, int nthreads)
127 {
128         struct perf_counts *counts;
129
130         counts = perf_counts__new(ncpus, nthreads);
131         if (counts)
132                 evsel->prev_raw_counts = counts;
133
134         return counts ? 0 : -ENOMEM;
135 }
136
137 void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
138 {
139         perf_counts__delete(evsel->prev_raw_counts);
140         evsel->prev_raw_counts = NULL;
141 }
142
143 int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
144 {
145         int ncpus = perf_evsel__nr_cpus(evsel);
146         int nthreads = thread_map__nr(evsel->threads);
147
148         if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
149             perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
150             (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
151                 return -ENOMEM;
152
153         return 0;
154 }
155
156 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
157 {
158         struct perf_evsel *evsel;
159
160         evlist__for_each(evlist, evsel) {
161                 if (perf_evsel__alloc_stats(evsel, alloc_raw))
162                         goto out_free;
163         }
164
165         return 0;
166
167 out_free:
168         perf_evlist__free_stats(evlist);
169         return -1;
170 }
171
172 void perf_evlist__free_stats(struct perf_evlist *evlist)
173 {
174         struct perf_evsel *evsel;
175
176         evlist__for_each(evlist, evsel) {
177                 perf_evsel__free_stat_priv(evsel);
178                 perf_evsel__free_counts(evsel);
179                 perf_evsel__free_prev_raw_counts(evsel);
180         }
181 }
182
183 void perf_evlist__reset_stats(struct perf_evlist *evlist)
184 {
185         struct perf_evsel *evsel;
186
187         evlist__for_each(evlist, evsel) {
188                 perf_evsel__reset_stat_priv(evsel);
189                 perf_evsel__reset_counts(evsel);
190         }
191 }
192
193 static void zero_per_pkg(struct perf_evsel *counter)
194 {
195         if (counter->per_pkg_mask)
196                 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
197 }
198
199 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
200 {
201         unsigned long *mask = counter->per_pkg_mask;
202         struct cpu_map *cpus = perf_evsel__cpus(counter);
203         int s;
204
205         *skip = false;
206
207         if (!counter->per_pkg)
208                 return 0;
209
210         if (cpu_map__empty(cpus))
211                 return 0;
212
213         if (!mask) {
214                 mask = zalloc(MAX_NR_CPUS);
215                 if (!mask)
216                         return -ENOMEM;
217
218                 counter->per_pkg_mask = mask;
219         }
220
221         s = cpu_map__get_socket(cpus, cpu);
222         if (s < 0)
223                 return -1;
224
225         *skip = test_and_set_bit(s, mask) == 1;
226         return 0;
227 }
228
229 static int
230 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
231                        int cpu, int thread,
232                        struct perf_counts_values *count)
233 {
234         struct perf_counts_values *aggr = &evsel->counts->aggr;
235         static struct perf_counts_values zero;
236         bool skip = false;
237
238         if (check_per_pkg(evsel, cpu, &skip)) {
239                 pr_err("failed to read per-pkg counter\n");
240                 return -1;
241         }
242
243         if (skip)
244                 count = &zero;
245
246         switch (config->aggr_mode) {
247         case AGGR_THREAD:
248         case AGGR_CORE:
249         case AGGR_SOCKET:
250         case AGGR_NONE:
251                 if (!evsel->snapshot)
252                         perf_evsel__compute_deltas(evsel, cpu, thread, count);
253                 perf_counts_values__scale(count, config->scale, NULL);
254                 if (config->aggr_mode == AGGR_NONE)
255                         perf_stat__update_shadow_stats(evsel, count->values, cpu);
256                 break;
257         case AGGR_GLOBAL:
258                 aggr->val += count->val;
259                 if (config->scale) {
260                         aggr->ena += count->ena;
261                         aggr->run += count->run;
262                 }
263         default:
264                 break;
265         }
266
267         return 0;
268 }
269
270 static int process_counter_maps(struct perf_stat_config *config,
271                                 struct perf_evsel *counter)
272 {
273         int nthreads = thread_map__nr(counter->threads);
274         int ncpus = perf_evsel__nr_cpus(counter);
275         int cpu, thread;
276
277         if (counter->system_wide)
278                 nthreads = 1;
279
280         for (thread = 0; thread < nthreads; thread++) {
281                 for (cpu = 0; cpu < ncpus; cpu++) {
282                         if (process_counter_values(config, counter, cpu, thread,
283                                                    perf_counts(counter->counts, cpu, thread)))
284                                 return -1;
285                 }
286         }
287
288         return 0;
289 }
290
291 int perf_stat_process_counter(struct perf_stat_config *config,
292                               struct perf_evsel *counter)
293 {
294         struct perf_counts_values *aggr = &counter->counts->aggr;
295         struct perf_stat *ps = counter->priv;
296         u64 *count = counter->counts->aggr.values;
297         int i, ret;
298
299         aggr->val = aggr->ena = aggr->run = 0;
300         init_stats(ps->res_stats);
301
302         if (counter->per_pkg)
303                 zero_per_pkg(counter);
304
305         ret = process_counter_maps(config, counter);
306         if (ret)
307                 return ret;
308
309         if (config->aggr_mode != AGGR_GLOBAL)
310                 return 0;
311
312         if (!counter->snapshot)
313                 perf_evsel__compute_deltas(counter, -1, -1, aggr);
314         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
315
316         for (i = 0; i < 3; i++)
317                 update_stats(&ps->res_stats[i], count[i]);
318
319         if (verbose) {
320                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
321                         perf_evsel__name(counter), count[0], count[1], count[2]);
322         }
323
324         /*
325          * Save the full runtime - to allow normalization during printout:
326          */
327         perf_stat__update_shadow_stats(counter, count, 0);
328
329         return 0;
330 }