tools/perf/util/evsel.c

   1 #include "evsel.h"
   2 #include "../perf.h"
   3 #include "util.h"
   4 #include "cpumap.h"
   5 #include "thread.h"
   6
   7 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
   8
   9 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
  10 {
  11         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
  12
  13         if (evsel != NULL) {
  14                 evsel->idx         = idx;
  15                 evsel->attr        = *attr;
  16                 INIT_LIST_HEAD(&evsel->node);
  17         }
  18
  19         return evsel;
  20 }
  21
  22 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
  23 {
  24         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
  25         return evsel->fd != NULL ? 0 : -ENOMEM;
  26 }
  27
  28 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
  29 {
  30         evsel->counts = zalloc((sizeof(*evsel->counts) +
  31                                 (ncpus * sizeof(struct perf_counts_values))));
  32         return evsel->counts != NULL ? 0 : -ENOMEM;
  33 }
  34
  35 void perf_evsel__free_fd(struct perf_evsel *evsel)
  36 {
  37         xyarray__delete(evsel->fd);
  38         evsel->fd = NULL;
  39 }
  40
  41 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
  42 {
  43         int cpu, thread;
  44
  45         for (cpu = 0; cpu < ncpus; cpu++)
  46                 for (thread = 0; thread < nthreads; ++thread) {
  47                         close(FD(evsel, cpu, thread));
  48                         FD(evsel, cpu, thread) = -1;
  49                 }
  50 }
  51
  52 void perf_evsel__delete(struct perf_evsel *evsel)
  53 {
  54         assert(list_empty(&evsel->node));
  55         xyarray__delete(evsel->fd);
  56         free(evsel);
  57 }
  58
  59 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
  60                               int cpu, int thread, bool scale)
  61 {
  62         struct perf_counts_values count;
  63         size_t nv = scale ? 3 : 1;
  64
  65         if (FD(evsel, cpu, thread) < 0)
  66                 return -EINVAL;
  67
  68         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
  69                 return -ENOMEM;
  70
  71         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
  72                 return -errno;
  73
  74         if (scale) {
  75                 if (count.run == 0)
  76                         count.val = 0;
  77                 else if (count.run < count.ena)
  78                         count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
  79         } else
  80                 count.ena = count.run = 0;
  81
  82         evsel->counts->cpu[cpu] = count;
  83         return 0;
  84 }
  85
  86 int __perf_evsel__read(struct perf_evsel *evsel,
  87                        int ncpus, int nthreads, bool scale)
  88 {
  89         size_t nv = scale ? 3 : 1;
  90         int cpu, thread;
  91         struct perf_counts_values *aggr = &evsel->counts->aggr, count;
  92
  93         aggr->val = aggr->ena = aggr->run = 0;
  94
  95         for (cpu = 0; cpu < ncpus; cpu++) {
  96                 for (thread = 0; thread < nthreads; thread++) {
  97                         if (FD(evsel, cpu, thread) < 0)
  98                                 continue;
  99
 100                         if (readn(FD(evsel, cpu, thread),
 101                                   &count, nv * sizeof(u64)) < 0)
 102                                 return -errno;
 103
 104                         aggr->val += count.val;
 105                         if (scale) {
 106                                 aggr->ena += count.ena;
 107                                 aggr->run += count.run;
 108                         }
 109                 }
 110         }
 111
 112         evsel->counts->scaled = 0;
 113         if (scale) {
 114                 if (aggr->run == 0) {
 115                         evsel->counts->scaled = -1;
 116                         aggr->val = 0;
 117                         return 0;
 118                 }
 119
 120                 if (aggr->run < aggr->ena) {
 121                         evsel->counts->scaled = 1;
 122                         aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
 123                 }
 124         } else
 125                 aggr->ena = aggr->run = 0;
 126
 127         return 0;
 128 }
 129
 130 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 131                               struct thread_map *threads)
 132 {
 133         int cpu, thread;
 134
 135         if (evsel->fd == NULL &&
 136             perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
 137                 return -1;
 138
 139         for (cpu = 0; cpu < cpus->nr; cpu++) {
 140                 for (thread = 0; thread < threads->nr; thread++) {
 141                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
 142                                                                      threads->map[thread],
 143                                                                      cpus->map[cpu], -1, 0);
 144                         if (FD(evsel, cpu, thread) < 0)
 145                                 goto out_close;
 146                 }
 147         }
 148
 149         return 0;
 150
 151 out_close:
 152         do {
 153                 while (--thread >= 0) {
 154                         close(FD(evsel, cpu, thread));
 155                         FD(evsel, cpu, thread) = -1;
 156                 }
 157                 thread = threads->nr;
 158         } while (--cpu >= 0);
 159         return -1;
 160 }
 161
 162 static struct {
 163         struct cpu_map map;
 164         int cpus[1];
 165 } empty_cpu_map = {
 166         .map.nr = 1,
 167         .cpus   = { -1, },
 168 };
 169
 170 static struct {
 171         struct thread_map map;
 172         int threads[1];
 173 } empty_thread_map = {
 174         .map.nr  = 1,
 175         .threads = { -1, },
 176 };
 177
 178 int perf_evsel__open(struct perf_evsel *evsel,
 179                      struct cpu_map *cpus, struct thread_map *threads)
 180 {
 181
 182         if (cpus == NULL) {
 183                 /* Work around old compiler warnings about strict aliasing */
 184                 cpus = &empty_cpu_map.map;
 185         }
 186
 187         if (threads == NULL)
 188                 threads = &empty_thread_map.map;
 189
 190         return __perf_evsel__open(evsel, cpus, threads);
 191 }
 192
 193 int perf_evsel__open_per_cpu(struct perf_evsel *evsel, struct cpu_map *cpus)
 194 {
 195         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
 196 }
 197
 198 int perf_evsel__open_per_thread(struct perf_evsel *evsel, struct thread_map *threads)
 199 {
 200         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
 201 }