pktgen: correctly handle failures when adding a device
[linux-2.6.git] / tools / perf / builtin-test.c
1 /*
2  * builtin-test.c
3  *
4  * Builtin regression testing command: ever growing number of sanity tests
5  */
6 #include "builtin.h"
7
8 #include "util/cache.h"
9 #include "util/debug.h"
10 #include "util/debugfs.h"
11 #include "util/evlist.h"
12 #include "util/parse-options.h"
13 #include "util/parse-events.h"
14 #include "util/symbol.h"
15 #include "util/thread_map.h"
16 #include "util/pmu.h"
17 #include "../../include/linux/hw_breakpoint.h"
18
19 #include <sys/mman.h>
20
21 static int vmlinux_matches_kallsyms_filter(struct map *map __used, struct symbol *sym)
22 {
23         bool *visited = symbol__priv(sym);
24         *visited = true;
25         return 0;
26 }
27
28 static int test__vmlinux_matches_kallsyms(void)
29 {
30         int err = -1;
31         struct rb_node *nd;
32         struct symbol *sym;
33         struct map *kallsyms_map, *vmlinux_map;
34         struct machine kallsyms, vmlinux;
35         enum map_type type = MAP__FUNCTION;
36         long page_size = sysconf(_SC_PAGE_SIZE);
37         struct ref_reloc_sym ref_reloc_sym = { .name = "_stext", };
38
39         /*
40          * Step 1:
41          *
42          * Init the machines that will hold kernel, modules obtained from
43          * both vmlinux + .ko files and from /proc/kallsyms split by modules.
44          */
45         machine__init(&kallsyms, "", HOST_KERNEL_ID);
46         machine__init(&vmlinux, "", HOST_KERNEL_ID);
47
48         /*
49          * Step 2:
50          *
51          * Create the kernel maps for kallsyms and the DSO where we will then
52          * load /proc/kallsyms. Also create the modules maps from /proc/modules
53          * and find the .ko files that match them in /lib/modules/`uname -r`/.
54          */
55         if (machine__create_kernel_maps(&kallsyms) < 0) {
56                 pr_debug("machine__create_kernel_maps ");
57                 return -1;
58         }
59
60         /*
61          * Step 3:
62          *
63          * Load and split /proc/kallsyms into multiple maps, one per module.
64          */
65         if (machine__load_kallsyms(&kallsyms, "/proc/kallsyms", type, NULL) <= 0) {
66                 pr_debug("dso__load_kallsyms ");
67                 goto out;
68         }
69
70         /*
71          * Step 4:
72          *
73          * kallsyms will be internally on demand sorted by name so that we can
74          * find the reference relocation * symbol, i.e. the symbol we will use
75          * to see if the running kernel was relocated by checking if it has the
76          * same value in the vmlinux file we load.
77          */
78         kallsyms_map = machine__kernel_map(&kallsyms, type);
79
80         sym = map__find_symbol_by_name(kallsyms_map, ref_reloc_sym.name, NULL);
81         if (sym == NULL) {
82                 pr_debug("dso__find_symbol_by_name ");
83                 goto out;
84         }
85
86         ref_reloc_sym.addr = sym->start;
87
88         /*
89          * Step 5:
90          *
91          * Now repeat step 2, this time for the vmlinux file we'll auto-locate.
92          */
93         if (machine__create_kernel_maps(&vmlinux) < 0) {
94                 pr_debug("machine__create_kernel_maps ");
95                 goto out;
96         }
97
98         vmlinux_map = machine__kernel_map(&vmlinux, type);
99         map__kmap(vmlinux_map)->ref_reloc_sym = &ref_reloc_sym;
100
101         /*
102          * Step 6:
103          *
104          * Locate a vmlinux file in the vmlinux path that has a buildid that
105          * matches the one of the running kernel.
106          *
107          * While doing that look if we find the ref reloc symbol, if we find it
108          * we'll have its ref_reloc_symbol.unrelocated_addr and then
109          * maps__reloc_vmlinux will notice and set proper ->[un]map_ip routines
110          * to fixup the symbols.
111          */
112         if (machine__load_vmlinux_path(&vmlinux, type,
113                                        vmlinux_matches_kallsyms_filter) <= 0) {
114                 pr_debug("machine__load_vmlinux_path ");
115                 goto out;
116         }
117
118         err = 0;
119         /*
120          * Step 7:
121          *
122          * Now look at the symbols in the vmlinux DSO and check if we find all of them
123          * in the kallsyms dso. For the ones that are in both, check its names and
124          * end addresses too.
125          */
126         for (nd = rb_first(&vmlinux_map->dso->symbols[type]); nd; nd = rb_next(nd)) {
127                 struct symbol *pair, *first_pair;
128                 bool backwards = true;
129
130                 sym  = rb_entry(nd, struct symbol, rb_node);
131
132                 if (sym->start == sym->end)
133                         continue;
134
135                 first_pair = machine__find_kernel_symbol(&kallsyms, type, sym->start, NULL, NULL);
136                 pair = first_pair;
137
138                 if (pair && pair->start == sym->start) {
139 next_pair:
140                         if (strcmp(sym->name, pair->name) == 0) {
141                                 /*
142                                  * kallsyms don't have the symbol end, so we
143                                  * set that by using the next symbol start - 1,
144                                  * in some cases we get this up to a page
145                                  * wrong, trace_kmalloc when I was developing
146                                  * this code was one such example, 2106 bytes
147                                  * off the real size. More than that and we
148                                  * _really_ have a problem.
149                                  */
150                                 s64 skew = sym->end - pair->end;
151                                 if (llabs(skew) < page_size)
152                                         continue;
153
154                                 pr_debug("%#" PRIx64 ": diff end addr for %s v: %#" PRIx64 " k: %#" PRIx64 "\n",
155                                          sym->start, sym->name, sym->end, pair->end);
156                         } else {
157                                 struct rb_node *nnd;
158 detour:
159                                 nnd = backwards ? rb_prev(&pair->rb_node) :
160                                                   rb_next(&pair->rb_node);
161                                 if (nnd) {
162                                         struct symbol *next = rb_entry(nnd, struct symbol, rb_node);
163
164                                         if (next->start == sym->start) {
165                                                 pair = next;
166                                                 goto next_pair;
167                                         }
168                                 }
169
170                                 if (backwards) {
171                                         backwards = false;
172                                         pair = first_pair;
173                                         goto detour;
174                                 }
175
176                                 pr_debug("%#" PRIx64 ": diff name v: %s k: %s\n",
177                                          sym->start, sym->name, pair->name);
178                         }
179                 } else
180                         pr_debug("%#" PRIx64 ": %s not on kallsyms\n", sym->start, sym->name);
181
182                 err = -1;
183         }
184
185         if (!verbose)
186                 goto out;
187
188         pr_info("Maps only in vmlinux:\n");
189
190         for (nd = rb_first(&vmlinux.kmaps.maps[type]); nd; nd = rb_next(nd)) {
191                 struct map *pos = rb_entry(nd, struct map, rb_node), *pair;
192                 /*
193                  * If it is the kernel, kallsyms is always "[kernel.kallsyms]", while
194                  * the kernel will have the path for the vmlinux file being used,
195                  * so use the short name, less descriptive but the same ("[kernel]" in
196                  * both cases.
197                  */
198                 pair = map_groups__find_by_name(&kallsyms.kmaps, type,
199                                                 (pos->dso->kernel ?
200                                                         pos->dso->short_name :
201                                                         pos->dso->name));
202                 if (pair)
203                         pair->priv = 1;
204                 else
205                         map__fprintf(pos, stderr);
206         }
207
208         pr_info("Maps in vmlinux with a different name in kallsyms:\n");
209
210         for (nd = rb_first(&vmlinux.kmaps.maps[type]); nd; nd = rb_next(nd)) {
211                 struct map *pos = rb_entry(nd, struct map, rb_node), *pair;
212
213                 pair = map_groups__find(&kallsyms.kmaps, type, pos->start);
214                 if (pair == NULL || pair->priv)
215                         continue;
216
217                 if (pair->start == pos->start) {
218                         pair->priv = 1;
219                         pr_info(" %" PRIx64 "-%" PRIx64 " %" PRIx64 " %s in kallsyms as",
220                                 pos->start, pos->end, pos->pgoff, pos->dso->name);
221                         if (pos->pgoff != pair->pgoff || pos->end != pair->end)
222                                 pr_info(": \n*%" PRIx64 "-%" PRIx64 " %" PRIx64 "",
223                                         pair->start, pair->end, pair->pgoff);
224                         pr_info(" %s\n", pair->dso->name);
225                         pair->priv = 1;
226                 }
227         }
228
229         pr_info("Maps only in kallsyms:\n");
230
231         for (nd = rb_first(&kallsyms.kmaps.maps[type]);
232              nd; nd = rb_next(nd)) {
233                 struct map *pos = rb_entry(nd, struct map, rb_node);
234
235                 if (!pos->priv)
236                         map__fprintf(pos, stderr);
237         }
238 out:
239         return err;
240 }
241
242 #include "util/cpumap.h"
243 #include "util/evsel.h"
244 #include <sys/types.h>
245
246 static int trace_event__id(const char *evname)
247 {
248         char *filename;
249         int err = -1, fd;
250
251         if (asprintf(&filename,
252                      "%s/syscalls/%s/id",
253                      tracing_events_path, evname) < 0)
254                 return -1;
255
256         fd = open(filename, O_RDONLY);
257         if (fd >= 0) {
258                 char id[16];
259                 if (read(fd, id, sizeof(id)) > 0)
260                         err = atoi(id);
261                 close(fd);
262         }
263
264         free(filename);
265         return err;
266 }
267
268 static int test__open_syscall_event(void)
269 {
270         int err = -1, fd;
271         struct thread_map *threads;
272         struct perf_evsel *evsel;
273         struct perf_event_attr attr;
274         unsigned int nr_open_calls = 111, i;
275         int id = trace_event__id("sys_enter_open");
276
277         if (id < 0) {
278                 pr_debug("is debugfs mounted on /sys/kernel/debug?\n");
279                 return -1;
280         }
281
282         threads = thread_map__new(-1, getpid(), UINT_MAX);
283         if (threads == NULL) {
284                 pr_debug("thread_map__new\n");
285                 return -1;
286         }
287
288         memset(&attr, 0, sizeof(attr));
289         attr.type = PERF_TYPE_TRACEPOINT;
290         attr.config = id;
291         evsel = perf_evsel__new(&attr, 0);
292         if (evsel == NULL) {
293                 pr_debug("perf_evsel__new\n");
294                 goto out_thread_map_delete;
295         }
296
297         if (perf_evsel__open_per_thread(evsel, threads, false, NULL) < 0) {
298                 pr_debug("failed to open counter: %s, "
299                          "tweak /proc/sys/kernel/perf_event_paranoid?\n",
300                          strerror(errno));
301                 goto out_evsel_delete;
302         }
303
304         for (i = 0; i < nr_open_calls; ++i) {
305                 fd = open("/etc/passwd", O_RDONLY);
306                 close(fd);
307         }
308
309         if (perf_evsel__read_on_cpu(evsel, 0, 0) < 0) {
310                 pr_debug("perf_evsel__read_on_cpu\n");
311                 goto out_close_fd;
312         }
313
314         if (evsel->counts->cpu[0].val != nr_open_calls) {
315                 pr_debug("perf_evsel__read_on_cpu: expected to intercept %d calls, got %" PRIu64 "\n",
316                          nr_open_calls, evsel->counts->cpu[0].val);
317                 goto out_close_fd;
318         }
319         
320         err = 0;
321 out_close_fd:
322         perf_evsel__close_fd(evsel, 1, threads->nr);
323 out_evsel_delete:
324         perf_evsel__delete(evsel);
325 out_thread_map_delete:
326         thread_map__delete(threads);
327         return err;
328 }
329
330 #include <sched.h>
331
332 static int test__open_syscall_event_on_all_cpus(void)
333 {
334         int err = -1, fd, cpu;
335         struct thread_map *threads;
336         struct cpu_map *cpus;
337         struct perf_evsel *evsel;
338         struct perf_event_attr attr;
339         unsigned int nr_open_calls = 111, i;
340         cpu_set_t cpu_set;
341         int id = trace_event__id("sys_enter_open");
342
343         if (id < 0) {
344                 pr_debug("is debugfs mounted on /sys/kernel/debug?\n");
345                 return -1;
346         }
347
348         threads = thread_map__new(-1, getpid(), UINT_MAX);
349         if (threads == NULL) {
350                 pr_debug("thread_map__new\n");
351                 return -1;
352         }
353
354         cpus = cpu_map__new(NULL);
355         if (cpus == NULL) {
356                 pr_debug("cpu_map__new\n");
357                 goto out_thread_map_delete;
358         }
359
360
361         CPU_ZERO(&cpu_set);
362
363         memset(&attr, 0, sizeof(attr));
364         attr.type = PERF_TYPE_TRACEPOINT;
365         attr.config = id;
366         evsel = perf_evsel__new(&attr, 0);
367         if (evsel == NULL) {
368                 pr_debug("perf_evsel__new\n");
369                 goto out_thread_map_delete;
370         }
371
372         if (perf_evsel__open(evsel, cpus, threads, false, NULL) < 0) {
373                 pr_debug("failed to open counter: %s, "
374                          "tweak /proc/sys/kernel/perf_event_paranoid?\n",
375                          strerror(errno));
376                 goto out_evsel_delete;
377         }
378
379         for (cpu = 0; cpu < cpus->nr; ++cpu) {
380                 unsigned int ncalls = nr_open_calls + cpu;
381                 /*
382                  * XXX eventually lift this restriction in a way that
383                  * keeps perf building on older glibc installations
384                  * without CPU_ALLOC. 1024 cpus in 2010 still seems
385                  * a reasonable upper limit tho :-)
386                  */
387                 if (cpus->map[cpu] >= CPU_SETSIZE) {
388                         pr_debug("Ignoring CPU %d\n", cpus->map[cpu]);
389                         continue;
390                 }
391
392                 CPU_SET(cpus->map[cpu], &cpu_set);
393                 if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) {
394                         pr_debug("sched_setaffinity() failed on CPU %d: %s ",
395                                  cpus->map[cpu],
396                                  strerror(errno));
397                         goto out_close_fd;
398                 }
399                 for (i = 0; i < ncalls; ++i) {
400                         fd = open("/etc/passwd", O_RDONLY);
401                         close(fd);
402                 }
403                 CPU_CLR(cpus->map[cpu], &cpu_set);
404         }
405
406         /*
407          * Here we need to explicitely preallocate the counts, as if
408          * we use the auto allocation it will allocate just for 1 cpu,
409          * as we start by cpu 0.
410          */
411         if (perf_evsel__alloc_counts(evsel, cpus->nr) < 0) {
412                 pr_debug("perf_evsel__alloc_counts(ncpus=%d)\n", cpus->nr);
413                 goto out_close_fd;
414         }
415
416         err = 0;
417
418         for (cpu = 0; cpu < cpus->nr; ++cpu) {
419                 unsigned int expected;
420
421                 if (cpus->map[cpu] >= CPU_SETSIZE)
422                         continue;
423
424                 if (perf_evsel__read_on_cpu(evsel, cpu, 0) < 0) {
425                         pr_debug("perf_evsel__read_on_cpu\n");
426                         err = -1;
427                         break;
428                 }
429
430                 expected = nr_open_calls + cpu;
431                 if (evsel->counts->cpu[cpu].val != expected) {
432                         pr_debug("perf_evsel__read_on_cpu: expected to intercept %d calls on cpu %d, got %" PRIu64 "\n",
433                                  expected, cpus->map[cpu], evsel->counts->cpu[cpu].val);
434                         err = -1;
435                 }
436         }
437
438 out_close_fd:
439         perf_evsel__close_fd(evsel, 1, threads->nr);
440 out_evsel_delete:
441         perf_evsel__delete(evsel);
442 out_thread_map_delete:
443         thread_map__delete(threads);
444         return err;
445 }
446
447 /*
448  * This test will generate random numbers of calls to some getpid syscalls,
449  * then establish an mmap for a group of events that are created to monitor
450  * the syscalls.
451  *
452  * It will receive the events, using mmap, use its PERF_SAMPLE_ID generated
453  * sample.id field to map back to its respective perf_evsel instance.
454  *
455  * Then it checks if the number of syscalls reported as perf events by
456  * the kernel corresponds to the number of syscalls made.
457  */
458 static int test__basic_mmap(void)
459 {
460         int err = -1;
461         union perf_event *event;
462         struct thread_map *threads;
463         struct cpu_map *cpus;
464         struct perf_evlist *evlist;
465         struct perf_event_attr attr = {
466                 .type           = PERF_TYPE_TRACEPOINT,
467                 .read_format    = PERF_FORMAT_ID,
468                 .sample_type    = PERF_SAMPLE_ID,
469                 .watermark      = 0,
470         };
471         cpu_set_t cpu_set;
472         const char *syscall_names[] = { "getsid", "getppid", "getpgrp",
473                                         "getpgid", };
474         pid_t (*syscalls[])(void) = { (void *)getsid, getppid, getpgrp,
475                                       (void*)getpgid };
476 #define nsyscalls ARRAY_SIZE(syscall_names)
477         int ids[nsyscalls];
478         unsigned int nr_events[nsyscalls],
479                      expected_nr_events[nsyscalls], i, j;
480         struct perf_evsel *evsels[nsyscalls], *evsel;
481         int sample_size = __perf_evsel__sample_size(attr.sample_type);
482
483         for (i = 0; i < nsyscalls; ++i) {
484                 char name[64];
485
486                 snprintf(name, sizeof(name), "sys_enter_%s", syscall_names[i]);
487                 ids[i] = trace_event__id(name);
488                 if (ids[i] < 0) {
489                         pr_debug("Is debugfs mounted on /sys/kernel/debug?\n");
490                         return -1;
491                 }
492                 nr_events[i] = 0;
493                 expected_nr_events[i] = random() % 257;
494         }
495
496         threads = thread_map__new(-1, getpid(), UINT_MAX);
497         if (threads == NULL) {
498                 pr_debug("thread_map__new\n");
499                 return -1;
500         }
501
502         cpus = cpu_map__new(NULL);
503         if (cpus == NULL) {
504                 pr_debug("cpu_map__new\n");
505                 goto out_free_threads;
506         }
507
508         CPU_ZERO(&cpu_set);
509         CPU_SET(cpus->map[0], &cpu_set);
510         sched_setaffinity(0, sizeof(cpu_set), &cpu_set);
511         if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) {
512                 pr_debug("sched_setaffinity() failed on CPU %d: %s ",
513                          cpus->map[0], strerror(errno));
514                 goto out_free_cpus;
515         }
516
517         evlist = perf_evlist__new(cpus, threads);
518         if (evlist == NULL) {
519                 pr_debug("perf_evlist__new\n");
520                 goto out_free_cpus;
521         }
522
523         /* anonymous union fields, can't be initialized above */
524         attr.wakeup_events = 1;
525         attr.sample_period = 1;
526
527         for (i = 0; i < nsyscalls; ++i) {
528                 attr.config = ids[i];
529                 evsels[i] = perf_evsel__new(&attr, i);
530                 if (evsels[i] == NULL) {
531                         pr_debug("perf_evsel__new\n");
532                         goto out_free_evlist;
533                 }
534
535                 perf_evlist__add(evlist, evsels[i]);
536
537                 if (perf_evsel__open(evsels[i], cpus, threads, false, NULL) < 0) {
538                         pr_debug("failed to open counter: %s, "
539                                  "tweak /proc/sys/kernel/perf_event_paranoid?\n",
540                                  strerror(errno));
541                         goto out_close_fd;
542                 }
543         }
544
545         if (perf_evlist__mmap(evlist, 128, true) < 0) {
546                 pr_debug("failed to mmap events: %d (%s)\n", errno,
547                          strerror(errno));
548                 goto out_close_fd;
549         }
550
551         for (i = 0; i < nsyscalls; ++i)
552                 for (j = 0; j < expected_nr_events[i]; ++j) {
553                         int foo = syscalls[i]();
554                         ++foo;
555                 }
556
557         while ((event = perf_evlist__mmap_read(evlist, 0)) != NULL) {
558                 struct perf_sample sample;
559
560                 if (event->header.type != PERF_RECORD_SAMPLE) {
561                         pr_debug("unexpected %s event\n",
562                                  perf_event__name(event->header.type));
563                         goto out_munmap;
564                 }
565
566                 err = perf_event__parse_sample(event, attr.sample_type, sample_size,
567                                                false, &sample, false);
568                 if (err) {
569                         pr_err("Can't parse sample, err = %d\n", err);
570                         goto out_munmap;
571                 }
572
573                 evsel = perf_evlist__id2evsel(evlist, sample.id);
574                 if (evsel == NULL) {
575                         pr_debug("event with id %" PRIu64
576                                  " doesn't map to an evsel\n", sample.id);
577                         goto out_munmap;
578                 }
579                 nr_events[evsel->idx]++;
580         }
581
582         list_for_each_entry(evsel, &evlist->entries, node) {
583                 if (nr_events[evsel->idx] != expected_nr_events[evsel->idx]) {
584                         pr_debug("expected %d %s events, got %d\n",
585                                  expected_nr_events[evsel->idx],
586                                  event_name(evsel), nr_events[evsel->idx]);
587                         goto out_munmap;
588                 }
589         }
590
591         err = 0;
592 out_munmap:
593         perf_evlist__munmap(evlist);
594 out_close_fd:
595         for (i = 0; i < nsyscalls; ++i)
596                 perf_evsel__close_fd(evsels[i], 1, threads->nr);
597 out_free_evlist:
598         perf_evlist__delete(evlist);
599 out_free_cpus:
600         cpu_map__delete(cpus);
601 out_free_threads:
602         thread_map__delete(threads);
603         return err;
604 #undef nsyscalls
605 }
606
607 #define TEST_ASSERT_VAL(text, cond) \
608 do { \
609         if (!(cond)) { \
610                 pr_debug("FAILED %s:%d %s\n", __FILE__, __LINE__, text); \
611                 return -1; \
612         } \
613 } while (0)
614
615 static int test__checkevent_tracepoint(struct perf_evlist *evlist)
616 {
617         struct perf_evsel *evsel = list_entry(evlist->entries.next,
618                                               struct perf_evsel, node);
619
620         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
621         TEST_ASSERT_VAL("wrong type", PERF_TYPE_TRACEPOINT == evsel->attr.type);
622         TEST_ASSERT_VAL("wrong sample_type",
623                 (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | PERF_SAMPLE_CPU) ==
624                 evsel->attr.sample_type);
625         TEST_ASSERT_VAL("wrong sample_period", 1 == evsel->attr.sample_period);
626         return 0;
627 }
628
629 static int test__checkevent_tracepoint_multi(struct perf_evlist *evlist)
630 {
631         struct perf_evsel *evsel;
632
633         TEST_ASSERT_VAL("wrong number of entries", evlist->nr_entries > 1);
634
635         list_for_each_entry(evsel, &evlist->entries, node) {
636                 TEST_ASSERT_VAL("wrong type",
637                         PERF_TYPE_TRACEPOINT == evsel->attr.type);
638                 TEST_ASSERT_VAL("wrong sample_type",
639                         (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | PERF_SAMPLE_CPU)
640                         == evsel->attr.sample_type);
641                 TEST_ASSERT_VAL("wrong sample_period",
642                         1 == evsel->attr.sample_period);
643         }
644         return 0;
645 }
646
647 static int test__checkevent_raw(struct perf_evlist *evlist)
648 {
649         struct perf_evsel *evsel = list_entry(evlist->entries.next,
650                                               struct perf_evsel, node);
651
652         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
653         TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->attr.type);
654         TEST_ASSERT_VAL("wrong config", 0x1a == evsel->attr.config);
655         return 0;
656 }
657
658 static int test__checkevent_numeric(struct perf_evlist *evlist)
659 {
660         struct perf_evsel *evsel = list_entry(evlist->entries.next,
661                                               struct perf_evsel, node);
662
663         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
664         TEST_ASSERT_VAL("wrong type", 1 == evsel->attr.type);
665         TEST_ASSERT_VAL("wrong config", 1 == evsel->attr.config);
666         return 0;
667 }
668
669 static int test__checkevent_symbolic_name(struct perf_evlist *evlist)
670 {
671         struct perf_evsel *evsel = list_entry(evlist->entries.next,
672                                               struct perf_evsel, node);
673
674         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
675         TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE == evsel->attr.type);
676         TEST_ASSERT_VAL("wrong config",
677                         PERF_COUNT_HW_INSTRUCTIONS == evsel->attr.config);
678         return 0;
679 }
680
681 static int test__checkevent_symbolic_name_config(struct perf_evlist *evlist)
682 {
683         struct perf_evsel *evsel = list_entry(evlist->entries.next,
684                                               struct perf_evsel, node);
685
686         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
687         TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE == evsel->attr.type);
688         TEST_ASSERT_VAL("wrong config",
689                         PERF_COUNT_HW_CPU_CYCLES == evsel->attr.config);
690         TEST_ASSERT_VAL("wrong period",
691                         100000 == evsel->attr.sample_period);
692         TEST_ASSERT_VAL("wrong config1",
693                         0 == evsel->attr.config1);
694         TEST_ASSERT_VAL("wrong config2",
695                         1 == evsel->attr.config2);
696         return 0;
697 }
698
699 static int test__checkevent_symbolic_alias(struct perf_evlist *evlist)
700 {
701         struct perf_evsel *evsel = list_entry(evlist->entries.next,
702                                               struct perf_evsel, node);
703
704         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
705         TEST_ASSERT_VAL("wrong type", PERF_TYPE_SOFTWARE == evsel->attr.type);
706         TEST_ASSERT_VAL("wrong config",
707                         PERF_COUNT_SW_PAGE_FAULTS == evsel->attr.config);
708         return 0;
709 }
710
711 static int test__checkevent_genhw(struct perf_evlist *evlist)
712 {
713         struct perf_evsel *evsel = list_entry(evlist->entries.next,
714                                               struct perf_evsel, node);
715
716         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
717         TEST_ASSERT_VAL("wrong type", PERF_TYPE_HW_CACHE == evsel->attr.type);
718         TEST_ASSERT_VAL("wrong config", (1 << 16) == evsel->attr.config);
719         return 0;
720 }
721
722 static int test__checkevent_breakpoint(struct perf_evlist *evlist)
723 {
724         struct perf_evsel *evsel = list_entry(evlist->entries.next,
725                                               struct perf_evsel, node);
726
727         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
728         TEST_ASSERT_VAL("wrong type", PERF_TYPE_BREAKPOINT == evsel->attr.type);
729         TEST_ASSERT_VAL("wrong config", 0 == evsel->attr.config);
730         TEST_ASSERT_VAL("wrong bp_type", (HW_BREAKPOINT_R | HW_BREAKPOINT_W) ==
731                                          evsel->attr.bp_type);
732         TEST_ASSERT_VAL("wrong bp_len", HW_BREAKPOINT_LEN_4 ==
733                                         evsel->attr.bp_len);
734         return 0;
735 }
736
737 static int test__checkevent_breakpoint_x(struct perf_evlist *evlist)
738 {
739         struct perf_evsel *evsel = list_entry(evlist->entries.next,
740                                               struct perf_evsel, node);
741
742         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
743         TEST_ASSERT_VAL("wrong type", PERF_TYPE_BREAKPOINT == evsel->attr.type);
744         TEST_ASSERT_VAL("wrong config", 0 == evsel->attr.config);
745         TEST_ASSERT_VAL("wrong bp_type",
746                         HW_BREAKPOINT_X == evsel->attr.bp_type);
747         TEST_ASSERT_VAL("wrong bp_len", sizeof(long) == evsel->attr.bp_len);
748         return 0;
749 }
750
751 static int test__checkevent_breakpoint_r(struct perf_evlist *evlist)
752 {
753         struct perf_evsel *evsel = list_entry(evlist->entries.next,
754                                               struct perf_evsel, node);
755
756         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
757         TEST_ASSERT_VAL("wrong type",
758                         PERF_TYPE_BREAKPOINT == evsel->attr.type);
759         TEST_ASSERT_VAL("wrong config", 0 == evsel->attr.config);
760         TEST_ASSERT_VAL("wrong bp_type",
761                         HW_BREAKPOINT_R == evsel->attr.bp_type);
762         TEST_ASSERT_VAL("wrong bp_len",
763                         HW_BREAKPOINT_LEN_4 == evsel->attr.bp_len);
764         return 0;
765 }
766
767 static int test__checkevent_breakpoint_w(struct perf_evlist *evlist)
768 {
769         struct perf_evsel *evsel = list_entry(evlist->entries.next,
770                                               struct perf_evsel, node);
771
772         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
773         TEST_ASSERT_VAL("wrong type",
774                         PERF_TYPE_BREAKPOINT == evsel->attr.type);
775         TEST_ASSERT_VAL("wrong config", 0 == evsel->attr.config);
776         TEST_ASSERT_VAL("wrong bp_type",
777                         HW_BREAKPOINT_W == evsel->attr.bp_type);
778         TEST_ASSERT_VAL("wrong bp_len",
779                         HW_BREAKPOINT_LEN_4 == evsel->attr.bp_len);
780         return 0;
781 }
782
783 static int test__checkevent_tracepoint_modifier(struct perf_evlist *evlist)
784 {
785         struct perf_evsel *evsel = list_entry(evlist->entries.next,
786                                               struct perf_evsel, node);
787
788         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
789         TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->attr.exclude_kernel);
790         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
791         TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
792
793         return test__checkevent_tracepoint(evlist);
794 }
795
796 static int
797 test__checkevent_tracepoint_multi_modifier(struct perf_evlist *evlist)
798 {
799         struct perf_evsel *evsel;
800
801         TEST_ASSERT_VAL("wrong number of entries", evlist->nr_entries > 1);
802
803         list_for_each_entry(evsel, &evlist->entries, node) {
804                 TEST_ASSERT_VAL("wrong exclude_user",
805                                 !evsel->attr.exclude_user);
806                 TEST_ASSERT_VAL("wrong exclude_kernel",
807                                 evsel->attr.exclude_kernel);
808                 TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
809                 TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
810         }
811
812         return test__checkevent_tracepoint_multi(evlist);
813 }
814
815 static int test__checkevent_raw_modifier(struct perf_evlist *evlist)
816 {
817         struct perf_evsel *evsel = list_entry(evlist->entries.next,
818                                               struct perf_evsel, node);
819
820         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
821         TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->attr.exclude_kernel);
822         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
823         TEST_ASSERT_VAL("wrong precise_ip", evsel->attr.precise_ip);
824
825         return test__checkevent_raw(evlist);
826 }
827
828 static int test__checkevent_numeric_modifier(struct perf_evlist *evlist)
829 {
830         struct perf_evsel *evsel = list_entry(evlist->entries.next,
831                                               struct perf_evsel, node);
832
833         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
834         TEST_ASSERT_VAL("wrong exclude_kernel", evsel->attr.exclude_kernel);
835         TEST_ASSERT_VAL("wrong exclude_hv", !evsel->attr.exclude_hv);
836         TEST_ASSERT_VAL("wrong precise_ip", evsel->attr.precise_ip);
837
838         return test__checkevent_numeric(evlist);
839 }
840
841 static int test__checkevent_symbolic_name_modifier(struct perf_evlist *evlist)
842 {
843         struct perf_evsel *evsel = list_entry(evlist->entries.next,
844                                               struct perf_evsel, node);
845
846         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
847         TEST_ASSERT_VAL("wrong exclude_kernel", evsel->attr.exclude_kernel);
848         TEST_ASSERT_VAL("wrong exclude_hv", !evsel->attr.exclude_hv);
849         TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
850
851         return test__checkevent_symbolic_name(evlist);
852 }
853
854 static int test__checkevent_exclude_host_modifier(struct perf_evlist *evlist)
855 {
856         struct perf_evsel *evsel = list_entry(evlist->entries.next,
857                                               struct perf_evsel, node);
858
859         TEST_ASSERT_VAL("wrong exclude guest", !evsel->attr.exclude_guest);
860         TEST_ASSERT_VAL("wrong exclude host", evsel->attr.exclude_host);
861
862         return test__checkevent_symbolic_name(evlist);
863 }
864
865 static int test__checkevent_exclude_guest_modifier(struct perf_evlist *evlist)
866 {
867         struct perf_evsel *evsel = list_entry(evlist->entries.next,
868                                               struct perf_evsel, node);
869
870         TEST_ASSERT_VAL("wrong exclude guest", evsel->attr.exclude_guest);
871         TEST_ASSERT_VAL("wrong exclude host", !evsel->attr.exclude_host);
872
873         return test__checkevent_symbolic_name(evlist);
874 }
875
876 static int test__checkevent_symbolic_alias_modifier(struct perf_evlist *evlist)
877 {
878         struct perf_evsel *evsel = list_entry(evlist->entries.next,
879                                               struct perf_evsel, node);
880
881         TEST_ASSERT_VAL("wrong exclude_user", !evsel->attr.exclude_user);
882         TEST_ASSERT_VAL("wrong exclude_kernel", evsel->attr.exclude_kernel);
883         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
884         TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
885
886         return test__checkevent_symbolic_alias(evlist);
887 }
888
889 static int test__checkevent_genhw_modifier(struct perf_evlist *evlist)
890 {
891         struct perf_evsel *evsel = list_entry(evlist->entries.next,
892                                               struct perf_evsel, node);
893
894         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
895         TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->attr.exclude_kernel);
896         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
897         TEST_ASSERT_VAL("wrong precise_ip", evsel->attr.precise_ip);
898
899         return test__checkevent_genhw(evlist);
900 }
901
902 static int test__checkevent_breakpoint_modifier(struct perf_evlist *evlist)
903 {
904         struct perf_evsel *evsel = list_entry(evlist->entries.next,
905                                               struct perf_evsel, node);
906
907         TEST_ASSERT_VAL("wrong exclude_user", !evsel->attr.exclude_user);
908         TEST_ASSERT_VAL("wrong exclude_kernel", evsel->attr.exclude_kernel);
909         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
910         TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
911
912         return test__checkevent_breakpoint(evlist);
913 }
914
915 static int test__checkevent_breakpoint_x_modifier(struct perf_evlist *evlist)
916 {
917         struct perf_evsel *evsel = list_entry(evlist->entries.next,
918                                               struct perf_evsel, node);
919
920         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
921         TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->attr.exclude_kernel);
922         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
923         TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
924
925         return test__checkevent_breakpoint_x(evlist);
926 }
927
928 static int test__checkevent_breakpoint_r_modifier(struct perf_evlist *evlist)
929 {
930         struct perf_evsel *evsel = list_entry(evlist->entries.next,
931                                               struct perf_evsel, node);
932
933         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
934         TEST_ASSERT_VAL("wrong exclude_kernel", evsel->attr.exclude_kernel);
935         TEST_ASSERT_VAL("wrong exclude_hv", !evsel->attr.exclude_hv);
936         TEST_ASSERT_VAL("wrong precise_ip", evsel->attr.precise_ip);
937
938         return test__checkevent_breakpoint_r(evlist);
939 }
940
941 static int test__checkevent_breakpoint_w_modifier(struct perf_evlist *evlist)
942 {
943         struct perf_evsel *evsel = list_entry(evlist->entries.next,
944                                               struct perf_evsel, node);
945
946         TEST_ASSERT_VAL("wrong exclude_user", !evsel->attr.exclude_user);
947         TEST_ASSERT_VAL("wrong exclude_kernel", evsel->attr.exclude_kernel);
948         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
949         TEST_ASSERT_VAL("wrong precise_ip", evsel->attr.precise_ip);
950
951         return test__checkevent_breakpoint_w(evlist);
952 }
953
954 static int test__checkevent_pmu(struct perf_evlist *evlist)
955 {
956
957         struct perf_evsel *evsel = list_entry(evlist->entries.next,
958                                               struct perf_evsel, node);
959
960         TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->nr_entries);
961         TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->attr.type);
962         TEST_ASSERT_VAL("wrong config",    10 == evsel->attr.config);
963         TEST_ASSERT_VAL("wrong config1",    1 == evsel->attr.config1);
964         TEST_ASSERT_VAL("wrong config2",    3 == evsel->attr.config2);
965         TEST_ASSERT_VAL("wrong period",  1000 == evsel->attr.sample_period);
966
967         return 0;
968 }
969
970 static int test__checkevent_list(struct perf_evlist *evlist)
971 {
972         struct perf_evsel *evsel;
973
974         TEST_ASSERT_VAL("wrong number of entries", 3 == evlist->nr_entries);
975
976         /* r1 */
977         evsel = list_entry(evlist->entries.next, struct perf_evsel, node);
978         TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->attr.type);
979         TEST_ASSERT_VAL("wrong config", 1 == evsel->attr.config);
980         TEST_ASSERT_VAL("wrong config1", 0 == evsel->attr.config1);
981         TEST_ASSERT_VAL("wrong config2", 0 == evsel->attr.config2);
982         TEST_ASSERT_VAL("wrong exclude_user", !evsel->attr.exclude_user);
983         TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->attr.exclude_kernel);
984         TEST_ASSERT_VAL("wrong exclude_hv", !evsel->attr.exclude_hv);
985         TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
986
987         /* syscalls:sys_enter_open:k */
988         evsel = list_entry(evsel->node.next, struct perf_evsel, node);
989         TEST_ASSERT_VAL("wrong type", PERF_TYPE_TRACEPOINT == evsel->attr.type);
990         TEST_ASSERT_VAL("wrong sample_type",
991                 (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | PERF_SAMPLE_CPU) ==
992                 evsel->attr.sample_type);
993         TEST_ASSERT_VAL("wrong sample_period", 1 == evsel->attr.sample_period);
994         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
995         TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->attr.exclude_kernel);
996         TEST_ASSERT_VAL("wrong exclude_hv", evsel->attr.exclude_hv);
997         TEST_ASSERT_VAL("wrong precise_ip", !evsel->attr.precise_ip);
998
999         /* 1:1:hp */
1000         evsel = list_entry(evsel->node.next, struct perf_evsel, node);
1001         TEST_ASSERT_VAL("wrong type", 1 == evsel->attr.type);
1002         TEST_ASSERT_VAL("wrong config", 1 == evsel->attr.config);
1003         TEST_ASSERT_VAL("wrong exclude_user", evsel->attr.exclude_user);
1004         TEST_ASSERT_VAL("wrong exclude_kernel", evsel->attr.exclude_kernel);
1005         TEST_ASSERT_VAL("wrong exclude_hv", !evsel->attr.exclude_hv);
1006         TEST_ASSERT_VAL("wrong precise_ip", evsel->attr.precise_ip);
1007
1008         return 0;
1009 }
1010
1011 static struct test__event_st {
1012         const char *name;
1013         __u32 type;
1014         int (*check)(struct perf_evlist *evlist);
1015 } test__events[] = {
1016         {
1017                 .name  = "syscalls:sys_enter_open",
1018                 .check = test__checkevent_tracepoint,
1019         },
1020         {
1021                 .name  = "syscalls:*",
1022                 .check = test__checkevent_tracepoint_multi,
1023         },
1024         {
1025                 .name  = "r1a",
1026                 .check = test__checkevent_raw,
1027         },
1028         {
1029                 .name  = "1:1",
1030                 .check = test__checkevent_numeric,
1031         },
1032         {
1033                 .name  = "instructions",
1034                 .check = test__checkevent_symbolic_name,
1035         },
1036         {
1037                 .name  = "cycles/period=100000,config2/",
1038                 .check = test__checkevent_symbolic_name_config,
1039         },
1040         {
1041                 .name  = "faults",
1042                 .check = test__checkevent_symbolic_alias,
1043         },
1044         {
1045                 .name  = "L1-dcache-load-miss",
1046                 .check = test__checkevent_genhw,
1047         },
1048         {
1049                 .name  = "mem:0",
1050                 .check = test__checkevent_breakpoint,
1051         },
1052         {
1053                 .name  = "mem:0:x",
1054                 .check = test__checkevent_breakpoint_x,
1055         },
1056         {
1057                 .name  = "mem:0:r",
1058                 .check = test__checkevent_breakpoint_r,
1059         },
1060         {
1061                 .name  = "mem:0:w",
1062                 .check = test__checkevent_breakpoint_w,
1063         },
1064         {
1065                 .name  = "syscalls:sys_enter_open:k",
1066                 .check = test__checkevent_tracepoint_modifier,
1067         },
1068         {
1069                 .name  = "syscalls:*:u",
1070                 .check = test__checkevent_tracepoint_multi_modifier,
1071         },
1072         {
1073                 .name  = "r1a:kp",
1074                 .check = test__checkevent_raw_modifier,
1075         },
1076         {
1077                 .name  = "1:1:hp",
1078                 .check = test__checkevent_numeric_modifier,
1079         },
1080         {
1081                 .name  = "instructions:h",
1082                 .check = test__checkevent_symbolic_name_modifier,
1083         },
1084         {
1085                 .name  = "faults:u",
1086                 .check = test__checkevent_symbolic_alias_modifier,
1087         },
1088         {
1089                 .name  = "L1-dcache-load-miss:kp",
1090                 .check = test__checkevent_genhw_modifier,
1091         },
1092         {
1093                 .name  = "mem:0:u",
1094                 .check = test__checkevent_breakpoint_modifier,
1095         },
1096         {
1097                 .name  = "mem:0:x:k",
1098                 .check = test__checkevent_breakpoint_x_modifier,
1099         },
1100         {
1101                 .name  = "mem:0:r:hp",
1102                 .check = test__checkevent_breakpoint_r_modifier,
1103         },
1104         {
1105                 .name  = "mem:0:w:up",
1106                 .check = test__checkevent_breakpoint_w_modifier,
1107         },
1108         {
1109                 .name  = "cpu/config=10,config1,config2=3,period=1000/u",
1110                 .check = test__checkevent_pmu,
1111         },
1112         {
1113                 .name  = "r1,syscalls:sys_enter_open:k,1:1:hp",
1114                 .check = test__checkevent_list,
1115         },
1116         {
1117                 .name  = "instructions:G",
1118                 .check = test__checkevent_exclude_host_modifier,
1119         },
1120         {
1121                 .name  = "instructions:H",
1122                 .check = test__checkevent_exclude_guest_modifier,
1123         },
1124 };
1125
1126 #define TEST__EVENTS_CNT (sizeof(test__events) / sizeof(struct test__event_st))
1127
1128 static int test__parse_events(void)
1129 {
1130         struct perf_evlist *evlist;
1131         u_int i;
1132         int ret = 0;
1133
1134         for (i = 0; i < TEST__EVENTS_CNT; i++) {
1135                 struct test__event_st *e = &test__events[i];
1136
1137                 evlist = perf_evlist__new(NULL, NULL);
1138                 if (evlist == NULL)
1139                         break;
1140
1141                 ret = parse_events(evlist, e->name, 0);
1142                 if (ret) {
1143                         pr_debug("failed to parse event '%s', err %d\n",
1144                                  e->name, ret);
1145                         break;
1146                 }
1147
1148                 ret = e->check(evlist);
1149                 perf_evlist__delete(evlist);
1150                 if (ret)
1151                         break;
1152         }
1153
1154         return ret;
1155 }
1156
1157 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
1158 {
1159         int i, cpu = -1, nrcpus = 1024;
1160 realloc:
1161         CPU_ZERO(maskp);
1162
1163         if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
1164                 if (errno == EINVAL && nrcpus < (1024 << 8)) {
1165                         nrcpus = nrcpus << 2;
1166                         goto realloc;
1167                 }
1168                 perror("sched_getaffinity");
1169                         return -1;
1170         }
1171
1172         for (i = 0; i < nrcpus; i++) {
1173                 if (CPU_ISSET(i, maskp)) {
1174                         if (cpu == -1)
1175                                 cpu = i;
1176                         else
1177                                 CPU_CLR(i, maskp);
1178                 }
1179         }
1180
1181         return cpu;
1182 }
1183
1184 static int test__PERF_RECORD(void)
1185 {
1186         struct perf_record_opts opts = {
1187                 .no_delay   = true,
1188                 .freq       = 10,
1189                 .mmap_pages = 256,
1190         };
1191         cpu_set_t cpu_mask;
1192         size_t cpu_mask_size = sizeof(cpu_mask);
1193         struct perf_evlist *evlist = perf_evlist__new(NULL, NULL);
1194         struct perf_evsel *evsel;
1195         struct perf_sample sample;
1196         const char *cmd = "sleep";
1197         const char *argv[] = { cmd, "1", NULL, };
1198         char *bname;
1199         u64 sample_type, prev_time = 0;
1200         bool found_cmd_mmap = false,
1201              found_libc_mmap = false,
1202              found_vdso_mmap = false,
1203              found_ld_mmap = false;
1204         int err = -1, errs = 0, i, wakeups = 0, sample_size;
1205         u32 cpu;
1206         int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
1207
1208         if (evlist == NULL || argv == NULL) {
1209                 pr_debug("Not enough memory to create evlist\n");
1210                 goto out;
1211         }
1212
1213         /*
1214          * We need at least one evsel in the evlist, use the default
1215          * one: "cycles".
1216          */
1217         err = perf_evlist__add_default(evlist);
1218         if (err < 0) {
1219                 pr_debug("Not enough memory to create evsel\n");
1220                 goto out_delete_evlist;
1221         }
1222
1223         /*
1224          * Create maps of threads and cpus to monitor. In this case
1225          * we start with all threads and cpus (-1, -1) but then in
1226          * perf_evlist__prepare_workload we'll fill in the only thread
1227          * we're monitoring, the one forked there.
1228          */
1229         err = perf_evlist__create_maps(evlist, opts.target_pid,
1230                                        opts.target_tid, UINT_MAX, opts.cpu_list);
1231         if (err < 0) {
1232                 pr_debug("Not enough memory to create thread/cpu maps\n");
1233                 goto out_delete_evlist;
1234         }
1235
1236         /*
1237          * Prepare the workload in argv[] to run, it'll fork it, and then wait
1238          * for perf_evlist__start_workload() to exec it. This is done this way
1239          * so that we have time to open the evlist (calling sys_perf_event_open
1240          * on all the fds) and then mmap them.
1241          */
1242         err = perf_evlist__prepare_workload(evlist, &opts, argv);
1243         if (err < 0) {
1244                 pr_debug("Couldn't run the workload!\n");
1245                 goto out_delete_evlist;
1246         }
1247
1248         /*
1249          * Config the evsels, setting attr->comm on the first one, etc.
1250          */
1251         evsel = list_entry(evlist->entries.next, struct perf_evsel, node);
1252         evsel->attr.sample_type |= PERF_SAMPLE_CPU;
1253         evsel->attr.sample_type |= PERF_SAMPLE_TID;
1254         evsel->attr.sample_type |= PERF_SAMPLE_TIME;
1255         perf_evlist__config_attrs(evlist, &opts);
1256
1257         err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
1258         if (err < 0) {
1259                 pr_debug("sched__get_first_possible_cpu: %s\n", strerror(errno));
1260                 goto out_delete_evlist;
1261         }
1262
1263         cpu = err;
1264
1265         /*
1266          * So that we can check perf_sample.cpu on all the samples.
1267          */
1268         if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) {
1269                 pr_debug("sched_setaffinity: %s\n", strerror(errno));
1270                 goto out_delete_evlist;
1271         }
1272
1273         /*
1274          * Call sys_perf_event_open on all the fds on all the evsels,
1275          * grouping them if asked to.
1276          */
1277         err = perf_evlist__open(evlist, opts.group);
1278         if (err < 0) {
1279                 pr_debug("perf_evlist__open: %s\n", strerror(errno));
1280                 goto out_delete_evlist;
1281         }
1282
1283         /*
1284          * mmap the first fd on a given CPU and ask for events for the other
1285          * fds in the same CPU to be injected in the same mmap ring buffer
1286          * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
1287          */
1288         err = perf_evlist__mmap(evlist, opts.mmap_pages, false);
1289         if (err < 0) {
1290                 pr_debug("perf_evlist__mmap: %s\n", strerror(errno));
1291                 goto out_delete_evlist;
1292         }
1293
1294         /*
1295          * We'll need these two to parse the PERF_SAMPLE_* fields in each
1296          * event.
1297          */
1298         sample_type = perf_evlist__sample_type(evlist);
1299         sample_size = __perf_evsel__sample_size(sample_type);
1300
1301         /*
1302          * Now that all is properly set up, enable the events, they will
1303          * count just on workload.pid, which will start...
1304          */
1305         perf_evlist__enable(evlist);
1306
1307         /*
1308          * Now!
1309          */
1310         perf_evlist__start_workload(evlist);
1311
1312         while (1) {
1313                 int before = total_events;
1314
1315                 for (i = 0; i < evlist->nr_mmaps; i++) {
1316                         union perf_event *event;
1317
1318                         while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
1319                                 const u32 type = event->header.type;
1320                                 const char *name = perf_event__name(type);
1321
1322                                 ++total_events;
1323                                 if (type < PERF_RECORD_MAX)
1324                                         nr_events[type]++;
1325
1326                                 err = perf_event__parse_sample(event, sample_type,
1327                                                                sample_size, true,
1328                                                                &sample, false);
1329                                 if (err < 0) {
1330                                         if (verbose)
1331                                                 perf_event__fprintf(event, stderr);
1332                                         pr_debug("Couldn't parse sample\n");
1333                                         goto out_err;
1334                                 }
1335
1336                                 if (verbose) {
1337                                         pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
1338                                         perf_event__fprintf(event, stderr);
1339                                 }
1340
1341                                 if (prev_time > sample.time) {
1342                                         pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
1343                                                  name, prev_time, sample.time);
1344                                         ++errs;
1345                                 }
1346
1347                                 prev_time = sample.time;
1348
1349                                 if (sample.cpu != cpu) {
1350                                         pr_debug("%s with unexpected cpu, expected %d, got %d\n",
1351                                                  name, cpu, sample.cpu);
1352                                         ++errs;
1353                                 }
1354
1355                                 if ((pid_t)sample.pid != evlist->workload.pid) {
1356                                         pr_debug("%s with unexpected pid, expected %d, got %d\n",
1357                                                  name, evlist->workload.pid, sample.pid);
1358                                         ++errs;
1359                                 }
1360
1361                                 if ((pid_t)sample.tid != evlist->workload.pid) {
1362                                         pr_debug("%s with unexpected tid, expected %d, got %d\n",
1363                                                  name, evlist->workload.pid, sample.tid);
1364                                         ++errs;
1365                                 }
1366
1367                                 if ((type == PERF_RECORD_COMM ||
1368                                      type == PERF_RECORD_MMAP ||
1369                                      type == PERF_RECORD_FORK ||
1370                                      type == PERF_RECORD_EXIT) &&
1371                                      (pid_t)event->comm.pid != evlist->workload.pid) {
1372                                         pr_debug("%s with unexpected pid/tid\n", name);
1373                                         ++errs;
1374                                 }
1375
1376                                 if ((type == PERF_RECORD_COMM ||
1377                                      type == PERF_RECORD_MMAP) &&
1378                                      event->comm.pid != event->comm.tid) {
1379                                         pr_debug("%s with different pid/tid!\n", name);
1380                                         ++errs;
1381                                 }
1382
1383                                 switch (type) {
1384                                 case PERF_RECORD_COMM:
1385                                         if (strcmp(event->comm.comm, cmd)) {
1386                                                 pr_debug("%s with unexpected comm!\n", name);
1387                                                 ++errs;
1388                                         }
1389                                         break;
1390                                 case PERF_RECORD_EXIT:
1391                                         goto found_exit;
1392                                 case PERF_RECORD_MMAP:
1393                                         bname = strrchr(event->mmap.filename, '/');
1394                                         if (bname != NULL) {
1395                                                 if (!found_cmd_mmap)
1396                                                         found_cmd_mmap = !strcmp(bname + 1, cmd);
1397                                                 if (!found_libc_mmap)
1398                                                         found_libc_mmap = !strncmp(bname + 1, "libc", 4);
1399                                                 if (!found_ld_mmap)
1400                                                         found_ld_mmap = !strncmp(bname + 1, "ld", 2);
1401                                         } else if (!found_vdso_mmap)
1402                                                 found_vdso_mmap = !strcmp(event->mmap.filename, "[vdso]");
1403                                         break;
1404
1405                                 case PERF_RECORD_SAMPLE:
1406                                         /* Just ignore samples for now */
1407                                         break;
1408                                 default:
1409                                         pr_debug("Unexpected perf_event->header.type %d!\n",
1410                                                  type);
1411                                         ++errs;
1412                                 }
1413                         }
1414                 }
1415
1416                 /*
1417                  * We don't use poll here because at least at 3.1 times the
1418                  * PERF_RECORD_{!SAMPLE} events don't honour
1419                  * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
1420                  */
1421                 if (total_events == before && false)
1422                         poll(evlist->pollfd, evlist->nr_fds, -1);
1423
1424                 sleep(1);
1425                 if (++wakeups > 5) {
1426                         pr_debug("No PERF_RECORD_EXIT event!\n");
1427                         break;
1428                 }
1429         }
1430
1431 found_exit:
1432         if (nr_events[PERF_RECORD_COMM] > 1) {
1433                 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
1434                 ++errs;
1435         }
1436
1437         if (nr_events[PERF_RECORD_COMM] == 0) {
1438                 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
1439                 ++errs;
1440         }
1441
1442         if (!found_cmd_mmap) {
1443                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
1444                 ++errs;
1445         }
1446
1447         if (!found_libc_mmap) {
1448                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
1449                 ++errs;
1450         }
1451
1452         if (!found_ld_mmap) {
1453                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
1454                 ++errs;
1455         }
1456
1457         if (!found_vdso_mmap) {
1458                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
1459                 ++errs;
1460         }
1461 out_err:
1462         perf_evlist__munmap(evlist);
1463 out_delete_evlist:
1464         perf_evlist__delete(evlist);
1465 out:
1466         return (err < 0 || errs > 0) ? -1 : 0;
1467 }
1468
1469
1470 #if defined(__x86_64__) || defined(__i386__)
1471
1472 #define barrier() asm volatile("" ::: "memory")
1473
1474 static u64 rdpmc(unsigned int counter)
1475 {
1476         unsigned int low, high;
1477
1478         asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter));
1479
1480         return low | ((u64)high) << 32;
1481 }
1482
1483 static u64 rdtsc(void)
1484 {
1485         unsigned int low, high;
1486
1487         asm volatile("rdtsc" : "=a" (low), "=d" (high));
1488
1489         return low | ((u64)high) << 32;
1490 }
1491
1492 static u64 mmap_read_self(void *addr)
1493 {
1494         struct perf_event_mmap_page *pc = addr;
1495         u32 seq, idx, time_mult = 0, time_shift = 0;
1496         u64 count, cyc = 0, time_offset = 0, enabled, running, delta;
1497
1498         do {
1499                 seq = pc->lock;
1500                 barrier();
1501
1502                 enabled = pc->time_enabled;
1503                 running = pc->time_running;
1504
1505                 if (enabled != running) {
1506                         cyc = rdtsc();
1507                         time_mult = pc->time_mult;
1508                         time_shift = pc->time_shift;
1509                         time_offset = pc->time_offset;
1510                 }
1511
1512                 idx = pc->index;
1513                 count = pc->offset;
1514                 if (idx)
1515                         count += rdpmc(idx - 1);
1516
1517                 barrier();
1518         } while (pc->lock != seq);
1519
1520         if (enabled != running) {
1521                 u64 quot, rem;
1522
1523                 quot = (cyc >> time_shift);
1524                 rem = cyc & ((1 << time_shift) - 1);
1525                 delta = time_offset + quot * time_mult +
1526                         ((rem * time_mult) >> time_shift);
1527
1528                 enabled += delta;
1529                 if (idx)
1530                         running += delta;
1531
1532                 quot = count / running;
1533                 rem = count % running;
1534                 count = quot * enabled + (rem * enabled) / running;
1535         }
1536
1537         return count;
1538 }
1539
1540 /*
1541  * If the RDPMC instruction faults then signal this back to the test parent task:
1542  */
1543 static void segfault_handler(int sig __used, siginfo_t *info __used, void *uc __used)
1544 {
1545         exit(-1);
1546 }
1547
1548 static int __test__rdpmc(void)
1549 {
1550         long page_size = sysconf(_SC_PAGE_SIZE);
1551         volatile int tmp = 0;
1552         u64 i, loops = 1000;
1553         int n;
1554         int fd;
1555         void *addr;
1556         struct perf_event_attr attr = {
1557                 .type = PERF_TYPE_HARDWARE,
1558                 .config = PERF_COUNT_HW_INSTRUCTIONS,
1559                 .exclude_kernel = 1,
1560         };
1561         u64 delta_sum = 0;
1562         struct sigaction sa;
1563
1564         sigfillset(&sa.sa_mask);
1565         sa.sa_sigaction = segfault_handler;
1566         sigaction(SIGSEGV, &sa, NULL);
1567
1568         fprintf(stderr, "\n\n");
1569
1570         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
1571         if (fd < 0) {
1572                 die("Error: sys_perf_event_open() syscall returned "
1573                     "with %d (%s)\n", fd, strerror(errno));
1574         }
1575
1576         addr = mmap(NULL, page_size, PROT_READ, MAP_SHARED, fd, 0);
1577         if (addr == (void *)(-1)) {
1578                 die("Error: mmap() syscall returned "
1579                     "with (%s)\n", strerror(errno));
1580         }
1581
1582         for (n = 0; n < 6; n++) {
1583                 u64 stamp, now, delta;
1584
1585                 stamp = mmap_read_self(addr);
1586
1587                 for (i = 0; i < loops; i++)
1588                         tmp++;
1589
1590                 now = mmap_read_self(addr);
1591                 loops *= 10;
1592
1593                 delta = now - stamp;
1594                 fprintf(stderr, "%14d: %14Lu\n", n, (long long)delta);
1595
1596                 delta_sum += delta;
1597         }
1598
1599         munmap(addr, page_size);
1600         close(fd);
1601
1602         fprintf(stderr, "   ");
1603
1604         if (!delta_sum)
1605                 return -1;
1606
1607         return 0;
1608 }
1609
1610 static int test__rdpmc(void)
1611 {
1612         int status = 0;
1613         int wret = 0;
1614         int ret;
1615         int pid;
1616
1617         pid = fork();
1618         if (pid < 0)
1619                 return -1;
1620
1621         if (!pid) {
1622                 ret = __test__rdpmc();
1623
1624                 exit(ret);
1625         }
1626
1627         wret = waitpid(pid, &status, 0);
1628         if (wret < 0 || status)
1629                 return -1;
1630
1631         return 0;
1632 }
1633
1634 #endif
1635
1636 static int test__perf_pmu(void)
1637 {
1638         return perf_pmu__test();
1639 }
1640
1641 static struct test {
1642         const char *desc;
1643         int (*func)(void);
1644 } tests[] = {
1645         {
1646                 .desc = "vmlinux symtab matches kallsyms",
1647                 .func = test__vmlinux_matches_kallsyms,
1648         },
1649         {
1650                 .desc = "detect open syscall event",
1651                 .func = test__open_syscall_event,
1652         },
1653         {
1654                 .desc = "detect open syscall event on all cpus",
1655                 .func = test__open_syscall_event_on_all_cpus,
1656         },
1657         {
1658                 .desc = "read samples using the mmap interface",
1659                 .func = test__basic_mmap,
1660         },
1661         {
1662                 .desc = "parse events tests",
1663                 .func = test__parse_events,
1664         },
1665 #if defined(__x86_64__) || defined(__i386__)
1666         {
1667                 .desc = "x86 rdpmc test",
1668                 .func = test__rdpmc,
1669         },
1670 #endif
1671         {
1672                 .desc = "Validate PERF_RECORD_* events & perf_sample fields",
1673                 .func = test__PERF_RECORD,
1674         },
1675         {
1676                 .desc = "Test perf pmu format parsing",
1677                 .func = test__perf_pmu,
1678         },
1679         {
1680                 .func = NULL,
1681         },
1682 };
1683
1684 static bool perf_test__matches(int curr, int argc, const char *argv[])
1685 {
1686         int i;
1687
1688         if (argc == 0)
1689                 return true;
1690
1691         for (i = 0; i < argc; ++i) {
1692                 char *end;
1693                 long nr = strtoul(argv[i], &end, 10);
1694
1695                 if (*end == '\0') {
1696                         if (nr == curr + 1)
1697                                 return true;
1698                         continue;
1699                 }
1700
1701                 if (strstr(tests[curr].desc, argv[i]))
1702                         return true;
1703         }
1704
1705         return false;
1706 }
1707
1708 static int __cmd_test(int argc, const char *argv[])
1709 {
1710         int i = 0;
1711
1712         while (tests[i].func) {
1713                 int curr = i++, err;
1714
1715                 if (!perf_test__matches(curr, argc, argv))
1716                         continue;
1717
1718                 pr_info("%2d: %s:", i, tests[curr].desc);
1719                 pr_debug("\n--- start ---\n");
1720                 err = tests[curr].func();
1721                 pr_debug("---- end ----\n%s:", tests[curr].desc);
1722                 pr_info(" %s\n", err ? "FAILED!\n" : "Ok");
1723         }
1724
1725         return 0;
1726 }
1727
1728 static int perf_test__list(int argc, const char **argv)
1729 {
1730         int i = 0;
1731
1732         while (tests[i].func) {
1733                 int curr = i++;
1734
1735                 if (argc > 1 && !strstr(tests[curr].desc, argv[1]))
1736                         continue;
1737
1738                 pr_info("%2d: %s\n", i, tests[curr].desc);
1739         }
1740
1741         return 0;
1742 }
1743
1744 int cmd_test(int argc, const char **argv, const char *prefix __used)
1745 {
1746         const char * const test_usage[] = {
1747         "perf test [<options>] [{list <test-name-fragment>|[<test-name-fragments>|<test-numbers>]}]",
1748         NULL,
1749         };
1750         const struct option test_options[] = {
1751         OPT_INCR('v', "verbose", &verbose,
1752                     "be more verbose (show symbol address, etc)"),
1753         OPT_END()
1754         };
1755
1756         argc = parse_options(argc, argv, test_options, test_usage, 0);
1757         if (argc >= 1 && !strcmp(argv[0], "list"))
1758                 return perf_test__list(argc, argv);
1759
1760         symbol_conf.priv_size = sizeof(int);
1761         symbol_conf.sort_by_name = true;
1762         symbol_conf.try_vmlinux_path = true;
1763
1764         if (symbol__init() < 0)
1765                 return -1;
1766
1767         return __cmd_test(argc, argv);
1768 }