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/parse-options.h"
  11 #include "util/session.h"
  12 #include "util/symbol.h"
  13 #include "util/thread.h"
  14
  15 static long page_size;
  16
  17 static int vmlinux_matches_kallsyms_filter(struct map *map __used, struct symbol *sym)
  18 {
  19         bool *visited = symbol__priv(sym);
  20         *visited = true;
  21         return 0;
  22 }
  23
  24 static int test__vmlinux_matches_kallsyms(void)
  25 {
  26         int err = -1;
  27         struct rb_node *nd;
  28         struct symbol *sym;
  29         struct map *kallsyms_map, *vmlinux_map;
  30         struct machine kallsyms, vmlinux;
  31         enum map_type type = MAP__FUNCTION;
  32         struct ref_reloc_sym ref_reloc_sym = { .name = "_stext", };
  33
  34         /*
  35          * Step 1:
  36          *
  37          * Init the machines that will hold kernel, modules obtained from
  38          * both vmlinux + .ko files and from /proc/kallsyms split by modules.
  39          */
  40         machine__init(&kallsyms, "", HOST_KERNEL_ID);
  41         machine__init(&vmlinux, "", HOST_KERNEL_ID);
  42
  43         /*
  44          * Step 2:
  45          *
  46          * Create the kernel maps for kallsyms and the DSO where we will then
  47          * load /proc/kallsyms. Also create the modules maps from /proc/modules
  48          * and find the .ko files that match them in /lib/modules/`uname -r`/.
  49          */
  50         if (machine__create_kernel_maps(&kallsyms) < 0) {
  51                 pr_debug("machine__create_kernel_maps ");
  52                 return -1;
  53         }
  54
  55         /*
  56          * Step 3:
  57          *
  58          * Load and split /proc/kallsyms into multiple maps, one per module.
  59          */
  60         if (machine__load_kallsyms(&kallsyms, "/proc/kallsyms", type, NULL) <= 0) {
  61                 pr_debug("dso__load_kallsyms ");
  62                 goto out;
  63         }
  64
  65         /*
  66          * Step 4:
  67          *
  68          * kallsyms will be internally on demand sorted by name so that we can
  69          * find the reference relocation * symbol, i.e. the symbol we will use
  70          * to see if the running kernel was relocated by checking if it has the
  71          * same value in the vmlinux file we load.
  72          */
  73         kallsyms_map = machine__kernel_map(&kallsyms, type);
  74
  75         sym = map__find_symbol_by_name(kallsyms_map, ref_reloc_sym.name, NULL);
  76         if (sym == NULL) {
  77                 pr_debug("dso__find_symbol_by_name ");
  78                 goto out;
  79         }
  80
  81         ref_reloc_sym.addr = sym->start;
  82
  83         /*
  84          * Step 5:
  85          *
  86          * Now repeat step 2, this time for the vmlinux file we'll auto-locate.
  87          */
  88         if (machine__create_kernel_maps(&vmlinux) < 0) {
  89                 pr_debug("machine__create_kernel_maps ");
  90                 goto out;
  91         }
  92
  93         vmlinux_map = machine__kernel_map(&vmlinux, type);
  94         map__kmap(vmlinux_map)->ref_reloc_sym = &ref_reloc_sym;
  95
  96         /*
  97          * Step 6:
  98          *
  99          * Locate a vmlinux file in the vmlinux path that has a buildid that
 100          * matches the one of the running kernel.
 101          *
 102          * While doing that look if we find the ref reloc symbol, if we find it
 103          * we'll have its ref_reloc_symbol.unrelocated_addr and then
 104          * maps__reloc_vmlinux will notice and set proper ->[un]map_ip routines
 105          * to fixup the symbols.
 106          */
 107         if (machine__load_vmlinux_path(&vmlinux, type,
 108                                        vmlinux_matches_kallsyms_filter) <= 0) {
 109                 pr_debug("machine__load_vmlinux_path ");
 110                 goto out;
 111         }
 112
 113         err = 0;
 114         /*
 115          * Step 7:
 116          *
 117          * Now look at the symbols in the vmlinux DSO and check if we find all of them
 118          * in the kallsyms dso. For the ones that are in both, check its names and
 119          * end addresses too.
 120          */
 121         for (nd = rb_first(&vmlinux_map->dso->symbols[type]); nd; nd = rb_next(nd)) {
 122                 struct symbol *pair, *first_pair;
 123                 bool backwards = true;
 124
 125                 sym  = rb_entry(nd, struct symbol, rb_node);
 126
 127                 if (sym->start == sym->end)
 128                         continue;
 129
 130                 first_pair = machine__find_kernel_symbol(&kallsyms, type, sym->start, NULL, NULL);
 131                 pair = first_pair;
 132
 133                 if (pair && pair->start == sym->start) {
 134 next_pair:
 135                         if (strcmp(sym->name, pair->name) == 0) {
 136                                 /*
 137                                  * kallsyms don't have the symbol end, so we
 138                                  * set that by using the next symbol start - 1,
 139                                  * in some cases we get this up to a page
 140                                  * wrong, trace_kmalloc when I was developing
 141                                  * this code was one such example, 2106 bytes
 142                                  * off the real size. More than that and we
 143                                  * _really_ have a problem.
 144                                  */
 145                                 s64 skew = sym->end - pair->end;
 146                                 if (llabs(skew) < page_size)
 147                                         continue;
 148
 149                                 pr_debug("%#" PRIx64 ": diff end addr for %s v: %#" PRIx64 " k: %#" PRIx64 "\n",
 150                                          sym->start, sym->name, sym->end, pair->end);
 151                         } else {
 152                                 struct rb_node *nnd;
 153 detour:
 154                                 nnd = backwards ? rb_prev(&pair->rb_node) :
 155                                                   rb_next(&pair->rb_node);
 156                                 if (nnd) {
 157                                         struct symbol *next = rb_entry(nnd, struct symbol, rb_node);
 158
 159                                         if (next->start == sym->start) {
 160                                                 pair = next;
 161                                                 goto next_pair;
 162                                         }
 163                                 }
 164
 165                                 if (backwards) {
 166                                         backwards = false;
 167                                         pair = first_pair;
 168                                         goto detour;
 169                                 }
 170
 171                                 pr_debug("%#" PRIx64 ": diff name v: %s k: %s\n",
 172                                          sym->start, sym->name, pair->name);
 173                         }
 174                 } else
 175                         pr_debug("%#" PRIx64 ": %s not on kallsyms\n", sym->start, sym->name);
 176
 177                 err = -1;
 178         }
 179
 180         if (!verbose)
 181                 goto out;
 182
 183         pr_info("Maps only in vmlinux:\n");
 184
 185         for (nd = rb_first(&vmlinux.kmaps.maps[type]); nd; nd = rb_next(nd)) {
 186                 struct map *pos = rb_entry(nd, struct map, rb_node), *pair;
 187                 /*
 188                  * If it is the kernel, kallsyms is always "[kernel.kallsyms]", while
 189                  * the kernel will have the path for the vmlinux file being used,
 190                  * so use the short name, less descriptive but the same ("[kernel]" in
 191                  * both cases.
 192                  */
 193                 pair = map_groups__find_by_name(&kallsyms.kmaps, type,
 194                                                 (pos->dso->kernel ?
 195                                                         pos->dso->short_name :
 196                                                         pos->dso->name));
 197                 if (pair)
 198                         pair->priv = 1;
 199                 else
 200                         map__fprintf(pos, stderr);
 201         }
 202
 203         pr_info("Maps in vmlinux with a different name in kallsyms:\n");
 204
 205         for (nd = rb_first(&vmlinux.kmaps.maps[type]); nd; nd = rb_next(nd)) {
 206                 struct map *pos = rb_entry(nd, struct map, rb_node), *pair;
 207
 208                 pair = map_groups__find(&kallsyms.kmaps, type, pos->start);
 209                 if (pair == NULL || pair->priv)
 210                         continue;
 211
 212                 if (pair->start == pos->start) {
 213                         pair->priv = 1;
 214                         pr_info(" %" PRIx64 "-%" PRIx64 " %" PRIx64 " %s in kallsyms as",
 215                                 pos->start, pos->end, pos->pgoff, pos->dso->name);
 216                         if (pos->pgoff != pair->pgoff || pos->end != pair->end)
 217                                 pr_info(": \n*%" PRIx64 "-%" PRIx64 " %" PRIx64 "",
 218                                         pair->start, pair->end, pair->pgoff);
 219                         pr_info(" %s\n", pair->dso->name);
 220                         pair->priv = 1;
 221                 }
 222         }
 223
 224         pr_info("Maps only in kallsyms:\n");
 225
 226         for (nd = rb_first(&kallsyms.kmaps.maps[type]);
 227              nd; nd = rb_next(nd)) {
 228                 struct map *pos = rb_entry(nd, struct map, rb_node);
 229
 230                 if (!pos->priv)
 231                         map__fprintf(pos, stderr);
 232         }
 233 out:
 234         return err;
 235 }
 236
 237 #include "util/cpumap.h"
 238 #include "util/evsel.h"
 239 #include <sys/types.h>
 240
 241 static int trace_event__id(const char *event_name)
 242 {
 243         char *filename;
 244         int err = -1, fd;
 245
 246         if (asprintf(&filename,
 247                      "/sys/kernel/debug/tracing/events/syscalls/%s/id",
 248                      event_name) < 0)
 249                 return -1;
 250
 251         fd = open(filename, O_RDONLY);
 252         if (fd >= 0) {
 253                 char id[16];
 254                 if (read(fd, id, sizeof(id)) > 0)
 255                         err = atoi(id);
 256                 close(fd);
 257         }
 258
 259         free(filename);
 260         return err;
 261 }
 262
 263 static int test__open_syscall_event(void)
 264 {
 265         int err = -1, fd;
 266         struct thread_map *threads;
 267         struct perf_evsel *evsel;
 268         struct perf_event_attr attr;
 269         unsigned int nr_open_calls = 111, i;
 270         int id = trace_event__id("sys_enter_open");
 271
 272         if (id < 0) {
 273                 pr_debug("is debugfs mounted on /sys/kernel/debug?\n");
 274                 return -1;
 275         }
 276
 277         threads = thread_map__new(-1, getpid());
 278         if (threads == NULL) {
 279                 pr_debug("thread_map__new\n");
 280                 return -1;
 281         }
 282
 283         memset(&attr, 0, sizeof(attr));
 284         attr.type = PERF_TYPE_TRACEPOINT;
 285         attr.config = id;
 286         evsel = perf_evsel__new(&attr, 0);
 287         if (evsel == NULL) {
 288                 pr_debug("perf_evsel__new\n");
 289                 goto out_thread_map_delete;
 290         }
 291
 292         if (perf_evsel__open_per_thread(evsel, threads) < 0) {
 293                 pr_debug("failed to open counter: %s, "
 294                          "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 295                          strerror(errno));
 296                 goto out_evsel_delete;
 297         }
 298
 299         for (i = 0; i < nr_open_calls; ++i) {
 300                 fd = open("/etc/passwd", O_RDONLY);
 301                 close(fd);
 302         }
 303
 304         if (perf_evsel__read_on_cpu(evsel, 0, 0) < 0) {
 305                 pr_debug("perf_evsel__open_read_on_cpu\n");
 306                 goto out_close_fd;
 307         }
 308
 309         if (evsel->counts->cpu[0].val != nr_open_calls) {
 310                 pr_debug("perf_evsel__read_on_cpu: expected to intercept %d calls, got %" PRIu64 "\n",
 311                          nr_open_calls, evsel->counts->cpu[0].val);
 312                 goto out_close_fd;
 313         }
 314
 315         err = 0;
 316 out_close_fd:
 317         perf_evsel__close_fd(evsel, 1, threads->nr);
 318 out_evsel_delete:
 319         perf_evsel__delete(evsel);
 320 out_thread_map_delete:
 321         thread_map__delete(threads);
 322         return err;
 323 }
 324
 325 #include <sched.h>
 326
 327 static int test__open_syscall_event_on_all_cpus(void)
 328 {
 329         int err = -1, fd, cpu;
 330         struct thread_map *threads;
 331         struct cpu_map *cpus;
 332         struct perf_evsel *evsel;
 333         struct perf_event_attr attr;
 334         unsigned int nr_open_calls = 111, i;
 335         cpu_set_t cpu_set;
 336         int id = trace_event__id("sys_enter_open");
 337
 338         if (id < 0) {
 339                 pr_debug("is debugfs mounted on /sys/kernel/debug?\n");
 340                 return -1;
 341         }
 342
 343         threads = thread_map__new(-1, getpid());
 344         if (threads == NULL) {
 345                 pr_debug("thread_map__new\n");
 346                 return -1;
 347         }
 348
 349         cpus = cpu_map__new(NULL);
 350         if (threads == NULL) {
 351                 pr_debug("thread_map__new\n");
 352                 return -1;
 353         }
 354
 355
 356         CPU_ZERO(&cpu_set);
 357
 358         memset(&attr, 0, sizeof(attr));
 359         attr.type = PERF_TYPE_TRACEPOINT;
 360         attr.config = id;
 361         evsel = perf_evsel__new(&attr, 0);
 362         if (evsel == NULL) {
 363                 pr_debug("perf_evsel__new\n");
 364                 goto out_thread_map_delete;
 365         }
 366
 367         if (perf_evsel__open(evsel, cpus, threads) < 0) {
 368                 pr_debug("failed to open counter: %s, "
 369                          "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 370                          strerror(errno));
 371                 goto out_evsel_delete;
 372         }
 373
 374         for (cpu = 0; cpu < cpus->nr; ++cpu) {
 375                 unsigned int ncalls = nr_open_calls + cpu;
 376                 /*
 377                  * XXX eventually lift this restriction in a way that
 378                  * keeps perf building on older glibc installations
 379                  * without CPU_ALLOC. 1024 cpus in 2010 still seems
 380                  * a reasonable upper limit tho :-)
 381                  */
 382                 if (cpus->map[cpu] >= CPU_SETSIZE) {
 383                         pr_debug("Ignoring CPU %d\n", cpus->map[cpu]);
 384                         continue;
 385                 }
 386
 387                 CPU_SET(cpus->map[cpu], &cpu_set);
 388                 if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) {
 389                         pr_debug("sched_setaffinity() failed on CPU %d: %s ",
 390                                  cpus->map[cpu],
 391                                  strerror(errno));
 392                         goto out_close_fd;
 393                 }
 394                 for (i = 0; i < ncalls; ++i) {
 395                         fd = open("/etc/passwd", O_RDONLY);
 396                         close(fd);
 397                 }
 398                 CPU_CLR(cpus->map[cpu], &cpu_set);
 399         }
 400
 401         /*
 402          * Here we need to explicitely preallocate the counts, as if
 403          * we use the auto allocation it will allocate just for 1 cpu,
 404          * as we start by cpu 0.
 405          */
 406         if (perf_evsel__alloc_counts(evsel, cpus->nr) < 0) {
 407                 pr_debug("perf_evsel__alloc_counts(ncpus=%d)\n", cpus->nr);
 408                 goto out_close_fd;
 409         }
 410
 411         for (cpu = 0; cpu < cpus->nr; ++cpu) {
 412                 unsigned int expected;
 413
 414                 if (cpus->map[cpu] >= CPU_SETSIZE)
 415                         continue;
 416
 417                 if (perf_evsel__read_on_cpu(evsel, cpu, 0) < 0) {
 418                         pr_debug("perf_evsel__open_read_on_cpu\n");
 419                         goto out_close_fd;
 420                 }
 421
 422                 expected = nr_open_calls + cpu;
 423                 if (evsel->counts->cpu[cpu].val != expected) {
 424                         pr_debug("perf_evsel__read_on_cpu: expected to intercept %d calls on cpu %d, got %" PRIu64 "\n",
 425                                  expected, cpus->map[cpu], evsel->counts->cpu[cpu].val);
 426                         goto out_close_fd;
 427                 }
 428         }
 429
 430         err = 0;
 431 out_close_fd:
 432         perf_evsel__close_fd(evsel, 1, threads->nr);
 433 out_evsel_delete:
 434         perf_evsel__delete(evsel);
 435 out_thread_map_delete:
 436         thread_map__delete(threads);
 437         return err;
 438 }
 439
 440 static struct test {
 441         const char *desc;
 442         int (*func)(void);
 443 } tests[] = {
 444         {
 445                 .desc = "vmlinux symtab matches kallsyms",
 446                 .func = test__vmlinux_matches_kallsyms,
 447         },
 448         {
 449                 .desc = "detect open syscall event",
 450                 .func = test__open_syscall_event,
 451         },
 452         {
 453                 .desc = "detect open syscall event on all cpus",
 454                 .func = test__open_syscall_event_on_all_cpus,
 455         },
 456         {
 457                 .func = NULL,
 458         },
 459 };
 460
 461 static int __cmd_test(void)
 462 {
 463         int i = 0;
 464
 465         page_size = sysconf(_SC_PAGE_SIZE);
 466
 467         while (tests[i].func) {
 468                 int err;
 469                 pr_info("%2d: %s:", i + 1, tests[i].desc);
 470                 pr_debug("\n--- start ---\n");
 471                 err = tests[i].func();
 472                 pr_debug("---- end ----\n%s:", tests[i].desc);
 473                 pr_info(" %s\n", err ? "FAILED!\n" : "Ok");
 474                 ++i;
 475         }
 476
 477         return 0;
 478 }
 479
 480 static const char * const test_usage[] = {
 481         "perf test [<options>]",
 482         NULL,
 483 };
 484
 485 static const struct option test_options[] = {
 486         OPT_INTEGER('v', "verbose", &verbose,
 487                     "be more verbose (show symbol address, etc)"),
 488         OPT_END()
 489 };
 490
 491 int cmd_test(int argc, const char **argv, const char *prefix __used)
 492 {
 493         argc = parse_options(argc, argv, test_options, test_usage, 0);
 494         if (argc)
 495                 usage_with_options(test_usage, test_options);
 496
 497         symbol_conf.priv_size = sizeof(int);
 498         symbol_conf.sort_by_name = true;
 499         symbol_conf.try_vmlinux_path = true;
 500
 501         if (symbol__init() < 0)
 502                 return -1;
 503
 504         setup_pager();
 505
 506         return __cmd_test();
 507 }