granicus.if.org Git - strace/blob - perf.c

   1 /*
   2  * Copyright (c) 2013 Ben Noordhuis <info@bnoordhuis.nl>
   3  * Copyright (c) 2013-2015 Dmitry V. Levin <ldv@altlinux.org>
   4  * Copyright (c) 2016 Eugene Syromyatnikov <evgsyr@gmail.com>
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. The name of the author may not be used to endorse or promote products
  16  *    derived from this software without specific prior written permission.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  28  */
  29
  30 #include "defs.h"
  31
  32 #include "perf_event_struct.h"
  33
  34 #include "xlat/hw_breakpoint_len.h"
  35 #include "xlat/hw_breakpoint_type.h"
  36 #include "xlat/perf_attr_size.h"
  37 #include "xlat/perf_branch_sample_type.h"
  38 #include "xlat/perf_event_open_flags.h"
  39 #include "xlat/perf_event_read_format.h"
  40 #include "xlat/perf_event_sample_format.h"
  41 #include "xlat/perf_hw_cache_id.h"
  42 #include "xlat/perf_hw_cache_op_id.h"
  43 #include "xlat/perf_hw_cache_op_result_id.h"
  44 #include "xlat/perf_hw_id.h"
  45 #include "xlat/perf_sw_ids.h"
  46 #include "xlat/perf_type_id.h"
  47
  48 struct pea_desc {
  49         struct perf_event_attr *attr;
  50         uint32_t size;
  51 };
  52
  53 static void
  54 free_pea_desc(void *pea_desc_ptr)
  55 {
  56         struct pea_desc *desc = pea_desc_ptr;
  57
  58         free(desc->attr);
  59         free(desc);
  60 }
  61
  62 static int
  63 fetch_perf_event_attr(struct tcb *const tcp, const kernel_ulong_t addr)
  64 {
  65         struct pea_desc *desc;
  66         struct perf_event_attr *attr;
  67         uint32_t size;
  68
  69         if (umove(tcp, addr + offsetof(struct perf_event_attr, size), &size)) {
  70                 printaddr(addr);
  71                 return 1;
  72         }
  73
  74         if (size > sizeof(*attr))
  75                 size = sizeof(*attr);
  76
  77         if (!size)
  78                 size = PERF_ATTR_SIZE_VER0;
  79
  80         /*
  81          * Kernel (rightfully) deems invalid attribute structures with size less
  82          * than first published format size, and we do the same.
  83          */
  84         if (size < PERF_ATTR_SIZE_VER0) {
  85                 printaddr(addr);
  86                 return 1;
  87         }
  88
  89         if (abbrev(tcp))
  90                 size = offsetofend(struct perf_event_attr, config);
  91
  92         /* Size should be multiple of 8, but kernel doesn't check for it */
  93         /* size &= ~7; */
  94
  95         attr = xcalloc(1, sizeof(*attr));
  96
  97         if (umoven_or_printaddr(tcp, addr, size, attr)) {
  98                 free(attr);
  99
 100                 return 1;
 101         }
 102
 103         desc = xmalloc(sizeof(*desc));
 104
 105         desc->attr = attr;
 106         desc->size = size;
 107
 108         set_tcb_priv_data(tcp, desc, free_pea_desc);
 109
 110         return 0;
 111 }
 112
 113 #define PRINT_XLAT(prefix, xlat, x, dflt) \
 114         do { \
 115                 tprints(prefix); \
 116                 printxval_search(xlat, x, dflt); \
 117         } while (0)
 118
 119 static void
 120 print_perf_event_attr(struct tcb *const tcp, const kernel_ulong_t addr)
 121 {
 122         static const char *precise_ip_desc[] = {
 123                 "arbitrary skid",
 124                 "constant skid",
 125                 "requested to have 0 skid",
 126                 "must have 0 skid",
 127         };
 128
 129         struct pea_desc *desc;
 130         struct perf_event_attr *attr;
 131         uint32_t size;
 132         uint32_t new_size;
 133         int use_new_size = 0;
 134
 135         /*
 136          * Amusingly, kernel accepts structures with only part of the field
 137          * present, so we making check like this (instead of checking
 138          * offsetofend against size) in order to print fields as kernel sees
 139          * them. This also should work great on big endian architectures.
 140          */
 141         #define _PERF_CHECK_FIELD(_field) \
 142                 do { \
 143                         if (offsetof(struct perf_event_attr, _field) >= size) \
 144                                 goto print_perf_event_attr_out; \
 145                 } while (0)
 146
 147         desc = get_tcb_priv_data(tcp);
 148
 149         attr = desc->attr;
 150         size = desc->size;
 151
 152         /* The only error which expected to change size field currently */
 153         if (tcp->u_error == E2BIG) {
 154                 if (umove(tcp, addr + offsetof(struct perf_event_attr, size),
 155                     &new_size))
 156                         use_new_size = -1;
 157                 else
 158                         use_new_size = 1;
 159         }
 160
 161         PRINT_XLAT("{type=", perf_type_id, attr->type, "PERF_TYPE_???");
 162         tprints(", size=");
 163         printxval(perf_attr_size, attr->size, "PERF_ATTR_SIZE_???");
 164
 165         if (use_new_size) {
 166                 tprints(" => ");
 167
 168                 if (use_new_size > 0)
 169                         printxval(perf_attr_size, new_size,
 170                                   "PERF_ATTR_SIZE_???");
 171                 else
 172                         tprints("???");
 173         }
 174
 175         switch (attr->type) {
 176         case PERF_TYPE_HARDWARE:
 177                 PRINT_XLAT(", config=", perf_hw_id, attr->config,
 178                            "PERF_COUNT_HW_???");
 179                 break;
 180         case PERF_TYPE_SOFTWARE:
 181                 PRINT_XLAT(", config=", perf_sw_ids, attr->config,
 182                            "PERF_COUNT_SW_???");
 183                 break;
 184         case PERF_TYPE_TRACEPOINT:
 185                 /*
 186                  * "The value to use in config can be obtained from under
 187                  * debugfs tracing/events/../../id if ftrace is enabled in the
 188                  * kernel."
 189                  */
 190                 tprintf(", config=%" PRIu64, attr->config);
 191                 break;
 192         case PERF_TYPE_HW_CACHE:
 193                 /*
 194                  * (perf_hw_cache_id) | (perf_hw_cache_op_id << 8) |
 195                  * (perf_hw_cache_op_result_id << 16)
 196                  */
 197                 PRINT_XLAT(", config=", perf_hw_cache_id, attr->config & 0xFF,
 198                            "PERF_COUNT_HW_CACHE_???");
 199                 PRINT_XLAT("|", perf_hw_cache_op_id, (attr->config >> 8) & 0xFF,
 200                            "PERF_COUNT_HW_CACHE_OP_???");
 201                 /*
 202                  * Current code (see set_ext_hw_attr in arch/x86/events/core.c,
 203                  * tile_map_cache_event in arch/tile/kernel/perf_event.c,
 204                  * arc_pmu_cache_event in arch/arc/kernel/perf_event.c,
 205                  * hw_perf_cache_event in arch/blackfin/kernel/perf_event.c,
 206                  * _hw_perf_cache_event in arch/metag/kernel/perf/perf_event.c,
 207                  * mipspmu_map_cache_event in arch/mips/kernel/perf_event_mipsxx.c,
 208                  * hw_perf_cache_event in arch/powerpc/perf/core-book3s.c,
 209                  * hw_perf_cache_event in arch/powerpc/perf/core-fsl-emb.c,
 210                  * hw_perf_cache_event in arch/sh/kernel/perf_event.c,
 211                  * sparc_map_cache_event in arch/sparc/kernel/perf_event.c,
 212                  * xtensa_pmu_cache_event in arch/xtensa/kernel/perf_event.c,
 213                  * armpmu_map_cache_event in drivers/perf/arm_pmu.c) assumes
 214                  * that cache result is 8 bits in size.
 215                  */
 216                 PRINT_XLAT("<<8|", perf_hw_cache_op_result_id,
 217                            (attr->config >> 16) & 0xFF,
 218                            "PERF_COUNT_HW_CACHE_RESULT_???");
 219                 tprints("<<16");
 220                 if (attr->config >> 24)
 221                         tprintf("|%#" PRIx64 "<<24 "
 222                                 "/* PERF_COUNT_HW_CACHE_??? */",
 223                                 attr->config >> 24);
 224                 break;
 225         case PERF_TYPE_RAW:
 226                 /*
 227                  * "If type is PERF_TYPE_RAW, then a custom "raw" config
 228                  * value is needed. Most CPUs support events that are not
 229                  * covered by the "generalized" events. These are
 230                  * implementation defined; see your CPU manual (for example the
 231                  * Intel Volume 3B documentation or the AMD BIOS and Kernel
 232                  * Developer Guide). The libpfm4 library can be used to
 233                  * translate from the name in the architectural manuals
 234                  * to the raw hex value perf_event_open() expects in this
 235                  * field."
 236                  */
 237         case PERF_TYPE_BREAKPOINT:
 238                 /*
 239                  * "If type is PERF_TYPE_BREAKPOINT, then leave config set
 240                  * to zero. Its parameters are set in other places."
 241                  */
 242         default:
 243                 tprintf(", config=%#" PRIx64, attr->config);
 244                 break;
 245         }
 246
 247         if (abbrev(tcp))
 248                 goto print_perf_event_attr_out;
 249
 250         if (attr->freq)
 251                 tprintf(", sample_freq=%" PRIu64, attr->sample_freq);
 252         else
 253                 tprintf(", sample_period=%" PRIu64, attr->sample_period);
 254
 255         tprints(", sample_type=");
 256         printflags64(perf_event_sample_format, attr->sample_type,
 257                 "PERF_SAMPLE_???");
 258
 259         tprints(", read_format=");
 260         printflags64(perf_event_read_format, attr->read_format,
 261                 "PERF_FORMAT_???");
 262
 263         tprintf(", disabled=%u"
 264                 ", inherit=%u"
 265                 ", pinned=%u"
 266                 ", exclusive=%u"
 267                 ", exclusive_user=%u"
 268                 ", exclude_kernel=%u"
 269                 ", exclude_hv=%u"
 270                 ", exclude_idle=%u"
 271                 ", mmap=%u"
 272                 ", comm=%u"
 273                 ", freq=%u"
 274                 ", inherit_stat=%u"
 275                 ", enable_on_exec=%u"
 276                 ", task=%u"
 277                 ", watermark=%u"
 278                 ", precise_ip=%u /* %s */"
 279                 ", mmap_data=%u"
 280                 ", sample_id_all=%u"
 281                 ", exclude_host=%u"
 282                 ", exclude_guest=%u"
 283                 ", exclude_callchain_kernel=%u"
 284                 ", exclude_callchain_user=%u"
 285                 ", mmap2=%u"
 286                 ", comm_exec=%u"
 287                 ", use_clockid=%u"
 288                 ", context_switch=%u"
 289                 ", write_backward=%u",
 290                 attr->disabled,
 291                 attr->inherit,
 292                 attr->pinned,
 293                 attr->exclusive,
 294                 attr->exclude_user,
 295                 attr->exclude_kernel,
 296                 attr->exclude_hv,
 297                 attr->exclude_idle,
 298                 attr->mmap,
 299                 attr->comm,
 300                 attr->freq,
 301                 attr->inherit_stat,
 302                 attr->enable_on_exec,
 303                 attr->task,
 304                 attr->watermark,
 305                 attr->precise_ip, precise_ip_desc[attr->precise_ip],
 306                 attr->mmap_data,
 307                 attr->sample_id_all,
 308                 attr->exclude_host,
 309                 attr->exclude_guest,
 310                 attr->exclude_callchain_kernel,
 311                 attr->exclude_callchain_user,
 312                 attr->mmap2,
 313                 attr->comm_exec,
 314                 attr->use_clockid,
 315                 attr->context_switch,
 316                 attr->write_backward);
 317
 318         /*
 319          * Print it only in case it is non-zero, since it may contain flags we
 320          * are not aware about.
 321          */
 322         if (attr->__reserved_1)
 323                 tprintf(", __reserved_1=%#" PRIx64 " /* Bits 63..28 */",
 324                         (uint64_t) attr->__reserved_1);
 325
 326         if (attr->watermark)
 327                 tprintf(", wakeup_watermark=%u", attr->wakeup_watermark);
 328         else
 329                 tprintf(", wakeup_events=%u", attr->wakeup_events);
 330
 331         if (attr->type == PERF_TYPE_BREAKPOINT)
 332                 /* Any combination of R/W with X is deemed invalid */
 333                 PRINT_XLAT(", bp_type=", hw_breakpoint_type, attr->bp_type,
 334                            (attr->bp_type <=
 335                                    (HW_BREAKPOINT_X | HW_BREAKPOINT_RW)) ?
 336                                            "HW_BREAKPOINT_INVALID" :
 337                                            "HW_BREAKPOINT_???");
 338
 339         if (attr->type == PERF_TYPE_BREAKPOINT)
 340                 tprintf(", bp_addr=%#" PRIx64, attr->bp_addr);
 341         else
 342                 tprintf(", config1=%#" PRIx64, attr->config1);
 343
 344         /*
 345          * Fields after bp_addr/config1 are optional and may not present; check
 346          * against size is needed.
 347          */
 348
 349         _PERF_CHECK_FIELD(bp_len);
 350         if (attr->type == PERF_TYPE_BREAKPOINT)
 351                 tprintf(", bp_len=%" PRIu64, attr->bp_len);
 352         else
 353                 tprintf(", config2=%#" PRIx64, attr->config2);
 354
 355         _PERF_CHECK_FIELD(branch_sample_type);
 356         if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
 357                 tprints(", branch_sample_type=");
 358                 printflags64(perf_branch_sample_type, attr->branch_sample_type,
 359                              "PERF_SAMPLE_BRANCH_???");
 360         }
 361
 362         _PERF_CHECK_FIELD(sample_regs_user);
 363         /*
 364          * "This bit mask defines the set of user CPU registers to dump on
 365          * samples. The layout of the register mask is architecture-specific and
 366          * described in the kernel header
 367          * arch/ARCH/include/uapi/asm/perf_regs.h."
 368          */
 369         tprintf(", sample_regs_user=%#" PRIx64, attr->sample_regs_user);
 370
 371         _PERF_CHECK_FIELD(sample_stack_user);
 372         /*
 373          * "size of the user stack to dump if PERF_SAMPLE_STACK_USER is
 374          * specified."
 375          */
 376         if (attr->sample_type & PERF_SAMPLE_STACK_USER)
 377                 tprintf(", sample_stack_user=%#" PRIx32,
 378                         attr->sample_stack_user);
 379
 380         if (attr->use_clockid) {
 381                 _PERF_CHECK_FIELD(clockid);
 382                 tprints(", clockid=");
 383                 printxval(clocknames, attr->clockid, "CLOCK_???");
 384         }
 385
 386         _PERF_CHECK_FIELD(sample_regs_intr);
 387         tprintf(", sample_regs_intr=%#" PRIx64, attr->sample_regs_intr);
 388
 389         _PERF_CHECK_FIELD(aux_watermark);
 390         tprintf(", aux_watermark=%" PRIu32, attr->aux_watermark);
 391
 392         _PERF_CHECK_FIELD(sample_max_stack);
 393         tprintf(", sample_max_stack=%" PRIu16, attr->sample_max_stack);
 394
 395         /* _PERF_CHECK_FIELD(__reserved_2);
 396         tprintf(", __reserved2=%" PRIu16, attr->__reserved_2); */
 397
 398 print_perf_event_attr_out:
 399         if ((attr->size && (attr->size > size)) ||
 400             (!attr->size && (size < PERF_ATTR_SIZE_VER0)))
 401                 tprints(", ...");
 402
 403         tprints("}");
 404 }
 405
 406 SYS_FUNC(perf_event_open)
 407 {
 408         /*
 409          * We try to copy out the whole structure on entering in order to check
 410          * size value on exiting. We do not check the rest of the fields because
 411          * they shouldn't be changed, but copy the whole structure instead
 412          * of just size field because they could.
 413          */
 414         if (entering(tcp)) {
 415                 if (!fetch_perf_event_attr(tcp, tcp->u_arg[0]))
 416                         return 0;
 417         } else {
 418                 print_perf_event_attr(tcp, tcp->u_arg[0]);
 419         }
 420
 421         tprintf(", %d, %d, %d, ",
 422                 (int) tcp->u_arg[1],
 423                 (int) tcp->u_arg[2],
 424                 (int) tcp->u_arg[3]);
 425         printflags64(perf_event_open_flags, tcp->u_arg[4], "PERF_FLAG_???");
 426
 427         return RVAL_DECODED | RVAL_FD;
 428 }