usertrace-fast/ltt-usertrace-fast.c

   1 /* LTTng user-space "fast" library
   2  *
   3  * This daemon is spawned by each traced thread (to share the mmap).
   4  *
   5  * Its job is to dump periodically this buffer to disk (when it receives a
   6  * SIGUSR1 from its parent).
   7  *
   8  * It uses the control information in the shared memory area (producer/consumer
   9  * count).
  10  *
  11  * When the parent thread dies (yes, those thing may happen) ;) , this daemon
  12  * will flush the last buffer and write it to disk.
  13  *
  14  * Supplement note for streaming : the daemon is responsible for flushing
  15  * periodically the buffer if it is streaming data.
  16  *
  17  *
  18  * Notes :
  19  * shm memory is typically limited to 4096 units (system wide limit SHMMNI in
  20  * /proc/sys/kernel/shmmni). As it requires computation time upon creation, we
  21  * do not use it : we will use a shared mmap() instead which is passed through
  22  * the fork().
  23  * MAP_SHARED mmap segment. Updated when msync or munmap are called.
  24  * MAP_ANONYMOUS.
  25  * Memory  mapped  by  mmap()  is  preserved across fork(2), with the same
  26  *   attributes.
  27  *
  28  * Eventually, there will be two mode :
  29  * * Slow thread spawn : a fork() is done for each new thread. If the process
  30  *   dies, the data is not lost.
  31  * * Fast thread spawn : a pthread_create() is done by the application for each
  32  *   new thread.
  33  *
  34  * We use a timer to check periodically if the parent died. I think it is less
  35  * intrusive than a ptrace() on the parent, which would get every signal. The
  36  * side effect of this is that we won't be notified if the parent does an
  37  * exec(). In this case, we will just sit there until the parent exits.
  38  *
  39  *
  40  * Copyright 2006 Mathieu Desnoyers
  41  *
  42  */
  43
  44 #define inline inline __attribute__((always_inline))
  45
  46 #define _GNU_SOURCE
  47 #define LTT_TRACE
  48 #include <sys/types.h>
  49 #include <sys/wait.h>
  50 #include <unistd.h>
  51 #include <stdlib.h>
  52 #include <stdio.h>
  53 #include <signal.h>
  54 #include <syscall.h>
  55 #include <features.h>
  56 #include <pthread.h>
  57 #include <malloc.h>
  58 #include <string.h>
  59 #include <sys/mman.h>
  60 #include <signal.h>
  61 #include <sys/stat.h>
  62 #include <fcntl.h>
  63 #include <stdlib.h>
  64 #include <sys/param.h>
  65 #include <sys/time.h>
  66 #include <errno.h>
  67
  68 #include <asm/atomic.h>
  69 #include <asm/timex.h>  //for get_cycles()
  70
  71 _syscall0(pid_t,gettid)
  72
  73 #include <ltt/ltt-usertrace-fast.h>
  74
  75 #ifdef LTT_SHOW_DEBUG
  76 #define dbg_printf(...) dbg_printf(__VA_ARGS__)
  77 #else
  78 #define dbg_printf(...)
  79 #endif //LTT_SHOW_DEBUG
  80
  81
  82 enum force_switch_mode { FORCE_ACTIVE, FORCE_FLUSH };
  83
  84 /* Writer (the traced application) */
  85
  86 __thread struct ltt_trace_info *thread_trace_info = NULL;
  87
  88 void ltt_usertrace_fast_buffer_switch(void)
  89 {
  90         struct ltt_trace_info *tmp = thread_trace_info;
  91         if(tmp)
  92                 kill(tmp->daemon_id, SIGUSR1);
  93 }
  94
  95 /* The cleanup should never be called from a signal handler */
  96 static void ltt_usertrace_fast_cleanup(void *arg)
  97 {
  98         struct ltt_trace_info *tmp = thread_trace_info;
  99         if(tmp) {
 100                 thread_trace_info = NULL;
 101                 kill(tmp->daemon_id, SIGUSR2);
 102                 munmap(tmp, sizeof(*tmp));
 103         }
 104 }
 105
 106 /* Reader (the disk dumper daemon) */
 107
 108 static pid_t traced_pid = 0;
 109 static pid_t traced_tid = 0;
 110 static int parent_exited = 0;
 111
 112 /* signal handling */
 113 static void handler_sigusr1(int signo)
 114 {
 115         dbg_printf("LTT Signal %d received : parent buffer switch.\n", signo);
 116 }
 117
 118 static void handler_sigusr2(int signo)
 119 {
 120         dbg_printf("LTT Signal %d received : parent exited.\n", signo);
 121         parent_exited = 1;
 122 }
 123
 124 static void handler_sigalarm(int signo)
 125 {
 126         dbg_printf("LTT Signal %d received\n", signo);
 127
 128         if(getppid() != traced_pid) {
 129                 /* Parent died */
 130                 dbg_printf("LTT Parent %lu died, cleaning up\n", traced_pid);
 131                 traced_pid = 0;
 132         }
 133         alarm(3);
 134 }
 135
 136 /* Do a buffer switch. Don't switch if buffer is completely empty */
 137 static void flush_buffer(struct ltt_buf *ltt_buf, enum force_switch_mode mode)
 138 {
 139         uint64_t tsc;
 140         int offset_begin, offset_end, offset_old;
 141         int reserve_commit_diff;
 142         int consumed_old, consumed_new;
 143         int commit_count, reserve_count;
 144         int end_switch_old;
 145
 146         do {
 147                 offset_old = atomic_read(&ltt_buf->offset);
 148                 offset_begin = offset_old;
 149                 end_switch_old = 0;
 150                 tsc = ltt_get_timestamp();
 151                 if(tsc == 0) {
 152                         /* Error in getting the timestamp : should not happen : it would
 153                          * mean we are called from an NMI during a write seqlock on xtime. */
 154                         return;
 155                 }
 156
 157                 if(SUBBUF_OFFSET(offset_begin, ltt_buf) != 0) {
 158                         offset_begin = SUBBUF_ALIGN(offset_begin, ltt_buf);
 159                         end_switch_old = 1;
 160                 } else {
 161       /* we do not have to switch : buffer is empty */
 162       return;
 163     }
 164                 if(mode == FORCE_ACTIVE)
 165                         offset_begin += ltt_subbuf_header_len(ltt_buf);
 166                 /* Always begin_switch in FORCE_ACTIVE mode */
 167
 168                 /* Test new buffer integrity */
 169                 reserve_commit_diff =
 170                         atomic_read(
 171                                 &ltt_buf->reserve_count[SUBBUF_INDEX(offset_begin, ltt_buf)])
 172                         - atomic_read(
 173                                         &ltt_buf->commit_count[SUBBUF_INDEX(offset_begin, ltt_buf)]);
 174                 if(reserve_commit_diff == 0) {
 175       /* Next buffer not corrupted. */
 176       if(mode == FORCE_ACTIVE
 177                                 && (offset_begin-atomic_read(&ltt_buf->consumed))
 178                                                                                         >= ltt_buf->alloc_size) {
 179         /* We do not overwrite non consumed buffers and we are full : ignore
 180                  switch while tracing is active. */
 181         return;
 182       }
 183     } else {
 184       /* Next subbuffer corrupted. Force pushing reader even in normal mode */
 185     }
 186
 187                 offset_end = offset_begin;
 188         } while(atomic_cmpxchg(&ltt_buf->offset, offset_old, offset_end)
 189                                                         != offset_old);
 190
 191
 192         if(mode == FORCE_ACTIVE) {
 193                 /* Push the reader if necessary */
 194                 do {
 195                         consumed_old = atomic_read(&ltt_buf->consumed);
 196                         /* If buffer is in overwrite mode, push the reader consumed count if
 197                                  the write position has reached it and we are not at the first
 198                                  iteration (don't push the reader farther than the writer).
 199                                  This operation can be done concurrently by many writers in the
 200                                  same buffer, the writer being at the fartest write position sub-buffer
 201                                  index in the buffer being the one which will win this loop. */
 202                         /* If the buffer is not in overwrite mode, pushing the reader only
 203                                  happen if a sub-buffer is corrupted */
 204                         if((SUBBUF_TRUNC(offset_end, ltt_buf)
 205                                         - SUBBUF_TRUNC(consumed_old, ltt_buf))
 206                                                         >= ltt_buf->alloc_size)
 207                                 consumed_new = SUBBUF_ALIGN(consumed_old, ltt_buf);
 208                         else {
 209                                 consumed_new = consumed_old;
 210                                 break;
 211                         }
 212                 } while(atomic_cmpxchg(&ltt_buf->consumed, consumed_old, consumed_new)
 213                                 != consumed_old);
 214
 215                 if(consumed_old != consumed_new) {
 216                         /* Reader pushed : we are the winner of the push, we can therefore
 217                                  reequilibrate reserve and commit. Atomic increment of the commit
 218                                  count permits other writers to play around with this variable
 219                                  before us. We keep track of corrupted_subbuffers even in overwrite
 220                                  mode :
 221                                  we never want to write over a non completely committed sub-buffer :
 222                                  possible causes : the buffer size is too low compared to the unordered
 223                                  data input, or there is a writer who died between the reserve and the
 224                                  commit. */
 225                         if(reserve_commit_diff) {
 226                                 /* We have to alter the sub-buffer commit count : a sub-buffer is
 227                                          corrupted */
 228                                 atomic_add(reserve_commit_diff,
 229                                                                 &ltt_buf->commit_count[SUBBUF_INDEX(offset_begin, ltt_buf)]);
 230                                 atomic_inc(&ltt_buf->corrupted_subbuffers);
 231                         }
 232                 }
 233         }
 234
 235         /* Always switch */
 236
 237         if(end_switch_old) {
 238                 /* old subbuffer */
 239                 /* Concurrency safe because we are the last and only thread to alter this
 240                          sub-buffer. As long as it is not delivered and read, no other thread can
 241                          alter the offset, alter the reserve_count or call the
 242                          client_buffer_end_callback on this sub-buffer.
 243                          The only remaining threads could be the ones with pending commits. They
 244                          will have to do the deliver themself.
 245                          Not concurrency safe in overwrite mode. We detect corrupted subbuffers with
 246                          commit and reserve counts. We keep a corrupted sub-buffers count and push
 247                          the readers across these sub-buffers.
 248                          Not concurrency safe if a writer is stalled in a subbuffer and
 249                          another writer switches in, finding out it's corrupted. The result will be
 250                          than the old (uncommited) subbuffer will be declared corrupted, and that
 251                          the new subbuffer will be declared corrupted too because of the commit
 252                          count adjustment.
 253                          Offset old should never be 0. */
 254                 ltt_buffer_end_callback(ltt_buf, tsc, offset_old,
 255                                 SUBBUF_INDEX((offset_old), ltt_buf));
 256                 /* Setting this reserve_count will allow the sub-buffer to be delivered by
 257                          the last committer. */
 258                 reserve_count = atomic_add_return((SUBBUF_OFFSET((offset_old-1),
 259                                                       ltt_buf) + 1),
 260                                                                                 &ltt_buf->reserve_count[SUBBUF_INDEX((offset_old),
 261                                                           ltt_buf)]);
 262                 if(reserve_count == atomic_read(
 263                                 &ltt_buf->commit_count[SUBBUF_INDEX((offset_old), ltt_buf)])) {
 264                         ltt_deliver_callback(ltt_buf, SUBBUF_INDEX((offset_old), ltt_buf), NULL);
 265                 }
 266         }
 267
 268         if(mode == FORCE_ACTIVE) {
 269                 /* New sub-buffer */
 270                 /* This code can be executed unordered : writers may already have written
 271                          to the sub-buffer before this code gets executed, caution. */
 272                 /* The commit makes sure that this code is executed before the deliver
 273                          of this sub-buffer */
 274                 ltt_buffer_begin_callback(ltt_buf, tsc, SUBBUF_INDEX(offset_begin, ltt_buf));
 275                 commit_count = atomic_add_return(ltt_subbuf_header_len(ltt_buf),
 276                                                                  &ltt_buf->commit_count[SUBBUF_INDEX(offset_begin, ltt_buf)]);
 277                 /* Check if the written buffer has to be delivered */
 278                 if(commit_count == atomic_read(
 279                                         &ltt_buf->reserve_count[SUBBUF_INDEX(offset_begin, ltt_buf)])) {
 280                         ltt_deliver_callback(ltt_buf, SUBBUF_INDEX(offset_begin, ltt_buf), NULL);
 281                 }
 282         }
 283
 284 }
 285
 286 static inline int ltt_buffer_get(struct ltt_buf *ltt_buf,
 287                 unsigned int *offset)
 288 {
 289         unsigned int consumed_old, consumed_idx;
 290         consumed_old = atomic_read(&ltt_buf->consumed);
 291         consumed_idx = SUBBUF_INDEX(consumed_old, ltt_buf);
 292
 293         if(atomic_read(&ltt_buf->commit_count[consumed_idx])
 294                 != atomic_read(&ltt_buf->reserve_count[consumed_idx])) {
 295                 return -EAGAIN;
 296         }
 297         if((SUBBUF_TRUNC(atomic_read(&ltt_buf->offset), ltt_buf)
 298                                 -SUBBUF_TRUNC(consumed_old, ltt_buf)) == 0) {
 299                 return -EAGAIN;
 300         }
 301
 302         *offset = consumed_old;
 303
 304         return 0;
 305 }
 306
 307 static inline int ltt_buffer_put(struct ltt_buf *ltt_buf,
 308                 unsigned int offset)
 309 {
 310         unsigned int consumed_old, consumed_new;
 311         int ret;
 312
 313         consumed_old = offset;
 314         consumed_new = SUBBUF_ALIGN(consumed_old, ltt_buf);
 315         if(atomic_cmpxchg(&ltt_buf->consumed, consumed_old, consumed_new)
 316                         != consumed_old) {
 317                 /* We have been pushed by the writer : the last buffer read _is_
 318                  * corrupted!
 319                  * It can also happen if this is a buffer we never got. */
 320                 return -EIO;
 321         } else {
 322                 if(atomic_inc_return(&ltt_buf->writer_futex) <= 0) {
 323                         atomic_set(&ltt_buf->writer_futex, 1);
 324                         /* tell the client that buffer is now unfull */
 325                         ret = futex((unsigned long)&ltt_buf->writer_futex,
 326                                         FUTEX_WAKE, 1, 0, 0, 0);
 327                         if(ret != 1) {
 328                                 dbg_printf("LTT warning : race condition : writer not waiting or too many writers\n");
 329                         }
 330                 }
 331         }
 332 }
 333
 334 static int read_subbuffer(struct ltt_buf *ltt_buf, int fd)
 335 {
 336         unsigned int consumed_old;
 337         int err;
 338         dbg_printf("LTT read buffer\n");
 339
 340
 341         err = ltt_buffer_get(ltt_buf, &consumed_old);
 342         if(err != 0) {
 343                 if(err != -EAGAIN) dbg_printf("LTT Reserving sub buffer failed\n");
 344                 goto get_error;
 345         }
 346
 347         err = TEMP_FAILURE_RETRY(write(fd,
 348                                 ltt_buf->start
 349                                         + (consumed_old & ((ltt_buf->alloc_size)-1)),
 350                                 ltt_buf->subbuf_size));
 351
 352         if(err < 0) {
 353                 perror("Error in writing to file");
 354                 goto write_error;
 355         }
 356 #if 0
 357         err = fsync(pair->trace);
 358         if(err < 0) {
 359                 ret = errno;
 360                 perror("Error in writing to file");
 361                 goto write_error;
 362         }
 363 #endif //0
 364 write_error:
 365         err = ltt_buffer_put(ltt_buf, consumed_old);
 366
 367         if(err != 0) {
 368                 if(err == -EIO) {
 369                         dbg_printf("Reader has been pushed by the writer, last subbuffer corrupted.\n");
 370                         /* FIXME : we may delete the last written buffer if we wish. */
 371                 }
 372                 goto get_error;
 373         }
 374
 375 get_error:
 376         return err;
 377 }
 378
 379 /* This function is called by ltt_rw_init which has signals blocked */
 380 static void ltt_usertrace_fast_daemon(struct ltt_trace_info *shared_trace_info,
 381                 sigset_t oldset, pid_t l_traced_pid, pthread_t l_traced_tid)
 382 {
 383         struct sigaction act;
 384         int ret;
 385         int fd_process;
 386         char outfile_name[PATH_MAX];
 387         char identifier_name[PATH_MAX];
 388
 389
 390         traced_pid = l_traced_pid;
 391         traced_tid = l_traced_tid;
 392
 393         dbg_printf("LTT ltt_usertrace_fast_daemon : init is %d, pid is %lu, traced_pid is %lu, traced_tid is %lu\n",
 394                         shared_trace_info->init, getpid(), traced_pid, traced_tid);
 395
 396         act.sa_handler = handler_sigusr1;
 397         act.sa_flags = 0;
 398         sigemptyset(&(act.sa_mask));
 399         sigaddset(&(act.sa_mask), SIGUSR1);
 400         sigaction(SIGUSR1, &act, NULL);
 401
 402         act.sa_handler = handler_sigusr2;
 403         act.sa_flags = 0;
 404         sigemptyset(&(act.sa_mask));
 405         sigaddset(&(act.sa_mask), SIGUSR2);
 406         sigaction(SIGUSR2, &act, NULL);
 407
 408         act.sa_handler = handler_sigalarm;
 409         act.sa_flags = 0;
 410         sigemptyset(&(act.sa_mask));
 411         sigaddset(&(act.sa_mask), SIGALRM);
 412         sigaction(SIGALRM, &act, NULL);
 413
 414         /* Enable signals */
 415         ret = pthread_sigmask(SIG_SETMASK, &oldset, NULL);
 416         if(ret) {
 417                 dbg_printf("LTT Error in pthread_sigmask\n");
 418         }
 419
 420         alarm(3);
 421
 422         /* Open output files */
 423         umask(00000);
 424         ret = mkdir(LTT_USERTRACE_ROOT, 0777);
 425         if(ret < 0 && errno != EEXIST) {
 426                 perror("LTT Error in creating output (mkdir)");
 427                 exit(-1);
 428         }
 429         ret = chdir(LTT_USERTRACE_ROOT);
 430         if(ret < 0) {
 431                 perror("LTT Error in creating output (chdir)");
 432                 exit(-1);
 433         }
 434         snprintf(identifier_name, PATH_MAX-1,   "%lu.%lu.%llu",
 435                         traced_tid, traced_pid, get_cycles());
 436         snprintf(outfile_name, PATH_MAX-1,      "process-%s", identifier_name);
 437 #ifndef LTT_NULL_OUTPUT_TEST
 438         fd_process = creat(outfile_name, 0644);
 439 #else
 440         /* NULL test */
 441         ret = symlink("/dev/null", outfile_name);
 442         if(ret < 0) {
 443                 perror("error in symlink");
 444         }
 445         fd_process = open(outfile_name, O_WRONLY);
 446         if(fd_process < 0) {
 447                 perror("Error in open");
 448         }
 449 #endif //LTT_NULL_OUTPUT_TEST
 450
 451         while(1) {
 452                 pause();
 453                 if(traced_pid == 0) break; /* parent died */
 454                 if(parent_exited) break;
 455                 dbg_printf("LTT Doing a buffer switch read. pid is : %lu\n", getpid());
 456
 457                 do {
 458                         ret = read_subbuffer(&shared_trace_info->channel.process, fd_process);
 459                 } while(ret == 0);
 460         }
 461
 462         /* The parent thread is dead and we have finished with the buffer */
 463
 464         /* Buffer force switch (flush). Using FLUSH instead of ACTIVE because we know
 465          * there is no writer. */
 466         flush_buffer(&shared_trace_info->channel.process, FORCE_FLUSH);
 467         do {
 468                 ret = read_subbuffer(&shared_trace_info->channel.process, fd_process);
 469         } while(ret == 0);
 470
 471
 472         close(fd_process);
 473
 474         munmap(shared_trace_info, sizeof(*shared_trace_info));
 475
 476         exit(0);
 477 }
 478
 479
 480 /* Reader-writer initialization */
 481
 482 static enum ltt_process_role { LTT_ROLE_WRITER, LTT_ROLE_READER }
 483         role = LTT_ROLE_WRITER;
 484
 485
 486 void ltt_rw_init(void)
 487 {
 488         pid_t pid;
 489         struct ltt_trace_info *shared_trace_info;
 490         int ret;
 491         sigset_t set, oldset;
 492         pid_t l_traced_pid = getpid();
 493         pid_t l_traced_tid = gettid();
 494
 495         /* parent : create the shared memory map */
 496         shared_trace_info = mmap(0, sizeof(*thread_trace_info),
 497                         PROT_READ|PROT_WRITE, MAP_SHARED|MAP_ANONYMOUS, 0, 0);
 498         shared_trace_info->init=0;
 499         shared_trace_info->filter=0;
 500         shared_trace_info->daemon_id=0;
 501         shared_trace_info->nesting=0;
 502         memset(&shared_trace_info->channel.process, 0,
 503                         sizeof(shared_trace_info->channel.process));
 504         /* Tricky semaphore : is in a shared memory space, so it's ok for a fast
 505          * mutex (futex). */
 506         atomic_set(&shared_trace_info->channel.process.writer_futex, LTT_N_SUBBUFS);
 507         shared_trace_info->channel.process.alloc_size = LTT_BUF_SIZE_PROCESS;
 508         shared_trace_info->channel.process.subbuf_size = LTT_SUBBUF_SIZE_PROCESS;
 509         shared_trace_info->channel.process.start =
 510                                                 shared_trace_info->channel.process_buf;
 511         ltt_buffer_begin_callback(&shared_trace_info->channel.process,
 512                         ltt_get_timestamp(), 0);
 513
 514         shared_trace_info->init = 1;
 515
 516         /* Disable signals */
 517   ret = sigfillset(&set);
 518   if(ret) {
 519     dbg_printf("LTT Error in sigfillset\n");
 520   }
 521
 522
 523   ret = pthread_sigmask(SIG_BLOCK, &set, &oldset);
 524   if(ret) {
 525     dbg_printf("LTT Error in pthread_sigmask\n");
 526   }
 527
 528         pid = fork();
 529         if(pid > 0) {
 530                 /* Parent */
 531                 shared_trace_info->daemon_id = pid;
 532                 thread_trace_info = shared_trace_info;
 533
 534                 /* Enable signals */
 535                 ret = pthread_sigmask(SIG_SETMASK, &oldset, NULL);
 536                 if(ret) {
 537                         dbg_printf("LTT Error in pthread_sigmask\n");
 538                 }
 539         } else if(pid == 0) {
 540                 pid_t sid;
 541                 /* Child */
 542                 role = LTT_ROLE_READER;
 543                 sid = setsid();
 544                 //Not a good idea to renice, unless futex wait eventually implement
 545                 //priority inheritence.
 546                 //ret = nice(1);
 547                 //if(ret < 0) {
 548                 //      perror("Error in nice");
 549                 //}
 550                 if(sid < 0) {
 551                         perror("Error setting sid");
 552                 }
 553                 ltt_usertrace_fast_daemon(shared_trace_info, oldset, l_traced_pid,
 554                                         l_traced_tid);
 555                 /* Should never return */
 556                 exit(-1);
 557         } else if(pid < 0) {
 558                 /* fork error */
 559                 perror("LTT Error in forking ltt-usertrace-fast");
 560         }
 561 }
 562
 563 static __thread struct _pthread_cleanup_buffer cleanup_buffer;
 564
 565 void ltt_thread_init(void)
 566 {
 567         _pthread_cleanup_push(&cleanup_buffer, ltt_usertrace_fast_cleanup, NULL);
 568         ltt_rw_init();
 569 }
 570
 571 void __attribute__((constructor)) __ltt_usertrace_fast_init(void)
 572 {
 573   dbg_printf("LTT usertrace-fast init\n");
 574
 575         ltt_rw_init();
 576 }
 577
 578 void __attribute__((destructor)) __ltt_usertrace_fast_fini(void)
 579 {
 580         if(role == LTT_ROLE_WRITER) {
 581           dbg_printf("LTT usertrace-fast fini\n");
 582                 ltt_usertrace_fast_cleanup(NULL);
 583         }
 584 }
 585