src/common/kernel-consumer/kernel-consumer.c

   1 /*
   2  * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
   3  *                      Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License, version 2 only,
   7  * as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along
  15  * with this program; if not, write to the Free Software Foundation, Inc.,
  16  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  17  */
  18
  19 #define _GNU_SOURCE
  20 #include <assert.h>
  21 #include <poll.h>
  22 #include <pthread.h>
  23 #include <stdlib.h>
  24 #include <string.h>
  25 #include <sys/mman.h>
  26 #include <sys/socket.h>
  27 #include <sys/types.h>
  28 #include <inttypes.h>
  29 #include <unistd.h>
  30 #include <sys/stat.h>
  31
  32 #include <common/common.h>
  33 #include <common/kernel-ctl/kernel-ctl.h>
  34 #include <common/sessiond-comm/sessiond-comm.h>
  35 #include <common/sessiond-comm/relayd.h>
  36 #include <common/compat/fcntl.h>
  37 #include <common/pipe.h>
  38 #include <common/relayd/relayd.h>
  39 #include <common/utils.h>
  40
  41 #include "kernel-consumer.h"
  42
  43 extern struct lttng_consumer_global_data consumer_data;
  44 extern int consumer_poll_timeout;
  45 extern volatile int consumer_quit;
  46
  47 /*
  48  * Take a snapshot for a specific fd
  49  *
  50  * Returns 0 on success, < 0 on error
  51  */
  52 int lttng_kconsumer_take_snapshot(struct lttng_consumer_stream *stream)
  53 {
  54         int ret = 0;
  55         int infd = stream->wait_fd;
  56
  57         ret = kernctl_snapshot(infd);
  58         if (ret != 0) {
  59                 errno = -ret;
  60                 perror("Getting sub-buffer snapshot.");
  61         }
  62
  63         return ret;
  64 }
  65
  66 /*
  67  * Get the produced position
  68  *
  69  * Returns 0 on success, < 0 on error
  70  */
  71 int lttng_kconsumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
  72                 unsigned long *pos)
  73 {
  74         int ret;
  75         int infd = stream->wait_fd;
  76
  77         ret = kernctl_snapshot_get_produced(infd, pos);
  78         if (ret != 0) {
  79                 errno = -ret;
  80                 perror("kernctl_snapshot_get_produced");
  81         }
  82
  83         return ret;
  84 }
  85
  86 /*
  87  * Receive command from session daemon and process it.
  88  *
  89  * Return 1 on success else a negative value or 0.
  90  */
  91 int lttng_kconsumer_recv_cmd(struct lttng_consumer_local_data *ctx,
  92                 int sock, struct pollfd *consumer_sockpoll)
  93 {
  94         ssize_t ret;
  95         enum lttng_error_code ret_code = LTTNG_OK;
  96         struct lttcomm_consumer_msg msg;
  97
  98         ret = lttcomm_recv_unix_sock(sock, &msg, sizeof(msg));
  99         if (ret != sizeof(msg)) {
 100                 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_CMD);
 101                 if (ret > 0) {
 102                         ret = -1;
 103                 }
 104                 return ret;
 105         }
 106         if (msg.cmd_type == LTTNG_CONSUMER_STOP) {
 107                 /*
 108                  * Notify the session daemon that the command is completed.
 109                  *
 110                  * On transport layer error, the function call will print an error
 111                  * message so handling the returned code is a bit useless since we
 112                  * return an error code anyway.
 113                  */
 114                 (void) consumer_send_status_msg(sock, ret_code);
 115                 return -ENOENT;
 116         }
 117
 118         /* relayd needs RCU read-side protection */
 119         rcu_read_lock();
 120
 121         switch (msg.cmd_type) {
 122         case LTTNG_CONSUMER_ADD_RELAYD_SOCKET:
 123         {
 124                 /* Session daemon status message are handled in the following call. */
 125                 ret = consumer_add_relayd_socket(msg.u.relayd_sock.net_index,
 126                                 msg.u.relayd_sock.type, ctx, sock, consumer_sockpoll,
 127                                 &msg.u.relayd_sock.sock, msg.u.relayd_sock.session_id);
 128                 goto end_nosignal;
 129         }
 130         case LTTNG_CONSUMER_ADD_CHANNEL:
 131         {
 132                 struct lttng_consumer_channel *new_channel;
 133                 int ret_recv;
 134
 135                 /* First send a status message before receiving the fds. */
 136                 ret = consumer_send_status_msg(sock, ret_code);
 137                 if (ret < 0) {
 138                         /* Somehow, the session daemon is not responding anymore. */
 139                         goto error_fatal;
 140                 }
 141                 DBG("consumer_add_channel %" PRIu64, msg.u.channel.channel_key);
 142                 new_channel = consumer_allocate_channel(msg.u.channel.channel_key,
 143                                 msg.u.channel.session_id, msg.u.channel.pathname,
 144                                 msg.u.channel.name, msg.u.channel.uid, msg.u.channel.gid,
 145                                 msg.u.channel.relayd_id, msg.u.channel.output,
 146                                 msg.u.channel.tracefile_size,
 147                                 msg.u.channel.tracefile_count);
 148                 if (new_channel == NULL) {
 149                         lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR);
 150                         goto end_nosignal;
 151                 }
 152                 new_channel->nb_init_stream_left = msg.u.channel.nb_init_streams;
 153
 154                 /* Translate and save channel type. */
 155                 switch (msg.u.channel.type) {
 156                 case CONSUMER_CHANNEL_TYPE_DATA:
 157                 case CONSUMER_CHANNEL_TYPE_METADATA:
 158                         new_channel->type = msg.u.channel.type;
 159                         break;
 160                 default:
 161                         assert(0);
 162                         goto end_nosignal;
 163                 };
 164
 165                 if (ctx->on_recv_channel != NULL) {
 166                         ret_recv = ctx->on_recv_channel(new_channel);
 167                         if (ret_recv == 0) {
 168                                 ret = consumer_add_channel(new_channel, ctx);
 169                         } else if (ret_recv < 0) {
 170                                 goto end_nosignal;
 171                         }
 172                 } else {
 173                         ret = consumer_add_channel(new_channel, ctx);
 174                 }
 175
 176                 /* If we received an error in add_channel, we need to report it. */
 177                 if (ret < 0) {
 178                         ret = consumer_send_status_msg(sock, ret);
 179                         if (ret < 0) {
 180                                 goto error_fatal;
 181                         }
 182                         goto end_nosignal;
 183                 }
 184
 185                 goto end_nosignal;
 186         }
 187         case LTTNG_CONSUMER_ADD_STREAM:
 188         {
 189                 int fd;
 190                 struct lttng_pipe *stream_pipe;
 191                 struct consumer_relayd_sock_pair *relayd = NULL;
 192                 struct lttng_consumer_stream *new_stream;
 193                 struct lttng_consumer_channel *channel;
 194                 int alloc_ret = 0;
 195
 196                 /*
 197                  * Get stream's channel reference. Needed when adding the stream to the
 198                  * global hash table.
 199                  */
 200                 channel = consumer_find_channel(msg.u.stream.channel_key);
 201                 if (!channel) {
 202                         /*
 203                          * We could not find the channel. Can happen if cpu hotplug
 204                          * happens while tearing down.
 205                          */
 206                         ERR("Unable to find channel key %" PRIu64, msg.u.stream.channel_key);
 207                         ret_code = LTTNG_ERR_KERN_CHAN_NOT_FOUND;
 208                 }
 209
 210                 /* First send a status message before receiving the fds. */
 211                 ret = consumer_send_status_msg(sock, ret_code);
 212                 if (ret < 0) {
 213                         /*
 214                          * Somehow, the session daemon is not responding
 215                          * anymore.
 216                          */
 217                         goto error_fatal;
 218                 }
 219                 if (ret_code != LTTNG_OK) {
 220                         /*
 221                          * Channel was not found.
 222                          */
 223                         goto end_nosignal;
 224                 }
 225
 226                 /* block */
 227                 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
 228                         rcu_read_unlock();
 229                         return -EINTR;
 230                 }
 231
 232                 /* Get stream file descriptor from socket */
 233                 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
 234                 if (ret != sizeof(fd)) {
 235                         lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
 236                         rcu_read_unlock();
 237                         return ret;
 238                 }
 239
 240                 /*
 241                  * Send status code to session daemon only if the recv works. If the
 242                  * above recv() failed, the session daemon is notified through the
 243                  * error socket and the teardown is eventually done.
 244                  */
 245                 ret = consumer_send_status_msg(sock, ret_code);
 246                 if (ret < 0) {
 247                         /* Somehow, the session daemon is not responding anymore. */
 248                         goto end_nosignal;
 249                 }
 250
 251                 new_stream = consumer_allocate_stream(channel->key,
 252                                 fd,
 253                                 LTTNG_CONSUMER_ACTIVE_STREAM,
 254                                 channel->name,
 255                                 channel->uid,
 256                                 channel->gid,
 257                                 channel->relayd_id,
 258                                 channel->session_id,
 259                                 msg.u.stream.cpu,
 260                                 &alloc_ret,
 261                                 channel->type);
 262                 if (new_stream == NULL) {
 263                         switch (alloc_ret) {
 264                         case -ENOMEM:
 265                         case -EINVAL:
 266                         default:
 267                                 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR);
 268                                 break;
 269                         }
 270                         goto end_nosignal;
 271                 }
 272                 new_stream->chan = channel;
 273                 new_stream->wait_fd = fd;
 274
 275                 /*
 276                  * We've just assigned the channel to the stream so increment the
 277                  * refcount right now.
 278                  */
 279                 uatomic_inc(&new_stream->chan->refcount);
 280
 281                 /*
 282                  * The buffer flush is done on the session daemon side for the kernel
 283                  * so no need for the stream "hangup_flush_done" variable to be
 284                  * tracked. This is important for a kernel stream since we don't rely
 285                  * on the flush state of the stream to read data. It's not the case for
 286                  * user space tracing.
 287                  */
 288                 new_stream->hangup_flush_done = 0;
 289
 290                 /* The stream is not metadata. Get relayd reference if exists. */
 291                 relayd = consumer_find_relayd(new_stream->net_seq_idx);
 292                 if (relayd != NULL) {
 293                         /* Add stream on the relayd */
 294                         pthread_mutex_lock(&relayd->ctrl_sock_mutex);
 295                         ret = relayd_add_stream(&relayd->control_sock,
 296                                         new_stream->name, new_stream->chan->pathname,
 297                                         &new_stream->relayd_stream_id,
 298                                         new_stream->chan->tracefile_size,
 299                                         new_stream->chan->tracefile_count);
 300                         pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
 301                         if (ret < 0) {
 302                                 consumer_del_stream(new_stream, NULL);
 303                                 goto end_nosignal;
 304                         }
 305                 } else if (new_stream->net_seq_idx != (uint64_t) -1ULL) {
 306                         ERR("Network sequence index %" PRIu64 " unknown. Not adding stream.",
 307                                         new_stream->net_seq_idx);
 308                         consumer_del_stream(new_stream, NULL);
 309                         goto end_nosignal;
 310                 }
 311
 312                 if (ctx->on_recv_stream) {
 313                         ret = ctx->on_recv_stream(new_stream);
 314                         if (ret < 0) {
 315                                 consumer_del_stream(new_stream, NULL);
 316                                 goto end_nosignal;
 317                         }
 318                 }
 319
 320                 /* Get the right pipe where the stream will be sent. */
 321                 if (new_stream->metadata_flag) {
 322                         stream_pipe = ctx->consumer_metadata_pipe;
 323                 } else {
 324                         stream_pipe = ctx->consumer_data_pipe;
 325                 }
 326
 327                 ret = lttng_pipe_write(stream_pipe, &new_stream, sizeof(new_stream));
 328                 if (ret < 0) {
 329                         ERR("Consumer write %s stream to pipe %d",
 330                                         new_stream->metadata_flag ? "metadata" : "data",
 331                                         lttng_pipe_get_writefd(stream_pipe));
 332                         consumer_del_stream(new_stream, NULL);
 333                         goto end_nosignal;
 334                 }
 335
 336                 DBG("Kernel consumer ADD_STREAM %s (fd: %d) with relayd id %" PRIu64,
 337                                 new_stream->name, fd, new_stream->relayd_stream_id);
 338                 break;
 339         }
 340         case LTTNG_CONSUMER_UPDATE_STREAM:
 341         {
 342                 rcu_read_unlock();
 343                 return -ENOSYS;
 344         }
 345         case LTTNG_CONSUMER_DESTROY_RELAYD:
 346         {
 347                 uint64_t index = msg.u.destroy_relayd.net_seq_idx;
 348                 struct consumer_relayd_sock_pair *relayd;
 349
 350                 DBG("Kernel consumer destroying relayd %" PRIu64, index);
 351
 352                 /* Get relayd reference if exists. */
 353                 relayd = consumer_find_relayd(index);
 354                 if (relayd == NULL) {
 355                         DBG("Unable to find relayd %" PRIu64, index);
 356                         ret_code = LTTNG_ERR_NO_CONSUMER;
 357                 }
 358
 359                 /*
 360                  * Each relayd socket pair has a refcount of stream attached to it
 361                  * which tells if the relayd is still active or not depending on the
 362                  * refcount value.
 363                  *
 364                  * This will set the destroy flag of the relayd object and destroy it
 365                  * if the refcount reaches zero when called.
 366                  *
 367                  * The destroy can happen either here or when a stream fd hangs up.
 368                  */
 369                 if (relayd) {
 370                         consumer_flag_relayd_for_destroy(relayd);
 371                 }
 372
 373                 ret = consumer_send_status_msg(sock, ret_code);
 374                 if (ret < 0) {
 375                         /* Somehow, the session daemon is not responding anymore. */
 376                         goto error_fatal;
 377                 }
 378
 379                 goto end_nosignal;
 380         }
 381         case LTTNG_CONSUMER_DATA_PENDING:
 382         {
 383                 int32_t ret;
 384                 uint64_t id = msg.u.data_pending.session_id;
 385
 386                 DBG("Kernel consumer data pending command for id %" PRIu64, id);
 387
 388                 ret = consumer_data_pending(id);
 389
 390                 /* Send back returned value to session daemon */
 391                 ret = lttcomm_send_unix_sock(sock, &ret, sizeof(ret));
 392                 if (ret < 0) {
 393                         PERROR("send data pending ret code");
 394                         goto error_fatal;
 395                 }
 396
 397                 /*
 398                  * No need to send back a status message since the data pending
 399                  * returned value is the response.
 400                  */
 401                 break;
 402         }
 403         default:
 404                 goto end_nosignal;
 405         }
 406
 407 end_nosignal:
 408         rcu_read_unlock();
 409
 410         /*
 411          * Return 1 to indicate success since the 0 value can be a socket
 412          * shutdown during the recv() or send() call.
 413          */
 414         return 1;
 415
 416 error_fatal:
 417         rcu_read_unlock();
 418         /* This will issue a consumer stop. */
 419         return -1;
 420 }
 421
 422 /*
 423  * Consume data on a file descriptor and write it on a trace file.
 424  */
 425 ssize_t lttng_kconsumer_read_subbuffer(struct lttng_consumer_stream *stream,
 426                 struct lttng_consumer_local_data *ctx)
 427 {
 428         unsigned long len, subbuf_size, padding;
 429         int err;
 430         ssize_t ret = 0;
 431         int infd = stream->wait_fd;
 432
 433         DBG("In read_subbuffer (infd : %d)", infd);
 434         /* Get the next subbuffer */
 435         err = kernctl_get_next_subbuf(infd);
 436         if (err != 0) {
 437                 ret = err;
 438                 /*
 439                  * This is a debug message even for single-threaded consumer,
 440                  * because poll() have more relaxed criterions than get subbuf,
 441                  * so get_subbuf may fail for short race windows where poll()
 442                  * would issue wakeups.
 443                  */
 444                 DBG("Reserving sub buffer failed (everything is normal, "
 445                                 "it is due to concurrency)");
 446                 goto end;
 447         }
 448
 449         /* Get the full subbuffer size including padding */
 450         err = kernctl_get_padded_subbuf_size(infd, &len);
 451         if (err != 0) {
 452                 errno = -err;
 453                 perror("Getting sub-buffer len failed.");
 454                 ret = err;
 455                 goto end;
 456         }
 457
 458         switch (stream->chan->output) {
 459         case LTTNG_EVENT_SPLICE:
 460
 461                 /*
 462                  * XXX: The lttng-modules splice "actor" does not handle copying
 463                  * partial pages hence only using the subbuffer size without the
 464                  * padding makes the splice fail.
 465                  */
 466                 subbuf_size = len;
 467                 padding = 0;
 468
 469                 /* splice the subbuffer to the tracefile */
 470                 ret = lttng_consumer_on_read_subbuffer_splice(ctx, stream, subbuf_size,
 471                                 padding);
 472                 /*
 473                  * XXX: Splice does not support network streaming so the return value
 474                  * is simply checked against subbuf_size and not like the mmap() op.
 475                  */
 476                 if (ret != subbuf_size) {
 477                         /*
 478                          * display the error but continue processing to try
 479                          * to release the subbuffer
 480                          */
 481                         ERR("Error splicing to tracefile (ret: %zd != len: %lu)",
 482                                         ret, subbuf_size);
 483                 }
 484                 break;
 485         case LTTNG_EVENT_MMAP:
 486                 /* Get subbuffer size without padding */
 487                 err = kernctl_get_subbuf_size(infd, &subbuf_size);
 488                 if (err != 0) {
 489                         errno = -err;
 490                         perror("Getting sub-buffer len failed.");
 491                         ret = err;
 492                         goto end;
 493                 }
 494
 495                 /* Make sure the tracer is not gone mad on us! */
 496                 assert(len >= subbuf_size);
 497
 498                 padding = len - subbuf_size;
 499
 500                 /* write the subbuffer to the tracefile */
 501                 ret = lttng_consumer_on_read_subbuffer_mmap(ctx, stream, subbuf_size,
 502                                 padding);
 503                 /*
 504                  * The mmap operation should write subbuf_size amount of data when
 505                  * network streaming or the full padding (len) size when we are _not_
 506                  * streaming.
 507                  */
 508                 if ((ret != subbuf_size && stream->net_seq_idx != (uint64_t) -1ULL) ||
 509                                 (ret != len && stream->net_seq_idx == (uint64_t) -1ULL)) {
 510                         /*
 511                          * Display the error but continue processing to try to release the
 512                          * subbuffer
 513                          */
 514                         ERR("Error writing to tracefile "
 515                                         "(ret: %zd != len: %lu != subbuf_size: %lu)",
 516                                         ret, len, subbuf_size);
 517                 }
 518                 break;
 519         default:
 520                 ERR("Unknown output method");
 521                 ret = -1;
 522         }
 523
 524         err = kernctl_put_next_subbuf(infd);
 525         if (err != 0) {
 526                 errno = -err;
 527                 if (errno == EFAULT) {
 528                         perror("Error in unreserving sub buffer\n");
 529                 } else if (errno == EIO) {
 530                         /* Should never happen with newer LTTng versions */
 531                         perror("Reader has been pushed by the writer, last sub-buffer corrupted.");
 532                 }
 533
 534                 ret = -err;
 535                 goto end;
 536         }
 537
 538 end:
 539         return ret;
 540 }
 541
 542 int lttng_kconsumer_on_recv_stream(struct lttng_consumer_stream *stream)
 543 {
 544         int ret;
 545
 546         assert(stream);
 547
 548         /* Don't create anything if this is set for streaming. */
 549         if (stream->net_seq_idx == (uint64_t) -1ULL) {
 550                 ret = utils_create_stream_file(stream->chan->pathname, stream->name,
 551                                 stream->chan->tracefile_size, stream->tracefile_count_current,
 552                                 stream->uid, stream->gid);
 553                 if (ret < 0) {
 554                         goto error;
 555                 }
 556                 stream->out_fd = ret;
 557                 stream->tracefile_size_current = 0;
 558         }
 559
 560         if (stream->output == LTTNG_EVENT_MMAP) {
 561                 /* get the len of the mmap region */
 562                 unsigned long mmap_len;
 563
 564                 ret = kernctl_get_mmap_len(stream->wait_fd, &mmap_len);
 565                 if (ret != 0) {
 566                         errno = -ret;
 567                         PERROR("kernctl_get_mmap_len");
 568                         goto error_close_fd;
 569                 }
 570                 stream->mmap_len = (size_t) mmap_len;
 571
 572                 stream->mmap_base = mmap(NULL, stream->mmap_len, PROT_READ,
 573                                 MAP_PRIVATE, stream->wait_fd, 0);
 574                 if (stream->mmap_base == MAP_FAILED) {
 575                         PERROR("Error mmaping");
 576                         ret = -1;
 577                         goto error_close_fd;
 578                 }
 579         }
 580
 581         /* we return 0 to let the library handle the FD internally */
 582         return 0;
 583
 584 error_close_fd:
 585         {
 586                 int err;
 587
 588                 err = close(stream->out_fd);
 589                 assert(!err);
 590         }
 591 error:
 592         return ret;
 593 }
 594
 595 /*
 596  * Check if data is still being extracted from the buffers for a specific
 597  * stream. Consumer data lock MUST be acquired before calling this function
 598  * and the stream lock.
 599  *
 600  * Return 1 if the traced data are still getting read else 0 meaning that the
 601  * data is available for trace viewer reading.
 602  */
 603 int lttng_kconsumer_data_pending(struct lttng_consumer_stream *stream)
 604 {
 605         int ret;
 606
 607         assert(stream);
 608
 609         ret = kernctl_get_next_subbuf(stream->wait_fd);
 610         if (ret == 0) {
 611                 /* There is still data so let's put back this subbuffer. */
 612                 ret = kernctl_put_subbuf(stream->wait_fd);
 613                 assert(ret == 0);
 614                 ret = 1;   /* Data is pending */
 615                 goto end;
 616         }
 617
 618         /* Data is NOT pending and ready to be read. */
 619         ret = 0;
 620
 621 end:
 622         return ret;
 623 }