liblttngkconsumerd/lttngkconsumerd.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
   6  * modify it under the terms of the GNU General Public License
   7  * as published by the Free Software Foundation; only version 2
   8  * of the License.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  18  */
  19
  20 #define _GNU_SOURCE
  21 #include <assert.h>
  22 #include <fcntl.h>
  23 #include <poll.h>
  24 #include <pthread.h>
  25 #include <stdlib.h>
  26 #include <string.h>
  27 #include <sys/mman.h>
  28 #include <sys/socket.h>
  29 #include <sys/types.h>
  30 #include <unistd.h>
  31
  32 #include <lttng-kernel-ctl.h>
  33 #include <lttng-sessiond-comm.h>
  34 #include <lttng/lttng-kconsumerd.h>
  35 #include <lttngerr.h>
  36
  37 static struct lttng_kconsumerd_global_data {
  38         /*
  39          * kconsumerd_data.lock protects kconsumerd_data.fd_list,
  40          * kconsumerd_data.fds_count, and kconsumerd_data.need_update. It ensures
  41          * the count matches the number of items in the fd_list. It ensures the
  42          * list updates *always* trigger an fd_array update (therefore need to make
  43          * list update vs kconsumerd_data.need_update flag update atomic, and also
  44          * flag read, fd array and flag clear atomic).
  45          */
  46         pthread_mutex_t lock;
  47         /*
  48          * Number of element for the list below. Protected by kconsumerd_data.lock.
  49          */
  50         unsigned int fds_count;
  51         /*
  52          * List of FDs. Protected by kconsumerd_data.lock.
  53          */
  54         struct lttng_kconsumerd_fd_list fd_list;
  55         /*
  56          * Flag specifying if the local array of FDs needs update in the poll
  57          * function. Protected by kconsumerd_data.lock.
  58          */
  59         unsigned int need_update;
  60 } kconsumerd_data = {
  61         .fd_list.head = CDS_LIST_HEAD_INIT(kconsumerd_data.fd_list.head),
  62         .fds_count = 0,
  63         .need_update = 1,
  64 };
  65
  66 /* timeout parameter, to control the polling thread grace period. */
  67 static int kconsumerd_poll_timeout = -1;
  68
  69 /*
  70  * Flag to inform the polling thread to quit when all fd hung up. Updated by
  71  * the kconsumerd_thread_receive_fds when it notices that all fds has hung up.
  72  * Also updated by the signal handler (kconsumerd_should_exit()). Read by the
  73  * polling threads.
  74  */
  75 static volatile int kconsumerd_quit = 0;
  76
  77 /*
  78  * Find a session fd in the global list. The kconsumerd_data.lock must be
  79  * locked during this call.
  80  *
  81  * Return 1 if found else 0.
  82  */
  83 static int kconsumerd_find_session_fd(int fd)
  84 {
  85         struct lttng_kconsumerd_fd *iter;
  86
  87         cds_list_for_each_entry(iter, &kconsumerd_data.fd_list.head, list) {
  88                 if (iter->sessiond_fd == fd) {
  89                         DBG("Duplicate session fd %d", fd);
  90                         return 1;
  91                 }
  92         }
  93
  94         return 0;
  95 }
  96
  97 /*
  98  * Remove a fd from the global list protected by a mutex.
  99  */
 100 static void kconsumerd_del_fd(struct lttng_kconsumerd_fd *lcf)
 101 {
 102         int ret;
 103         pthread_mutex_lock(&kconsumerd_data.lock);
 104         cds_list_del(&lcf->list);
 105         if (kconsumerd_data.fds_count > 0) {
 106                 kconsumerd_data.fds_count--;
 107                 if (lcf != NULL) {
 108                         if (lcf->mmap_base != NULL) {
 109                                 ret = munmap(lcf->mmap_base, lcf->mmap_len);
 110                                 if (ret != 0) {
 111                                         perror("munmap");
 112                                 }
 113                         }
 114                         if (lcf->out_fd != 0) {
 115                                 close(lcf->out_fd);
 116                         }
 117                         close(lcf->consumerd_fd);
 118                         free(lcf);
 119                         lcf = NULL;
 120                 }
 121         }
 122         kconsumerd_data.need_update = 1;
 123         pthread_mutex_unlock(&kconsumerd_data.lock);
 124 }
 125
 126 /*
 127  * Create a struct lttcomm_kconsumerd_msg from the
 128  * information received on the receiving socket
 129  */
 130 struct lttng_kconsumerd_fd *kconsumerd_allocate_fd(
 131                 struct lttcomm_kconsumerd_msg *buf,
 132                 int consumerd_fd)
 133 {
 134         struct lttng_kconsumerd_fd *tmp_fd;
 135
 136         tmp_fd = malloc(sizeof(struct lttng_kconsumerd_fd));
 137         if (tmp_fd == NULL) {
 138                 perror("malloc struct lttng_kconsumerd_fd");
 139                 goto end;
 140         }
 141
 142         tmp_fd->sessiond_fd = buf->fd;
 143         tmp_fd->consumerd_fd = consumerd_fd;
 144         tmp_fd->state = buf->state;
 145         tmp_fd->max_sb_size = buf->max_sb_size;
 146         tmp_fd->out_fd = 0;
 147         tmp_fd->out_fd_offset = 0;
 148         tmp_fd->mmap_len = 0;
 149         tmp_fd->mmap_base = NULL;
 150         tmp_fd->output = buf->output;
 151         strncpy(tmp_fd->path_name, buf->path_name, PATH_MAX);
 152         tmp_fd->path_name[PATH_MAX - 1] = '\0';
 153         DBG("Allocated %s (sessiond_fd %d, consumerd_fd %d, out_fd %d)",
 154                         tmp_fd->path_name, tmp_fd->sessiond_fd,
 155                         tmp_fd->consumerd_fd, tmp_fd->out_fd);
 156
 157 end:
 158         return tmp_fd;
 159 }
 160
 161 /*
 162  * Add a fd to the global list protected by a mutex.
 163  */
 164 static int kconsumerd_add_fd(struct lttng_kconsumerd_fd *tmp_fd)
 165 {
 166         int ret;
 167
 168         pthread_mutex_lock(&kconsumerd_data.lock);
 169         /* Check if already exist */
 170         ret = kconsumerd_find_session_fd(tmp_fd->sessiond_fd);
 171         if (ret == 1) {
 172                 goto end;
 173         }
 174         cds_list_add(&tmp_fd->list, &kconsumerd_data.fd_list.head);
 175         kconsumerd_data.fds_count++;
 176         kconsumerd_data.need_update = 1;
 177
 178 end:
 179         pthread_mutex_unlock(&kconsumerd_data.lock);
 180         return ret;
 181 }
 182
 183 /*
 184  * Update a fd according to what we just received.
 185  */
 186 static void kconsumerd_change_fd_state(int sessiond_fd,
 187                 enum lttng_kconsumerd_fd_state state)
 188 {
 189         struct lttng_kconsumerd_fd *iter;
 190
 191         pthread_mutex_lock(&kconsumerd_data.lock);
 192         cds_list_for_each_entry(iter, &kconsumerd_data.fd_list.head, list) {
 193                 if (iter->sessiond_fd == sessiond_fd) {
 194                         iter->state = state;
 195                         break;
 196                 }
 197         }
 198         kconsumerd_data.need_update = 1;
 199         pthread_mutex_unlock(&kconsumerd_data.lock);
 200 }
 201
 202 /*
 203  * Allocate the pollfd structure and the local view of the out fds to avoid
 204  * doing a lookup in the linked list and concurrency issues when writing is
 205  * needed. Called with kconsumerd_data.lock held.
 206  *
 207  * Returns the number of fds in the structures.
 208  */
 209 static int kconsumerd_update_poll_array(
 210                 struct lttng_kconsumerd_local_data *ctx, struct pollfd **pollfd,
 211                 struct lttng_kconsumerd_fd **local_kconsumerd_fd)
 212 {
 213         struct lttng_kconsumerd_fd *iter;
 214         int i = 0;
 215
 216         DBG("Updating poll fd array");
 217         cds_list_for_each_entry(iter, &kconsumerd_data.fd_list.head, list) {
 218                 if (iter->state == ACTIVE_FD) {
 219                         DBG("Active FD %d", iter->consumerd_fd);
 220                         (*pollfd)[i].fd = iter->consumerd_fd;
 221                         (*pollfd)[i].events = POLLIN | POLLPRI;
 222                         local_kconsumerd_fd[i] = iter;
 223                         i++;
 224                 }
 225         }
 226
 227         /*
 228          * Insert the kconsumerd_poll_pipe at the end of the array and don't
 229          * increment i so nb_fd is the number of real FD.
 230          */
 231         (*pollfd)[i].fd = ctx->kconsumerd_poll_pipe[0];
 232         (*pollfd)[i].events = POLLIN;
 233         return i;
 234 }
 235
 236 /*
 237  * Receives an array of file descriptors and the associated structures
 238  * describing each fd (path name).
 239  *
 240  * Returns the size of received data
 241  */
 242 static int kconsumerd_consumerd_recv_fd(
 243                 struct lttng_kconsumerd_local_data *ctx, int sfd,
 244                 struct pollfd *kconsumerd_sockpoll, int size,
 245                 enum lttng_kconsumerd_command cmd_type)
 246 {
 247         struct iovec iov[1];
 248         int ret = 0, i, tmp2;
 249         struct cmsghdr *cmsg;
 250         int nb_fd;
 251         char recv_fd[CMSG_SPACE(sizeof(int))];
 252         struct lttcomm_kconsumerd_msg lkm;
 253         struct lttng_kconsumerd_fd *new_fd;
 254
 255         /* the number of fds we are about to receive */
 256         nb_fd = size / sizeof(struct lttcomm_kconsumerd_msg);
 257
 258         /*
 259          * nb_fd is the number of fds we receive. One fd per recvmsg.
 260          */
 261         for (i = 0; i < nb_fd; i++) {
 262                 struct msghdr msg = { 0 };
 263
 264                 /* Prepare to receive the structures */
 265                 iov[0].iov_base = &lkm;
 266                 iov[0].iov_len = sizeof(lkm);
 267                 msg.msg_iov = iov;
 268                 msg.msg_iovlen = 1;
 269
 270                 msg.msg_control = recv_fd;
 271                 msg.msg_controllen = sizeof(recv_fd);
 272
 273                 DBG("Waiting to receive fd");
 274                 if (lttng_kconsumerd_poll_socket(kconsumerd_sockpoll) < 0) {
 275                         goto end;
 276                 }
 277
 278                 if ((ret = recvmsg(sfd, &msg, 0)) < 0) {
 279                         perror("recvmsg");
 280                         continue;
 281                 }
 282
 283                 if (ret != (size / nb_fd)) {
 284                         ERR("Received only %d, expected %d", ret, size);
 285                         lttng_kconsumerd_send_error(ctx, KCONSUMERD_ERROR_RECV_FD);
 286                         goto end;
 287                 }
 288
 289                 cmsg = CMSG_FIRSTHDR(&msg);
 290                 if (!cmsg) {
 291                         ERR("Invalid control message header");
 292                         ret = -1;
 293                         lttng_kconsumerd_send_error(ctx, KCONSUMERD_ERROR_RECV_FD);
 294                         goto end;
 295                 }
 296
 297                 /* if we received fds */
 298                 if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS) {
 299                         switch (cmd_type) {
 300                                 case ADD_STREAM:
 301                                         DBG("kconsumerd_add_fd %s (%d)", lkm.path_name,
 302                                                         ((int *) CMSG_DATA(cmsg))[0]);
 303
 304                                         new_fd = kconsumerd_allocate_fd(&lkm, ((int *) CMSG_DATA(cmsg))[0]);
 305                                         if (new_fd == NULL) {
 306                                                 lttng_kconsumerd_send_error(ctx, KCONSUMERD_OUTFD_ERROR);
 307                                                 goto end;
 308                                         }
 309
 310                                         if (ctx->on_recv_fd != NULL) {
 311                                                 ret = ctx->on_recv_fd(new_fd);
 312                                                 if (ret == 0) {
 313                                                         kconsumerd_add_fd(new_fd);
 314                                                 } else if (ret < 0) {
 315                                                         goto end;
 316                                                 }
 317                                         } else {
 318                                                 kconsumerd_add_fd(new_fd);
 319                                         }
 320                                         break;
 321                                 case UPDATE_STREAM:
 322                                         if (ctx->on_update_fd != NULL) {
 323                                                 ret = ctx->on_update_fd(lkm.fd, lkm.state);
 324                                                 if (ret == 0) {
 325                                                         kconsumerd_change_fd_state(lkm.fd, lkm.state);
 326                                                 } else if (ret < 0) {
 327                                                         goto end;
 328                                                 }
 329                                         } else {
 330                                                 kconsumerd_change_fd_state(lkm.fd, lkm.state);
 331                                         }
 332                                         break;
 333                                 default:
 334                                         break;
 335                         }
 336                         /* signal the poll thread */
 337                         tmp2 = write(ctx->kconsumerd_poll_pipe[1], "4", 1);
 338                         if (tmp2 < 0) {
 339                                 perror("write kconsumerd poll");
 340                         }
 341                 } else {
 342                         ERR("Didn't received any fd");
 343                         lttng_kconsumerd_send_error(ctx, KCONSUMERD_ERROR_RECV_FD);
 344                         ret = -1;
 345                         goto end;
 346                 }
 347         }
 348
 349 end:
 350         return ret;
 351 }
 352
 353 /*
 354  * Set the error socket.
 355  */
 356 void lttng_kconsumerd_set_error_sock(
 357                 struct lttng_kconsumerd_local_data *ctx, int sock)
 358 {
 359         ctx->kconsumerd_error_socket = sock;
 360 }
 361
 362 /*
 363  * Set the command socket path.
 364  */
 365
 366 void lttng_kconsumerd_set_command_sock_path(
 367                 struct lttng_kconsumerd_local_data *ctx, char *sock)
 368 {
 369         ctx->kconsumerd_command_sock_path = sock;
 370 }
 371
 372 static void lttng_kconsumerd_sync_trace_file(
 373                 struct lttng_kconsumerd_fd *kconsumerd_fd, off_t orig_offset)
 374 {
 375         int outfd = kconsumerd_fd->out_fd;
 376         /*
 377          * This does a blocking write-and-wait on any page that belongs to the
 378          * subbuffer prior to the one we just wrote.
 379          * Don't care about error values, as these are just hints and ways to
 380          * limit the amount of page cache used.
 381          */
 382         if (orig_offset >= kconsumerd_fd->max_sb_size) {
 383                 sync_file_range(outfd, orig_offset - kconsumerd_fd->max_sb_size,
 384                                 kconsumerd_fd->max_sb_size,
 385                                 SYNC_FILE_RANGE_WAIT_BEFORE
 386                                 | SYNC_FILE_RANGE_WRITE
 387                                 | SYNC_FILE_RANGE_WAIT_AFTER);
 388                 /*
 389                  * Give hints to the kernel about how we access the file:
 390                  * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
 391                  * we write it.
 392                  *
 393                  * We need to call fadvise again after the file grows because the
 394                  * kernel does not seem to apply fadvise to non-existing parts of the
 395                  * file.
 396                  *
 397                  * Call fadvise _after_ having waited for the page writeback to
 398                  * complete because the dirty page writeback semantic is not well
 399                  * defined. So it can be expected to lead to lower throughput in
 400                  * streaming.
 401                  */
 402                 posix_fadvise(outfd, orig_offset - kconsumerd_fd->max_sb_size,
 403                                 kconsumerd_fd->max_sb_size, POSIX_FADV_DONTNEED);
 404         }
 405 }
 406
 407
 408 /*
 409  * Mmap the ring buffer, read it and write the data to the tracefile.
 410  *
 411  * Returns the number of bytes written
 412  */
 413 int lttng_kconsumerd_on_read_subbuffer_mmap(
 414                 struct lttng_kconsumerd_local_data *ctx,
 415                 struct lttng_kconsumerd_fd *kconsumerd_fd, unsigned long len)
 416 {
 417         unsigned long mmap_offset;
 418         long ret = 0;
 419         off_t orig_offset = kconsumerd_fd->out_fd_offset;
 420         int fd = kconsumerd_fd->consumerd_fd;
 421         int outfd = kconsumerd_fd->out_fd;
 422
 423         /* get the offset inside the fd to mmap */
 424         ret = kernctl_get_mmap_read_offset(fd, &mmap_offset);
 425         if (ret != 0) {
 426                 ret = errno;
 427                 perror("kernctl_get_mmap_read_offset");
 428                 goto end;
 429         }
 430
 431         while (len > 0) {
 432                 ret = write(outfd, kconsumerd_fd->mmap_base + mmap_offset, len);
 433                 if (ret >= len) {
 434                         len = 0;
 435                 } else if (ret < 0) {
 436                         ret = errno;
 437                         perror("Error in file write");
 438                         goto end;
 439                 }
 440                 /* This won't block, but will start writeout asynchronously */
 441                 sync_file_range(outfd, kconsumerd_fd->out_fd_offset, ret,
 442                                 SYNC_FILE_RANGE_WRITE);
 443                 kconsumerd_fd->out_fd_offset += ret;
 444         }
 445
 446         lttng_kconsumerd_sync_trace_file(kconsumerd_fd, orig_offset);
 447
 448         goto end;
 449
 450 end:
 451         return ret;
 452 }
 453
 454 /*
 455  * Splice the data from the ring buffer to the tracefile.
 456  *
 457  * Returns the number of bytes spliced.
 458  */
 459 int lttng_kconsumerd_on_read_subbuffer_splice(
 460                 struct lttng_kconsumerd_local_data *ctx,
 461                 struct lttng_kconsumerd_fd *kconsumerd_fd, unsigned long len)
 462 {
 463         long ret = 0;
 464         loff_t offset = 0;
 465         off_t orig_offset = kconsumerd_fd->out_fd_offset;
 466         int fd = kconsumerd_fd->consumerd_fd;
 467         int outfd = kconsumerd_fd->out_fd;
 468
 469         while (len > 0) {
 470                 DBG("splice chan to pipe offset %lu (fd : %d)",
 471                                 (unsigned long)offset, fd);
 472                 ret = splice(fd, &offset, ctx->kconsumerd_thread_pipe[1], NULL, len,
 473                                 SPLICE_F_MOVE | SPLICE_F_MORE);
 474                 DBG("splice chan to pipe ret %ld", ret);
 475                 if (ret < 0) {
 476                         ret = errno;
 477                         perror("Error in relay splice");
 478                         goto splice_error;
 479                 }
 480
 481                 ret = splice(ctx->kconsumerd_thread_pipe[0], NULL, outfd, NULL, ret,
 482                                 SPLICE_F_MOVE | SPLICE_F_MORE);
 483                 DBG("splice pipe to file %ld", ret);
 484                 if (ret < 0) {
 485                         ret = errno;
 486                         perror("Error in file splice");
 487                         goto splice_error;
 488                 }
 489                 len -= ret;
 490                 /* This won't block, but will start writeout asynchronously */
 491                 sync_file_range(outfd, kconsumerd_fd->out_fd_offset, ret,
 492                                 SYNC_FILE_RANGE_WRITE);
 493                 kconsumerd_fd->out_fd_offset += ret;
 494         }
 495         lttng_kconsumerd_sync_trace_file(kconsumerd_fd, orig_offset);
 496
 497         goto end;
 498
 499 splice_error:
 500         /* send the appropriate error description to sessiond */
 501         switch(ret) {
 502         case EBADF:
 503                 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_EBADF);
 504                 break;
 505         case EINVAL:
 506                 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_EINVAL);
 507                 break;
 508         case ENOMEM:
 509                 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_ENOMEM);
 510                 break;
 511         case ESPIPE:
 512                 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_ESPIPE);
 513                 break;
 514         }
 515
 516 end:
 517         return ret;
 518 }
 519
 520 /*
 521  * Take a snapshot for a specific fd
 522  *
 523  * Returns 0 on success, < 0 on error
 524  */
 525 int lttng_kconsumerd_take_snapshot(struct lttng_kconsumerd_local_data *ctx,
 526                 struct lttng_kconsumerd_fd *kconsumerd_fd)
 527 {
 528         int ret = 0;
 529         int infd = kconsumerd_fd->consumerd_fd;
 530
 531         ret = kernctl_snapshot(infd);
 532         if (ret != 0) {
 533                 ret = errno;
 534                 perror("Getting sub-buffer snapshot.");
 535         }
 536
 537         return ret;
 538 }
 539
 540 /*
 541  * Get the produced position
 542  *
 543  * Returns 0 on success, < 0 on error
 544  */
 545 int lttng_kconsumerd_get_produced_snapshot(
 546                 struct lttng_kconsumerd_local_data *ctx,
 547                 struct lttng_kconsumerd_fd *kconsumerd_fd,
 548                 unsigned long *pos)
 549 {
 550         int ret;
 551         int infd = kconsumerd_fd->consumerd_fd;
 552
 553         ret = kernctl_snapshot_get_produced(infd, pos);
 554         if (ret != 0) {
 555                 ret = errno;
 556                 perror("kernctl_snapshot_get_produced");
 557         }
 558
 559         return ret;
 560 }
 561
 562 /*
 563  * Poll on the should_quit pipe and the command socket return -1 on error and
 564  * should exit, 0 if data is available on the command socket
 565  */
 566 int lttng_kconsumerd_poll_socket(struct pollfd *kconsumerd_sockpoll)
 567 {
 568         int num_rdy;
 569
 570         num_rdy = poll(kconsumerd_sockpoll, 2, -1);
 571         if (num_rdy == -1) {
 572                 perror("Poll error");
 573                 goto exit;
 574         }
 575         if (kconsumerd_sockpoll[0].revents == POLLIN) {
 576                 DBG("kconsumerd_should_quit wake up");
 577                 goto exit;
 578         }
 579         return 0;
 580
 581 exit:
 582         return -1;
 583 }
 584
 585 /*
 586  * This thread polls the fds in the ltt_fd_list to consume the data and write
 587  * it to tracefile if necessary.
 588  */
 589 void *lttng_kconsumerd_thread_poll_fds(void *data)
 590 {
 591         int num_rdy, num_hup, high_prio, ret, i;
 592         struct pollfd *pollfd = NULL;
 593         /* local view of the fds */
 594         struct lttng_kconsumerd_fd **local_kconsumerd_fd = NULL;
 595         /* local view of kconsumerd_data.fds_count */
 596         int nb_fd = 0;
 597         char tmp;
 598         int tmp2;
 599         struct lttng_kconsumerd_local_data *ctx = data;
 600
 601
 602         local_kconsumerd_fd = malloc(sizeof(struct lttng_kconsumerd_fd));
 603
 604         while (1) {
 605                 high_prio = 0;
 606                 num_hup = 0;
 607
 608                 /*
 609                  * the ltt_fd_list has been updated, we need to update our
 610                  * local array as well
 611                  */
 612                 pthread_mutex_lock(&kconsumerd_data.lock);
 613                 if (kconsumerd_data.need_update) {
 614                         if (pollfd != NULL) {
 615                                 free(pollfd);
 616                                 pollfd = NULL;
 617                         }
 618                         if (local_kconsumerd_fd != NULL) {
 619                                 free(local_kconsumerd_fd);
 620                                 local_kconsumerd_fd = NULL;
 621                         }
 622
 623                         /* allocate for all fds + 1 for the kconsumerd_poll_pipe */
 624                         pollfd = malloc((kconsumerd_data.fds_count + 1) * sizeof(struct pollfd));
 625                         if (pollfd == NULL) {
 626                                 perror("pollfd malloc");
 627                                 pthread_mutex_unlock(&kconsumerd_data.lock);
 628                                 goto end;
 629                         }
 630
 631                         /* allocate for all fds + 1 for the kconsumerd_poll_pipe */
 632                         local_kconsumerd_fd = malloc((kconsumerd_data.fds_count + 1) *
 633                                         sizeof(struct lttng_kconsumerd_fd));
 634                         if (local_kconsumerd_fd == NULL) {
 635                                 perror("local_kconsumerd_fd malloc");
 636                                 pthread_mutex_unlock(&kconsumerd_data.lock);
 637                                 goto end;
 638                         }
 639                         ret = kconsumerd_update_poll_array(ctx, &pollfd, local_kconsumerd_fd);
 640                         if (ret < 0) {
 641                                 ERR("Error in allocating pollfd or local_outfds");
 642                                 lttng_kconsumerd_send_error(ctx, KCONSUMERD_POLL_ERROR);
 643                                 pthread_mutex_unlock(&kconsumerd_data.lock);
 644                                 goto end;
 645                         }
 646                         nb_fd = ret;
 647                         kconsumerd_data.need_update = 0;
 648                 }
 649                 pthread_mutex_unlock(&kconsumerd_data.lock);
 650
 651                 /* poll on the array of fds */
 652                 DBG("polling on %d fd", nb_fd + 1);
 653                 num_rdy = poll(pollfd, nb_fd + 1, kconsumerd_poll_timeout);
 654                 DBG("poll num_rdy : %d", num_rdy);
 655                 if (num_rdy == -1) {
 656                         perror("Poll error");
 657                         lttng_kconsumerd_send_error(ctx, KCONSUMERD_POLL_ERROR);
 658                         goto end;
 659                 } else if (num_rdy == 0) {
 660                         DBG("Polling thread timed out");
 661                         goto end;
 662                 }
 663
 664                 /* No FDs and kconsumerd_quit, kconsumerd_cleanup the thread */
 665                 if (nb_fd == 0 && kconsumerd_quit == 1) {
 666                         goto end;
 667                 }
 668
 669                 /*
 670                  * If the kconsumerd_poll_pipe triggered poll go
 671                  * directly to the beginning of the loop to update the
 672                  * array. We want to prioritize array update over
 673                  * low-priority reads.
 674                  */
 675                 if (pollfd[nb_fd].revents == POLLIN) {
 676                         DBG("kconsumerd_poll_pipe wake up");
 677                         tmp2 = read(ctx->kconsumerd_poll_pipe[0], &tmp, 1);
 678                         if (tmp2 < 0) {
 679                                 perror("read kconsumerd poll");
 680                         }
 681                         continue;
 682                 }
 683
 684                 /* Take care of high priority channels first. */
 685                 for (i = 0; i < nb_fd; i++) {
 686                         switch(pollfd[i].revents) {
 687                         case POLLERR:
 688                                 ERR("Error returned in polling fd %d.", pollfd[i].fd);
 689                                 kconsumerd_del_fd(local_kconsumerd_fd[i]);
 690                                 num_hup++;
 691                                 break;
 692                         case POLLHUP:
 693                                 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
 694                                 kconsumerd_del_fd(local_kconsumerd_fd[i]);
 695                                 num_hup++;
 696                                 break;
 697                         case POLLNVAL:
 698                                 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
 699                                 kconsumerd_del_fd(local_kconsumerd_fd[i]);
 700                                 num_hup++;
 701                                 break;
 702                         case POLLPRI:
 703                                 DBG("Urgent read on fd %d", pollfd[i].fd);
 704                                 high_prio = 1;
 705                                 ret = ctx->on_buffer_ready(local_kconsumerd_fd[i]);
 706                                 /* it's ok to have an unavailable sub-buffer */
 707                                 if (ret == EAGAIN) {
 708                                         ret = 0;
 709                                 }
 710                                 break;
 711                         }
 712                 }
 713
 714                 /* If every buffer FD has hung up, we end the read loop here */
 715                 if (nb_fd > 0 && num_hup == nb_fd) {
 716                         DBG("every buffer FD has hung up\n");
 717                         if (kconsumerd_quit == 1) {
 718                                 goto end;
 719                         }
 720                         continue;
 721                 }
 722
 723                 /* Take care of low priority channels. */
 724                 if (high_prio == 0) {
 725                         for (i = 0; i < nb_fd; i++) {
 726                                 if (pollfd[i].revents == POLLIN) {
 727                                         DBG("Normal read on fd %d", pollfd[i].fd);
 728                                         ret = ctx->on_buffer_ready(local_kconsumerd_fd[i]);
 729                                         /* it's ok to have an unavailable subbuffer */
 730                                         if (ret == EAGAIN) {
 731                                                 ret = 0;
 732                                         }
 733                                 }
 734                         }
 735                 }
 736         }
 737 end:
 738         DBG("polling thread exiting");
 739         if (pollfd != NULL) {
 740                 free(pollfd);
 741                 pollfd = NULL;
 742         }
 743         if (local_kconsumerd_fd != NULL) {
 744                 free(local_kconsumerd_fd);
 745                 local_kconsumerd_fd = NULL;
 746         }
 747         return NULL;
 748 }
 749
 750 /*
 751  * Initialise the necessary environnement :
 752  * - create a new context
 753  * - create the poll_pipe
 754  * - create the should_quit pipe (for signal handler)
 755  * - create the thread pipe (for splice)
 756  *
 757  * Takes a function pointer as argument, this function is called when data is
 758  * available on a buffer. This function is responsible to do the
 759  * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
 760  * buffer configuration and then kernctl_put_next_subbuf at the end.
 761  *
 762  * Returns a pointer to the new context or NULL on error.
 763  */
 764 struct lttng_kconsumerd_local_data *lttng_kconsumerd_create(
 765                 int (*buffer_ready)(struct lttng_kconsumerd_fd *kconsumerd_fd),
 766                 int (*recv_fd)(struct lttng_kconsumerd_fd *kconsumerd_fd),
 767                 int (*update_fd)(int sessiond_fd, uint32_t state))
 768 {
 769         int ret, i;
 770         struct lttng_kconsumerd_local_data *ctx;
 771
 772         ctx = malloc(sizeof(struct lttng_kconsumerd_local_data));
 773         if (ctx == NULL) {
 774                 perror("allocating context");
 775                 goto error;
 776         }
 777
 778         ctx->kconsumerd_error_socket = -1;
 779         /* assign the callbacks */
 780         ctx->on_buffer_ready = buffer_ready;
 781         ctx->on_recv_fd = recv_fd;
 782         ctx->on_update_fd = update_fd;
 783
 784         ret = pipe(ctx->kconsumerd_poll_pipe);
 785         if (ret < 0) {
 786                 perror("Error creating poll pipe");
 787                 goto error_poll_pipe;
 788         }
 789
 790         ret = pipe(ctx->kconsumerd_should_quit);
 791         if (ret < 0) {
 792                 perror("Error creating recv pipe");
 793                 goto error_quit_pipe;
 794         }
 795
 796         ret = pipe(ctx->kconsumerd_thread_pipe);
 797         if (ret < 0) {
 798                 perror("Error creating thread pipe");
 799                 goto error_thread_pipe;
 800         }
 801
 802         return ctx;
 803
 804
 805 error_thread_pipe:
 806         for (i = 0; i < 2; i++) {
 807                 int err;
 808
 809                 err = close(ctx->kconsumerd_should_quit[i]);
 810                 assert(!err);
 811         }
 812 error_quit_pipe:
 813         for (i = 0; i < 2; i++) {
 814                 int err;
 815
 816                 err = close(ctx->kconsumerd_poll_pipe[i]);
 817                 assert(!err);
 818         }
 819 error_poll_pipe:
 820         free(ctx);
 821 error:
 822         return NULL;
 823 }
 824
 825 /*
 826  * Close all fds associated with the instance and free the context.
 827  */
 828 void lttng_kconsumerd_destroy(struct lttng_kconsumerd_local_data *ctx)
 829 {
 830         close(ctx->kconsumerd_error_socket);
 831         close(ctx->kconsumerd_thread_pipe[0]);
 832         close(ctx->kconsumerd_thread_pipe[1]);
 833         close(ctx->kconsumerd_poll_pipe[0]);
 834         close(ctx->kconsumerd_poll_pipe[1]);
 835         close(ctx->kconsumerd_should_quit[0]);
 836         close(ctx->kconsumerd_should_quit[1]);
 837         unlink(ctx->kconsumerd_command_sock_path);
 838         free(ctx);
 839         ctx = NULL;
 840 }
 841
 842 /*
 843  * This thread listens on the consumerd socket and receives the file
 844  * descriptors from the session daemon.
 845  */
 846 void *lttng_kconsumerd_thread_receive_fds(void *data)
 847 {
 848         int sock, client_socket, ret;
 849         struct lttcomm_kconsumerd_header tmp;
 850         /*
 851          * structure to poll for incoming data on communication socket avoids
 852          * making blocking sockets.
 853          */
 854         struct pollfd kconsumerd_sockpoll[2];
 855         struct lttng_kconsumerd_local_data *ctx = data;
 856
 857
 858         DBG("Creating command socket %s", ctx->kconsumerd_command_sock_path);
 859         unlink(ctx->kconsumerd_command_sock_path);
 860         client_socket = lttcomm_create_unix_sock(ctx->kconsumerd_command_sock_path);
 861         if (client_socket < 0) {
 862                 ERR("Cannot create command socket");
 863                 goto end;
 864         }
 865
 866         ret = lttcomm_listen_unix_sock(client_socket);
 867         if (ret < 0) {
 868                 goto end;
 869         }
 870
 871         DBG("Sending ready command to ltt-sessiond");
 872         ret = lttng_kconsumerd_send_error(ctx, KCONSUMERD_COMMAND_SOCK_READY);
 873         /* return < 0 on error, but == 0 is not fatal */
 874         if (ret < 0) {
 875                 ERR("Error sending ready command to ltt-sessiond");
 876                 goto end;
 877         }
 878
 879         ret = fcntl(client_socket, F_SETFL, O_NONBLOCK);
 880         if (ret < 0) {
 881                 perror("fcntl O_NONBLOCK");
 882                 goto end;
 883         }
 884
 885         /* prepare the FDs to poll : to client socket and the should_quit pipe */
 886         kconsumerd_sockpoll[0].fd = ctx->kconsumerd_should_quit[0];
 887         kconsumerd_sockpoll[0].events = POLLIN | POLLPRI;
 888         kconsumerd_sockpoll[1].fd = client_socket;
 889         kconsumerd_sockpoll[1].events = POLLIN | POLLPRI;
 890
 891         if (lttng_kconsumerd_poll_socket(kconsumerd_sockpoll) < 0) {
 892                 goto end;
 893         }
 894         DBG("Connection on client_socket");
 895
 896         /* Blocking call, waiting for transmission */
 897         sock = lttcomm_accept_unix_sock(client_socket);
 898         if (sock <= 0) {
 899                 WARN("On accept");
 900                 goto end;
 901         }
 902         ret = fcntl(sock, F_SETFL, O_NONBLOCK);
 903         if (ret < 0) {
 904                 perror("fcntl O_NONBLOCK");
 905                 goto end;
 906         }
 907
 908         /* update the polling structure to poll on the established socket */
 909         kconsumerd_sockpoll[1].fd = sock;
 910         kconsumerd_sockpoll[1].events = POLLIN | POLLPRI;
 911
 912         while (1) {
 913                 if (lttng_kconsumerd_poll_socket(kconsumerd_sockpoll) < 0) {
 914                         goto end;
 915                 }
 916                 DBG("Incoming fds on sock");
 917
 918                 /* We first get the number of fd we are about to receive */
 919                 ret = lttcomm_recv_unix_sock(sock, &tmp,
 920                                 sizeof(struct lttcomm_kconsumerd_header));
 921                 if (ret <= 0) {
 922                         ERR("Communication interrupted on command socket");
 923                         goto end;
 924                 }
 925                 if (tmp.cmd_type == STOP) {
 926                         DBG("Received STOP command");
 927                         goto end;
 928                 }
 929                 if (kconsumerd_quit) {
 930                         DBG("kconsumerd_thread_receive_fds received quit from signal");
 931                         goto end;
 932                 }
 933
 934                 /* we received a command to add or update fds */
 935                 ret = kconsumerd_consumerd_recv_fd(ctx, sock, kconsumerd_sockpoll,
 936                                 tmp.payload_size, tmp.cmd_type);
 937                 if (ret < 0) {
 938                         ERR("Receiving the FD, exiting");
 939                         goto end;
 940                 }
 941                 DBG("received fds on sock");
 942         }
 943
 944 end:
 945         DBG("kconsumerd_thread_receive_fds exiting");
 946
 947         /*
 948          * when all fds have hung up, the polling thread
 949          * can exit cleanly
 950          */
 951         kconsumerd_quit = 1;
 952
 953         /*
 954          * 2s of grace period, if no polling events occur during
 955          * this period, the polling thread will exit even if there
 956          * are still open FDs (should not happen, but safety mechanism).
 957          */
 958         kconsumerd_poll_timeout = LTTNG_KCONSUMERD_POLL_GRACE_PERIOD;
 959
 960         /* wake up the polling thread */
 961         ret = write(ctx->kconsumerd_poll_pipe[1], "4", 1);
 962         if (ret < 0) {
 963                 perror("poll pipe write");
 964         }
 965         return NULL;
 966 }
 967
 968 /*
 969  * Close all the tracefiles and stream fds, should be called when all instances
 970  * are destroyed.
 971  */
 972 void lttng_kconsumerd_cleanup(void)
 973 {
 974         struct lttng_kconsumerd_fd *iter, *tmp;
 975
 976         /*
 977          * close all outfd. Called when there are no more threads
 978          * running (after joining on the threads), no need to protect
 979          * list iteration with mutex.
 980          */
 981         cds_list_for_each_entry_safe(iter, tmp,
 982                         &kconsumerd_data.fd_list.head, list) {
 983                 kconsumerd_del_fd(iter);
 984         }
 985 }
 986
 987 /*
 988  * Called from signal handler.
 989  */
 990 void lttng_kconsumerd_should_exit(struct lttng_kconsumerd_local_data *ctx)
 991 {
 992         int ret;
 993         kconsumerd_quit = 1;
 994         ret = write(ctx->kconsumerd_should_quit[1], "4", 1);
 995         if (ret < 0) {
 996                 perror("write kconsumerd quit");
 997         }
 998 }
 999
1000 /*
1001  * Send return code to the session daemon.
1002  * If the socket is not defined, we return 0, it is not a fatal error
1003  */
1004 int lttng_kconsumerd_send_error(
1005                 struct lttng_kconsumerd_local_data *ctx, int cmd)
1006 {
1007         if (ctx->kconsumerd_error_socket > 0) {
1008                 return lttcomm_send_unix_sock(ctx->kconsumerd_error_socket, &cmd,
1009                                 sizeof(enum lttcomm_sessiond_command));
1010         }
1011
1012         return 0;
1013 }