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