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