69ef9a0e8b65173e9d6544bea875e6a9b6e9753c
[lttng-tools.git] / 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 static void lttng_kconsumerd_sync_trace_file(
365 struct lttng_kconsumerd_fd *kconsumerd_fd, off_t orig_offset)
366 {
367 int outfd = kconsumerd_fd->out_fd;
368 /*
369 * This does a blocking write-and-wait on any page that belongs to the
370 * subbuffer prior to the one we just wrote.
371 * Don't care about error values, as these are just hints and ways to
372 * limit the amount of page cache used.
373 */
374 if (orig_offset >= kconsumerd_fd->max_sb_size) {
375 sync_file_range(outfd, orig_offset - kconsumerd_fd->max_sb_size,
376 kconsumerd_fd->max_sb_size,
377 SYNC_FILE_RANGE_WAIT_BEFORE
378 | SYNC_FILE_RANGE_WRITE
379 | SYNC_FILE_RANGE_WAIT_AFTER);
380 /*
381 * Give hints to the kernel about how we access the file:
382 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
383 * we write it.
384 *
385 * We need to call fadvise again after the file grows because the
386 * kernel does not seem to apply fadvise to non-existing parts of the
387 * file.
388 *
389 * Call fadvise _after_ having waited for the page writeback to
390 * complete because the dirty page writeback semantic is not well
391 * defined. So it can be expected to lead to lower throughput in
392 * streaming.
393 */
394 posix_fadvise(outfd, orig_offset - kconsumerd_fd->max_sb_size,
395 kconsumerd_fd->max_sb_size, POSIX_FADV_DONTNEED);
396 }
397 }
398
399
400 /*
401 * Mmap the ring buffer, read it and write the data to the tracefile.
402 *
403 * Returns the number of bytes written
404 */
405 int lttng_kconsumerd_on_read_subbuffer_mmap(
406 struct lttng_kconsumerd_local_data *ctx,
407 struct lttng_kconsumerd_fd *kconsumerd_fd, unsigned long len)
408 {
409 unsigned long mmap_offset;
410 long ret = 0;
411 off_t orig_offset = kconsumerd_fd->out_fd_offset;
412 int fd = kconsumerd_fd->consumerd_fd;
413 int outfd = kconsumerd_fd->out_fd;
414
415 /* get the offset inside the fd to mmap */
416 ret = kernctl_get_mmap_read_offset(fd, &mmap_offset);
417 if (ret != 0) {
418 ret = errno;
419 perror("kernctl_get_mmap_read_offset");
420 goto end;
421 }
422
423 while (len > 0) {
424 ret = write(outfd, kconsumerd_fd->mmap_base + mmap_offset, len);
425 if (ret >= len) {
426 len = 0;
427 } else if (ret < 0) {
428 ret = errno;
429 perror("Error in file write");
430 goto end;
431 }
432 /* This won't block, but will start writeout asynchronously */
433 sync_file_range(outfd, kconsumerd_fd->out_fd_offset, ret,
434 SYNC_FILE_RANGE_WRITE);
435 kconsumerd_fd->out_fd_offset += ret;
436 }
437
438 lttng_kconsumerd_sync_trace_file(kconsumerd_fd, orig_offset);
439
440 goto end;
441
442 end:
443 return ret;
444 }
445
446 /*
447 * Splice the data from the ring buffer to the tracefile.
448 *
449 * Returns the number of bytes spliced.
450 */
451 int lttng_kconsumerd_on_read_subbuffer_splice(
452 struct lttng_kconsumerd_local_data *ctx,
453 struct lttng_kconsumerd_fd *kconsumerd_fd, unsigned long len)
454 {
455 long ret = 0;
456 loff_t offset = 0;
457 off_t orig_offset = kconsumerd_fd->out_fd_offset;
458 int fd = kconsumerd_fd->consumerd_fd;
459 int outfd = kconsumerd_fd->out_fd;
460
461 while (len > 0) {
462 DBG("splice chan to pipe offset %lu (fd : %d)",
463 (unsigned long)offset, fd);
464 ret = splice(fd, &offset, ctx->kconsumerd_thread_pipe[1], NULL, len,
465 SPLICE_F_MOVE | SPLICE_F_MORE);
466 DBG("splice chan to pipe ret %ld", ret);
467 if (ret < 0) {
468 ret = errno;
469 perror("Error in relay splice");
470 goto splice_error;
471 }
472
473 ret = splice(ctx->kconsumerd_thread_pipe[0], NULL, outfd, NULL, ret,
474 SPLICE_F_MOVE | SPLICE_F_MORE);
475 DBG("splice pipe to file %ld", ret);
476 if (ret < 0) {
477 ret = errno;
478 perror("Error in file splice");
479 goto splice_error;
480 }
481 len -= ret;
482 /* This won't block, but will start writeout asynchronously */
483 sync_file_range(outfd, kconsumerd_fd->out_fd_offset, ret,
484 SYNC_FILE_RANGE_WRITE);
485 kconsumerd_fd->out_fd_offset += ret;
486 }
487 lttng_kconsumerd_sync_trace_file(kconsumerd_fd, orig_offset);
488
489 goto end;
490
491 splice_error:
492 /* send the appropriate error description to sessiond */
493 switch(ret) {
494 case EBADF:
495 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_EBADF);
496 break;
497 case EINVAL:
498 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_EINVAL);
499 break;
500 case ENOMEM:
501 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_ENOMEM);
502 break;
503 case ESPIPE:
504 lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_ESPIPE);
505 break;
506 }
507
508 end:
509 return ret;
510 }
511
512 /*
513 * Take a snapshot for a specific fd
514 *
515 * Returns 0 on success, < 0 on error
516 */
517 int lttng_kconsumerd_take_snapshot(struct lttng_kconsumerd_local_data *ctx,
518 struct lttng_kconsumerd_fd *kconsumerd_fd)
519 {
520 int ret = 0;
521 int infd = kconsumerd_fd->consumerd_fd;
522
523 ret = kernctl_snapshot(infd);
524 if (ret != 0) {
525 ret = errno;
526 perror("Getting sub-buffer snapshot.");
527 }
528
529 return ret;
530 }
531
532 /*
533 * Get the produced position
534 *
535 * Returns 0 on success, < 0 on error
536 */
537 int lttng_kconsumerd_get_produced_snapshot(
538 struct lttng_kconsumerd_local_data *ctx,
539 struct lttng_kconsumerd_fd *kconsumerd_fd,
540 unsigned long *pos)
541 {
542 int ret;
543 int infd = kconsumerd_fd->consumerd_fd;
544
545 ret = kernctl_snapshot_get_produced(infd, pos);
546 if (ret != 0) {
547 ret = errno;
548 perror("kernctl_snapshot_get_produced");
549 }
550
551 return ret;
552 }
553
554 /*
555 * Poll on the should_quit pipe and the command socket return -1 on error and
556 * should exit, 0 if data is available on the command socket
557 */
558 int lttng_kconsumerd_poll_socket(struct pollfd *kconsumerd_sockpoll)
559 {
560 int num_rdy;
561
562 num_rdy = poll(kconsumerd_sockpoll, 2, -1);
563 if (num_rdy == -1) {
564 perror("Poll error");
565 goto exit;
566 }
567 if (kconsumerd_sockpoll[0].revents == POLLIN) {
568 DBG("kconsumerd_should_quit wake up");
569 goto exit;
570 }
571 return 0;
572
573 exit:
574 return -1;
575 }
576
577 /*
578 * This thread polls the fds in the ltt_fd_list to consume the data and write
579 * it to tracefile if necessary.
580 */
581 void *lttng_kconsumerd_thread_poll_fds(void *data)
582 {
583 int num_rdy, num_hup, high_prio, ret, i;
584 struct pollfd *pollfd = NULL;
585 /* local view of the fds */
586 struct lttng_kconsumerd_fd **local_kconsumerd_fd = NULL;
587 /* local view of kconsumerd_data.fds_count */
588 int nb_fd = 0;
589 char tmp;
590 int tmp2;
591 struct lttng_kconsumerd_local_data *ctx = data;
592
593
594 local_kconsumerd_fd = malloc(sizeof(struct lttng_kconsumerd_fd));
595
596 while (1) {
597 high_prio = 0;
598 num_hup = 0;
599
600 /*
601 * the ltt_fd_list has been updated, we need to update our
602 * local array as well
603 */
604 pthread_mutex_lock(&kconsumerd_data.lock);
605 if (kconsumerd_data.need_update) {
606 if (pollfd != NULL) {
607 free(pollfd);
608 pollfd = NULL;
609 }
610 if (local_kconsumerd_fd != NULL) {
611 free(local_kconsumerd_fd);
612 local_kconsumerd_fd = NULL;
613 }
614
615 /* allocate for all fds + 1 for the kconsumerd_poll_pipe */
616 pollfd = malloc((kconsumerd_data.fds_count + 1) * sizeof(struct pollfd));
617 if (pollfd == NULL) {
618 perror("pollfd malloc");
619 pthread_mutex_unlock(&kconsumerd_data.lock);
620 goto end;
621 }
622
623 /* allocate for all fds + 1 for the kconsumerd_poll_pipe */
624 local_kconsumerd_fd = malloc((kconsumerd_data.fds_count + 1) *
625 sizeof(struct lttng_kconsumerd_fd));
626 if (local_kconsumerd_fd == NULL) {
627 perror("local_kconsumerd_fd malloc");
628 pthread_mutex_unlock(&kconsumerd_data.lock);
629 goto end;
630 }
631 ret = kconsumerd_update_poll_array(ctx, &pollfd, local_kconsumerd_fd);
632 if (ret < 0) {
633 ERR("Error in allocating pollfd or local_outfds");
634 lttng_kconsumerd_send_error(ctx, KCONSUMERD_POLL_ERROR);
635 pthread_mutex_unlock(&kconsumerd_data.lock);
636 goto end;
637 }
638 nb_fd = ret;
639 kconsumerd_data.need_update = 0;
640 }
641 pthread_mutex_unlock(&kconsumerd_data.lock);
642
643 /* poll on the array of fds */
644 DBG("polling on %d fd", nb_fd + 1);
645 num_rdy = poll(pollfd, nb_fd + 1, kconsumerd_poll_timeout);
646 DBG("poll num_rdy : %d", num_rdy);
647 if (num_rdy == -1) {
648 perror("Poll error");
649 lttng_kconsumerd_send_error(ctx, KCONSUMERD_POLL_ERROR);
650 goto end;
651 } else if (num_rdy == 0) {
652 DBG("Polling thread timed out");
653 goto end;
654 }
655
656 /* No FDs and kconsumerd_quit, kconsumerd_cleanup the thread */
657 if (nb_fd == 0 && kconsumerd_quit == 1) {
658 goto end;
659 }
660
661 /*
662 * If the kconsumerd_poll_pipe triggered poll go
663 * directly to the beginning of the loop to update the
664 * array. We want to prioritize array update over
665 * low-priority reads.
666 */
667 if (pollfd[nb_fd].revents == POLLIN) {
668 DBG("kconsumerd_poll_pipe wake up");
669 tmp2 = read(ctx->kconsumerd_poll_pipe[0], &tmp, 1);
670 if (tmp2 < 0) {
671 perror("read kconsumerd poll");
672 }
673 continue;
674 }
675
676 /* Take care of high priority channels first. */
677 for (i = 0; i < nb_fd; i++) {
678 switch(pollfd[i].revents) {
679 case POLLERR:
680 ERR("Error returned in polling fd %d.", pollfd[i].fd);
681 kconsumerd_del_fd(local_kconsumerd_fd[i]);
682 num_hup++;
683 break;
684 case POLLHUP:
685 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
686 kconsumerd_del_fd(local_kconsumerd_fd[i]);
687 num_hup++;
688 break;
689 case POLLNVAL:
690 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
691 kconsumerd_del_fd(local_kconsumerd_fd[i]);
692 num_hup++;
693 break;
694 case POLLPRI:
695 DBG("Urgent read on fd %d", pollfd[i].fd);
696 high_prio = 1;
697 ret = ctx->on_buffer_ready(local_kconsumerd_fd[i]);
698 /* it's ok to have an unavailable sub-buffer */
699 if (ret == EAGAIN) {
700 ret = 0;
701 }
702 break;
703 }
704 }
705
706 /* If every buffer FD has hung up, we end the read loop here */
707 if (nb_fd > 0 && num_hup == nb_fd) {
708 DBG("every buffer FD has hung up\n");
709 if (kconsumerd_quit == 1) {
710 goto end;
711 }
712 continue;
713 }
714
715 /* Take care of low priority channels. */
716 if (high_prio == 0) {
717 for (i = 0; i < nb_fd; i++) {
718 if (pollfd[i].revents == POLLIN) {
719 DBG("Normal read on fd %d", pollfd[i].fd);
720 ret = ctx->on_buffer_ready(local_kconsumerd_fd[i]);
721 /* it's ok to have an unavailable subbuffer */
722 if (ret == EAGAIN) {
723 ret = 0;
724 }
725 }
726 }
727 }
728 }
729 end:
730 DBG("polling thread exiting");
731 if (pollfd != NULL) {
732 free(pollfd);
733 pollfd = NULL;
734 }
735 if (local_kconsumerd_fd != NULL) {
736 free(local_kconsumerd_fd);
737 local_kconsumerd_fd = NULL;
738 }
739 return NULL;
740 }
741
742 /*
743 * Initialise the necessary environnement :
744 * - create a new context
745 * - create the poll_pipe
746 * - create the should_quit pipe (for signal handler)
747 * - create the thread pipe (for splice)
748 *
749 * Takes a function pointer as argument, this function is called when data is
750 * available on a buffer. This function is responsible to do the
751 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
752 * buffer configuration and then kernctl_put_next_subbuf at the end.
753 *
754 * Returns a pointer to the new context or NULL on error.
755 */
756 struct lttng_kconsumerd_local_data *lttng_kconsumerd_create(
757 int (*buffer_ready)(struct lttng_kconsumerd_fd *kconsumerd_fd))
758 {
759 int ret;
760 struct lttng_kconsumerd_local_data *ctx;
761
762 ctx = malloc(sizeof(struct lttng_kconsumerd_local_data));
763 if (ctx == NULL) {
764 perror("allocating context");
765 goto end;
766 }
767
768 ctx->on_buffer_ready = buffer_ready;
769
770 ret = pipe(ctx->kconsumerd_poll_pipe);
771 if (ret < 0) {
772 perror("Error creating poll pipe");
773 ctx = NULL;
774 goto end;
775 }
776
777 ret = pipe(ctx->kconsumerd_should_quit);
778 if (ret < 0) {
779 perror("Error creating recv pipe");
780 ctx = NULL;
781 goto end;
782 }
783
784 ret = pipe(ctx->kconsumerd_thread_pipe);
785 if (ret < 0) {
786 perror("Error creating thread pipe");
787 ctx = NULL;
788 goto end;
789 }
790
791 end:
792 return ctx;
793 }
794
795 /*
796 * Close all fds associated with the instance and free the context.
797 */
798 void lttng_kconsumerd_destroy(struct lttng_kconsumerd_local_data *ctx)
799 {
800 close(ctx->kconsumerd_error_socket);
801 close(ctx->kconsumerd_thread_pipe[0]);
802 close(ctx->kconsumerd_thread_pipe[1]);
803 close(ctx->kconsumerd_poll_pipe[0]);
804 close(ctx->kconsumerd_poll_pipe[1]);
805 close(ctx->kconsumerd_should_quit[0]);
806 close(ctx->kconsumerd_should_quit[1]);
807 unlink(ctx->kconsumerd_command_sock_path);
808 free(ctx);
809 ctx = NULL;
810 }
811
812 /*
813 * This thread listens on the consumerd socket and receives the file
814 * descriptors from the session daemon.
815 */
816 void *lttng_kconsumerd_thread_receive_fds(void *data)
817 {
818 int sock, client_socket, ret;
819 struct lttcomm_kconsumerd_header tmp;
820 /*
821 * structure to poll for incoming data on communication socket avoids
822 * making blocking sockets.
823 */
824 struct pollfd kconsumerd_sockpoll[2];
825 struct lttng_kconsumerd_local_data *ctx = data;
826
827
828 DBG("Creating command socket %s", ctx->kconsumerd_command_sock_path);
829 unlink(ctx->kconsumerd_command_sock_path);
830 client_socket = lttcomm_create_unix_sock(ctx->kconsumerd_command_sock_path);
831 if (client_socket < 0) {
832 ERR("Cannot create command socket");
833 goto end;
834 }
835
836 ret = lttcomm_listen_unix_sock(client_socket);
837 if (ret < 0) {
838 goto end;
839 }
840
841 DBG("Sending ready command to ltt-sessiond");
842 ret = lttng_kconsumerd_send_error(ctx, KCONSUMERD_COMMAND_SOCK_READY);
843 if (ret < 0) {
844 ERR("Error sending ready command to ltt-sessiond");
845 goto end;
846 }
847
848 ret = fcntl(client_socket, F_SETFL, O_NONBLOCK);
849 if (ret < 0) {
850 perror("fcntl O_NONBLOCK");
851 goto end;
852 }
853
854 /* prepare the FDs to poll : to client socket and the should_quit pipe */
855 kconsumerd_sockpoll[0].fd = ctx->kconsumerd_should_quit[0];
856 kconsumerd_sockpoll[0].events = POLLIN | POLLPRI;
857 kconsumerd_sockpoll[1].fd = client_socket;
858 kconsumerd_sockpoll[1].events = POLLIN | POLLPRI;
859
860 if (lttng_kconsumerd_poll_socket(kconsumerd_sockpoll) < 0) {
861 goto end;
862 }
863 DBG("Connection on client_socket");
864
865 /* Blocking call, waiting for transmission */
866 sock = lttcomm_accept_unix_sock(client_socket);
867 if (sock <= 0) {
868 WARN("On accept");
869 goto end;
870 }
871 ret = fcntl(sock, F_SETFL, O_NONBLOCK);
872 if (ret < 0) {
873 perror("fcntl O_NONBLOCK");
874 goto end;
875 }
876
877 /* update the polling structure to poll on the established socket */
878 kconsumerd_sockpoll[1].fd = sock;
879 kconsumerd_sockpoll[1].events = POLLIN | POLLPRI;
880
881 while (1) {
882 if (lttng_kconsumerd_poll_socket(kconsumerd_sockpoll) < 0) {
883 goto end;
884 }
885 DBG("Incoming fds on sock");
886
887 /* We first get the number of fd we are about to receive */
888 ret = lttcomm_recv_unix_sock(sock, &tmp,
889 sizeof(struct lttcomm_kconsumerd_header));
890 if (ret <= 0) {
891 ERR("Communication interrupted on command socket");
892 goto end;
893 }
894 if (tmp.cmd_type == STOP) {
895 DBG("Received STOP command");
896 goto end;
897 }
898 if (kconsumerd_quit) {
899 DBG("kconsumerd_thread_receive_fds received quit from signal");
900 goto end;
901 }
902
903 /* we received a command to add or update fds */
904 ret = kconsumerd_consumerd_recv_fd(ctx, sock, kconsumerd_sockpoll,
905 tmp.payload_size, tmp.cmd_type);
906 if (ret < 0) {
907 ERR("Receiving the FD, exiting");
908 goto end;
909 }
910 DBG("received fds on sock");
911 }
912
913 end:
914 DBG("kconsumerd_thread_receive_fds exiting");
915
916 /*
917 * when all fds have hung up, the polling thread
918 * can exit cleanly
919 */
920 kconsumerd_quit = 1;
921
922 /*
923 * 2s of grace period, if no polling events occur during
924 * this period, the polling thread will exit even if there
925 * are still open FDs (should not happen, but safety mechanism).
926 */
927 kconsumerd_poll_timeout = LTTNG_KCONSUMERD_POLL_GRACE_PERIOD;
928
929 /* wake up the polling thread */
930 ret = write(ctx->kconsumerd_poll_pipe[1], "4", 1);
931 if (ret < 0) {
932 perror("poll pipe write");
933 }
934 return NULL;
935 }
936
937 /*
938 * Close all the tracefiles and stream fds, should be called when all instances
939 * are destroyed.
940 */
941 void lttng_kconsumerd_cleanup(void)
942 {
943 struct lttng_kconsumerd_fd *iter, *tmp;
944
945 /*
946 * close all outfd. Called when there are no more threads
947 * running (after joining on the threads), no need to protect
948 * list iteration with mutex.
949 */
950 cds_list_for_each_entry_safe(iter, tmp,
951 &kconsumerd_data.fd_list.head, list) {
952 kconsumerd_del_fd(iter);
953 }
954 }
955
956 /*
957 * Called from signal handler.
958 */
959 void lttng_kconsumerd_should_exit(struct lttng_kconsumerd_local_data *ctx)
960 {
961 int ret;
962 kconsumerd_quit = 1;
963 ret = write(ctx->kconsumerd_should_quit[1], "4", 1);
964 if (ret < 0) {
965 perror("write kconsumerd quit");
966 }
967 }
968
969 /*
970 * Send return code to the session daemon.
971 */
972 int lttng_kconsumerd_send_error(
973 struct lttng_kconsumerd_local_data *ctx, int cmd)
974 {
975 if (ctx->kconsumerd_error_socket > 0) {
976 return lttcomm_send_unix_sock(ctx->kconsumerd_error_socket, &cmd,
977 sizeof(enum lttcomm_sessiond_command));
978 }
979
980 return 0;
981 }
This page took 0.069526 seconds and 4 git commands to generate.