Commit | Line | Data |
---|---|---|
1ce86c9a JD |
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 | |
82a3637f DG |
7 | * as published by the Free Software Foundation; only version 2 |
8 | * of the License. | |
1ce86c9a JD |
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> | |
6533b585 DG |
31 | |
32 | #include <lttng/lttng-kconsumerd.h> | |
1ce86c9a | 33 | |
50ecdf72 | 34 | #include "kernelctl.h" |
1ce86c9a | 35 | #include "lttngerr.h" |
6533b585 | 36 | #include "lttng-sessiond-comm.h" |
1ce86c9a | 37 | |
6533b585 | 38 | static struct lttng_kconsumerd_global_data { |
242cd187 MD |
39 | /* |
40 | * kconsumerd_data.lock protects kconsumerd_data.fd_list, | |
6533b585 DG |
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). | |
242cd187 MD |
46 | */ |
47 | pthread_mutex_t lock; | |
48 | /* | |
6533b585 | 49 | * Number of element for the list below. Protected by kconsumerd_data.lock. |
242cd187 MD |
50 | */ |
51 | unsigned int fds_count; | |
52 | /* | |
6533b585 | 53 | * List of FDs. Protected by kconsumerd_data.lock. |
242cd187 | 54 | */ |
6533b585 | 55 | struct lttng_kconsumerd_fd_list fd_list; |
242cd187 | 56 | /* |
6533b585 DG |
57 | * Flag specifying if the local array of FDs needs update in the poll |
58 | * function. Protected by kconsumerd_data.lock. | |
242cd187 MD |
59 | */ |
60 | unsigned int need_update; | |
61 | } kconsumerd_data = { | |
62 | .fd_list.head = CDS_LIST_HEAD_INIT(kconsumerd_data.fd_list.head), | |
cb040cc1 JD |
63 | .fds_count = 0, |
64 | .need_update = 1, | |
1ce86c9a JD |
65 | }; |
66 | ||
6533b585 | 67 | /* timeout parameter, to control the polling thread grace period. */ |
1ce86c9a JD |
68 | static int kconsumerd_poll_timeout = -1; |
69 | ||
3dcd2721 | 70 | /* |
6533b585 DG |
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. | |
3dcd2721 MD |
75 | */ |
76 | static volatile int kconsumerd_quit = 0; | |
1ce86c9a JD |
77 | |
78 | /* | |
6533b585 DG |
79 | * Find a session fd in the global list. The kconsumerd_data.lock must be |
80 | * locked during this call. | |
38079a1b | 81 | * |
6533b585 | 82 | * Return 1 if found else 0. |
38079a1b DG |
83 | */ |
84 | static int kconsumerd_find_session_fd(int fd) | |
85 | { | |
6533b585 | 86 | struct lttng_kconsumerd_fd *iter; |
38079a1b | 87 | |
242cd187 | 88 | cds_list_for_each_entry(iter, &kconsumerd_data.fd_list.head, list) { |
38079a1b DG |
89 | if (iter->sessiond_fd == fd) { |
90 | DBG("Duplicate session fd %d", fd); | |
38079a1b DG |
91 | return 1; |
92 | } | |
93 | } | |
38079a1b DG |
94 | |
95 | return 0; | |
96 | } | |
97 | ||
1ce86c9a | 98 | /* |
6533b585 | 99 | * Remove a fd from the global list protected by a mutex. |
1ce86c9a | 100 | */ |
6533b585 | 101 | static void kconsumerd_del_fd(struct lttng_kconsumerd_fd *lcf) |
1ce86c9a | 102 | { |
8b270bdb | 103 | int ret; |
242cd187 | 104 | pthread_mutex_lock(&kconsumerd_data.lock); |
1ce86c9a | 105 | cds_list_del(&lcf->list); |
242cd187 MD |
106 | if (kconsumerd_data.fds_count > 0) { |
107 | kconsumerd_data.fds_count--; | |
1ce86c9a | 108 | if (lcf != NULL) { |
8b270bdb JD |
109 | if (lcf->mmap_base != NULL) { |
110 | ret = munmap(lcf->mmap_base, lcf->mmap_len); | |
111 | if (ret != 0) { | |
112 | perror("munmap"); | |
113 | } | |
114 | } | |
cb040cc1 JD |
115 | if (lcf->out_fd != 0) { |
116 | close(lcf->out_fd); | |
117 | } | |
1ce86c9a JD |
118 | close(lcf->consumerd_fd); |
119 | free(lcf); | |
120 | lcf = NULL; | |
121 | } | |
122 | } | |
242cd187 MD |
123 | kconsumerd_data.need_update = 1; |
124 | pthread_mutex_unlock(&kconsumerd_data.lock); | |
1ce86c9a JD |
125 | } |
126 | ||
127 | /* | |
6533b585 | 128 | * Add a fd to the global list protected by a mutex. |
1ce86c9a | 129 | */ |
6533b585 DG |
130 | static int kconsumerd_add_fd(struct lttcomm_kconsumerd_msg *buf, |
131 | int consumerd_fd) | |
1ce86c9a | 132 | { |
6533b585 | 133 | struct lttng_kconsumerd_fd *tmp_fd; |
cb040cc1 | 134 | int ret = 0; |
38079a1b | 135 | |
242cd187 | 136 | pthread_mutex_lock(&kconsumerd_data.lock); |
38079a1b DG |
137 | /* Check if already exist */ |
138 | ret = kconsumerd_find_session_fd(buf->fd); | |
139 | if (ret == 1) { | |
140 | goto end; | |
141 | } | |
1ce86c9a | 142 | |
6533b585 | 143 | tmp_fd = malloc(sizeof(struct lttng_kconsumerd_fd)); |
1ce86c9a JD |
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; | |
cb040cc1 JD |
148 | tmp_fd->out_fd = 0; |
149 | tmp_fd->out_fd_offset = 0; | |
8b270bdb JD |
150 | tmp_fd->mmap_len = 0; |
151 | tmp_fd->mmap_base = NULL; | |
152 | tmp_fd->output = buf->output; | |
1ce86c9a | 153 | strncpy(tmp_fd->path_name, buf->path_name, PATH_MAX); |
99497cd0 | 154 | tmp_fd->path_name[PATH_MAX - 1] = '\0'; |
1ce86c9a JD |
155 | |
156 | /* Opening the tracefile in write mode */ | |
cb040cc1 JD |
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); | |
1ce86c9a | 168 | } |
1ce86c9a | 169 | |
8b270bdb JD |
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 | ||
242cd187 MD |
188 | cds_list_add(&tmp_fd->list, &kconsumerd_data.fd_list.head); |
189 | kconsumerd_data.fds_count++; | |
190 | kconsumerd_data.need_update = 1; | |
1ce86c9a | 191 | end: |
242cd187 | 192 | pthread_mutex_unlock(&kconsumerd_data.lock); |
1ce86c9a JD |
193 | return ret; |
194 | } | |
195 | ||
196 | /* | |
6533b585 | 197 | * Update a fd according to what we just received. |
1ce86c9a JD |
198 | */ |
199 | static void kconsumerd_change_fd_state(int sessiond_fd, | |
6533b585 | 200 | enum lttng_kconsumerd_fd_state state) |
1ce86c9a | 201 | { |
6533b585 | 202 | struct lttng_kconsumerd_fd *iter; |
0237248c | 203 | |
242cd187 MD |
204 | pthread_mutex_lock(&kconsumerd_data.lock); |
205 | cds_list_for_each_entry(iter, &kconsumerd_data.fd_list.head, list) { | |
1ce86c9a JD |
206 | if (iter->sessiond_fd == sessiond_fd) { |
207 | iter->state = state; | |
208 | break; | |
209 | } | |
210 | } | |
242cd187 MD |
211 | kconsumerd_data.need_update = 1; |
212 | pthread_mutex_unlock(&kconsumerd_data.lock); | |
1ce86c9a JD |
213 | } |
214 | ||
215 | /* | |
6533b585 DG |
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. | |
1ce86c9a | 219 | * |
6533b585 | 220 | * Returns the number of fds in the structures. |
1ce86c9a | 221 | */ |
6533b585 DG |
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) | |
1ce86c9a | 225 | { |
6533b585 | 226 | struct lttng_kconsumerd_fd *iter; |
1ce86c9a JD |
227 | int i = 0; |
228 | ||
229 | DBG("Updating poll fd array"); | |
242cd187 | 230 | cds_list_for_each_entry(iter, &kconsumerd_data.fd_list.head, list) { |
1ce86c9a JD |
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 | /* | |
6533b585 DG |
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. | |
1ce86c9a | 243 | */ |
cb040cc1 | 244 | (*pollfd)[i].fd = ctx->kconsumerd_poll_pipe[0]; |
1ce86c9a | 245 | (*pollfd)[i].events = POLLIN; |
1ce86c9a JD |
246 | return i; |
247 | } | |
248 | ||
6533b585 DG |
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 | } | |
1ce86c9a JD |
363 | |
364 | /* | |
6533b585 | 365 | * Mmap the ring buffer, read it and write the data to the tracefile. |
1ce86c9a | 366 | * |
1ce86c9a JD |
367 | * Returns the number of bytes written |
368 | */ | |
6533b585 DG |
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) | |
1ce86c9a | 372 | { |
8b270bdb | 373 | unsigned long mmap_offset; |
1ce86c9a JD |
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 | ||
1ce86c9a JD |
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 | ||
1ce86c9a | 388 | while (len > 0) { |
8b270bdb | 389 | ret = write(outfd, kconsumerd_fd->mmap_base + mmap_offset, len); |
1ce86c9a JD |
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 | ||
1ce86c9a JD |
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 | /* | |
1ce86c9a | 443 | * Splice the data from the ring buffer to the tracefile. |
6533b585 DG |
444 | * |
445 | * Returns the number of bytes spliced. | |
1ce86c9a | 446 | */ |
6533b585 DG |
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) | |
1ce86c9a JD |
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); | |
cb040cc1 | 460 | ret = splice(fd, &offset, ctx->kconsumerd_thread_pipe[1], NULL, len, |
1ce86c9a JD |
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 | ||
cb040cc1 | 469 | ret = splice(ctx->kconsumerd_thread_pipe[0], NULL, outfd, NULL, ret, |
1ce86c9a JD |
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 | } | |
751667bd | 477 | len -= ret; |
1ce86c9a JD |
478 | /* This won't block, but will start writeout asynchronously */ |
479 | sync_file_range(outfd, kconsumerd_fd->out_fd_offset, ret, | |
480 | SYNC_FILE_RANGE_WRITE); | |
481 | kconsumerd_fd->out_fd_offset += ret; | |
482 | } | |
483 | ||
484 | /* | |
485 | * This does a blocking write-and-wait on any page that belongs to the | |
486 | * subbuffer prior to the one we just wrote. | |
487 | * Don't care about error values, as these are just hints and ways to | |
488 | * limit the amount of page cache used. | |
489 | */ | |
490 | if (orig_offset >= kconsumerd_fd->max_sb_size) { | |
491 | sync_file_range(outfd, orig_offset - kconsumerd_fd->max_sb_size, | |
492 | kconsumerd_fd->max_sb_size, | |
493 | SYNC_FILE_RANGE_WAIT_BEFORE | |
494 | | SYNC_FILE_RANGE_WRITE | |
495 | | SYNC_FILE_RANGE_WAIT_AFTER); | |
496 | /* | |
497 | * Give hints to the kernel about how we access the file: | |
498 | * POSIX_FADV_DONTNEED : we won't re-access data in a near future after | |
499 | * we write it. | |
500 | * | |
501 | * We need to call fadvise again after the file grows because the | |
502 | * kernel does not seem to apply fadvise to non-existing parts of the | |
503 | * file. | |
504 | * | |
505 | * Call fadvise _after_ having waited for the page writeback to | |
506 | * complete because the dirty page writeback semantic is not well | |
507 | * defined. So it can be expected to lead to lower throughput in | |
508 | * streaming. | |
509 | */ | |
510 | posix_fadvise(outfd, orig_offset - kconsumerd_fd->max_sb_size, | |
511 | kconsumerd_fd->max_sb_size, POSIX_FADV_DONTNEED); | |
512 | } | |
513 | goto end; | |
514 | ||
515 | splice_error: | |
516 | /* send the appropriate error description to sessiond */ | |
517 | switch(ret) { | |
518 | case EBADF: | |
6533b585 | 519 | lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_EBADF); |
1ce86c9a JD |
520 | break; |
521 | case EINVAL: | |
6533b585 | 522 | lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_EINVAL); |
1ce86c9a JD |
523 | break; |
524 | case ENOMEM: | |
6533b585 | 525 | lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_ENOMEM); |
1ce86c9a JD |
526 | break; |
527 | case ESPIPE: | |
6533b585 | 528 | lttng_kconsumerd_send_error(ctx, KCONSUMERD_SPLICE_ESPIPE); |
1ce86c9a | 529 | break; |
1ce86c9a JD |
530 | } |
531 | ||
532 | end: | |
533 | return ret; | |
534 | } | |
535 | ||
4de84ad9 | 536 | /* |
6533b585 DG |
537 | * Poll on the should_quit pipe and the command socket return -1 on error and |
538 | * should exit, 0 if data is available on the command socket | |
4de84ad9 | 539 | */ |
6533b585 | 540 | int lttng_kconsumerd_poll_socket(struct pollfd *kconsumerd_sockpoll) |
4de84ad9 JD |
541 | { |
542 | int num_rdy; | |
543 | ||
544 | num_rdy = poll(kconsumerd_sockpoll, 2, -1); | |
545 | if (num_rdy == -1) { | |
546 | perror("Poll error"); | |
547 | goto exit; | |
548 | } | |
549 | if (kconsumerd_sockpoll[0].revents == POLLIN) { | |
550 | DBG("kconsumerd_should_quit wake up"); | |
551 | goto exit; | |
552 | } | |
553 | return 0; | |
554 | ||
555 | exit: | |
556 | return -1; | |
557 | } | |
558 | ||
1ce86c9a | 559 | /* |
6533b585 DG |
560 | * This thread polls the fds in the ltt_fd_list to consume the data and write |
561 | * it to tracefile if necessary. | |
1ce86c9a | 562 | */ |
6533b585 | 563 | void *lttng_kconsumerd_thread_poll_fds(void *data) |
1ce86c9a JD |
564 | { |
565 | int num_rdy, num_hup, high_prio, ret, i; | |
566 | struct pollfd *pollfd = NULL; | |
567 | /* local view of the fds */ | |
6533b585 | 568 | struct lttng_kconsumerd_fd **local_kconsumerd_fd = NULL; |
242cd187 | 569 | /* local view of kconsumerd_data.fds_count */ |
1ce86c9a JD |
570 | int nb_fd = 0; |
571 | char tmp; | |
572 | int tmp2; | |
6533b585 | 573 | struct lttng_kconsumerd_local_data *ctx = data; |
1ce86c9a | 574 | |
1ce86c9a | 575 | |
6533b585 | 576 | local_kconsumerd_fd = malloc(sizeof(struct lttng_kconsumerd_fd)); |
1ce86c9a JD |
577 | |
578 | while (1) { | |
579 | high_prio = 0; | |
580 | num_hup = 0; | |
581 | ||
582 | /* | |
583 | * the ltt_fd_list has been updated, we need to update our | |
584 | * local array as well | |
585 | */ | |
242cd187 MD |
586 | pthread_mutex_lock(&kconsumerd_data.lock); |
587 | if (kconsumerd_data.need_update) { | |
1ce86c9a JD |
588 | if (pollfd != NULL) { |
589 | free(pollfd); | |
590 | pollfd = NULL; | |
591 | } | |
592 | if (local_kconsumerd_fd != NULL) { | |
593 | free(local_kconsumerd_fd); | |
594 | local_kconsumerd_fd = NULL; | |
595 | } | |
0237248c | 596 | |
1ce86c9a | 597 | /* allocate for all fds + 1 for the kconsumerd_poll_pipe */ |
242cd187 | 598 | pollfd = malloc((kconsumerd_data.fds_count + 1) * sizeof(struct pollfd)); |
1ce86c9a JD |
599 | if (pollfd == NULL) { |
600 | perror("pollfd malloc"); | |
242cd187 | 601 | pthread_mutex_unlock(&kconsumerd_data.lock); |
1ce86c9a JD |
602 | goto end; |
603 | } | |
0237248c | 604 | |
1ce86c9a | 605 | /* allocate for all fds + 1 for the kconsumerd_poll_pipe */ |
242cd187 | 606 | local_kconsumerd_fd = malloc((kconsumerd_data.fds_count + 1) * |
6533b585 | 607 | sizeof(struct lttng_kconsumerd_fd)); |
1ce86c9a JD |
608 | if (local_kconsumerd_fd == NULL) { |
609 | perror("local_kconsumerd_fd malloc"); | |
242cd187 | 610 | pthread_mutex_unlock(&kconsumerd_data.lock); |
1ce86c9a JD |
611 | goto end; |
612 | } | |
cb040cc1 | 613 | ret = kconsumerd_update_poll_array(ctx, &pollfd, local_kconsumerd_fd); |
1ce86c9a JD |
614 | if (ret < 0) { |
615 | ERR("Error in allocating pollfd or local_outfds"); | |
6533b585 | 616 | lttng_kconsumerd_send_error(ctx, KCONSUMERD_POLL_ERROR); |
242cd187 | 617 | pthread_mutex_unlock(&kconsumerd_data.lock); |
1ce86c9a JD |
618 | goto end; |
619 | } | |
620 | nb_fd = ret; | |
242cd187 | 621 | kconsumerd_data.need_update = 0; |
1ce86c9a | 622 | } |
242cd187 | 623 | pthread_mutex_unlock(&kconsumerd_data.lock); |
1ce86c9a JD |
624 | |
625 | /* poll on the array of fds */ | |
626 | DBG("polling on %d fd", nb_fd + 1); | |
627 | num_rdy = poll(pollfd, nb_fd + 1, kconsumerd_poll_timeout); | |
628 | DBG("poll num_rdy : %d", num_rdy); | |
629 | if (num_rdy == -1) { | |
630 | perror("Poll error"); | |
6533b585 | 631 | lttng_kconsumerd_send_error(ctx, KCONSUMERD_POLL_ERROR); |
1ce86c9a JD |
632 | goto end; |
633 | } else if (num_rdy == 0) { | |
634 | DBG("Polling thread timed out"); | |
635 | goto end; | |
636 | } | |
637 | ||
638 | /* No FDs and kconsumerd_quit, kconsumerd_cleanup the thread */ | |
639 | if (nb_fd == 0 && kconsumerd_quit == 1) { | |
640 | goto end; | |
641 | } | |
642 | ||
643 | /* | |
242cd187 MD |
644 | * If the kconsumerd_poll_pipe triggered poll go |
645 | * directly to the beginning of the loop to update the | |
646 | * array. We want to prioritize array update over | |
647 | * low-priority reads. | |
1ce86c9a | 648 | */ |
242cd187 | 649 | if (pollfd[nb_fd].revents == POLLIN) { |
1ce86c9a | 650 | DBG("kconsumerd_poll_pipe wake up"); |
cb040cc1 | 651 | tmp2 = read(ctx->kconsumerd_poll_pipe[0], &tmp, 1); |
f40799e8 DG |
652 | if (tmp2 < 0) { |
653 | perror("read kconsumerd poll"); | |
654 | } | |
1ce86c9a JD |
655 | continue; |
656 | } | |
657 | ||
658 | /* Take care of high priority channels first. */ | |
659 | for (i = 0; i < nb_fd; i++) { | |
660 | switch(pollfd[i].revents) { | |
661 | case POLLERR: | |
662 | ERR("Error returned in polling fd %d.", pollfd[i].fd); | |
663 | kconsumerd_del_fd(local_kconsumerd_fd[i]); | |
1ce86c9a JD |
664 | num_hup++; |
665 | break; | |
666 | case POLLHUP: | |
667 | DBG("Polling fd %d tells it has hung up.", pollfd[i].fd); | |
668 | kconsumerd_del_fd(local_kconsumerd_fd[i]); | |
1ce86c9a JD |
669 | num_hup++; |
670 | break; | |
671 | case POLLNVAL: | |
672 | ERR("Polling fd %d tells fd is not open.", pollfd[i].fd); | |
673 | kconsumerd_del_fd(local_kconsumerd_fd[i]); | |
1ce86c9a JD |
674 | num_hup++; |
675 | break; | |
676 | case POLLPRI: | |
677 | DBG("Urgent read on fd %d", pollfd[i].fd); | |
678 | high_prio = 1; | |
cb040cc1 | 679 | ret = ctx->on_buffer_ready(local_kconsumerd_fd[i]); |
1ce86c9a JD |
680 | /* it's ok to have an unavailable sub-buffer */ |
681 | if (ret == EAGAIN) { | |
682 | ret = 0; | |
683 | } | |
684 | break; | |
685 | } | |
686 | } | |
687 | ||
688 | /* If every buffer FD has hung up, we end the read loop here */ | |
689 | if (nb_fd > 0 && num_hup == nb_fd) { | |
690 | DBG("every buffer FD has hung up\n"); | |
691 | if (kconsumerd_quit == 1) { | |
692 | goto end; | |
693 | } | |
694 | continue; | |
695 | } | |
696 | ||
697 | /* Take care of low priority channels. */ | |
698 | if (high_prio == 0) { | |
699 | for (i = 0; i < nb_fd; i++) { | |
700 | if (pollfd[i].revents == POLLIN) { | |
701 | DBG("Normal read on fd %d", pollfd[i].fd); | |
cb040cc1 | 702 | ret = ctx->on_buffer_ready(local_kconsumerd_fd[i]); |
1ce86c9a JD |
703 | /* it's ok to have an unavailable subbuffer */ |
704 | if (ret == EAGAIN) { | |
705 | ret = 0; | |
706 | } | |
707 | } | |
708 | } | |
709 | } | |
710 | } | |
711 | end: | |
712 | DBG("polling thread exiting"); | |
713 | if (pollfd != NULL) { | |
714 | free(pollfd); | |
715 | pollfd = NULL; | |
716 | } | |
717 | if (local_kconsumerd_fd != NULL) { | |
718 | free(local_kconsumerd_fd); | |
719 | local_kconsumerd_fd = NULL; | |
720 | } | |
1ce86c9a JD |
721 | return NULL; |
722 | } | |
723 | ||
724 | /* | |
6533b585 | 725 | * Initialise the necessary environnement : |
cb040cc1 | 726 | * - create a new context |
4de84ad9 JD |
727 | * - create the poll_pipe |
728 | * - create the should_quit pipe (for signal handler) | |
cb040cc1 | 729 | * - create the thread pipe (for splice) |
6533b585 | 730 | * |
cb040cc1 JD |
731 | * Takes a function pointer as argument, this function is called when data is |
732 | * available on a buffer. This function is responsible to do the | |
733 | * kernctl_get_next_subbuf, read the data with mmap or splice depending on the | |
734 | * buffer configuration and then kernctl_put_next_subbuf at the end. | |
6533b585 | 735 | * |
cb040cc1 | 736 | * Returns a pointer to the new context or NULL on error. |
1ce86c9a | 737 | */ |
6533b585 DG |
738 | struct lttng_kconsumerd_local_data *lttng_kconsumerd_create( |
739 | int (*buffer_ready)(struct lttng_kconsumerd_fd *kconsumerd_fd)) | |
1ce86c9a | 740 | { |
4de84ad9 | 741 | int ret; |
6533b585 | 742 | struct lttng_kconsumerd_local_data *ctx; |
4de84ad9 | 743 | |
6533b585 | 744 | ctx = malloc(sizeof(struct lttng_kconsumerd_local_data)); |
cb040cc1 JD |
745 | if (ctx == NULL) { |
746 | perror("allocating context"); | |
747 | goto end; | |
748 | } | |
749 | ||
750 | ctx->on_buffer_ready = buffer_ready; | |
4de84ad9 | 751 | |
cb040cc1 | 752 | ret = pipe(ctx->kconsumerd_poll_pipe); |
4de84ad9 JD |
753 | if (ret < 0) { |
754 | perror("Error creating poll pipe"); | |
cb040cc1 | 755 | ctx = NULL; |
4de84ad9 JD |
756 | goto end; |
757 | } | |
758 | ||
cb040cc1 | 759 | ret = pipe(ctx->kconsumerd_should_quit); |
4de84ad9 JD |
760 | if (ret < 0) { |
761 | perror("Error creating recv pipe"); | |
cb040cc1 JD |
762 | ctx = NULL; |
763 | goto end; | |
764 | } | |
765 | ||
766 | ret = pipe(ctx->kconsumerd_thread_pipe); | |
767 | if (ret < 0) { | |
768 | perror("Error creating thread pipe"); | |
769 | ctx = NULL; | |
4de84ad9 JD |
770 | goto end; |
771 | } | |
772 | ||
773 | end: | |
cb040cc1 JD |
774 | return ctx; |
775 | } | |
776 | ||
777 | /* | |
6533b585 | 778 | * Close all fds associated with the instance and free the context. |
cb040cc1 | 779 | */ |
6533b585 | 780 | void lttng_kconsumerd_destroy(struct lttng_kconsumerd_local_data *ctx) |
cb040cc1 JD |
781 | { |
782 | close(ctx->kconsumerd_error_socket); | |
783 | close(ctx->kconsumerd_thread_pipe[0]); | |
784 | close(ctx->kconsumerd_thread_pipe[1]); | |
785 | close(ctx->kconsumerd_poll_pipe[0]); | |
786 | close(ctx->kconsumerd_poll_pipe[1]); | |
787 | close(ctx->kconsumerd_should_quit[0]); | |
788 | close(ctx->kconsumerd_should_quit[1]); | |
789 | unlink(ctx->kconsumerd_command_sock_path); | |
790 | free(ctx); | |
791 | ctx = NULL; | |
1ce86c9a JD |
792 | } |
793 | ||
794 | /* | |
6533b585 DG |
795 | * This thread listens on the consumerd socket and receives the file |
796 | * descriptors from the session daemon. | |
1ce86c9a | 797 | */ |
6533b585 | 798 | void *lttng_kconsumerd_thread_receive_fds(void *data) |
1ce86c9a JD |
799 | { |
800 | int sock, client_socket, ret; | |
801 | struct lttcomm_kconsumerd_header tmp; | |
4de84ad9 | 802 | /* |
6533b585 DG |
803 | * structure to poll for incoming data on communication socket avoids |
804 | * making blocking sockets. | |
4de84ad9 JD |
805 | */ |
806 | struct pollfd kconsumerd_sockpoll[2]; | |
6533b585 | 807 | struct lttng_kconsumerd_local_data *ctx = data; |
4de84ad9 | 808 | |
1ce86c9a | 809 | |
cb040cc1 JD |
810 | DBG("Creating command socket %s", ctx->kconsumerd_command_sock_path); |
811 | unlink(ctx->kconsumerd_command_sock_path); | |
812 | client_socket = lttcomm_create_unix_sock(ctx->kconsumerd_command_sock_path); | |
1ce86c9a JD |
813 | if (client_socket < 0) { |
814 | ERR("Cannot create command socket"); | |
815 | goto end; | |
816 | } | |
817 | ||
818 | ret = lttcomm_listen_unix_sock(client_socket); | |
819 | if (ret < 0) { | |
820 | goto end; | |
821 | } | |
822 | ||
823 | DBG("Sending ready command to ltt-sessiond"); | |
6533b585 | 824 | ret = lttng_kconsumerd_send_error(ctx, KCONSUMERD_COMMAND_SOCK_READY); |
1ce86c9a JD |
825 | if (ret < 0) { |
826 | ERR("Error sending ready command to ltt-sessiond"); | |
827 | goto end; | |
828 | } | |
829 | ||
4de84ad9 JD |
830 | ret = fcntl(client_socket, F_SETFL, O_NONBLOCK); |
831 | if (ret < 0) { | |
832 | perror("fcntl O_NONBLOCK"); | |
833 | goto end; | |
834 | } | |
835 | ||
836 | /* prepare the FDs to poll : to client socket and the should_quit pipe */ | |
cb040cc1 | 837 | kconsumerd_sockpoll[0].fd = ctx->kconsumerd_should_quit[0]; |
4de84ad9 JD |
838 | kconsumerd_sockpoll[0].events = POLLIN | POLLPRI; |
839 | kconsumerd_sockpoll[1].fd = client_socket; | |
840 | kconsumerd_sockpoll[1].events = POLLIN | POLLPRI; | |
841 | ||
6533b585 | 842 | if (lttng_kconsumerd_poll_socket(kconsumerd_sockpoll) < 0) { |
4de84ad9 JD |
843 | goto end; |
844 | } | |
845 | DBG("Connection on client_socket"); | |
846 | ||
1ce86c9a JD |
847 | /* Blocking call, waiting for transmission */ |
848 | sock = lttcomm_accept_unix_sock(client_socket); | |
849 | if (sock <= 0) { | |
850 | WARN("On accept"); | |
851 | goto end; | |
852 | } | |
4de84ad9 JD |
853 | ret = fcntl(sock, F_SETFL, O_NONBLOCK); |
854 | if (ret < 0) { | |
855 | perror("fcntl O_NONBLOCK"); | |
856 | goto end; | |
857 | } | |
858 | ||
859 | /* update the polling structure to poll on the established socket */ | |
860 | kconsumerd_sockpoll[1].fd = sock; | |
861 | kconsumerd_sockpoll[1].events = POLLIN | POLLPRI; | |
862 | ||
1ce86c9a | 863 | while (1) { |
6533b585 | 864 | if (lttng_kconsumerd_poll_socket(kconsumerd_sockpoll) < 0) { |
4de84ad9 JD |
865 | goto end; |
866 | } | |
867 | DBG("Incoming fds on sock"); | |
868 | ||
1ce86c9a JD |
869 | /* We first get the number of fd we are about to receive */ |
870 | ret = lttcomm_recv_unix_sock(sock, &tmp, | |
871 | sizeof(struct lttcomm_kconsumerd_header)); | |
872 | if (ret <= 0) { | |
873 | ERR("Communication interrupted on command socket"); | |
874 | goto end; | |
875 | } | |
876 | if (tmp.cmd_type == STOP) { | |
877 | DBG("Received STOP command"); | |
878 | goto end; | |
879 | } | |
3dcd2721 MD |
880 | if (kconsumerd_quit) { |
881 | DBG("kconsumerd_thread_receive_fds received quit from signal"); | |
882 | goto end; | |
883 | } | |
4de84ad9 | 884 | |
1ce86c9a | 885 | /* we received a command to add or update fds */ |
cb040cc1 | 886 | ret = kconsumerd_consumerd_recv_fd(ctx, sock, kconsumerd_sockpoll, |
4de84ad9 | 887 | tmp.payload_size, tmp.cmd_type); |
8b270bdb | 888 | if (ret < 0) { |
1ce86c9a JD |
889 | ERR("Receiving the FD, exiting"); |
890 | goto end; | |
891 | } | |
4de84ad9 | 892 | DBG("received fds on sock"); |
1ce86c9a JD |
893 | } |
894 | ||
895 | end: | |
896 | DBG("kconsumerd_thread_receive_fds exiting"); | |
897 | ||
898 | /* | |
899 | * when all fds have hung up, the polling thread | |
900 | * can exit cleanly | |
901 | */ | |
902 | kconsumerd_quit = 1; | |
903 | ||
904 | /* | |
905 | * 2s of grace period, if no polling events occur during | |
906 | * this period, the polling thread will exit even if there | |
907 | * are still open FDs (should not happen, but safety mechanism). | |
908 | */ | |
6533b585 | 909 | kconsumerd_poll_timeout = LTTNG_KCONSUMERD_POLL_GRACE_PERIOD; |
1ce86c9a JD |
910 | |
911 | /* wake up the polling thread */ | |
cb040cc1 | 912 | ret = write(ctx->kconsumerd_poll_pipe[1], "4", 1); |
1ce86c9a JD |
913 | if (ret < 0) { |
914 | perror("poll pipe write"); | |
915 | } | |
916 | return NULL; | |
917 | } | |
918 | ||
919 | /* | |
6533b585 DG |
920 | * Close all the tracefiles and stream fds, should be called when all instances |
921 | * are destroyed. | |
1ce86c9a | 922 | */ |
6533b585 | 923 | void lttng_kconsumerd_cleanup(void) |
1ce86c9a | 924 | { |
6533b585 | 925 | struct lttng_kconsumerd_fd *iter, *tmp; |
1ce86c9a | 926 | |
3dcd2721 MD |
927 | /* |
928 | * close all outfd. Called when there are no more threads | |
929 | * running (after joining on the threads), no need to protect | |
930 | * list iteration with mutex. | |
931 | */ | |
6533b585 DG |
932 | cds_list_for_each_entry_safe(iter, tmp, |
933 | &kconsumerd_data.fd_list.head, list) { | |
1ce86c9a JD |
934 | kconsumerd_del_fd(iter); |
935 | } | |
936 | } | |
937 | ||
3dcd2721 MD |
938 | /* |
939 | * Called from signal handler. | |
940 | */ | |
6533b585 | 941 | void lttng_kconsumerd_should_exit(struct lttng_kconsumerd_local_data *ctx) |
3dcd2721 | 942 | { |
4de84ad9 | 943 | int ret; |
3dcd2721 | 944 | kconsumerd_quit = 1; |
cb040cc1 | 945 | ret = write(ctx->kconsumerd_should_quit[1], "4", 1); |
f40799e8 DG |
946 | if (ret < 0) { |
947 | perror("write kconsumerd quit"); | |
948 | } | |
3dcd2721 MD |
949 | } |
950 | ||
1ce86c9a | 951 | /* |
6533b585 | 952 | * Send return code to the session daemon. |
1ce86c9a | 953 | */ |
6533b585 DG |
954 | int lttng_kconsumerd_send_error( |
955 | struct lttng_kconsumerd_local_data *ctx, int cmd) | |
1ce86c9a | 956 | { |
cb040cc1 JD |
957 | if (ctx->kconsumerd_error_socket > 0) { |
958 | return lttcomm_send_unix_sock(ctx->kconsumerd_error_socket, &cmd, | |
1ce86c9a JD |
959 | sizeof(enum lttcomm_sessiond_command)); |
960 | } | |
961 | ||
962 | return 0; | |
963 | } |