Commit | Line | Data |
---|---|---|
b5b073e2 PMF |
1 | /* |
2 | * buffers.c | |
3 | * LTTng userspace tracer buffering system | |
4 | * | |
5 | * Copyright (C) 2009 - Pierre-Marc Fournier (pierre-marc dot fournier at polymtl dot ca) | |
d4d511d9 | 6 | * Copyright (C) 2008-2011 - Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca) |
b5b073e2 PMF |
7 | * |
8 | * This library is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
4fbf9cd1 MD |
10 | * License as published by the Free Software Foundation; version 2.1 of |
11 | * the License. | |
b5b073e2 PMF |
12 | * |
13 | * This library is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | * Lesser General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU Lesser General Public | |
19 | * License along with this library; if not, write to the Free Software | |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
21 | */ | |
22 | ||
d4d511d9 MD |
23 | /* |
24 | * Note: this code does not support the ref/noref flag and reader-owned | |
283f1158 MD |
25 | * subbuffer scheme. Therefore, flight recorder mode uses a mechanism |
26 | * where the reader can read corrupted data (and detect this), thus | |
27 | * returning -EIO. | |
d4d511d9 MD |
28 | */ |
29 | ||
204141ee | 30 | #include <unistd.h> |
b5b073e2 PMF |
31 | #include <sys/mman.h> |
32 | #include <sys/ipc.h> | |
33 | #include <sys/shm.h> | |
34 | #include <fcntl.h> | |
909bc43f | 35 | #include <stdlib.h> |
518d7abb PMF |
36 | |
37 | #include <ust/clock.h> | |
38 | ||
b5b073e2 PMF |
39 | #include "buffers.h" |
40 | #include "channels.h" | |
41 | #include "tracer.h" | |
42 | #include "tracercore.h" | |
1dbfff0c | 43 | #include <ust/usterr-signal-safe.h> |
b5b073e2 | 44 | |
b73a4c47 PMF |
45 | struct ltt_reserve_switch_offsets { |
46 | long begin, end, old; | |
47 | long begin_switch, end_switch_current, end_switch_old; | |
48 | size_t before_hdr_pad, size; | |
49 | }; | |
50 | ||
51 | ||
b5b073e2 | 52 | static DEFINE_MUTEX(ust_buffers_channels_mutex); |
0222e121 | 53 | static CDS_LIST_HEAD(ust_buffers_channels); |
b5b073e2 | 54 | |
d4d511d9 MD |
55 | static void ltt_force_switch(struct ust_buffer *buf, |
56 | enum force_switch_mode mode); | |
57 | ||
bb3132c8 MD |
58 | /** |
59 | * _ust_buffers_strncpy_fixup - Fix an incomplete string in a ltt_relay buffer. | |
60 | * @buf : buffer | |
61 | * @offset : offset within the buffer | |
62 | * @len : length to write | |
63 | * @copied: string actually copied | |
64 | * @terminated: does string end with \0 | |
b73a4c47 | 65 | * |
bb3132c8 | 66 | * Fills string with "X" if incomplete. |
b73a4c47 | 67 | */ |
bb3132c8 MD |
68 | void _ust_buffers_strncpy_fixup(struct ust_buffer *buf, size_t offset, |
69 | size_t len, size_t copied, int terminated) | |
b73a4c47 | 70 | { |
bb3132c8 MD |
71 | size_t buf_offset, cpy; |
72 | ||
73 | if (copied == len) { | |
74 | /* | |
75 | * Deal with non-terminated string. | |
76 | */ | |
77 | assert(!terminated); | |
78 | offset += copied - 1; | |
79 | buf_offset = BUFFER_OFFSET(offset, buf->chan); | |
80 | /* | |
81 | * Underlying layer should never ask for writes across | |
82 | * subbuffers. | |
83 | */ | |
84 | assert(buf_offset | |
85 | < buf->chan->subbuf_size*buf->chan->subbuf_cnt); | |
86 | ust_buffers_do_memset(buf->buf_data + buf_offset, '\0', 1); | |
87 | return; | |
88 | } | |
89 | ||
90 | /* | |
91 | * Deal with incomplete string. | |
92 | * Overwrite string's \0 with X too. | |
93 | */ | |
94 | cpy = copied - 1; | |
95 | assert(terminated); | |
96 | len -= cpy; | |
97 | offset += cpy; | |
98 | buf_offset = BUFFER_OFFSET(offset, buf->chan); | |
99 | ||
100 | /* | |
101 | * Underlying layer should never ask for writes across subbuffers. | |
102 | */ | |
103 | assert(buf_offset | |
104 | < buf->chan->subbuf_size*buf->chan->subbuf_cnt); | |
b73a4c47 | 105 | |
bb3132c8 MD |
106 | ust_buffers_do_memset(buf->buf_data + buf_offset, |
107 | 'X', len); | |
b73a4c47 | 108 | |
bb3132c8 MD |
109 | /* |
110 | * Overwrite last 'X' with '\0'. | |
111 | */ | |
112 | offset += len - 1; | |
113 | buf_offset = BUFFER_OFFSET(offset, buf->chan); | |
114 | /* | |
115 | * Underlying layer should never ask for writes across subbuffers. | |
116 | */ | |
117 | assert(buf_offset | |
118 | < buf->chan->subbuf_size*buf->chan->subbuf_cnt); | |
119 | ust_buffers_do_memset(buf->buf_data + buf_offset, '\0', 1); | |
b73a4c47 PMF |
120 | } |
121 | ||
ef15e552 NC |
122 | static void ltt_buffer_begin(struct ust_buffer *buf, |
123 | u64 tsc, unsigned int subbuf_idx) | |
124 | { | |
125 | struct ust_channel *channel = buf->chan; | |
126 | struct ltt_subbuffer_header *header = | |
127 | (struct ltt_subbuffer_header *) | |
128 | ust_buffers_offset_address(buf, | |
129 | subbuf_idx * buf->chan->subbuf_size); | |
b5b073e2 | 130 | |
ef15e552 NC |
131 | header->cycle_count_begin = tsc; |
132 | header->data_size = 0xFFFFFFFF; /* for recognizing crashed buffers */ | |
133 | header->sb_size = 0xFFFFFFFF; /* for recognizing crashed buffers */ | |
c74e3560 MD |
134 | /* |
135 | * No memory barrier needed to order data_data/sb_size vs commit count | |
136 | * update, because commit count update contains a compiler barrier that | |
137 | * ensures the order of the writes are OK from a program POV. It only | |
138 | * matters for crash dump recovery which is not executed concurrently, | |
139 | * so memory write order does not matter. | |
140 | */ | |
ef15e552 NC |
141 | ltt_write_trace_header(channel->trace, header); |
142 | } | |
143 | ||
144 | static int map_buf_data(struct ust_buffer *buf, size_t *size) | |
b5b073e2 PMF |
145 | { |
146 | void *ptr; | |
147 | int result; | |
148 | ||
149 | *size = PAGE_ALIGN(*size); | |
150 | ||
151 | result = buf->shmid = shmget(getpid(), *size, IPC_CREAT | IPC_EXCL | 0700); | |
ef15e552 | 152 | if (result < 0 && errno == EINVAL) { |
b5b073e2 PMF |
153 | ERR("shmget() returned EINVAL; maybe /proc/sys/kernel/shmmax should be increased."); |
154 | return -1; | |
ef15e552 | 155 | } else if (result < 0) { |
b5b073e2 PMF |
156 | PERROR("shmget"); |
157 | return -1; | |
158 | } | |
159 | ||
160 | ptr = shmat(buf->shmid, NULL, 0); | |
ef15e552 | 161 | if (ptr == (void *) -1) { |
b5b073e2 PMF |
162 | perror("shmat"); |
163 | goto destroy_shmem; | |
164 | } | |
165 | ||
166 | /* Already mark the shared memory for destruction. This will occur only | |
167 | * when all users have detached. | |
168 | */ | |
169 | result = shmctl(buf->shmid, IPC_RMID, NULL); | |
170 | if(result == -1) { | |
171 | perror("shmctl"); | |
172 | return -1; | |
173 | } | |
174 | ||
175 | buf->buf_data = ptr; | |
176 | buf->buf_size = *size; | |
177 | ||
178 | return 0; | |
179 | ||
ef15e552 | 180 | destroy_shmem: |
b5b073e2 PMF |
181 | result = shmctl(buf->shmid, IPC_RMID, NULL); |
182 | if(result == -1) { | |
183 | perror("shmctl"); | |
184 | } | |
185 | ||
186 | return -1; | |
187 | } | |
188 | ||
ef15e552 | 189 | static int open_buf(struct ust_channel *chan, int cpu) |
b5b073e2 | 190 | { |
ef15e552 NC |
191 | int result, fds[2]; |
192 | unsigned int j; | |
193 | struct ust_trace *trace = chan->trace; | |
194 | struct ust_buffer *buf = chan->buf[cpu]; | |
195 | unsigned int n_subbufs = chan->subbuf_cnt; | |
b5b073e2 | 196 | |
ef15e552 NC |
197 | |
198 | result = map_buf_data(buf, &chan->alloc_size); | |
199 | if (result < 0) | |
204141ee | 200 | return -1; |
b5b073e2 | 201 | |
ef15e552 NC |
202 | buf->commit_count = |
203 | zmalloc(sizeof(*buf->commit_count) * n_subbufs); | |
204 | if (!buf->commit_count) | |
205 | goto unmap_buf; | |
b5b073e2 | 206 | |
ef15e552 NC |
207 | result = pipe(fds); |
208 | if (result < 0) { | |
209 | PERROR("pipe"); | |
210 | goto free_commit_count; | |
211 | } | |
212 | buf->data_ready_fd_read = fds[0]; | |
213 | buf->data_ready_fd_write = fds[1]; | |
b5b073e2 | 214 | |
ef15e552 NC |
215 | buf->cpu = cpu; |
216 | buf->chan = chan; | |
b5b073e2 | 217 | |
ef15e552 NC |
218 | uatomic_set(&buf->offset, ltt_subbuffer_header_size()); |
219 | uatomic_set(&buf->consumed, 0); | |
220 | uatomic_set(&buf->active_readers, 0); | |
221 | for (j = 0; j < n_subbufs; j++) { | |
222 | uatomic_set(&buf->commit_count[j].cc, 0); | |
223 | uatomic_set(&buf->commit_count[j].cc_sb, 0); | |
b5b073e2 PMF |
224 | } |
225 | ||
ef15e552 | 226 | ltt_buffer_begin(buf, trace->start_tsc, 0); |
b5b073e2 | 227 | |
ef15e552 NC |
228 | uatomic_add(&buf->commit_count[0].cc, ltt_subbuffer_header_size()); |
229 | ||
230 | uatomic_set(&buf->events_lost, 0); | |
231 | uatomic_set(&buf->corrupted_subbuffers, 0); | |
232 | ||
233 | memset(buf->commit_seq, 0, sizeof(buf->commit_seq[0]) * n_subbufs); | |
234 | ||
235 | return 0; | |
236 | ||
237 | free_commit_count: | |
238 | free(buf->commit_count); | |
239 | ||
240 | unmap_buf: | |
241 | if (shmdt(buf->buf_data) < 0) { | |
242 | PERROR("shmdt failed"); | |
243 | } | |
244 | ||
245 | return -1; | |
b5b073e2 PMF |
246 | } |
247 | ||
ef15e552 | 248 | static void close_buf(struct ust_buffer *buf) |
b5b073e2 | 249 | { |
204141ee | 250 | int result; |
b5b073e2 | 251 | |
ef15e552 NC |
252 | result = shmdt(buf->buf_data); |
253 | if (result < 0) { | |
254 | PERROR("shmdt"); | |
255 | } | |
b5b073e2 | 256 | |
ef15e552 NC |
257 | result = close(buf->data_ready_fd_read); |
258 | if (result < 0) { | |
259 | PERROR("close"); | |
260 | } | |
b5b073e2 | 261 | |
ef15e552 NC |
262 | result = close(buf->data_ready_fd_write); |
263 | if (result < 0 && errno != EBADF) { | |
264 | PERROR("close"); | |
265 | } | |
b5b073e2 PMF |
266 | } |
267 | ||
b5b073e2 | 268 | |
ef15e552 NC |
269 | static int open_channel(struct ust_channel *chan, size_t subbuf_size, |
270 | size_t subbuf_cnt) | |
b5b073e2 | 271 | { |
204141ee PMF |
272 | int i; |
273 | int result; | |
274 | ||
b5b073e2 PMF |
275 | if(subbuf_size == 0 || subbuf_cnt == 0) |
276 | return -1; | |
277 | ||
b73a4c47 PMF |
278 | /* Check that the subbuffer size is larger than a page. */ |
279 | WARN_ON_ONCE(subbuf_size < PAGE_SIZE); | |
280 | ||
281 | /* | |
282 | * Make sure the number of subbuffers and subbuffer size are power of 2. | |
283 | */ | |
284 | WARN_ON_ONCE(hweight32(subbuf_size) != 1); | |
285 | WARN_ON(hweight32(subbuf_cnt) != 1); | |
286 | ||
b5b073e2 PMF |
287 | chan->version = UST_CHANNEL_VERSION; |
288 | chan->subbuf_cnt = subbuf_cnt; | |
289 | chan->subbuf_size = subbuf_size; | |
290 | chan->subbuf_size_order = get_count_order(subbuf_size); | |
b73a4c47 | 291 | chan->alloc_size = subbuf_size * subbuf_cnt; |
204141ee | 292 | |
f7b16408 | 293 | pthread_mutex_lock(&ust_buffers_channels_mutex); |
ef15e552 NC |
294 | for (i=0; i < chan->n_cpus; i++) { |
295 | result = open_buf(chan, i); | |
204141ee PMF |
296 | if (result == -1) |
297 | goto error; | |
298 | } | |
0222e121 | 299 | cds_list_add(&chan->list, &ust_buffers_channels); |
f7b16408 | 300 | pthread_mutex_unlock(&ust_buffers_channels_mutex); |
b5b073e2 PMF |
301 | |
302 | return 0; | |
303 | ||
b2ccc231 | 304 | /* Error handling */ |
204141ee | 305 | error: |
b2ccc231 MD |
306 | for(i--; i >= 0; i--) |
307 | close_buf(chan->buf[i]); | |
204141ee | 308 | |
f7b16408 | 309 | pthread_mutex_unlock(&ust_buffers_channels_mutex); |
b5b073e2 PMF |
310 | return -1; |
311 | } | |
312 | ||
ef15e552 | 313 | static void close_channel(struct ust_channel *chan) |
b5b073e2 | 314 | { |
204141ee PMF |
315 | int i; |
316 | if(!chan) | |
b5b073e2 PMF |
317 | return; |
318 | ||
f7b16408 | 319 | pthread_mutex_lock(&ust_buffers_channels_mutex); |
d4d511d9 MD |
320 | /* |
321 | * checking for chan->buf[i] being NULL or not is useless in | |
322 | * practice because we allocate buffers for all possible cpus. | |
323 | * However, should we decide to change this and only allocate | |
324 | * for online cpus, this check becomes useful. | |
325 | */ | |
326 | for (i=0; i<chan->n_cpus; i++) { | |
327 | if (chan->buf[i]) | |
ef15e552 | 328 | close_buf(chan->buf[i]); |
204141ee | 329 | } |
b5b073e2 | 330 | |
0222e121 | 331 | cds_list_del(&chan->list); |
ef15e552 | 332 | |
f7b16408 | 333 | pthread_mutex_unlock(&ust_buffers_channels_mutex); |
b5b073e2 PMF |
334 | } |
335 | ||
b5b073e2 PMF |
336 | /* |
337 | * offset is assumed to never be 0 here : never deliver a completely empty | |
338 | * subbuffer. The lost size is between 0 and subbuf_size-1. | |
339 | */ | |
b73a4c47 | 340 | static notrace void ltt_buffer_end(struct ust_buffer *buf, |
b5b073e2 PMF |
341 | u64 tsc, unsigned int offset, unsigned int subbuf_idx) |
342 | { | |
343 | struct ltt_subbuffer_header *header = | |
344 | (struct ltt_subbuffer_header *) | |
b73a4c47 | 345 | ust_buffers_offset_address(buf, |
b5b073e2 | 346 | subbuf_idx * buf->chan->subbuf_size); |
8c36d1ee | 347 | u32 data_size = SUBBUF_OFFSET(offset - 1, buf->chan) + 1; |
b5b073e2 | 348 | |
8c36d1ee | 349 | header->sb_size = PAGE_ALIGN(data_size); |
b5b073e2 | 350 | header->cycle_count_end = tsc; |
b102c2b0 PMF |
351 | header->events_lost = uatomic_read(&buf->events_lost); |
352 | header->subbuf_corrupt = uatomic_read(&buf->corrupted_subbuffers); | |
719569e4 PMF |
353 | if(unlikely(header->events_lost > 0)) { |
354 | DBG("Some events (%d) were lost in %s_%d", header->events_lost, buf->chan->channel_name, buf->cpu); | |
355 | } | |
c74e3560 MD |
356 | /* |
357 | * Makes sure data_size write happens after write of the rest of the | |
358 | * buffer end data, because data_size is used to identify a completely | |
359 | * written subbuffer in a crash dump. | |
360 | */ | |
361 | cmm_barrier(); | |
362 | header->data_size = data_size; | |
b5b073e2 PMF |
363 | } |
364 | ||
365 | /* | |
366 | * This function should not be called from NMI interrupt context | |
367 | */ | |
368 | static notrace void ltt_buf_unfull(struct ust_buffer *buf, | |
369 | unsigned int subbuf_idx, | |
370 | long offset) | |
371 | { | |
b5b073e2 PMF |
372 | } |
373 | ||
b73a4c47 | 374 | /* |
0222e121 | 375 | * Promote compiler cmm_barrier to a smp_mb(). |
b73a4c47 PMF |
376 | * For the specific LTTng case, this IPI call should be removed if the |
377 | * architecture does not reorder writes. This should eventually be provided by | |
378 | * a separate architecture-specific infrastructure. | |
379 | */ | |
e17571a5 PMF |
380 | //ust// static void remote_mb(void *info) |
381 | //ust// { | |
382 | //ust// smp_mb(); | |
383 | //ust// } | |
b73a4c47 PMF |
384 | |
385 | int ust_buffers_get_subbuf(struct ust_buffer *buf, long *consumed) | |
b5b073e2 PMF |
386 | { |
387 | struct ust_channel *channel = buf->chan; | |
388 | long consumed_old, consumed_idx, commit_count, write_offset; | |
b73a4c47 PMF |
389 | //ust// int retval; |
390 | ||
b102c2b0 | 391 | consumed_old = uatomic_read(&buf->consumed); |
b5b073e2 | 392 | consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan); |
b102c2b0 | 393 | commit_count = uatomic_read(&buf->commit_count[consumed_idx].cc_sb); |
b5b073e2 PMF |
394 | /* |
395 | * Make sure we read the commit count before reading the buffer | |
396 | * data and the write offset. Correct consumed offset ordering | |
397 | * wrt commit count is insured by the use of cmpxchg to update | |
398 | * the consumed offset. | |
b73a4c47 | 399 | */ |
d4d511d9 | 400 | |
b73a4c47 PMF |
401 | /* |
402 | * Local rmb to match the remote wmb to read the commit count before the | |
403 | * buffer data and the write offset. | |
b5b073e2 | 404 | */ |
0222e121 | 405 | cmm_smp_rmb(); |
b73a4c47 | 406 | |
b102c2b0 | 407 | write_offset = uatomic_read(&buf->offset); |
b5b073e2 PMF |
408 | /* |
409 | * Check that the subbuffer we are trying to consume has been | |
410 | * already fully committed. | |
411 | */ | |
412 | if (((commit_count - buf->chan->subbuf_size) | |
413 | & channel->commit_count_mask) | |
414 | - (BUFFER_TRUNC(consumed_old, buf->chan) | |
415 | >> channel->n_subbufs_order) | |
416 | != 0) { | |
417 | return -EAGAIN; | |
418 | } | |
419 | /* | |
420 | * Check that we are not about to read the same subbuffer in | |
421 | * which the writer head is. | |
422 | */ | |
423 | if ((SUBBUF_TRUNC(write_offset, buf->chan) | |
424 | - SUBBUF_TRUNC(consumed_old, buf->chan)) | |
425 | == 0) { | |
426 | return -EAGAIN; | |
427 | } | |
b73a4c47 PMF |
428 | *consumed = consumed_old; |
429 | ||
b5b073e2 PMF |
430 | return 0; |
431 | } | |
432 | ||
b73a4c47 | 433 | int ust_buffers_put_subbuf(struct ust_buffer *buf, unsigned long uconsumed_old) |
b5b073e2 PMF |
434 | { |
435 | long consumed_new, consumed_old; | |
436 | ||
b102c2b0 | 437 | consumed_old = uatomic_read(&buf->consumed); |
b5b073e2 PMF |
438 | consumed_old = consumed_old & (~0xFFFFFFFFL); |
439 | consumed_old = consumed_old | uconsumed_old; | |
440 | consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan); | |
441 | ||
b102c2b0 | 442 | if (uatomic_cmpxchg(&buf->consumed, consumed_old, |
b5b073e2 PMF |
443 | consumed_new) |
444 | != consumed_old) { | |
445 | /* We have been pushed by the writer : the last | |
446 | * buffer read _is_ corrupted! It can also | |
447 | * happen if this is a buffer we never got. */ | |
b5b073e2 PMF |
448 | return -EIO; |
449 | } else { | |
450 | /* tell the client that buffer is now unfull */ | |
451 | int index; | |
452 | long data; | |
453 | index = SUBBUF_INDEX(consumed_old, buf->chan); | |
454 | data = BUFFER_OFFSET(consumed_old, buf->chan); | |
455 | ltt_buf_unfull(buf, index, data); | |
b5b073e2 PMF |
456 | } |
457 | return 0; | |
458 | } | |
459 | ||
ef15e552 | 460 | static int map_buf_structs(struct ust_channel *chan) |
b5b073e2 PMF |
461 | { |
462 | void *ptr; | |
463 | int result; | |
204141ee PMF |
464 | size_t size; |
465 | int i; | |
b5b073e2 | 466 | |
204141ee | 467 | size = PAGE_ALIGN(1); |
b5b073e2 | 468 | |
204141ee | 469 | for(i=0; i<chan->n_cpus; i++) { |
b5b073e2 | 470 | |
204141ee PMF |
471 | result = chan->buf_struct_shmids[i] = shmget(getpid(), size, IPC_CREAT | IPC_EXCL | 0700); |
472 | if(result == -1) { | |
473 | PERROR("shmget"); | |
474 | goto destroy_previous; | |
475 | } | |
b5b073e2 | 476 | |
204141ee PMF |
477 | ptr = shmat(chan->buf_struct_shmids[i], NULL, 0); |
478 | if(ptr == (void *) -1) { | |
479 | perror("shmat"); | |
480 | goto destroy_shm; | |
481 | } | |
482 | ||
483 | /* Already mark the shared memory for destruction. This will occur only | |
484 | * when all users have detached. | |
485 | */ | |
486 | result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL); | |
487 | if(result == -1) { | |
488 | perror("shmctl"); | |
489 | goto destroy_previous; | |
490 | } | |
491 | ||
492 | chan->buf[i] = ptr; | |
b5b073e2 PMF |
493 | } |
494 | ||
204141ee | 495 | return 0; |
b5b073e2 | 496 | |
204141ee PMF |
497 | /* Jumping inside this loop occurs from within the other loop above with i as |
498 | * counter, so it unallocates the structures for the cpu = current_i down to | |
499 | * zero. */ | |
500 | for(; i>=0; i--) { | |
501 | destroy_shm: | |
502 | result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL); | |
503 | if(result == -1) { | |
504 | perror("shmctl"); | |
505 | } | |
b5b073e2 | 506 | |
204141ee PMF |
507 | destroy_previous: |
508 | continue; | |
b5b073e2 PMF |
509 | } |
510 | ||
204141ee | 511 | return -1; |
b5b073e2 PMF |
512 | } |
513 | ||
ef15e552 NC |
514 | static int unmap_buf_structs(struct ust_channel *chan) |
515 | { | |
516 | int i; | |
517 | ||
518 | for (i=0; i < chan->n_cpus; i++) { | |
519 | if (shmdt(chan->buf[i]) < 0) { | |
520 | PERROR("shmdt"); | |
521 | } | |
522 | } | |
fbae86d6 | 523 | return 0; |
ef15e552 NC |
524 | } |
525 | ||
b5b073e2 PMF |
526 | /* |
527 | * Create channel. | |
528 | */ | |
ef15e552 NC |
529 | static int create_channel(const char *trace_name, struct ust_trace *trace, |
530 | const char *channel_name, struct ust_channel *chan, | |
b5b073e2 PMF |
531 | unsigned int subbuf_size, unsigned int n_subbufs, int overwrite) |
532 | { | |
ef15e552 | 533 | int i, result; |
b5b073e2 | 534 | |
ef15e552 NC |
535 | chan->trace = trace; |
536 | chan->overwrite = overwrite; | |
537 | chan->n_subbufs_order = get_count_order(n_subbufs); | |
538 | chan->commit_count_mask = (~0UL >> chan->n_subbufs_order); | |
539 | chan->n_cpus = get_n_cpus(); | |
b5b073e2 | 540 | |
ef15e552 NC |
541 | /* These mappings should ideall be per-cpu, if somebody can do that |
542 | * from userspace, that would be cool! | |
543 | */ | |
544 | chan->buf = (void *) zmalloc(chan->n_cpus * sizeof(void *)); | |
545 | if(chan->buf == NULL) { | |
204141ee PMF |
546 | goto error; |
547 | } | |
ef15e552 NC |
548 | chan->buf_struct_shmids = (int *) zmalloc(chan->n_cpus * sizeof(int)); |
549 | if(chan->buf_struct_shmids == NULL) | |
204141ee | 550 | goto free_buf; |
b5b073e2 | 551 | |
ef15e552 | 552 | result = map_buf_structs(chan); |
204141ee PMF |
553 | if(result != 0) { |
554 | goto free_buf_struct_shmids; | |
555 | } | |
b5b073e2 | 556 | |
ef15e552 | 557 | result = open_channel(chan, subbuf_size, n_subbufs); |
204141ee | 558 | if (result != 0) { |
c1f20530 | 559 | ERR("Cannot open channel for trace %s", trace_name); |
ef15e552 | 560 | goto unmap_buf_structs; |
b5b073e2 PMF |
561 | } |
562 | ||
204141ee PMF |
563 | return 0; |
564 | ||
ef15e552 NC |
565 | unmap_buf_structs: |
566 | for (i=0; i < chan->n_cpus; i++) { | |
567 | if (shmdt(chan->buf[i]) < 0) { | |
568 | PERROR("shmdt bufstruct"); | |
569 | } | |
570 | } | |
204141ee PMF |
571 | |
572 | free_buf_struct_shmids: | |
ef15e552 | 573 | free(chan->buf_struct_shmids); |
b5b073e2 | 574 | |
204141ee | 575 | free_buf: |
ef15e552 | 576 | free(chan->buf); |
204141ee PMF |
577 | |
578 | error: | |
579 | return -1; | |
b5b073e2 PMF |
580 | } |
581 | ||
ef15e552 NC |
582 | |
583 | static void remove_channel(struct ust_channel *chan) | |
584 | { | |
585 | close_channel(chan); | |
586 | ||
587 | unmap_buf_structs(chan); | |
588 | ||
589 | free(chan->buf_struct_shmids); | |
590 | ||
591 | free(chan->buf); | |
ef15e552 NC |
592 | } |
593 | ||
b5b073e2 PMF |
594 | static void ltt_relay_async_wakeup_chan(struct ust_channel *ltt_channel) |
595 | { | |
b5b073e2 PMF |
596 | } |
597 | ||
204141ee | 598 | static void ltt_relay_finish_buffer(struct ust_channel *channel, unsigned int cpu) |
b5b073e2 | 599 | { |
204141ee PMF |
600 | if (channel->buf[cpu]) { |
601 | struct ust_buffer *buf = channel->buf[cpu]; | |
97c10252 | 602 | ltt_force_switch(buf, FORCE_FLUSH); |
ef15e552 | 603 | |
b5b073e2 | 604 | /* closing the pipe tells the consumer the buffer is finished */ |
b5b073e2 PMF |
605 | close(buf->data_ready_fd_write); |
606 | } | |
607 | } | |
608 | ||
609 | ||
ef15e552 | 610 | static void finish_channel(struct ust_channel *channel) |
b5b073e2 | 611 | { |
204141ee | 612 | unsigned int i; |
b5b073e2 | 613 | |
d4d511d9 | 614 | for (i=0; i<channel->n_cpus; i++) { |
204141ee PMF |
615 | ltt_relay_finish_buffer(channel, i); |
616 | } | |
b5b073e2 PMF |
617 | } |
618 | ||
b5b073e2 | 619 | |
b5b073e2 | 620 | /* |
b73a4c47 PMF |
621 | * ltt_reserve_switch_old_subbuf: switch old subbuffer |
622 | * | |
623 | * Concurrency safe because we are the last and only thread to alter this | |
624 | * sub-buffer. As long as it is not delivered and read, no other thread can | |
625 | * alter the offset, alter the reserve_count or call the | |
626 | * client_buffer_end_callback on this sub-buffer. | |
627 | * | |
628 | * The only remaining threads could be the ones with pending commits. They will | |
629 | * have to do the deliver themselves. Not concurrency safe in overwrite mode. | |
630 | * We detect corrupted subbuffers with commit and reserve counts. We keep a | |
631 | * corrupted sub-buffers count and push the readers across these sub-buffers. | |
632 | * | |
633 | * Not concurrency safe if a writer is stalled in a subbuffer and another writer | |
634 | * switches in, finding out it's corrupted. The result will be than the old | |
635 | * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer | |
636 | * will be declared corrupted too because of the commit count adjustment. | |
637 | * | |
638 | * Note : offset_old should never be 0 here. | |
b5b073e2 | 639 | */ |
b73a4c47 PMF |
640 | static void ltt_reserve_switch_old_subbuf( |
641 | struct ust_channel *chan, struct ust_buffer *buf, | |
642 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) | |
b5b073e2 | 643 | { |
b73a4c47 PMF |
644 | long oldidx = SUBBUF_INDEX(offsets->old - 1, chan); |
645 | long commit_count, padding_size; | |
b5b073e2 | 646 | |
b73a4c47 PMF |
647 | padding_size = chan->subbuf_size |
648 | - (SUBBUF_OFFSET(offsets->old - 1, chan) + 1); | |
649 | ltt_buffer_end(buf, *tsc, offsets->old, oldidx); | |
b5b073e2 | 650 | |
b73a4c47 PMF |
651 | /* |
652 | * Must write slot data before incrementing commit count. | |
ef15e552 | 653 | * This compiler barrier is upgraded into a cmm_smp_wmb() by the IPI |
0222e121 | 654 | * sent by get_subbuf() when it does its cmm_smp_rmb(). |
b73a4c47 | 655 | */ |
0222e121 | 656 | cmm_smp_wmb(); |
b102c2b0 PMF |
657 | uatomic_add(&buf->commit_count[oldidx].cc, padding_size); |
658 | commit_count = uatomic_read(&buf->commit_count[oldidx].cc); | |
b73a4c47 | 659 | ltt_check_deliver(chan, buf, offsets->old - 1, commit_count, oldidx); |
1e8c9e7b | 660 | ltt_write_commit_counter(chan, buf, oldidx, |
b73a4c47 PMF |
661 | offsets->old, commit_count, padding_size); |
662 | } | |
b5b073e2 | 663 | |
b73a4c47 PMF |
664 | /* |
665 | * ltt_reserve_switch_new_subbuf: Populate new subbuffer. | |
666 | * | |
667 | * This code can be executed unordered : writers may already have written to the | |
668 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
669 | * that this code is executed before the deliver of this sub-buffer. | |
670 | */ | |
671 | static void ltt_reserve_switch_new_subbuf( | |
672 | struct ust_channel *chan, struct ust_buffer *buf, | |
673 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) | |
674 | { | |
675 | long beginidx = SUBBUF_INDEX(offsets->begin, chan); | |
676 | long commit_count; | |
b5b073e2 | 677 | |
b73a4c47 | 678 | ltt_buffer_begin(buf, *tsc, beginidx); |
b5b073e2 | 679 | |
b73a4c47 PMF |
680 | /* |
681 | * Must write slot data before incrementing commit count. | |
ef15e552 | 682 | * This compiler barrier is upgraded into a cmm_smp_wmb() by the IPI |
0222e121 | 683 | * sent by get_subbuf() when it does its cmm_smp_rmb(). |
b73a4c47 | 684 | */ |
0222e121 | 685 | cmm_smp_wmb(); |
b102c2b0 PMF |
686 | uatomic_add(&buf->commit_count[beginidx].cc, ltt_subbuffer_header_size()); |
687 | commit_count = uatomic_read(&buf->commit_count[beginidx].cc); | |
b73a4c47 PMF |
688 | /* Check if the written buffer has to be delivered */ |
689 | ltt_check_deliver(chan, buf, offsets->begin, commit_count, beginidx); | |
1e8c9e7b | 690 | ltt_write_commit_counter(chan, buf, beginidx, |
b73a4c47 PMF |
691 | offsets->begin, commit_count, ltt_subbuffer_header_size()); |
692 | } | |
b5b073e2 | 693 | |
b73a4c47 PMF |
694 | /* |
695 | * ltt_reserve_end_switch_current: finish switching current subbuffer | |
696 | * | |
697 | * Concurrency safe because we are the last and only thread to alter this | |
698 | * sub-buffer. As long as it is not delivered and read, no other thread can | |
699 | * alter the offset, alter the reserve_count or call the | |
700 | * client_buffer_end_callback on this sub-buffer. | |
701 | * | |
702 | * The only remaining threads could be the ones with pending commits. They will | |
703 | * have to do the deliver themselves. Not concurrency safe in overwrite mode. | |
704 | * We detect corrupted subbuffers with commit and reserve counts. We keep a | |
705 | * corrupted sub-buffers count and push the readers across these sub-buffers. | |
706 | * | |
707 | * Not concurrency safe if a writer is stalled in a subbuffer and another writer | |
708 | * switches in, finding out it's corrupted. The result will be than the old | |
709 | * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer | |
710 | * will be declared corrupted too because of the commit count adjustment. | |
711 | */ | |
712 | static void ltt_reserve_end_switch_current( | |
713 | struct ust_channel *chan, | |
714 | struct ust_buffer *buf, | |
715 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) | |
716 | { | |
717 | long endidx = SUBBUF_INDEX(offsets->end - 1, chan); | |
718 | long commit_count, padding_size; | |
719 | ||
720 | padding_size = chan->subbuf_size | |
721 | - (SUBBUF_OFFSET(offsets->end - 1, chan) + 1); | |
722 | ||
723 | ltt_buffer_end(buf, *tsc, offsets->end, endidx); | |
724 | ||
725 | /* | |
726 | * Must write slot data before incrementing commit count. | |
ef15e552 | 727 | * This compiler barrier is upgraded into a cmm_smp_wmb() by the IPI |
0222e121 | 728 | * sent by get_subbuf() when it does its cmm_smp_rmb(). |
b73a4c47 | 729 | */ |
0222e121 | 730 | cmm_smp_wmb(); |
b102c2b0 PMF |
731 | uatomic_add(&buf->commit_count[endidx].cc, padding_size); |
732 | commit_count = uatomic_read(&buf->commit_count[endidx].cc); | |
b73a4c47 PMF |
733 | ltt_check_deliver(chan, buf, |
734 | offsets->end - 1, commit_count, endidx); | |
1e8c9e7b | 735 | ltt_write_commit_counter(chan, buf, endidx, |
b73a4c47 | 736 | offsets->end, commit_count, padding_size); |
b5b073e2 PMF |
737 | } |
738 | ||
739 | /* | |
740 | * Returns : | |
741 | * 0 if ok | |
742 | * !0 if execution must be aborted. | |
743 | */ | |
b73a4c47 | 744 | static int ltt_relay_try_switch_slow( |
b5b073e2 | 745 | enum force_switch_mode mode, |
b73a4c47 | 746 | struct ust_channel *chan, |
b5b073e2 PMF |
747 | struct ust_buffer *buf, |
748 | struct ltt_reserve_switch_offsets *offsets, | |
749 | u64 *tsc) | |
750 | { | |
751 | long subbuf_index; | |
b73a4c47 | 752 | long reserve_commit_diff; |
b5b073e2 | 753 | |
b102c2b0 | 754 | offsets->begin = uatomic_read(&buf->offset); |
b5b073e2 PMF |
755 | offsets->old = offsets->begin; |
756 | offsets->begin_switch = 0; | |
757 | offsets->end_switch_old = 0; | |
758 | ||
759 | *tsc = trace_clock_read64(); | |
760 | ||
761 | if (SUBBUF_OFFSET(offsets->begin, buf->chan) != 0) { | |
762 | offsets->begin = SUBBUF_ALIGN(offsets->begin, buf->chan); | |
763 | offsets->end_switch_old = 1; | |
764 | } else { | |
765 | /* we do not have to switch : buffer is empty */ | |
766 | return -1; | |
767 | } | |
768 | if (mode == FORCE_ACTIVE) | |
769 | offsets->begin += ltt_subbuffer_header_size(); | |
770 | /* | |
771 | * Always begin_switch in FORCE_ACTIVE mode. | |
772 | * Test new buffer integrity | |
773 | */ | |
774 | subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); | |
b73a4c47 | 775 | reserve_commit_diff = |
b5b073e2 | 776 | (BUFFER_TRUNC(offsets->begin, buf->chan) |
b73a4c47 | 777 | >> chan->n_subbufs_order) |
b102c2b0 | 778 | - (uatomic_read(&buf->commit_count[subbuf_index].cc_sb) |
b73a4c47 PMF |
779 | & chan->commit_count_mask); |
780 | if (reserve_commit_diff == 0) { | |
b5b073e2 PMF |
781 | /* Next buffer not corrupted. */ |
782 | if (mode == FORCE_ACTIVE | |
b73a4c47 | 783 | && !chan->overwrite |
b102c2b0 | 784 | && offsets->begin - uatomic_read(&buf->consumed) |
b73a4c47 | 785 | >= chan->alloc_size) { |
b5b073e2 PMF |
786 | /* |
787 | * We do not overwrite non consumed buffers and we are | |
788 | * full : ignore switch while tracing is active. | |
789 | */ | |
790 | return -1; | |
791 | } | |
792 | } else { | |
793 | /* | |
794 | * Next subbuffer corrupted. Force pushing reader even in normal | |
795 | * mode | |
796 | */ | |
797 | } | |
798 | offsets->end = offsets->begin; | |
799 | return 0; | |
800 | } | |
801 | ||
b5b073e2 | 802 | /* |
b73a4c47 PMF |
803 | * Force a sub-buffer switch for a per-cpu buffer. This operation is |
804 | * completely reentrant : can be called while tracing is active with | |
805 | * absolutely no lock held. | |
b5b073e2 | 806 | */ |
b73a4c47 PMF |
807 | void ltt_force_switch_lockless_slow(struct ust_buffer *buf, |
808 | enum force_switch_mode mode) | |
b5b073e2 | 809 | { |
b73a4c47 | 810 | struct ust_channel *chan = buf->chan; |
b5b073e2 | 811 | struct ltt_reserve_switch_offsets offsets; |
b73a4c47 | 812 | u64 tsc; |
b5b073e2 | 813 | |
b5b073e2 PMF |
814 | offsets.size = 0; |
815 | ||
10dd3941 | 816 | DBG("Switching (forced) %s_%d", chan->channel_name, buf->cpu); |
b5b073e2 PMF |
817 | /* |
818 | * Perform retryable operations. | |
819 | */ | |
b5b073e2 | 820 | do { |
b73a4c47 PMF |
821 | if (ltt_relay_try_switch_slow(mode, chan, buf, |
822 | &offsets, &tsc)) | |
823 | return; | |
b102c2b0 | 824 | } while (uatomic_cmpxchg(&buf->offset, offsets.old, |
b5b073e2 PMF |
825 | offsets.end) != offsets.old); |
826 | ||
827 | /* | |
828 | * Atomically update last_tsc. This update races against concurrent | |
829 | * atomic updates, but the race will always cause supplementary full TSC | |
830 | * events, never the opposite (missing a full TSC event when it would be | |
831 | * needed). | |
832 | */ | |
b73a4c47 | 833 | save_last_tsc(buf, tsc); |
b5b073e2 PMF |
834 | |
835 | /* | |
836 | * Push the reader if necessary | |
837 | */ | |
b73a4c47 PMF |
838 | if (mode == FORCE_ACTIVE) { |
839 | ltt_reserve_push_reader(chan, buf, offsets.end - 1); | |
b73a4c47 | 840 | } |
b5b073e2 PMF |
841 | |
842 | /* | |
843 | * Switch old subbuffer if needed. | |
844 | */ | |
b73a4c47 | 845 | if (offsets.end_switch_old) { |
b73a4c47 PMF |
846 | ltt_reserve_switch_old_subbuf(chan, buf, &offsets, &tsc); |
847 | } | |
b5b073e2 PMF |
848 | |
849 | /* | |
850 | * Populate new subbuffer. | |
851 | */ | |
b73a4c47 PMF |
852 | if (mode == FORCE_ACTIVE) |
853 | ltt_reserve_switch_new_subbuf(chan, buf, &offsets, &tsc); | |
854 | } | |
b5b073e2 | 855 | |
b73a4c47 PMF |
856 | /* |
857 | * Returns : | |
858 | * 0 if ok | |
859 | * !0 if execution must be aborted. | |
860 | */ | |
861 | static int ltt_relay_try_reserve_slow(struct ust_channel *chan, struct ust_buffer *buf, | |
862 | struct ltt_reserve_switch_offsets *offsets, size_t data_size, | |
863 | u64 *tsc, unsigned int *rflags, int largest_align) | |
864 | { | |
865 | long reserve_commit_diff; | |
b5b073e2 | 866 | |
b102c2b0 | 867 | offsets->begin = uatomic_read(&buf->offset); |
b73a4c47 PMF |
868 | offsets->old = offsets->begin; |
869 | offsets->begin_switch = 0; | |
870 | offsets->end_switch_current = 0; | |
871 | offsets->end_switch_old = 0; | |
872 | ||
873 | *tsc = trace_clock_read64(); | |
874 | if (last_tsc_overflow(buf, *tsc)) | |
875 | *rflags = LTT_RFLAG_ID_SIZE_TSC; | |
876 | ||
877 | if (unlikely(SUBBUF_OFFSET(offsets->begin, buf->chan) == 0)) { | |
878 | offsets->begin_switch = 1; /* For offsets->begin */ | |
879 | } else { | |
880 | offsets->size = ust_get_header_size(chan, | |
881 | offsets->begin, data_size, | |
882 | &offsets->before_hdr_pad, *rflags); | |
883 | offsets->size += ltt_align(offsets->begin + offsets->size, | |
884 | largest_align) | |
885 | + data_size; | |
886 | if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan) + | |
887 | offsets->size) > buf->chan->subbuf_size)) { | |
888 | offsets->end_switch_old = 1; /* For offsets->old */ | |
889 | offsets->begin_switch = 1; /* For offsets->begin */ | |
890 | } | |
891 | } | |
892 | if (unlikely(offsets->begin_switch)) { | |
893 | long subbuf_index; | |
894 | ||
895 | /* | |
896 | * We are typically not filling the previous buffer completely. | |
897 | */ | |
898 | if (likely(offsets->end_switch_old)) | |
899 | offsets->begin = SUBBUF_ALIGN(offsets->begin, | |
900 | buf->chan); | |
901 | offsets->begin = offsets->begin + ltt_subbuffer_header_size(); | |
902 | /* Test new buffer integrity */ | |
903 | subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); | |
904 | reserve_commit_diff = | |
905 | (BUFFER_TRUNC(offsets->begin, buf->chan) | |
906 | >> chan->n_subbufs_order) | |
b102c2b0 | 907 | - (uatomic_read(&buf->commit_count[subbuf_index].cc_sb) |
b73a4c47 PMF |
908 | & chan->commit_count_mask); |
909 | if (likely(reserve_commit_diff == 0)) { | |
910 | /* Next buffer not corrupted. */ | |
911 | if (unlikely(!chan->overwrite && | |
912 | (SUBBUF_TRUNC(offsets->begin, buf->chan) | |
b102c2b0 | 913 | - SUBBUF_TRUNC(uatomic_read( |
b73a4c47 PMF |
914 | &buf->consumed), |
915 | buf->chan)) | |
916 | >= chan->alloc_size)) { | |
917 | /* | |
918 | * We do not overwrite non consumed buffers | |
919 | * and we are full : event is lost. | |
920 | */ | |
b102c2b0 | 921 | uatomic_inc(&buf->events_lost); |
b73a4c47 PMF |
922 | return -1; |
923 | } else { | |
924 | /* | |
925 | * next buffer not corrupted, we are either in | |
926 | * overwrite mode or the buffer is not full. | |
927 | * It's safe to write in this new subbuffer. | |
928 | */ | |
929 | } | |
930 | } else { | |
931 | /* | |
932 | * Next subbuffer corrupted. Drop event in normal and | |
933 | * overwrite mode. Caused by either a writer OOPS or | |
934 | * too many nested writes over a reserve/commit pair. | |
935 | */ | |
b102c2b0 | 936 | uatomic_inc(&buf->events_lost); |
b73a4c47 PMF |
937 | return -1; |
938 | } | |
939 | offsets->size = ust_get_header_size(chan, | |
940 | offsets->begin, data_size, | |
941 | &offsets->before_hdr_pad, *rflags); | |
942 | offsets->size += ltt_align(offsets->begin + offsets->size, | |
943 | largest_align) | |
944 | + data_size; | |
945 | if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan) | |
946 | + offsets->size) > buf->chan->subbuf_size)) { | |
947 | /* | |
948 | * Event too big for subbuffers, report error, don't | |
949 | * complete the sub-buffer switch. | |
950 | */ | |
b102c2b0 | 951 | uatomic_inc(&buf->events_lost); |
b73a4c47 PMF |
952 | return -1; |
953 | } else { | |
954 | /* | |
955 | * We just made a successful buffer switch and the event | |
956 | * fits in the new subbuffer. Let's write. | |
957 | */ | |
958 | } | |
959 | } else { | |
960 | /* | |
961 | * Event fits in the current buffer and we are not on a switch | |
962 | * boundary. It's safe to write. | |
963 | */ | |
964 | } | |
965 | offsets->end = offsets->begin + offsets->size; | |
966 | ||
967 | if (unlikely((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0)) { | |
968 | /* | |
969 | * The offset_end will fall at the very beginning of the next | |
970 | * subbuffer. | |
971 | */ | |
972 | offsets->end_switch_current = 1; /* For offsets->begin */ | |
973 | } | |
b5b073e2 PMF |
974 | return 0; |
975 | } | |
976 | ||
b73a4c47 PMF |
977 | /** |
978 | * ltt_relay_reserve_slot_lockless_slow - Atomic slot reservation in a buffer. | |
979 | * @trace: the trace structure to log to. | |
980 | * @ltt_channel: channel structure | |
981 | * @transport_data: data structure specific to ltt relay | |
982 | * @data_size: size of the variable length data to log. | |
983 | * @slot_size: pointer to total size of the slot (out) | |
984 | * @buf_offset : pointer to reserved buffer offset (out) | |
985 | * @tsc: pointer to the tsc at the slot reservation (out) | |
986 | * @cpu: cpuid | |
b5b073e2 | 987 | * |
b73a4c47 PMF |
988 | * Return : -ENOSPC if not enough space, else returns 0. |
989 | * It will take care of sub-buffer switching. | |
b5b073e2 | 990 | */ |
12e81b07 PMF |
991 | int ltt_reserve_slot_lockless_slow(struct ust_channel *chan, |
992 | struct ust_trace *trace, size_t data_size, | |
993 | int largest_align, int cpu, | |
994 | struct ust_buffer **ret_buf, | |
995 | size_t *slot_size, long *buf_offset, | |
996 | u64 *tsc, unsigned int *rflags) | |
b5b073e2 | 997 | { |
12e81b07 | 998 | struct ust_buffer *buf = *ret_buf = chan->buf[cpu]; |
b5b073e2 | 999 | struct ltt_reserve_switch_offsets offsets; |
b5b073e2 | 1000 | |
b5b073e2 PMF |
1001 | offsets.size = 0; |
1002 | ||
b5b073e2 | 1003 | do { |
b73a4c47 PMF |
1004 | if (unlikely(ltt_relay_try_reserve_slow(chan, buf, &offsets, |
1005 | data_size, tsc, rflags, largest_align))) | |
1006 | return -ENOSPC; | |
b102c2b0 | 1007 | } while (unlikely(uatomic_cmpxchg(&buf->offset, offsets.old, |
b73a4c47 | 1008 | offsets.end) != offsets.old)); |
b5b073e2 PMF |
1009 | |
1010 | /* | |
1011 | * Atomically update last_tsc. This update races against concurrent | |
1012 | * atomic updates, but the race will always cause supplementary full TSC | |
1013 | * events, never the opposite (missing a full TSC event when it would be | |
1014 | * needed). | |
1015 | */ | |
b73a4c47 | 1016 | save_last_tsc(buf, *tsc); |
b5b073e2 PMF |
1017 | |
1018 | /* | |
1019 | * Push the reader if necessary | |
1020 | */ | |
b73a4c47 PMF |
1021 | ltt_reserve_push_reader(chan, buf, offsets.end - 1); |
1022 | ||
b5b073e2 PMF |
1023 | /* |
1024 | * Switch old subbuffer if needed. | |
1025 | */ | |
b73a4c47 | 1026 | if (unlikely(offsets.end_switch_old)) { |
b73a4c47 | 1027 | ltt_reserve_switch_old_subbuf(chan, buf, &offsets, tsc); |
10dd3941 | 1028 | DBG("Switching %s_%d", chan->channel_name, cpu); |
b73a4c47 | 1029 | } |
b5b073e2 PMF |
1030 | |
1031 | /* | |
1032 | * Populate new subbuffer. | |
1033 | */ | |
b73a4c47 PMF |
1034 | if (unlikely(offsets.begin_switch)) |
1035 | ltt_reserve_switch_new_subbuf(chan, buf, &offsets, tsc); | |
1036 | ||
1037 | if (unlikely(offsets.end_switch_current)) | |
1038 | ltt_reserve_end_switch_current(chan, buf, &offsets, tsc); | |
1039 | ||
1040 | *slot_size = offsets.size; | |
1041 | *buf_offset = offsets.begin + offsets.before_hdr_pad; | |
1042 | return 0; | |
b5b073e2 PMF |
1043 | } |
1044 | ||
b5b073e2 PMF |
1045 | static struct ltt_transport ust_relay_transport = { |
1046 | .name = "ustrelay", | |
1047 | .ops = { | |
ef15e552 NC |
1048 | .create_channel = create_channel, |
1049 | .finish_channel = finish_channel, | |
1050 | .remove_channel = remove_channel, | |
b5b073e2 | 1051 | .wakeup_channel = ltt_relay_async_wakeup_chan, |
b5b073e2 PMF |
1052 | }, |
1053 | }; | |
1054 | ||
b5b073e2 PMF |
1055 | static char initialized = 0; |
1056 | ||
1057 | void __attribute__((constructor)) init_ustrelay_transport(void) | |
1058 | { | |
5e96a467 | 1059 | init_usterr(); |
b5b073e2 PMF |
1060 | if(!initialized) { |
1061 | ltt_transport_register(&ust_relay_transport); | |
1062 | initialized = 1; | |
1063 | } | |
1064 | } | |
1065 | ||
b73a4c47 | 1066 | static void __attribute__((destructor)) ust_buffers_exit(void) |
b5b073e2 PMF |
1067 | { |
1068 | ltt_transport_unregister(&ust_relay_transport); | |
1069 | } | |
b73a4c47 | 1070 | |
12e81b07 | 1071 | size_t ltt_write_event_header_slow(struct ust_channel *channel, |
b73a4c47 PMF |
1072 | struct ust_buffer *buf, long buf_offset, |
1073 | u16 eID, u32 event_size, | |
1074 | u64 tsc, unsigned int rflags) | |
1075 | { | |
1076 | struct ltt_event_header header; | |
1077 | u16 small_size; | |
1078 | ||
1079 | switch (rflags) { | |
1080 | case LTT_RFLAG_ID_SIZE_TSC: | |
1081 | header.id_time = 29 << LTT_TSC_BITS; | |
1082 | break; | |
1083 | case LTT_RFLAG_ID_SIZE: | |
1084 | header.id_time = 30 << LTT_TSC_BITS; | |
1085 | break; | |
1086 | case LTT_RFLAG_ID: | |
1087 | header.id_time = 31 << LTT_TSC_BITS; | |
1088 | break; | |
e2b46575 DG |
1089 | default: |
1090 | WARN_ON_ONCE(1); | |
1091 | header.id_time = 0; | |
1092 | break; | |
b73a4c47 PMF |
1093 | } |
1094 | ||
1095 | header.id_time |= (u32)tsc & LTT_TSC_MASK; | |
1096 | ust_buffers_write(buf, buf_offset, &header, sizeof(header)); | |
1097 | buf_offset += sizeof(header); | |
1098 | ||
1099 | switch (rflags) { | |
1100 | case LTT_RFLAG_ID_SIZE_TSC: | |
1101 | small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE); | |
1102 | ust_buffers_write(buf, buf_offset, | |
1103 | &eID, sizeof(u16)); | |
1104 | buf_offset += sizeof(u16); | |
1105 | ust_buffers_write(buf, buf_offset, | |
1106 | &small_size, sizeof(u16)); | |
1107 | buf_offset += sizeof(u16); | |
1108 | if (small_size == LTT_MAX_SMALL_SIZE) { | |
1109 | ust_buffers_write(buf, buf_offset, | |
1110 | &event_size, sizeof(u32)); | |
1111 | buf_offset += sizeof(u32); | |
1112 | } | |
1113 | buf_offset += ltt_align(buf_offset, sizeof(u64)); | |
1114 | ust_buffers_write(buf, buf_offset, | |
1115 | &tsc, sizeof(u64)); | |
1116 | buf_offset += sizeof(u64); | |
1117 | break; | |
1118 | case LTT_RFLAG_ID_SIZE: | |
1119 | small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE); | |
1120 | ust_buffers_write(buf, buf_offset, | |
1121 | &eID, sizeof(u16)); | |
1122 | buf_offset += sizeof(u16); | |
1123 | ust_buffers_write(buf, buf_offset, | |
1124 | &small_size, sizeof(u16)); | |
1125 | buf_offset += sizeof(u16); | |
1126 | if (small_size == LTT_MAX_SMALL_SIZE) { | |
1127 | ust_buffers_write(buf, buf_offset, | |
1128 | &event_size, sizeof(u32)); | |
1129 | buf_offset += sizeof(u32); | |
1130 | } | |
1131 | break; | |
1132 | case LTT_RFLAG_ID: | |
1133 | ust_buffers_write(buf, buf_offset, | |
1134 | &eID, sizeof(u16)); | |
1135 | buf_offset += sizeof(u16); | |
1136 | break; | |
1137 | } | |
1138 | ||
1139 | return buf_offset; | |
1140 | } |