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) | |
6 | * Copyright (C) 2008 - Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca) | |
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 | |
10 | * License as published by the Free Software Foundation; either | |
11 | * version 2.1 of the License, or (at your option) any later version. | |
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 | ||
204141ee | 23 | #include <unistd.h> |
b5b073e2 PMF |
24 | #include <sys/mman.h> |
25 | #include <sys/ipc.h> | |
26 | #include <sys/shm.h> | |
27 | #include <fcntl.h> | |
909bc43f | 28 | #include <stdlib.h> |
518d7abb PMF |
29 | |
30 | #include <ust/clock.h> | |
31 | ||
b5b073e2 PMF |
32 | #include "buffers.h" |
33 | #include "channels.h" | |
34 | #include "tracer.h" | |
35 | #include "tracercore.h" | |
36 | #include "usterr.h" | |
37 | ||
b73a4c47 PMF |
38 | struct ltt_reserve_switch_offsets { |
39 | long begin, end, old; | |
40 | long begin_switch, end_switch_current, end_switch_old; | |
41 | size_t before_hdr_pad, size; | |
42 | }; | |
43 | ||
44 | ||
b5b073e2 PMF |
45 | static DEFINE_MUTEX(ust_buffers_channels_mutex); |
46 | static LIST_HEAD(ust_buffers_channels); | |
47 | ||
204141ee PMF |
48 | static int get_n_cpus(void) |
49 | { | |
50 | int result; | |
51 | static int n_cpus = 0; | |
52 | ||
c7dc133c PMF |
53 | if(!n_cpus) { |
54 | /* On Linux, when some processors are offline | |
55 | * _SC_NPROCESSORS_CONF counts the offline | |
56 | * processors, whereas _SC_NPROCESSORS_ONLN | |
57 | * does not. If we used _SC_NPROCESSORS_ONLN, | |
58 | * getcpu() could return a value greater than | |
59 | * this sysconf, in which case the arrays | |
60 | * indexed by processor would overflow. | |
61 | */ | |
62 | result = sysconf(_SC_NPROCESSORS_CONF); | |
63 | if(result == -1) { | |
64 | return -1; | |
65 | } | |
66 | ||
67 | n_cpus = result; | |
204141ee PMF |
68 | } |
69 | ||
c7dc133c | 70 | return n_cpus; |
204141ee PMF |
71 | } |
72 | ||
b73a4c47 PMF |
73 | /* _ust_buffers_write() |
74 | * | |
75 | * @buf: destination buffer | |
76 | * @offset: offset in destination | |
77 | * @src: source buffer | |
78 | * @len: length of source | |
79 | * @cpy: already copied | |
80 | */ | |
81 | ||
82 | void _ust_buffers_write(struct ust_buffer *buf, size_t offset, | |
83 | const void *src, size_t len, ssize_t cpy) | |
84 | { | |
85 | do { | |
86 | len -= cpy; | |
87 | src += cpy; | |
88 | offset += cpy; | |
89 | ||
90 | WARN_ON(offset >= buf->buf_size); | |
91 | ||
92 | cpy = min_t(size_t, len, buf->buf_size - offset); | |
93 | ust_buffers_do_copy(buf->buf_data + offset, src, cpy); | |
94 | } while (unlikely(len != cpy)); | |
95 | } | |
96 | ||
97 | static int ust_buffers_init_buffer(struct ust_trace *trace, | |
b5b073e2 PMF |
98 | struct ust_channel *ltt_chan, |
99 | struct ust_buffer *buf, | |
100 | unsigned int n_subbufs); | |
101 | ||
102 | static int ust_buffers_alloc_buf(struct ust_buffer *buf, size_t *size) | |
103 | { | |
104 | void *ptr; | |
105 | int result; | |
106 | ||
107 | *size = PAGE_ALIGN(*size); | |
108 | ||
109 | result = buf->shmid = shmget(getpid(), *size, IPC_CREAT | IPC_EXCL | 0700); | |
110 | if(result == -1 && errno == EINVAL) { | |
111 | ERR("shmget() returned EINVAL; maybe /proc/sys/kernel/shmmax should be increased."); | |
112 | return -1; | |
113 | } | |
114 | else if(result == -1) { | |
115 | PERROR("shmget"); | |
116 | return -1; | |
117 | } | |
118 | ||
204141ee | 119 | /* FIXME: should have matching call to shmdt */ |
b5b073e2 PMF |
120 | ptr = shmat(buf->shmid, NULL, 0); |
121 | if(ptr == (void *) -1) { | |
122 | perror("shmat"); | |
123 | goto destroy_shmem; | |
124 | } | |
125 | ||
126 | /* Already mark the shared memory for destruction. This will occur only | |
127 | * when all users have detached. | |
128 | */ | |
129 | result = shmctl(buf->shmid, IPC_RMID, NULL); | |
130 | if(result == -1) { | |
131 | perror("shmctl"); | |
132 | return -1; | |
133 | } | |
134 | ||
135 | buf->buf_data = ptr; | |
136 | buf->buf_size = *size; | |
137 | ||
138 | return 0; | |
139 | ||
140 | destroy_shmem: | |
141 | result = shmctl(buf->shmid, IPC_RMID, NULL); | |
142 | if(result == -1) { | |
143 | perror("shmctl"); | |
144 | } | |
145 | ||
146 | return -1; | |
147 | } | |
148 | ||
204141ee | 149 | int ust_buffers_create_buf(struct ust_channel *channel, int cpu) |
b5b073e2 PMF |
150 | { |
151 | int result; | |
204141ee | 152 | struct ust_buffer *buf = channel->buf[cpu]; |
b5b073e2 | 153 | |
204141ee PMF |
154 | buf->cpu = cpu; |
155 | result = ust_buffers_alloc_buf(buf, &channel->alloc_size); | |
b5b073e2 | 156 | if(result) |
204141ee | 157 | return -1; |
b5b073e2 | 158 | |
204141ee | 159 | buf->chan = channel; |
b5b073e2 | 160 | kref_get(&channel->kref); |
204141ee | 161 | return 0; |
b5b073e2 PMF |
162 | } |
163 | ||
164 | static void ust_buffers_destroy_channel(struct kref *kref) | |
165 | { | |
166 | struct ust_channel *chan = container_of(kref, struct ust_channel, kref); | |
167 | free(chan); | |
168 | } | |
169 | ||
170 | static void ust_buffers_destroy_buf(struct ust_buffer *buf) | |
171 | { | |
172 | struct ust_channel *chan = buf->chan; | |
173 | int result; | |
174 | ||
175 | result = munmap(buf->buf_data, buf->buf_size); | |
176 | if(result == -1) { | |
177 | PERROR("munmap"); | |
178 | } | |
179 | ||
204141ee | 180 | //ust// chan->buf[buf->cpu] = NULL; |
b5b073e2 PMF |
181 | free(buf); |
182 | kref_put(&chan->kref, ust_buffers_destroy_channel); | |
183 | } | |
184 | ||
185 | /* called from kref_put */ | |
186 | static void ust_buffers_remove_buf(struct kref *kref) | |
187 | { | |
188 | struct ust_buffer *buf = container_of(kref, struct ust_buffer, kref); | |
189 | ust_buffers_destroy_buf(buf); | |
190 | } | |
191 | ||
204141ee | 192 | int ust_buffers_open_buf(struct ust_channel *chan, int cpu) |
b5b073e2 | 193 | { |
204141ee | 194 | int result; |
b5b073e2 | 195 | |
204141ee PMF |
196 | result = ust_buffers_create_buf(chan, cpu); |
197 | if (result == -1) | |
198 | return -1; | |
b5b073e2 | 199 | |
204141ee | 200 | kref_init(&chan->buf[cpu]->kref); |
b5b073e2 | 201 | |
204141ee PMF |
202 | result = ust_buffers_init_buffer(chan->trace, chan, chan->buf[cpu], chan->subbuf_cnt); |
203 | if(result == -1) | |
204 | return -1; | |
b5b073e2 | 205 | |
204141ee | 206 | return 0; |
b5b073e2 PMF |
207 | |
208 | /* FIXME: decrementally destroy on error? */ | |
209 | } | |
210 | ||
211 | /** | |
212 | * ust_buffers_close_buf - close a channel buffer | |
213 | * @buf: buffer | |
214 | */ | |
215 | static void ust_buffers_close_buf(struct ust_buffer *buf) | |
216 | { | |
217 | kref_put(&buf->kref, ust_buffers_remove_buf); | |
218 | } | |
219 | ||
220 | int ust_buffers_channel_open(struct ust_channel *chan, size_t subbuf_size, size_t subbuf_cnt) | |
221 | { | |
204141ee PMF |
222 | int i; |
223 | int result; | |
224 | ||
b5b073e2 PMF |
225 | if(subbuf_size == 0 || subbuf_cnt == 0) |
226 | return -1; | |
227 | ||
b73a4c47 PMF |
228 | /* Check that the subbuffer size is larger than a page. */ |
229 | WARN_ON_ONCE(subbuf_size < PAGE_SIZE); | |
230 | ||
231 | /* | |
232 | * Make sure the number of subbuffers and subbuffer size are power of 2. | |
233 | */ | |
234 | WARN_ON_ONCE(hweight32(subbuf_size) != 1); | |
235 | WARN_ON(hweight32(subbuf_cnt) != 1); | |
236 | ||
b5b073e2 PMF |
237 | chan->version = UST_CHANNEL_VERSION; |
238 | chan->subbuf_cnt = subbuf_cnt; | |
239 | chan->subbuf_size = subbuf_size; | |
240 | chan->subbuf_size_order = get_count_order(subbuf_size); | |
b73a4c47 | 241 | chan->alloc_size = subbuf_size * subbuf_cnt; |
204141ee | 242 | |
b5b073e2 PMF |
243 | kref_init(&chan->kref); |
244 | ||
f7b16408 | 245 | pthread_mutex_lock(&ust_buffers_channels_mutex); |
204141ee PMF |
246 | for(i=0; i<chan->n_cpus; i++) { |
247 | result = ust_buffers_open_buf(chan, i); | |
248 | if (result == -1) | |
249 | goto error; | |
250 | } | |
b5b073e2 | 251 | list_add(&chan->list, &ust_buffers_channels); |
f7b16408 | 252 | pthread_mutex_unlock(&ust_buffers_channels_mutex); |
b5b073e2 PMF |
253 | |
254 | return 0; | |
255 | ||
204141ee PMF |
256 | /* Jump directly inside the loop to close the buffers that were already |
257 | * opened. */ | |
258 | for(; i>=0; i--) { | |
259 | ust_buffers_close_buf(chan->buf[i]); | |
260 | error: | |
120b0ec3 | 261 | do {} while(0); |
204141ee PMF |
262 | } |
263 | ||
b5b073e2 | 264 | kref_put(&chan->kref, ust_buffers_destroy_channel); |
f7b16408 | 265 | pthread_mutex_unlock(&ust_buffers_channels_mutex); |
b5b073e2 PMF |
266 | return -1; |
267 | } | |
268 | ||
269 | void ust_buffers_channel_close(struct ust_channel *chan) | |
270 | { | |
204141ee PMF |
271 | int i; |
272 | if(!chan) | |
b5b073e2 PMF |
273 | return; |
274 | ||
f7b16408 | 275 | pthread_mutex_lock(&ust_buffers_channels_mutex); |
204141ee PMF |
276 | for(i=0; i<chan->n_cpus; i++) { |
277 | /* FIXME: if we make it here, then all buffers were necessarily allocated. Moreover, we don't | |
278 | * initialize to NULL so we cannot use this check. Should we? */ | |
279 | //ust// if (chan->buf[i]) | |
280 | ust_buffers_close_buf(chan->buf[i]); | |
281 | } | |
b5b073e2 PMF |
282 | |
283 | list_del(&chan->list); | |
284 | kref_put(&chan->kref, ust_buffers_destroy_channel); | |
f7b16408 | 285 | pthread_mutex_unlock(&ust_buffers_channels_mutex); |
b5b073e2 PMF |
286 | } |
287 | ||
b5b073e2 PMF |
288 | /* |
289 | * ------- | |
290 | */ | |
291 | ||
204141ee | 292 | static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu); |
b5b073e2 PMF |
293 | |
294 | static void ltt_force_switch(struct ust_buffer *buf, | |
295 | enum force_switch_mode mode); | |
296 | ||
297 | /* | |
298 | * Trace callbacks | |
299 | */ | |
b73a4c47 | 300 | static void ltt_buffer_begin(struct ust_buffer *buf, |
b5b073e2 PMF |
301 | u64 tsc, unsigned int subbuf_idx) |
302 | { | |
303 | struct ust_channel *channel = buf->chan; | |
304 | struct ltt_subbuffer_header *header = | |
305 | (struct ltt_subbuffer_header *) | |
b73a4c47 | 306 | ust_buffers_offset_address(buf, |
b5b073e2 PMF |
307 | subbuf_idx * buf->chan->subbuf_size); |
308 | ||
309 | header->cycle_count_begin = tsc; | |
02af3e60 PMF |
310 | header->data_size = 0xFFFFFFFF; /* for recognizing crashed buffers */ |
311 | header->sb_size = 0xFFFFFFFF; /* for recognizing crashed buffers */ | |
312 | /* FIXME: add memory barrier? */ | |
b5b073e2 PMF |
313 | ltt_write_trace_header(channel->trace, header); |
314 | } | |
315 | ||
316 | /* | |
317 | * offset is assumed to never be 0 here : never deliver a completely empty | |
318 | * subbuffer. The lost size is between 0 and subbuf_size-1. | |
319 | */ | |
b73a4c47 | 320 | static notrace void ltt_buffer_end(struct ust_buffer *buf, |
b5b073e2 PMF |
321 | u64 tsc, unsigned int offset, unsigned int subbuf_idx) |
322 | { | |
323 | struct ltt_subbuffer_header *header = | |
324 | (struct ltt_subbuffer_header *) | |
b73a4c47 | 325 | ust_buffers_offset_address(buf, |
b5b073e2 | 326 | subbuf_idx * buf->chan->subbuf_size); |
8c36d1ee | 327 | u32 data_size = SUBBUF_OFFSET(offset - 1, buf->chan) + 1; |
b5b073e2 | 328 | |
8c36d1ee PMF |
329 | header->data_size = data_size; |
330 | header->sb_size = PAGE_ALIGN(data_size); | |
b5b073e2 | 331 | header->cycle_count_end = tsc; |
b102c2b0 PMF |
332 | header->events_lost = uatomic_read(&buf->events_lost); |
333 | header->subbuf_corrupt = uatomic_read(&buf->corrupted_subbuffers); | |
719569e4 PMF |
334 | if(unlikely(header->events_lost > 0)) { |
335 | DBG("Some events (%d) were lost in %s_%d", header->events_lost, buf->chan->channel_name, buf->cpu); | |
336 | } | |
b5b073e2 PMF |
337 | } |
338 | ||
339 | /* | |
340 | * This function should not be called from NMI interrupt context | |
341 | */ | |
342 | static notrace void ltt_buf_unfull(struct ust_buffer *buf, | |
343 | unsigned int subbuf_idx, | |
344 | long offset) | |
345 | { | |
b5b073e2 PMF |
346 | } |
347 | ||
b73a4c47 PMF |
348 | /* |
349 | * Promote compiler barrier to a smp_mb(). | |
350 | * For the specific LTTng case, this IPI call should be removed if the | |
351 | * architecture does not reorder writes. This should eventually be provided by | |
352 | * a separate architecture-specific infrastructure. | |
353 | */ | |
e17571a5 PMF |
354 | //ust// static void remote_mb(void *info) |
355 | //ust// { | |
356 | //ust// smp_mb(); | |
357 | //ust// } | |
b73a4c47 PMF |
358 | |
359 | int ust_buffers_get_subbuf(struct ust_buffer *buf, long *consumed) | |
b5b073e2 PMF |
360 | { |
361 | struct ust_channel *channel = buf->chan; | |
362 | long consumed_old, consumed_idx, commit_count, write_offset; | |
b73a4c47 PMF |
363 | //ust// int retval; |
364 | ||
b102c2b0 | 365 | consumed_old = uatomic_read(&buf->consumed); |
b5b073e2 | 366 | consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan); |
b102c2b0 | 367 | commit_count = uatomic_read(&buf->commit_count[consumed_idx].cc_sb); |
b5b073e2 PMF |
368 | /* |
369 | * Make sure we read the commit count before reading the buffer | |
370 | * data and the write offset. Correct consumed offset ordering | |
371 | * wrt commit count is insured by the use of cmpxchg to update | |
372 | * the consumed offset. | |
b73a4c47 PMF |
373 | * smp_call_function_single can fail if the remote CPU is offline, |
374 | * this is OK because then there is no wmb to execute there. | |
375 | * If our thread is executing on the same CPU as the on the buffers | |
376 | * belongs to, we don't have to synchronize it at all. If we are | |
377 | * migrated, the scheduler will take care of the memory barriers. | |
378 | * Normally, smp_call_function_single() should ensure program order when | |
379 | * executing the remote function, which implies that it surrounds the | |
380 | * function execution with : | |
381 | * smp_mb() | |
382 | * send IPI | |
383 | * csd_lock_wait | |
384 | * recv IPI | |
385 | * smp_mb() | |
386 | * exec. function | |
387 | * smp_mb() | |
388 | * csd unlock | |
389 | * smp_mb() | |
390 | * | |
391 | * However, smp_call_function_single() does not seem to clearly execute | |
392 | * such barriers. It depends on spinlock semantic to provide the barrier | |
393 | * before executing the IPI and, when busy-looping, csd_lock_wait only | |
394 | * executes smp_mb() when it has to wait for the other CPU. | |
395 | * | |
396 | * I don't trust this code. Therefore, let's add the smp_mb() sequence | |
397 | * required ourself, even if duplicated. It has no performance impact | |
398 | * anyway. | |
399 | * | |
400 | * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs | |
401 | * read and write vs write. They do not ensure core synchronization. We | |
402 | * really have to ensure total order between the 3 barriers running on | |
403 | * the 2 CPUs. | |
404 | */ | |
405 | //ust// #ifdef LTT_NO_IPI_BARRIER | |
406 | /* | |
407 | * Local rmb to match the remote wmb to read the commit count before the | |
408 | * buffer data and the write offset. | |
b5b073e2 PMF |
409 | */ |
410 | smp_rmb(); | |
b73a4c47 PMF |
411 | //ust// #else |
412 | //ust// if (raw_smp_processor_id() != buf->cpu) { | |
413 | //ust// smp_mb(); /* Total order with IPI handler smp_mb() */ | |
414 | //ust// smp_call_function_single(buf->cpu, remote_mb, NULL, 1); | |
415 | //ust// smp_mb(); /* Total order with IPI handler smp_mb() */ | |
416 | //ust// } | |
417 | //ust// #endif | |
418 | ||
b102c2b0 | 419 | write_offset = uatomic_read(&buf->offset); |
b5b073e2 PMF |
420 | /* |
421 | * Check that the subbuffer we are trying to consume has been | |
422 | * already fully committed. | |
423 | */ | |
424 | if (((commit_count - buf->chan->subbuf_size) | |
425 | & channel->commit_count_mask) | |
426 | - (BUFFER_TRUNC(consumed_old, buf->chan) | |
427 | >> channel->n_subbufs_order) | |
428 | != 0) { | |
429 | return -EAGAIN; | |
430 | } | |
431 | /* | |
432 | * Check that we are not about to read the same subbuffer in | |
433 | * which the writer head is. | |
434 | */ | |
435 | if ((SUBBUF_TRUNC(write_offset, buf->chan) | |
436 | - SUBBUF_TRUNC(consumed_old, buf->chan)) | |
437 | == 0) { | |
438 | return -EAGAIN; | |
439 | } | |
440 | ||
b73a4c47 PMF |
441 | /* FIXME: is this ok to disable the reading feature? */ |
442 | //ust// retval = update_read_sb_index(buf, consumed_idx); | |
443 | //ust// if (retval) | |
444 | //ust// return retval; | |
445 | ||
446 | *consumed = consumed_old; | |
447 | ||
b5b073e2 PMF |
448 | return 0; |
449 | } | |
450 | ||
b73a4c47 | 451 | int ust_buffers_put_subbuf(struct ust_buffer *buf, unsigned long uconsumed_old) |
b5b073e2 PMF |
452 | { |
453 | long consumed_new, consumed_old; | |
454 | ||
b102c2b0 | 455 | consumed_old = uatomic_read(&buf->consumed); |
b5b073e2 PMF |
456 | consumed_old = consumed_old & (~0xFFFFFFFFL); |
457 | consumed_old = consumed_old | uconsumed_old; | |
458 | consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan); | |
459 | ||
460 | //ust// spin_lock(<t_buf->full_lock); | |
b102c2b0 | 461 | if (uatomic_cmpxchg(&buf->consumed, consumed_old, |
b5b073e2 PMF |
462 | consumed_new) |
463 | != consumed_old) { | |
464 | /* We have been pushed by the writer : the last | |
465 | * buffer read _is_ corrupted! It can also | |
466 | * happen if this is a buffer we never got. */ | |
467 | //ust// spin_unlock(<t_buf->full_lock); | |
468 | return -EIO; | |
469 | } else { | |
470 | /* tell the client that buffer is now unfull */ | |
471 | int index; | |
472 | long data; | |
473 | index = SUBBUF_INDEX(consumed_old, buf->chan); | |
474 | data = BUFFER_OFFSET(consumed_old, buf->chan); | |
475 | ltt_buf_unfull(buf, index, data); | |
476 | //ust// spin_unlock(<t_buf->full_lock); | |
477 | } | |
478 | return 0; | |
479 | } | |
480 | ||
481 | static void ltt_relay_print_subbuffer_errors( | |
482 | struct ust_channel *channel, | |
204141ee | 483 | long cons_off, int cpu) |
b5b073e2 | 484 | { |
204141ee | 485 | struct ust_buffer *ltt_buf = channel->buf[cpu]; |
b73a4c47 | 486 | long cons_idx, commit_count, commit_count_sb, write_offset; |
b5b073e2 PMF |
487 | |
488 | cons_idx = SUBBUF_INDEX(cons_off, channel); | |
b102c2b0 PMF |
489 | commit_count = uatomic_read(<t_buf->commit_count[cons_idx].cc); |
490 | commit_count_sb = uatomic_read(<t_buf->commit_count[cons_idx].cc_sb); | |
b73a4c47 | 491 | |
b5b073e2 PMF |
492 | /* |
493 | * No need to order commit_count and write_offset reads because we | |
494 | * execute after trace is stopped when there are no readers left. | |
495 | */ | |
b102c2b0 | 496 | write_offset = uatomic_read(<t_buf->offset); |
b5b073e2 | 497 | WARN( "LTT : unread channel %s offset is %ld " |
b73a4c47 PMF |
498 | "and cons_off : %ld (cpu %d)\n", |
499 | channel->channel_name, write_offset, cons_off, cpu); | |
b5b073e2 PMF |
500 | /* Check each sub-buffer for non filled commit count */ |
501 | if (((commit_count - channel->subbuf_size) & channel->commit_count_mask) | |
502 | - (BUFFER_TRUNC(cons_off, channel) >> channel->n_subbufs_order) != 0) { | |
503 | ERR("LTT : %s : subbuffer %lu has non filled " | |
b73a4c47 PMF |
504 | "commit count [cc, cc_sb] [%lu,%lu].\n", |
505 | channel->channel_name, cons_idx, commit_count, commit_count_sb); | |
b5b073e2 PMF |
506 | } |
507 | ERR("LTT : %s : commit count : %lu, subbuf size %zd\n", | |
508 | channel->channel_name, commit_count, | |
509 | channel->subbuf_size); | |
510 | } | |
511 | ||
b73a4c47 | 512 | static void ltt_relay_print_errors(struct ust_trace *trace, |
204141ee | 513 | struct ust_channel *channel, int cpu) |
b5b073e2 | 514 | { |
204141ee | 515 | struct ust_buffer *ltt_buf = channel->buf[cpu]; |
b5b073e2 PMF |
516 | long cons_off; |
517 | ||
4292ed8a PMF |
518 | /* |
519 | * Can be called in the error path of allocation when | |
520 | * trans_channel_data is not yet set. | |
521 | */ | |
522 | if (!channel) | |
523 | return; | |
524 | ||
e17571a5 PMF |
525 | //ust// for (cons_off = 0; cons_off < rchan->alloc_size; |
526 | //ust// cons_off = SUBBUF_ALIGN(cons_off, rchan)) | |
527 | //ust// ust_buffers_print_written(ltt_chan, cons_off, cpu); | |
b102c2b0 PMF |
528 | for (cons_off = uatomic_read(<t_buf->consumed); |
529 | (SUBBUF_TRUNC(uatomic_read(<t_buf->offset), | |
b5b073e2 PMF |
530 | channel) |
531 | - cons_off) > 0; | |
532 | cons_off = SUBBUF_ALIGN(cons_off, channel)) | |
204141ee | 533 | ltt_relay_print_subbuffer_errors(channel, cons_off, cpu); |
b5b073e2 PMF |
534 | } |
535 | ||
204141ee | 536 | static void ltt_relay_print_buffer_errors(struct ust_channel *channel, int cpu) |
b5b073e2 | 537 | { |
b73a4c47 | 538 | struct ust_trace *trace = channel->trace; |
204141ee | 539 | struct ust_buffer *ltt_buf = channel->buf[cpu]; |
b5b073e2 | 540 | |
b102c2b0 | 541 | if (uatomic_read(<t_buf->events_lost)) |
b73a4c47 | 542 | ERR("channel %s: %ld events lost (cpu %d)", |
b5b073e2 | 543 | channel->channel_name, |
b102c2b0 PMF |
544 | uatomic_read(<t_buf->events_lost), cpu); |
545 | if (uatomic_read(<t_buf->corrupted_subbuffers)) | |
b73a4c47 | 546 | ERR("channel %s : %ld corrupted subbuffers (cpu %d)", |
b5b073e2 | 547 | channel->channel_name, |
b102c2b0 | 548 | uatomic_read(<t_buf->corrupted_subbuffers), cpu); |
b5b073e2 | 549 | |
204141ee | 550 | ltt_relay_print_errors(trace, channel, cpu); |
b5b073e2 PMF |
551 | } |
552 | ||
553 | static void ltt_relay_release_channel(struct kref *kref) | |
554 | { | |
555 | struct ust_channel *ltt_chan = container_of(kref, | |
556 | struct ust_channel, kref); | |
557 | free(ltt_chan->buf); | |
558 | } | |
559 | ||
560 | /* | |
561 | * Create ltt buffer. | |
562 | */ | |
b73a4c47 | 563 | //ust// static int ltt_relay_create_buffer(struct ust_trace *trace, |
b5b073e2 PMF |
564 | //ust// struct ltt_channel_struct *ltt_chan, struct rchan_buf *buf, |
565 | //ust// unsigned int cpu, unsigned int n_subbufs) | |
566 | //ust// { | |
567 | //ust// struct ltt_channel_buf_struct *ltt_buf = | |
568 | //ust// percpu_ptr(ltt_chan->buf, cpu); | |
569 | //ust// unsigned int j; | |
b73a4c47 | 570 | //ust// |
b5b073e2 PMF |
571 | //ust// ltt_buf->commit_count = |
572 | //ust// kzalloc_node(sizeof(ltt_buf->commit_count) * n_subbufs, | |
573 | //ust// GFP_KERNEL, cpu_to_node(cpu)); | |
574 | //ust// if (!ltt_buf->commit_count) | |
575 | //ust// return -ENOMEM; | |
576 | //ust// kref_get(&trace->kref); | |
577 | //ust// kref_get(&trace->ltt_transport_kref); | |
578 | //ust// kref_get(<t_chan->kref); | |
b102c2b0 PMF |
579 | //ust// uatomic_set(<t_buf->offset, ltt_subbuffer_header_size()); |
580 | //ust// uatomic_set(<t_buf->consumed, 0); | |
581 | //ust// uatomic_set(<t_buf->active_readers, 0); | |
b5b073e2 | 582 | //ust// for (j = 0; j < n_subbufs; j++) |
b102c2b0 | 583 | //ust// uatomic_set(<t_buf->commit_count[j], 0); |
b5b073e2 | 584 | //ust// init_waitqueue_head(<t_buf->write_wait); |
b102c2b0 | 585 | //ust// uatomic_set(<t_buf->wakeup_readers, 0); |
b5b073e2 | 586 | //ust// spin_lock_init(<t_buf->full_lock); |
b73a4c47 | 587 | //ust// |
b5b073e2 PMF |
588 | //ust// ltt_buffer_begin_callback(buf, trace->start_tsc, 0); |
589 | //ust// /* atomic_add made on local variable on data that belongs to | |
590 | //ust// * various CPUs : ok because tracing not started (for this cpu). */ | |
b102c2b0 | 591 | //ust// uatomic_add(<t_buf->commit_count[0], ltt_subbuffer_header_size()); |
b73a4c47 | 592 | //ust// |
b102c2b0 PMF |
593 | //ust// uatomic_set(<t_buf->events_lost, 0); |
594 | //ust// uatomic_set(<t_buf->corrupted_subbuffers, 0); | |
b73a4c47 | 595 | //ust// |
b5b073e2 PMF |
596 | //ust// return 0; |
597 | //ust// } | |
598 | ||
b73a4c47 | 599 | static int ust_buffers_init_buffer(struct ust_trace *trace, |
b5b073e2 PMF |
600 | struct ust_channel *ltt_chan, struct ust_buffer *buf, |
601 | unsigned int n_subbufs) | |
602 | { | |
603 | unsigned int j; | |
604 | int fds[2]; | |
605 | int result; | |
606 | ||
607 | buf->commit_count = | |
b73a4c47 | 608 | zmalloc(sizeof(*buf->commit_count) * n_subbufs); |
b5b073e2 PMF |
609 | if (!buf->commit_count) |
610 | return -ENOMEM; | |
611 | kref_get(&trace->kref); | |
612 | kref_get(&trace->ltt_transport_kref); | |
613 | kref_get(<t_chan->kref); | |
b102c2b0 PMF |
614 | uatomic_set(&buf->offset, ltt_subbuffer_header_size()); |
615 | uatomic_set(&buf->consumed, 0); | |
616 | uatomic_set(&buf->active_readers, 0); | |
b73a4c47 | 617 | for (j = 0; j < n_subbufs; j++) { |
b102c2b0 PMF |
618 | uatomic_set(&buf->commit_count[j].cc, 0); |
619 | uatomic_set(&buf->commit_count[j].cc_sb, 0); | |
b73a4c47 | 620 | } |
b5b073e2 | 621 | //ust// init_waitqueue_head(&buf->write_wait); |
b102c2b0 | 622 | //ust// uatomic_set(&buf->wakeup_readers, 0); |
b5b073e2 PMF |
623 | //ust// spin_lock_init(&buf->full_lock); |
624 | ||
b73a4c47 | 625 | ltt_buffer_begin(buf, trace->start_tsc, 0); |
b5b073e2 | 626 | |
b102c2b0 | 627 | uatomic_add(&buf->commit_count[0].cc, ltt_subbuffer_header_size()); |
b5b073e2 | 628 | |
b102c2b0 PMF |
629 | uatomic_set(&buf->events_lost, 0); |
630 | uatomic_set(&buf->corrupted_subbuffers, 0); | |
b5b073e2 PMF |
631 | |
632 | result = pipe(fds); | |
633 | if(result == -1) { | |
634 | PERROR("pipe"); | |
635 | return -1; | |
636 | } | |
637 | buf->data_ready_fd_read = fds[0]; | |
638 | buf->data_ready_fd_write = fds[1]; | |
639 | ||
b5b073e2 PMF |
640 | //ust// buf->commit_seq = malloc(sizeof(buf->commit_seq) * n_subbufs); |
641 | //ust// if(!ltt_buf->commit_seq) { | |
642 | //ust// return -1; | |
643 | //ust// } | |
37315729 | 644 | memset(buf->commit_seq, 0, sizeof(buf->commit_seq[0]) * n_subbufs); |
b5b073e2 PMF |
645 | |
646 | /* FIXME: decrementally destroy on error */ | |
647 | ||
648 | return 0; | |
649 | } | |
650 | ||
651 | /* FIXME: use this function */ | |
204141ee | 652 | static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu) |
b5b073e2 | 653 | { |
b73a4c47 | 654 | struct ust_trace *trace = ltt_chan->trace; |
204141ee | 655 | struct ust_buffer *ltt_buf = ltt_chan->buf[cpu]; |
b5b073e2 PMF |
656 | |
657 | kref_put(<t_chan->trace->ltt_transport_kref, | |
658 | ltt_release_transport); | |
204141ee | 659 | ltt_relay_print_buffer_errors(ltt_chan, cpu); |
b5b073e2 | 660 | //ust// free(ltt_buf->commit_seq); |
909bc43f | 661 | free(ltt_buf->commit_count); |
b5b073e2 PMF |
662 | ltt_buf->commit_count = NULL; |
663 | kref_put(<t_chan->kref, ltt_relay_release_channel); | |
664 | kref_put(&trace->kref, ltt_release_trace); | |
665 | //ust// wake_up_interruptible(&trace->kref_wq); | |
666 | } | |
667 | ||
204141ee | 668 | static int ust_buffers_alloc_channel_buf_structs(struct ust_channel *chan) |
b5b073e2 PMF |
669 | { |
670 | void *ptr; | |
671 | int result; | |
204141ee PMF |
672 | size_t size; |
673 | int i; | |
b5b073e2 | 674 | |
204141ee | 675 | size = PAGE_ALIGN(1); |
b5b073e2 | 676 | |
204141ee | 677 | for(i=0; i<chan->n_cpus; i++) { |
b5b073e2 | 678 | |
204141ee PMF |
679 | result = chan->buf_struct_shmids[i] = shmget(getpid(), size, IPC_CREAT | IPC_EXCL | 0700); |
680 | if(result == -1) { | |
681 | PERROR("shmget"); | |
682 | goto destroy_previous; | |
683 | } | |
b5b073e2 | 684 | |
204141ee PMF |
685 | /* FIXME: should have matching call to shmdt */ |
686 | ptr = shmat(chan->buf_struct_shmids[i], NULL, 0); | |
687 | if(ptr == (void *) -1) { | |
688 | perror("shmat"); | |
689 | goto destroy_shm; | |
690 | } | |
691 | ||
692 | /* Already mark the shared memory for destruction. This will occur only | |
693 | * when all users have detached. | |
694 | */ | |
695 | result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL); | |
696 | if(result == -1) { | |
697 | perror("shmctl"); | |
698 | goto destroy_previous; | |
699 | } | |
700 | ||
701 | chan->buf[i] = ptr; | |
b5b073e2 PMF |
702 | } |
703 | ||
204141ee | 704 | return 0; |
b5b073e2 | 705 | |
204141ee PMF |
706 | /* Jumping inside this loop occurs from within the other loop above with i as |
707 | * counter, so it unallocates the structures for the cpu = current_i down to | |
708 | * zero. */ | |
709 | for(; i>=0; i--) { | |
710 | destroy_shm: | |
711 | result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL); | |
712 | if(result == -1) { | |
713 | perror("shmctl"); | |
714 | } | |
b5b073e2 | 715 | |
204141ee PMF |
716 | destroy_previous: |
717 | continue; | |
b5b073e2 PMF |
718 | } |
719 | ||
204141ee | 720 | return -1; |
b5b073e2 PMF |
721 | } |
722 | ||
723 | /* | |
724 | * Create channel. | |
725 | */ | |
b73a4c47 | 726 | static int ust_buffers_create_channel(const char *trace_name, struct ust_trace *trace, |
b5b073e2 PMF |
727 | const char *channel_name, struct ust_channel *ltt_chan, |
728 | unsigned int subbuf_size, unsigned int n_subbufs, int overwrite) | |
729 | { | |
b5b073e2 PMF |
730 | int result; |
731 | ||
732 | kref_init(<t_chan->kref); | |
733 | ||
734 | ltt_chan->trace = trace; | |
b5b073e2 PMF |
735 | ltt_chan->overwrite = overwrite; |
736 | ltt_chan->n_subbufs_order = get_count_order(n_subbufs); | |
737 | ltt_chan->commit_count_mask = (~0UL >> ltt_chan->n_subbufs_order); | |
204141ee | 738 | ltt_chan->n_cpus = get_n_cpus(); |
b5b073e2 | 739 | //ust// ltt_chan->buf = percpu_alloc_mask(sizeof(struct ltt_channel_buf_struct), GFP_KERNEL, cpu_possible_map); |
204141ee PMF |
740 | ltt_chan->buf = (void *) malloc(ltt_chan->n_cpus * sizeof(void *)); |
741 | if(ltt_chan->buf == NULL) { | |
742 | goto error; | |
743 | } | |
744 | ltt_chan->buf_struct_shmids = (int *) malloc(ltt_chan->n_cpus * sizeof(int)); | |
745 | if(ltt_chan->buf_struct_shmids == NULL) | |
746 | goto free_buf; | |
b5b073e2 | 747 | |
204141ee PMF |
748 | result = ust_buffers_alloc_channel_buf_structs(ltt_chan); |
749 | if(result != 0) { | |
750 | goto free_buf_struct_shmids; | |
751 | } | |
b5b073e2 | 752 | |
b5b073e2 | 753 | result = ust_buffers_channel_open(ltt_chan, subbuf_size, n_subbufs); |
204141ee | 754 | if (result != 0) { |
c1f20530 | 755 | ERR("Cannot open channel for trace %s", trace_name); |
204141ee | 756 | goto unalloc_buf_structs; |
b5b073e2 PMF |
757 | } |
758 | ||
204141ee PMF |
759 | return 0; |
760 | ||
761 | unalloc_buf_structs: | |
762 | /* FIXME: put a call here to unalloc the buf structs! */ | |
763 | ||
764 | free_buf_struct_shmids: | |
765 | free(ltt_chan->buf_struct_shmids); | |
b5b073e2 | 766 | |
204141ee PMF |
767 | free_buf: |
768 | free(ltt_chan->buf); | |
769 | ||
770 | error: | |
771 | return -1; | |
b5b073e2 PMF |
772 | } |
773 | ||
b5b073e2 PMF |
774 | static void ltt_relay_async_wakeup_chan(struct ust_channel *ltt_channel) |
775 | { | |
776 | //ust// unsigned int i; | |
777 | //ust// struct rchan *rchan = ltt_channel->trans_channel_data; | |
778 | //ust// | |
779 | //ust// for_each_possible_cpu(i) { | |
780 | //ust// struct ltt_channel_buf_struct *ltt_buf = | |
781 | //ust// percpu_ptr(ltt_channel->buf, i); | |
782 | //ust// | |
b102c2b0 PMF |
783 | //ust// if (uatomic_read(<t_buf->wakeup_readers) == 1) { |
784 | //ust// uatomic_set(<t_buf->wakeup_readers, 0); | |
b5b073e2 PMF |
785 | //ust// wake_up_interruptible(&rchan->buf[i]->read_wait); |
786 | //ust// } | |
787 | //ust// } | |
788 | } | |
789 | ||
204141ee | 790 | static void ltt_relay_finish_buffer(struct ust_channel *channel, unsigned int cpu) |
b5b073e2 PMF |
791 | { |
792 | // int result; | |
793 | ||
204141ee PMF |
794 | if (channel->buf[cpu]) { |
795 | struct ust_buffer *buf = channel->buf[cpu]; | |
97c10252 | 796 | ltt_force_switch(buf, FORCE_FLUSH); |
b5b073e2 PMF |
797 | //ust// ltt_relay_wake_writers(ltt_buf); |
798 | /* closing the pipe tells the consumer the buffer is finished */ | |
799 | ||
800 | //result = write(ltt_buf->data_ready_fd_write, "D", 1); | |
801 | //if(result == -1) { | |
802 | // PERROR("write (in ltt_relay_finish_buffer)"); | |
803 | // ERR("this should never happen!"); | |
804 | //} | |
805 | close(buf->data_ready_fd_write); | |
806 | } | |
807 | } | |
808 | ||
809 | ||
810 | static void ltt_relay_finish_channel(struct ust_channel *channel) | |
811 | { | |
204141ee | 812 | unsigned int i; |
b5b073e2 | 813 | |
204141ee PMF |
814 | for(i=0; i<channel->n_cpus; i++) { |
815 | ltt_relay_finish_buffer(channel, i); | |
816 | } | |
b5b073e2 PMF |
817 | } |
818 | ||
819 | static void ltt_relay_remove_channel(struct ust_channel *channel) | |
820 | { | |
821 | ust_buffers_channel_close(channel); | |
822 | kref_put(&channel->kref, ltt_relay_release_channel); | |
823 | } | |
824 | ||
b5b073e2 | 825 | /* |
b73a4c47 PMF |
826 | * ltt_reserve_switch_old_subbuf: switch old subbuffer |
827 | * | |
828 | * Concurrency safe because we are the last and only thread to alter this | |
829 | * sub-buffer. As long as it is not delivered and read, no other thread can | |
830 | * alter the offset, alter the reserve_count or call the | |
831 | * client_buffer_end_callback on this sub-buffer. | |
832 | * | |
833 | * The only remaining threads could be the ones with pending commits. They will | |
834 | * have to do the deliver themselves. Not concurrency safe in overwrite mode. | |
835 | * We detect corrupted subbuffers with commit and reserve counts. We keep a | |
836 | * corrupted sub-buffers count and push the readers across these sub-buffers. | |
837 | * | |
838 | * Not concurrency safe if a writer is stalled in a subbuffer and another writer | |
839 | * switches in, finding out it's corrupted. The result will be than the old | |
840 | * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer | |
841 | * will be declared corrupted too because of the commit count adjustment. | |
842 | * | |
843 | * Note : offset_old should never be 0 here. | |
b5b073e2 | 844 | */ |
b73a4c47 PMF |
845 | static void ltt_reserve_switch_old_subbuf( |
846 | struct ust_channel *chan, struct ust_buffer *buf, | |
847 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) | |
b5b073e2 | 848 | { |
b73a4c47 PMF |
849 | long oldidx = SUBBUF_INDEX(offsets->old - 1, chan); |
850 | long commit_count, padding_size; | |
b5b073e2 | 851 | |
b73a4c47 PMF |
852 | padding_size = chan->subbuf_size |
853 | - (SUBBUF_OFFSET(offsets->old - 1, chan) + 1); | |
854 | ltt_buffer_end(buf, *tsc, offsets->old, oldidx); | |
b5b073e2 | 855 | |
b73a4c47 PMF |
856 | /* |
857 | * Must write slot data before incrementing commit count. | |
858 | * This compiler barrier is upgraded into a smp_wmb() by the IPI | |
859 | * sent by get_subbuf() when it does its smp_rmb(). | |
860 | */ | |
861 | barrier(); | |
b102c2b0 PMF |
862 | uatomic_add(&buf->commit_count[oldidx].cc, padding_size); |
863 | commit_count = uatomic_read(&buf->commit_count[oldidx].cc); | |
b73a4c47 | 864 | ltt_check_deliver(chan, buf, offsets->old - 1, commit_count, oldidx); |
1e8c9e7b | 865 | ltt_write_commit_counter(chan, buf, oldidx, |
b73a4c47 PMF |
866 | offsets->old, commit_count, padding_size); |
867 | } | |
b5b073e2 | 868 | |
b73a4c47 PMF |
869 | /* |
870 | * ltt_reserve_switch_new_subbuf: Populate new subbuffer. | |
871 | * | |
872 | * This code can be executed unordered : writers may already have written to the | |
873 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
874 | * that this code is executed before the deliver of this sub-buffer. | |
875 | */ | |
876 | static void ltt_reserve_switch_new_subbuf( | |
877 | struct ust_channel *chan, struct ust_buffer *buf, | |
878 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) | |
879 | { | |
880 | long beginidx = SUBBUF_INDEX(offsets->begin, chan); | |
881 | long commit_count; | |
b5b073e2 | 882 | |
b73a4c47 | 883 | ltt_buffer_begin(buf, *tsc, beginidx); |
b5b073e2 | 884 | |
b73a4c47 PMF |
885 | /* |
886 | * Must write slot data before incrementing commit count. | |
887 | * This compiler barrier is upgraded into a smp_wmb() by the IPI | |
888 | * sent by get_subbuf() when it does its smp_rmb(). | |
889 | */ | |
890 | barrier(); | |
b102c2b0 PMF |
891 | uatomic_add(&buf->commit_count[beginidx].cc, ltt_subbuffer_header_size()); |
892 | commit_count = uatomic_read(&buf->commit_count[beginidx].cc); | |
b73a4c47 PMF |
893 | /* Check if the written buffer has to be delivered */ |
894 | ltt_check_deliver(chan, buf, offsets->begin, commit_count, beginidx); | |
1e8c9e7b | 895 | ltt_write_commit_counter(chan, buf, beginidx, |
b73a4c47 PMF |
896 | offsets->begin, commit_count, ltt_subbuffer_header_size()); |
897 | } | |
b5b073e2 | 898 | |
b73a4c47 PMF |
899 | /* |
900 | * ltt_reserve_end_switch_current: finish switching current subbuffer | |
901 | * | |
902 | * Concurrency safe because we are the last and only thread to alter this | |
903 | * sub-buffer. As long as it is not delivered and read, no other thread can | |
904 | * alter the offset, alter the reserve_count or call the | |
905 | * client_buffer_end_callback on this sub-buffer. | |
906 | * | |
907 | * The only remaining threads could be the ones with pending commits. They will | |
908 | * have to do the deliver themselves. Not concurrency safe in overwrite mode. | |
909 | * We detect corrupted subbuffers with commit and reserve counts. We keep a | |
910 | * corrupted sub-buffers count and push the readers across these sub-buffers. | |
911 | * | |
912 | * Not concurrency safe if a writer is stalled in a subbuffer and another writer | |
913 | * switches in, finding out it's corrupted. The result will be than the old | |
914 | * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer | |
915 | * will be declared corrupted too because of the commit count adjustment. | |
916 | */ | |
917 | static void ltt_reserve_end_switch_current( | |
918 | struct ust_channel *chan, | |
919 | struct ust_buffer *buf, | |
920 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) | |
921 | { | |
922 | long endidx = SUBBUF_INDEX(offsets->end - 1, chan); | |
923 | long commit_count, padding_size; | |
924 | ||
925 | padding_size = chan->subbuf_size | |
926 | - (SUBBUF_OFFSET(offsets->end - 1, chan) + 1); | |
927 | ||
928 | ltt_buffer_end(buf, *tsc, offsets->end, endidx); | |
929 | ||
930 | /* | |
931 | * Must write slot data before incrementing commit count. | |
932 | * This compiler barrier is upgraded into a smp_wmb() by the IPI | |
933 | * sent by get_subbuf() when it does its smp_rmb(). | |
934 | */ | |
935 | barrier(); | |
b102c2b0 PMF |
936 | uatomic_add(&buf->commit_count[endidx].cc, padding_size); |
937 | commit_count = uatomic_read(&buf->commit_count[endidx].cc); | |
b73a4c47 PMF |
938 | ltt_check_deliver(chan, buf, |
939 | offsets->end - 1, commit_count, endidx); | |
1e8c9e7b | 940 | ltt_write_commit_counter(chan, buf, endidx, |
b73a4c47 | 941 | offsets->end, commit_count, padding_size); |
b5b073e2 PMF |
942 | } |
943 | ||
944 | /* | |
945 | * Returns : | |
946 | * 0 if ok | |
947 | * !0 if execution must be aborted. | |
948 | */ | |
b73a4c47 | 949 | static int ltt_relay_try_switch_slow( |
b5b073e2 | 950 | enum force_switch_mode mode, |
b73a4c47 | 951 | struct ust_channel *chan, |
b5b073e2 PMF |
952 | struct ust_buffer *buf, |
953 | struct ltt_reserve_switch_offsets *offsets, | |
954 | u64 *tsc) | |
955 | { | |
956 | long subbuf_index; | |
b73a4c47 | 957 | long reserve_commit_diff; |
b5b073e2 | 958 | |
b102c2b0 | 959 | offsets->begin = uatomic_read(&buf->offset); |
b5b073e2 PMF |
960 | offsets->old = offsets->begin; |
961 | offsets->begin_switch = 0; | |
962 | offsets->end_switch_old = 0; | |
963 | ||
964 | *tsc = trace_clock_read64(); | |
965 | ||
966 | if (SUBBUF_OFFSET(offsets->begin, buf->chan) != 0) { | |
967 | offsets->begin = SUBBUF_ALIGN(offsets->begin, buf->chan); | |
968 | offsets->end_switch_old = 1; | |
969 | } else { | |
970 | /* we do not have to switch : buffer is empty */ | |
971 | return -1; | |
972 | } | |
973 | if (mode == FORCE_ACTIVE) | |
974 | offsets->begin += ltt_subbuffer_header_size(); | |
975 | /* | |
976 | * Always begin_switch in FORCE_ACTIVE mode. | |
977 | * Test new buffer integrity | |
978 | */ | |
979 | subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); | |
b73a4c47 | 980 | reserve_commit_diff = |
b5b073e2 | 981 | (BUFFER_TRUNC(offsets->begin, buf->chan) |
b73a4c47 | 982 | >> chan->n_subbufs_order) |
b102c2b0 | 983 | - (uatomic_read(&buf->commit_count[subbuf_index].cc_sb) |
b73a4c47 PMF |
984 | & chan->commit_count_mask); |
985 | if (reserve_commit_diff == 0) { | |
b5b073e2 PMF |
986 | /* Next buffer not corrupted. */ |
987 | if (mode == FORCE_ACTIVE | |
b73a4c47 | 988 | && !chan->overwrite |
b102c2b0 | 989 | && offsets->begin - uatomic_read(&buf->consumed) |
b73a4c47 | 990 | >= chan->alloc_size) { |
b5b073e2 PMF |
991 | /* |
992 | * We do not overwrite non consumed buffers and we are | |
993 | * full : ignore switch while tracing is active. | |
994 | */ | |
995 | return -1; | |
996 | } | |
997 | } else { | |
998 | /* | |
999 | * Next subbuffer corrupted. Force pushing reader even in normal | |
1000 | * mode | |
1001 | */ | |
1002 | } | |
1003 | offsets->end = offsets->begin; | |
1004 | return 0; | |
1005 | } | |
1006 | ||
b5b073e2 | 1007 | /* |
b73a4c47 PMF |
1008 | * Force a sub-buffer switch for a per-cpu buffer. This operation is |
1009 | * completely reentrant : can be called while tracing is active with | |
1010 | * absolutely no lock held. | |
b5b073e2 | 1011 | */ |
b73a4c47 PMF |
1012 | void ltt_force_switch_lockless_slow(struct ust_buffer *buf, |
1013 | enum force_switch_mode mode) | |
b5b073e2 | 1014 | { |
b73a4c47 | 1015 | struct ust_channel *chan = buf->chan; |
b5b073e2 | 1016 | struct ltt_reserve_switch_offsets offsets; |
b73a4c47 | 1017 | u64 tsc; |
b5b073e2 | 1018 | |
b5b073e2 PMF |
1019 | offsets.size = 0; |
1020 | ||
10dd3941 | 1021 | DBG("Switching (forced) %s_%d", chan->channel_name, buf->cpu); |
b5b073e2 PMF |
1022 | /* |
1023 | * Perform retryable operations. | |
1024 | */ | |
b5b073e2 | 1025 | do { |
b73a4c47 PMF |
1026 | if (ltt_relay_try_switch_slow(mode, chan, buf, |
1027 | &offsets, &tsc)) | |
1028 | return; | |
b102c2b0 | 1029 | } while (uatomic_cmpxchg(&buf->offset, offsets.old, |
b5b073e2 PMF |
1030 | offsets.end) != offsets.old); |
1031 | ||
1032 | /* | |
1033 | * Atomically update last_tsc. This update races against concurrent | |
1034 | * atomic updates, but the race will always cause supplementary full TSC | |
1035 | * events, never the opposite (missing a full TSC event when it would be | |
1036 | * needed). | |
1037 | */ | |
b73a4c47 | 1038 | save_last_tsc(buf, tsc); |
b5b073e2 PMF |
1039 | |
1040 | /* | |
1041 | * Push the reader if necessary | |
1042 | */ | |
b73a4c47 PMF |
1043 | if (mode == FORCE_ACTIVE) { |
1044 | ltt_reserve_push_reader(chan, buf, offsets.end - 1); | |
1045 | //ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan)); | |
1046 | } | |
b5b073e2 PMF |
1047 | |
1048 | /* | |
1049 | * Switch old subbuffer if needed. | |
1050 | */ | |
b73a4c47 PMF |
1051 | if (offsets.end_switch_old) { |
1052 | //ust// ltt_clear_noref_flag(rchan, buf, SUBBUF_INDEX(offsets.old - 1, rchan)); | |
1053 | ltt_reserve_switch_old_subbuf(chan, buf, &offsets, &tsc); | |
1054 | } | |
b5b073e2 PMF |
1055 | |
1056 | /* | |
1057 | * Populate new subbuffer. | |
1058 | */ | |
b73a4c47 PMF |
1059 | if (mode == FORCE_ACTIVE) |
1060 | ltt_reserve_switch_new_subbuf(chan, buf, &offsets, &tsc); | |
1061 | } | |
b5b073e2 | 1062 | |
b73a4c47 PMF |
1063 | /* |
1064 | * Returns : | |
1065 | * 0 if ok | |
1066 | * !0 if execution must be aborted. | |
1067 | */ | |
1068 | static int ltt_relay_try_reserve_slow(struct ust_channel *chan, struct ust_buffer *buf, | |
1069 | struct ltt_reserve_switch_offsets *offsets, size_t data_size, | |
1070 | u64 *tsc, unsigned int *rflags, int largest_align) | |
1071 | { | |
1072 | long reserve_commit_diff; | |
b5b073e2 | 1073 | |
b102c2b0 | 1074 | offsets->begin = uatomic_read(&buf->offset); |
b73a4c47 PMF |
1075 | offsets->old = offsets->begin; |
1076 | offsets->begin_switch = 0; | |
1077 | offsets->end_switch_current = 0; | |
1078 | offsets->end_switch_old = 0; | |
1079 | ||
1080 | *tsc = trace_clock_read64(); | |
1081 | if (last_tsc_overflow(buf, *tsc)) | |
1082 | *rflags = LTT_RFLAG_ID_SIZE_TSC; | |
1083 | ||
1084 | if (unlikely(SUBBUF_OFFSET(offsets->begin, buf->chan) == 0)) { | |
1085 | offsets->begin_switch = 1; /* For offsets->begin */ | |
1086 | } else { | |
1087 | offsets->size = ust_get_header_size(chan, | |
1088 | offsets->begin, data_size, | |
1089 | &offsets->before_hdr_pad, *rflags); | |
1090 | offsets->size += ltt_align(offsets->begin + offsets->size, | |
1091 | largest_align) | |
1092 | + data_size; | |
1093 | if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan) + | |
1094 | offsets->size) > buf->chan->subbuf_size)) { | |
1095 | offsets->end_switch_old = 1; /* For offsets->old */ | |
1096 | offsets->begin_switch = 1; /* For offsets->begin */ | |
1097 | } | |
1098 | } | |
1099 | if (unlikely(offsets->begin_switch)) { | |
1100 | long subbuf_index; | |
1101 | ||
1102 | /* | |
1103 | * We are typically not filling the previous buffer completely. | |
1104 | */ | |
1105 | if (likely(offsets->end_switch_old)) | |
1106 | offsets->begin = SUBBUF_ALIGN(offsets->begin, | |
1107 | buf->chan); | |
1108 | offsets->begin = offsets->begin + ltt_subbuffer_header_size(); | |
1109 | /* Test new buffer integrity */ | |
1110 | subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); | |
1111 | reserve_commit_diff = | |
1112 | (BUFFER_TRUNC(offsets->begin, buf->chan) | |
1113 | >> chan->n_subbufs_order) | |
b102c2b0 | 1114 | - (uatomic_read(&buf->commit_count[subbuf_index].cc_sb) |
b73a4c47 PMF |
1115 | & chan->commit_count_mask); |
1116 | if (likely(reserve_commit_diff == 0)) { | |
1117 | /* Next buffer not corrupted. */ | |
1118 | if (unlikely(!chan->overwrite && | |
1119 | (SUBBUF_TRUNC(offsets->begin, buf->chan) | |
b102c2b0 | 1120 | - SUBBUF_TRUNC(uatomic_read( |
b73a4c47 PMF |
1121 | &buf->consumed), |
1122 | buf->chan)) | |
1123 | >= chan->alloc_size)) { | |
1124 | /* | |
1125 | * We do not overwrite non consumed buffers | |
1126 | * and we are full : event is lost. | |
1127 | */ | |
b102c2b0 | 1128 | uatomic_inc(&buf->events_lost); |
b73a4c47 PMF |
1129 | return -1; |
1130 | } else { | |
1131 | /* | |
1132 | * next buffer not corrupted, we are either in | |
1133 | * overwrite mode or the buffer is not full. | |
1134 | * It's safe to write in this new subbuffer. | |
1135 | */ | |
1136 | } | |
1137 | } else { | |
1138 | /* | |
1139 | * Next subbuffer corrupted. Drop event in normal and | |
1140 | * overwrite mode. Caused by either a writer OOPS or | |
1141 | * too many nested writes over a reserve/commit pair. | |
1142 | */ | |
b102c2b0 | 1143 | uatomic_inc(&buf->events_lost); |
b73a4c47 PMF |
1144 | return -1; |
1145 | } | |
1146 | offsets->size = ust_get_header_size(chan, | |
1147 | offsets->begin, data_size, | |
1148 | &offsets->before_hdr_pad, *rflags); | |
1149 | offsets->size += ltt_align(offsets->begin + offsets->size, | |
1150 | largest_align) | |
1151 | + data_size; | |
1152 | if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan) | |
1153 | + offsets->size) > buf->chan->subbuf_size)) { | |
1154 | /* | |
1155 | * Event too big for subbuffers, report error, don't | |
1156 | * complete the sub-buffer switch. | |
1157 | */ | |
b102c2b0 | 1158 | uatomic_inc(&buf->events_lost); |
b73a4c47 PMF |
1159 | return -1; |
1160 | } else { | |
1161 | /* | |
1162 | * We just made a successful buffer switch and the event | |
1163 | * fits in the new subbuffer. Let's write. | |
1164 | */ | |
1165 | } | |
1166 | } else { | |
1167 | /* | |
1168 | * Event fits in the current buffer and we are not on a switch | |
1169 | * boundary. It's safe to write. | |
1170 | */ | |
1171 | } | |
1172 | offsets->end = offsets->begin + offsets->size; | |
1173 | ||
1174 | if (unlikely((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0)) { | |
1175 | /* | |
1176 | * The offset_end will fall at the very beginning of the next | |
1177 | * subbuffer. | |
1178 | */ | |
1179 | offsets->end_switch_current = 1; /* For offsets->begin */ | |
1180 | } | |
b5b073e2 PMF |
1181 | return 0; |
1182 | } | |
1183 | ||
b73a4c47 PMF |
1184 | /** |
1185 | * ltt_relay_reserve_slot_lockless_slow - Atomic slot reservation in a buffer. | |
1186 | * @trace: the trace structure to log to. | |
1187 | * @ltt_channel: channel structure | |
1188 | * @transport_data: data structure specific to ltt relay | |
1189 | * @data_size: size of the variable length data to log. | |
1190 | * @slot_size: pointer to total size of the slot (out) | |
1191 | * @buf_offset : pointer to reserved buffer offset (out) | |
1192 | * @tsc: pointer to the tsc at the slot reservation (out) | |
1193 | * @cpu: cpuid | |
b5b073e2 | 1194 | * |
b73a4c47 PMF |
1195 | * Return : -ENOSPC if not enough space, else returns 0. |
1196 | * It will take care of sub-buffer switching. | |
b5b073e2 | 1197 | */ |
12e81b07 PMF |
1198 | int ltt_reserve_slot_lockless_slow(struct ust_channel *chan, |
1199 | struct ust_trace *trace, size_t data_size, | |
1200 | int largest_align, int cpu, | |
1201 | struct ust_buffer **ret_buf, | |
1202 | size_t *slot_size, long *buf_offset, | |
1203 | u64 *tsc, unsigned int *rflags) | |
b5b073e2 | 1204 | { |
12e81b07 | 1205 | struct ust_buffer *buf = *ret_buf = chan->buf[cpu]; |
b5b073e2 | 1206 | struct ltt_reserve_switch_offsets offsets; |
b5b073e2 | 1207 | |
b5b073e2 PMF |
1208 | offsets.size = 0; |
1209 | ||
b5b073e2 | 1210 | do { |
b73a4c47 PMF |
1211 | if (unlikely(ltt_relay_try_reserve_slow(chan, buf, &offsets, |
1212 | data_size, tsc, rflags, largest_align))) | |
1213 | return -ENOSPC; | |
b102c2b0 | 1214 | } while (unlikely(uatomic_cmpxchg(&buf->offset, offsets.old, |
b73a4c47 | 1215 | offsets.end) != offsets.old)); |
b5b073e2 PMF |
1216 | |
1217 | /* | |
1218 | * Atomically update last_tsc. This update races against concurrent | |
1219 | * atomic updates, but the race will always cause supplementary full TSC | |
1220 | * events, never the opposite (missing a full TSC event when it would be | |
1221 | * needed). | |
1222 | */ | |
b73a4c47 | 1223 | save_last_tsc(buf, *tsc); |
b5b073e2 PMF |
1224 | |
1225 | /* | |
1226 | * Push the reader if necessary | |
1227 | */ | |
b73a4c47 PMF |
1228 | ltt_reserve_push_reader(chan, buf, offsets.end - 1); |
1229 | ||
1230 | /* | |
1231 | * Clear noref flag for this subbuffer. | |
1232 | */ | |
1233 | //ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan)); | |
b5b073e2 PMF |
1234 | |
1235 | /* | |
1236 | * Switch old subbuffer if needed. | |
1237 | */ | |
b73a4c47 PMF |
1238 | if (unlikely(offsets.end_switch_old)) { |
1239 | //ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.old - 1, chan)); | |
1240 | ltt_reserve_switch_old_subbuf(chan, buf, &offsets, tsc); | |
10dd3941 | 1241 | DBG("Switching %s_%d", chan->channel_name, cpu); |
b73a4c47 | 1242 | } |
b5b073e2 PMF |
1243 | |
1244 | /* | |
1245 | * Populate new subbuffer. | |
1246 | */ | |
b73a4c47 PMF |
1247 | if (unlikely(offsets.begin_switch)) |
1248 | ltt_reserve_switch_new_subbuf(chan, buf, &offsets, tsc); | |
1249 | ||
1250 | if (unlikely(offsets.end_switch_current)) | |
1251 | ltt_reserve_end_switch_current(chan, buf, &offsets, tsc); | |
1252 | ||
1253 | *slot_size = offsets.size; | |
1254 | *buf_offset = offsets.begin + offsets.before_hdr_pad; | |
1255 | return 0; | |
b5b073e2 PMF |
1256 | } |
1257 | ||
b5b073e2 PMF |
1258 | static struct ltt_transport ust_relay_transport = { |
1259 | .name = "ustrelay", | |
1260 | .ops = { | |
1261 | .create_channel = ust_buffers_create_channel, | |
1262 | .finish_channel = ltt_relay_finish_channel, | |
1263 | .remove_channel = ltt_relay_remove_channel, | |
1264 | .wakeup_channel = ltt_relay_async_wakeup_chan, | |
b5b073e2 PMF |
1265 | }, |
1266 | }; | |
1267 | ||
b5b073e2 PMF |
1268 | static char initialized = 0; |
1269 | ||
1270 | void __attribute__((constructor)) init_ustrelay_transport(void) | |
1271 | { | |
1272 | if(!initialized) { | |
1273 | ltt_transport_register(&ust_relay_transport); | |
1274 | initialized = 1; | |
1275 | } | |
1276 | } | |
1277 | ||
b73a4c47 | 1278 | static void __attribute__((destructor)) ust_buffers_exit(void) |
b5b073e2 PMF |
1279 | { |
1280 | ltt_transport_unregister(&ust_relay_transport); | |
1281 | } | |
b73a4c47 | 1282 | |
12e81b07 | 1283 | size_t ltt_write_event_header_slow(struct ust_channel *channel, |
b73a4c47 PMF |
1284 | struct ust_buffer *buf, long buf_offset, |
1285 | u16 eID, u32 event_size, | |
1286 | u64 tsc, unsigned int rflags) | |
1287 | { | |
1288 | struct ltt_event_header header; | |
1289 | u16 small_size; | |
1290 | ||
1291 | switch (rflags) { | |
1292 | case LTT_RFLAG_ID_SIZE_TSC: | |
1293 | header.id_time = 29 << LTT_TSC_BITS; | |
1294 | break; | |
1295 | case LTT_RFLAG_ID_SIZE: | |
1296 | header.id_time = 30 << LTT_TSC_BITS; | |
1297 | break; | |
1298 | case LTT_RFLAG_ID: | |
1299 | header.id_time = 31 << LTT_TSC_BITS; | |
1300 | break; | |
1301 | } | |
1302 | ||
1303 | header.id_time |= (u32)tsc & LTT_TSC_MASK; | |
1304 | ust_buffers_write(buf, buf_offset, &header, sizeof(header)); | |
1305 | buf_offset += sizeof(header); | |
1306 | ||
1307 | switch (rflags) { | |
1308 | case LTT_RFLAG_ID_SIZE_TSC: | |
1309 | small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE); | |
1310 | ust_buffers_write(buf, buf_offset, | |
1311 | &eID, sizeof(u16)); | |
1312 | buf_offset += sizeof(u16); | |
1313 | ust_buffers_write(buf, buf_offset, | |
1314 | &small_size, sizeof(u16)); | |
1315 | buf_offset += sizeof(u16); | |
1316 | if (small_size == LTT_MAX_SMALL_SIZE) { | |
1317 | ust_buffers_write(buf, buf_offset, | |
1318 | &event_size, sizeof(u32)); | |
1319 | buf_offset += sizeof(u32); | |
1320 | } | |
1321 | buf_offset += ltt_align(buf_offset, sizeof(u64)); | |
1322 | ust_buffers_write(buf, buf_offset, | |
1323 | &tsc, sizeof(u64)); | |
1324 | buf_offset += sizeof(u64); | |
1325 | break; | |
1326 | case LTT_RFLAG_ID_SIZE: | |
1327 | small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE); | |
1328 | ust_buffers_write(buf, buf_offset, | |
1329 | &eID, sizeof(u16)); | |
1330 | buf_offset += sizeof(u16); | |
1331 | ust_buffers_write(buf, buf_offset, | |
1332 | &small_size, sizeof(u16)); | |
1333 | buf_offset += sizeof(u16); | |
1334 | if (small_size == LTT_MAX_SMALL_SIZE) { | |
1335 | ust_buffers_write(buf, buf_offset, | |
1336 | &event_size, sizeof(u32)); | |
1337 | buf_offset += sizeof(u32); | |
1338 | } | |
1339 | break; | |
1340 | case LTT_RFLAG_ID: | |
1341 | ust_buffers_write(buf, buf_offset, | |
1342 | &eID, sizeof(u16)); | |
1343 | buf_offset += sizeof(u16); | |
1344 | break; | |
1345 | } | |
1346 | ||
1347 | return buf_offset; | |
1348 | } |