| 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 | |
| 23 | #include <unistd.h> |
| 24 | #include <sys/mman.h> |
| 25 | #include <sys/ipc.h> |
| 26 | #include <sys/shm.h> |
| 27 | #include <fcntl.h> |
| 28 | #include <ust/kernelcompat.h> |
| 29 | #include <kcompat/kref.h> |
| 30 | #include "buffers.h" |
| 31 | #include "channels.h" |
| 32 | #include "tracer.h" |
| 33 | #include "tracercore.h" |
| 34 | #include "usterr.h" |
| 35 | |
| 36 | static DEFINE_MUTEX(ust_buffers_channels_mutex); |
| 37 | static LIST_HEAD(ust_buffers_channels); |
| 38 | |
| 39 | static int get_n_cpus(void) |
| 40 | { |
| 41 | int result; |
| 42 | static int n_cpus = 0; |
| 43 | |
| 44 | if(n_cpus) { |
| 45 | return n_cpus; |
| 46 | } |
| 47 | |
| 48 | result = sysconf(_SC_NPROCESSORS_ONLN); |
| 49 | if(result == -1) { |
| 50 | return -1; |
| 51 | } |
| 52 | |
| 53 | n_cpus = result; |
| 54 | |
| 55 | return result; |
| 56 | } |
| 57 | |
| 58 | static int ust_buffers_init_buffer(struct ltt_trace_struct *trace, |
| 59 | struct ust_channel *ltt_chan, |
| 60 | struct ust_buffer *buf, |
| 61 | unsigned int n_subbufs); |
| 62 | |
| 63 | static int ust_buffers_alloc_buf(struct ust_buffer *buf, size_t *size) |
| 64 | { |
| 65 | void *ptr; |
| 66 | int result; |
| 67 | |
| 68 | *size = PAGE_ALIGN(*size); |
| 69 | |
| 70 | result = buf->shmid = shmget(getpid(), *size, IPC_CREAT | IPC_EXCL | 0700); |
| 71 | if(result == -1 && errno == EINVAL) { |
| 72 | ERR("shmget() returned EINVAL; maybe /proc/sys/kernel/shmmax should be increased."); |
| 73 | return -1; |
| 74 | } |
| 75 | else if(result == -1) { |
| 76 | PERROR("shmget"); |
| 77 | return -1; |
| 78 | } |
| 79 | |
| 80 | /* FIXME: should have matching call to shmdt */ |
| 81 | ptr = shmat(buf->shmid, NULL, 0); |
| 82 | if(ptr == (void *) -1) { |
| 83 | perror("shmat"); |
| 84 | goto destroy_shmem; |
| 85 | } |
| 86 | |
| 87 | /* Already mark the shared memory for destruction. This will occur only |
| 88 | * when all users have detached. |
| 89 | */ |
| 90 | result = shmctl(buf->shmid, IPC_RMID, NULL); |
| 91 | if(result == -1) { |
| 92 | perror("shmctl"); |
| 93 | return -1; |
| 94 | } |
| 95 | |
| 96 | buf->buf_data = ptr; |
| 97 | buf->buf_size = *size; |
| 98 | |
| 99 | return 0; |
| 100 | |
| 101 | destroy_shmem: |
| 102 | result = shmctl(buf->shmid, IPC_RMID, NULL); |
| 103 | if(result == -1) { |
| 104 | perror("shmctl"); |
| 105 | } |
| 106 | |
| 107 | return -1; |
| 108 | } |
| 109 | |
| 110 | int ust_buffers_create_buf(struct ust_channel *channel, int cpu) |
| 111 | { |
| 112 | int result; |
| 113 | struct ust_buffer *buf = channel->buf[cpu]; |
| 114 | |
| 115 | buf->cpu = cpu; |
| 116 | result = ust_buffers_alloc_buf(buf, &channel->alloc_size); |
| 117 | if(result) |
| 118 | return -1; |
| 119 | |
| 120 | buf->chan = channel; |
| 121 | kref_get(&channel->kref); |
| 122 | return 0; |
| 123 | } |
| 124 | |
| 125 | static void ust_buffers_destroy_channel(struct kref *kref) |
| 126 | { |
| 127 | struct ust_channel *chan = container_of(kref, struct ust_channel, kref); |
| 128 | free(chan); |
| 129 | } |
| 130 | |
| 131 | static void ust_buffers_destroy_buf(struct ust_buffer *buf) |
| 132 | { |
| 133 | struct ust_channel *chan = buf->chan; |
| 134 | int result; |
| 135 | |
| 136 | result = munmap(buf->buf_data, buf->buf_size); |
| 137 | if(result == -1) { |
| 138 | PERROR("munmap"); |
| 139 | } |
| 140 | |
| 141 | //ust// chan->buf[buf->cpu] = NULL; |
| 142 | free(buf); |
| 143 | kref_put(&chan->kref, ust_buffers_destroy_channel); |
| 144 | } |
| 145 | |
| 146 | /* called from kref_put */ |
| 147 | static void ust_buffers_remove_buf(struct kref *kref) |
| 148 | { |
| 149 | struct ust_buffer *buf = container_of(kref, struct ust_buffer, kref); |
| 150 | ust_buffers_destroy_buf(buf); |
| 151 | } |
| 152 | |
| 153 | int ust_buffers_open_buf(struct ust_channel *chan, int cpu) |
| 154 | { |
| 155 | int result; |
| 156 | |
| 157 | result = ust_buffers_create_buf(chan, cpu); |
| 158 | if (result == -1) |
| 159 | return -1; |
| 160 | |
| 161 | kref_init(&chan->buf[cpu]->kref); |
| 162 | |
| 163 | result = ust_buffers_init_buffer(chan->trace, chan, chan->buf[cpu], chan->subbuf_cnt); |
| 164 | if(result == -1) |
| 165 | return -1; |
| 166 | |
| 167 | return 0; |
| 168 | |
| 169 | /* FIXME: decrementally destroy on error? */ |
| 170 | } |
| 171 | |
| 172 | /** |
| 173 | * ust_buffers_close_buf - close a channel buffer |
| 174 | * @buf: buffer |
| 175 | */ |
| 176 | static void ust_buffers_close_buf(struct ust_buffer *buf) |
| 177 | { |
| 178 | kref_put(&buf->kref, ust_buffers_remove_buf); |
| 179 | } |
| 180 | |
| 181 | int ust_buffers_channel_open(struct ust_channel *chan, size_t subbuf_size, size_t subbuf_cnt) |
| 182 | { |
| 183 | int i; |
| 184 | int result; |
| 185 | |
| 186 | if(subbuf_size == 0 || subbuf_cnt == 0) |
| 187 | return -1; |
| 188 | |
| 189 | chan->version = UST_CHANNEL_VERSION; |
| 190 | chan->subbuf_cnt = subbuf_cnt; |
| 191 | chan->subbuf_size = subbuf_size; |
| 192 | chan->subbuf_size_order = get_count_order(subbuf_size); |
| 193 | chan->alloc_size = FIX_SIZE(subbuf_size * subbuf_cnt); |
| 194 | |
| 195 | kref_init(&chan->kref); |
| 196 | |
| 197 | mutex_lock(&ust_buffers_channels_mutex); |
| 198 | for(i=0; i<chan->n_cpus; i++) { |
| 199 | result = ust_buffers_open_buf(chan, i); |
| 200 | if (result == -1) |
| 201 | goto error; |
| 202 | } |
| 203 | list_add(&chan->list, &ust_buffers_channels); |
| 204 | mutex_unlock(&ust_buffers_channels_mutex); |
| 205 | |
| 206 | return 0; |
| 207 | |
| 208 | /* Jump directly inside the loop to close the buffers that were already |
| 209 | * opened. */ |
| 210 | for(; i>=0; i--) { |
| 211 | ust_buffers_close_buf(chan->buf[i]); |
| 212 | error: |
| 213 | } |
| 214 | |
| 215 | kref_put(&chan->kref, ust_buffers_destroy_channel); |
| 216 | mutex_unlock(&ust_buffers_channels_mutex); |
| 217 | return -1; |
| 218 | } |
| 219 | |
| 220 | void ust_buffers_channel_close(struct ust_channel *chan) |
| 221 | { |
| 222 | int i; |
| 223 | if(!chan) |
| 224 | return; |
| 225 | |
| 226 | mutex_lock(&ust_buffers_channels_mutex); |
| 227 | for(i=0; i<chan->n_cpus; i++) { |
| 228 | /* FIXME: if we make it here, then all buffers were necessarily allocated. Moreover, we don't |
| 229 | * initialize to NULL so we cannot use this check. Should we? */ |
| 230 | //ust// if (chan->buf[i]) |
| 231 | ust_buffers_close_buf(chan->buf[i]); |
| 232 | } |
| 233 | |
| 234 | list_del(&chan->list); |
| 235 | kref_put(&chan->kref, ust_buffers_destroy_channel); |
| 236 | mutex_unlock(&ust_buffers_channels_mutex); |
| 237 | } |
| 238 | |
| 239 | /* _ust_buffers_write() |
| 240 | * |
| 241 | * @buf: destination buffer |
| 242 | * @offset: offset in destination |
| 243 | * @src: source buffer |
| 244 | * @len: length of source |
| 245 | * @cpy: already copied |
| 246 | */ |
| 247 | |
| 248 | void _ust_buffers_write(struct ust_buffer *buf, size_t offset, |
| 249 | const void *src, size_t len, ssize_t cpy) |
| 250 | { |
| 251 | do { |
| 252 | len -= cpy; |
| 253 | src += cpy; |
| 254 | offset += cpy; |
| 255 | |
| 256 | WARN_ON(offset >= buf->buf_size); |
| 257 | |
| 258 | cpy = min_t(size_t, len, buf->buf_size - offset); |
| 259 | ust_buffers_do_copy(buf->buf_data + offset, src, cpy); |
| 260 | } while (unlikely(len != cpy)); |
| 261 | } |
| 262 | |
| 263 | void *ltt_buffers_offset_address(struct ust_buffer *buf, size_t offset) |
| 264 | { |
| 265 | return ((char *)buf->buf_data)+offset; |
| 266 | } |
| 267 | |
| 268 | /* |
| 269 | * ------- |
| 270 | */ |
| 271 | |
| 272 | /* |
| 273 | * Last TSC comparison functions. Check if the current TSC overflows |
| 274 | * LTT_TSC_BITS bits from the last TSC read. Reads and writes last_tsc |
| 275 | * atomically. |
| 276 | */ |
| 277 | |
| 278 | /* FIXME: does this test work properly? */ |
| 279 | #if (BITS_PER_LONG == 32) |
| 280 | static inline void save_last_tsc(struct ust_buffer *ltt_buf, |
| 281 | u64 tsc) |
| 282 | { |
| 283 | ltt_buf->last_tsc = (unsigned long)(tsc >> LTT_TSC_BITS); |
| 284 | } |
| 285 | |
| 286 | static inline int last_tsc_overflow(struct ust_buffer *ltt_buf, |
| 287 | u64 tsc) |
| 288 | { |
| 289 | unsigned long tsc_shifted = (unsigned long)(tsc >> LTT_TSC_BITS); |
| 290 | |
| 291 | if (unlikely((tsc_shifted - ltt_buf->last_tsc))) |
| 292 | return 1; |
| 293 | else |
| 294 | return 0; |
| 295 | } |
| 296 | #else |
| 297 | static inline void save_last_tsc(struct ust_buffer *ltt_buf, |
| 298 | u64 tsc) |
| 299 | { |
| 300 | ltt_buf->last_tsc = (unsigned long)tsc; |
| 301 | } |
| 302 | |
| 303 | static inline int last_tsc_overflow(struct ust_buffer *ltt_buf, |
| 304 | u64 tsc) |
| 305 | { |
| 306 | if (unlikely((tsc - ltt_buf->last_tsc) >> LTT_TSC_BITS)) |
| 307 | return 1; |
| 308 | else |
| 309 | return 0; |
| 310 | } |
| 311 | #endif |
| 312 | |
| 313 | /* |
| 314 | * A switch is done during tracing or as a final flush after tracing (so it |
| 315 | * won't write in the new sub-buffer). |
| 316 | */ |
| 317 | enum force_switch_mode { FORCE_ACTIVE, FORCE_FLUSH }; |
| 318 | |
| 319 | static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu); |
| 320 | |
| 321 | static void ltt_force_switch(struct ust_buffer *buf, |
| 322 | enum force_switch_mode mode); |
| 323 | |
| 324 | /* |
| 325 | * Trace callbacks |
| 326 | */ |
| 327 | static void ltt_buffer_begin_callback(struct ust_buffer *buf, |
| 328 | u64 tsc, unsigned int subbuf_idx) |
| 329 | { |
| 330 | struct ust_channel *channel = buf->chan; |
| 331 | struct ltt_subbuffer_header *header = |
| 332 | (struct ltt_subbuffer_header *) |
| 333 | ltt_buffers_offset_address(buf, |
| 334 | subbuf_idx * buf->chan->subbuf_size); |
| 335 | |
| 336 | header->cycle_count_begin = tsc; |
| 337 | header->lost_size = 0xFFFFFFFF; /* for debugging */ |
| 338 | header->buf_size = buf->chan->subbuf_size; |
| 339 | ltt_write_trace_header(channel->trace, header); |
| 340 | } |
| 341 | |
| 342 | /* |
| 343 | * offset is assumed to never be 0 here : never deliver a completely empty |
| 344 | * subbuffer. The lost size is between 0 and subbuf_size-1. |
| 345 | */ |
| 346 | static notrace void ltt_buffer_end_callback(struct ust_buffer *buf, |
| 347 | u64 tsc, unsigned int offset, unsigned int subbuf_idx) |
| 348 | { |
| 349 | struct ltt_subbuffer_header *header = |
| 350 | (struct ltt_subbuffer_header *) |
| 351 | ltt_buffers_offset_address(buf, |
| 352 | subbuf_idx * buf->chan->subbuf_size); |
| 353 | |
| 354 | header->lost_size = SUBBUF_OFFSET((buf->chan->subbuf_size - offset), |
| 355 | buf->chan); |
| 356 | header->cycle_count_end = tsc; |
| 357 | header->events_lost = local_read(&buf->events_lost); |
| 358 | header->subbuf_corrupt = local_read(&buf->corrupted_subbuffers); |
| 359 | |
| 360 | } |
| 361 | |
| 362 | void (*wake_consumer)(void *, int) = NULL; |
| 363 | |
| 364 | void relay_set_wake_consumer(void (*wake)(void *, int)) |
| 365 | { |
| 366 | wake_consumer = wake; |
| 367 | } |
| 368 | |
| 369 | void relay_wake_consumer(void *arg, int finished) |
| 370 | { |
| 371 | if(wake_consumer) |
| 372 | wake_consumer(arg, finished); |
| 373 | } |
| 374 | |
| 375 | static notrace void ltt_deliver(struct ust_buffer *buf, unsigned int subbuf_idx, |
| 376 | long commit_count) |
| 377 | { |
| 378 | int result; |
| 379 | |
| 380 | //ust// #ifdef CONFIG_LTT_VMCORE |
| 381 | local_set(&buf->commit_seq[subbuf_idx], commit_count); |
| 382 | //ust// #endif |
| 383 | |
| 384 | /* wakeup consumer */ |
| 385 | result = write(buf->data_ready_fd_write, "1", 1); |
| 386 | if(result == -1) { |
| 387 | PERROR("write (in ltt_relay_buffer_flush)"); |
| 388 | ERR("this should never happen!"); |
| 389 | } |
| 390 | //ust// atomic_set(<t_buf->wakeup_readers, 1); |
| 391 | } |
| 392 | |
| 393 | /* |
| 394 | * This function should not be called from NMI interrupt context |
| 395 | */ |
| 396 | static notrace void ltt_buf_unfull(struct ust_buffer *buf, |
| 397 | unsigned int subbuf_idx, |
| 398 | long offset) |
| 399 | { |
| 400 | //ust// struct ltt_channel_struct *ltt_channel = |
| 401 | //ust// (struct ltt_channel_struct *)buf->chan->private_data; |
| 402 | //ust// struct ltt_channel_buf_struct *ltt_buf = ltt_channel->buf; |
| 403 | //ust// |
| 404 | //ust// ltt_relay_wake_writers(ltt_buf); |
| 405 | } |
| 406 | |
| 407 | int ust_buffers_do_get_subbuf(struct ust_buffer *buf, long *pconsumed_old) |
| 408 | { |
| 409 | struct ust_channel *channel = buf->chan; |
| 410 | long consumed_old, consumed_idx, commit_count, write_offset; |
| 411 | consumed_old = atomic_long_read(&buf->consumed); |
| 412 | consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan); |
| 413 | commit_count = local_read(&buf->commit_count[consumed_idx]); |
| 414 | /* |
| 415 | * Make sure we read the commit count before reading the buffer |
| 416 | * data and the write offset. Correct consumed offset ordering |
| 417 | * wrt commit count is insured by the use of cmpxchg to update |
| 418 | * the consumed offset. |
| 419 | */ |
| 420 | smp_rmb(); |
| 421 | write_offset = local_read(&buf->offset); |
| 422 | /* |
| 423 | * Check that the subbuffer we are trying to consume has been |
| 424 | * already fully committed. |
| 425 | */ |
| 426 | if (((commit_count - buf->chan->subbuf_size) |
| 427 | & channel->commit_count_mask) |
| 428 | - (BUFFER_TRUNC(consumed_old, buf->chan) |
| 429 | >> channel->n_subbufs_order) |
| 430 | != 0) { |
| 431 | return -EAGAIN; |
| 432 | } |
| 433 | /* |
| 434 | * Check that we are not about to read the same subbuffer in |
| 435 | * which the writer head is. |
| 436 | */ |
| 437 | if ((SUBBUF_TRUNC(write_offset, buf->chan) |
| 438 | - SUBBUF_TRUNC(consumed_old, buf->chan)) |
| 439 | == 0) { |
| 440 | return -EAGAIN; |
| 441 | } |
| 442 | |
| 443 | *pconsumed_old = consumed_old; |
| 444 | return 0; |
| 445 | } |
| 446 | |
| 447 | int ust_buffers_do_put_subbuf(struct ust_buffer *buf, u32 uconsumed_old) |
| 448 | { |
| 449 | long consumed_new, consumed_old; |
| 450 | |
| 451 | consumed_old = atomic_long_read(&buf->consumed); |
| 452 | consumed_old = consumed_old & (~0xFFFFFFFFL); |
| 453 | consumed_old = consumed_old | uconsumed_old; |
| 454 | consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan); |
| 455 | |
| 456 | //ust// spin_lock(<t_buf->full_lock); |
| 457 | if (atomic_long_cmpxchg(&buf->consumed, consumed_old, |
| 458 | consumed_new) |
| 459 | != consumed_old) { |
| 460 | /* We have been pushed by the writer : the last |
| 461 | * buffer read _is_ corrupted! It can also |
| 462 | * happen if this is a buffer we never got. */ |
| 463 | //ust// spin_unlock(<t_buf->full_lock); |
| 464 | return -EIO; |
| 465 | } else { |
| 466 | /* tell the client that buffer is now unfull */ |
| 467 | int index; |
| 468 | long data; |
| 469 | index = SUBBUF_INDEX(consumed_old, buf->chan); |
| 470 | data = BUFFER_OFFSET(consumed_old, buf->chan); |
| 471 | ltt_buf_unfull(buf, index, data); |
| 472 | //ust// spin_unlock(<t_buf->full_lock); |
| 473 | } |
| 474 | return 0; |
| 475 | } |
| 476 | |
| 477 | static void ltt_relay_print_subbuffer_errors( |
| 478 | struct ust_channel *channel, |
| 479 | long cons_off, int cpu) |
| 480 | { |
| 481 | struct ust_buffer *ltt_buf = channel->buf[cpu]; |
| 482 | long cons_idx, commit_count, write_offset; |
| 483 | |
| 484 | cons_idx = SUBBUF_INDEX(cons_off, channel); |
| 485 | commit_count = local_read(<t_buf->commit_count[cons_idx]); |
| 486 | /* |
| 487 | * No need to order commit_count and write_offset reads because we |
| 488 | * execute after trace is stopped when there are no readers left. |
| 489 | */ |
| 490 | write_offset = local_read(<t_buf->offset); |
| 491 | WARN( "LTT : unread channel %s offset is %ld " |
| 492 | "and cons_off : %ld\n", |
| 493 | channel->channel_name, write_offset, cons_off); |
| 494 | /* Check each sub-buffer for non filled commit count */ |
| 495 | if (((commit_count - channel->subbuf_size) & channel->commit_count_mask) |
| 496 | - (BUFFER_TRUNC(cons_off, channel) >> channel->n_subbufs_order) != 0) { |
| 497 | ERR("LTT : %s : subbuffer %lu has non filled " |
| 498 | "commit count %lu.\n", |
| 499 | channel->channel_name, cons_idx, commit_count); |
| 500 | } |
| 501 | ERR("LTT : %s : commit count : %lu, subbuf size %zd\n", |
| 502 | channel->channel_name, commit_count, |
| 503 | channel->subbuf_size); |
| 504 | } |
| 505 | |
| 506 | static void ltt_relay_print_errors(struct ltt_trace_struct *trace, |
| 507 | struct ust_channel *channel, int cpu) |
| 508 | { |
| 509 | struct ust_buffer *ltt_buf = channel->buf[cpu]; |
| 510 | long cons_off; |
| 511 | |
| 512 | /* |
| 513 | * Can be called in the error path of allocation when |
| 514 | * trans_channel_data is not yet set. |
| 515 | */ |
| 516 | if (!channel) |
| 517 | return; |
| 518 | |
| 519 | for (cons_off = atomic_long_read(<t_buf->consumed); |
| 520 | (SUBBUF_TRUNC(local_read(<t_buf->offset), |
| 521 | channel) |
| 522 | - cons_off) > 0; |
| 523 | cons_off = SUBBUF_ALIGN(cons_off, channel)) |
| 524 | ltt_relay_print_subbuffer_errors(channel, cons_off, cpu); |
| 525 | } |
| 526 | |
| 527 | static void ltt_relay_print_buffer_errors(struct ust_channel *channel, int cpu) |
| 528 | { |
| 529 | struct ltt_trace_struct *trace = channel->trace; |
| 530 | struct ust_buffer *ltt_buf = channel->buf[cpu]; |
| 531 | |
| 532 | if (local_read(<t_buf->events_lost)) |
| 533 | ERR("channel %s: %ld events lost", |
| 534 | channel->channel_name, |
| 535 | local_read(<t_buf->events_lost)); |
| 536 | if (local_read(<t_buf->corrupted_subbuffers)) |
| 537 | ERR("channel %s : %ld corrupted subbuffers", |
| 538 | channel->channel_name, |
| 539 | local_read(<t_buf->corrupted_subbuffers)); |
| 540 | |
| 541 | ltt_relay_print_errors(trace, channel, cpu); |
| 542 | } |
| 543 | |
| 544 | static void ltt_relay_release_channel(struct kref *kref) |
| 545 | { |
| 546 | struct ust_channel *ltt_chan = container_of(kref, |
| 547 | struct ust_channel, kref); |
| 548 | free(ltt_chan->buf); |
| 549 | } |
| 550 | |
| 551 | /* |
| 552 | * Create ltt buffer. |
| 553 | */ |
| 554 | //ust// static int ltt_relay_create_buffer(struct ltt_trace_struct *trace, |
| 555 | //ust// struct ltt_channel_struct *ltt_chan, struct rchan_buf *buf, |
| 556 | //ust// unsigned int cpu, unsigned int n_subbufs) |
| 557 | //ust// { |
| 558 | //ust// struct ltt_channel_buf_struct *ltt_buf = |
| 559 | //ust// percpu_ptr(ltt_chan->buf, cpu); |
| 560 | //ust// unsigned int j; |
| 561 | //ust// |
| 562 | //ust// ltt_buf->commit_count = |
| 563 | //ust// kzalloc_node(sizeof(ltt_buf->commit_count) * n_subbufs, |
| 564 | //ust// GFP_KERNEL, cpu_to_node(cpu)); |
| 565 | //ust// if (!ltt_buf->commit_count) |
| 566 | //ust// return -ENOMEM; |
| 567 | //ust// kref_get(&trace->kref); |
| 568 | //ust// kref_get(&trace->ltt_transport_kref); |
| 569 | //ust// kref_get(<t_chan->kref); |
| 570 | //ust// local_set(<t_buf->offset, ltt_subbuffer_header_size()); |
| 571 | //ust// atomic_long_set(<t_buf->consumed, 0); |
| 572 | //ust// atomic_long_set(<t_buf->active_readers, 0); |
| 573 | //ust// for (j = 0; j < n_subbufs; j++) |
| 574 | //ust// local_set(<t_buf->commit_count[j], 0); |
| 575 | //ust// init_waitqueue_head(<t_buf->write_wait); |
| 576 | //ust// atomic_set(<t_buf->wakeup_readers, 0); |
| 577 | //ust// spin_lock_init(<t_buf->full_lock); |
| 578 | //ust// |
| 579 | //ust// ltt_buffer_begin_callback(buf, trace->start_tsc, 0); |
| 580 | //ust// /* atomic_add made on local variable on data that belongs to |
| 581 | //ust// * various CPUs : ok because tracing not started (for this cpu). */ |
| 582 | //ust// local_add(ltt_subbuffer_header_size(), <t_buf->commit_count[0]); |
| 583 | //ust// |
| 584 | //ust// local_set(<t_buf->events_lost, 0); |
| 585 | //ust// local_set(<t_buf->corrupted_subbuffers, 0); |
| 586 | //ust// |
| 587 | //ust// return 0; |
| 588 | //ust// } |
| 589 | |
| 590 | static int ust_buffers_init_buffer(struct ltt_trace_struct *trace, |
| 591 | struct ust_channel *ltt_chan, struct ust_buffer *buf, |
| 592 | unsigned int n_subbufs) |
| 593 | { |
| 594 | unsigned int j; |
| 595 | int fds[2]; |
| 596 | int result; |
| 597 | |
| 598 | buf->commit_count = |
| 599 | zmalloc(sizeof(buf->commit_count) * n_subbufs); |
| 600 | if (!buf->commit_count) |
| 601 | return -ENOMEM; |
| 602 | kref_get(&trace->kref); |
| 603 | kref_get(&trace->ltt_transport_kref); |
| 604 | kref_get(<t_chan->kref); |
| 605 | local_set(&buf->offset, ltt_subbuffer_header_size()); |
| 606 | atomic_long_set(&buf->consumed, 0); |
| 607 | atomic_long_set(&buf->active_readers, 0); |
| 608 | for (j = 0; j < n_subbufs; j++) |
| 609 | local_set(&buf->commit_count[j], 0); |
| 610 | //ust// init_waitqueue_head(&buf->write_wait); |
| 611 | //ust// atomic_set(&buf->wakeup_readers, 0); |
| 612 | //ust// spin_lock_init(&buf->full_lock); |
| 613 | |
| 614 | ltt_buffer_begin_callback(buf, trace->start_tsc, 0); |
| 615 | |
| 616 | local_add(ltt_subbuffer_header_size(), &buf->commit_count[0]); |
| 617 | |
| 618 | local_set(&buf->events_lost, 0); |
| 619 | local_set(&buf->corrupted_subbuffers, 0); |
| 620 | |
| 621 | result = pipe(fds); |
| 622 | if(result == -1) { |
| 623 | PERROR("pipe"); |
| 624 | return -1; |
| 625 | } |
| 626 | buf->data_ready_fd_read = fds[0]; |
| 627 | buf->data_ready_fd_write = fds[1]; |
| 628 | |
| 629 | /* FIXME: do we actually need this? */ |
| 630 | result = fcntl(fds[0], F_SETFL, O_NONBLOCK); |
| 631 | if(result == -1) { |
| 632 | PERROR("fcntl"); |
| 633 | } |
| 634 | |
| 635 | //ust// buf->commit_seq = malloc(sizeof(buf->commit_seq) * n_subbufs); |
| 636 | //ust// if(!ltt_buf->commit_seq) { |
| 637 | //ust// return -1; |
| 638 | //ust// } |
| 639 | |
| 640 | /* FIXME: decrementally destroy on error */ |
| 641 | |
| 642 | return 0; |
| 643 | } |
| 644 | |
| 645 | /* FIXME: use this function */ |
| 646 | static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu) |
| 647 | { |
| 648 | struct ltt_trace_struct *trace = ltt_chan->trace; |
| 649 | struct ust_buffer *ltt_buf = ltt_chan->buf[cpu]; |
| 650 | |
| 651 | kref_put(<t_chan->trace->ltt_transport_kref, |
| 652 | ltt_release_transport); |
| 653 | ltt_relay_print_buffer_errors(ltt_chan, cpu); |
| 654 | //ust// free(ltt_buf->commit_seq); |
| 655 | kfree(ltt_buf->commit_count); |
| 656 | ltt_buf->commit_count = NULL; |
| 657 | kref_put(<t_chan->kref, ltt_relay_release_channel); |
| 658 | kref_put(&trace->kref, ltt_release_trace); |
| 659 | //ust// wake_up_interruptible(&trace->kref_wq); |
| 660 | } |
| 661 | |
| 662 | static int ust_buffers_alloc_channel_buf_structs(struct ust_channel *chan) |
| 663 | { |
| 664 | void *ptr; |
| 665 | int result; |
| 666 | size_t size; |
| 667 | int i; |
| 668 | |
| 669 | size = PAGE_ALIGN(1); |
| 670 | |
| 671 | for(i=0; i<chan->n_cpus; i++) { |
| 672 | |
| 673 | result = chan->buf_struct_shmids[i] = shmget(getpid(), size, IPC_CREAT | IPC_EXCL | 0700); |
| 674 | if(result == -1) { |
| 675 | PERROR("shmget"); |
| 676 | goto destroy_previous; |
| 677 | } |
| 678 | |
| 679 | /* FIXME: should have matching call to shmdt */ |
| 680 | ptr = shmat(chan->buf_struct_shmids[i], NULL, 0); |
| 681 | if(ptr == (void *) -1) { |
| 682 | perror("shmat"); |
| 683 | goto destroy_shm; |
| 684 | } |
| 685 | |
| 686 | /* Already mark the shared memory for destruction. This will occur only |
| 687 | * when all users have detached. |
| 688 | */ |
| 689 | result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL); |
| 690 | if(result == -1) { |
| 691 | perror("shmctl"); |
| 692 | goto destroy_previous; |
| 693 | } |
| 694 | |
| 695 | chan->buf[i] = ptr; |
| 696 | } |
| 697 | |
| 698 | return 0; |
| 699 | |
| 700 | /* Jumping inside this loop occurs from within the other loop above with i as |
| 701 | * counter, so it unallocates the structures for the cpu = current_i down to |
| 702 | * zero. */ |
| 703 | for(; i>=0; i--) { |
| 704 | destroy_shm: |
| 705 | result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL); |
| 706 | if(result == -1) { |
| 707 | perror("shmctl"); |
| 708 | } |
| 709 | |
| 710 | destroy_previous: |
| 711 | continue; |
| 712 | } |
| 713 | |
| 714 | return -1; |
| 715 | } |
| 716 | |
| 717 | /* |
| 718 | * Create channel. |
| 719 | */ |
| 720 | static int ust_buffers_create_channel(const char *trace_name, struct ltt_trace_struct *trace, |
| 721 | const char *channel_name, struct ust_channel *ltt_chan, |
| 722 | unsigned int subbuf_size, unsigned int n_subbufs, int overwrite) |
| 723 | { |
| 724 | int result; |
| 725 | |
| 726 | kref_init(<t_chan->kref); |
| 727 | |
| 728 | ltt_chan->trace = trace; |
| 729 | ltt_chan->buffer_begin = ltt_buffer_begin_callback; |
| 730 | ltt_chan->buffer_end = ltt_buffer_end_callback; |
| 731 | ltt_chan->overwrite = overwrite; |
| 732 | ltt_chan->n_subbufs_order = get_count_order(n_subbufs); |
| 733 | ltt_chan->commit_count_mask = (~0UL >> ltt_chan->n_subbufs_order); |
| 734 | ltt_chan->n_cpus = get_n_cpus(); |
| 735 | //ust// ltt_chan->buf = percpu_alloc_mask(sizeof(struct ltt_channel_buf_struct), GFP_KERNEL, cpu_possible_map); |
| 736 | ltt_chan->buf = (void *) malloc(ltt_chan->n_cpus * sizeof(void *)); |
| 737 | if(ltt_chan->buf == NULL) { |
| 738 | goto error; |
| 739 | } |
| 740 | ltt_chan->buf_struct_shmids = (int *) malloc(ltt_chan->n_cpus * sizeof(int)); |
| 741 | if(ltt_chan->buf_struct_shmids == NULL) |
| 742 | goto free_buf; |
| 743 | |
| 744 | result = ust_buffers_alloc_channel_buf_structs(ltt_chan); |
| 745 | if(result != 0) { |
| 746 | goto free_buf_struct_shmids; |
| 747 | } |
| 748 | |
| 749 | result = ust_buffers_channel_open(ltt_chan, subbuf_size, n_subbufs); |
| 750 | if (result != 0) { |
| 751 | ERR("Cannot open channel for trace %s", trace_name); |
| 752 | goto unalloc_buf_structs; |
| 753 | } |
| 754 | |
| 755 | return 0; |
| 756 | |
| 757 | unalloc_buf_structs: |
| 758 | /* FIXME: put a call here to unalloc the buf structs! */ |
| 759 | |
| 760 | free_buf_struct_shmids: |
| 761 | free(ltt_chan->buf_struct_shmids); |
| 762 | |
| 763 | free_buf: |
| 764 | free(ltt_chan->buf); |
| 765 | |
| 766 | error: |
| 767 | return -1; |
| 768 | } |
| 769 | |
| 770 | /* |
| 771 | * LTTng channel flush function. |
| 772 | * |
| 773 | * Must be called when no tracing is active in the channel, because of |
| 774 | * accesses across CPUs. |
| 775 | */ |
| 776 | static notrace void ltt_relay_buffer_flush(struct ust_buffer *buf) |
| 777 | { |
| 778 | int result; |
| 779 | |
| 780 | //ust// buf->finalized = 1; |
| 781 | ltt_force_switch(buf, FORCE_FLUSH); |
| 782 | |
| 783 | result = write(buf->data_ready_fd_write, "1", 1); |
| 784 | if(result == -1) { |
| 785 | PERROR("write (in ltt_relay_buffer_flush)"); |
| 786 | ERR("this should never happen!"); |
| 787 | } |
| 788 | } |
| 789 | |
| 790 | static void ltt_relay_async_wakeup_chan(struct ust_channel *ltt_channel) |
| 791 | { |
| 792 | //ust// unsigned int i; |
| 793 | //ust// struct rchan *rchan = ltt_channel->trans_channel_data; |
| 794 | //ust// |
| 795 | //ust// for_each_possible_cpu(i) { |
| 796 | //ust// struct ltt_channel_buf_struct *ltt_buf = |
| 797 | //ust// percpu_ptr(ltt_channel->buf, i); |
| 798 | //ust// |
| 799 | //ust// if (atomic_read(<t_buf->wakeup_readers) == 1) { |
| 800 | //ust// atomic_set(<t_buf->wakeup_readers, 0); |
| 801 | //ust// wake_up_interruptible(&rchan->buf[i]->read_wait); |
| 802 | //ust// } |
| 803 | //ust// } |
| 804 | } |
| 805 | |
| 806 | static void ltt_relay_finish_buffer(struct ust_channel *channel, unsigned int cpu) |
| 807 | { |
| 808 | // int result; |
| 809 | |
| 810 | if (channel->buf[cpu]) { |
| 811 | struct ust_buffer *buf = channel->buf[cpu]; |
| 812 | ltt_relay_buffer_flush(buf); |
| 813 | //ust// ltt_relay_wake_writers(ltt_buf); |
| 814 | /* closing the pipe tells the consumer the buffer is finished */ |
| 815 | |
| 816 | //result = write(ltt_buf->data_ready_fd_write, "D", 1); |
| 817 | //if(result == -1) { |
| 818 | // PERROR("write (in ltt_relay_finish_buffer)"); |
| 819 | // ERR("this should never happen!"); |
| 820 | //} |
| 821 | close(buf->data_ready_fd_write); |
| 822 | } |
| 823 | } |
| 824 | |
| 825 | |
| 826 | static void ltt_relay_finish_channel(struct ust_channel *channel) |
| 827 | { |
| 828 | unsigned int i; |
| 829 | |
| 830 | for(i=0; i<channel->n_cpus; i++) { |
| 831 | ltt_relay_finish_buffer(channel, i); |
| 832 | } |
| 833 | } |
| 834 | |
| 835 | static void ltt_relay_remove_channel(struct ust_channel *channel) |
| 836 | { |
| 837 | ust_buffers_channel_close(channel); |
| 838 | kref_put(&channel->kref, ltt_relay_release_channel); |
| 839 | } |
| 840 | |
| 841 | struct ltt_reserve_switch_offsets { |
| 842 | long begin, end, old; |
| 843 | long begin_switch, end_switch_current, end_switch_old; |
| 844 | long commit_count, reserve_commit_diff; |
| 845 | size_t before_hdr_pad, size; |
| 846 | }; |
| 847 | |
| 848 | /* |
| 849 | * Returns : |
| 850 | * 0 if ok |
| 851 | * !0 if execution must be aborted. |
| 852 | */ |
| 853 | static inline int ltt_relay_try_reserve( |
| 854 | struct ust_channel *channel, struct ust_buffer *buf, |
| 855 | struct ltt_reserve_switch_offsets *offsets, size_t data_size, |
| 856 | u64 *tsc, unsigned int *rflags, int largest_align) |
| 857 | { |
| 858 | offsets->begin = local_read(&buf->offset); |
| 859 | offsets->old = offsets->begin; |
| 860 | offsets->begin_switch = 0; |
| 861 | offsets->end_switch_current = 0; |
| 862 | offsets->end_switch_old = 0; |
| 863 | |
| 864 | *tsc = trace_clock_read64(); |
| 865 | if (last_tsc_overflow(buf, *tsc)) |
| 866 | *rflags = LTT_RFLAG_ID_SIZE_TSC; |
| 867 | |
| 868 | if (SUBBUF_OFFSET(offsets->begin, buf->chan) == 0) { |
| 869 | offsets->begin_switch = 1; /* For offsets->begin */ |
| 870 | } else { |
| 871 | offsets->size = ust_get_header_size(channel, |
| 872 | offsets->begin, data_size, |
| 873 | &offsets->before_hdr_pad, *rflags); |
| 874 | offsets->size += ltt_align(offsets->begin + offsets->size, |
| 875 | largest_align) |
| 876 | + data_size; |
| 877 | if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size) |
| 878 | > buf->chan->subbuf_size) { |
| 879 | offsets->end_switch_old = 1; /* For offsets->old */ |
| 880 | offsets->begin_switch = 1; /* For offsets->begin */ |
| 881 | } |
| 882 | } |
| 883 | if (offsets->begin_switch) { |
| 884 | long subbuf_index; |
| 885 | |
| 886 | if (offsets->end_switch_old) |
| 887 | offsets->begin = SUBBUF_ALIGN(offsets->begin, |
| 888 | buf->chan); |
| 889 | offsets->begin = offsets->begin + ltt_subbuffer_header_size(); |
| 890 | /* Test new buffer integrity */ |
| 891 | subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); |
| 892 | offsets->reserve_commit_diff = |
| 893 | (BUFFER_TRUNC(offsets->begin, buf->chan) |
| 894 | >> channel->n_subbufs_order) |
| 895 | - (local_read(&buf->commit_count[subbuf_index]) |
| 896 | & channel->commit_count_mask); |
| 897 | if (offsets->reserve_commit_diff == 0) { |
| 898 | long consumed; |
| 899 | |
| 900 | consumed = atomic_long_read(&buf->consumed); |
| 901 | |
| 902 | /* Next buffer not corrupted. */ |
| 903 | if (!channel->overwrite && |
| 904 | (SUBBUF_TRUNC(offsets->begin, buf->chan) |
| 905 | - SUBBUF_TRUNC(consumed, buf->chan)) |
| 906 | >= channel->alloc_size) { |
| 907 | |
| 908 | long consumed_idx = SUBBUF_INDEX(consumed, buf->chan); |
| 909 | long commit_count = local_read(&buf->commit_count[consumed_idx]); |
| 910 | if(((commit_count - buf->chan->subbuf_size) & channel->commit_count_mask) - (BUFFER_TRUNC(consumed, buf->chan) >> channel->n_subbufs_order) != 0) { |
| 911 | WARN("Event dropped. Caused by non-committed event."); |
| 912 | } |
| 913 | else { |
| 914 | WARN("Event dropped. Caused by non-consumed buffer."); |
| 915 | } |
| 916 | /* |
| 917 | * We do not overwrite non consumed buffers |
| 918 | * and we are full : event is lost. |
| 919 | */ |
| 920 | local_inc(&buf->events_lost); |
| 921 | return -1; |
| 922 | } else { |
| 923 | /* |
| 924 | * next buffer not corrupted, we are either in |
| 925 | * overwrite mode or the buffer is not full. |
| 926 | * It's safe to write in this new subbuffer. |
| 927 | */ |
| 928 | } |
| 929 | } else { |
| 930 | /* |
| 931 | * Next subbuffer corrupted. Force pushing reader even |
| 932 | * in normal mode. It's safe to write in this new |
| 933 | * subbuffer. |
| 934 | */ |
| 935 | } |
| 936 | offsets->size = ust_get_header_size(channel, |
| 937 | offsets->begin, data_size, |
| 938 | &offsets->before_hdr_pad, *rflags); |
| 939 | offsets->size += ltt_align(offsets->begin + offsets->size, |
| 940 | largest_align) |
| 941 | + data_size; |
| 942 | if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size) |
| 943 | > buf->chan->subbuf_size) { |
| 944 | /* |
| 945 | * Event too big for subbuffers, report error, don't |
| 946 | * complete the sub-buffer switch. |
| 947 | */ |
| 948 | local_inc(&buf->events_lost); |
| 949 | return -1; |
| 950 | } else { |
| 951 | /* |
| 952 | * We just made a successful buffer switch and the event |
| 953 | * fits in the new subbuffer. Let's write. |
| 954 | */ |
| 955 | } |
| 956 | } else { |
| 957 | /* |
| 958 | * Event fits in the current buffer and we are not on a switch |
| 959 | * boundary. It's safe to write. |
| 960 | */ |
| 961 | } |
| 962 | offsets->end = offsets->begin + offsets->size; |
| 963 | |
| 964 | if ((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0) { |
| 965 | /* |
| 966 | * The offset_end will fall at the very beginning of the next |
| 967 | * subbuffer. |
| 968 | */ |
| 969 | offsets->end_switch_current = 1; /* For offsets->begin */ |
| 970 | } |
| 971 | return 0; |
| 972 | } |
| 973 | |
| 974 | /* |
| 975 | * Returns : |
| 976 | * 0 if ok |
| 977 | * !0 if execution must be aborted. |
| 978 | */ |
| 979 | static inline int ltt_relay_try_switch( |
| 980 | enum force_switch_mode mode, |
| 981 | struct ust_channel *channel, |
| 982 | struct ust_buffer *buf, |
| 983 | struct ltt_reserve_switch_offsets *offsets, |
| 984 | u64 *tsc) |
| 985 | { |
| 986 | long subbuf_index; |
| 987 | |
| 988 | offsets->begin = local_read(&buf->offset); |
| 989 | offsets->old = offsets->begin; |
| 990 | offsets->begin_switch = 0; |
| 991 | offsets->end_switch_old = 0; |
| 992 | |
| 993 | *tsc = trace_clock_read64(); |
| 994 | |
| 995 | if (SUBBUF_OFFSET(offsets->begin, buf->chan) != 0) { |
| 996 | offsets->begin = SUBBUF_ALIGN(offsets->begin, buf->chan); |
| 997 | offsets->end_switch_old = 1; |
| 998 | } else { |
| 999 | /* we do not have to switch : buffer is empty */ |
| 1000 | return -1; |
| 1001 | } |
| 1002 | if (mode == FORCE_ACTIVE) |
| 1003 | offsets->begin += ltt_subbuffer_header_size(); |
| 1004 | /* |
| 1005 | * Always begin_switch in FORCE_ACTIVE mode. |
| 1006 | * Test new buffer integrity |
| 1007 | */ |
| 1008 | subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); |
| 1009 | offsets->reserve_commit_diff = |
| 1010 | (BUFFER_TRUNC(offsets->begin, buf->chan) |
| 1011 | >> channel->n_subbufs_order) |
| 1012 | - (local_read(&buf->commit_count[subbuf_index]) |
| 1013 | & channel->commit_count_mask); |
| 1014 | if (offsets->reserve_commit_diff == 0) { |
| 1015 | /* Next buffer not corrupted. */ |
| 1016 | if (mode == FORCE_ACTIVE |
| 1017 | && !channel->overwrite |
| 1018 | && offsets->begin - atomic_long_read(&buf->consumed) |
| 1019 | >= channel->alloc_size) { |
| 1020 | /* |
| 1021 | * We do not overwrite non consumed buffers and we are |
| 1022 | * full : ignore switch while tracing is active. |
| 1023 | */ |
| 1024 | return -1; |
| 1025 | } |
| 1026 | } else { |
| 1027 | /* |
| 1028 | * Next subbuffer corrupted. Force pushing reader even in normal |
| 1029 | * mode |
| 1030 | */ |
| 1031 | } |
| 1032 | offsets->end = offsets->begin; |
| 1033 | return 0; |
| 1034 | } |
| 1035 | |
| 1036 | static inline void ltt_reserve_push_reader( |
| 1037 | struct ust_channel *channel, |
| 1038 | struct ust_buffer *buf, |
| 1039 | struct ltt_reserve_switch_offsets *offsets) |
| 1040 | { |
| 1041 | long consumed_old, consumed_new; |
| 1042 | |
| 1043 | do { |
| 1044 | consumed_old = atomic_long_read(&buf->consumed); |
| 1045 | /* |
| 1046 | * If buffer is in overwrite mode, push the reader consumed |
| 1047 | * count if the write position has reached it and we are not |
| 1048 | * at the first iteration (don't push the reader farther than |
| 1049 | * the writer). This operation can be done concurrently by many |
| 1050 | * writers in the same buffer, the writer being at the farthest |
| 1051 | * write position sub-buffer index in the buffer being the one |
| 1052 | * which will win this loop. |
| 1053 | * If the buffer is not in overwrite mode, pushing the reader |
| 1054 | * only happens if a sub-buffer is corrupted. |
| 1055 | */ |
| 1056 | if ((SUBBUF_TRUNC(offsets->end-1, buf->chan) |
| 1057 | - SUBBUF_TRUNC(consumed_old, buf->chan)) |
| 1058 | >= channel->alloc_size) |
| 1059 | consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan); |
| 1060 | else { |
| 1061 | consumed_new = consumed_old; |
| 1062 | break; |
| 1063 | } |
| 1064 | } while (atomic_long_cmpxchg(&buf->consumed, consumed_old, |
| 1065 | consumed_new) != consumed_old); |
| 1066 | |
| 1067 | if (consumed_old != consumed_new) { |
| 1068 | /* |
| 1069 | * Reader pushed : we are the winner of the push, we can |
| 1070 | * therefore reequilibrate reserve and commit. Atomic increment |
| 1071 | * of the commit count permits other writers to play around |
| 1072 | * with this variable before us. We keep track of |
| 1073 | * corrupted_subbuffers even in overwrite mode : |
| 1074 | * we never want to write over a non completely committed |
| 1075 | * sub-buffer : possible causes : the buffer size is too low |
| 1076 | * compared to the unordered data input, or there is a writer |
| 1077 | * that died between the reserve and the commit. |
| 1078 | */ |
| 1079 | if (offsets->reserve_commit_diff) { |
| 1080 | /* |
| 1081 | * We have to alter the sub-buffer commit count. |
| 1082 | * We do not deliver the previous subbuffer, given it |
| 1083 | * was either corrupted or not consumed (overwrite |
| 1084 | * mode). |
| 1085 | */ |
| 1086 | local_add(offsets->reserve_commit_diff, |
| 1087 | &buf->commit_count[ |
| 1088 | SUBBUF_INDEX(offsets->begin, |
| 1089 | buf->chan)]); |
| 1090 | if (!channel->overwrite |
| 1091 | || offsets->reserve_commit_diff |
| 1092 | != channel->subbuf_size) { |
| 1093 | /* |
| 1094 | * The reserve commit diff was not subbuf_size : |
| 1095 | * it means the subbuffer was partly written to |
| 1096 | * and is therefore corrupted. If it is multiple |
| 1097 | * of subbuffer size and we are in flight |
| 1098 | * recorder mode, we are skipping over a whole |
| 1099 | * subbuffer. |
| 1100 | */ |
| 1101 | local_inc(&buf->corrupted_subbuffers); |
| 1102 | } |
| 1103 | } |
| 1104 | } |
| 1105 | } |
| 1106 | |
| 1107 | |
| 1108 | /* |
| 1109 | * ltt_reserve_switch_old_subbuf: switch old subbuffer |
| 1110 | * |
| 1111 | * Concurrency safe because we are the last and only thread to alter this |
| 1112 | * sub-buffer. As long as it is not delivered and read, no other thread can |
| 1113 | * alter the offset, alter the reserve_count or call the |
| 1114 | * client_buffer_end_callback on this sub-buffer. |
| 1115 | * |
| 1116 | * The only remaining threads could be the ones with pending commits. They will |
| 1117 | * have to do the deliver themselves. Not concurrency safe in overwrite mode. |
| 1118 | * We detect corrupted subbuffers with commit and reserve counts. We keep a |
| 1119 | * corrupted sub-buffers count and push the readers across these sub-buffers. |
| 1120 | * |
| 1121 | * Not concurrency safe if a writer is stalled in a subbuffer and another writer |
| 1122 | * switches in, finding out it's corrupted. The result will be than the old |
| 1123 | * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer |
| 1124 | * will be declared corrupted too because of the commit count adjustment. |
| 1125 | * |
| 1126 | * Note : offset_old should never be 0 here. |
| 1127 | */ |
| 1128 | static inline void ltt_reserve_switch_old_subbuf( |
| 1129 | struct ust_channel *channel, |
| 1130 | struct ust_buffer *buf, |
| 1131 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) |
| 1132 | { |
| 1133 | long oldidx = SUBBUF_INDEX(offsets->old - 1, channel); |
| 1134 | |
| 1135 | channel->buffer_end(buf, *tsc, offsets->old, oldidx); |
| 1136 | /* Must write buffer end before incrementing commit count */ |
| 1137 | smp_wmb(); |
| 1138 | offsets->commit_count = |
| 1139 | local_add_return(channel->subbuf_size |
| 1140 | - (SUBBUF_OFFSET(offsets->old - 1, channel) |
| 1141 | + 1), |
| 1142 | &buf->commit_count[oldidx]); |
| 1143 | if ((BUFFER_TRUNC(offsets->old - 1, channel) |
| 1144 | >> channel->n_subbufs_order) |
| 1145 | - ((offsets->commit_count - channel->subbuf_size) |
| 1146 | & channel->commit_count_mask) == 0) |
| 1147 | ltt_deliver(buf, oldidx, offsets->commit_count); |
| 1148 | } |
| 1149 | |
| 1150 | /* |
| 1151 | * ltt_reserve_switch_new_subbuf: Populate new subbuffer. |
| 1152 | * |
| 1153 | * This code can be executed unordered : writers may already have written to the |
| 1154 | * sub-buffer before this code gets executed, caution. The commit makes sure |
| 1155 | * that this code is executed before the deliver of this sub-buffer. |
| 1156 | */ |
| 1157 | static /*inline*/ void ltt_reserve_switch_new_subbuf( |
| 1158 | struct ust_channel *channel, |
| 1159 | struct ust_buffer *buf, |
| 1160 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) |
| 1161 | { |
| 1162 | long beginidx = SUBBUF_INDEX(offsets->begin, channel); |
| 1163 | |
| 1164 | channel->buffer_begin(buf, *tsc, beginidx); |
| 1165 | /* Must write buffer end before incrementing commit count */ |
| 1166 | smp_wmb(); |
| 1167 | offsets->commit_count = local_add_return(ltt_subbuffer_header_size(), |
| 1168 | &buf->commit_count[beginidx]); |
| 1169 | /* Check if the written buffer has to be delivered */ |
| 1170 | if ((BUFFER_TRUNC(offsets->begin, channel) |
| 1171 | >> channel->n_subbufs_order) |
| 1172 | - ((offsets->commit_count - channel->subbuf_size) |
| 1173 | & channel->commit_count_mask) == 0) |
| 1174 | ltt_deliver(buf, beginidx, offsets->commit_count); |
| 1175 | } |
| 1176 | |
| 1177 | |
| 1178 | /* |
| 1179 | * ltt_reserve_end_switch_current: finish switching current subbuffer |
| 1180 | * |
| 1181 | * Concurrency safe because we are the last and only thread to alter this |
| 1182 | * sub-buffer. As long as it is not delivered and read, no other thread can |
| 1183 | * alter the offset, alter the reserve_count or call the |
| 1184 | * client_buffer_end_callback on this sub-buffer. |
| 1185 | * |
| 1186 | * The only remaining threads could be the ones with pending commits. They will |
| 1187 | * have to do the deliver themselves. Not concurrency safe in overwrite mode. |
| 1188 | * We detect corrupted subbuffers with commit and reserve counts. We keep a |
| 1189 | * corrupted sub-buffers count and push the readers across these sub-buffers. |
| 1190 | * |
| 1191 | * Not concurrency safe if a writer is stalled in a subbuffer and another writer |
| 1192 | * switches in, finding out it's corrupted. The result will be than the old |
| 1193 | * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer |
| 1194 | * will be declared corrupted too because of the commit count adjustment. |
| 1195 | */ |
| 1196 | static inline void ltt_reserve_end_switch_current( |
| 1197 | struct ust_channel *channel, |
| 1198 | struct ust_buffer *buf, |
| 1199 | struct ltt_reserve_switch_offsets *offsets, u64 *tsc) |
| 1200 | { |
| 1201 | long endidx = SUBBUF_INDEX(offsets->end - 1, channel); |
| 1202 | |
| 1203 | channel->buffer_end(buf, *tsc, offsets->end, endidx); |
| 1204 | /* Must write buffer begin before incrementing commit count */ |
| 1205 | smp_wmb(); |
| 1206 | offsets->commit_count = |
| 1207 | local_add_return(channel->subbuf_size |
| 1208 | - (SUBBUF_OFFSET(offsets->end - 1, channel) |
| 1209 | + 1), |
| 1210 | &buf->commit_count[endidx]); |
| 1211 | if ((BUFFER_TRUNC(offsets->end - 1, channel) |
| 1212 | >> channel->n_subbufs_order) |
| 1213 | - ((offsets->commit_count - channel->subbuf_size) |
| 1214 | & channel->commit_count_mask) == 0) |
| 1215 | ltt_deliver(buf, endidx, offsets->commit_count); |
| 1216 | } |
| 1217 | |
| 1218 | /** |
| 1219 | * ltt_relay_reserve_slot - Atomic slot reservation in a LTTng buffer. |
| 1220 | * @trace: the trace structure to log to. |
| 1221 | * @ltt_channel: channel structure |
| 1222 | * @transport_data: data structure specific to ltt relay |
| 1223 | * @data_size: size of the variable length data to log. |
| 1224 | * @slot_size: pointer to total size of the slot (out) |
| 1225 | * @buf_offset : pointer to reserved buffer offset (out) |
| 1226 | * @tsc: pointer to the tsc at the slot reservation (out) |
| 1227 | * @cpu: cpuid |
| 1228 | * |
| 1229 | * Return : -ENOSPC if not enough space, else returns 0. |
| 1230 | * It will take care of sub-buffer switching. |
| 1231 | */ |
| 1232 | static notrace int ltt_relay_reserve_slot(struct ltt_trace_struct *trace, |
| 1233 | struct ust_channel *channel, void **transport_data, |
| 1234 | size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc, |
| 1235 | unsigned int *rflags, int largest_align, int cpu) |
| 1236 | { |
| 1237 | struct ust_buffer *buf = *transport_data = channel->buf[cpu]; |
| 1238 | struct ltt_reserve_switch_offsets offsets; |
| 1239 | |
| 1240 | offsets.reserve_commit_diff = 0; |
| 1241 | offsets.size = 0; |
| 1242 | |
| 1243 | /* |
| 1244 | * Perform retryable operations. |
| 1245 | */ |
| 1246 | if (ltt_nesting > 4) { |
| 1247 | local_inc(&buf->events_lost); |
| 1248 | return -EPERM; |
| 1249 | } |
| 1250 | do { |
| 1251 | if (ltt_relay_try_reserve(channel, buf, &offsets, data_size, tsc, rflags, |
| 1252 | largest_align)) |
| 1253 | return -ENOSPC; |
| 1254 | } while (local_cmpxchg(&buf->offset, offsets.old, |
| 1255 | offsets.end) != offsets.old); |
| 1256 | |
| 1257 | /* |
| 1258 | * Atomically update last_tsc. This update races against concurrent |
| 1259 | * atomic updates, but the race will always cause supplementary full TSC |
| 1260 | * events, never the opposite (missing a full TSC event when it would be |
| 1261 | * needed). |
| 1262 | */ |
| 1263 | save_last_tsc(buf, *tsc); |
| 1264 | |
| 1265 | /* |
| 1266 | * Push the reader if necessary |
| 1267 | */ |
| 1268 | ltt_reserve_push_reader(channel, buf, &offsets); |
| 1269 | |
| 1270 | /* |
| 1271 | * Switch old subbuffer if needed. |
| 1272 | */ |
| 1273 | if (offsets.end_switch_old) |
| 1274 | ltt_reserve_switch_old_subbuf(channel, buf, &offsets, tsc); |
| 1275 | |
| 1276 | /* |
| 1277 | * Populate new subbuffer. |
| 1278 | */ |
| 1279 | if (offsets.begin_switch) |
| 1280 | ltt_reserve_switch_new_subbuf(channel, buf, &offsets, tsc); |
| 1281 | |
| 1282 | if (offsets.end_switch_current) |
| 1283 | ltt_reserve_end_switch_current(channel, buf, &offsets, tsc); |
| 1284 | |
| 1285 | *slot_size = offsets.size; |
| 1286 | *buf_offset = offsets.begin + offsets.before_hdr_pad; |
| 1287 | return 0; |
| 1288 | } |
| 1289 | |
| 1290 | /* |
| 1291 | * Force a sub-buffer switch for a per-cpu buffer. This operation is |
| 1292 | * completely reentrant : can be called while tracing is active with |
| 1293 | * absolutely no lock held. |
| 1294 | * |
| 1295 | * Note, however, that as a local_cmpxchg is used for some atomic |
| 1296 | * operations, this function must be called from the CPU which owns the buffer |
| 1297 | * for a ACTIVE flush. |
| 1298 | */ |
| 1299 | static notrace void ltt_force_switch(struct ust_buffer *buf, |
| 1300 | enum force_switch_mode mode) |
| 1301 | { |
| 1302 | struct ust_channel *channel = buf->chan; |
| 1303 | struct ltt_reserve_switch_offsets offsets; |
| 1304 | u64 tsc; |
| 1305 | |
| 1306 | offsets.reserve_commit_diff = 0; |
| 1307 | offsets.size = 0; |
| 1308 | |
| 1309 | /* |
| 1310 | * Perform retryable operations. |
| 1311 | */ |
| 1312 | do { |
| 1313 | if (ltt_relay_try_switch(mode, channel, buf, &offsets, &tsc)) |
| 1314 | return; |
| 1315 | } while (local_cmpxchg(&buf->offset, offsets.old, |
| 1316 | offsets.end) != offsets.old); |
| 1317 | |
| 1318 | /* |
| 1319 | * Atomically update last_tsc. This update races against concurrent |
| 1320 | * atomic updates, but the race will always cause supplementary full TSC |
| 1321 | * events, never the opposite (missing a full TSC event when it would be |
| 1322 | * needed). |
| 1323 | */ |
| 1324 | save_last_tsc(buf, tsc); |
| 1325 | |
| 1326 | /* |
| 1327 | * Push the reader if necessary |
| 1328 | */ |
| 1329 | if (mode == FORCE_ACTIVE) |
| 1330 | ltt_reserve_push_reader(channel, buf, &offsets); |
| 1331 | |
| 1332 | /* |
| 1333 | * Switch old subbuffer if needed. |
| 1334 | */ |
| 1335 | if (offsets.end_switch_old) |
| 1336 | ltt_reserve_switch_old_subbuf(channel, buf, &offsets, &tsc); |
| 1337 | |
| 1338 | /* |
| 1339 | * Populate new subbuffer. |
| 1340 | */ |
| 1341 | if (mode == FORCE_ACTIVE) |
| 1342 | ltt_reserve_switch_new_subbuf(channel, buf, &offsets, &tsc); |
| 1343 | } |
| 1344 | |
| 1345 | static struct ltt_transport ust_relay_transport = { |
| 1346 | .name = "ustrelay", |
| 1347 | .ops = { |
| 1348 | .create_channel = ust_buffers_create_channel, |
| 1349 | .finish_channel = ltt_relay_finish_channel, |
| 1350 | .remove_channel = ltt_relay_remove_channel, |
| 1351 | .wakeup_channel = ltt_relay_async_wakeup_chan, |
| 1352 | // .commit_slot = ltt_relay_commit_slot, |
| 1353 | .reserve_slot = ltt_relay_reserve_slot, |
| 1354 | }, |
| 1355 | }; |
| 1356 | |
| 1357 | /* |
| 1358 | * for flight recording. must be called after relay_commit. |
| 1359 | * This function decrements de subbuffer's lost_size each time the commit count |
| 1360 | * reaches back the reserve offset (module subbuffer size). It is useful for |
| 1361 | * crash dump. |
| 1362 | */ |
| 1363 | static /* inline */ void ltt_write_commit_counter(struct ust_buffer *buf, |
| 1364 | struct ust_buffer *ltt_buf, |
| 1365 | long idx, long buf_offset, long commit_count, size_t data_size) |
| 1366 | { |
| 1367 | long offset; |
| 1368 | long commit_seq_old; |
| 1369 | |
| 1370 | offset = buf_offset + data_size; |
| 1371 | |
| 1372 | /* |
| 1373 | * SUBBUF_OFFSET includes commit_count_mask. We can simply |
| 1374 | * compare the offsets within the subbuffer without caring about |
| 1375 | * buffer full/empty mismatch because offset is never zero here |
| 1376 | * (subbuffer header and event headers have non-zero length). |
| 1377 | */ |
| 1378 | if (unlikely(SUBBUF_OFFSET(offset - commit_count, buf->chan))) |
| 1379 | return; |
| 1380 | |
| 1381 | commit_seq_old = local_read(<t_buf->commit_seq[idx]); |
| 1382 | while (commit_seq_old < commit_count) |
| 1383 | commit_seq_old = local_cmpxchg(<t_buf->commit_seq[idx], |
| 1384 | commit_seq_old, commit_count); |
| 1385 | } |
| 1386 | |
| 1387 | /* |
| 1388 | * Atomic unordered slot commit. Increments the commit count in the |
| 1389 | * specified sub-buffer, and delivers it if necessary. |
| 1390 | * |
| 1391 | * Parameters: |
| 1392 | * |
| 1393 | * @ltt_channel : channel structure |
| 1394 | * @transport_data: transport-specific data |
| 1395 | * @buf_offset : offset following the event header. |
| 1396 | * @data_size : size of the event data. |
| 1397 | * @slot_size : size of the reserved slot. |
| 1398 | */ |
| 1399 | /* FIXME: make this function static inline in the .h! */ |
| 1400 | /*static*/ /* inline */ notrace void ltt_commit_slot( |
| 1401 | struct ust_channel *channel, |
| 1402 | void **transport_data, long buf_offset, |
| 1403 | size_t data_size, size_t slot_size) |
| 1404 | { |
| 1405 | struct ust_buffer *buf = *transport_data; |
| 1406 | long offset_end = buf_offset; |
| 1407 | long endidx = SUBBUF_INDEX(offset_end - 1, channel); |
| 1408 | long commit_count; |
| 1409 | |
| 1410 | /* Must write slot data before incrementing commit count */ |
| 1411 | smp_wmb(); |
| 1412 | commit_count = local_add_return(slot_size, |
| 1413 | &buf->commit_count[endidx]); |
| 1414 | /* Check if all commits have been done */ |
| 1415 | if ((BUFFER_TRUNC(offset_end - 1, channel) |
| 1416 | >> channel->n_subbufs_order) |
| 1417 | - ((commit_count - channel->subbuf_size) |
| 1418 | & channel->commit_count_mask) == 0) |
| 1419 | ltt_deliver(buf, endidx, commit_count); |
| 1420 | /* |
| 1421 | * Update lost_size for each commit. It's needed only for extracting |
| 1422 | * ltt buffers from vmcore, after crash. |
| 1423 | */ |
| 1424 | ltt_write_commit_counter(buf, buf, endidx, |
| 1425 | buf_offset, commit_count, data_size); |
| 1426 | } |
| 1427 | |
| 1428 | |
| 1429 | static char initialized = 0; |
| 1430 | |
| 1431 | void __attribute__((constructor)) init_ustrelay_transport(void) |
| 1432 | { |
| 1433 | if(!initialized) { |
| 1434 | ltt_transport_register(&ust_relay_transport); |
| 1435 | initialized = 1; |
| 1436 | } |
| 1437 | } |
| 1438 | |
| 1439 | static void __attribute__((destructor)) ltt_relay_exit(void) |
| 1440 | { |
| 1441 | ltt_transport_unregister(&ust_relay_transport); |
| 1442 | } |