| 1 | /* SPDX-License-Identifier: (GPL-2.0 OR LGPL-2.1) |
| 2 | * |
| 3 | * ring_buffer_iterator.c |
| 4 | * |
| 5 | * Ring buffer and channel iterators. Get each event of a channel in order. Uses |
| 6 | * a prio heap for per-cpu buffers, giving a O(log(NR_CPUS)) algorithmic |
| 7 | * complexity for the "get next event" operation. |
| 8 | * |
| 9 | * Copyright (C) 2010-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 10 | */ |
| 11 | |
| 12 | #include <wrapper/ringbuffer/iterator.h> |
| 13 | #include <wrapper/file.h> |
| 14 | #include <wrapper/uaccess.h> |
| 15 | #include <linux/jiffies.h> |
| 16 | #include <linux/delay.h> |
| 17 | #include <linux/module.h> |
| 18 | |
| 19 | /* |
| 20 | * Safety factor taking into account internal kernel interrupt latency. |
| 21 | * Assuming 250ms worse-case latency. |
| 22 | */ |
| 23 | #define MAX_SYSTEM_LATENCY 250 |
| 24 | |
| 25 | /* |
| 26 | * Maximum delta expected between trace clocks. At most 1 jiffy delta. |
| 27 | */ |
| 28 | #define MAX_CLOCK_DELTA (jiffies_to_usecs(1) * 1000) |
| 29 | |
| 30 | /** |
| 31 | * lib_ring_buffer_get_next_record - Get the next record in a buffer. |
| 32 | * @chan: channel |
| 33 | * @buf: buffer |
| 34 | * |
| 35 | * Returns the size of the event read, -EAGAIN if buffer is empty, -ENODATA if |
| 36 | * buffer is empty and finalized. The buffer must already be opened for reading. |
| 37 | */ |
| 38 | ssize_t lib_ring_buffer_get_next_record(struct channel *chan, |
| 39 | struct lib_ring_buffer *buf) |
| 40 | { |
| 41 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 42 | struct lib_ring_buffer_iter *iter = &buf->iter; |
| 43 | int ret; |
| 44 | |
| 45 | restart: |
| 46 | switch (iter->state) { |
| 47 | case ITER_GET_SUBBUF: |
| 48 | ret = lib_ring_buffer_get_next_subbuf(buf); |
| 49 | if (ret && !READ_ONCE(buf->finalized) |
| 50 | && config->alloc == RING_BUFFER_ALLOC_GLOBAL) { |
| 51 | /* |
| 52 | * Use "pull" scheme for global buffers. The reader |
| 53 | * itself flushes the buffer to "pull" data not visible |
| 54 | * to readers yet. Flush current subbuffer and re-try. |
| 55 | * |
| 56 | * Per-CPU buffers rather use a "push" scheme because |
| 57 | * the IPI needed to flush all CPU's buffers is too |
| 58 | * costly. In the "push" scheme, the reader waits for |
| 59 | * the writer periodic timer to flush the |
| 60 | * buffers (keeping track of a quiescent state |
| 61 | * timestamp). Therefore, the writer "pushes" data out |
| 62 | * of the buffers rather than letting the reader "pull" |
| 63 | * data from the buffer. |
| 64 | */ |
| 65 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); |
| 66 | ret = lib_ring_buffer_get_next_subbuf(buf); |
| 67 | } |
| 68 | if (ret) |
| 69 | return ret; |
| 70 | iter->consumed = buf->cons_snapshot; |
| 71 | iter->data_size = lib_ring_buffer_get_read_data_size(config, buf); |
| 72 | iter->read_offset = iter->consumed; |
| 73 | /* skip header */ |
| 74 | iter->read_offset += config->cb.subbuffer_header_size(); |
| 75 | iter->state = ITER_TEST_RECORD; |
| 76 | goto restart; |
| 77 | case ITER_TEST_RECORD: |
| 78 | if (iter->read_offset - iter->consumed >= iter->data_size) { |
| 79 | iter->state = ITER_PUT_SUBBUF; |
| 80 | } else { |
| 81 | CHAN_WARN_ON(chan, !config->cb.record_get); |
| 82 | config->cb.record_get(config, chan, buf, |
| 83 | iter->read_offset, |
| 84 | &iter->header_len, |
| 85 | &iter->payload_len, |
| 86 | &iter->timestamp); |
| 87 | iter->read_offset += iter->header_len; |
| 88 | subbuffer_consume_record(config, &buf->backend); |
| 89 | iter->state = ITER_NEXT_RECORD; |
| 90 | return iter->payload_len; |
| 91 | } |
| 92 | goto restart; |
| 93 | case ITER_NEXT_RECORD: |
| 94 | iter->read_offset += iter->payload_len; |
| 95 | iter->state = ITER_TEST_RECORD; |
| 96 | goto restart; |
| 97 | case ITER_PUT_SUBBUF: |
| 98 | lib_ring_buffer_put_next_subbuf(buf); |
| 99 | iter->state = ITER_GET_SUBBUF; |
| 100 | goto restart; |
| 101 | default: |
| 102 | CHAN_WARN_ON(chan, 1); /* Should not happen */ |
| 103 | return -EPERM; |
| 104 | } |
| 105 | } |
| 106 | EXPORT_SYMBOL_GPL(lib_ring_buffer_get_next_record); |
| 107 | |
| 108 | static int buf_is_higher(void *a, void *b) |
| 109 | { |
| 110 | struct lib_ring_buffer *bufa = a; |
| 111 | struct lib_ring_buffer *bufb = b; |
| 112 | |
| 113 | /* Consider lowest timestamps to be at the top of the heap */ |
| 114 | return (bufa->iter.timestamp < bufb->iter.timestamp); |
| 115 | } |
| 116 | |
| 117 | static |
| 118 | void lib_ring_buffer_get_empty_buf_records(const struct lib_ring_buffer_config *config, |
| 119 | struct channel *chan) |
| 120 | { |
| 121 | struct lttng_ptr_heap *heap = &chan->iter.heap; |
| 122 | struct lib_ring_buffer *buf, *tmp; |
| 123 | ssize_t len; |
| 124 | |
| 125 | list_for_each_entry_safe(buf, tmp, &chan->iter.empty_head, |
| 126 | iter.empty_node) { |
| 127 | len = lib_ring_buffer_get_next_record(chan, buf); |
| 128 | |
| 129 | /* |
| 130 | * Deal with -EAGAIN and -ENODATA. |
| 131 | * len >= 0 means record contains data. |
| 132 | * -EBUSY should never happen, because we support only one |
| 133 | * reader. |
| 134 | */ |
| 135 | switch (len) { |
| 136 | case -EAGAIN: |
| 137 | /* Keep node in empty list */ |
| 138 | break; |
| 139 | case -ENODATA: |
| 140 | /* |
| 141 | * Buffer is finalized. Don't add to list of empty |
| 142 | * buffer, because it has no more data to provide, ever. |
| 143 | */ |
| 144 | list_del(&buf->iter.empty_node); |
| 145 | break; |
| 146 | case -EBUSY: |
| 147 | CHAN_WARN_ON(chan, 1); |
| 148 | break; |
| 149 | default: |
| 150 | /* |
| 151 | * Insert buffer into the heap, remove from empty buffer |
| 152 | * list. |
| 153 | */ |
| 154 | CHAN_WARN_ON(chan, len < 0); |
| 155 | list_del(&buf->iter.empty_node); |
| 156 | CHAN_WARN_ON(chan, lttng_heap_insert(heap, buf)); |
| 157 | } |
| 158 | } |
| 159 | } |
| 160 | |
| 161 | static |
| 162 | void lib_ring_buffer_wait_for_qs(const struct lib_ring_buffer_config *config, |
| 163 | struct channel *chan) |
| 164 | { |
| 165 | u64 timestamp_qs; |
| 166 | unsigned long wait_msecs; |
| 167 | |
| 168 | /* |
| 169 | * No need to wait if no empty buffers are present. |
| 170 | */ |
| 171 | if (list_empty(&chan->iter.empty_head)) |
| 172 | return; |
| 173 | |
| 174 | timestamp_qs = config->cb.ring_buffer_clock_read(chan); |
| 175 | /* |
| 176 | * We need to consider previously empty buffers. |
| 177 | * Do a get next buf record on each of them. Add them to |
| 178 | * the heap if they have data. If at least one of them |
| 179 | * don't have data, we need to wait for |
| 180 | * switch_timer_interval + MAX_SYSTEM_LATENCY (so we are sure the |
| 181 | * buffers have been switched either by the timer or idle entry) and |
| 182 | * check them again, adding them if they have data. |
| 183 | */ |
| 184 | lib_ring_buffer_get_empty_buf_records(config, chan); |
| 185 | |
| 186 | /* |
| 187 | * No need to wait if no empty buffers are present. |
| 188 | */ |
| 189 | if (list_empty(&chan->iter.empty_head)) |
| 190 | return; |
| 191 | |
| 192 | /* |
| 193 | * We need to wait for the buffer switch timer to run. If the |
| 194 | * CPU is idle, idle entry performed the switch. |
| 195 | * TODO: we could optimize further by skipping the sleep if all |
| 196 | * empty buffers belong to idle or offline cpus. |
| 197 | */ |
| 198 | wait_msecs = jiffies_to_msecs(chan->switch_timer_interval); |
| 199 | wait_msecs += MAX_SYSTEM_LATENCY; |
| 200 | msleep(wait_msecs); |
| 201 | lib_ring_buffer_get_empty_buf_records(config, chan); |
| 202 | /* |
| 203 | * Any buffer still in the empty list here cannot possibly |
| 204 | * contain an event with a timestamp prior to "timestamp_qs". |
| 205 | * The new quiescent state timestamp is the one we grabbed |
| 206 | * before waiting for buffer data. It is therefore safe to |
| 207 | * ignore empty buffers up to last_qs timestamp for fusion |
| 208 | * merge. |
| 209 | */ |
| 210 | chan->iter.last_qs = timestamp_qs; |
| 211 | } |
| 212 | |
| 213 | /** |
| 214 | * channel_get_next_record - Get the next record in a channel. |
| 215 | * @chan: channel |
| 216 | * @ret_buf: the buffer in which the event is located (output) |
| 217 | * |
| 218 | * Returns the size of new current event, -EAGAIN if all buffers are empty, |
| 219 | * -ENODATA if all buffers are empty and finalized. The channel must already be |
| 220 | * opened for reading. |
| 221 | */ |
| 222 | |
| 223 | ssize_t channel_get_next_record(struct channel *chan, |
| 224 | struct lib_ring_buffer **ret_buf) |
| 225 | { |
| 226 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 227 | struct lib_ring_buffer *buf; |
| 228 | struct lttng_ptr_heap *heap; |
| 229 | ssize_t len; |
| 230 | |
| 231 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { |
| 232 | *ret_buf = channel_get_ring_buffer(config, chan, 0); |
| 233 | return lib_ring_buffer_get_next_record(chan, *ret_buf); |
| 234 | } |
| 235 | |
| 236 | heap = &chan->iter.heap; |
| 237 | |
| 238 | /* |
| 239 | * get next record for topmost buffer. |
| 240 | */ |
| 241 | buf = lttng_heap_maximum(heap); |
| 242 | if (buf) { |
| 243 | len = lib_ring_buffer_get_next_record(chan, buf); |
| 244 | /* |
| 245 | * Deal with -EAGAIN and -ENODATA. |
| 246 | * len >= 0 means record contains data. |
| 247 | */ |
| 248 | switch (len) { |
| 249 | case -EAGAIN: |
| 250 | buf->iter.timestamp = 0; |
| 251 | list_add(&buf->iter.empty_node, &chan->iter.empty_head); |
| 252 | /* Remove topmost buffer from the heap */ |
| 253 | CHAN_WARN_ON(chan, lttng_heap_remove(heap) != buf); |
| 254 | break; |
| 255 | case -ENODATA: |
| 256 | /* |
| 257 | * Buffer is finalized. Remove buffer from heap and |
| 258 | * don't add to list of empty buffer, because it has no |
| 259 | * more data to provide, ever. |
| 260 | */ |
| 261 | CHAN_WARN_ON(chan, lttng_heap_remove(heap) != buf); |
| 262 | break; |
| 263 | case -EBUSY: |
| 264 | CHAN_WARN_ON(chan, 1); |
| 265 | break; |
| 266 | default: |
| 267 | /* |
| 268 | * Reinsert buffer into the heap. Note that heap can be |
| 269 | * partially empty, so we need to use |
| 270 | * lttng_heap_replace_max(). |
| 271 | */ |
| 272 | CHAN_WARN_ON(chan, len < 0); |
| 273 | CHAN_WARN_ON(chan, lttng_heap_replace_max(heap, buf) != buf); |
| 274 | break; |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | buf = lttng_heap_maximum(heap); |
| 279 | if (!buf || buf->iter.timestamp > chan->iter.last_qs) { |
| 280 | /* |
| 281 | * Deal with buffers previously showing no data. |
| 282 | * Add buffers containing data to the heap, update |
| 283 | * last_qs. |
| 284 | */ |
| 285 | lib_ring_buffer_wait_for_qs(config, chan); |
| 286 | } |
| 287 | |
| 288 | *ret_buf = buf = lttng_heap_maximum(heap); |
| 289 | if (buf) { |
| 290 | /* |
| 291 | * If this warning triggers, you probably need to check your |
| 292 | * system interrupt latency. Typical causes: too many printk() |
| 293 | * output going to a serial console with interrupts off. |
| 294 | * Allow for MAX_CLOCK_DELTA ns timestamp delta going backward. |
| 295 | * Observed on SMP KVM setups with trace_clock(). |
| 296 | */ |
| 297 | if (chan->iter.last_timestamp |
| 298 | > (buf->iter.timestamp + MAX_CLOCK_DELTA)) { |
| 299 | printk(KERN_WARNING "ring_buffer: timestamps going " |
| 300 | "backward. Last time %llu ns, cpu %d, " |
| 301 | "current time %llu ns, cpu %d, " |
| 302 | "delta %llu ns.\n", |
| 303 | chan->iter.last_timestamp, chan->iter.last_cpu, |
| 304 | buf->iter.timestamp, buf->backend.cpu, |
| 305 | chan->iter.last_timestamp - buf->iter.timestamp); |
| 306 | CHAN_WARN_ON(chan, 1); |
| 307 | } |
| 308 | chan->iter.last_timestamp = buf->iter.timestamp; |
| 309 | chan->iter.last_cpu = buf->backend.cpu; |
| 310 | return buf->iter.payload_len; |
| 311 | } else { |
| 312 | /* Heap is empty */ |
| 313 | if (list_empty(&chan->iter.empty_head)) |
| 314 | return -ENODATA; /* All buffers finalized */ |
| 315 | else |
| 316 | return -EAGAIN; /* Temporarily empty */ |
| 317 | } |
| 318 | } |
| 319 | EXPORT_SYMBOL_GPL(channel_get_next_record); |
| 320 | |
| 321 | static |
| 322 | void lib_ring_buffer_iterator_init(struct channel *chan, struct lib_ring_buffer *buf) |
| 323 | { |
| 324 | if (buf->iter.allocated) |
| 325 | return; |
| 326 | |
| 327 | buf->iter.allocated = 1; |
| 328 | if (chan->iter.read_open && !buf->iter.read_open) { |
| 329 | CHAN_WARN_ON(chan, lib_ring_buffer_open_read(buf) != 0); |
| 330 | buf->iter.read_open = 1; |
| 331 | } |
| 332 | |
| 333 | /* Add to list of buffers without any current record */ |
| 334 | if (chan->backend.config.alloc == RING_BUFFER_ALLOC_PER_CPU) |
| 335 | list_add(&buf->iter.empty_node, &chan->iter.empty_head); |
| 336 | } |
| 337 | |
| 338 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
| 339 | |
| 340 | int lttng_cpuhp_rb_iter_online(unsigned int cpu, |
| 341 | struct lttng_cpuhp_node *node) |
| 342 | { |
| 343 | struct channel *chan = container_of(node, struct channel, |
| 344 | cpuhp_iter_online); |
| 345 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); |
| 346 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 347 | |
| 348 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); |
| 349 | |
| 350 | lib_ring_buffer_iterator_init(chan, buf); |
| 351 | return 0; |
| 352 | } |
| 353 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_iter_online); |
| 354 | |
| 355 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
| 356 | |
| 357 | #ifdef CONFIG_HOTPLUG_CPU |
| 358 | static |
| 359 | int channel_iterator_cpu_hotplug(struct notifier_block *nb, |
| 360 | unsigned long action, |
| 361 | void *hcpu) |
| 362 | { |
| 363 | unsigned int cpu = (unsigned long)hcpu; |
| 364 | struct channel *chan = container_of(nb, struct channel, |
| 365 | hp_iter_notifier); |
| 366 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); |
| 367 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 368 | |
| 369 | if (!chan->hp_iter_enable) |
| 370 | return NOTIFY_DONE; |
| 371 | |
| 372 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); |
| 373 | |
| 374 | switch (action) { |
| 375 | case CPU_DOWN_FAILED: |
| 376 | case CPU_DOWN_FAILED_FROZEN: |
| 377 | case CPU_ONLINE: |
| 378 | case CPU_ONLINE_FROZEN: |
| 379 | lib_ring_buffer_iterator_init(chan, buf); |
| 380 | return NOTIFY_OK; |
| 381 | default: |
| 382 | return NOTIFY_DONE; |
| 383 | } |
| 384 | } |
| 385 | #endif |
| 386 | |
| 387 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
| 388 | |
| 389 | int channel_iterator_init(struct channel *chan) |
| 390 | { |
| 391 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 392 | struct lib_ring_buffer *buf; |
| 393 | |
| 394 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
| 395 | int ret; |
| 396 | |
| 397 | INIT_LIST_HEAD(&chan->iter.empty_head); |
| 398 | ret = lttng_heap_init(&chan->iter.heap, |
| 399 | num_possible_cpus(), |
| 400 | GFP_KERNEL, buf_is_higher); |
| 401 | if (ret) |
| 402 | return ret; |
| 403 | |
| 404 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
| 405 | chan->cpuhp_iter_online.component = LTTNG_RING_BUFFER_ITER; |
| 406 | ret = cpuhp_state_add_instance(lttng_rb_hp_online, |
| 407 | &chan->cpuhp_iter_online.node); |
| 408 | if (ret) |
| 409 | return ret; |
| 410 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
| 411 | { |
| 412 | int cpu; |
| 413 | |
| 414 | /* |
| 415 | * In case of non-hotplug cpu, if the ring-buffer is allocated |
| 416 | * in early initcall, it will not be notified of secondary cpus. |
| 417 | * In that off case, we need to allocate for all possible cpus. |
| 418 | */ |
| 419 | #ifdef CONFIG_HOTPLUG_CPU |
| 420 | chan->hp_iter_notifier.notifier_call = |
| 421 | channel_iterator_cpu_hotplug; |
| 422 | chan->hp_iter_notifier.priority = 10; |
| 423 | register_cpu_notifier(&chan->hp_iter_notifier); |
| 424 | |
| 425 | get_online_cpus(); |
| 426 | for_each_online_cpu(cpu) { |
| 427 | buf = per_cpu_ptr(chan->backend.buf, cpu); |
| 428 | lib_ring_buffer_iterator_init(chan, buf); |
| 429 | } |
| 430 | chan->hp_iter_enable = 1; |
| 431 | put_online_cpus(); |
| 432 | #else |
| 433 | for_each_possible_cpu(cpu) { |
| 434 | buf = per_cpu_ptr(chan->backend.buf, cpu); |
| 435 | lib_ring_buffer_iterator_init(chan, buf); |
| 436 | } |
| 437 | #endif |
| 438 | } |
| 439 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
| 440 | } else { |
| 441 | buf = channel_get_ring_buffer(config, chan, 0); |
| 442 | lib_ring_buffer_iterator_init(chan, buf); |
| 443 | } |
| 444 | return 0; |
| 445 | } |
| 446 | |
| 447 | void channel_iterator_unregister_notifiers(struct channel *chan) |
| 448 | { |
| 449 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 450 | |
| 451 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
| 452 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
| 453 | { |
| 454 | int ret; |
| 455 | |
| 456 | ret = cpuhp_state_remove_instance(lttng_rb_hp_online, |
| 457 | &chan->cpuhp_iter_online.node); |
| 458 | WARN_ON(ret); |
| 459 | } |
| 460 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
| 461 | chan->hp_iter_enable = 0; |
| 462 | unregister_cpu_notifier(&chan->hp_iter_notifier); |
| 463 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
| 464 | } |
| 465 | } |
| 466 | |
| 467 | void channel_iterator_free(struct channel *chan) |
| 468 | { |
| 469 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 470 | |
| 471 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
| 472 | lttng_heap_free(&chan->iter.heap); |
| 473 | } |
| 474 | |
| 475 | int lib_ring_buffer_iterator_open(struct lib_ring_buffer *buf) |
| 476 | { |
| 477 | struct channel *chan = buf->backend.chan; |
| 478 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 479 | CHAN_WARN_ON(chan, config->output != RING_BUFFER_ITERATOR); |
| 480 | return lib_ring_buffer_open_read(buf); |
| 481 | } |
| 482 | EXPORT_SYMBOL_GPL(lib_ring_buffer_iterator_open); |
| 483 | |
| 484 | /* |
| 485 | * Note: Iterators must not be mixed with other types of outputs, because an |
| 486 | * iterator can leave the buffer in "GET" state, which is not consistent with |
| 487 | * other types of output (mmap, splice, raw data read). |
| 488 | */ |
| 489 | void lib_ring_buffer_iterator_release(struct lib_ring_buffer *buf) |
| 490 | { |
| 491 | lib_ring_buffer_release_read(buf); |
| 492 | } |
| 493 | EXPORT_SYMBOL_GPL(lib_ring_buffer_iterator_release); |
| 494 | |
| 495 | int channel_iterator_open(struct channel *chan) |
| 496 | { |
| 497 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 498 | struct lib_ring_buffer *buf; |
| 499 | int ret = 0, cpu; |
| 500 | |
| 501 | CHAN_WARN_ON(chan, config->output != RING_BUFFER_ITERATOR); |
| 502 | |
| 503 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
| 504 | get_online_cpus(); |
| 505 | /* Allow CPU hotplug to keep track of opened reader */ |
| 506 | chan->iter.read_open = 1; |
| 507 | for_each_channel_cpu(cpu, chan) { |
| 508 | buf = channel_get_ring_buffer(config, chan, cpu); |
| 509 | ret = lib_ring_buffer_iterator_open(buf); |
| 510 | if (ret) |
| 511 | goto error; |
| 512 | buf->iter.read_open = 1; |
| 513 | } |
| 514 | put_online_cpus(); |
| 515 | } else { |
| 516 | buf = channel_get_ring_buffer(config, chan, 0); |
| 517 | ret = lib_ring_buffer_iterator_open(buf); |
| 518 | } |
| 519 | return ret; |
| 520 | error: |
| 521 | /* Error should always happen on CPU 0, hence no close is required. */ |
| 522 | CHAN_WARN_ON(chan, cpu != 0); |
| 523 | put_online_cpus(); |
| 524 | return ret; |
| 525 | } |
| 526 | EXPORT_SYMBOL_GPL(channel_iterator_open); |
| 527 | |
| 528 | void channel_iterator_release(struct channel *chan) |
| 529 | { |
| 530 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 531 | struct lib_ring_buffer *buf; |
| 532 | int cpu; |
| 533 | |
| 534 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
| 535 | get_online_cpus(); |
| 536 | for_each_channel_cpu(cpu, chan) { |
| 537 | buf = channel_get_ring_buffer(config, chan, cpu); |
| 538 | if (buf->iter.read_open) { |
| 539 | lib_ring_buffer_iterator_release(buf); |
| 540 | buf->iter.read_open = 0; |
| 541 | } |
| 542 | } |
| 543 | chan->iter.read_open = 0; |
| 544 | put_online_cpus(); |
| 545 | } else { |
| 546 | buf = channel_get_ring_buffer(config, chan, 0); |
| 547 | lib_ring_buffer_iterator_release(buf); |
| 548 | } |
| 549 | } |
| 550 | EXPORT_SYMBOL_GPL(channel_iterator_release); |
| 551 | |
| 552 | void lib_ring_buffer_iterator_reset(struct lib_ring_buffer *buf) |
| 553 | { |
| 554 | struct channel *chan = buf->backend.chan; |
| 555 | |
| 556 | if (buf->iter.state != ITER_GET_SUBBUF) |
| 557 | lib_ring_buffer_put_next_subbuf(buf); |
| 558 | buf->iter.state = ITER_GET_SUBBUF; |
| 559 | /* Remove from heap (if present). */ |
| 560 | if (lttng_heap_cherrypick(&chan->iter.heap, buf)) |
| 561 | list_add(&buf->iter.empty_node, &chan->iter.empty_head); |
| 562 | buf->iter.timestamp = 0; |
| 563 | buf->iter.header_len = 0; |
| 564 | buf->iter.payload_len = 0; |
| 565 | buf->iter.consumed = 0; |
| 566 | buf->iter.read_offset = 0; |
| 567 | buf->iter.data_size = 0; |
| 568 | /* Don't reset allocated and read_open */ |
| 569 | } |
| 570 | |
| 571 | void channel_iterator_reset(struct channel *chan) |
| 572 | { |
| 573 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 574 | struct lib_ring_buffer *buf; |
| 575 | int cpu; |
| 576 | |
| 577 | /* Empty heap, put into empty_head */ |
| 578 | while ((buf = lttng_heap_remove(&chan->iter.heap)) != NULL) |
| 579 | list_add(&buf->iter.empty_node, &chan->iter.empty_head); |
| 580 | |
| 581 | for_each_channel_cpu(cpu, chan) { |
| 582 | buf = channel_get_ring_buffer(config, chan, cpu); |
| 583 | lib_ring_buffer_iterator_reset(buf); |
| 584 | } |
| 585 | /* Don't reset read_open */ |
| 586 | chan->iter.last_qs = 0; |
| 587 | chan->iter.last_timestamp = 0; |
| 588 | chan->iter.last_cpu = 0; |
| 589 | chan->iter.len_left = 0; |
| 590 | } |
| 591 | |
| 592 | /* |
| 593 | * Ring buffer payload extraction read() implementation. |
| 594 | */ |
| 595 | static |
| 596 | ssize_t channel_ring_buffer_file_read(struct file *filp, |
| 597 | char __user *user_buf, |
| 598 | size_t count, |
| 599 | loff_t *ppos, |
| 600 | struct channel *chan, |
| 601 | struct lib_ring_buffer *buf, |
| 602 | int fusionmerge) |
| 603 | { |
| 604 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 605 | size_t read_count = 0, read_offset; |
| 606 | ssize_t len; |
| 607 | |
| 608 | might_sleep(); |
| 609 | if (!lttng_access_ok(VERIFY_WRITE, user_buf, count)) |
| 610 | return -EFAULT; |
| 611 | |
| 612 | /* Finish copy of previous record */ |
| 613 | if (*ppos != 0) { |
| 614 | if (read_count < count) { |
| 615 | len = chan->iter.len_left; |
| 616 | read_offset = *ppos; |
| 617 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU |
| 618 | && fusionmerge) |
| 619 | buf = lttng_heap_maximum(&chan->iter.heap); |
| 620 | CHAN_WARN_ON(chan, !buf); |
| 621 | goto skip_get_next; |
| 622 | } |
| 623 | } |
| 624 | |
| 625 | while (read_count < count) { |
| 626 | size_t copy_len, space_left; |
| 627 | |
| 628 | if (fusionmerge) |
| 629 | len = channel_get_next_record(chan, &buf); |
| 630 | else |
| 631 | len = lib_ring_buffer_get_next_record(chan, buf); |
| 632 | len_test: |
| 633 | if (len < 0) { |
| 634 | /* |
| 635 | * Check if buffer is finalized (end of file). |
| 636 | */ |
| 637 | if (len == -ENODATA) { |
| 638 | /* A 0 read_count will tell about end of file */ |
| 639 | goto nodata; |
| 640 | } |
| 641 | if (filp->f_flags & O_NONBLOCK) { |
| 642 | if (!read_count) |
| 643 | read_count = -EAGAIN; |
| 644 | goto nodata; |
| 645 | } else { |
| 646 | int error; |
| 647 | |
| 648 | /* |
| 649 | * No data available at the moment, return what |
| 650 | * we got. |
| 651 | */ |
| 652 | if (read_count) |
| 653 | goto nodata; |
| 654 | |
| 655 | /* |
| 656 | * Wait for returned len to be >= 0 or -ENODATA. |
| 657 | */ |
| 658 | if (fusionmerge) |
| 659 | error = wait_event_interruptible( |
| 660 | chan->read_wait, |
| 661 | ((len = channel_get_next_record(chan, |
| 662 | &buf)), len != -EAGAIN)); |
| 663 | else |
| 664 | error = wait_event_interruptible( |
| 665 | buf->read_wait, |
| 666 | ((len = lib_ring_buffer_get_next_record( |
| 667 | chan, buf)), len != -EAGAIN)); |
| 668 | CHAN_WARN_ON(chan, len == -EBUSY); |
| 669 | if (error) { |
| 670 | read_count = error; |
| 671 | goto nodata; |
| 672 | } |
| 673 | CHAN_WARN_ON(chan, len < 0 && len != -ENODATA); |
| 674 | goto len_test; |
| 675 | } |
| 676 | } |
| 677 | read_offset = buf->iter.read_offset; |
| 678 | skip_get_next: |
| 679 | space_left = count - read_count; |
| 680 | if (len <= space_left) { |
| 681 | copy_len = len; |
| 682 | chan->iter.len_left = 0; |
| 683 | *ppos = 0; |
| 684 | } else { |
| 685 | copy_len = space_left; |
| 686 | chan->iter.len_left = len - copy_len; |
| 687 | *ppos = read_offset + copy_len; |
| 688 | } |
| 689 | if (__lib_ring_buffer_copy_to_user(&buf->backend, read_offset, |
| 690 | &user_buf[read_count], |
| 691 | copy_len)) { |
| 692 | /* |
| 693 | * Leave the len_left and ppos values at their current |
| 694 | * state, as we currently have a valid event to read. |
| 695 | */ |
| 696 | return -EFAULT; |
| 697 | } |
| 698 | read_count += copy_len; |
| 699 | }; |
| 700 | return read_count; |
| 701 | |
| 702 | nodata: |
| 703 | *ppos = 0; |
| 704 | chan->iter.len_left = 0; |
| 705 | return read_count; |
| 706 | } |
| 707 | |
| 708 | /** |
| 709 | * lib_ring_buffer_file_read - Read buffer record payload. |
| 710 | * @filp: file structure pointer. |
| 711 | * @buffer: user buffer to read data into. |
| 712 | * @count: number of bytes to read. |
| 713 | * @ppos: file read position. |
| 714 | * |
| 715 | * Returns a negative value on error, or the number of bytes read on success. |
| 716 | * ppos is used to save the position _within the current record_ between calls |
| 717 | * to read(). |
| 718 | */ |
| 719 | static |
| 720 | ssize_t lib_ring_buffer_file_read(struct file *filp, |
| 721 | char __user *user_buf, |
| 722 | size_t count, |
| 723 | loff_t *ppos) |
| 724 | { |
| 725 | struct inode *inode = filp->lttng_f_dentry->d_inode; |
| 726 | struct lib_ring_buffer *buf = inode->i_private; |
| 727 | struct channel *chan = buf->backend.chan; |
| 728 | |
| 729 | return channel_ring_buffer_file_read(filp, user_buf, count, ppos, |
| 730 | chan, buf, 0); |
| 731 | } |
| 732 | |
| 733 | /** |
| 734 | * channel_file_read - Read channel record payload. |
| 735 | * @filp: file structure pointer. |
| 736 | * @buffer: user buffer to read data into. |
| 737 | * @count: number of bytes to read. |
| 738 | * @ppos: file read position. |
| 739 | * |
| 740 | * Returns a negative value on error, or the number of bytes read on success. |
| 741 | * ppos is used to save the position _within the current record_ between calls |
| 742 | * to read(). |
| 743 | */ |
| 744 | static |
| 745 | ssize_t channel_file_read(struct file *filp, |
| 746 | char __user *user_buf, |
| 747 | size_t count, |
| 748 | loff_t *ppos) |
| 749 | { |
| 750 | struct inode *inode = filp->lttng_f_dentry->d_inode; |
| 751 | struct channel *chan = inode->i_private; |
| 752 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
| 753 | |
| 754 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
| 755 | return channel_ring_buffer_file_read(filp, user_buf, count, |
| 756 | ppos, chan, NULL, 1); |
| 757 | else { |
| 758 | struct lib_ring_buffer *buf = |
| 759 | channel_get_ring_buffer(config, chan, 0); |
| 760 | return channel_ring_buffer_file_read(filp, user_buf, count, |
| 761 | ppos, chan, buf, 0); |
| 762 | } |
| 763 | } |
| 764 | |
| 765 | static |
| 766 | int lib_ring_buffer_file_open(struct inode *inode, struct file *file) |
| 767 | { |
| 768 | struct lib_ring_buffer *buf = inode->i_private; |
| 769 | int ret; |
| 770 | |
| 771 | ret = lib_ring_buffer_iterator_open(buf); |
| 772 | if (ret) |
| 773 | return ret; |
| 774 | |
| 775 | file->private_data = buf; |
| 776 | ret = nonseekable_open(inode, file); |
| 777 | if (ret) |
| 778 | goto release_iter; |
| 779 | return 0; |
| 780 | |
| 781 | release_iter: |
| 782 | lib_ring_buffer_iterator_release(buf); |
| 783 | return ret; |
| 784 | } |
| 785 | |
| 786 | static |
| 787 | int lib_ring_buffer_file_release(struct inode *inode, struct file *file) |
| 788 | { |
| 789 | struct lib_ring_buffer *buf = inode->i_private; |
| 790 | |
| 791 | lib_ring_buffer_iterator_release(buf); |
| 792 | return 0; |
| 793 | } |
| 794 | |
| 795 | static |
| 796 | int channel_file_open(struct inode *inode, struct file *file) |
| 797 | { |
| 798 | struct channel *chan = inode->i_private; |
| 799 | int ret; |
| 800 | |
| 801 | ret = channel_iterator_open(chan); |
| 802 | if (ret) |
| 803 | return ret; |
| 804 | |
| 805 | file->private_data = chan; |
| 806 | ret = nonseekable_open(inode, file); |
| 807 | if (ret) |
| 808 | goto release_iter; |
| 809 | return 0; |
| 810 | |
| 811 | release_iter: |
| 812 | channel_iterator_release(chan); |
| 813 | return ret; |
| 814 | } |
| 815 | |
| 816 | static |
| 817 | int channel_file_release(struct inode *inode, struct file *file) |
| 818 | { |
| 819 | struct channel *chan = inode->i_private; |
| 820 | |
| 821 | channel_iterator_release(chan); |
| 822 | return 0; |
| 823 | } |
| 824 | |
| 825 | const struct file_operations channel_payload_file_operations = { |
| 826 | .owner = THIS_MODULE, |
| 827 | .open = channel_file_open, |
| 828 | .release = channel_file_release, |
| 829 | .read = channel_file_read, |
| 830 | .llseek = vfs_lib_ring_buffer_no_llseek, |
| 831 | }; |
| 832 | EXPORT_SYMBOL_GPL(channel_payload_file_operations); |
| 833 | |
| 834 | const struct file_operations lib_ring_buffer_payload_file_operations = { |
| 835 | .owner = THIS_MODULE, |
| 836 | .open = lib_ring_buffer_file_open, |
| 837 | .release = lib_ring_buffer_file_release, |
| 838 | .read = lib_ring_buffer_file_read, |
| 839 | .llseek = vfs_lib_ring_buffer_no_llseek, |
| 840 | }; |
| 841 | EXPORT_SYMBOL_GPL(lib_ring_buffer_payload_file_operations); |