| 1 | #ifndef _LIB_RING_BUFFER_FRONTEND_INTERNAL_H |
| 2 | #define _LIB_RING_BUFFER_FRONTEND_INTERNAL_H |
| 3 | |
| 4 | /* |
| 5 | * linux/ringbuffer/frontend_internal.h |
| 6 | * |
| 7 | * Ring Buffer Library Synchronization Header (internal helpers). |
| 8 | * |
| 9 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 10 | * |
| 11 | * This library is free software; you can redistribute it and/or |
| 12 | * modify it under the terms of the GNU Lesser General Public |
| 13 | * License as published by the Free Software Foundation; only |
| 14 | * version 2.1 of the License. |
| 15 | * |
| 16 | * This library is distributed in the hope that it will be useful, |
| 17 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 18 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 19 | * Lesser General Public License for more details. |
| 20 | * |
| 21 | * You should have received a copy of the GNU Lesser General Public |
| 22 | * License along with this library; if not, write to the Free Software |
| 23 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 24 | * |
| 25 | * Author: |
| 26 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 27 | * |
| 28 | * See ring_buffer_frontend.c for more information on wait-free algorithms. |
| 29 | */ |
| 30 | |
| 31 | #include "../../wrapper/ringbuffer/config.h" |
| 32 | #include "../../wrapper/ringbuffer/backend_types.h" |
| 33 | #include "../../wrapper/ringbuffer/frontend_types.h" |
| 34 | #include "../../lib/prio_heap/lttng_prio_heap.h" /* For per-CPU read-side iterator */ |
| 35 | |
| 36 | /* Buffer offset macros */ |
| 37 | |
| 38 | /* buf_trunc mask selects only the buffer number. */ |
| 39 | static inline |
| 40 | unsigned long buf_trunc(unsigned long offset, struct channel *chan) |
| 41 | { |
| 42 | return offset & ~(chan->backend.buf_size - 1); |
| 43 | |
| 44 | } |
| 45 | |
| 46 | /* Select the buffer number value (counter). */ |
| 47 | static inline |
| 48 | unsigned long buf_trunc_val(unsigned long offset, struct channel *chan) |
| 49 | { |
| 50 | return buf_trunc(offset, chan) >> chan->backend.buf_size_order; |
| 51 | } |
| 52 | |
| 53 | /* buf_offset mask selects only the offset within the current buffer. */ |
| 54 | static inline |
| 55 | unsigned long buf_offset(unsigned long offset, struct channel *chan) |
| 56 | { |
| 57 | return offset & (chan->backend.buf_size - 1); |
| 58 | } |
| 59 | |
| 60 | /* subbuf_offset mask selects the offset within the current subbuffer. */ |
| 61 | static inline |
| 62 | unsigned long subbuf_offset(unsigned long offset, struct channel *chan) |
| 63 | { |
| 64 | return offset & (chan->backend.subbuf_size - 1); |
| 65 | } |
| 66 | |
| 67 | /* subbuf_trunc mask selects the subbuffer number. */ |
| 68 | static inline |
| 69 | unsigned long subbuf_trunc(unsigned long offset, struct channel *chan) |
| 70 | { |
| 71 | return offset & ~(chan->backend.subbuf_size - 1); |
| 72 | } |
| 73 | |
| 74 | /* subbuf_align aligns the offset to the next subbuffer. */ |
| 75 | static inline |
| 76 | unsigned long subbuf_align(unsigned long offset, struct channel *chan) |
| 77 | { |
| 78 | return (offset + chan->backend.subbuf_size) |
| 79 | & ~(chan->backend.subbuf_size - 1); |
| 80 | } |
| 81 | |
| 82 | /* subbuf_index returns the index of the current subbuffer within the buffer. */ |
| 83 | static inline |
| 84 | unsigned long subbuf_index(unsigned long offset, struct channel *chan) |
| 85 | { |
| 86 | return buf_offset(offset, chan) >> chan->backend.subbuf_size_order; |
| 87 | } |
| 88 | |
| 89 | /* |
| 90 | * Last TSC comparison functions. Check if the current TSC overflows tsc_bits |
| 91 | * bits from the last TSC read. When overflows are detected, the full 64-bit |
| 92 | * timestamp counter should be written in the record header. Reads and writes |
| 93 | * last_tsc atomically. |
| 94 | */ |
| 95 | |
| 96 | #if (BITS_PER_LONG == 32) |
| 97 | static inline |
| 98 | void save_last_tsc(const struct lib_ring_buffer_config *config, |
| 99 | struct lib_ring_buffer *buf, u64 tsc) |
| 100 | { |
| 101 | if (config->tsc_bits == 0 || config->tsc_bits == 64) |
| 102 | return; |
| 103 | |
| 104 | /* |
| 105 | * Ensure the compiler performs this update in a single instruction. |
| 106 | */ |
| 107 | v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits)); |
| 108 | } |
| 109 | |
| 110 | static inline |
| 111 | int last_tsc_overflow(const struct lib_ring_buffer_config *config, |
| 112 | struct lib_ring_buffer *buf, u64 tsc) |
| 113 | { |
| 114 | unsigned long tsc_shifted; |
| 115 | |
| 116 | if (config->tsc_bits == 0 || config->tsc_bits == 64) |
| 117 | return 0; |
| 118 | |
| 119 | tsc_shifted = (unsigned long)(tsc >> config->tsc_bits); |
| 120 | if (unlikely(tsc_shifted |
| 121 | - (unsigned long)v_read(config, &buf->last_tsc))) |
| 122 | return 1; |
| 123 | else |
| 124 | return 0; |
| 125 | } |
| 126 | #else |
| 127 | static inline |
| 128 | void save_last_tsc(const struct lib_ring_buffer_config *config, |
| 129 | struct lib_ring_buffer *buf, u64 tsc) |
| 130 | { |
| 131 | if (config->tsc_bits == 0 || config->tsc_bits == 64) |
| 132 | return; |
| 133 | |
| 134 | v_set(config, &buf->last_tsc, (unsigned long)tsc); |
| 135 | } |
| 136 | |
| 137 | static inline |
| 138 | int last_tsc_overflow(const struct lib_ring_buffer_config *config, |
| 139 | struct lib_ring_buffer *buf, u64 tsc) |
| 140 | { |
| 141 | if (config->tsc_bits == 0 || config->tsc_bits == 64) |
| 142 | return 0; |
| 143 | |
| 144 | if (unlikely((tsc - v_read(config, &buf->last_tsc)) |
| 145 | >> config->tsc_bits)) |
| 146 | return 1; |
| 147 | else |
| 148 | return 0; |
| 149 | } |
| 150 | #endif |
| 151 | |
| 152 | extern |
| 153 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx); |
| 154 | |
| 155 | extern |
| 156 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, |
| 157 | enum switch_mode mode); |
| 158 | |
| 159 | extern |
| 160 | void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf); |
| 161 | |
| 162 | /* Buffer write helpers */ |
| 163 | |
| 164 | static inline |
| 165 | void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf, |
| 166 | struct channel *chan, |
| 167 | unsigned long offset) |
| 168 | { |
| 169 | unsigned long consumed_old, consumed_new; |
| 170 | |
| 171 | do { |
| 172 | consumed_old = atomic_long_read(&buf->consumed); |
| 173 | /* |
| 174 | * If buffer is in overwrite mode, push the reader consumed |
| 175 | * count if the write position has reached it and we are not |
| 176 | * at the first iteration (don't push the reader farther than |
| 177 | * the writer). This operation can be done concurrently by many |
| 178 | * writers in the same buffer, the writer being at the farthest |
| 179 | * write position sub-buffer index in the buffer being the one |
| 180 | * which will win this loop. |
| 181 | */ |
| 182 | if (unlikely(subbuf_trunc(offset, chan) |
| 183 | - subbuf_trunc(consumed_old, chan) |
| 184 | >= chan->backend.buf_size)) |
| 185 | consumed_new = subbuf_align(consumed_old, chan); |
| 186 | else |
| 187 | return; |
| 188 | } while (unlikely(atomic_long_cmpxchg(&buf->consumed, consumed_old, |
| 189 | consumed_new) != consumed_old)); |
| 190 | } |
| 191 | |
| 192 | static inline |
| 193 | void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config, |
| 194 | struct lib_ring_buffer *buf, |
| 195 | unsigned long commit_count, |
| 196 | unsigned long idx) |
| 197 | { |
| 198 | if (config->oops == RING_BUFFER_OOPS_CONSISTENCY) |
| 199 | v_set(config, &buf->commit_hot[idx].seq, commit_count); |
| 200 | } |
| 201 | |
| 202 | static inline |
| 203 | int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config, |
| 204 | struct lib_ring_buffer *buf, |
| 205 | struct channel *chan) |
| 206 | { |
| 207 | unsigned long consumed_old, consumed_idx, commit_count, write_offset; |
| 208 | |
| 209 | consumed_old = atomic_long_read(&buf->consumed); |
| 210 | consumed_idx = subbuf_index(consumed_old, chan); |
| 211 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); |
| 212 | /* |
| 213 | * No memory barrier here, since we are only interested |
| 214 | * in a statistically correct polling result. The next poll will |
| 215 | * get the data is we are racing. The mb() that ensures correct |
| 216 | * memory order is in get_subbuf. |
| 217 | */ |
| 218 | write_offset = v_read(config, &buf->offset); |
| 219 | |
| 220 | /* |
| 221 | * Check that the subbuffer we are trying to consume has been |
| 222 | * already fully committed. |
| 223 | */ |
| 224 | |
| 225 | if (((commit_count - chan->backend.subbuf_size) |
| 226 | & chan->commit_count_mask) |
| 227 | - (buf_trunc(consumed_old, chan) |
| 228 | >> chan->backend.num_subbuf_order) |
| 229 | != 0) |
| 230 | return 0; |
| 231 | |
| 232 | /* |
| 233 | * Check that we are not about to read the same subbuffer in |
| 234 | * which the writer head is. |
| 235 | */ |
| 236 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) |
| 237 | == 0) |
| 238 | return 0; |
| 239 | |
| 240 | return 1; |
| 241 | |
| 242 | } |
| 243 | |
| 244 | static inline |
| 245 | int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config, |
| 246 | struct lib_ring_buffer *buf, |
| 247 | struct channel *chan) |
| 248 | { |
| 249 | return !!subbuf_offset(v_read(config, &buf->offset), chan); |
| 250 | } |
| 251 | |
| 252 | static inline |
| 253 | unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config, |
| 254 | struct lib_ring_buffer *buf, |
| 255 | unsigned long idx) |
| 256 | { |
| 257 | return subbuffer_get_data_size(config, &buf->backend, idx); |
| 258 | } |
| 259 | |
| 260 | /* |
| 261 | * Check if all space reservation in a buffer have been committed. This helps |
| 262 | * knowing if an execution context is nested (for per-cpu buffers only). |
| 263 | * This is a very specific ftrace use-case, so we keep this as "internal" API. |
| 264 | */ |
| 265 | static inline |
| 266 | int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config, |
| 267 | struct lib_ring_buffer *buf, |
| 268 | struct channel *chan) |
| 269 | { |
| 270 | unsigned long offset, idx, commit_count; |
| 271 | |
| 272 | CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU); |
| 273 | CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU); |
| 274 | |
| 275 | /* |
| 276 | * Read offset and commit count in a loop so they are both read |
| 277 | * atomically wrt interrupts. By deal with interrupt concurrency by |
| 278 | * restarting both reads if the offset has been pushed. Note that given |
| 279 | * we only have to deal with interrupt concurrency here, an interrupt |
| 280 | * modifying the commit count will also modify "offset", so it is safe |
| 281 | * to only check for offset modifications. |
| 282 | */ |
| 283 | do { |
| 284 | offset = v_read(config, &buf->offset); |
| 285 | idx = subbuf_index(offset, chan); |
| 286 | commit_count = v_read(config, &buf->commit_hot[idx].cc); |
| 287 | } while (offset != v_read(config, &buf->offset)); |
| 288 | |
| 289 | return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order) |
| 290 | - (commit_count & chan->commit_count_mask) == 0); |
| 291 | } |
| 292 | |
| 293 | /* |
| 294 | * Receive end of subbuffer TSC as parameter. It has been read in the |
| 295 | * space reservation loop of either reserve or switch, which ensures it |
| 296 | * progresses monotonically with event records in the buffer. Therefore, |
| 297 | * it ensures that the end timestamp of a subbuffer is <= begin |
| 298 | * timestamp of the following subbuffers. |
| 299 | */ |
| 300 | static inline |
| 301 | void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config, |
| 302 | struct lib_ring_buffer *buf, |
| 303 | struct channel *chan, |
| 304 | unsigned long offset, |
| 305 | unsigned long commit_count, |
| 306 | unsigned long idx, |
| 307 | u64 tsc) |
| 308 | { |
| 309 | unsigned long old_commit_count = commit_count |
| 310 | - chan->backend.subbuf_size; |
| 311 | |
| 312 | /* Check if all commits have been done */ |
| 313 | if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order) |
| 314 | - (old_commit_count & chan->commit_count_mask) == 0)) { |
| 315 | /* |
| 316 | * If we succeeded at updating cc_sb below, we are the subbuffer |
| 317 | * writer delivering the subbuffer. Deals with concurrent |
| 318 | * updates of the "cc" value without adding a add_return atomic |
| 319 | * operation to the fast path. |
| 320 | * |
| 321 | * We are doing the delivery in two steps: |
| 322 | * - First, we cmpxchg() cc_sb to the new value |
| 323 | * old_commit_count + 1. This ensures that we are the only |
| 324 | * subbuffer user successfully filling the subbuffer, but we |
| 325 | * do _not_ set the cc_sb value to "commit_count" yet. |
| 326 | * Therefore, other writers that would wrap around the ring |
| 327 | * buffer and try to start writing to our subbuffer would |
| 328 | * have to drop records, because it would appear as |
| 329 | * non-filled. |
| 330 | * We therefore have exclusive access to the subbuffer control |
| 331 | * structures. This mutual exclusion with other writers is |
| 332 | * crucially important to perform record overruns count in |
| 333 | * flight recorder mode locklessly. |
| 334 | * - When we are ready to release the subbuffer (either for |
| 335 | * reading or for overrun by other writers), we simply set the |
| 336 | * cc_sb value to "commit_count" and perform delivery. |
| 337 | * |
| 338 | * The subbuffer size is least 2 bytes (minimum size: 1 page). |
| 339 | * This guarantees that old_commit_count + 1 != commit_count. |
| 340 | */ |
| 341 | |
| 342 | /* |
| 343 | * Order prior updates to reserve count prior to the |
| 344 | * commit_cold cc_sb update. |
| 345 | */ |
| 346 | smp_wmb(); |
| 347 | if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb, |
| 348 | old_commit_count, old_commit_count + 1) |
| 349 | == old_commit_count)) { |
| 350 | /* |
| 351 | * Start of exclusive subbuffer access. We are |
| 352 | * guaranteed to be the last writer in this subbuffer |
| 353 | * and any other writer trying to access this subbuffer |
| 354 | * in this state is required to drop records. |
| 355 | */ |
| 356 | v_add(config, |
| 357 | subbuffer_get_records_count(config, |
| 358 | &buf->backend, idx), |
| 359 | &buf->records_count); |
| 360 | v_add(config, |
| 361 | subbuffer_count_records_overrun(config, |
| 362 | &buf->backend, |
| 363 | idx), |
| 364 | &buf->records_overrun); |
| 365 | config->cb.buffer_end(buf, tsc, idx, |
| 366 | lib_ring_buffer_get_data_size(config, |
| 367 | buf, |
| 368 | idx)); |
| 369 | |
| 370 | /* |
| 371 | * Set noref flag and offset for this subbuffer id. |
| 372 | * Contains a memory barrier that ensures counter stores |
| 373 | * are ordered before set noref and offset. |
| 374 | */ |
| 375 | lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, |
| 376 | buf_trunc_val(offset, chan)); |
| 377 | |
| 378 | /* |
| 379 | * Order set_noref and record counter updates before the |
| 380 | * end of subbuffer exclusive access. Orders with |
| 381 | * respect to writers coming into the subbuffer after |
| 382 | * wrap around, and also order wrt concurrent readers. |
| 383 | */ |
| 384 | smp_mb(); |
| 385 | /* End of exclusive subbuffer access */ |
| 386 | v_set(config, &buf->commit_cold[idx].cc_sb, |
| 387 | commit_count); |
| 388 | /* |
| 389 | * Order later updates to reserve count after |
| 390 | * the commit_cold cc_sb update. |
| 391 | */ |
| 392 | smp_wmb(); |
| 393 | lib_ring_buffer_vmcore_check_deliver(config, buf, |
| 394 | commit_count, idx); |
| 395 | |
| 396 | /* |
| 397 | * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. |
| 398 | */ |
| 399 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER |
| 400 | && atomic_long_read(&buf->active_readers) |
| 401 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { |
| 402 | wake_up_interruptible(&buf->read_wait); |
| 403 | wake_up_interruptible(&chan->read_wait); |
| 404 | } |
| 405 | |
| 406 | } |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | /* |
| 411 | * lib_ring_buffer_write_commit_counter |
| 412 | * |
| 413 | * For flight recording. must be called after commit. |
| 414 | * This function increments the subbuffer's commit_seq counter each time the |
| 415 | * commit count reaches back the reserve offset (modulo subbuffer size). It is |
| 416 | * useful for crash dump. |
| 417 | */ |
| 418 | static inline |
| 419 | void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config, |
| 420 | struct lib_ring_buffer *buf, |
| 421 | struct channel *chan, |
| 422 | unsigned long idx, |
| 423 | unsigned long buf_offset, |
| 424 | unsigned long commit_count) |
| 425 | { |
| 426 | unsigned long commit_seq_old; |
| 427 | |
| 428 | if (config->oops != RING_BUFFER_OOPS_CONSISTENCY) |
| 429 | return; |
| 430 | |
| 431 | /* |
| 432 | * subbuf_offset includes commit_count_mask. We can simply |
| 433 | * compare the offsets within the subbuffer without caring about |
| 434 | * buffer full/empty mismatch because offset is never zero here |
| 435 | * (subbuffer header and record headers have non-zero length). |
| 436 | */ |
| 437 | if (unlikely(subbuf_offset(buf_offset - commit_count, chan))) |
| 438 | return; |
| 439 | |
| 440 | commit_seq_old = v_read(config, &buf->commit_hot[idx].seq); |
| 441 | while ((long) (commit_seq_old - commit_count) < 0) |
| 442 | commit_seq_old = v_cmpxchg(config, &buf->commit_hot[idx].seq, |
| 443 | commit_seq_old, commit_count); |
| 444 | } |
| 445 | |
| 446 | extern int lib_ring_buffer_create(struct lib_ring_buffer *buf, |
| 447 | struct channel_backend *chanb, int cpu); |
| 448 | extern void lib_ring_buffer_free(struct lib_ring_buffer *buf); |
| 449 | |
| 450 | /* Keep track of trap nesting inside ring buffer code */ |
| 451 | DECLARE_PER_CPU(unsigned int, lib_ring_buffer_nesting); |
| 452 | |
| 453 | #endif /* _LIB_RING_BUFFER_FRONTEND_INTERNAL_H */ |