libringbuffer/frontend_internal.h

   1 #ifndef _LTTNG_RING_BUFFER_FRONTEND_INTERNAL_H
   2 #define _LTTNG_RING_BUFFER_FRONTEND_INTERNAL_H
   3
   4 /*
   5  * libringbuffer/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  *
  26  * Author:
  27  *      Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  28  *
  29  * See ring_buffer_frontend.c for more information on wait-free algorithms.
  30  *
  31  * Dual LGPL v2.1/GPL v2 license.
  32  */
  33
  34 #include <urcu/compiler.h>
  35
  36 #include <lttng/ringbuffer-config.h>
  37 #include "backend_types.h"
  38 #include "frontend_types.h"
  39 #include "shm.h"
  40
  41 /* Buffer offset macros */
  42
  43 /* buf_trunc mask selects only the buffer number. */
  44 static inline
  45 unsigned long buf_trunc(unsigned long offset, struct channel *chan)
  46 {
  47         return offset & ~(chan->backend.buf_size - 1);
  48
  49 }
  50
  51 /* Select the buffer number value (counter). */
  52 static inline
  53 unsigned long buf_trunc_val(unsigned long offset, struct channel *chan)
  54 {
  55         return buf_trunc(offset, chan) >> chan->backend.buf_size_order;
  56 }
  57
  58 /* buf_offset mask selects only the offset within the current buffer. */
  59 static inline
  60 unsigned long buf_offset(unsigned long offset, struct channel *chan)
  61 {
  62         return offset & (chan->backend.buf_size - 1);
  63 }
  64
  65 /* subbuf_offset mask selects the offset within the current subbuffer. */
  66 static inline
  67 unsigned long subbuf_offset(unsigned long offset, struct channel *chan)
  68 {
  69         return offset & (chan->backend.subbuf_size - 1);
  70 }
  71
  72 /* subbuf_trunc mask selects the subbuffer number. */
  73 static inline
  74 unsigned long subbuf_trunc(unsigned long offset, struct channel *chan)
  75 {
  76         return offset & ~(chan->backend.subbuf_size - 1);
  77 }
  78
  79 /* subbuf_align aligns the offset to the next subbuffer. */
  80 static inline
  81 unsigned long subbuf_align(unsigned long offset, struct channel *chan)
  82 {
  83         return (offset + chan->backend.subbuf_size)
  84                & ~(chan->backend.subbuf_size - 1);
  85 }
  86
  87 /* subbuf_index returns the index of the current subbuffer within the buffer. */
  88 static inline
  89 unsigned long subbuf_index(unsigned long offset, struct channel *chan)
  90 {
  91         return buf_offset(offset, chan) >> chan->backend.subbuf_size_order;
  92 }
  93
  94 /*
  95  * Last TSC comparison functions. Check if the current TSC overflows tsc_bits
  96  * bits from the last TSC read. When overflows are detected, the full 64-bit
  97  * timestamp counter should be written in the record header. Reads and writes
  98  * last_tsc atomically.
  99  */
 100
 101 #if (CAA_BITS_PER_LONG == 32)
 102 static inline
 103 void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
 104                    struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 105 {
 106         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 107                 return;
 108
 109         /*
 110          * Ensure the compiler performs this update in a single instruction.
 111          */
 112         v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits));
 113 }
 114
 115 static inline
 116 int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
 117                       struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 118 {
 119         unsigned long tsc_shifted;
 120
 121         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 122                 return 0;
 123
 124         tsc_shifted = (unsigned long)(tsc >> config->tsc_bits);
 125         if (caa_unlikely(tsc_shifted
 126                      - (unsigned long)v_read(config, &buf->last_tsc)))
 127                 return 1;
 128         else
 129                 return 0;
 130 }
 131 #else
 132 static inline
 133 void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
 134                    struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 135 {
 136         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 137                 return;
 138
 139         v_set(config, &buf->last_tsc, (unsigned long)tsc);
 140 }
 141
 142 static inline
 143 int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
 144                       struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 145 {
 146         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 147                 return 0;
 148
 149         if (caa_unlikely((tsc - v_read(config, &buf->last_tsc))
 150                      >> config->tsc_bits))
 151                 return 1;
 152         else
 153                 return 0;
 154 }
 155 #endif
 156
 157 extern
 158 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx);
 159
 160 extern
 161 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf,
 162                                  enum switch_mode mode,
 163                                  struct lttng_ust_shm_handle *handle);
 164
 165 /* Buffer write helpers */
 166
 167 static inline
 168 void lib_ring_buffer_reserve_push_reader(struct lttng_ust_lib_ring_buffer *buf,
 169                                          struct channel *chan,
 170                                          unsigned long offset)
 171 {
 172         unsigned long consumed_old, consumed_new;
 173
 174         do {
 175                 consumed_old = uatomic_read(&buf->consumed);
 176                 /*
 177                  * If buffer is in overwrite mode, push the reader consumed
 178                  * count if the write position has reached it and we are not
 179                  * at the first iteration (don't push the reader farther than
 180                  * the writer). This operation can be done concurrently by many
 181                  * writers in the same buffer, the writer being at the farthest
 182                  * write position sub-buffer index in the buffer being the one
 183                  * which will win this loop.
 184                  */
 185                 if (caa_unlikely(subbuf_trunc(offset, chan)
 186                               - subbuf_trunc(consumed_old, chan)
 187                              >= chan->backend.buf_size))
 188                         consumed_new = subbuf_align(consumed_old, chan);
 189                 else
 190                         return;
 191         } while (caa_unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old,
 192                                               consumed_new) != consumed_old));
 193 }
 194
 195 static inline
 196 void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
 197                                           struct lttng_ust_lib_ring_buffer *buf,
 198                                           unsigned long commit_count,
 199                                           unsigned long idx,
 200                                           struct lttng_ust_shm_handle *handle)
 201 {
 202         if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
 203                 v_set(config, &shmp_index(handle, buf->commit_hot, idx)->seq, commit_count);
 204 }
 205
 206 static inline
 207 int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
 208                                  struct lttng_ust_lib_ring_buffer *buf,
 209                                  struct channel *chan,
 210                                  struct lttng_ust_shm_handle *handle)
 211 {
 212         unsigned long consumed_old, consumed_idx, commit_count, write_offset;
 213
 214         consumed_old = uatomic_read(&buf->consumed);
 215         consumed_idx = subbuf_index(consumed_old, chan);
 216         commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
 217         /*
 218          * No memory barrier here, since we are only interested
 219          * in a statistically correct polling result. The next poll will
 220          * get the data is we are racing. The mb() that ensures correct
 221          * memory order is in get_subbuf.
 222          */
 223         write_offset = v_read(config, &buf->offset);
 224
 225         /*
 226          * Check that the subbuffer we are trying to consume has been
 227          * already fully committed.
 228          */
 229
 230         if (((commit_count - chan->backend.subbuf_size)
 231              & chan->commit_count_mask)
 232             - (buf_trunc(consumed_old, chan)
 233                >> chan->backend.num_subbuf_order)
 234             != 0)
 235                 return 0;
 236
 237         /*
 238          * Check that we are not about to read the same subbuffer in
 239          * which the writer head is.
 240          */
 241         if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
 242             == 0)
 243                 return 0;
 244
 245         return 1;
 246
 247 }
 248
 249 static inline
 250 int lib_ring_buffer_pending_data(const struct lttng_ust_lib_ring_buffer_config *config,
 251                                  struct lttng_ust_lib_ring_buffer *buf,
 252                                  struct channel *chan)
 253 {
 254         return !!subbuf_offset(v_read(config, &buf->offset), chan);
 255 }
 256
 257 static inline
 258 unsigned long lib_ring_buffer_get_data_size(const struct lttng_ust_lib_ring_buffer_config *config,
 259                                             struct lttng_ust_lib_ring_buffer *buf,
 260                                             unsigned long idx,
 261                                             struct lttng_ust_shm_handle *handle)
 262 {
 263         return subbuffer_get_data_size(config, &buf->backend, idx, handle);
 264 }
 265
 266 /*
 267  * Check if all space reservation in a buffer have been committed. This helps
 268  * knowing if an execution context is nested (for per-cpu buffers only).
 269  * This is a very specific ftrace use-case, so we keep this as "internal" API.
 270  */
 271 static inline
 272 int lib_ring_buffer_reserve_committed(const struct lttng_ust_lib_ring_buffer_config *config,
 273                                       struct lttng_ust_lib_ring_buffer *buf,
 274                                       struct channel *chan,
 275                                       struct lttng_ust_shm_handle *handle)
 276 {
 277         unsigned long offset, idx, commit_count;
 278
 279         CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
 280         CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
 281
 282         /*
 283          * Read offset and commit count in a loop so they are both read
 284          * atomically wrt interrupts. By deal with interrupt concurrency by
 285          * restarting both reads if the offset has been pushed. Note that given
 286          * we only have to deal with interrupt concurrency here, an interrupt
 287          * modifying the commit count will also modify "offset", so it is safe
 288          * to only check for offset modifications.
 289          */
 290         do {
 291                 offset = v_read(config, &buf->offset);
 292                 idx = subbuf_index(offset, chan);
 293                 commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->cc);
 294         } while (offset != v_read(config, &buf->offset));
 295
 296         return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 297                      - (commit_count & chan->commit_count_mask) == 0);
 298 }
 299
 300 static inline
 301 void lib_ring_buffer_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
 302                                    struct lttng_ust_lib_ring_buffer *buf,
 303                                    struct channel *chan,
 304                                    unsigned long offset,
 305                                    unsigned long commit_count,
 306                                    unsigned long idx,
 307                                    struct lttng_ust_shm_handle *handle)
 308 {
 309         unsigned long old_commit_count = commit_count
 310                                          - chan->backend.subbuf_size;
 311         uint64_t tsc;
 312
 313         /* Check if all commits have been done */
 314         if (caa_unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 315                      - (old_commit_count & chan->commit_count_mask) == 0)) {
 316                 /*
 317                  * If we succeeded at updating cc_sb below, we are the subbuffer
 318                  * writer delivering the subbuffer. Deals with concurrent
 319                  * updates of the "cc" value without adding a add_return atomic
 320                  * operation to the fast path.
 321                  *
 322                  * We are doing the delivery in two steps:
 323                  * - First, we cmpxchg() cc_sb to the new value
 324                  *   old_commit_count + 1. This ensures that we are the only
 325                  *   subbuffer user successfully filling the subbuffer, but we
 326                  *   do _not_ set the cc_sb value to "commit_count" yet.
 327                  *   Therefore, other writers that would wrap around the ring
 328                  *   buffer and try to start writing to our subbuffer would
 329                  *   have to drop records, because it would appear as
 330                  *   non-filled.
 331                  *   We therefore have exclusive access to the subbuffer control
 332                  *   structures.  This mutual exclusion with other writers is
 333                  *   crucially important to perform record overruns count in
 334                  *   flight recorder mode locklessly.
 335                  * - When we are ready to release the subbuffer (either for
 336                  *   reading or for overrun by other writers), we simply set the
 337                  *   cc_sb value to "commit_count" and perform delivery.
 338                  *
 339                  * The subbuffer size is least 2 bytes (minimum size: 1 page).
 340                  * This guarantees that old_commit_count + 1 != commit_count.
 341                  */
 342                 if (caa_likely(v_cmpxchg(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
 343                                          old_commit_count, old_commit_count + 1)
 344                            == old_commit_count)) {
 345                         /*
 346                          * Start of exclusive subbuffer access. We are
 347                          * guaranteed to be the last writer in this subbuffer
 348                          * and any other writer trying to access this subbuffer
 349                          * in this state is required to drop records.
 350                          */
 351                         tsc = config->cb.ring_buffer_clock_read(chan);
 352                         v_add(config,
 353                               subbuffer_get_records_count(config,
 354                                                           &buf->backend,
 355                                                           idx, handle),
 356                               &buf->records_count);
 357                         v_add(config,
 358                               subbuffer_count_records_overrun(config,
 359                                                               &buf->backend,
 360                                                               idx, handle),
 361                               &buf->records_overrun);
 362                         config->cb.buffer_end(buf, tsc, idx,
 363                                               lib_ring_buffer_get_data_size(config,
 364                                                                         buf,
 365                                                                         idx,
 366                                                                         handle),
 367                                               handle);
 368
 369                         /*
 370                          * Set noref flag and offset for this subbuffer id.
 371                          * Contains a memory barrier that ensures counter stores
 372                          * are ordered before set noref and offset.
 373                          */
 374                         lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
 375                                                          buf_trunc_val(offset, chan), handle);
 376
 377                         /*
 378                          * Order set_noref and record counter updates before the
 379                          * end of subbuffer exclusive access. Orders with
 380                          * respect to writers coming into the subbuffer after
 381                          * wrap around, and also order wrt concurrent readers.
 382                          */
 383                         cmm_smp_mb();
 384                         /* End of exclusive subbuffer access */
 385                         v_set(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
 386                               commit_count);
 387                         lib_ring_buffer_vmcore_check_deliver(config, buf,
 388                                                  commit_count, idx, handle);
 389
 390                         /*
 391                          * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
 392                          */
 393                         if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
 394                             && (uatomic_read(&buf->active_readers)
 395                                 || uatomic_read(&buf->active_shadow_readers))
 396                             && lib_ring_buffer_poll_deliver(config, buf, chan, handle)) {
 397                                 int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);
 398
 399                                 if (wakeup_fd >= 0) {
 400                                         int ret;
 401                                         /*
 402                                          * Wake-up the other end by
 403                                          * writing a null byte in the
 404                                          * pipe (non-blocking).
 405                                          * Important note: Because
 406                                          * writing into the pipe is
 407                                          * non-blocking (and therefore
 408                                          * we allow dropping wakeup
 409                                          * data, as long as there is
 410                                          * wakeup data present in the
 411                                          * pipe buffer to wake up the
 412                                          * consumer), the consumer
 413                                          * should perform the following
 414                                          * sequence for waiting:
 415                                          * 1) empty the pipe (reads).
 416                                          * 2) check if there is data in
 417                                          *    the buffer.
 418                                          * 3) wait on the pipe (poll).
 419                                          */
 420                                         do {
 421                                                 ret = write(wakeup_fd, "", 1);
 422                                         } while (ret == -1L && errno == EINTR);
 423                                 }
 424                         }
 425
 426                 }
 427         }
 428 }
 429
 430 /*
 431  * lib_ring_buffer_write_commit_counter
 432  *
 433  * For flight recording. must be called after commit.
 434  * This function increments the subbuffer's commit_seq counter each time the
 435  * commit count reaches back the reserve offset (modulo subbuffer size). It is
 436  * useful for crash dump.
 437  */
 438 static inline
 439 void lib_ring_buffer_write_commit_counter(const struct lttng_ust_lib_ring_buffer_config *config,
 440                                           struct lttng_ust_lib_ring_buffer *buf,
 441                                           struct channel *chan,
 442                                           unsigned long idx,
 443                                           unsigned long buf_offset,
 444                                           unsigned long commit_count,
 445                                           size_t slot_size,
 446                                           struct lttng_ust_shm_handle *handle)
 447 {
 448         unsigned long offset, commit_seq_old;
 449
 450         if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
 451                 return;
 452
 453         offset = buf_offset + slot_size;
 454
 455         /*
 456          * subbuf_offset includes commit_count_mask. We can simply
 457          * compare the offsets within the subbuffer without caring about
 458          * buffer full/empty mismatch because offset is never zero here
 459          * (subbuffer header and record headers have non-zero length).
 460          */
 461         if (caa_unlikely(subbuf_offset(offset - commit_count, chan)))
 462                 return;
 463
 464         commit_seq_old = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->seq);
 465         while ((long) (commit_seq_old - commit_count) < 0)
 466                 commit_seq_old = v_cmpxchg(config, &shmp_index(handle, buf->commit_hot, idx)->seq,
 467                                            commit_seq_old, commit_count);
 468 }
 469
 470 extern int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf,
 471                                   struct channel_backend *chanb, int cpu,
 472                                   struct lttng_ust_shm_handle *handle,
 473                                   struct shm_object *shmobj);
 474 extern void lib_ring_buffer_free(struct lttng_ust_lib_ring_buffer *buf,
 475                                  struct lttng_ust_shm_handle *handle);
 476
 477 /* Keep track of trap nesting inside ring buffer code */
 478 extern __thread unsigned int lib_ring_buffer_nesting;
 479
 480 #endif /* _LTTNG_RING_BUFFER_FRONTEND_INTERNAL_H */