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 #include <urcu/tls-compat.h>
  36 #include <signal.h>
  37 #include <pthread.h>
  38
  39 #include <lttng/ringbuffer-config.h>
  40 #include "backend_types.h"
  41 #include "frontend_types.h"
  42 #include "shm.h"
  43
  44 /* Buffer offset macros */
  45
  46 /* buf_trunc mask selects only the buffer number. */
  47 static inline
  48 unsigned long buf_trunc(unsigned long offset, struct channel *chan)
  49 {
  50         return offset & ~(chan->backend.buf_size - 1);
  51
  52 }
  53
  54 /* Select the buffer number value (counter). */
  55 static inline
  56 unsigned long buf_trunc_val(unsigned long offset, struct channel *chan)
  57 {
  58         return buf_trunc(offset, chan) >> chan->backend.buf_size_order;
  59 }
  60
  61 /* buf_offset mask selects only the offset within the current buffer. */
  62 static inline
  63 unsigned long buf_offset(unsigned long offset, struct channel *chan)
  64 {
  65         return offset & (chan->backend.buf_size - 1);
  66 }
  67
  68 /* subbuf_offset mask selects the offset within the current subbuffer. */
  69 static inline
  70 unsigned long subbuf_offset(unsigned long offset, struct channel *chan)
  71 {
  72         return offset & (chan->backend.subbuf_size - 1);
  73 }
  74
  75 /* subbuf_trunc mask selects the subbuffer number. */
  76 static inline
  77 unsigned long subbuf_trunc(unsigned long offset, struct channel *chan)
  78 {
  79         return offset & ~(chan->backend.subbuf_size - 1);
  80 }
  81
  82 /* subbuf_align aligns the offset to the next subbuffer. */
  83 static inline
  84 unsigned long subbuf_align(unsigned long offset, struct channel *chan)
  85 {
  86         return (offset + chan->backend.subbuf_size)
  87                & ~(chan->backend.subbuf_size - 1);
  88 }
  89
  90 /* subbuf_index returns the index of the current subbuffer within the buffer. */
  91 static inline
  92 unsigned long subbuf_index(unsigned long offset, struct channel *chan)
  93 {
  94         return buf_offset(offset, chan) >> chan->backend.subbuf_size_order;
  95 }
  96
  97 /*
  98  * Last TSC comparison functions. Check if the current TSC overflows tsc_bits
  99  * bits from the last TSC read. When overflows are detected, the full 64-bit
 100  * timestamp counter should be written in the record header. Reads and writes
 101  * last_tsc atomically.
 102  */
 103
 104 #if (CAA_BITS_PER_LONG == 32)
 105 static inline
 106 void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
 107                    struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 108 {
 109         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 110                 return;
 111
 112         /*
 113          * Ensure the compiler performs this update in a single instruction.
 114          */
 115         v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits));
 116 }
 117
 118 static inline
 119 int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
 120                       struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 121 {
 122         unsigned long tsc_shifted;
 123
 124         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 125                 return 0;
 126
 127         tsc_shifted = (unsigned long)(tsc >> config->tsc_bits);
 128         if (caa_unlikely(tsc_shifted
 129                      - (unsigned long)v_read(config, &buf->last_tsc)))
 130                 return 1;
 131         else
 132                 return 0;
 133 }
 134 #else
 135 static inline
 136 void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
 137                    struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 138 {
 139         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 140                 return;
 141
 142         v_set(config, &buf->last_tsc, (unsigned long)tsc);
 143 }
 144
 145 static inline
 146 int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
 147                       struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc)
 148 {
 149         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 150                 return 0;
 151
 152         if (caa_unlikely((tsc - v_read(config, &buf->last_tsc))
 153                      >> config->tsc_bits))
 154                 return 1;
 155         else
 156                 return 0;
 157 }
 158 #endif
 159
 160 extern
 161 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx);
 162
 163 extern
 164 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf,
 165                                  enum switch_mode mode,
 166                                  struct lttng_ust_shm_handle *handle);
 167
 168 /* Buffer write helpers */
 169
 170 static inline
 171 void lib_ring_buffer_reserve_push_reader(struct lttng_ust_lib_ring_buffer *buf,
 172                                          struct channel *chan,
 173                                          unsigned long offset)
 174 {
 175         unsigned long consumed_old, consumed_new;
 176
 177         do {
 178                 consumed_old = uatomic_read(&buf->consumed);
 179                 /*
 180                  * If buffer is in overwrite mode, push the reader consumed
 181                  * count if the write position has reached it and we are not
 182                  * at the first iteration (don't push the reader farther than
 183                  * the writer). This operation can be done concurrently by many
 184                  * writers in the same buffer, the writer being at the farthest
 185                  * write position sub-buffer index in the buffer being the one
 186                  * which will win this loop.
 187                  */
 188                 if (caa_unlikely(subbuf_trunc(offset, chan)
 189                               - subbuf_trunc(consumed_old, chan)
 190                              >= chan->backend.buf_size))
 191                         consumed_new = subbuf_align(consumed_old, chan);
 192                 else
 193                         return;
 194         } while (caa_unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old,
 195                                               consumed_new) != consumed_old));
 196 }
 197
 198 static inline
 199 void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
 200                                           struct lttng_ust_lib_ring_buffer *buf,
 201                                           unsigned long commit_count,
 202                                           unsigned long idx,
 203                                           struct lttng_ust_shm_handle *handle)
 204 {
 205         if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
 206                 v_set(config, &shmp_index(handle, buf->commit_hot, idx)->seq, commit_count);
 207 }
 208
 209 static inline
 210 int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
 211                                  struct lttng_ust_lib_ring_buffer *buf,
 212                                  struct channel *chan,
 213                                  struct lttng_ust_shm_handle *handle)
 214 {
 215         unsigned long consumed_old, consumed_idx, commit_count, write_offset;
 216
 217         consumed_old = uatomic_read(&buf->consumed);
 218         consumed_idx = subbuf_index(consumed_old, chan);
 219         commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
 220         /*
 221          * No memory barrier here, since we are only interested
 222          * in a statistically correct polling result. The next poll will
 223          * get the data is we are racing. The mb() that ensures correct
 224          * memory order is in get_subbuf.
 225          */
 226         write_offset = v_read(config, &buf->offset);
 227
 228         /*
 229          * Check that the subbuffer we are trying to consume has been
 230          * already fully committed.
 231          */
 232
 233         if (((commit_count - chan->backend.subbuf_size)
 234              & chan->commit_count_mask)
 235             - (buf_trunc(consumed_old, chan)
 236                >> chan->backend.num_subbuf_order)
 237             != 0)
 238                 return 0;
 239
 240         /*
 241          * Check that we are not about to read the same subbuffer in
 242          * which the writer head is.
 243          */
 244         if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
 245             == 0)
 246                 return 0;
 247
 248         return 1;
 249
 250 }
 251
 252 static inline
 253 int lib_ring_buffer_pending_data(const struct lttng_ust_lib_ring_buffer_config *config,
 254                                  struct lttng_ust_lib_ring_buffer *buf,
 255                                  struct channel *chan)
 256 {
 257         return !!subbuf_offset(v_read(config, &buf->offset), chan);
 258 }
 259
 260 static inline
 261 unsigned long lib_ring_buffer_get_data_size(const struct lttng_ust_lib_ring_buffer_config *config,
 262                                             struct lttng_ust_lib_ring_buffer *buf,
 263                                             unsigned long idx,
 264                                             struct lttng_ust_shm_handle *handle)
 265 {
 266         return subbuffer_get_data_size(config, &buf->backend, idx, handle);
 267 }
 268
 269 /*
 270  * Check if all space reservation in a buffer have been committed. This helps
 271  * knowing if an execution context is nested (for per-cpu buffers only).
 272  * This is a very specific ftrace use-case, so we keep this as "internal" API.
 273  */
 274 static inline
 275 int lib_ring_buffer_reserve_committed(const struct lttng_ust_lib_ring_buffer_config *config,
 276                                       struct lttng_ust_lib_ring_buffer *buf,
 277                                       struct channel *chan,
 278                                       struct lttng_ust_shm_handle *handle)
 279 {
 280         unsigned long offset, idx, commit_count;
 281
 282         CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
 283         CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
 284
 285         /*
 286          * Read offset and commit count in a loop so they are both read
 287          * atomically wrt interrupts. By deal with interrupt concurrency by
 288          * restarting both reads if the offset has been pushed. Note that given
 289          * we only have to deal with interrupt concurrency here, an interrupt
 290          * modifying the commit count will also modify "offset", so it is safe
 291          * to only check for offset modifications.
 292          */
 293         do {
 294                 offset = v_read(config, &buf->offset);
 295                 idx = subbuf_index(offset, chan);
 296                 commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->cc);
 297         } while (offset != v_read(config, &buf->offset));
 298
 299         return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 300                      - (commit_count & chan->commit_count_mask) == 0);
 301 }
 302
 303 static inline
 304 void lib_ring_buffer_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
 305                                    struct lttng_ust_lib_ring_buffer *buf,
 306                                    struct channel *chan,
 307                                    unsigned long offset,
 308                                    unsigned long commit_count,
 309                                    unsigned long idx,
 310                                    struct lttng_ust_shm_handle *handle)
 311 {
 312         unsigned long old_commit_count = commit_count
 313                                          - chan->backend.subbuf_size;
 314         uint64_t tsc;
 315
 316         /* Check if all commits have been done */
 317         if (caa_unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 318                      - (old_commit_count & chan->commit_count_mask) == 0)) {
 319                 /*
 320                  * If we succeeded at updating cc_sb below, we are the subbuffer
 321                  * writer delivering the subbuffer. Deals with concurrent
 322                  * updates of the "cc" value without adding a add_return atomic
 323                  * operation to the fast path.
 324                  *
 325                  * We are doing the delivery in two steps:
 326                  * - First, we cmpxchg() cc_sb to the new value
 327                  *   old_commit_count + 1. This ensures that we are the only
 328                  *   subbuffer user successfully filling the subbuffer, but we
 329                  *   do _not_ set the cc_sb value to "commit_count" yet.
 330                  *   Therefore, other writers that would wrap around the ring
 331                  *   buffer and try to start writing to our subbuffer would
 332                  *   have to drop records, because it would appear as
 333                  *   non-filled.
 334                  *   We therefore have exclusive access to the subbuffer control
 335                  *   structures.  This mutual exclusion with other writers is
 336                  *   crucially important to perform record overruns count in
 337                  *   flight recorder mode locklessly.
 338                  * - When we are ready to release the subbuffer (either for
 339                  *   reading or for overrun by other writers), we simply set the
 340                  *   cc_sb value to "commit_count" and perform delivery.
 341                  *
 342                  * The subbuffer size is least 2 bytes (minimum size: 1 page).
 343                  * This guarantees that old_commit_count + 1 != commit_count.
 344                  */
 345                 if (caa_likely(v_cmpxchg(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
 346                                          old_commit_count, old_commit_count + 1)
 347                            == old_commit_count)) {
 348                         /*
 349                          * Start of exclusive subbuffer access. We are
 350                          * guaranteed to be the last writer in this subbuffer
 351                          * and any other writer trying to access this subbuffer
 352                          * in this state is required to drop records.
 353                          */
 354                         tsc = config->cb.ring_buffer_clock_read(chan);
 355                         v_add(config,
 356                               subbuffer_get_records_count(config,
 357                                                           &buf->backend,
 358                                                           idx, handle),
 359                               &buf->records_count);
 360                         v_add(config,
 361                               subbuffer_count_records_overrun(config,
 362                                                               &buf->backend,
 363                                                               idx, handle),
 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                                                                         handle),
 370                                               handle);
 371
 372                         /*
 373                          * Set noref flag and offset for this subbuffer id.
 374                          * Contains a memory barrier that ensures counter stores
 375                          * are ordered before set noref and offset.
 376                          */
 377                         lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
 378                                                          buf_trunc_val(offset, chan), handle);
 379
 380                         /*
 381                          * Order set_noref and record counter updates before the
 382                          * end of subbuffer exclusive access. Orders with
 383                          * respect to writers coming into the subbuffer after
 384                          * wrap around, and also order wrt concurrent readers.
 385                          */
 386                         cmm_smp_mb();
 387                         /* End of exclusive subbuffer access */
 388                         v_set(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
 389                               commit_count);
 390                         lib_ring_buffer_vmcore_check_deliver(config, buf,
 391                                                  commit_count, idx, handle);
 392
 393                         /*
 394                          * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
 395                          */
 396                         if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
 397                             && uatomic_read(&buf->active_readers)
 398                             && lib_ring_buffer_poll_deliver(config, buf, chan, handle)) {
 399                                 int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);
 400
 401                                 if (wakeup_fd >= 0) {
 402                                         sigset_t sigpipe_set, pending_set, old_set;
 403                                         int ret, sigpipe_was_pending = 0;
 404
 405                                         /*
 406                                          * Wake-up the other end by
 407                                          * writing a null byte in the
 408                                          * pipe (non-blocking).
 409                                          * Important note: Because
 410                                          * writing into the pipe is
 411                                          * non-blocking (and therefore
 412                                          * we allow dropping wakeup
 413                                          * data, as long as there is
 414                                          * wakeup data present in the
 415                                          * pipe buffer to wake up the
 416                                          * consumer), the consumer
 417                                          * should perform the following
 418                                          * sequence for waiting:
 419                                          * 1) empty the pipe (reads).
 420                                          * 2) check if there is data in
 421                                          *    the buffer.
 422                                          * 3) wait on the pipe (poll).
 423                                          *
 424                                          * Discard the SIGPIPE from write(), not
 425                                          * disturbing any SIGPIPE that might be
 426                                          * already pending. If a bogus SIGPIPE
 427                                          * is sent to the entire process
 428                                          * concurrently by a malicious user, it
 429                                          * may be simply discarded.
 430                                          */
 431                                         ret = sigemptyset(&pending_set);
 432                                         assert(!ret);
 433                                         /*
 434                                          * sigpending returns the mask
 435                                          * of signals that are _both_
 436                                          * blocked for the thread _and_
 437                                          * pending for either the thread
 438                                          * or the entire process.
 439                                          */
 440                                         ret = sigpending(&pending_set);
 441                                         assert(!ret);
 442                                         sigpipe_was_pending = sigismember(&pending_set, SIGPIPE);
 443                                         /*
 444                                          * If sigpipe was pending, it
 445                                          * means it was already blocked,
 446                                          * so no need to block it.
 447                                          */
 448                                         if (!sigpipe_was_pending) {
 449                                                 ret = sigemptyset(&sigpipe_set);
 450                                                 assert(!ret);
 451                                                 ret = sigaddset(&sigpipe_set, SIGPIPE);
 452                                                 assert(!ret);
 453                                                 ret = pthread_sigmask(SIG_BLOCK, &sigpipe_set, &old_set);
 454                                                 assert(!ret);
 455                                         }
 456                                         do {
 457                                                 ret = write(wakeup_fd, "", 1);
 458                                         } while (ret == -1L && errno == EINTR);
 459                                         if (ret == -1L && errno == EPIPE && !sigpipe_was_pending) {
 460                                                 struct timespec timeout = { 0, 0 };
 461                                                 do {
 462                                                         ret = sigtimedwait(&sigpipe_set, NULL,
 463                                                                 &timeout);
 464                                                 } while (ret == -1L && errno == EINTR);
 465                                         }
 466                                         if (!sigpipe_was_pending) {
 467                                                 ret = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
 468                                                 assert(!ret);
 469                                         }
 470                                 }
 471                         }
 472                 }
 473         }
 474 }
 475
 476 /*
 477  * lib_ring_buffer_write_commit_counter
 478  *
 479  * For flight recording. must be called after commit.
 480  * This function increments the subbuffer's commit_seq counter each time the
 481  * commit count reaches back the reserve offset (modulo subbuffer size). It is
 482  * useful for crash dump.
 483  */
 484 static inline
 485 void lib_ring_buffer_write_commit_counter(const struct lttng_ust_lib_ring_buffer_config *config,
 486                                           struct lttng_ust_lib_ring_buffer *buf,
 487                                           struct channel *chan,
 488                                           unsigned long idx,
 489                                           unsigned long buf_offset,
 490                                           unsigned long commit_count,
 491                                           size_t slot_size,
 492                                           struct lttng_ust_shm_handle *handle)
 493 {
 494         unsigned long offset, commit_seq_old;
 495
 496         if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
 497                 return;
 498
 499         offset = buf_offset + slot_size;
 500
 501         /*
 502          * subbuf_offset includes commit_count_mask. We can simply
 503          * compare the offsets within the subbuffer without caring about
 504          * buffer full/empty mismatch because offset is never zero here
 505          * (subbuffer header and record headers have non-zero length).
 506          */
 507         if (caa_unlikely(subbuf_offset(offset - commit_count, chan)))
 508                 return;
 509
 510         commit_seq_old = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->seq);
 511         while ((long) (commit_seq_old - commit_count) < 0)
 512                 commit_seq_old = v_cmpxchg(config, &shmp_index(handle, buf->commit_hot, idx)->seq,
 513                                            commit_seq_old, commit_count);
 514 }
 515
 516 extern int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf,
 517                                   struct channel_backend *chanb, int cpu,
 518                                   struct lttng_ust_shm_handle *handle,
 519                                   struct shm_object *shmobj);
 520 extern void lib_ring_buffer_free(struct lttng_ust_lib_ring_buffer *buf,
 521                                  struct lttng_ust_shm_handle *handle);
 522
 523 /* Keep track of trap nesting inside ring buffer code */
 524 extern DECLARE_URCU_TLS(unsigned int, lib_ring_buffer_nesting);
 525
 526 #endif /* _LTTNG_RING_BUFFER_FRONTEND_INTERNAL_H */